Tumor rim-specific computed tomography radiomics improves prediction of pathological complete response to neoadjuvant chemoimmunotherapy in non-small cell lung cancer

Introduction

In recent years, neoadjuvant therapies based on immunotherapy have made notable progress in resectable non-small cell lung cancer (NSCLC) (1-3). Neoadjuvant immunotherapy leverages the intact immune system to maximize its potency, enabling the elimination of tumor cells and distant micro-metastases. This approach reduces tumor burden prior to surgery, improves surgical outcomes, and lowers the risk of recurrence. In this context, neoadjuvant immunotherapy has transformed treatment strategies and has been progressively integrated into the clinical management of NSCLC (4).

Subsequent analyses building on data from the CheckMate 816 trial (5) and NADIM study (3) have demonstrated that pathological complete response (pCR) is associated with improved survival outcomes in resectable NSCLC (6,7), supporting its role as a surrogate marker for long-term survival. Although neoadjuvant immunotherapy has demonstrated a higher pCR rate compared to neoadjuvant chemotherapy (8,9), only a limited proportion of patients achieved pCR following neoadjuvant immunotherapy. Meanwhile, a more targeted lymphadenectomy approach, which spares lymph nodes essential for immune function, might be more appropriate for those patients with pCR, since evidence has shown that an elevated dissected lymph node count was associated with poorer immunotherapy efficacy in post-resectional recurred NSCLC (10). In such instances, an effective method to predict pCR prior to neoadjuvant immunotherapy is therefore urgently required to optimize treatment strategies and provide more personalized therapies.

Radiomics (11) extracts quantitative data from medical images through high-throughput techniques, enabling advanced analysis beyond human observation limits. In clinical practice, computed tomography (CT) remains the primary imaging method used for evaluating tumor response to immunotherapy in lung cancer patients. Several studies have utilized CT-derived tumor features to develop radiomics models to predict the pathological response following neoadjuvant immunotherapy in NSCLC, with encouraging outcomes (12,13). However, most of these studies predominantly focus on the tumor itself, relying on intratumoral regions to assess pathological responses to treatment. In contrast, the peritumoral microenvironment significantly influences the immunotherapy outcomes (14), including tumor-infiltrating lymphocytes (TILs). T and B cell populations are typically restricted in solid tumors due to immune suppression and their confinement surrounding the tumor microenvironment (15). Emerging evidence has demonstrated that the peritumoral radiomic features were correlated with TILs density and makeup (16-18). Further investigation is needed to explore the impact of peritumor microenvironment specificity on the effectiveness of neoadjuvant immunotherapy.

This retrospective study seeks to establish and validate radiomics models utilizing different volumes of interest (VOIs) around tumor for predicting pCR in resectable NSCLC patients undergoing neoadjuvant immunotherapy. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-259/rc).

Methods

Patients

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Review Board of Tianjin Medical University Cancer Institute and Hospital (No. EK20240001). Because of the retrospective nature, informed consent was waived. Medical records of resectable NSCLC patients who received neoadjuvant immunotherapy at Tianjin Medical University Cancer Institute and Hospital between August 2018 and May 2024 were retrieved through the institutional electronic medical system. Patients were consecutively selected based on specific inclusion criteria: (I) pathologically confirmed NSCLC diagnosed via biopsy before treatment; (II) clinically staged as IB to IIIC; (III) with measurable lung lesions on base-line CT images, defined according to Response Evaluation Criteria in Solid Tumors version 1.1 (RECIST 1.1) as lesions with maximum diameter ≥10 mm; (IV) without epidermal growth factor receptor (EGFR), anaplastic lymphoma kinase (ALK), or other driver gene mutations; (V) receiving neoadjuvant immunotherapy followed by surgery. Patients were excluded if they had any current or previous malignancy other than lung cancer, underwent any treatment for lung cancer before neoadjuvant immunotherapy, unsatisfactory CT image quality due to respiratory artifact, or time interval between chest CT and treatment initiation exceeded one month. A total of 229 patients were included in the analysis, with 137 assigned to the training cohort and 92 to the validation cohort, following a 6:4 allocation ratio. The process of patient selection is presented in Figure 1.

Figure 1 Flowchart illustrating the patient selection process. ALK, anaplastic lymphoma kinase; CT, computed tomography; EGFR, epidermal growth factor receptor; NSCLC, non-small cell lung cancer.

Clinical data [age, sex, smoking history, clinical tumor-node-metastasis (TNM) stage, regimen and cycle of neoadjuvant immunotherapy] and peripheral blood laboratory data (complete blood count with differential and routine tumor markers of lung cancer) at baseline (within 2 weeks before neoadjuvant immunotherapy administration) were extracted from the electronic medical record system. Inflammatory biomarkers such as neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), lymphocyte-to-monocyte ratio (LMR), derived NLR (dNLR) and systemic immune-inflammation index (SII) were calculated. All tumors were staged according to the TNM classification of malignant tumors, ninth edition (19). All enrolled patients received at least two cycles of neoadjuvant immunotherapy, either monotherapy with programmed death-1 (PD-1) immune checkpoint inhibitors or in combination with platinum-based chemotherapy, administered in 21-day cycles. Surgical resection and lymph node dissection were carried out within 4–6 weeks following the completion of neoadjuvant immunotherapy.

Histopathological evaluation

The resected tumor and lymph nodes were examined by two senior pathologists independently to determine the pathological response. The pCR was defined as the absence of residual viable tumor cells in both the primary tumor bed and lymph nodes across all examined slides.

CT acquisition and VOIs segmentation

CT imaging data were collected from the lung apex down to the base. Chest CT scans were performed using equipment from three manufacturers: GE Medical Systems (Chicago, USA; including Discovery CT750 HD, Optima 680 Expert, and Revolution EVO), Siemens (Munich, Germany; SOMATOM Definition AS+ and SOMATOM Drive), and Philips (Amsterdam, Netherlands; IQon spectral CT). Imaging was performed with a tube voltage of 120 kVp and a current ranging from 150 to 200 mA, or automatically adjusted. A 512×512 matrix was used, and reconstruction thicknesses of 1.25 or 1.5 mm were applied with standard kernel algorithms (e.g., Std for GE, Y-detail for Philips, B30f for Siemens). All patients underwent standardized breath-hold training prior to CT examination and were instructed to hold their breath at the end of inspiration during image acquisition to minimize respiratory motion artifacts and ensure consistency across scans. All images were resampled to 1×1×1 mm³ isotropic voxels using linear interpolation, a standard preprocessing step to harmonize spatial resolution across diverse CT scanners (20).

The primary lung tumor was designated as the target lesion for each patient. Gross tumor volume (GTV) refers to the tumor confined within its visible boundary. The tumor rim volume (TRV) is defined as a 6-mm-wide region spanning the tumor border, comprising 3 mm inward into the tumor parenchyma and 3 mm outward into the surrounding tissue. Additionally, the peritumoral volumes (PTVs) are categorized as PTV3 (3 mm beyond the tumor margin) and PTV6 (6 mm beyond the tumor margin), as illustrated in Figure 2. The initial VOIs were delineated on CT images using a lung-specific CT window setting (width, 1,600 HU; level, −600 HU). Tumor borders were identified by a trained radiologist (Reader A), who was blinded to clinical information and pathology results, and GTV was semi-automatically segmented using the Painting tool in 3D Slicer software (v5.6.1; https://download.slicer.org) with threshold settings of −150 to 875 intensity range. Subsequently, TRV, PTV3, and PTV6 were automatically generated based on GTV using RIAS software (v1.0.0; https://riasml.notion.site/RIAS-916ad7256e1e472985d4b11c8ebf0fe0). When the VOI exceeded the lung boundaries (e.g., into the mediastinum or chest wall), manual adjustments were made to ensure the contour stayed within the lung field. To evaluate whether combining intratumoral and peritumoral features offers superior predictive performance, GPTV3 (GTV combined with PTV3) and GPTV6 (GTV combined with PTV6) were constructed by integrating GTV with PTV to create combined VOIs.

Figure 2 Schematic illustrating the overall study workflow. GTV, gross tumor volume; GPTV3, gross tumor volume combined with peritumoral volume of 3 mm beyond the tumor margin; GPTV6, gross tumor volume combined with peritumoral volume of 6 mm beyond the tumor margin; PTV3, peritumoral volume of 3 mm beyond the tumor margin; PTV6, peritumoral volume of 6 mm beyond the tumor margin; RNA, ribonucleic acid; ROC, receiver operating characteristic; TRV, tumor rim volume.

Intra- and inter-observer variability were assessed in 30 randomly selected cases. A second radiologist (Reader B) independently conducted tumor segmentation using the same method. After a 3-month washout period, the initial radiologist (Reader A) repeated tumor segmentation for the same 30 patients. The stability and reproducibility of features were assessed using the intraclass correlation coefficient (ICC).

Feature selection and radiomics signature (Radscore) development

The study schematic is illustrated in Figure 2. From each VOI, 944 radiomics features were extracted using FeAture Explorer (FAE) software (v0.5.16), an open-source tool built on PyRadiomics (21). Features were derived from original images and processed with Wavelet Transform and Laplacian of Gaussian. A comprehensive list of features is provided in Table S1. Only features with ICC ≥0.75 were included in subsequent analyses. To reduce dimensionality, one feature from each pair with a Pearson correlation coefficient (PCC) >0.990 was randomly removed. Recursive feature elimination (RFE) was employed for feature selection. A Radscore was constructed using linear discriminant analysis (LDA), a method that applies Bayes’ rule to fit class-conditional densities, followed by 10-fold cross-validation. After dimensionality reduction and feature selection, combinations of two to nine features were evaluated in the training cohort. Model performance was evaluated using receiver operating characteristic (ROC) curve analysis, and the signature exhibiting the highest area under the curve (AUC) in the validation cohort was chosen as the optimal model. Finally, SHapley Additive exPlanations (SHAP) was used to interpret and visualize the prediction process of the selected model.

Validation in a neoadjuvant chemotherapy cohort

To further evaluate the treatment specificity of the optimal model, we incorporated an independent validation cohort comprising patients with resectable NSCLC who received neoadjuvant chemotherapy exclusively. This cohort was recruited during the same time period and from the same medical center with comparable baseline clinical characteristics.

Biologic basis exploration

To investigate the biological mechanisms underlying the optimal radiomics model, we performed genetic analysis on 36 patients with clinical TNM stage IB–IIIC NSCLC from The Cancer Imaging Archive (TCIA) dataset (22). These patients had both RNA-sequencing data and high-quality preoperative CT images. Using the optimal cutoff from the maximum Youden index in the training set, patients were classified into high- and low-Radscore subgroups. Differential gene expression profiling was conducted using limma (v3.62.1), applying a log fold change threshold of >1 and P value <0.05. Gene set enrichment analysis (GSEA) and single-sample GSEA (ssGSEA) were then performed to examine genetic pathway heterogeneity and tumor-infiltrating immune cell composition across subgroups.

Statistical analysis

For continuous variables, the Student’s t-test or Mann-Whitney test was applied, whereas categorical variables were analyzed with the chi-square or Fisher’s exact test. Subsequently, univariate logistic regression was conducted on each variable, with those showing P value <0.05 incorporated into the multivariate logistic regression model. The ROC curves were generated using the pROC package (v1.18.5), with the cutoff selected to maximize the Youden index. To assess model performance, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were computed. AUCs were compared using the DeLong test, and model calibration was assessed with the Hosmer-Lemeshow test (23) based on 1,000 bootstrap samples. Decision curve analysis (DCA) was employed to verify the clinical value of the optimal model by evaluating net benefit at different threshold probabilities.

To assess the incremental value of the radiomics model over an existing model, integrated discrimination improvement (IDI) and net reclassification index (NRI) were calculated at cutoffs of 0, 0.4, and 1, using the PredictABEL package (v1.2). All statistical analyses were performed using SPSS (v27.0, IBM SPSS Statistics) and R (v4.4.1, http://www.R-project.org). P values were derived from two-tailed tests with a significance level of 0.05.

Results

Baseline information of the study cohorts

The clinicopathological characteristics of the 229 enrolled patients are summarized in Table 1. The mean age of the entire cohort was 62 years and 83.8% (n=192) of participants were male. In terms of histologic type, squamous cell carcinoma accounted for the largest proportion (n=159, 69.4%), followed by adenocarcinoma (n=56, 24.5%). With respect to pretreatment clinical staging, the majority of patients were evaluated to be T2 (n=86, 37.6%), N2 (n=106, 46.3%) and stage III (n=156, 68.1%) disease. Most of the patients (n=224, 97.8%) received 2–4 cycles of neoadjuvant immunotherapy regimens. R0 resections were obtained for all included patients, and pCR was achieved in 97 patients, resulting in a pCR rate of 42.4%. No remarkable difference was observed in the distribution of patients with and without pCR between the training and validation cohorts. Similarly, baseline clinical and pathological characteristics did not differ significantly between these two cohorts (all P>0.05).

In the training cohort, pCR was more likely to occur in male patients (P=0.02), and dNLR was slightly lower in pCR group compared to non-pCR group (P=0.03) (Table 2). Multivariable logistic regression analysis identified sex of male [odds ratio (OR) =3.323; 95% confidence interval (CI): 1.126, 9.802; P=0.03] and lower dNLR (OR =0.000; 95% CI: 0.000, 0.719; P=0.04) as independent predictors of pCR after neoadjuvant immunotherapy (Table S2). Based on these two clinical predictors, a clinical model incorporating sex and dNLR was constructed. In the validation cohort, the clinical model showed an AUC of 0.579 (95% CI: 0.466–0.694) (Figure 3), with an accuracy of 56.5%, sensitivity of 45.0%, specificity of 65.4%, PPV of 50.0%, and NPV of 60.7% (Table 3).

Figure 3 ROC curves and radar chart visualization evaluating discriminative ability of different prediction models in the training and validation cohorts. AUC, area under the curve; GPTV3, gross tumor volume combined with peritumoral volume of 3 mm beyond the tumor margin; GPTV6, gross tumor volume combined with peritumoral volume of 6 mm beyond the tumor margin; GTV, gross tumor volume; NPV, negative predictive value; PPV, positive predictive value; PTV3, peritumoral volume of 3 mm beyond the tumor margin; PTV6, peritumoral volume of 6 mm beyond the tumor margin; ROC, receiver operating characteristic; TRV, tumor rim volume.

Radiomics feature selection and model development

Of the 944 radiomic features extracted from each VOI, 786, 651, 697 and 822 features were initially retained for TRV, GTV, PTV3 and PTV6, respectively, based on high stability (inter-observer and intra-observer ICC ≥0.75). Following the combination of GTV with PTV3 and PTV6, GPTV3 and GPTV6 contained 1,348 and 1,473 features, respectively. Thereafter, 573, 500, 509, 616, 1,005 and 1,113 features were retained in the TRV, GTV, PTV3, PTV6, GPTV3 and GPTV6 using PCC. Six radiomics models were constructed using the RFE algorithm and LDA classifier, the top-ranked features were selected. Specifically, the TRV model with nine features, the GTV model with seven features, the PTV3 model with six features, the PTV6 model with eight features, the GPTV3 model with six features, and the GPTV6 model with nine features were selected as the optimal model for each VOI (Tables S3-S8). Although several volume-related features, including those directly reflecting VOI size, were initially included in the feature pool, these features were not retained in the final predictive radiomics model following feature selection, indicating their limited association with treatment response in this cohort. Of note, the Radscore was higher in the pCR group compared to the non-pCR group across all radiomics models (all P<0.05) (Figure S1).

Comparative analysis of predictive model efficiency

The TRV radiomics model yielded an AUC of 0.827 (95% CI: 0.742–0.913) in predicting pCR in the validation cohort, outperforming the clinical model and all other radiomics models (Figure 3). The Delong test revealed a significant improvement in AUC of the TRV model compared to other models (Table S9), except for the GPTV6 radiomics model (P=0.08). Using the optimal cutoff, accuracy, sensitivity, specificity, PPV and NPV of the TRV radiomics model were 76.1%, 72.5%, 78.9%, 72.5% and 78.9%, respectively, which demonstrated the highest predictive performance (Table 3 and Figure S2). Additionally, the SHAP bar chart (Figure 4A) illustrates the contribution of the nine features in the TRV model, and the SHAP violin chart (Figure 4B) highlights the impact of each feature on the prediction probabilities.

Figure 4 Visualization of the TRV model using SHAP. (A) SHAP bar chart displaying the weights of features in the TRV model. (B) SHAP violin plot illustrating the positive and negative impacts of each feature on prediction probability. GLCM, gray-level co-occurrence matrix; GLDM, gray-level dependence matrix; GLRLM, gray-level run-length matrix; GLSZM, gray-level size zone matrix; LoG, Laplacian of Gaussian; MCC, maximal correlation coefficient; SHAP, SHapley Additive exPlanations; TRV, tumor rim volume.

Subgroup analysis and clinical usefulness of radiomics models

To further assess the robustness of the predictive model, subgroup analysis was performed based on treatment regimen in the entire cohort. TRV radiomics model demonstrated relatively stable and satisfactory predictive performance (Figure S3) and Delong test indicated no significant differences between different treatment regimen subgroups (P=0.75).

The calibration curve of all radiomics models in the training cohort showed good agreement between the predicted and actual outcome of pCR (Figure 5A), and Hosmer-Lemeshow test for goodness-of-fit showed no significant discrepancy between the observed and expected values (all P>0.05) in the training and validation cohorts. In the validation cohort, TRV radiomics model demonstrated higher net benefits than other models at most risk thresholds (Figure 5B). Additionally, the enhanced ability of the TRV radiomics model to identify pCR was further confirmed by NRI tests (all P<0.05, except for comparisons with the PTV6 and GPTV6 radiomics models) and IDI tests (all P<0.05, except for the comparison with the GPTV6 radiomics model) (Table 4). Two representative cases diagnosed using the TRV radiomics model are displayed in Figure 6.

Figure 5 Performance assessment of different prediction models in the training and validation cohorts. (A) Calibration curves comparing predicted and observed outcomes. (B) DCA curves assessing clinical utility. DCA, decision curve analysis; GPTV3, gross tumor volume combined with peritumoral volume of 3 mm beyond the tumor margin; GPTV6, gross tumor volume combined with peritumoral volume of 6 mm beyond the tumor margin; GTV, gross tumor volume; PTV3, peritumoral volume of 3 mm beyond the tumor margin; PTV6, peritumoral volume of 6 mm beyond the tumor margin; TRV, tumor rim volume.

Figure 6 Examples of application of the TRV radiomics model in two patients with comparable clinical profiles. A 57-year-old man with adenocarcinoma in the right upper lobe (Patient A). Restaging with CT after four cycles of chemoimmunotherapy revealed a 31.3% reduction in total tumor burden, which indicates PR according to RECIST. Postoperative specimens revealed pCR. Based on the TRV model prediction result, this patient has a high probability of achieving pCR, which is consistent with pathological findings. The clinical model predicts a pCR probability of 0.462 for this patient, which is below the cutoff value, classifying the patient as non-pCR. Another 51-year-old man with squamous cell carcinoma in the right lower lobe (Patient B). The patient underwent three cycles of chemoimmunotherapy, and he was evaluated by CT as having PR. Postoperative specimens revealed a major pathologic response (9% residual viable tumor in the primary tumor). According to the TRV model, the patient has a low likelihood of achieving pCR. This predictive outcome aligns with the pathological findings, whereas the clinical model predicts that the patient could achieve pCR. ADC, adenocarcinoma; CT, computed tomography; pCR, pathological complete response; PR, partial response; RECIST, Response Evaluation Criteria in Solid Tumors; SCC, squamous cell carcinoma; TRV, tumor rim volume.

Model performance in the neoadjuvant chemotherapy validation cohort

In the neoadjuvant chemotherapy cohort (n=62), the pCR rate was low, with only 5 patients (8.1%) achieving pCR, which is significantly lower than that observed in the neoadjuvant chemoimmunotherapy cohort (P<0.001). There were no significant differences in baseline clinical characteristics between these two cohorts (Table S10). The TRV model exhibited limited predictive ability in the neoadjuvant chemotherapy validation cohort, with an AUC of 0.519 (95% CI: 0.281–0.758), accuracy of 56.5%, sensitivity of 40.0%, and specificity of 57.9%. These metrics are significantly lower than those observed in the neoadjuvant chemoimmunotherapy validation cohort.

Biologic basis of radiomics prediction

Based on the optimal cutoff determined for TRV radiomics model in the training cohort, 36 patients from TCIA dataset were stratified into different groups, 24 patients as low-Radscore and the remaining 12 patients as high-Radscore. The volcano plot (Figure 7A) illustrates the differential gene expression between low- and high-Radscore groups, highlighting 83 genes upregulated and 282 genes downregulated in the high-Radscore group. In the GSEA analysis (Figure 7B), immune regulation, signal transduction, and metabolic processes were significantly upregulated in patients with high TRV radiomics scores, indicating active alterations in the tumor microenvironment and immune response of the high-score group. Additionally, analysis of the tumor microenvironment reveals greater infiltration of central memory CD8⁺ T cells, effector memory CD4⁺ T cells, natural killer cells, eosinophils, and mast cells in tumors with high-Radscore (Figure 7C).

Figure 7 Biologic basis based on TRV radiomics model. (A) Volcano plot depicting differential gene expressions across low- and high-Radscore groups. (B) Ridge plots illustrating the top 10 upregulated pathways enriched in the set in differentially expressed genes for high-Radscore patients. (C) Box plot comparing immune cell type abundances between low- and high-Radscore groups. “ns” indicates no significant difference, and “*” indicates a significant difference (P<0.05). BMP, bone morphogenetic protein; FC, fold change; MDSC, myeloid-derived suppressor cells; NES, normalized enrichment score; pCR, pathological complete response; PD-1, programmed death-1; TCR, T cell receptor; TRV, tumor rim volume.

Discussion

This study quantitatively analyzed different VOIs from CT images encompassing the intratumoral region, peritumoral region, and the tumor invasion margin transition zone, for predicting pCR in NSCLC patients treated with neoadjuvant immunotherapy. Our findings demonstrated that the optimal tumor invasion margin transition zone (TRV) provides superior predictive performance, achieving the highest AUC compared to all evaluated models.

Previous studies have explored intratumoral radiomics models based on GTV for predicting pathological responses to neoadjuvant immunotherapy using pre-treatment CT images (24-27). In our study, although the GTV radiomics model could screen out the patients with pCR to a certain extent, the sensitivity only reached 52.5% in the validation cohort, indicating its limited ability to accurately identify patients achieving complete pathological remission. This suggests that relying solely on the intratumoral region for radiomics analysis may fail to capture crucial biological and immunological interactions occurring in the surrounding tumor microenvironment. To address these issues, we extended our analysis to peritumoral regions, given their recognized importance in immune activity. Notably, peritumoral radiomics models have been successfully used to predict major pathological response (MPR), achieving improved performance over GTV models with higher AUCs (range, 0.662 to 0.741) and sensitivity (range, 0.670 to 0.947) (28). In the current study, we evaluated peritumoral models at different expanding volumes (PTV3 and PTV6) as well as combined intratumoral-peritumoral models (e.g., GPTV3 and GPTV6). While these models demonstrated some theoretical potential, PTV3 and PTV6 radiomics models still could not accurately imply pCR after neoadjuvant immunotherapy. This may be due to manual segmentation variability, where the expanded peritumoral regions (3 and 6 mm) may not accurately correspond to the true pathological tumor margins, leading to potential mismatches between the extracted radiomics features and the biologically relevant tumor microenvironment. Additionally, combining intratumoral and peritumoral features (e.g., GPTV3 and GPTV6) did not lead to significant performance improvements, suggesting inherent limitations in feature robustness of these models.

A prior investigation quantifying microscopic extension (ME) in surgical resection specimens revealed that NSCLC tended to infiltrate into adjacent tissues, with adenocarcinoma exhibiting an average extension of 2.69 mm and squamous cell carcinoma exhibiting an average extension of 1.48 mm (29). Based on these findings, microscopic tumor extension is approximately 3 mm beyond the radiologic tumor margin to encompass the majority of histologic types in the current cohort. A previous histopathological study have identified the tumor rim, which encompasses both the invasive front and the adjacent peritumoral lung tissue, as a biologically aggressive region where cancerous islets and activated microvasculature are frequently observed (30). This zone often represents the actual true leading edge of tumor infiltration and is critical for evaluating tumor aggressiveness and treatment response. With evidence that the region adjacent to the visible tumor edge might provide information of tumor aggressiveness (31), we attempted to extract radiomics features from an annular region surrounding the tumor margin that encompassed tumor-normal interface, which was defined as the area extending 3 mm both inside and outside the tumor boundary. The TRV model demonstrated superior predictive power, with a significantly higher AUC compared to other models in the validation cohort (Delong test, P<0.05 for all comparisons except GPTV6). TRV capitalizes on the distinct biological characteristics of this region, which include immune cell infiltration, inflammatory responses, and microvascular changes (32,33). Studies have further linked peritumoral texture features to TILs, which are strongly associated with enhanced immune system activation and tumor suppression (34,35). Given that immunotherapy primarily works by modulating the tumor microenvironment to boost immune recognition and attack of tumor cells (36), TRV effectively captures this immune response, positioning it as a robust biomarker for predicting pCR. Furthermore, the peritumoral region, with higher vascular permeability and greater exposure, is more likely to reveal pathological responses (37).

In addition, the interpretability of the TRV model is enhanced by the explainability of its key radiomics features. We further investigated the most contributive features based on SHAP analysis. The most contributive feature, Laplacian of Gaussian (LoG) gray-level dependence matrix (GLDM) DependenceEntropy, quantifies the complexity and heterogeneity of local intensity dependencies, which may reflect microenvironmental changes such as immune cell infiltration and extracellular matrix remodeling triggered by immunotherapy. The second and third most important features, LoG gray-level run-length matrix (GLRLM) GrayLevelNonUniformityNormalized and LoG GLRLM RunEntropy, represent the variability and randomness of gray-level distributions, which may correspond to the spatial heterogeneity of immune responses and vascular alterations in the tumor rim. These high-contributing features further support the hypothesis that the tumor rim captures critical biological signals that reflect treatment response.

Systemic inflammation contributes significantly to tumor development and progression by promoting genetic mutations, enhancing angiogenesis, and affecting immune surveillance and treatment response (38,39). Various inflammation-related hematological markers are recognized as reliable indicators of immunotherapy response and patient prognosis across different cancers (40-43). Alessi et al. (44) found that patients with low pretreatment dNLR exhibited a distinct immune tumor microenvironment, characterized by higher counts of tumor-associated CD8⁺, FOXP3⁺, PD-1⁺ immune cells, and PD-1⁺ CD8⁺ T cells. This might explain the underlying biological mechanisms by which dNLR serves as a biomarker. Similarly, Li et al. (45) observed that patients with MPR had significantly lower on-treatment dNLR levels in their study cohort. In our study, patients exhibiting lower dNLR levels had a higher likelihood of achieving pCR, which aligns with previous research findings. The clinical model showed limited predictive ability, possibly due to the small number of available clinical factors associated with pCR. In contrast, the TRV radiomics model could non-invasively capture complex spatial and textural features of the tumor and peritumoral regions, offering superior prognostic value.

The main challenge in radiomics is feature interpretability, with tumor-specific metabolic and immune changes hypothesized to link to oncological outcomes. To investigate the underlying mechanisms behind the predictive ability of the Radscore for pCR, gene analyses were performed utilizing public dataset. The TRV radiomics features were linked to upregulated PD-1 signaling pathways. PD-1, an immune checkpoint receptor on T cells, inhibits TCR signaling, enabling tumor immune evasion and progression (46). In our study, all enrolled patients received neoadjuvant immunotherapy regimens based on PD-1 immune checkpoint inhibitors. The observed PD-1 pathway upregulation is therefore biologically relevant to the therapeutic mechanism of the administered treatments. This upregulation may reflect enhanced PD-1 pathway engagement, potentially indicating restored T-cell activation and improved anti-tumor immune responses. These findings align with the CheckMate 159 study (47), where post-nivolumab treatment showed increased CD8⁺ and PD-1⁺ immune cell infiltration in patients achieving MPR. Additionally, pathways such as “Translocation of ZAP-70 to the Immunological Synapse” and “Phosphorylation of CD3 and TCR zeta chains” (48,49) were identified as playing critical roles in T-cell activation and proliferation. Their upregulation indicates an enhanced response of T cells to tumor antigens, contributing to stronger anti-tumor immune activation. These observations, based on exploratory genomic analyses from a publicly available cohort independent of our radiomics dataset, collectively elucidate the potential biological mechanisms through which the TRV radiomics framework demonstrates prognostic capability in assessing pathological responses to neoadjuvant immunotherapy.

This study has several limitations. Primarily, the retrospective nature of the analysis introduces potential selection bias, which could impact the robustness and replicability of the results. Secondly, variability in CT scanners and image acquisition protocols reflects diverse clinical practices, which may enhance the model’s applicability in real-world settings but could also affect the results. CT images were standardized through preprocessing steps to minimize scanner-related variability. Finally, the limited sample size and single-center design restrict the broader applicability of these results. Follow-up studies involving larger cohorts and multi-center collaboration are planned to confirm these findings.

Conclusions

In summary, the TRV-based radiomics model, focusing on the 6 mm tumor rim, shows superior accuracy in predicting pCR, positioning it as a valuable non-invasive imaging biomarker for personalized neoadjuvant immunotherapy in NSCLC patients. The model’s predictive capability is supported by underlying biological pathways that are related to immune regulation, cellular metabolism, and enhanced antitumor immune infiltration within the tumor microenvironment.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-259/rc

Data Sharing Statement: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-259/dss

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-259/prf

Funding: This study was supported by the National Natural Science Foundation of China (grant No. 82472063), The Chinese National Key Research and Development Project (grant No. 2021YFC2500400 and grant No. 2021YFC2500402), and Tianjin Key Medical Discipline (Specialty) Construction Project (grant No. TJYXZDXK-010A).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-259/coif). The 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. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Review Board of Tianjin Medical University Cancer Institute and Hospital (No. EK20240001) and individual consent for this retrospective analysis was waived.

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. Shu CA, Gainor JF, Awad MM, et al. Neoadjuvant atezolizumab and chemotherapy in patients with resectable non-small-cell lung cancer: an open-label, multicentre, single-arm, phase 2 trial. Lancet Oncol 2020;21:786-95. [Crossref] [PubMed]
2. O'Brien M, Paz-Ares L, Marreaud S, et al. Pembrolizumab versus placebo as adjuvant therapy for completely resected stage IB-IIIA non-small-cell lung cancer (PEARLS/KEYNOTE-091): an interim analysis of a randomised, triple-blind, phase 3 trial. Lancet Oncol 2022;23:1274-86. [Crossref] [PubMed]
3. Provencio M, Nadal E, Insa A, et al. Neoadjuvant chemotherapy and nivolumab in resectable non-small-cell lung cancer (NADIM): an open-label, multicentre, single-arm, phase 2 trial. Lancet Oncol 2020;21:1413-22. [Crossref] [PubMed]
4. National Comprehensive Cancer Network. Non-Small Cell Lung Cancer (Version 1.2025). Accessed December 20, 2024. Available online: https://www.nccn.org/professionals/physician_gls/pdf/nscl.pdf
5. Forde PM, Spicer J, Lu S, et al. Neoadjuvant Nivolumab plus Chemotherapy in Resectable Lung Cancer. N Engl J Med 2022;386:1973-85. [Crossref] [PubMed]
6. Deutsch JS, Cimino-Mathews A, Thompson E, et al. Association between pathologic response and survival after neoadjuvant therapy in lung cancer. Nat Med 2024;30:218-28. [Crossref] [PubMed]
7. Provencio M, Nadal E, Insa A, et al. Perioperative chemotherapy and nivolumab in non-small-cell lung cancer (NADIM): 5-year clinical outcomes from a multicentre, single-arm, phase 2 trial. Lancet Oncol 2024;25:1453-64. [Crossref] [PubMed]
8. Banna GL, Hassan MA, Signori A, et al. Neoadjuvant Chemo-Immunotherapy for Early-Stage Non-Small Cell Lung Cancer: A Systematic Review and Meta-Analysis. JAMA Netw Open 2024;7:e246837. [Crossref] [PubMed]
9. Yu S, Zhai S, Gong Q, et al. Neoadjuvant Immunotherapy and Non-Small Cell Lung Cancer: A Systematic Review and Meta-analysis of Randomized Controlled Trials. Am J Clin Oncol 2023;46:517-28. [Crossref] [PubMed]
10. Deng H, Zhou J, Chen H, et al. Impact of lymphadenectomy extent on immunotherapy efficacy in postresectional recurred non-small cell lung cancer: a multi-institutional retrospective cohort study. Int J Surg 2024;110:238-52. [Crossref] [PubMed]
11. Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012;48:441-6. [Crossref] [PubMed]
12. 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]
13. She Y, He B, Wang F, et al. Deep learning for predicting major pathological response to neoadjuvant chemoimmunotherapy in non-small cell lung cancer: A multicentre study. EBioMedicine 2022;86:104364. [Crossref] [PubMed]
14. Zhang S, Regan K, Najera J, et al. The peritumor microenvironment: physics and immunity. Trends Cancer 2023;9:609-23. [Crossref] [PubMed]
15. Salmon H, Franciszkiewicz K, Damotte D, et al. Matrix architecture defines the preferential localization and migration of T cells into the stroma of human lung tumors. J Clin Invest 2012;122:899-910. [Crossref] [PubMed]
16. Khorrami M, Prasanna P, Gupta A, et al. Changes in CT Radiomic Features Associated with Lymphocyte Distribution Predict Overall Survival and Response to Immunotherapy in Non-Small Cell Lung Cancer. Cancer Immunol Res 2020;8:108-19. [Crossref] [PubMed]
17. Ramlee S, Hulse D, Bernatowicz K, et al. Radiomic Signatures Associated with CD8(+) Tumour-Infiltrating Lymphocytes: A Systematic Review and Quality Assessment Study. Cancers (Basel) 2022;14:3656. [Crossref] [PubMed]
18. Han X, Guo Y, Ye H, et al. Development of a machine learning-based radiomics signature for estimating breast cancer TME phenotypes and predicting anti-PD-1/PD-L1 immunotherapy response. Breast Cancer Res 2024;26:18. [Crossref] [PubMed]
19. Asamura H, Nishimura KK, Giroux DJ, et al. IASLC Lung Cancer Staging Project: The New Database to Inform Revisions in the Ninth Edition of the TNM Classification of Lung Cancer. J Thorac Oncol 2023;18:564-75.
20. van Timmeren JE, Cester D, Tanadini-Lang S, et al. Radiomics in medical imaging-"how-to" guide and critical reflection. Insights Imaging 2020;11:91. [Crossref] [PubMed]
21. Song Y, Zhang J, Zhang YD, et al. FeAture Explorer (FAE): A tool for developing and comparing radiomics models. PLoS One 2020;15:e0237587. [Crossref] [PubMed]
22. Data for NSCLC Radiogenomics (Version 4). The Cancer Imaging Archive. Accessed October 5, 2024. Available online: https://doi.org/10.7937/K9/TCIA.2017.7hs46erv
23. Huang Y, Li W, Macheret F, et al. A tutorial on calibration measurements and calibration models for clinical prediction models. J Am Med Inform Assoc 2020;27:621-33. [Crossref] [PubMed]
24. 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]
25. Yang N, Yue HL, Zhang BH, et al. Predicting pathological response to neoadjuvant or conversion chemoimmunotherapy in stage IB-III non-small cell lung cancer patients using radiomic features. Thorac Cancer 2023;14:2869-76. [Crossref] [PubMed]
26. Wang F, Yang H, Chen W, et al. A combined model using pre-treatment CT radiomics and clinicopathological features of non-small cell lung cancer to predict major pathological responses after neoadjuvant chemoimmunotherapy. Curr Probl Cancer 2024;50:101098. [Crossref] [PubMed]
27. Han X, Wang M, Zheng Y, et al. Delta-radiomics features for predicting the major pathological response to neoadjuvant chemoimmunotherapy in non-small cell lung cancer. Eur Radiol 2024;34:2716-26. [Crossref] [PubMed]
28. Huang D, Lin C, Jiang Y, et al. Radiomics model based on intratumoral and peritumoral features for predicting major pathological response in non-small cell lung cancer receiving neoadjuvant immunochemotherapy. Front Oncol 2024;14:1348678. [Crossref] [PubMed]
29. Giraud P, Antoine M, Larrouy A, et al. Evaluation of microscopic tumor extension in non-small-cell lung cancer for three-dimensional conformal radiotherapy planning. Int J Radiat Oncol Biol Phys 2000;48:1015-24. [Crossref] [PubMed]
30. Koukourakis MI, Giatromanolaki A, Thorpe PE, et al. Vascular endothelial growth factor/KDR activated microvessel density versus CD31 standard microvessel density in non-small cell lung cancer. Cancer Res 2000;60:3088-95.
31. Liu K, Li K, Wu T, et al. Improving the accuracy of prognosis for clinical stage I solid lung adenocarcinoma by radiomics models covering tumor per se and peritumoral changes on CT. Eur Radiol 2022;32:1065-77. [Crossref] [PubMed]
32. Wang JH, Wahid KA, van Dijk LV, et al. Radiomic biomarkers of tumor immune biology and immunotherapy response. Clin Transl Radiat Oncol 2021;28:97-115. [Crossref] [PubMed]
33. Lamplugh Z, Fan Y. Vascular Microenvironment, Tumor Immunity and Immunotherapy. Front Immunol 2021;12:811485. [Crossref] [PubMed]
34. Parra ER, Behrens C, Rodriguez-Canales J, et al. Image Analysis-based Assessment of PD-L1 and Tumor-Associated Immune Cells Density Supports Distinct Intratumoral Microenvironment Groups in Non-small Cell Lung Carcinoma Patients. Clin Cancer Res 2016;22:6278-89. [Crossref] [PubMed]
35. Presti D, Dall'Olio FG, Besse B, et al. Tumor infiltrating lymphocytes (TILs) as a predictive biomarker of response to checkpoint blockers in solid tumors: A systematic review. Crit Rev Oncol Hematol 2022;177:103773. [Crossref] [PubMed]
36. Murciano-Goroff YR, Warner AB, Wolchok JD. The future of cancer immunotherapy: microenvironment-targeting combinations. Cell Res 2020;30:507-19. [Crossref] [PubMed]
37. Vaidya P, Bera K, Gupta A, et al. CT derived radiomic score for predicting the added benefit of adjuvant chemotherapy following surgery in Stage I, II resectable Non-Small Cell Lung Cancer: a retrospective multi-cohort study for outcome prediction. Lancet Digit Health 2020;2:e116-28. [Crossref] [PubMed]
38. Crusz SM, Balkwill FR. Inflammation and cancer: advances and new agents. Nat Rev Clin Oncol 2015;12:584-96. [Crossref] [PubMed]
39. Galdiero MR, Marone G, Mantovani A. Cancer Inflammation and Cytokines. Cold Spring Harb Perspect Biol 2018;10:a028662. [Crossref] [PubMed]
40. Lei Y, Cao C, Tang R, et al. Peripheral blood inflammatory biomarkers neutrophil/ lymphocyte ratio, platelet/lymphocyte ratio and systemic immune-inflammation index/albumin ratio predict prognosis and efficacy in non-small cell lung cancer patients receiving immunotherapy and opioids. BMC Cancer 2025;25:664. [Crossref] [PubMed]
41. Hirasawa Y, Kubota Y, Mura E, et al. Maximum Efficacy of Immune Checkpoint Inhibitors Occurs in Esophageal Cancer Patients With a Low Neutrophil-to-Lymphocyte Ratio and Good Performance Status Prior to Treatment. Anticancer Res 2024;44:3397-407. [Crossref] [PubMed]
42. Liao S, Sun H, Lu H, et al. Neutrophil-to-lymphocyte ratio-based prognostic score can predict outcomes in patients with advanced non-small cell lung cancer treated with immunotherapy plus chemotherapy. BMC Cancer 2025;25:697. [Crossref] [PubMed]
43. Li Y, Xu J, Zhang L, et al. The association of immune-inflammation indices at multiple time points with treatment response and survival in advanced non-small cell lung cancer patients receiving immune checkpoint inhibitors. World J Surg Oncol 2025;23:195. [Crossref] [PubMed]
44. Alessi JV, Ricciuti B, Alden SL, et al. Low peripheral blood derived neutrophil-to-lymphocyte ratio (dNLR) is associated with increased tumor T-cell infiltration and favorable outcomes to first-line pembrolizumab in non-small cell lung cancer. J Immunother Cancer 2021;9:e003536. [Crossref] [PubMed]
45. Li C, Wu J, Jiang L, et al. The predictive value of inflammatory biomarkers for major pathological response in non-small cell lung cancer patients receiving neoadjuvant chemoimmunotherapy and its association with the immune-related tumor microenvironment: a multi-center study. Cancer Immunol Immunother 2023;72:783-94. [Crossref] [PubMed]
46. Gao M, Shi J, Xiao X, et al. PD-1 regulation in immune homeostasis and immunotherapy. Cancer Lett 2024;588:216726. [Crossref] [PubMed]
47. Forde PM, Chaft JE, Smith KN, et al. Neoadjuvant PD-1 Blockade in Resectable Lung Cancer. N Engl J Med 2018;378:1976-86. [Crossref] [PubMed]
48. Blanchard N, Di Bartolo V, Hivroz C. In the immune synapse, ZAP-70 controls T cell polarization and recruitment of signaling proteins but not formation of the synaptic pattern. Immunity 2002;17:389-99. [Crossref] [PubMed]
49. Steinberg M, Adjali O, Swainson L, et al. T-cell receptor-induced phosphorylation of the zeta chain is efficiently promoted by ZAP-70 but not Syk. Blood 2004;104:760-7. [Crossref] [PubMed]