A narrative review of imaging intratumor heterogeneity in non-small cell lung cancer: current advances

Introduction

Lung cancer is the leading cause of cancer-related deaths globally, accounting for 18.7% of all cancer deaths, and is also the most frequently diagnosed cancer (12.4%). Non-small cell lung cancer (NSCLC) constitutes approximately 85% of all lung cancer cases. Because of its asymptomatic nature in early stages, NSCLC is often diagnosed at an advanced stage (1). The discovery of oncogene addiction has laid the theoretical foundation for molecular targeted therapy, marking the advent of personalized treatment for advanced NSCLC. Current treatment strategies for advanced NSCLC involve combinations of chemotherapy, radiotherapy, targeted therapy, and immunotherapy. However, tumor heterogeneity remains a major obstacle to precision medicine in NSCLC (2). Tumor heterogeneity encompasses differences at multiple levels—molecular, cellular, morphological, and functional—either within the same tumor or among different tumors. Most tumors follow a Darwinian evolutionary trajectory, during which cancer cells acquire genetic and phenotypic alterations through repeated division and proliferation, ultimately leading to self-renewal, invasion, and metastasis. During this process, daughter cells gradually diverge into distinct genetic or phenotypic clones, exhibiting heterogeneity in growth rate, invasiveness, drug sensitivity, and clinical outcomes. Clonal mutations, detectable in all cancer cells, are thought to occur early in tumorigenesis, whereas subclonal mutations arise later and are confined to specific subpopulations of tumor cells. Hence, genetic heterogeneity exists within a single tumor, and a higher number or proportion of subclonal mutations is generally indicative of increased tumor heterogeneity. Tumor heterogeneity is classified into intertumor heterogeneity and intratumor heterogeneity (ITH). The former refers to genetically and phenotypically different tumor foci in different patients or at different sites in the same patient, and the latter refers to genetically and phenotypically different between different cell clones within the same tumor. Notably, ITH often manifests as variable sensitivity of subpopulations to treatment, directly contributing to therapy resistance and failure, posing a major challenge in lung cancer management (3-5). Currently, ITH research has become a global focus, with molecular analysis techniques such as next-generation sequencing (NGS) advancing its investigation. While histopathological and molecular analyses remain the gold standard for evaluating ITH, they require invasive procedures such as surgical resection or needle biopsy, which carry risks such as infection. In addition to being invasive, these methods are susceptible to sampling bias and may fail to represent the overall tumor landscape (6,7). Consequently, numerous studies suggest that non-invasive imaging methods can quantitatively assess ITH using specific imaging parameters, aiding clinical decision-making and prognostic evaluation in the context of precision medicine (8-10). This review is presented as a narrative review that focuses on imaging studies of ITH in NSCLC, aiming to provide an up-to-date overview of its clinical relevance and potential applications in diagnosis, treatment response prediction, and prognosis. We present this article in accordance with the Narrative Review reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-506/rc).

Methods

To comprehensively capture the current techniques, applications, and challenges related to imaging ITH (IITH), we conducted a systematic literature search across major databases, including PubMed, Web of Science, EMBASE, and the Cochrane Library. The search scope covers publications up to June 1, 2025 to ensure extensive and timely coverage of relevant studies. Keywords included “intratumor heterogeneity”, “non-small cell lung cancer”, “habitat imaging”, “texture analysis”, “radiogenomics”, and related terms. Two authors independently screened the titles and abstracts, followed by full-text assessments for eligible studies. Only English-language articles were included. Any disagreements were resolved by a third reviewer. All authors then reviewed the included articles and discussed the findings to reach a consensus. The search strategy is summarized in Table 1. The detailed search formula using PubMed as an example is shown in Table S1.

Spatial and temporal heterogeneity of NSCLC

ITH, a fundamental biological characteristic of malignancies, displays complex spatiotemporal dynamics in NSCLC. Recent multiregional genomic sequencing of specific lung cancer subtypes has revealed marked genetic variability. These variations correlate with tumor location and stage, and differ across tumor regions, a phenomenon known as spatial heterogeneity (9,11). Furthermore, the characteristics of a single tumor lesion may undergo substantial temporal changes driven by genetic factors, a concept defined as temporal heterogeneity (4). The formation of ITH is primarily driven by three factors: genetic mutations, the tumor microenvironment (TME), and metabolic processes. Among them, genetic mutations underlie both temporal and spatial heterogeneity within the TME and metabolic networks of lung cancer (12). Therefore, a comprehensive understanding of ITH—from the genome, transcriptome, and proteome to macroscopic metabolic and immune features—is crucial for elucidating its mechanisms and guiding biomarker discovery in imaging histology. This section summarizes recent advances in the study of both spatial and temporal heterogeneity in lung cancer, providing essential insights for imaging-based ITH research. The brief overview of ITH is presented in Figure 1.

Figure 1 The brief overview of ITH. The ITH of non-small cell lung cancer is divided into spatial heterogeneity (blue part) and temporal heterogeneity (red part). The survival plot, which is illustrative and not based on real data, in the figure indicates that the level of ITH affects patients’ survival rate. In the figure, the x-axis and y-axis represent spatial dimensions, reflecting the differences in tumor cell characteristics in different physical spaces within the tumor. The cells of different colors and sizes represent the heterogeneity differences in both time and space within the tumor. The formation and development of ITH involve multiple omics aspects, such as genomics, transcriptomics, proteomics, epigenomics, and so on. ITH, intratumor heterogeneity; NSCLC, non-small cell lung cancer.

Spatial heterogeneity of NSCLC

The mutations and molecular phenotypes accumulated by cancer cells often vary from site to site due to differences in the TME or local specific stress. Studies have shown that local oxygen partial pressure gradients [e.g., differences in hypoxia-inducible factor-1α (HIF-1α) expression) and cytokine secretion profiles (typified by transforming growth factor β 1 (TGF-β1)] can significantly affect the rate of mutation accumulation (13). This microenvironmental pressure gradient leads to regional differential expression of immune checkpoint molecules such as programmed death-ligand 1 (PD-L1), which in turn creates a dynamic evolutionary selection pressure field, causing different intra-tumoral subpopulations to accumulate mutations at different rates and in different forms, and adapting to different microenvironments with different molecular phenotypes, which is manifested as spatial heterogeneity of tumors (12,14).

NSCLC genetic spatial heterogeneity

In the course of lung cancer development, the prognosis of patients with mutations in different regions of the driver genes within the tumor that show a high level of spatial ITH is usually not optimistic (15). For NSCLC, there is a difference between clonal and subclonal mutations in driver gene mutations. Driver mutations, represented by the epidermal growth factor receptor (EGFR), BRAF and MET genes, tend to be clonal and occur at an early stage of disease development. These clonal driver mutations are capable of affecting multiple sites of the disease and are often the direct target of action for targeted therapies. These mutations show diverse distribution patterns in different regions of the tumor. For example, EGFR mutations may be highly enriched in certain regions of the tumor, while in other regions they may be wild-type or have other types of mutations. In contrast, subclonal driver mutations occurring in genes such as phosphatidylinositol-4,5-bisphosphate 3-kinase catalytic subunit alpha (PIK3CA), neurofibromin 1 (NF1), Kirsten rat sarcoma viral oncogene homolog (KRAS), tumor protein 53 (TP53), and members of the notch Homolog (NOTCH) family are mostly present in a subclonal form. This means that these subclonal mutations are only present in some regions of the tumor and not in others. It is this difference in mutations in different regions within the tumor that makes the tumor response to targeted therapies show marked inconsistency. Tumor cells in some regions may be more sensitive to treatment, while tumor cells in other regions may exhibit resistance (16,17). This phenomenon fully reflects the significant differences in the spatial distribution of different driver mutations within the tumor. To address the many challenges of NSCLC due to spatial ITH, several approaches have been used to reveal the regional distribution of driver mutations. For example, multiregional whole-exome sequencing enables in-depth profiling of genetic variation information in different regions of the tumor (18,19). And liquid biopsy techniques, such as the detection of cell free DNA (cfDNA), circulating tumor DNA (ctDNA), and circulating tumor cells (CTCs) (20-22), also provide an important insight into the genetic characteristics of a tumor. These methods play a crucial role in helping physicians select appropriate and precise treatment options, monitor and accurately assess early lung tumor recurrence in real time, determine tumor drug resistance, and detect microscopic residual disease.

Spatial heterogeneity of the NSCLC microenvironment

The spatial heterogeneity of the TME in lung cancer is mainly manifested by three-dimensional (3D) differences in the immune landscape, stromal composition and vascular network (23). According to the characteristics of immune cell infiltration, tumor immune microenvironment (TIME) can be categorized into three phenotypes: “hot” (abundant infiltration of cytotoxic T lymphocytes, CTLs), “cold” (lack of T cells), and “excluded” (peripheral aggregation of T cells), and the differences in their distributions have a direct impact on the efficacy of immune checkpoint inhibitors (24). In addition to cytotoxic T-lymphocytes, a large number of immune cell subtypes show significant differences in regional numbers and functional distribution. Single-cell analysis revealed significant variability in the suppressive function of regulatory T cells Tregs in different regions, and this heterogeneity was mainly driven by factors such as TGF-β secreted by cancer-associated fibroblasts (CAFs). In addition, immunosuppressive cells such as tumor-associated macrophages (TAMs), dendritic cells (DCs), and myeloid-derived suppressor cells (MDSCs) showed obvious regional enrichment features and together with the gradient expression of immune checkpoints such as PD-L1/cytotoxic T lymphocyte antigen-4 (CTLA-4), they constituted an immune escape network (25). In terms of stromal heterogeneity, CAFs form a pro-metastatic microenvironment through secretion of vascular endothelial growth factor (VEGF) and remodeling of extracellular matrix (ECM) components, whereas tumor-derived exosomes (e.g., miR-210/CD81⁺) contribute to malignant progression through the epithelial-mesenchymal transition (EMT) and Wnt-PCP pathways (26). This multidimensional heterogeneous network provides new targets for the development of space-specific therapeutic strategies.

NSCLC metabolic spatial heterogeneity

Intra-tumor spatial heterogeneity in lung cancer is mainly manifested at the metabolic level by significant differences in glucose metabolism, lipid metabolism, and key signaling pathways (27) in different regions. For glucose metabolism, some regions exhibit glycolytic metabolism dominated by the Warburg effect, whereas other regions rely on oxidative phosphorylation, and this metabolic heterogeneity promotes tumor proliferation and drug resistance through the phosphatidylinositol-3-kinase/protein kinase B/mammalian target of rapamycin (PI3K/AKT/mTOR) pathway. With respect to lipid metabolism, regions of active endogenous fatty acid synthesis and regions of dominant exogenous lipid uptake coexist, and their metabolites, such as lysophosphatidic acid (LPA) and sphingosine-1-phosphate (S1P) affect tumor progression by regulating immune cell recruitment and neovascularization. Notably, regional activation differences in PI3K/AKT/mTOR and mitogen-activated protein kinases (MAPK) signaling pathways further reinforce this metabolic heterogeneity, providing a theoretical basis for the development of metabolically targeted precision therapeutic strategies (28).

Temporal heterogeneity of NSCLC

Tumor cells undergo significant biological changes during disease progression or therapeutic interventions through mechanisms such as gene mutation, clonal proliferation, and adaptive evolution. These changes cause tumors to exhibit distinct biological characteristics at various stages of disease development, and this dynamic difference based on the time dimension is defined as the temporal heterogeneity of tumors (3).

NSCLC genetic temporal heterogeneity

Disease recurrence after treatment constitutes a major clinical prognostic challenge in the disease process of NSCLC. In essence, this phenomenon stems from the dynamic evolutionary process of tumor cells under the pressure of therapeutic selection. Specifically, subclonal populations of specific driver genes gradually adapt to microenvironmental alterations through acquired genetic variants (e.g., EGFR T790M mutation), resulting in a treatment-resistant phenotype (29). In addition, studies have shown (30) that lung cancer patients treated with EGFR tyrosine kinase inhibitors (TKIs) after chemotherapy have significantly lower response rates than pre-treatment levels, as well as a significant decrease in EGFR gene mutation rates. Not only that, patients’ overall survival (OS) rate, median survival time, and time to disease evolution after EGFR TKIs treatment were all substantially shorter, fully reflecting the significant differences in tumor characteristics in the time dimension. In summary, the temporal heterogeneity of lung cancer is an important cause of a series of poor prognoses and high mortality, which has a great negative impact on patient survival.

Temporal heterogeneity of the NSCLC microenvironment

The TME of NSCLC is characterized by significant dynamic heterogeneity, mainly in the temporal evolution of immune and stromal components (31). Single-cell analysis revealed that immune cell subsets [including tumor infiltrating lymphocytes (TILs), DCs and TAMs] in the TME exhibit complex phenotypic diversity, e.g., CD8⁺ T cells exist in a continuum of functional states from effector to depletion, whereas TAMs exhibit a multidimensional activation profile that transcends traditional M1/M2 categorization. This dynamic change in the immune landscape together with CAFs-mediated ECM remodeling constitutes a synergistic network that promotes tumor progression (32,33). Notably, hypoxia-induced exosomes (e.g., miR-21+ exosomes) can elevate chemotherapy resistance by 2-3-fold through epigenetic regulation (34), while spatiotemporal specific activation of the kallikreins protease-activated receptors (KLKs-PARs) signaling axis further exacerbates microenvironmental heterogeneity (35). These findings provide a molecular basis for the development of temporal-specific therapeutic strategies.

NSCLC metabolic temporal heterogeneity

The dynamic evolution of metabolic heterogeneity characterizing lung cancer significantly increases therapeutic difficulty. This metabolic heterogeneity is not only reflected in the spatial and temporal differences in the underlying metabolic pathways, but also reinforced by a complex epigenetic regulatory network. In addition to the aforementioned mechanisms of metabolic heterogeneity, heterogeneous expression of key metabolic enzymes induces changes in DNA methylation patterns, while epigenetic mechanisms, such as chromatin remodeling, histone modification changes, and non-coding RNA regulation collectively shape the dynamic features of metabolic heterogeneity. Notably, this metabolic-epigenetic interaction also affects the immunogenicity of tumor cells, facilitating immune escape by modulating the expression of tumor-associated antigens and antigen presentation processes (27). An in-depth understanding of the dynamic features of this metabolic heterogeneity is of great clinical significance for optimizing therapeutic strategies and improving patient prognosis.

Imaging analysis methods for IITH

In cancer imaging, traditional imaging only focuses on the shape, edge, blood supply, density or signal of the tumor lesion, which makes it difficult to distinguish the internal heterogeneity of the tumor. In recent years, with the continuous development of new technologies, medical imaging has evolved from structural imaging to a new stage of functional imaging and even molecular imaging, which has made real-time evaluation of ITH by imaging possible. Quantitative features of tumors calculated from images of various modalities, such as computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) can not only provide spatial information, but also repeatable detection by analyzing and evaluating their biological features, such as tissue structure, blood flow, and metabolism. Therefore, the use of imaging-related images to analyze specific parameters can quantify the degree of tumor heterogeneity and provide guidance for the development of therapeutic options and evaluation of prognosis (8). Currently, there are two main categories of methods for quantifying IITH in lung cancer: one is the construction of radiomic features, which is used as a clinical indicator by identifying single or a combination of computed features, including intensity-based histograms, shape- and volume-based features, and measures, such as texture analysis. Another is habitat imaging, where quantitative imaging markers are used to cluster clusters of tumor cells with similar representations to locate high-risk subregions within a tumor. Since CT and PET/CT dominate lung cancer staging and follow-up, the imaging methods for IITH described in this review primarily focus on these two modalities (36). In summary, both CT and PET/CT play key roles in the quantitative imaging assessment of IITH in NSCLC, with complementary strengths across multiple analysis methods (8). CT-based radiomics mainly focuses on high-resolution structural and perfusion information, enabling detailed texture analysis, shape- and volume-based feature extraction, and habitat imaging using contrast-enhanced CT (CECT) to delineate subregions with distinct vascular or perfusion characteristics. PET/CT, by integrating metabolic data, is particularly valuable for quantifying functional heterogeneity through parameters, such as standardized uptake value (SUV), metabolic tumor volume (MTV), total lesion glycolysis (TLG), cumulative SUV-volume histograms (CSH), and for habitat analysis that identifies metabolic subregions within tumors. Notably, texture analysis, shape- and volume-based features, and habitat imaging can all be derived from both CT and PET/CT, providing complementary perspectives on tumor structure, metabolism, and microenvironmental heterogeneity. By combining the anatomical detail and functional mapping of CT with the metabolic insights from PET, clinicians and researchers can achieve a more comprehensive and non-invasive characterization of IITH, which is crucial for personalized treatment strategies and prognostic prediction. This section will provide an overview of the mathematical basis of common texture analysis methods, histogram methods, shape- and volume-based feature analysis methods and habitat imaging. A comparison of these methods is shown in Table 2.

Histogram method

18F-fluorodeoxyglucose (18F-FDG) is the most commonly used PET tracer in clinical practice. FDG uptake in tumors is related to glucose metabolism, and PET images can show non-uniformity of FDG uptake, which can be used to quantitatively characterize tumor heterogeneity. SUV is one of the most commonly used semi-quantitative indices for tumor diagnosis in PET images, and CSH converts complex 3D functional image grayscale information into SUV distribution curve for the region of interest (ROI). The abscissa represents the percentage of the maximum (SUV_max), while the ordinate represents the percentage of tumor volume greater than or equal to the corresponding SUV threshold, reflecting the distribution of SUV within the ROI. CSH can objectively reflect the degree of tumor heterogeneity. The lower the AUC value, the more dispersed the FDG distribution within the tumor, indicating higher heterogeneity (37,38).

Texture analysis method

Compared to histograms, texture analysis combines spatial information on top of grayscale analysis for greater robustness. Texture is a large number of very regular or weakly similar elements or graphical structures in an image, generally understood as spatial variations and repetitions of image gray levels, or recurring local patterns (texture units) and their rules of arrangement in an image (39). Texture analysis can be defined as the quantification of the spatial distribution of voxel gray levels. There are three common approaches to texture analysis: statistical analysis, modeling analysis and transformation analysis (8). Statistical-based methods group texture parameters according to increasing complexity by investigating the spatial correlation properties of gray levels, and mainly describe third-order statistical metrics. First-order statistical metrics describe the global characteristics of the ROI, which can be obtained directly from the histogram pixel intensities, such as mean, standard deviation, skewness, kurtosis, image energy, entropy, median, minimum, maximum, percentile, coefficient of variation (COV), and so on. First-order statistics have fast computing time and low requirements on hardware and software because of the simple formula and small computation volume. However, first-order statistics do not take into account the spatial interrelationships between gray values, which means that we cannot rely on first-order statistics alone to make clinical analysis. Second-order statistics are obtained by gray level co-occurrence matrix (GLCM), run-length matrix (RLM), which is the most common, widely used and effective method to describe texture by studying the spatial correlation of gray level. GLCM describes texture by studying the spatial correlation properties of gray scale, and is the most common and widely used texture analysis method with the best results, whose parameters include contrast, entropy, homogeneity, correlation, etc. RLM analyzes the texture mainly in a specific direction, and is a second-order statistic of the gray scale relationship of the image. Higher-order statistics are characterized by Neighborhood Greytone (intensity) Difference Matrices (NGTDMS) describing the local features, mainly checking the spatial relationship of ≥3 voxels. However, the computational complexity of higher order statistics is high, which is more suitable for texture analysis of linear structures (40). Model-based methods apply complex mathematical models (fractal model, autoregressive model, etc.) to identify texture features, obtain structural information by extracting the internal spatial arrangement of the texture, so as to construct a corresponding spatial structure model for matching, and find out the spatial structure of the arrangement of the primitives and the periodicity law, etc. Fractal analysis is a kind of pattern or geometric recognition. The fractal dimension, as an important feature and metric of fractal analysis, mainly measures the irregularity and roughness of the image. The larger the fractal dimension, the rougher the texture. This method combines the spatial information and gray scale information of the image in a simple and organic way, which can take into account the local randomness of the texture and the overall regularity, and has great flexibility, so it has attracted much attention in image processing, but the difficulty in solving the coefficients of the fractal dimension model has hindered its application (41). Transform-based methods are used to analyze texture properties in images in different spaces, such as different frequency spaces or scale spaces. For example, wavelet transform, Fourier transform and Gabor transform are more commonly used in transform analysis methods. This class can transform the entire image so that statistical features can be extracted using multi-scale methods. Among them wavelet transform is most widely used in transform analysis methods (39).

Shape- and volume-based feature analysis method

Shape and volume features are quantitative descriptors used to capture the geometric properties and spatial configuration of tumors (8). These features are independent of pixel or voxel intensity values and provide essential morphological information for evaluating ITH. Commonly used features include tumor volume, surface area, maximum 3D diameter, and bounding box dimensions, which reflect the overall size and spatial extent of the lesion. More refined descriptors—such as sphericity (the degree to which the tumor resembles a sphere), compactness (the efficiency of volume enclosure relative to surface area), elongation (the ratio of major to minor axis length), and flatness (the degree of compression along one axis)—offer insights into the irregularity and asymmetry of tumor growth. Variations in these parameters within or between tumor subregions may indicate underlying biological heterogeneity, such as differences in proliferative activity or invasive potential. In addition, volume-based metabolic parameters, such as MTV and TLG, often derived from PET imaging, are widely utilized. MTV quantifies the volume of metabolically active tumor tissue above a predefined SUV threshold, while TLG reflects the total glycolytic activity by integrating both tumor volume and metabolic intensity. These markers have demonstrated clinical relevance in risk stratification, prognostic assessment, and treatment response monitoring (42). When combined with texture or functional features, shape- and volume-based descriptors enhance the multidimensional characterization of tumor heterogeneity in medical imaging.

Habitat imaging

In past imaging histology studies, it was usually assumed that the internal heterogeneity of the tumor was well-mixed and that quantitative features were generated equally throughout the tumor. This approach ignores the spatial heterogeneity of tumors on imaging, so some researchers have focused on subregions within the tumor, i.e., tumor subzones. By delineating the subregions, a “habitat map” of the tumor is established to study the interrelationship between the quantitative features and the TME, and to non-invasively capture the microscale information hidden in medical imaging. Habitat analysis is an imaging histology technique to cluster tumor cells with similar biological characteristics into voxels in multimodal images based on differences in histopathology, hemodynamics, and molecular biology. It can focus on the internal spatial distribution of tumors and reflect the evolution of tumor cells over time under the action of microenvironment, and can effectively display and quantify ITH to characterize TME, thus providing a basis for validating the biological findings behind the heterogeneity of the images (43). Habitat imaging, firstly, it is necessary to select valuable quantitative imaging markers, cluster tumor cell clusters with similar characterization through the quantitative imaging markers, and divide them into multiple habitat subregions. Imaging markers are derived from imaging techniques, such as MRI, PET/CT (44,45). Non-quantitative structural MRI (sMRI) is able to show anatomical morphology, but its signal strength is affected by the magnetic field strength, scanning equipment and image acquisition parameters, and the signal strength error is difficult to completely eliminate or correct by the normalization step (46). Functional MRI (fMRI) and PET/CT can further visualize ITH by measuring quantitative parameters, such as apparent diffusion coefficient (ADC), cerebral blood volume (CBV), volumetric transfer constant (Ktrans), SUV, etc., which quantify the biological status of tumor diffusion, perfusion, hypoxia, and acidification. In addition, the difference/relative value between quantitative parameters can also characterize the dynamic process of ITH and can be clustered as potential quantitative imaging markers (47). Clustering is the process of partitioning a dataset into clusters according to a specific criterion (e.g., a fixed distance), such that the similarity of data within clusters and the difference of data in non-identical clusters are as large as possible. There are 2 main approaches to clustering: one is the Otsu thresholding method based on scientific imaging, which determines the threshold based on the distribution of pixel intensity values to maximize the differences between pixel classes (48); the other is the classical clustering methods, including Gaussian mixture model, hierarchical clustering, and K-means (49). Gaussian mixture model operates under the probabilistic framework that data derive from a linear combination of K Gaussian distributions, each characterized by mean (µ), covariance (Σ), and mixing coefficient (π). It employs the expectation-maximization (EM) algorithm to iteratively optimize parameters: the E-step estimates posterior probabilities of data belonging to each component, while the M-step refines parameters to maximize likelihood—this mechanism enables Gaussian mixture models to perform clustering by identifying Gaussian distributions in the data distribution and perform hard or soft clustering of data with different Gaussian distributions, particularly for overlapping tumor subregions where soft clustering captures probabilistic memberships Hierarchical clustering constructs a dendrogram by merging similar clusters (agglomerative) or splitting dissimilar ones (divisive), relying on distance metrics (e.g., Euclidean) to define proximity. This method preserves hierarchical relationships, making it suitable for individual-level data, and hierarchical clustering groups data points together based on local proximity and is only applicable to individual level data. K-means is the most widely used self-supervised algorithm, which iteratively assigns data points to a clustering “center of mass” (i.e., the mean of a set of neighboring data points), and continually updates the center of mass to minimize the squared distance between the data points and the corresponding centers of mass to which they are assigned. The center of mass is continuously updated to minimize the sum of the squared distances between each data point and its assigned corresponding center of mass, thus clustering the data. However, in tumors, too many clusters can reduce the habitat volume, thus limiting the possibility of cross-validating the habitat by tumor sampling at a later stage, so the optimal number of clusters can be specified in advance. The optimal number of clusters can be determined by specifying the degree of aberration by plotting an “elbow diagram” (50,51) or by averaging silhouettes and plotting a “tree diagram” (52,53). The workflow of habitat analysis is shown in Figure 2.

Figure 2 The workflow of habitat analysis. This figure illustrates the step-by-step process of habitat analysis, including image acquisition, delineation of the VOI, tumor subregion segmentation, and the splitting of subregions. It shows the extraction of various features such as texture, histogram, and wavelet features, followed by feature selection. The selected features are then used to construct and validate models using machine learning algorithms. KNN, K-nearest neighbors; LR, logistic regression; RF, random forest; SVM, support vector machine; VOI, volume of interest; XGB, XGBoost.

Clinical application of IITH in NSCLC

The evaluation of IITH in NSCLC is reflected in various aspects, including classification of lung cancer lesions and thoracic lymph nodes, classification of lung cancer subtypes, assessment of efficacy and prognosis, and determination of histopathologic and molecular features. Since CT or FDG PET plays a dominant role in lung cancer diagnosis and follow-up, the IITHs described in this section are all for CT or FDG PET. Representative literature is summarized in Tables 3,4.

Temporal IITH of NSCLC

Efficacy and prognostic assessment

Tumor stage, size, lymph node involvement, and histological grading are all important factors influencing the prognosis of lung cancer. Previous studies have suggested that methods for quantitative characterization of ITH using FDG PET can be conceptually divided into two categories: variability of voxel intensities within the tumor volume, and descriptors of irregularities in the shape of the tumor. Morphological features can complement texture and histogram analysis by quantifying spatial irregularity of the tumor (69). In terms of tumor-node-metastasis (TNM) staging, Apostolova et al. (70) retrospectively analyzed pre-treatment FDG PET quantitative morphologic features in 83 NSCLC patients and showed that the SUV_max, MTV, and the degree of non-sphericity (ASP) correlated with pathological T-stage and N-stage, whereas the degree of ASP also correlated with the M-stage and served as an independent predictor of progression-free survival (PFS). Karacavus et al. (54) explored the association between preoperative FDG PET textural features and tumor histopathological features in 67 cases of NSCLC. It was found that one texture feature obtained using the gray-level run length matrix method [gray level inhomogeneity (GLN)] and nine texture features obtained using the Laws method successfully differentiated all tumor stages. By incorporating the selected five features, the k-nearest neighbors (k-NN) classifier (k =3) and the support vector machine (SVM) had an accuracy of approximately 84% in the automated staging of tumors. In terms of T-staging, Li et al. (71) retrospectively evaluated 234 patients with NSCLC who underwent F-FDG PET/CT prior to treatment, targeting the primary foci to calculate the metabolic heterogeneity factor (MHF) derived from a volume threshold function ranging from 40% to 90%. Trend analysis showed a strong positive correlation between MHF and T stage (P<0.05). In addition, achieving reproducible and high-precision imaging histology analysis requires prudent feature extraction strategies, which are particularly important for tumor characterization in multilayer CT images. Currently, there exist two paradigms for ROI definition: a 3D volume covering all relevant layers (3D-ROI), and a two-dimensional monolayer selecting only the largest cross-sectional area (2D-ROI). 2D-ROI, although it has the advantages of easy operation and high computational efficiency, may lose the spatial heterogeneity information of tumors, while 3D-ROI theoretically more comprehensively reflects the 3D distribution of tumor cell population characteristics. It is also worth noting that most radiomics studies focus on feature mining in the intra-tumor region, often neglecting the key biological information that may be contained in the peritumor tissue. Zhang et al. (55) developed independent and combined radiomics models based on 3D, 2D, and peritumoral regions, incorporating features of shape, histogram, and texture in 508 NSCLC patients. The results showed that through five-fold cross-validation, the combined model achieved an average area under the curve (AUC) of 0.839 [95% confidence interval (CI): 0.829–0.849] and an accuracy of 71% for predicting pathological T-stage in the independent test set. Incorporating information about the peri-tumoral region improved the accuracy of the prediction tasks, but it was found that the 2D and 3D imaging histology features had comparable performance and discriminative power in these tasks. In terms of N staging, mediastinal lymph node metastasis is a key factor affecting the prognosis of NSCLC patients without distant metastases, and its precise staging directly determines the development of clinical treatment strategies. Currently, CT examination is routinely used for staging assessment by measuring the maximum shortest diameter of lymph nodes, but there are significant limitations in the diagnostic efficacy of this method: studies have shown that 28% of the enlarged lymph nodes diagnosed by CT were pathologically confirmed to be false-positive, and 20% of the metastatic lymph nodes did not show abnormal dimensions on CT (72). 18F-FDG PET/CT, as a molecular imaging technique that integrates morphological and functional information, has demonstrated significant value in the staging of NSCLC and effectiveness in its treatment. Systematic evaluation data showed that its efficiency in identifying mediastinal lymph node metastasis reached a sensitivity of 77% and a specificity of 86% (73). However, the technique still faces specificity challenges-benign lesions, such as inflammatory reactions or lymphoid tissue hyperplasia may lead to increased FDG uptake, producing false-positive results (74). This overlap of biological characteristics makes accurate identification of mediastinal lymph node metastasis a difficult problem in clinical practice. Therefore, studies have been conducted to predict lymph node metastasis by imaging histologic resolution of ITH. Existing studies have demonstrated that primary tumor texture analysis [≥2 high-risk CT features predicting metastasis with a risk ratio (RR) of 11.00, 95% CI: 1.560–77.800, P<0.001] (75) and PET/CT gradient variance [odds ratio (OR) 0.93, 95% CI: 0.870–1.000, P=0.03] (76) show particular promise in predicting mediastinal lymph node metastasis. Multiparametric models combining mediastinal lymph node SUV ratio, maximum short-axis diameter (D_min), and CoV exhibit superior performance (AUC up to 0.907) (56). Furthermore, a scoring system incorporating semiquantitative parameters—including lymph node size, SUV_max, MTV, and TLG—demonstrated through visual assessment that a score of ≥3 was associated with a 14.3-fold increased risk of metastasis (P<0.001) (77). Notably, metabolic parameters, such as MTV and heterogeneity factors (derived from volume threshold functions of 40–80% SUV_max) also provide histology-specific predictive value for occult nodal disease (78). For M staging, Hakkak Moghadam Torbati et al. (57) developed a machine learning model using 18F-FDG PET/CT texture features from primary tumors and metastatic lymph nodes in 79 advanced NSCLC patients (104 lesions in training, 78 in testing). The optimal model combined shape (sphericity, compactness) and metabolic (TLG, MTV) features with texture parameters (gray level run length matrix_run length non-uniformity, GLRLM_RLNU), achieving moderate predictive performance (accuracy 74.40%, AUC 0.630) in the test cohort using decision tree analysis.

In addition, texture analysis can assess the relationship between ITH in NSCLC and its survival outcomes. Win et al. (79) found in a prospective study that for patients undergoing FDG PET/CT combined with dynamic contrast-enhanced (DCE) CT, tumor texture heterogeneity in the CT part of PET/CT (P=0.02) and DCE CT permeability (P<0.001) were independent survival predictors. Andersen et al. (80) reported that normalized uniformity [hazard ratio (HR) 16.06, 95% CI: 3.861–66.788, P<0.001] and skewness (HR 1.914, 95% CI: 1.330–2.754, P<0.001) of texture parameters derived from CECT were prognostic factors for OS. Ohri et al. (81) analyzed the MTV, SUV_max, and texture features from pretreatment 18F-FDG PET in 201 patients with locally advanced NSCLC. Patients were divided into three categories based on median survival time: low tumor MTV group (median OS 22.6 months), high tumor MTV with high SUM_mean group (median OS 20.0 months), and high tumor MTV with low SUM_mean group (median OS 6.2 months; P<0.001). Fried et al. (82) demonstrated that incorporating quantitative PET/CT radiomic features (histogram, shape/volume, and co-occurrence matrix parameters) with conventional prognostic factors significantly improved OS prediction in stage III NSCLC compared to using clinical factors alone (concordance index 0.62 vs. 0.58, log-rank P<0.001 vs. 0.180). The model, developed through leave-one-out cross-validation in 195 patients, consistently selected tumor solidity and co-occurrence matrix energy as key predictive features across all validation folds (P=0.007 for improvement over clinical factors). Desseroit et al. (83) developed a nomogram integrating PET entropy, CT regional percentage, clinical stage, and functional volume, which improved stratification of stage II–III patients and identified those with the poorest prognosis (clinical stage III, large tumor volume, high PET heterogeneity, and low CT heterogeneity). In addition, Jimenez Londoño et al. (84) also explored the pattern of recurrence after treatment in patients with stage I–III NSCLC. The study found that in 173 patients, factors such as age, pathologic lymphovascular invasion and normalized SUV_max perimeter distance were considered risk factors for early recurrence. Adenocarcinoma histologic type was an independent variable for distant recurrence, while Patient age, number of mediastinal metastatic lymph nodes at staging, sphericity, normalized SUV_peak to centroid distance, entropy, low grayscale tour emphasis, and high grayscale tour emphasis were independent variables for multisite metastatic disease. Certain variables were associated with organ-specific recurrence.

It is worth noting that although some NSCLC patients are staged at an early stage, the risk of disease recurrence still exists. Taking stage I NSCLC patients as an example, according to relevant studies, as many as 20–30% of patients will experience recurrence after undergoing surgical treatment (1). In view of this grim situation, it is particularly critical to accurately identify prognostic factors that can effectively predict high-risk patients after surgical treatment. Nakajo et al. (58) suggested that preoperative 18F-FDG PET/CT-derived heterogeneity texture parameters, specifically intensity variability (HR 6.19, P=0.05), can independently predict postoperative progression. Although lobectomy serves as the current standard treatment for early-stage NSCLC, many patients are unable to undergo surgery due to the presence of comorbidities, especially those with chronic obstructive pulmonary disease (COPD). Stereotactic radiation therapy (SBRT) is an alternative treatment for these patients and has been shown to be effective in terms of OS and local control, with distant metastases being the main reason for its treatment failure. Early identification of patients at risk of recurrence is therefore of critical importance in guiding clinicians to take timely and early interventions to improve patient prognosis (85). Lovinfosse et al. (86) confirmed that the texture analysis parameter of baseline 18F-FDG PET/CT, dissimilarity, was a parameter associated with both the disease-free survival (HR =0.834, P<0.01) and disease-specific survival (HR =0.822, P=0.04) in NSCLC patients who received SBRT. Wu et al. (87) developed and validated a PET-based prognostic model for distant metastasis in early-stage NSCLC patients undergoing SBRT. Using LASSO regression analysis of 35 robust radiomic features (including texture, morphological and histogram parameters) in a discovery cohort (n=70), they identified an optimal model incorporating two heterogeneity-quantifying features and peak SUV. The model demonstrated superior performance in the independent validation cohort (n=31), with a concordance index of 0.71 compared to 0.67 for SUV_max alone, and effectively stratified patients into high- and low-risk groups (HR 4.80, P<0.05). Notably, predictive power improved further when combined with histologic type (HR 6.90, P=0.03, C-index 0.80). Franceschini et al. (59) developed and validated CT-based radiomic models for predicting lymph node recurrence and survival outcomes in early-stage NSCLC patients undergoing SBRT. Using a retrospective cohort of 102 patients (70 training, 32 validation), they constructed three distinct models through rigorous statistical approaches: (I) a linear discriminant analysis classifier for lymph node relapse risk demonstrated 85.40% accuracy (AUC 0.730 in validation); (II) Cox regression models for progression-free and disease-specific survival stratified patients into significant risk groups (high-risk: 44.60–52.70 vs. low-risk: 84.10–88.20 months survival). The models, based on 45 textural features reflecting tumor heterogeneity, employed elastic net regularization and backward selection to ensure robustness, with validation metrics showing strong concordance between training and test cohorts.

Radiotherapy and chemotherapy are the standard treatment modalities for advanced NSCLC; however, due to the highly heterogeneous nature of the patient population with advanced NSCLC, it is necessary to predict the treatment response and long-term prognosis of patients in the early stages of treatment, which is expected to improve tumor control and reduce side effects by stratifying patients with greater precision. Dong et al. (88) retrospectively analyzed the global and local scale texture features of pre-treatment 18F-FDG PET/CT in NSCLC patients receiving concurrent chemoradiotherapy (CCRT). The results showed that COV, AUC-CSH, contrast, and dissimilarity may have important values for predicting treatment response and survival in patients with locally advanced NSCLC. Cook et al. (89) found that texture features from the 18F-FDG PET/CT NGTDMs, including roughness, contrast, and busyness, were able to predict 12-week response evaluation criteria in solid tumors (RECIST) response in NSCLC patients receiving chemoradiotherapy. All 3 features were associated with PFS and localized PFS, whereas roughness was also associated with OS. CT-based radiomics also showed prognostic value, as Ahn et al. (90) reported that higher entropy, skewness and mean attenuation independently predicted worse 3-year survival in patients with NSCLC who underwent concomitant chemoradiotherapy. Fried et al. (91) further demonstrated that combining four-dimensional CT as well as CECT images texture features with clinical factors improved risk stratification for multiple endpoints, including locoregional control and distant metastasis-free survival in patients with stage III NSCLC who underwent radical chemoradiotherapy.

Histogram analysis provides a promising noninvasive test for ITH assessment, and its quantitative results are expected to serve as a prognostic reference for clinicians to develop individualized treatment plans. Virginia et al. (92) confirmed the clinical relevance of primary tumor entropy in 692 patients with advanced NSCLC using the filtered-histogram method, showing a significant association with survival (HR 1.30, 95% CI: 1.1–1.5, P=0.001). Subsequent studies validated the practicality of SUV-volume histogram metrics, with the cumulative SUV-volume histogram AUC being able to predict disease progression (93,94). Heterogeneity factors (HF) (95) (defined as the derivative of the volume threshold function between 20% and 80%) were found to identify high-risk recurrence in adenocarcinoma patients. However, most existing models use multiple parameters for feature extraction, making quantitative characterization difficult in a clinical setting. Therefore, obtaining simple and clinically applicable imaging parameters to describe ITH is crucial. The CoV [defined as the standard deviation divided by the mean SUV (SUV_mean)] can effectively stratify prognosis at different thresholds [patients with CoV ≤0.38 and total MTV (MTVTOT) >89.5 mL have the worst prognosis (96); lymph node CoV ≤0.29 and primary tumor CoV >0.38 have the best prognosis (97)], and it can also predict OS and PFS in centrally located lung cancer NSCLC patients. The novel H-index (98) (which quantifies metabolic heterogeneity by calculating the spatially weighted sum of SUV differences between voxels within a tumor region) outperforms traditional texture analysis, achieving complete discrimination in body-model studies (while GLCM only identified 1/6 of the 2D body models). It shows a stronger correlation with tumor volume changes after radiotherapy (R²=0.83, P<0.05) and is easier to implement clinically.

The rapid development of molecular targeted therapy and immunotherapy in recent years has significantly increased the diagnosis and treatment rate of patients with advanced lung cancer, especially the targeted therapeutic agents, such as EGFR, anaplastic lymphoma kinase (ALK) and ROS oncogene 1 (ROS1) inhibitors are widely used in patients with specific genetic mutations, resulting in prolonged PFS and OS in some patients. However, the biggest obstacle to targeted therapy is the inevitable emergence of drug resistance (2). Recent studies have shown that the emergence of such resistance is closely associated with ITH, and the difference in PFS remains even in lung cancer patients with the same characteristics after targeted therapy, suggesting that ITH may be an important determinant of TKI resistance and treatment failure (4). It is therefore crucial to identify non-responders or drug-resistant populations at an early stage by non-invasive and global means so as to minimize their exposure to potentially side-effecting and futile treatments. Park et al. (60) showed that texture parameter grayscale covariance matrix entropy obtained from FDG-PET/CT images had an HR of 6.41 (95% CI: 2.800–14.680, P<0.01) for predicting PFS in NSCLC patients with EGFR mutations who underwent TKI treatment. Similarly, Song et al. (99) showed that CT-extracted texture features, such as “cluster prominence of gray Level co-occurrence” (HR 2.13, 95% CI: 1.330–3.400, P=0.01) and “short run high gray level emphasis of run length” (HR 2.43, 95% CI: 1.46–4.05, P=0.005) could accurately predict the risk of disease recurrence of NSCLC patients with EGFR mutations treated with TKI. And Gary et al. (100) demonstrated the prognostic value of temporal PET texture analysis in 47 NSCLC patients treated with erlotinib. Their prospective study showed that reduced metabolic heterogeneity [measured by first-order standard deviation (P=0.01), entropy (P=0.001), and uniformity (P=0.001)] was significantly associated with RECIST response at 12 weeks. Importantly, two key parameters independently predicted outcomes: (I) higher-order contrast at 6 weeks (P=0.002) and (II) percentage change in first-order entropy (P=0.03) for OS, with the latter also predicting treatment response (P=0.01). These findings suggest that early changes in tumor metabolic heterogeneity may serve as valuable biomarkers for monitoring erlotinib efficacy.

The implementation of immunotherapy has demonstrated OS benefit compared to chemotherapy and is recommended as one of the treatments for patients with stage IV NSCLC without driver mutations. Specifically, for patients with high PD-L1 expression (≥50%), specific immune checkpoint inhibitors alone or in combination with chemotherapy are recommended as first-line treatment options; for patients with intermediate PD-L1 expression (1–49%), immune-combination chemotherapy regimens are recommended; and for PD-L1-negative (<1%) patients, immune-combination chemotherapy or chemotherapy-only regimens may be chosen (2). Notably, clinical observation revealed that about 10% of PD-L1-negative patients still produce therapeutic responses to programmed death-1 (PD-1)/PD-L1 inhibitors, which may be related to the spatiotemporal heterogeneity of PD-L1 expression in tumor tissues (101). In addition, changes in tumor size are often measured clinically by imaging to assess drug efficacy, e.g., RECIST provides a uniform and reproducible method for evaluating the efficacy of tumor treatments in clinical trials and clinical practice, which helps physicians and researchers to compare the efficacy of different treatment strategies. However, atypical treatment responses, such as an increase in tumor volume during the initial phase, are sometimes observed in patients receiving immunotherapy, a phenomenon known as pseudoprogression. This phenomenon can be misinterpreted as disease progression by traditional RECIST criteria, thus influencing the development of the patient’s next cycle of treatment (102). This clinical dilemma highlights the value of developing novel non-invasive biomarkers that can accurately predict PD-L1 expression status and identify early response to therapy, thus providing clinicians with a more accurate basis for treatment decisions. Multiple studies have demonstrated the potential of radiomic features to predict PD-L1 expression and immunotherapy response in NSCLC through different imaging modalities. Kim et al. (103) identified significant correlations between PET texture features (low gray-level run emphasis, short run low gray-level emphasis, and long run low gray-level emphasis, all P<0.001) and PD-L1 mRNA expression in TCGA data. Norikane et al. (104) further established that six specific PET texture features could differentiate PD-L1 expression levels across the full spectrum (negative, low, and high expression). For clinical application, Chang et al. (105) validated the CoV from 18F-FDG PET/CT as an independent predictor of PD-L1 status (AUC 0.649, P=0.05), with high specificity (82.10%) at the optimal cutoff of 28.90. Similarly, Bracci et al. (106) developed CT-based models that reliably stratified PD-L1 expression, achieving AUCs of 0.789–0.811 for high expression (≥50%) and 0.763–0.806 for intermediate expression (1–49%). Yamagata et al. (61) demonstrated that dual-energy CT parameters (skewness and extracellular volume) could predict high PD-L1 expression with excellent accuracy (AUC 0.830, 95% CI: 0.670–0.930, specificity 96.00%). Wang et al. (107) advanced this field by identifying two metabolic phenotypes through dynamic PET/CT, where the fast metabolic group, characterized by more homogeneous FDG uptake and higher immune cell infiltration, may represent ideal candidates for immunotherapy. These studies collectively highlight how advanced imaging analytics can non-invasively assess tumor biology and guide precision immunotherapy approaches.

Classification of lung lesions and chest lymph nodes

Pathologic biopsy can provide a reliable basis for the development of treatment plans for lung cancer patients; however, biopsy tissue samples may suffer from sampling bias and be affected by a number of factors, such as sampling and testing time, and these limitations may significantly interfere with clinical treatment decisions. In terms of noninvasive diagnostic techniques, CT imaging can provide a preliminary assessment of the probability of malignancy by analyzing the morphological features of the nodule (including maximum diameter, doubling time, calcification pattern, and distribution of ground-glass components). Although the sensitivity of CT for the detection of isolated pulmonary nodules can reach 98%, its specificity is only 58%, with an obvious imbalance in diagnostic efficacy. 18F-FDG PET/CT technology aids in differential diagnosis by quantifying the differences in glucose metabolism, with SUV as the key quantitative index. Clinical studies have suggested SUV_max >2.5 as the diagnostic threshold for malignant nodules, with the following limitations: (I) benign lesions, such as active granulomas and infected lesions may lead to false-positive results; and (II) non-infectious inflammatory responses may also cause increased FDG uptake (108). These factors suggest that single imaging parameters still suffer from insufficient diagnostic accuracy, and there is an urgent need to develop multimodal joint analysis strategies to enhance differential diagnostic efficacy. Chen et al. (62) confirmed that machine learning analysis of dual-phase FDG PET/CT texture features significantly improved the ability to differentiate between benign and malignant solid pulmonary nodules (SPNs) in areas with a high prevalence of granulomas. The SVM model constructed using delayed-phase PET features demonstrated superior diagnostic performance compared to traditional indicators, such as early SUV_max and visual assessment, with an AUC of up to 0.900, highlighting the additional value of temporal metabolic heterogeneity evaluation. Caruso et al. (109) retrospectively included patients with lung nodules showing a SUV <2.5 on FDG PET/CT. First-order texture analysis was performed on the initial CT images obtained for CT-guided lung biopsy. The results showed that in malignant lesions, CT texture analysis revealed lower kurtosis values and higher skewness values. Under moderate filtering conditions, the ROC curves showed significant AUC for kurtosis and skewness (0.654 and 0.642, respectively, P<0.001). Suo et al. (110) found that through whole-lesion texture analysis, the heterogeneity difference between the edge and center of lesions on enhanced CT can effectively distinguish between benign and malignant pulmonary nodules. Analysis of 48 solitary nodules showed that the mean attenuation value difference and entropy difference were significantly higher in malignant lesions, with the combined indicator achieving an AUC of 0.864. Miwa et al. (111) applied fractal analysis to FDG-PET, showing significant differences in morphological fractal dimension (m-FD), SUV_max and density FD (d-FD) between benign and malignant nodules (P<0.050). Combining m-FD with SUV_max or d-FD increased diagnostic accuracy to 92.60% and 94.40%, respectively. In addition, with the widespread use of low-dose CT (LDCT) in lung cancer screening, the detection rate of ground-glass nodules (GGNs) has increased significantly. However, current CT imaging techniques still face significant challenges in the differential diagnosis of the benign and malignant nature of pure GGNs (pGGNs). The core difficulty lies in the inability to reliably distinguish pGGNs manifesting as invasive adenocarcinoma (IAC) from those presenting with other histopathological types. Qi et al. (112) retrospectively enrolled 575 patients presenting with persistent pGGNs and diagnosed with lung adenocarcinoma on postoperative pathology. They developed an ITH score combining local radiomics features and global pixel distribution patterns to evaluate its effectiveness in predicting IAC before surgery. The results showed that the ITH score demonstrated better predictive efficacy than conventional imaging performance, with AUC of 0.860 and 0.867 in the training cohort and validation cohort, respectively. GGNs can be classified into four clinicopathologic types: atypical adenomatous hyperplasia (AAH), adenocarcinoma in situ (AIS), microinvasive adenocarcinoma (MIA), and IAC. The study showed that the 5-year disease-free survival rate for patients with AIS and MIA was close to 100%, significantly better than the 60–70% survival rate for patients with IAC (108). This difference highlights the clinical value of early and accurate assessment of the risk of malignancy of GGNs by CT examination and implementation of timely intervention. However, as the exact correspondence between this classification criterion and CT imaging features has not been fully elucidated, it remains a challenge to accurately identify different pathologic subtypes, particularly in diagnostic CT examinations. This cognitive gap emphasizes the urgency of conducting studies on the quantitative analysis of CT manifestations of GGNs, with a view to establishing a more accurate imaging-pathology control system to provide a reliable basis for clinical differential diagnosis. Li et al. (113) developed and validated an SVM model using quantitative imaging biomarkers derived from high-resolution CT of 248 GGNs. The results showed that the model demonstrated significantly higher accuracy than three radiology experts in: (I) classifying lesions into four categories: AAH, AIS, MIA, and IAC (accuracy: 70.90% vs. 39.60%); (II) distinguishing AIS from MIA (accuracy: 73.10% vs. 35.70%); (III) differentiating indolent lesions from invasive lesions (accuracy: 88.10% vs. 60.80%). In addition, texture analysis of PET/CT has shown potential value in differentiating histologic types in patients with NSCLC. Bianconi et al. (114) retrospectively analyzed baseline PET/CT scans of 81 patients with histologically confirmed NSCLC and showed that compared to other histologic subtypes, squamous cell carcinoma had a larger volume, higher uptake, greater PET variability, and lower homogeneity than other histologic subtypes. In contrast, adenocarcinomas had significantly lower uptake, lower variability and higher homogeneity than the other subtypes.

In the field of mediastinal lymph node pathologic evaluation, two main interventional techniques are currently used: transbronchial ultrasound-guided transbronchial needle aspiration biopsy and mediastinoscopy, both of which have a specificity close to the ideal of 100%. In terms of imaging assessment, CT demonstrated a diagnostic performance of 55% sensitivity and 81% specificity when using a short-axis diameter of ≥10 mm as the discriminatory threshold for benign and malignant lymph nodes. In contrast, FDG-PET/CT demonstrated superior diagnostic accuracy by quantifying focal glucose metabolic activity (36). However, existing diagnostic techniques still face important challenges: first, mediastinal lymph node enlargement may arise from different etiologies, such as inflammatory response or malignant lesions; second, there is room for improvement in the diagnostic accuracy of thoracic lymph node metastasis, both by CT morphologic assessment and PET metabolic activity analysis. These limitations highlight the clinical need for the development of novel automated analytical methods-through deep mining of radiomics features and the establishment of reproducible quantitative analytical models, it is expected to provide more reliable decision support for the identification of benign and malignant lymph nodes, and ultimately to improve clinical practice. Meyer et al. (115) showed that several texture features derived from CT correlated with the degree of malignancy of mediastinal lymph nodes. Pham et al. (116) performed CT texture analysis of mediastinal lymph nodes for lung cancer by using two complementary texture analysis methods, namely, GLCM and experimental semivariogram. The study found that a logistic regression model achieved 75% sensitivity, 90% specificity, and an AUC of 0.890 in the test population. Ten-fold cross validation with 70% accuracy was obtained using SVM as a pattern classifier to classify benign and malignant lymph nodes. Bayanati et al. (117) retrospectively analyzed 72 lymph nodes from 43 cases of primary lung cancer by extracting chest CT grayscale covariance matrix and tour matrix texture features as well as morphological features. The results showed that the combination of textural and morphological features identified malignant mediastinal lymph nodes with a sensitivity of 81% and a specificity of 80% (AUC 0.870, P<0.001). Budiawan et al. (118) represented ITH by calculating the CoV of preoperative 18F-FDG PET/CT. The results showed that the mean CoV was higher in metastatic lymph nodes than in inflammatory lymph nodes. Therefore, CoV of primary tumors may complement conventional imaging in providing lymph node information before biopsy. Shin et al. (119) performed CT texture analysis of 132 lymph nodes from 80 patients, and the results showed that tightness and normalized standard deviation differed between the benign and malignant groups. The ability to identify malignant lymph nodes using high normalized uptake values (SUV_max ≥5) on FDG PET/CT and normalized standard deviation on CT texture analysis was higher (AUC 0.739 and 0.742, respectively). However, normalized standard deviation had a very low sensitivity, despite its high specificity. Therefore, CT texture analysis may be helpful in identifying benign and malignant lymph nodes, but the diagnostic value is not high and needs to be evaluated in combination with other imaging methods.

Characterize tumor pathology and molecular features

Tumors exhibit significant heterogeneity at both the genetic and pathohistological levels, and imaging techniques are emerging as an important tool for quantifying ITH. However, the mechanism of correlation between quantitative imaging features and biological characteristics in CT or PET/CT images is still unclear. Radiomics combined with cellular features in pathology images forms radiopathology, and correlation with tumor genetics data constitutes radiogenomics, and both interdisciplinary research approaches can provide potential biological explanations for radiomics features. At the level of radiopathology, Aydos et al. (76) concluded that textural features were significantly correlated with histological tumor type, degree of pathological differentiation, tumor proliferation and necrosis rates. They retrospectively analyzed the textural features of primary tumors on preoperative PET/CT images in 90 patients with NSCLC. The results showed significant differences in metabolic parameters between adenocarcinomas and squamous cell carcinomas, with the squamous cell carcinoma group having a higher Ki-67 index and lower kurtosis values. In patients with adenocarcinoma, textural heterogeneity was higher in poorly differentiated tumors than in moderately differentiated tumors. The Ki-67 index, an indicator of tumor proliferation, was significantly correlated with metabolic parameters and kurtosis, and tumor necrosis rate was significantly correlated only with textural features. Bashir et al. (120) prospectively enrolled 22 patients with NSCLC, and showed that histopathological mean tumor cell density and voiding were correlated with several commonly used F-FDG PET-derived metrics, including SUV-voiding, metabolic activity tumor volume, SUV_mean, entropy, skewness, and kurtosis, and thus may explain the biological basis of F-FDG PET uptake heterogeneity in NSCLC. Multifactorial logistic regression analyses by Liu et al. (63) showed that textural features F33 (homogeneity 1) (P=0.005) and F38 (inverse variance) (P=0.03) on preoperative CECT images of lung adenocarcinoma were pathologic grading independent imaging metrics, and the AUC calculated by this model (AUC 0.834) was higher than that of the clinical model (AUC 0.615) (P<0.001). It was shown that features, such as tumor hypoxic microenvironment, necrotic foci formation, aberrant neovascularization, heterogeneous cell subpopulations, and differential proliferation rates were independently predictive of increased tumor aggressiveness and poor response to therapy. Ganeshan et al. (121) explored the correlation of texture parameters derived from preoperative CT images of NSCLC with tumor hypoxia and angiogenesis. Image filtering analysis revealed that areas of higher tumor heterogeneity (as evidenced by larger standard deviation, higher maximum probability pixel values and lower percentage of homogeneity) were significantly correlated with enhanced expression of endogenous (pimonidazole) and exogenous (Glut-1) markers of hypoxia. In addition, elevated maximum probability pixel values of highlighted objects tended to correlate negatively with attenuated expression of angiogenic markers (CD34) and reduced microvessel density. A study by Castello et al. (122) showed that higher entropy levels of texture features appeared to correlate with an increased density of CD8-TILs suggesting that it may be related to the presence of inactivated immune cell infiltration. In radiogenomics, Weiss et al. (123) selected cases of NSCLC patients with the presence of K-RAS mutations or 26 oncogenes and tumor suppressor genes that were pan-wild-type for CT quantitative texture analysis. The results showed that positive skewness and lower kurtosis were significantly associated with the presence of K-RAS mutations. A decision tree containing five nodes was 89.6% accurate in identifying K-RAS mutations and correlated with patient survival. Agazzi et al. (64) developed a CT texture-based model to predict EGFR-mutant, ALK-rearranged lung adenocarcinomas. A filter-histogram method was used to extract 17 texture features for the model construction, and the model achieved an average accuracy of 81.76% (95% CI: 81.450–82.060) in an independent nested validation set. Among them, the skewness on unfiltered images and fine texture were the most important features, which may be promising for non-invasive characterization of lung adenocarcinoma in terms of EGFR and ALK mutations. Zhang et al. (124) extracted 485 features that could reflect the tumor heterogeneity and phenotypes from CT images, including clinical cognitive features, image intensity features, texture features, and wavelet features. Multivariate analysis showed that the adiogenomics features had the potential to construct a model for predicting EGFR mutations in NSCLC patients, which was superior to predictive models constructed using only clinical variables. Multifactorial logistic regression analyses by Shi et al. (125) showed that low SUV_max, low SUV_mean, and high COV were significantly correlated with EGFR mutations. When combined with clinicopathological features, these parameters could improve the predictive value for EGFR mutations.

Spatial IITH of NSCLC

Habitat analysis aims to identify distinct tumor regions and cellular subpopulations. This approach quantitatively characterizes different functional areas within a tumor by segmenting it into voxel clusters with comparable radiogenomics features. Currently, habitat analysis has been widely used to display and quantify ITH, showing great potential in non-invasive and precise localization of lesions, prediction of tumor typing and prognosis, characterization of tumor pathology and molecular features, and identification of tumor progression and response to therapy, which further facilitates clinical diagnosis and treatment. In this section, we will give an overview of the clinical application of NSCLC in habitat analysis.

Habitat analysis can be used for classification of histopathological subtypes and prediction of molecular features. Shen et al. (65) successfully classified lung squamous carcinoma and lung adenocarcinoma using PET/CT images. The study extracted 3D feature vectors for each tumor voxel based on the PET value, CT value, and the CT local dominant orientation. Each tumor was classified into four subregions by K-means individual clustering and overall clustering, and finally seven classifiers were tested to obtain an optimized model for classification. The results showed the best performance when using subregion-based imaging histology features and an SVM-RBF classifier with an accuracy of 0.862 and an AUC of 0.916, respectively, under 5-fold cross-validation. Wu et al. (66) retrospectively included 438 NSCLC patients from four academic medical centers who underwent radical surgery. A radiomics column plot prediction model integrating deep learning, radiomics, and clinical variables was developed to identify EGFR mutation status in stage I NSCLC by extracting and analyzing radiomics features within the tumor, around the tumor, and in the imaging ecoregion region of CT images. The results showed that radiomics features based on imaging ecological loci demonstrated good discriminative efficacy in identifying EGFR mutant versus wild-type, with the highest AUC values of 0.917, 0.837, and 0.809 for the training, validation, and external test sets, respectively. Habitat analysis has also achieved some success in evaluating the response to immunotherapy for patients with lung cancer. Ye et al. (67) conducted a quantitative analysis of ITH in pre-treatment CT images and used K-means unsupervised clustering to determine the optimal division of tumor subregions for constructing an ITH habitat model. The results showed that the AUC value and accuracy of the habitat model were superior to traditional imaging in predicting pathological complete response (pCR) in NSCLC patients receiving neoadjuvant immunochemotherapy. Caii et al.(68) developed a non-invasive predictive model for immunotherapy response in advanced NSCLC patients based on the integration of deep learning and TME radiomics. Entropy and voxel metrics extracted from CT images were used to classify the volume of interest into different subregions. Tumor subregions were identified via the K-means method, and deep learning features of the entire tumor were extracted using a 3D ResNet algorithm. SVMs were applied to build models based on subregion radiomics features and whole-tumor deep learning features, respectively. The results showed that the combined model constructed from whole-tumor deep learning features and subregion radiomics features significantly improved the identification of high- and low-risk patients. Peng et al. (126) also used the K-means method to extract radiomics features from three subregions of the tumor ROI in CT images. The least absolute shrinkage and selection operator (LASSO) regression method was employed to select subregion radiomics features, and an SVM algorithm was used to construct and validate the subregion radiomics model. The results showed that the constructed subregional radiomics model predicted the treatment response of NSCLC patients treated with an immune checkpoint inhibitor in the training cohort with an AUC as high as 0.900 and was significantly correlated with PFS. In addition, significant correlations were observed between the subregional radiomics model and the three variant genes (H3C4, PAX5 and EGFR). Thus, subregional radiomics models can effectively stratify the risk of PFS in NSCLC patients receiving immunotherapy.

However, the habitat delineation method extracts statistical or textural features from specific subregions without taking into account the global distribution patterns of these features, thus leading to the loss of ITH-related information. In lung cancer, Li et al. (9) developed a novel metric called the ITH score. Specifically, for each pixel within an ROI, 104 radiomic features from every pixel in its surrounding neighborhood were recorded as its “local information”. Pixels were then clustered based on their “local information” and assigned corresponding colors. The resulting cluster label map visualized the global distribution pattern of the tumor, where clustered distributions reflected ITH, and the degree of dispersion calculated by a formula defined the ITH score. Analysis of CT images from 1,399 patients demonstrated that the ITH score was significantly associated with tumor staging, pathological grade, invasiveness (all P<0.001), and patient prognosis. The significance of the ITH score was further validated via gene set enrichment analysis (GSEA), which showed that differentially expressed genes in patients with high ITH scores were enriched in heterogeneity-related pathways. By integrating global and local imaging features, the proposed ITH score comprehensively quantifies ITH. It overcomes limitations of existing radiomic signatures (which lack characterization of regional variations) and tumor habitat analysis (which lacks global perspective), providing a more comprehensive and systematic depiction of ITH.

Evaluation of multi-omics studies contributing to ITH

The high-dimensional, multi-level characterization of NSCLC through genomics, epigenomics, transcriptomics, proteomics, and metabolomics has opened up new avenues for dissecting tumor heterogeneity. Radiogenomics is a key component of precision cancer medicine. By leveraging the rich database information from genomics, transcriptomics, and proteomics, as well as enrichment analysis algorithms, it provides possible explanations for the genes associated with imaging features and traits, and has been widely used to address the problem of tumor heterogeneity. In the future, with methodological advancements, the combination of radiomics with epigenomics and metabolomics will likely bring broader application value to radiomics, enabling multi-omics precise prediction of lung cancer before surgery with radiomics as a bridge, and bringing about a revolution in treatment regimens in the era of artificial intelligence (AI) (127). Ju et al. (128) collected 18F-FDG PET/CT images and RNA sequencing data from 93 NSCLC patients in The Cancer Imaging Archive (TCIA). They developed a graph neural network model to predict lymph node metastasis in NSCLC by integrating a protein-protein interaction (PPI) network derived from gene expression with PET/CT-derived texture features. Results demonstrated that incorporating PPI data from four metastasis-related genes with 18F-FDG PET/CT features increased prediction accuracy to 0.8545, outperforming models without imaging input.

Metzenmacher et al. (129) analyzed sequentially collected plasma and paired tumor samples from lung cancer patients (61 stage IV patients, 48 stage I-III patients) and 61 healthy samples through NGS, radiographic imaging, and droplet digital polymerase chain reaction (ddPCR) mutation and methylation detection. The results showed that the combined dynamic analysis of ctDNA mutation profiles and radiomics heterogeneity features significantly improved the sensitivity of recurrence monitoring in NSCLC (92% vs. 68%), and predicted drug resistance 8 weeks in advance, providing a multi-scale window for tumor evolution monitoring. In a multicenter study involving 394 NSCLC patients, Sujit et al. (130) utilized preprocessed CT and 18F-FDG PET to construct a habitat imaging framework for assessing intra-tumoral regional heterogeneity. This framework identified three distinct PET/CT subtypes that retained prognostic significance after adjustment for clinicopathological variables including tumor volume. Moreover, these imaging subtypes complemented ctDNA analysis and were predictive of disease recurrence. Radiogenomic analysis elucidated the molecular basis of these imaging subtypes, notably identifying downregulation of interferon-α and interferon-γ pathways in high-risk groups. In conclusion, the study demonstrates that habitat imaging subtypes enable effective risk stratification of NSCLC patients, offering valuable insights for personalized post-surgical or post-radiotherapy management.

Challenges and directions for future research

Previous studies at the molecular, biological, and macroscopic morphological levels have consistently demonstrated a high degree of ITH in NSCLC, posing significant challenges for its diagnosis and treatment. Given the temporal dynamics and spatial complexity of ITH in NSCLC, there is growing interest in developing tools that can accurately, intuitively, and comprehensively capture tumor heterogeneity in a clinically feasible manner. Such tools hold promise for overcoming the limitations of conventional diagnostic methods and advancing precision medicine. With continuous advances in computing technologies, quantitative imaging features derived from various imaging modalities—together with advanced analytical techniques, such as radiomics, texture analysis, and habitat imaging—have shown value not only for diagnosing and classifying NSCLC and its molecular subtypes, but also for assessing treatment response and prognosis. Overall, tumors with higher imaging-detected heterogeneity often exhibit more aggressive biological behavior and are associated with poorer outcomes. Notably, ITH as assessed by imaging typically decreases following treatment. However, translating these methods into routine clinical practice remains challenging. Key obstacles include the need for robust standardization of imaging acquisition and analysis protocols, reliable prospective validation in multi-center cohorts, and improvements in the interpretability of radiomic and texture-derived biomarkers.

Despite the enormous potential of AI in medical imaging, there are currently no AI tools specifically designed or approved for ITH characterization, highlighting the translational gap between research and clinical application and underscoring the need to develop domain-specific AI solutions for radiomic heterogeneity analysis. In addition, there is still no universally accepted, reliable, and standardized method for quantifying IITH, and various steps in image analysis—from acquisition and reconstruction to tumor segmentation, feature extraction, and post-processing—remain technically challenging. Many studies have shown that radiomic features are highly sensitive to variations in acquisition parameters, reconstruction algorithms, discretization schemes, and segmentation techniques. For example, Hosseini et al. (131) reported that in PET radiomic analyses of NSCLC, only 29% of features had a COV ≤5% under motion conditions, and many features were unstable when respiratory motion and reconstruction parameters interacted, emphasizing the crucial impact of image acquisition and motion control on feature robustness. Similarly, Yan et al. (132) found that only about 33% of texture features maintained a COV below 5% under different reconstruction settings (iterations, filters, grid size), with skewness, cluster shade, and zone percentage being particularly sensitive—grid size had the largest impact. Quantization methods also significantly affect feature reproducibility. In PET, fixed bin number often performs better, while in CT, fixed bin width yields more robust results. Observer variability in defining tumor ROIs is another important factor. Various segmentation methods—including thresholding, region-growing, and fuzzy locally adaptive Bayesian algorithms—have produced differing levels of reproducibility and clinical validity. Hatt et al. (133) highlighted that tumor size and uptake heterogeneity significantly influence PET segmentation results, with CT-based tumor volumes generally larger than those defined by PET. Especially for larger, more heterogeneous tumors, threshold-based methods often underestimate the true functional tumor extent. Therefore, optimal segmentation should be tailored to specific imaging workflows, scanners, and reconstruction settings, and semi-automatic or automatic methods may help reduce variability. Moreover, medical imaging inevitably faces challenges related to signal-to-noise ratio, partial volume effects, limited resolution, and image noise. Radiomics studies have used various filters (e.g., Wiener, Gaussian) for denoising and Laplacian of Gaussian filters to enhance edges. One study (134) proposed that integrating segmentation, denoising, and partial volume correction (PVC) as a combined process can significantly improve the accuracy of quantitative PET image analysis, and demonstrated its advantages for image quality and feature extraction across multiple datasets. The sheer number of image features presents another challenge, as many are linearly correlated and redundant. Thus, careful selection of a minimal, robust feature set is necessary, informed by the variability, study objectives, case number, and lesion size, or by applying dimensionality reduction techniques, such as principal component analysis to remove redundant features. Despite the surge in retrospective studies, prospective validation of radiomic biomarkers remains scarce. Dedicated prospective trials are needed to establish the reproducibility, clinical utility, and generalizability of imaging biomarkers—ideally within multi-center trials with predefined acquisition and analysis protocols. External validation, especially using open repositories, such as TCIA or The Cancer Genome Atlas (TCGA), is still uncommon, and the average radiomics quality score (RQS) for lung cancer studies remains low (135). Another major barrier is the fragmentation and limited size of available datasets. Although repositories like TCIA provide open-access cohorts, varying institutional protocols lead to significant inconsistencies in data quality. Cross-institutional collaboration—through federated learning frameworks or trusted third-party data custodians—will be essential to build large, well-annotated databases.

Artificial intelligence, especially machine learning and deep learning algorithms, offers an alternative to traditional hand-crafted radiomics by reducing reliance on precise tumor segmentation and feature definition, thereby enhancing feature consistency and reproducibility while easing data management. However, the use of deep learning in current literature remains limited due to its demand for large training datasets and the challenge of model interpretability. The perception of deep learning-based radiomics as a “black box” further hampers its clinical adoption due to the abstract nature of extracted features. Meanwhile, the rise of multi-omics research—including genomics, epigenomics, transcriptomics, proteomics, metabolomics, and single-cell sequencing—opens new avenues to integrate multi-omics with imaging features. Such integration could not only enhance the biological interpretability of radiomics but also facilitate the creation of powerful noninvasive predictive models, thereby advancing personalized treatment planning.

Conclusions

IITH correlates with aggressive behavior and poorer outcomes in NSCLC, typically decreasing post-treatment. However, translating quantitative imaging features into clinical practice faces major challenges. Key barriers include lack of standardized protocols, insufficient prospective validation, limited biomarker interpretability, and absence of approved AI tools for IITH characterization. Overcoming these hurdles alongside multi-omics integration is essential to advance NSCLC precision medicine.

Acknowledgments

The authors acknowledge that all figures presented in this study were created with BioGDP.com.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-506/rc

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-506/prf

Funding: This study was supported by the 1.3.5 project for disciplines of excellence, West China Hospital, Sichuan University (Nos. 2017-046, ZYJC21028 to W.L.); Key R & D plan of Sichuan Provincial Department of science and technology (No. 2021YFS0072 to W.L.); Young Scientists Fund of Natural Science Foundation of Sichuan (No. 2024NSFJQ0051 to W.L.) and Science and Technology Project of Sichuan (No. 2022ZDZX0018 to W.L.).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-506/coif). W.L. reports research funding from the 1.3.5 Project for Disciplines of Excellence (West China Hospital, Sichuan University), the Key R & D Plan of Sichuan Provincial Department of Science and Technology, the Young Scientists Fund of the Natural Science Foundation of Sichuan, and the Science and Technology Project of Sichuan. The other authors have no conflicts of interest to declare

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Su PL, Furuya N, Asrar A, et al. Recent advances in therapeutic strategies for non-small cell lung cancer. J Hematol Oncol 2025;18:35. [Crossref] [PubMed]
3. Hausser J, Alon U. Tumour heterogeneity and the evolutionary trade-offs of cancer. Nat Rev Cancer 2020;20:247-57. [Crossref] [PubMed]
4. Fu YC, Liang SB, Luo M, et al. Intratumoral heterogeneity and drug resistance in cancer. Cancer Cell Int 2025;25:103. [Crossref] [PubMed]
5. Goyette MA, Lipsyc-Sharf M, Polyak K. Clinical and translational relevance of intratumor heterogeneity. Trends Cancer 2023;9:726-37. [Crossref] [PubMed]
6. Martínez-Ruiz C, Black JRM, Puttick C, et al. Genomic-transcriptomic evolution in lung cancer and metastasis. Nature 2023;616:543-52. [Crossref] [PubMed]
7. Koo JM, Kim J, Lee J, et al. Deciphering the intratumoral histologic heterogeneity of lung adenocarcinoma using radiomics. Eur Radiol 2025;35:4861-72. [Crossref] [PubMed]
8. Chen M, Copley SJ, Viola P, et al. Radiomics and artificial intelligence for precision medicine in lung cancer treatment. Semin Cancer Biol 2023;93:97-113. [Crossref] [PubMed]
9. Li J, Qiu Z, Zhang C, et al. ITHscore: comprehensive quantification of intra-tumor heterogeneity in NSCLC by multi-scale radiomic features. Eur Radiol 2023;33:893-903. [Crossref] [PubMed]
10. Zhang L, Wang Y, Peng Z, et al. The progress of multimodal imaging combination and subregion based radiomics research of cancers. Int J Biol Sci 2022;18:3458-69. [Crossref] [PubMed]
11. Hynds RE, Huebner A, Pearce DR, et al. Representation of genomic intratumor heterogeneity in multi-region non-small cell lung cancer patient-derived xenograft models. Nat Commun 2024;15:4653. [Crossref] [PubMed]
12. Lv X, Mao Z, Sun X, et al. Intratumoral Heterogeneity in Lung Cancer. Cancers (Basel) 2023;15:2709. [Crossref] [PubMed]
13. Tauriello DVF, Sancho E, Batlle E. Overcoming TGFβ-mediated immune evasion in cancer. Nat Rev Cancer 2022;22:25-44. [Crossref] [PubMed]
14. Naso JR, Wang G, Pender A, et al. Intratumoral heterogeneity in programmed death-ligand 1 immunoreactivity is associated with variation in non-small cell lung carcinoma histotype. Histopathology 2020;76:394-403. [Crossref] [PubMed]
15. Gerlinger M, Rowan AJ, Horswell S, et al. Intratumor heterogeneity and branched evolution revealed by multiregion sequencing. N Engl J Med 2012;366:883-92. [Crossref] [PubMed]
16. Dentro SC, Leshchiner I, Haase K, et al. Characterizing genetic intra-tumor heterogeneity across 2,658 human cancer genomes. Cell 2021;184:2239-2254.e39. [Crossref] [PubMed]
17. Ou X, Gao G, Habaz IA, et al. Mechanisms of resistance to tyrosine kinase inhibitor-targeted therapy and overcoming strategies. MedComm (2020) 2024;5:e694.
18. Sinjab A, Han G, Treekitkarnmongkol W, et al. Resolving the Spatial and Cellular Architecture of Lung Adenocarcinoma by Multiregion Single-Cell Sequencing. Cancer Discov 2021;11:2506-23. [Crossref] [PubMed]
19. Jamal-Hanjani M, Wilson GA, McGranahan N, et al. Tracking the Evolution of Non-Small-Cell Lung Cancer. N Engl J Med 2017;376:2109-21. [Crossref] [PubMed]
20. Roosan MR, Mambetsariev I, Pharaon R, et al. Usefulness of Circulating Tumor DNA in Identifying Somatic Mutations and Tracking Tumor Evolution in Patients With Non-small Cell Lung Cancer. Chest 2021;160:1095-107. [Crossref] [PubMed]
21. Li C, Shao J, Li P, et al. Circulating tumor DNA as liquid biopsy in lung cancer: Biological characteristics and clinical integration. Cancer Lett 2023;577:216365. [Crossref] [PubMed]
22. Souza VGP, Forder A, Brockley LJ, et al. Liquid Biopsy in Lung Cancer: Biomarkers for the Management of Recurrence and Metastasis. Int J Mol Sci 2023;24:8894. [Crossref] [PubMed]
23. Rahal Z, El Darzi R, Moghaddam SJ, et al. Tumour and microenvironment crosstalk in NSCLC progression and response to therapy. Nat Rev Clin Oncol 2025;22:463-82. [Crossref] [PubMed]
24. Oliveira G, Wu CJ. Dynamics and specificities of T cells in cancer immunotherapy. Nat Rev Cancer 2023;23:295-316. [Crossref] [PubMed]
25. Hiam-Galvez KJ, Allen BM, Spitzer MH. Systemic immunity in cancer. Nat Rev Cancer 2021;21:345-59. [Crossref] [PubMed]
26. Luga V, Zhang L, Viloria-Petit AM, et al. Exosomes mediate stromal mobilization of autocrine Wnt-PCP signaling in breast cancer cell migration. Cell 2012;151:1542-56. [Crossref] [PubMed]
27. Tong Y, Gao WQ, Liu Y. Metabolic heterogeneity in cancer: An overview and therapeutic implications. Biochim Biophys Acta Rev Cancer 2020;1874:188421. [Crossref] [PubMed]
28. Demicco M, Liu XZ, Leithner K, et al. Metabolic heterogeneity in cancer. Nat Metab 2024;6:18-38. [Crossref] [PubMed]
29. Vaclova T, Grazini U, Ward L, et al. Clinical impact of subclonal EGFR T790M mutations in advanced-stage EGFR-mutant non-small-cell lung cancers. Nat Commun 2021;12:1780. [Crossref] [PubMed]
30. Luo YH, Chen YM. Influence of chemotherapy on EGFR mutation status. Transl Lung Cancer Res 2013;2:442-4. [Crossref] [PubMed]
31. Bailey C, Black JRM, Reading JL, et al. Tracking Cancer Evolution through the Disease Course. Cancer Discov 2021;11:916-32. [Crossref] [PubMed]
32. Wu F, Fan J, He Y, et al. Single-cell profiling of tumor heterogeneity and the microenvironment in advanced non-small cell lung cancer. Nat Commun 2021;12:2540. [Crossref] [PubMed]
33. Cui X, Liu S, Song H, et al. Single-cell and spatial transcriptomic analyses revealing tumor microenvironment remodeling after neoadjuvant chemoimmunotherapy in non-small cell lung cancer. Mol Cancer 2025;24:111. [Crossref] [PubMed]
34. Au Yeung CL, Co NN, Tsuruga T, et al. Exosomal transfer of stroma-derived miR21 confers paclitaxel resistance in ovarian cancer cells through targeting APAF1. Nat Commun 2016;7:11150. [Crossref] [PubMed]
35. Filippou PS, Karagiannis GS, Musrap N, et al. Kallikrein-related peptidases (KLKs) and the hallmarks of cancer. Crit Rev Clin Lab Sci 2016;53:277-91. [Crossref] [PubMed]
36. Chen P, Liu Y, Wen Y, et al. Non-small cell lung cancer in China. Cancer Commun (Lond) 2022;42:937-70. [Crossref] [PubMed]
37. Zhang J, Zhang X, Zhong Y, et al. PET radiomics for histologic subtype classification of non-small cell lung cancer: a systematic review and meta-analysis. Eur J Nucl Med Mol Imaging 2025;52:2212-24. [Crossref] [PubMed]
38. van Velden FH, Cheebsumon P, Yaqub M, et al. Evaluation of a cumulative SUV-volume histogram method for parameterizing heterogeneous intratumoural FDG uptake in non-small cell lung cancer PET studies. Eur J Nucl Med Mol Imaging 2011;38:1636-47. [Crossref] [PubMed]
39. Espinasse M, Pitre-Champagnat S, Charmettant B, et al. CT Texture Analysis Challenges: Influence of Acquisition and Reconstruction Parameters: A Comprehensive Review. Diagnostics (Basel) 2020;10:258. [Crossref] [PubMed]
40. Bashir U, Siddique MM, Mclean E, et al. Imaging Heterogeneity in Lung Cancer: Techniques, Applications, and Challenges. AJR Am J Roentgenol 2016;207:534-43. [Crossref] [PubMed]
41. Lennon FE, Cianci GC, Cipriani NA, et al. Lung cancer-a fractal viewpoint. Nat Rev Clin Oncol 2015;12:664-75. [Crossref] [PubMed]
42. Bera K, Braman N, Gupta A, et al. Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat Rev Clin Oncol 2022;19:132-46. [Crossref] [PubMed]
43. Napel S, Mu W, Jardim-Perassi BV, et al. Quantitative imaging of cancer in the postgenomic era: Radio(geno)mics, deep learning, and habitats. Cancer 2018;124:4633-49. [Crossref] [PubMed]
44. Molad J CT. Perfusion: More Than What You Thought. AJNR Am J Neuroradiol 2021;42:73-4. [Crossref] [PubMed]
45. Koopman MS, Hoving JW, Kappelhof M, et al. Association of Ischemic Core Imaging Biomarkers With Post-Thrombectomy Clinical Outcomes in the MR CLEAN Registry. Front Neurol 2021;12:771367. [Crossref] [PubMed]
46. McGuire SA, Wijtenburg SA, Sherman PM, et al. Reproducibility of quantitative structural and physiological MRI measurements. Brain Behav 2017;7:e00759. [Crossref] [PubMed]
47. Cho HH, Kim H, Nam SY, et al. Measurement of Perfusion Heterogeneity within Tumor Habitats on Magnetic Resonance Imaging and Its Association with Prognosis in Breast Cancer Patients. Cancers (Basel) 2022;14:1858. [Crossref] [PubMed]
48. Xue JH, Titterington DM. t-Tests, F-tests and Otsu's methods for image thresholding. IEEE Trans Image Process 2011;20:2392-6. [Crossref] [PubMed]
49. You D, Kim MM, Aryal MP, et al. Tumor image signatures and habitats: a processing pipeline of multimodality metabolic and physiological images. J Med Imaging (Bellingham) 2018;5:011009. [Crossref] [PubMed]
50. Xu X, Samaras D, Prasanna P. Radiologically Defined Tumor-habitat Adjacency as a Prognostic Biomarker in Glioblastoma. Annu Int Conf IEEE Eng Med Biol Soc 2021;2021:3998-4001. [Crossref] [PubMed]
51. Sammouda R, El-Zaart A. An Optimized Approach for Prostate Image Segmentation Using K-Means Clustering Algorithm with Elbow Method. Comput Intell Neurosci 2021;2021:4553832. [Crossref] [PubMed]
52. Syed AK, Whisenant JG, Barnes SL, et al. Multiparametric Analysis of Longitudinal Quantitative MRI data to Identify Distinct Tumor Habitats in Preclinical Models of Breast Cancer. Cancers (Basel) 2020;12:1682. [Crossref] [PubMed]
53. Kazerouni AS, Hormuth DA 2nd, Davis T, et al. Quantifying Tumor Heterogeneity via MRI Habitats to Characterize Microenvironmental Alterations in HER2+ Breast Cancer. Cancers (Basel) 2022;14:1837. [Crossref] [PubMed]
54. Karacavus S, Yılmaz B, Tasdemir A, et al. Can Laws Be a Potential PET Image Texture Analysis Approach for Evaluation of Tumor Heterogeneity and Histopathological Characteristics in NSCLC? J Digit Imaging 2018;31:210-23. [Crossref] [PubMed]
55. Zhang X, Zhang G, Qiu X, et al. Radiomics under 2D regions, 3D regions, and peritumoral regions reveal tumor heterogeneity in non-small cell lung cancer: a multicenter study. Radiol Med 2023;128:1079-92. [Crossref] [PubMed]
56. Hua J, Li L, Liu L, et al. The diagnostic value of metabolic, morphological and heterogeneous parameters of 18F-FDG PET/CT in mediastinal lymph node metastasis of non-small cell lung cancer. Nucl Med Commun 2021;42:1247-53. [Crossref] [PubMed]
57. Hakkak Moghadam Torbati A, Pellegrino S, Fonti R, et al. Machine Learning and Texture Analysis of [18F]FDG PET/CT Images for the Prediction of Distant Metastases in Non-Small-Cell Lung Cancer Patients. Biomedicines 2024;12:472.
58. Nakajo M, Jinguji M, Shinaji T, et al. A Pilot Study of Texture Analysis of Primary Tumor [18F]FDG Uptake to Predict Recurrence in Surgically Treated Patients with Non-small Cell Lung Cancer. Mol Imaging Biol 2019;21:771-80.
59. Franceschini D, Cozzi L, De Rose F, et al. A radiomic approach to predicting nodal relapse and disease-specific survival in patients treated with stereotactic body radiation therapy for early-stage non-small cell lung cancer. Strahlenther Onkol 2020;196:922-31. [Crossref] [PubMed]
60. Park S, Ha S, Lee SH, et al. Intratumoral heterogeneity characterized by pretreatment PET in non-small cell lung cancer patients predicts progression-free survival on EGFR tyrosine kinase inhibitor. PLoS One 2018;13:e0189766. [Crossref] [PubMed]
61. Yamagata K, Yanagawa M, Hata A, et al. Three-dimensional iodine mapping quantified by dual-energy CT for predicting programmed death-ligand 1 expression in invasive pulmonary adenocarcinoma. Sci Rep 2024;14:18310. [Crossref] [PubMed]
62. Chen S, Harmon S, Perk T, et al. Diagnostic classification of solitary pulmonary nodules using dual time (18)F-FDG PET/CT image texture features in granuloma-endemic regions. Sci Rep 2017;7:9370. [Crossref] [PubMed]
63. Liu Y, Liu S, Qu F, et al. Tumor heterogeneity assessed by texture analysis on contrast-enhanced CT in lung adenocarcinoma: association with pathologic grade. Oncotarget 2017;8:53664-74. [Crossref] [PubMed]
64. Agazzi GM, Ravanelli M, Roca E, et al. CT texture analysis for prediction of EGFR mutational status and ALK rearrangement in patients with non-small cell lung cancer. Radiol Med 2021;126:786-94. [Crossref] [PubMed]
65. Shen H, Chen L, Liu K, et al. A subregion-based positron emission tomography/computed tomography (PET/CT) radiomics model for the classification of non-small cell lung cancer histopathological subtypes. Quant Imaging Med Surg 2021;11:2918-32. [Crossref] [PubMed]
66. Wu J, Meng H, Zhou L, et al. Habitat radiomics and deep learning fusion nomogram to predict EGFR mutation status in stage I non-small cell lung cancer: a multicenter study. Sci Rep 2024;14:15877. [Crossref] [PubMed]
67. Ye G, Wu G, Zhang C, et al. CT-based quantification of intratumoral heterogeneity for predicting pathologic complete response to neoadjuvant immunochemotherapy in non-small cell lung cancer. Front Immunol 2024;15:1414954. [Crossref] [PubMed]
68. Caii W, Wu X, Guo K, et al. Integration of deep learning and habitat radiomics for predicting the response to immunotherapy in NSCLC patients. Cancer Immunol Immunother 2024;73:153. [Crossref] [PubMed]
69. Safarian A, Mirshahvalad SA, Nasrollahi H, et al. Impact of [18F]FDG PET/CT Radiomics and Artificial Intelligence in Clinical Decision Making in Lung Cancer: Its Current Role. Semin Nucl Med 2025;55:156-66.
70. Apostolova I, Ego K, Steffen IG, et al. The asphericity of the metabolic tumour volume in NSCLC: correlation with histopathology and molecular markers. Eur J Nucl Med Mol Imaging 2016;43:2360-73. [Crossref] [PubMed]
71. Li R, Lin J, Wang L, et al. The association between 18F-fluorodeoxyglucose PET intratumoral metabolic heterogeneity and pathological parameters in non-small cell lung cancer. Nucl Med Commun 2019;40:1022-8. [Crossref] [PubMed]
72. De Leyn P, Vansteenkiste J, Cuypers P, et al. Role of cervical mediastinoscopy in staging of non-small cell lung cancer without enlarged mediastinal lymph nodes on CT scan. Eur J Cardiothorac Surg 1997;12:706-12. [Crossref] [PubMed]
73. Dunne EG, Fick CN, Jones DR. Mediastinal Staging in Non-Small-Cell Lung Cancer: Saying Goodbye to Mediastinoscopy. J Clin Oncol 2023;41:3785-90. [Crossref] [PubMed]
74. Mazzone PJ, Lam L. Evaluating the Patient With a Pulmonary Nodule: A Review. JAMA 2022;327:264-73. [Crossref] [PubMed]
75. Craigie M, Squires J, Miles K. Can CT measures of tumour heterogeneity stratify risk for nodal metastasis in patients with non-small cell lung cancer? Clin Radiol 2017;72:899.e1-7. [Crossref] [PubMed]
76. Aydos U, Ünal ER, Özçelik M, et al. Texture features of primary tumor on (18)F-FDG PET images in non-small cell lung cancer: The relationship between imaging and histopathological parameters. Rev Esp Med Nucl Imagen Mol (Engl Ed) 2021;40:343-50. [Crossref] [PubMed]
77. Öztürk Ergür F, Öztürk A, Özdağ Ö, et al. When to Consider Invasive Lymph Node Staging in Non-Small-Cell Lung Cancer? A Novel Scoring System Utilising Metabolic Parameters in 18F-FDG PET/CT. Arch Bronconeumol 2024;60:S4-S12. [Crossref] [PubMed]
78. Ouyang ML, Xia HW, Xu MM, et al. Prediction of occult lymph node metastasis using SUV, volumetric parameters and intratumoral heterogeneity of the primary tumor in T1-2N0M0 lung cancer patients staged by PET/CT. Ann Nucl Med 2019;33:671-80. [Crossref] [PubMed]
79. Win T, Miles KA, Janes SM, et al. Tumor heterogeneity and permeability as measured on the CT component of PET/CT predict survival in patients with non-small cell lung cancer. Clin Cancer Res 2013;19:3591-9. [Crossref] [PubMed]
80. Andersen MB, Harders SW, Thygesen J, et al. Potential impact of texture analysis in contrast enhanced CT in non-small cell lung cancer as a marker of survival: A retrospective feasibility study. Medicine (Baltimore) 2022;101:e31855. [Crossref] [PubMed]
81. Ohri N, Duan F, Snyder BS, et al. Pretreatment 18F-FDG PET Textural Features in Locally Advanced Non-Small Cell Lung Cancer: Secondary Analysis of ACRIN 6668/RTOG 0235. J Nucl Med 2016;57:842-8. [Crossref] [PubMed]
82. Fried DV, Mawlawi O, Zhang L, et al. Stage III Non-Small Cell Lung Cancer: Prognostic Value of FDG PET Quantitative Imaging Features Combined with Clinical Prognostic Factors. Radiology 2016;278:214-22. [Crossref] [PubMed]
83. Desseroit MC, Visvikis D, Tixier F, et al. Development of a nomogram combining clinical staging with (18)F-FDG PET/CT image features in non-small-cell lung cancer stage I-III. Eur J Nucl Med Mol Imaging 2016;43:1477-85. [Crossref] [PubMed]
84. Jimenez Londoño GA, Pérez-Beteta J, Amo-Salas M, et al. Clinicopathologic and metabolic variables from (18)F-FDG PET/CT in the prediction of recurrence pattern in stage I-III non-small cell lung cancer after curative surgery. Ann Nucl Med 2025;39:476-505. [Crossref] [PubMed]
85. Andruska N, Stowe HB, Crockett C, et al. Stereotactic Radiation for Lung Cancer: A Practical Approach to Challenging Scenarios. J Thorac Oncol 2021;16:1075-85. [Crossref] [PubMed]
86. Lovinfosse P, Janvary ZL, Coucke P, et al. FDG PET/CT texture analysis for predicting the outcome of lung cancer treated by stereotactic body radiation therapy. Eur J Nucl Med Mol Imaging 2016;43:1453-60. [Crossref] [PubMed]
87. Wu J, Aguilera T, Shultz D, et al. Early-Stage Non-Small Cell Lung Cancer: Quantitative Imaging Characteristics of (18)F Fluorodeoxyglucose PET/CT Allow Prediction of Distant Metastasis. Radiology 2016;281:270-8. [Crossref] [PubMed]
88. Dong X, Sun X, Sun L, et al. Early Change in Metabolic Tumor Heterogeneity during Chemoradiotherapy and Its Prognostic Value for Patients with Locally Advanced Non-Small Cell Lung Cancer. PLoS One 2016;11:e0157836. [Crossref] [PubMed]
89. Cook GJ, Yip C, Siddique M, et al. Are pretreatment 18F-FDG PET tumor textural features in non-small cell lung cancer associated with response and survival after chemoradiotherapy? J Nucl Med 2013;54:19-26. [Crossref] [PubMed]
90. Ahn SY, Park CM, Park SJ, et al. Prognostic value of computed tomography texture features in non-small cell lung cancers treated with definitive concomitant chemoradiotherapy. Invest Radiol 2015;50:719-25. [Crossref] [PubMed]
91. Fried DV, Tucker SL, Zhou S, et al. Prognostic value and reproducibility of pretreatment CT texture features in stage III non-small cell lung cancer. Int J Radiat Oncol Biol Phys 2014;90:834-42. [Crossref] [PubMed]
92. Virginia BM, Laura F, Silvia R, et al. Prognostic value of histogram analysis in advanced non-small cell lung cancer: a radiomic study. Oncotarget 2018;9:1906-14. [Crossref] [PubMed]
93. Kang SR, Song HC, Byun BH, et al. Intratumoral Metabolic Heterogeneity for Prediction of Disease Progression After Concurrent Chemoradiotherapy in Patients with Inoperable Stage III Non-Small-Cell Lung Cancer. Nucl Med Mol Imaging 2014;48:16-25. [Crossref] [PubMed]
94. Carles M, Fechter T, Radicioni G, et al. FDG-PET Radiomics for Response Monitoring in Non-Small-Cell Lung Cancer Treated with Radiation Therapy. Cancers (Basel) 2021;13:814. [Crossref] [PubMed]
95. Kim DH, Jung JH, Son SH, et al. Prognostic Significance of Intratumoral Metabolic Heterogeneity on 18F-FDG PET/CT in Pathological N0 Non-Small Cell Lung Cancer. Clin Nucl Med 2015;40:708-14. [Crossref] [PubMed]
96. Pellegrino S, Fonti R, Hakkak Moghadam Torbati A, et al. Heterogeneity of Glycolytic Phenotype Determined by (18)F-FDG PET/CT Using Coefficient of Variation in Patients with Advanced Non-Small Cell Lung Cancer. Diagnostics (Basel) 2023;13:2448. [Crossref] [PubMed]
97. Pellegrino S, Fonti R, Vallone C, et al. Coefficient of Variation in Metastatic Lymph Nodes Determined by (18)F-FDG PET/CT in Patients with Advanced NSCLC: Combination with Coefficient of Variation in Primary Tumors. Cancers (Basel) 2024;16:279. [Crossref] [PubMed]
98. Wang P, Xu W, Sun J, et al. A new assessment model for tumor heterogeneity analysis with [18]F-FDG PET images. EXCLI J 2016;15:75-84.
99. Song J, Dong D, Huang Y, et al. Association between tumor heterogeneity and progression-free survival in non-small cell lung cancer patients with EGFR mutations undergoing tyrosine kinase inhibitors therapy. Annu Int Conf IEEE Eng Med Biol Soc 2016;2016:1268-71. [Crossref] [PubMed]
100. Cook GJ, O'Brien ME, Siddique M, et al. Non-Small Cell Lung Cancer Treated with Erlotinib: Heterogeneity of (18)F-FDG Uptake at PET-Association with Treatment Response and Prognosis. Radiology 2015;276:883-93. [Crossref] [PubMed]
101. Memon D, Schoenfeld AJ, Ye D, et al. Clinical and molecular features of acquired resistance to immunotherapy in non-small cell lung cancer. Cancer Cell 2024;42:209-224.e9. [Crossref] [PubMed]
102. Fang W, Jin H, Zhou H, et al. Intratumoral heterogeneity as a predictive biomarker in anti-PD-(L)1 therapies for non-small cell lung cancer. Mol Cancer 2021;20:37. [Crossref] [PubMed]
103. Kim BS, Kang J, Jun S, et al. Association between immunotherapy biomarkers and glucose metabolism from F-18 FDG PET. Eur Rev Med Pharmacol Sci 2020;24:8288-95. [Crossref] [PubMed]
104. Norikane T, Ishimura M, Mitamura K, et al. Texture Features of (18)F-Fluorodeoxyglucose Positron Emission Tomography for Predicting Programmed Death-Ligand-1 Levels in Non-Small Cell Lung Cancer. J Clin Med 2024;13:1625. [Crossref] [PubMed]
105. Chang R, Luo L, Shen C, et al. Using 18F-FDG PET/CT to predict programmed cell death ligand expression in non-small cell lung cancer via metabolic tumour heterogeneity. Br J Radiol 2025;98:715-20. [Crossref] [PubMed]
106. Bracci S, Dolciami M, Trobiani C, et al. Quantitative CT texture analysis in predicting PD-L1 expression in locally advanced or metastatic NSCLC patients. Radiol Med 2021;126:1425-33. [Crossref] [PubMed]
107. Wang D, Zhang X, Liu H, et al. Assessing dynamic metabolic heterogeneity in non-small cell lung cancer patients via ultra-high sensitivity total-body [18F]FDG PET/CT imaging: quantitative analysis of [18F]FDG uptake in primary tumors and metastatic lymph nodes. Eur J Nucl Med Mol Imaging 2022;49:4692-704.
108. Hendriks LEL, Remon J, Faivre-Finn C, et al. Non-small-cell lung cancer. Nat Rev Dis Primers 2024;10:71. [Crossref] [PubMed]
109. Caruso D, Zerunian M, Daffina J, et al. Radiomics and functional imaging in lung cancer: the importance of radiological heterogeneity beyond FDG PET/CT and lung biopsy. Eur J Radiol 2021;142:109874. [Crossref] [PubMed]
110. Suo S, Cheng J, Cao M, et al. Assessment of Heterogeneity Difference Between Edge and Core by Using Texture Analysis: Differentiation of Malignant From Inflammatory Pulmonary Nodules and Masses. Acad Radiol 2016;23:1115-22. [Crossref] [PubMed]
111. Miwa K, Inubushi M, Wagatsuma K, et al. FDG uptake heterogeneity evaluated by fractal analysis improves the differential diagnosis of pulmonary nodules. Eur J Radiol 2014;83:715-9. [Crossref] [PubMed]
112. Qi H, Zuo Z, Lin S, et al. Assessment of intratumor heterogeneity for preoperatively predicting the invasiveness of pulmonary adenocarcinomas manifesting as pure ground-glass nodules. Quant Imaging Med Surg 2025;15:272-86. [Crossref] [PubMed]
113. Li M, Narayan V, Gill RR, et al. Computer-Aided Diagnosis of Ground-Glass Opacity Nodules Using Open-Source Software for Quantifying Tumor Heterogeneity. AJR Am J Roentgenol 2017;209:1216-27. [Crossref] [PubMed]
114. Bianconi F, Palumbo I, Fravolini ML, et al. Texture Analysis on [18F]FDG PET/CT in Non-Small-Cell Lung Cancer: Correlations Between PET Features, CT Features, and Histological Types. Mol Imaging Biol 2019;21:1200-9.
115. Meyer HJ, Schnarkowski B, Pappisch J, et al. CT texture analysis and node-RADS CT score of mediastinal lymph nodes - diagnostic performance in lung cancer patients. Cancer Imaging 2022;22:75. [Crossref] [PubMed]
116. Pham TD, Watanabe Y, Higuchi M, et al. Texture Analysis and Synthesis of Malignant and Benign Mediastinal Lymph Nodes in Patients with Lung Cancer on Computed Tomography. Sci Rep 2017;7:43209. [Crossref] [PubMed]
117. Bayanati H, E, Thornhill R, Souza CA, et al. Quantitative CT texture and shape analysis: can it differentiate benign and malignant mediastinal lymph nodes in patients with primary lung cancer? Eur Radiol 2015;25:480-7. [Crossref] [PubMed]
118. Budiawan H, Cheon GJ, Im HJ, et al. Heterogeneity Analysis of (18)F-FDG Uptake in Differentiating Between Metastatic and Inflammatory Lymph Nodes in Adenocarcinoma of the Lung: Comparison with Other Parameters and its Application in a Clinical Setting. Nucl Med Mol Imaging 2013;47:232-41. [Crossref] [PubMed]
119. Shin SY, Hong IK, Jo YS. Quantitative computed tomography texture analysis: can it improve diagnostic accuracy to differentiate malignant lymph nodes? Cancer Imaging 2019;19:25. [Crossref] [PubMed]
120. Bashir U, Foot O, Wise O, et al. Investigating the histopathologic correlates of 18F-FDG PET heterogeneity in non-small-cell lung cancer. Nucl Med Commun 2018;39:1197-206. [Crossref] [PubMed]
121. Ganeshan B, Goh V, Mandeville HC, et al. Non-small cell lung cancer: histopathologic correlates for texture parameters at CT. Radiology 2013;266:326-36. [Crossref] [PubMed]
122. Castello A, Grizzi F, Toschi L, et al. Tumor heterogeneity, hypoxia, and immune markers in surgically resected non-small-cell lung cancer. Nucl Med Commun 2018;39:636-44. [Crossref] [PubMed]
123. Weiss GJ, Ganeshan B, Miles KA, et al. Noninvasive image texture analysis differentiates K-ras mutation from pan-wildtype NSCLC and is prognostic. PLoS One 2014;9:e100244. [Crossref] [PubMed]
124. Zhang L, Chen B, Liu X, et al. Quantitative Biomarkers for Prediction of Epidermal Growth Factor Receptor Mutation in Non-Small Cell Lung Cancer. Transl Oncol 2018;11:94-101. [Crossref] [PubMed]
125. Shi A, Wang J, Wang Y, et al. Predictive value of multiple metabolic and heterogeneity parameters of (18)F-FDG PET/CT for EGFR mutations in non-small cell lung cancer. Ann Nucl Med 2022;36:393-400. [Crossref] [PubMed]
126. Peng J, Zou D, Zhang X, et al. A novel sub-regional radiomics model to predict immunotherapy response in non-small cell lung carcinoma. J Transl Med 2024;22:87. [Crossref] [PubMed]
127. Liu Z, Duan T, Zhang Y, et al. Radiogenomics: a key component of precision cancer medicine. Br J Cancer 2023;129:741-53. [Crossref] [PubMed]
128. Ju H, Kim K, Kim BI, et al. Graph Neural Network Model for Prediction of Non-Small Cell Lung Cancer Lymph Node Metastasis Using Protein-Protein Interaction Network and (18)F-FDG PET/CT Radiomics. Int J Mol Sci 2024;25:698. [Crossref] [PubMed]
129. Metzenmacher M, Hegedüs B, Forster J, et al. Combined multimodal ctDNA analysis and radiological imaging for tumor surveillance in Non-small cell lung cancer. Transl Oncol 2022;15:101279. [Crossref] [PubMed]
130. Sujit SJ, Aminu M, Karpinets TV, et al. Enhancing NSCLC recurrence prediction with PET/CT habitat imaging, ctDNA, and integrative radiogenomics-blood insights. Nat Commun 2024;15:3152. [Crossref] [PubMed]
131. Hosseini SA, Shiri I, Hajianfar G, et al. Synergistic impact of motion and acquisition/reconstruction parameters on (18) F-FDG PET radiomic features in non-small cell lung cancer: Phantom and clinical studies. Med Phys 2022;49:3783-96. [Crossref] [PubMed]
132. Yan J, Chu-Shern JL, Loi HY, et al. Impact of Image Reconstruction Settings on Texture Features in 18F-FDG PET. J Nucl Med 2015;56:1667-73. [Crossref] [PubMed]
133. Hatt M, Cheze-le Rest C, van Baardwijk A, et al. Impact of tumor size and tracer uptake heterogeneity in (18)F-FDG PET and CT non-small cell lung cancer tumor delineation. J Nucl Med 2011;52:1690-7. [Crossref] [PubMed]
134. Xu Z, Gao M, Papadakis GZ, et al. Joint solution for PET image segmentation, denoising, and partial volume correction. Med Image Anal 2018;46:229-43. [Crossref] [PubMed]
135. Ladbury C, Amini A, Govindarajan A, et al. Integration of artificial intelligence in lung cancer: Rise of the machine. Cell Rep Med 2023;4:100933. [Crossref] [PubMed]