Enhancing the prediction of KRAS mutation status in Asian lung adenocarcinoma: a comprehensive approach combining clinical, dual-energy spectral computed tomography, and radiomics features

Introduction

Lung adenocarcinoma (LUAD), a predominant sub-type of non-small cell lung cancer (NSCLC), remains at the forefront of oncological research due to its global prevalence and the complex interplay of its genetic alterations (1). These alterations, often associated with specific clinical outcomes and therapeutic responses, play a crucial role in determining patients’ prognosis and overall survival rates. Among the myriad of genetic alterations associated with LUAD, the Kirsten rat sarcoma viral oncogene homolog (KRAS) mutations have been of particular interest to researchers and clinicians alike (2). KRAS is the common oncogene alteration of LUAD with smoking-history and happens in 20–30% Western and less than 10% in Asian patients with LUAD (2-4). For a long time, the KRAS mutation is associated with bad prognosis and tyrosine kinase inhibitors (TKIs) resistance, which has been considered unresponsive to targeted therapy (2). In recent years, targeted drugs for KRAS mutation have been a research hotspot and sotorasib and adagrasib show encouraging anticancer activity in patients with NSCLC harboring the KRAS p.G12C mutation (5-7). As such, the quest for accurate, efficient, and minimally or non-invasive detection methods for KRAS mutations has become a focal point in the realm of lung cancer research (4).

Historically, breakthroughs and challenges have marked the journey of understanding and detecting KRAS mutations in LUAD (5). In the early days, the primary detection method was through invasive tissue biopsies. While this method offers direct insights into the tumor’s genetic composition, it has drawbacks. The invasive nature of biopsies often leads to complication, ranging from infections to more severe medical conditions (8). Furthermore, the procedure’s invasiveness often deters patients, leading to delays in diagnosis and treatment (9). The inherent spatial heterogeneity of tumors further complicates the biopsy approach (10). A single biopsy might only represent a fraction of the tumor’s genetic landscape, leading to potential misdiagnoses (11). This spatial variability means that while one part of the tumor might harbor the KRAS mutation, another might not, leading to false negatives (12). With the advent of molecular biology and genomics, new methods of detecting genetic alterations have begun to emerge. Liquid biopsies, which analyze circulating tumor DNA (ctDNA) in the blood, have become a beacon of hope for non-invasive genetic alterations detection. The premise is simple yet revolutionary: if tumors shed their DNA into the bloodstream, a simple blood test could detect these genetic alterations. However, the reality has been proven more complex. The sensitivity of liquid biopsies is often compromised due to the low abundance of ctDNA, especially in the early stages of cancer (13).

The relationship between imaging features and LUAD oncogene phenotypes have been of particular interest to clinical doctors, even radiologists. Medical imaging-especially computed tomography (CT) is an overarching method for lung cancer screening and diagnosis. However, only a few studies have examined the correlation between the conventional CT characteristics and KRAS mutation status in LUAD. These studies showed that no or few inconsistent CT features were associated with KRAS mutation in LUAD (11,12,14). Over the past decade, dual-energy spectral CT (DESCT) has been a profound innovation in radiographic image domain and has the potential to be used in clinical routine. DESCT can provide more information than conventional single-energy CT and has the potential to be used for lung cancer diagnosis, differential diagnosis, staging of tumour, efficacy evaluation, postoperative follow-up (15). There has been only scarce previous description of DESCT features of tumours of KRAS mutation in LUAD (16,17). Radiomics, as another promising imaging technology, offers a new lens to view and understand tumours. By combining advanced imaging techniques with computational algorithms, radiomics extracts vast amounts of high throughput and higher-dimensional features from radiographic images, which cannot be observed in the human visual system. However, the study exploring radiomics to predict KRAS mutation in LUAD is rare.

In light of these challenges, opportunities and the ever-evolving landscape of KRAS, this study seeks to develop and validate a more accurate, non-invasive approach (combining clinical, DESCT, and radiological features) to predict KRAS mutations in Asian patients with LUAD. By juxtaposing different prediction methods, this study charts a path forward, on that holds the promise of revolutionizing patient care and therapeutic strategies. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-24-694/rc).

Methods

Patient selection

The study cohort comprised patients diagnosed with LUAD at the Cancer Hospital of the Chinese Academy of Medical Sciences in China, between May 2013 and December 2016. These participants were identified retrospectively from a meticulously maintained, prospectively collected database that recorded individuals presenting with lung nodules and masses during the specified period at our institution.

The inclusion criteria were:

• Histologically confirmed LUAD according to the 2021 World Health Organization (WHO) lung cancer classification (18).
• Availability of complete clinical data of patients with thin-slice CT (1.25 mm) images that can be found in the picture achieving and communication system (PACS).
• Patients with available KRAS mutation status were determined by next-generation sequencing (NGS) of surgical specimens.
• Patients who underwent DESCT imaging within two weeks before any therapeutic intervention.

The exclusion criteria were:

• Patients with other concurrent malignancies or those who had undergone prior lung surgeries.
• Patients without available KRAS mutation status.
• Patients with poor-quality CT images due to motion artifacts or other technical issues.

Patients were stratified into development and validation cohorts in a 7:3 ratio through random allocation. Baseline clinical and demographic data encompassed age, gender, smoking history (categorized into non-smokers or never-smokers and smokers which included both former and current smokers), CT morphological patterns of the tumour, tumor diameter, and KRAS mutation status, all of which were sourced from the PACS. The CT morphological patterns of the tumour were assessed in alignment with the criteria set by the Fleischner Society (19). This categorization included sub-solid masses, further delineated into ground-glass opacities (GGOs) and part-solid masses. KRAS mutation was determined using fluorescence polymerase chain reaction (PCR), with samples procured either from surgical resections or biopsy.

DESCT examination

All patients underwent a DESCT scan using Discovery CT 750 HD system (GE Healthcare, USA). This scan covered the region from the lung apex to the adrenal gland and was performed prior to any treatment. The scanning protocol utilized the gemstone spectral imaging (GSI) mode, which rapidly switches tube voltage between 80 and 140 keV in cycles of 0.5 ms. Additional scanning parameters included a tube current of 550 mA, a tube rotation time of 0.6 s, and a collimator width of 40 mm. Patients received an intravenous injection of Ultravist 300 (Bayer Pharma AG, Germany) using high-pressure injector for contrast enhancement. The injection rate was set at 2.5 mL/s, with a total volume ranging from 85–100 mL. Scanning commenced after a 35-second delay post-injection. The DESCT images were acquired with patients in a supine position during the end-inspiratory phase.

Tumor delineation and extraction for DESCT features

DESCT images were delineated using GSI Viewer software in the post-processing workstation (Advantage Work Station 4.6, GE Healthcare, Milwaukee, WI, USA). A radiologist with 8 years of expertise (J.W.M.), who were blinded to all clinical information, in thoracic imaging manually selected the axial CT slice, positioned the region of interest (ROI) at the center of the lesion, and depicted the maximum diameter of the primary tumour. The ROI range was drawn with no less than 2/3 of the area of the lesion. Cavities, vacuoles, calcification, blood vessels and pulmonary atelectasis were avoided. The DESCT features were extracted automatically by the GSI Viewer software.

Tumor delineation and extraction for radiomics features

Preoperative high-resolution CT images were imported into ITK-SNAP 3.8 (http://www.itksnap.org) using the medical Digital Imaging and Communication in Medicine (DICOM) format. This facilitated the three-dimensional (3D) manual segmentation of the ROIs. To ensure consistency and minimize inter-observer variability, the radiologist mentioned above delineated the ROI on the axial CT lung window image (width, 1,200 Hounsfield Unit [HU]; level, −500 HU). This delineation was subsequently cross-verified in both coronal and sagittal planes. Any inaccuracies in the ROI were manually corrected. The segmented regions from each slice were then amalgamated to form the volume of interest (VOI). An additional verification of each VOI was conducted by a second radiologist with 15 years of experience. Notably, both radiologists needed to be blinded to the clinical and histopathological data of patients. Radiomics features were automatically extracted using the Artificial Intelligence Kit software (GE Healthcare, version 3.3.0), which operate based on the open-source Python Package Pyradiomics (https://pyradiomics.readthedocs.io/en/Latest/).

Selection of clinical, DESCT and radiomics features

We comprehensively evaluated clinical features potentially associated KRAS mutation status using univariate analysis. Depending on the nature of the variable, the Chi-squared test, t-test or Fisher’s exact test was employed for the nominal variable. In contrast, the Mann-Whitney test was utilized for continuous variables exhibiting non-normal distribution.

Given the distinct nature of radiomics and DESCT features, we adopted separate feature selection methodologies for each. Initially, all features underwent standardization via z-score transformation. For DESCT features, the Max-Relevance and Min-Redundancy (MRMR) approach was employed to pinpoint the most representative features. Radiomics features with P values less than 0.05 were selected using either the analysis of variance (ANOVA) or the Wilcoxon signed-rank test. Subsequently, the minimum redundancy maximum relevance (mRMR) algorithm was invoked to identify features that exhibited high relevance to KRAS mutation and ensured minimal redundancy. Concluding this process, the least absolute shrinkage and selection operator (LASSO) regression model, complemented by 10-fold cross-validation, was employed to discern features with non-zero coefficients.

Model construction

We constructed four models: a clinical model, a CT-based radiomics model, a DESCT model, and a combined model incorporating all three aforementioned features, termed the C-S-R model. These models were developed using multivariable logistic regression (LR) analysis. The Figure 1 shows the whole analysis pipeline step.

Figure 1 Pipeline step diagram of predicting KRAS mutation status in lung adenocarcinoma by clinical, DESCT, and radiomics features. ROI, region of interest; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic; DCA, decision curve analysis; DESCT, dual-energy spectral computed tomography; KRAS, Kirsten rat sarcoma viral oncogene homolog.

Statistical analyses and model evaluation

All statistical analyses were performed with RStudio Server (Version 1.1.463; RStudio, Inc., Boston, MA, USA) and Python 3.8.1. Continuous variables were compared using the Student’s t-test or Mann-Whitney U tests, as appropriate. Categorical variables between cohorts were assessed using the Chi-squared test. The performance of different models was evaluated using the receiver operating characteristic (ROC) curve, from which metrics such as the area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV) and negative predictive value (NPV) were derived. The model boasting the highest AUC was deemed the most optimal. A nomogram was constructed based on the coefficient from multivariable LR analysis, incorporating clinical features as well as scores from both DESCT and radiomics models. The nomogram, aimed at predicting KRAS mutations, is illustrated. For both the development and validation cohorts, decision curve analysis (DCA) was employed to ascertain the net benefits of each model across varying threshold probabilities, thereby assessing the clinical utility of the model. The model’s fit was gauged using a calibration curve and the Hosmer-Lemeshow test. The Delong test facilitated comparisons of AUC values between models. Statistical significance was denoted by a two-tailed P value of less than 0.05.

Ethical considerations

This study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This study was approved by the Institutional Review Board of Cancer Hospital of the Chinese Academy of Medical Science, Beijing, China (No. NCC2016G-029). Given the retrospective nature of the study, the requirement for informed consent was waived. All the patient data were anonymized and de-identified before analysis to ensure confidentiality.

Results

Patient demographics and selection criteria

Based on 501 patients collected prospectively, a cohort of 172 patients met our criteria. These participants exhibited an average age, represented as mean ± standard deviation (SD), of 57.69±9.67 years, spanning an age range from 30 to 76 years. A closer examination of the gender distribution revealed the inclusion of 91 women (mean age ± SD, 57.07±9.41 years; age spectrum: 30–75 years) and 81 men (mean age ± SD, 58.38±9.96 years; age spectrum: 33–76 years). There were 120 and 52 patients in the developing and validation cohorts, respectively. The baseline characteristics in developing and validation groups have no significant difference (Table 1). The correlation between the KRAS mutations and clinical features and CT morphological patterns is shown in Table 2. KRAS mutations are significantly associated with smoking (P=0.02) and moderately correlated with CT pattern (P=0.07). The outcomes of these analyses are detailed in Tables 1,2. The research diagram is shown in Figure 1.

Feature selection in DESCT and radiomics

The DESCT features of the patient cohort were encapsulated through 24 variables. The LASSO dimensionality reduction model was employed to streamline the feature set and eliminate redundancy. A noteworthy observation was the absence of any features with a P value below 0.05, suggesting that no DESCT feature bore a statistically significant association with the KRAS mutation in LUAD. Nevertheless, three distinct features emerged as potential markers for discerning the KRAS mutation status in LUAD: CT value at 40 keV, effective Z, calcium water.

From the radiomics data, we extracted an expansive set of 1,316 features. This encompassed 28 first-order features and 75 texture features derived from metrics such as the Gray Level Co-occurrence Matrix (GLCM), Gray Level Size Zone Matrix (GLSZM), Gray Level Run Length Matrix (GLRLM), and Gray Level Dependence Matrix (GLDM). Furthermore, 1,209 transformed first-order and textual features were extracted including wavelet-decomposed, Laplacian of Gaussian (LoG) filtered with Sigma of 2.0 and 3.0 mm, and local binary pattern (LBP) filtered texture features. Post a comprehensive statistical evaluation, we zeroed in on the most pertinent features, which included wavelet-HHL_glcm_ClusterShade, wavelet-HHH_firstorder_Skewness, wavelet-LHH_glszm_SmallAreaLowGrayLevelEmphasis, wavelet-HHH_firstorder_Median. Figure 2 illustrated LASSO regression model, complemented by 10-fold cross-validation, was employed to discern features with non-zero coefficients.

Figure 2 The LASSO was used to further filter the most relevant features. (A) The LASSO model’s tuning parameter (λ) selection used 10-fold cross validation via minimum criterion. The vertical dashed lines indicate the optimal value of LASSO tuning parameter (λ). (B) LASSO coefficient profile plot with different log (λ) was shown. The vertical dashed line represents 4 radiomics features with nonzero coefficients selected with the optimal λ value. LASSO, least absolute shrinkage and selection operator.

Comparative performance analysis of clinical, DESCT, radiomics, and C-S-R models

The efficacy of the four proposed models was rigorously assessed both in the development and in validation cohort, utilizing the area under the ROC curves as the primary metric. Detailed visual representations of the ROC curves and the corresponding AUC values are delineated in Figure 3 and Table 3 for the development and validation cohorts, respectively. Side-by-side performance comparison of the quartet of models across distinct cohorts is presented in Table 4. Additionally, Figure 4 offers a graphical depiction of the RadScores distribution across the models for both cohorts.

Figure 3 The ROC curves of the four prediction models that indicate KRAS mutation status in different cohorts. (A) The developing cohort. (B) The validation cohort. DESCT, dual-energy spectral computed tomography; C_S_R_Model, the combined model incorporating clinical, DESCT and radiomics features; ROC, receiver operating characteristic; KRAS, Kirsten rat sarcoma viral oncogene homolog.

Figure 4 RadScore distribution in the developing and validation cohorts. Boxplot showed the predictive power of different models between the KRAS mutant and wild groups in the development (A) cohort and validation cohort (B). *, P<0.05; **, P<0.01; ***, P<0.001. DESCT, dual-energy spectral computed tomography; C_S_R_Model, the combined model incorporating clinical, DESCT and radiomics features; KRAS, Kirsten rat sarcoma viral oncogene homolog.

The C-S-R model, an amalgamation of clinical, DESCT, and radiomics features, emerged as the most efficacious one of the models evaluated. In the development cohort, this model achieved an AUC of 0.92, and accuracy rate of 91.9%, a sensitivity of 90.0%, and a specificity of 84.6%. When tested on the validation cohort, the model’s performance metrics were an AUC of 0.87, accuracy of 92.3%, sensitivity of 80%, and specificity of 93.6%. While the C-S-R model’s performance was on par with the radiomics model (P=0.28), it significantly outperformed the DESCT model (P=0.01). A prediction model nomogram was constructed to provide a visual representation of the optimal prediction model, as showcased in Figure 5.

Figure 5 The nomogram for predicting KRAS mutation. The nomogram was composed of the RadScore, smoking history and solid nodal. KRAS, Kirsten rat sarcoma viral oncogene homolog.

Discussion

Our study stands as a pioneering endeavor in the realm of medical research, particularly in the amalgamation of clinical, DESCT, and radiomics features to predict the elusive KRAS mutation in Asian patients with LUAD. The results that we procured were affirmative, underscoring the enhanced predictive value when CT-based radiomics, DESCT, and clinical features are synergistically combined.

A salient observation from our study was the discernible correlation between KRAS mutation and smoking history in patients with LUAD. This observation resonates with findings from prior research endeavors (2,3,20). Our clinical model showcased a commendable classification power, with AUC values of 0.70 and 0.69 for the development and validation cohorts, respectively. However, it is imperative to note that the precision of this model is yet to reach the zenith required for seamless clinical integration. While smoking history undeniably plays a pivotal role in KRAS mutation, it is not the sole determinant. A plethora of other mechanisms, some of which remain unexplored in this study, could be instrumental in the manifestation of the KRAS mutation in LUAD.

Our research delved into the relationship between conventional CT features and KRAS mutation in LUAD. A revelation from our study was the non-association of the presence of a solid nodule with the KRAS mutation in LUAD (P=0.07). This outcome could be attributed to the limited sample size and KRAS’s relatively low mutation rate in our study cohort. This hypothesis is further bolstered by previous studies that have identified a solid nodule as the quintessential CT feature of LUAD harboring a KRAS mutation (12,17).

DESCT, with its avant-garde fast-kVp switching technology, is adept at generating two congruent energy datasets at 80–140 kVp. The monochromatic images, spanning from 40 to 140 keV, are meticulously crafted through projection-based reconstruction. This offers a treasure trove of quantitative data about diverse materials, such as calcium, iodine, water, fat, etc., leveraging their unique linear attenuation coefficients (21). Despite its immense potential, the DESCT model in our study exhibited suboptimal predictive prowess, with AUC values of 0.73 and 0.67 for the developing and validation cohorts, respectively. Our research included subsolid nodules (SSN), encompassing mixed GGO and pure GGO. This classification led to a diminished CT values across all energy levels, subsequently impacting CT values in different energy levels, thereby influencing the predictive model’s performance. We also surmise that the performance was affected by the limited sample size and the subdued mutation rate of KRAS observed in our study. In the study conducted by Li et al. (17), the results indicated that a model integrating both clinical and DESCT features could effectively differentiate between KRAS and EGFR statuses in solid LUAD. Their predictive AUC, derived from two pivotal factors (CT number at 70 keV and smoking history) for distinguishing KRAS and EGFR mutations, stood at 0.841 (95% confidence interval: 0.717–0.965, P<0.001) with a demarcation point of 2.72. In a subsequent multi-center study, Li et al. (22) employed a dual-energy CT-based radiomics nomogram to predict the histological differentiation of head and neck squamous carcinoma. This approach yielded exemplary results, with AUC values of 0.987 and 0.968 for the development and validation cohorts. Given these findings, we postulate that DESCT features, either independently or in conjunction with radiomics features, could be instrumental in identifying the KRAS mutation in LUAD. The efficacy of DESCT features warrants further exploration. The association between KRAS mutation in LUAD and DESCT features has been infrequently documented in prior literature.

Radiomics, on the other hand, showcased a robust predictive capability in our study, with AUC values of 0.88 and 0.75 for the development and validation cohorts, respectively. Our rigorous analysis distilled the vast feature set to four paramount features: wavelet-HHL_glcm_ClusterShade, wavelet-HHH_firstorder_Skewness, wavelet-LHH_glszm_SmallAreaLowGrayLevelEmphasis, wavelet-HHH_firstorder_Median. Our literature review only identified a study that employed radiomics to predict KRAS mutation in LUAD. The unique research constructed a model using ten radiomics features to predict KRAS mutations in LUAD, achieving a modest AUC of 0.63 (20). The radiomics approach offers several advantages: non-invasive, rapid, cost-effective, and easily implemented. Compared with liquid biopsy, radiomics is more accessible, does not necessitate specialized equipment, and is significantly more economical, paving the way for broader adoption. Radiomics delves into the entire tumor’s texture instead of focusing on external elements like ctDNA or circulating tumour cells (CTCs). This ensures that the features discerned by radiomics offer a direct and robust representation of the tumour. Compared to surgical or biopsy specimens, radiomics provides a holistic view of the tumor, mitigating risks associate with incorrect sampling or transplant errors. Radiomics, with its non-invasive quantification of tumor heterogeneity, holds immense promise. However, its performance could be influenced by myriad factors, including spatial resolution, post-processing techniques, ROI delineation, feature definition, model algorithms, and the size of the development set. During the evaluations of different models in validation cohorts, we applied the pre-treatment CT scans of patients with LUAD, not that after chemotherapy and/or radiotherapy. Therefore, this underscores the necessity for future assessments of models’ stability under diverse conditions. Subsequent research could prioritize predicting the emergence of drug-resistant genetic alterations. Additionally, the genetic alteration statuses of prevalent driver genes, such as epidermal growth factor receptor (EGFR), anaplastic lymphoma kinase (ALK), and KRAS, along with infrequent gene modifications like rearranging transfection (RET) rearrangement, c-ros oncogene 1 (ROS-1), v-raf murine sarcoma viral oncogene homolog B1 (BRAF), and human epidermal growth factor receptor 2 (HER2) mutations, warrant exploration.

The C-S-R model, a harmonious blend of clinical, DESCT, and radiomics features, emerged as the paragon in our study. However, no statistically significant difference appeared when juxtaposed with the radiomics model, underscoring the need for further research with a larger sample size. Both the radiomics and C-S-R models outperformed clinical and DESCT model in two groups in our study. Concurrently, the model’s visualization is facilitated by a nomogram. The nomogram model might also unveil other tumour phenotypes, including tumour proliferative activity, hypoxia status, tumour microenvironment, treatment response, treatment resistance, and metastatic tendencies. Such insights could enrich diagnostic precision, treatment planning, and tumour recurrence monitoring. Enhancing the performance of the nomogram model could be achieved by expanding the sample size, incorporating diverse medical imaging techniques, refining the selection of radiomics and DESCT features, and optimizing the model development algorithm. The decision analysis curve (DAC) attests to the model’s commendable clinical applicability, offering clinicians a non-invasive, quantitative, convenient, and swift diagnostic tool. It’s worth noting that DAC serves as a mechanism to forecast clinical outcomes by juxtaposing various variables (23). This underscores the need for expansive research with augmented sample sizes. While the results are promising, the sensitivity, specificity, and accuracy metrics of the radiomics and C-S-R models have yet to reach tissue or plasma NGS for clinical deployment. We remain sanguine about the prospects of the radiomics approach in predicting KRAS mutation and envisage its integration into clinical paradigms in the foreseeable future.

In recent years, some studies have used CT radiomics-based deep learning methods to predict gene alterations and prognosis information in NSCLC. Zhang et al. found that the Deep-RadScore exhibits strong discrimination for KRAS mutation (AUC =0.896) (24). Morel et al. (25) used a standard prediction pipeline based on a convolutional neural network for the detection of cancer driver genomic alterations including KRAS mutation in LUAD. However, the performance is not satisfactory with an AUC of 0.57±0.049. Interestingly, the DESCT-based radiomics model has a higher predictive power (AUC =0.92) than the reported results for deep learning method. The difference in predictive performance may be many-sided, such as the gene alteration rate and sample size in the study. Although deep learning is a promising medical tool, the biological significance of deep network features remains to be further elucidated, such as seeking to understand the expression mechanisms of imaging biomarkers, establish radiogenomic features with causal relationships, and unveil the underlying biology driven by deep learning.

Our study boasts several distinct advantages

Uniformity in data collection

All our data were sourced from a singular CT scanner, ensuring uniformity in image acquisition parameters. This consistency effectively mitigates the inherent heterogeneity often found in images, which could otherwise jeopardize test-retest reproducibility. It is well-established that variations in CT acquisition parameters and reconstruction algorithms across different CT scanners can adversely affect reproducibility and stability (26).

Optimal ROI delineation

We employed a semi-automatic method for delineating the ROIs of tumours, a widely regarded superior technique. The methodology chosen for ROI delineation can significantly influence boundary description, subsequently impacting feature extraction. There’s a discernible shift towards automation, minimizing manual interventions. Ambiguous region contouring demands advanced algorithms capable of distinguishing between normal tissue and tumours. In the foreseeable future, ROI delineation might benefit from rapid artificial intelligence (AI) assistance.

Robust feature selection

Our approach incorporated a performance-driven feature selection strategy coupled with an apt algorithm for model development. This combination is resilient to overfitting, facilitating the creation of a high-performing, reliable model. The judicious selection of image features is pivotal to model performance. An excessive number of features can lead to overfitting, thereby diminishing model efficacy.

Precision in image resolution

Our models rely on contrast-enhanced CT images with a layer thickness of 1.25 mm. This fine resolution enables the discernment of minute structures and intricate details. During model construction, it ensures minimal loss of interlayer information, such as subtle structures, undefined features, and other interlayer data. After all, The layer thickness of CT scans can significantly influence the stability of quantitative image features (27).

Limitations

Our study, despite its groundbreaking findings, is not devoid of limitations. The retrospective design, confined to a single-center and predominantly encompassing an Eastern Asian demographic, potentially introduces biases and curtails the universal applicability of the results. The relatively diminutive sample size, especially for patients with KRAS mutations, could impede the precision of the DESCT and radiomics features in predicting KRAS mutations. We advocate for future research endeavors to adopt a multi-center approach, augmented by larger sample size, to validate and amplify the findings of this study.

Conclusions

In conclusion, our research represents a trailblazing initiative in synergizing clinical, DESCT, and radiomics features to predict the rare KRAS mutation in Asian patients with LUAD. The zenith achieved by the best model, with an AUC of 0.92, sensitivity of 80%, specificity of 94%, and accuracy of 92%, is commendable. We are optimistic about the potential of radiomics in mirroring the genetic heterogeneity inherent in tumors. However, the journey to achieving clinical-grade precision is still underway. We are confident that with continued research and innovation, the radiomics approach will evolve into an invaluable asset in clinical diagnostics.

Acknowledgments

Funding: The research was supported by the National Natural Science Foundation of China (grant Nos. 81601494 and 81701692).

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-24-694/rc

Data Sharing Statement: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-24-694/dss

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-24-694/prf

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-24-694/coif). Y.M.W. is an employee of GE Healthcare China company. 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. This study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This study was approved by the Institutional Review Board of Cancer Hospital of the Chinese Academy of Medical Science, Beijing, China (No. NCC2016G-029). Given the retrospective nature of the study, the requirement for informed consent was waived. All the patient data were anonymized and de-identified before analysis to ensure confidentiality.

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. Siegel RL, Miller KD, Wagle NS, et al. Cancer statistics, 2023. CA Cancer J Clin 2023;73:17-48. [Crossref] [PubMed]
2. Guan JL, Zhong WZ, An SJ, et al. KRAS mutation in patients with lung cancer: a predictor for poor prognosis but not for EGFR-TKIs or chemotherapy. Ann Surg Oncol 2013;20:1381-8. [Crossref] [PubMed]
3. Eberhard DA, Johnson BE, Amler LC, et al. Mutations in the epidermal growth factor receptor and in KRAS are predictive and prognostic indicators in patients with non-small-cell lung cancer treated with chemotherapy alone and in combination with erlotinib. J Clin Oncol 2005;23:5900-9. [Crossref] [PubMed]
4. Kohno T, Nakaoku T, Tsuta K, et al. Beyond ALK-RET, ROS1 and other oncogene fusions in lung cancer. Transl Lung Cancer Res 2015;4:156-64. [Crossref] [PubMed]
5. Hong DS, Fakih MG, Strickler JH, et al. KRAS(G12C) Inhibition with Sotorasib in Advanced Solid Tumors. N Engl J Med 2020;383:1207-17. [Crossref] [PubMed]
6. Jänne PA, Riely GJ, Gadgeel SM, et al. Adagrasib in Non-Small-Cell Lung Cancer Harboring a KRAS(G12C) Mutation. N Engl J Med 2022;387:120-31. [Crossref] [PubMed]
7. Sacher A, LoRusso P, Patel MR, et al. Single-Agent Divarasib (GDC-6036) in Solid Tumors with a KRAS G12C Mutation. N Engl J Med 2023;389:710-21. [Crossref] [PubMed]
8. Robertson EG, Baxter G. Tumour seeding following percutaneous needle biopsy: the real story! Clin Radiol 2011;66:1007-14. [Crossref] [PubMed]
9. Shyamala K, Girish HC, Murgod S. Risk of tumor cell seeding through biopsy and aspiration cytology. J Int Soc Prev Community Dent 2014;4:5-11. [Crossref] [PubMed]
10. Di Capua D, Bracken-Clarke D, Ronan K, et al. The Liquid Biopsy for Lung Cancer: State of the Art, Limitations and Future Developments. Cancers (Basel) 2021;13:3923. [Crossref] [PubMed]
11. Rizzo S, Petrella F, Buscarino V, et al. CT Radiogenomic Characterization of EGFR, K-RAS, and ALK Mutations in Non-Small Cell Lung Cancer. Eur Radiol 2016;26:32-42. [Crossref] [PubMed]
12. Park J, Kobayashi Y, Urayama KY, et al. Imaging Characteristics of Driver Mutations in EGFR, KRAS, and ALK among Treatment-Naïve Patients with Advanced Lung Adenocarcinoma. PLoS One 2016;11:e0161081. [Crossref] [PubMed]
13. Nikanjam M, Kato S, Kurzrock R. Liquid biopsy: current technology and clinical applications. J Hematol Oncol 2022;15:131. [Crossref] [PubMed]
14. Sugano M, Shimizu K, Nakano T, et al. Correlation between computed tomography findings and epidermal growth factor receptor and KRAS gene mutations in patients with pulmonary adenocarcinoma. Oncol Rep 2011;26:1205-11. [Crossref] [PubMed]
15. Kim C, Kim W, Park SJ, et al. Application of Dual-Energy Spectral Computed Tomography to Thoracic Oncology Imaging. Korean J Radiol 2020;21:838-50. [Crossref] [PubMed]
16. Li M, Zhang L, Tang W, et al. Quantitative features of dual-energy spectral computed tomography for solid lung adenocarcinoma with EGFR and KRAS mutations, and ALK rearrangement: a preliminary study. Transl Lung Cancer Res 2019;8:401-12. [Crossref] [PubMed]
17. Li M, Zhang L, Tang W, et al. Dual-energy spectral CT characteristics in surgically resected lung adenocarcinoma: comparison between Kirsten rat sarcoma viral oncogene mutations and epidermal growth factor receptor mutations. Cancer Imaging 2019;19:77. [Crossref] [PubMed]
18. Maleszewski JJ, Basso C, Bois MC, et al. The 2021 WHO Classification of Tumors of the Heart. J Thorac Oncol 2022;17:510-8. [Crossref] [PubMed]
19. MacMahon H, Naidich DP, Goo JM, et al. Guidelines for Management of Incidental Pulmonary Nodules Detected on CT Images: From the Fleischner Society 2017. Radiology 2017;284:228-43. [Crossref] [PubMed]
20. Rios Velazquez E, Parmar C, Liu Y, et al. Somatic Mutations Drive Distinct Imaging Phenotypes in Lung Cancer. Cancer Res 2017;77:3922-30. [Crossref] [PubMed]
21. Goo HW, Goo JM, Dual-Energy CT. New Horizon in Medical Imaging. Korean J Radiol 2017;18:555-69. [Crossref] [PubMed]
22. Li Z, Liu Z, Guo Y, et al. Dual-energy CT-based radiomics nomogram in predicting histological differentiation of head and neck squamous carcinoma: a multicenter study. Neuroradiology 2022;64:361-9. [Crossref] [PubMed]
23. Van Calster B, Wynants L, Verbeek JFM, et al. Reporting and Interpreting Decision Curve Analysis: A Guide for Investigators. Eur Urol 2018;74:796-804. [Crossref] [PubMed]
24. Zhang X, Zhang G, Qiu X, et al. Exploring non-invasive precision treatment in non-small cell lung cancer patients through deep learning radiomics across imaging features and molecular phenotypes. Biomark Res 2024;12:12. [Crossref] [PubMed]
25. Morel LO, Derangère V, Arnould L, et al. Preliminary evaluation of deep learning for first-line diagnostic prediction of tumor mutational status. Sci Rep 2023;13:6927. [Crossref] [PubMed]
26. Emaminejad N, Wahi-Anwar MW, Kim GHJ, et al. Reproducibility of lung nodule radiomic features: Multivariable and univariable investigations that account for interactions between CT acquisition and reconstruction parameters. Med Phys 2021;48:2906-19. [Crossref] [PubMed]
27. Zhao B, Tan Y, Tsai WY, et al. Reproducibility of radiomics for deciphering tumor phenotype with imaging. Sci Rep 2016;6:23428. [Crossref] [PubMed]