Non-invasive prediction of non-responders to neoadjuvant immunochemotherapy in NSCLC using intratumoral and peritumoral radiomics: a multicenter study

Neoadjuvant immunochemotherapy (NICT) has emerged as the standard perioperative treatment for patients with resectable nonsmall cell lung cancer (NSCLC). It can downstage tumors preoperatively, improve the R0 resection rate, and enhance the long-term survival of NSCLC patients. However, results from several phase III clinical trials[1–3] indicate that approximately 16–20% of NSCLC patients do not undergo surgery following NICT, primarily due to insufficient response, adverse events, or other reasons. Current biomarker research[4] on NICT response predominantly focuses on predicting pathological response postsurgery, yet neglects this significant subset of nonoperated patients. For these patients, we should adopt more targeted treatment strategies, including timely surgery or appropriate alternative therapies preoperatively. Therefore, identifying nonresponders before NICT holds greater clinical importance for precise pretreatment stratification, but research in this area is scarce. Radiomics, an emerging class of noninvasive biomarkers derived from medical imaging, has demonstrated considerable success in characterizing NSCLC[5] and predicting outcomes[6]. However, its potential for predicting treatment response to NICT remains underexplored. To address this gap, we propose developing an optimized radiomics model that integrates intratumoral region with the optimal peritumoral region to support personalized therapeutic decision-making. The work has been reported in line with the TITAN criteria[7]. Methods Two experienced radiologists independently delineated tumor regions using 3D Slicer (v5.6.2), followed by consensus discussion. The PyRadiomics open-source package (v3.0.1) was used for automatic peritumoral region expansion and radiomic feature extraction. After feature selection and dimensionality reduction, five machine learning models were employed to build and compare predictive models. Performance was evaluated using receiver operating characteristic (ROC) curves, with calculation of AUC, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Calibration curves assessed prediction accuracy, decision curve analysis (DCA) compared clinical utility, and DeLong tests statistically evaluated AUC differences. Shapley Additive exPlanations (SHAP) visualized the best model’s predictions. A comprehensive description of the methodology and the study design flowchart is shown in Supplemental digital content Figure S1, available at: https://links.lww.com/JS9/F601. Results This study ultimately selected 347 patients from a total of 4366 lung cancer (LC) patients at center A from January 2018 to August 2023, and 4078 LC patients at center B from January 2019 to February 2024. The flowchart of the patient recruitment pathway is shown in Supplemental digital content Figure S2, available at: https://links.lww.com/JS9/F601 Patients from center A were randomly assigned 7:3 to the training set (n = 150) and internal validation set (n = 65). The entire training subset was utilized for model development, while cases from center B served as an independent external cohort for performance evaluation. The baseline clinical characteristics of the patients are shown in Supplemental digital content Table S1, available at: https://links.lww.com/JS9/F601 To compare predictive performance with the radiomics model, we employed univariate and multivariate logistic regression to identify clinical features (Supplemental digital content Table S2, available at: https://links.lww.com/JS9/F601) associated with non-responders for constructing the clinical model. Radiomic feature selection and dimensionality reduction processes for different models are shown in Supplementa, digital content, Figures S3–S8. available at: https://links.lww.com/JS9/F601 To select the optimal machine learning model, we compared five methods across intratumoral models and combinations with varying peritumoral regions (Supplemental digital content Tables S3–S8, available at: https://links.lww.com/JS9/F601). The Gradient Boosting algorithm demonstrated superior overall performance and was selected for modeling (Supplemental digital content Figure S9, available at: https://links.lww.com/JS9/F601). Table 1 compares the performance comparison of the clinical, intratumoral, and combined intraperitumoral radiomics models across three datasets. The model integrating intratumoral and 3-mm peritumoral features (TP3) achieved higher AUCs than other models in the training, internal validation, and external test sets (Fig. 1A–C), with respective AUCs of 0.929, 0.892, and 0.855. The combined intraperitumoral radiomics model outperformed both traditional (intratumoral-only) radiomics models and the clinical model, which confirms that features from the peritumoral microenvironment can improve predictive performance. Predictive performance initially increased then decreased with expanding peritumoral regions across all datasets. Furthermore, DeLong tests (Fig. 1D-F) confirmed that the model combining intratumoral with 3-mm peritumoral features showed statistically significant differences versus other models in training and external validation sets. Thus, we identified that the 3-mm peritumoral extent was optimal for the combined model in predicting nonresponders to NICT in NSCLC patients. Calibration curves across datasets are presented in Figures 2A–C. The TP3 model exhibited a calibration curve that was closer to the ideal diagonal line in most cases, indicating better calibration performance and greater consistency between predicted probabilities and actual outcomes. Decision curve analysis (DCA) compared the predictive models to assess their clinical utility. As shown in Figure 2D-F, the DCA curve for the TP3 model exceeded those of other models across most risk thresholds in all datasets. This indicates that clinical decisions based on the TP3 model yield higher net benefit across most threshold probabilities. The SHAP value summary plot (Supplemental digital content, Figure 10A, available at: https://links.lww.com/JS9/F601) illustrates overall feature importance rankings, while force plots (Supplemental digital content, Figures 10B–C, available at: https://links.lww.com/JS9/F601) visualize individual feature contributions to predictions. Figure 1.: ROC curves of clinical tumor, TP1, TP2, TP3, TP4, and TP5 models in the training (A), internal validation (B), and external test cohorts (C). DeLong tests of these models in the training (D), internal validation (E), and external test cohorts (F). ROC curve, receiver operating characteristic curve; AUC, area under the curve; TP1 model, based on intratumoral and 1 mm peritumoral features; TP2 model, based on intratumoral and 2 mm peritumoral features; TP3 model, based on intratumoral and 3 mm peritumoral features; TP4 model, based on intratumoral and 4 mm peritumoral features; TP5 model, based on intratumoral and 5 mm peritumoral features; tumor model, based on traditional radiomic features from tumor volume; clinical model, based on clinical variables.Figure 2.: Calibration curves of clinical, tumor, TP1, TP2, TP3, TP4 and TP5 models in the training (A), internal validation (B), and external test cohorts (C). DCA curves of these models in the training (D), internal validation (E), and external test cohorts (F). DCA curves, decision curve analysis curves; TP1 model, based on intratumoral and 1 mm peritumoral features; TP2 model, based on intratumoral and 2 mm peritumoral features; TP3 model, based on intratumoral and 3 mm peritumoral features; TP4 model, based on intratumoral and 4 mm peritumoral features; TP5 model, based on intratumoral and 5 mm peritumoral features; tumor model, based on traditional radiomic features from tumor volume; clinical model, based on clinical variables. Table 1 - The comparison of predictive performance of different models in three datasets.. Cohort Model AUC AUC 95%CI Acc Acc 95%CI Sen Spe PPV NPV Train Clinical 0.781 0.6960–0.8650 0.773 0.6979–0.8376 0.87 0.58 0.806 0.690 Train Tumor 0.851 0.7877–0.9147 0.787 0.7124–0.8493 0.84 0.68 0.840 0.680 Train TP1 0.868 0.8087–0.9273 0.833 0.7639–0.8891 0.92 0.66 0.844 0.805 Train TP2 0.880 0.8171–0.9435 0.840 0.7714–0.8947 0.87 0.78 0.888 0.750 Train TP3 0.929 0.8849–0.9731 0.887 0.8248–0.9326 0.93 0.80 0.903 0.851 Train TP4 0.851 0.7865–0.9155 0.807 0.7343–0.8665 0.84 0.74 0.866 0.698 Train TP5 0.811 0.7408–0.8818 0.707 0.6269–0.7781 0.66 0.80 0.868 0.541 Val Clinical 0.636 0.4873–0.7854 0.708 0.5817–0.8140 0.818 0.476 0.766 0.556 Val Tumor 0.843 0.7477–0.9385 0.769 0.6481–0.8647 0.705 0.905 0.939 0.594 Val TP1 0.855 0.7571–0.9528 0.769 0.6481–0.8647 0.750 0.810 0.892 0.607 Val TP2 0.857 0.7644–0.9499 0.800 0.6823–0.8890 0.864 0.667 0.844 0.700 Val TP3 0.892 0.8136–0.9710 0.807 0.7661–0.8872 0.740 0.940 0.961 0.644 Val TP4 0.814 0.7117–0.9160 0.677 0.5495–0.7877 0.545 0.952 0.960 0.500 Val TP5 0.757 0.6382–0.8759 0.631 0.5020–0.7472 0.500 0.905 0.917 0.463 Test Clinical 0.542 0.4329–0.6513 0.636 0.5482–0.7183 0.795 0.318 0.700 0.437 Test Tumor 0.750 0.6589–0.8408 0.750 0.6672–0.8212 0.841 0.568 0.796 0.641 Test TP1 0.764 0.6760–0.8514 0.742 0.6591–0.8146 0.818 0.591 0.800 0.619 Test TP2 0.791 0.7077–0.8744 0.780 0.7000–0.8477 0.830 0.682 0.839 0.667 Test TP3 0.855 0.7907–0.9201 0.720 0.7438–0.8774 0.860 0.900 0.945 0.763 Test TP4 0.739 0.6424–0.8349 0.742 0.6591–0.8146 0.818 0.591 0.800 0.619 Test TP5 0.709 0.6134–0.8045 0.583 0.4944–0.6685 0.500 0.750 0.800 0.429 Discussion This study focuses on the clinical value of identifying potential nonresponders before NICT initiation. Unlike prior research predicting pathological response after neoadjuvant therapy[8], we aim to provide clinicians with earlier decision-making support for timely treatment adjustments. While strategies for identifying potential nonresponders to neoadjuvant therapy have been explored in other cancers, such as breast[9] and rectal cancer[10], these studies were predominantly single-center. A key strength of our study lies in the external validation cohort that strictly adhered to the same inclusion and exclusion criteria. This approach robustly assessed the model’s performance in an independent population and significantly enhanced its clinical applicability and generalizability. Furthermore, this study introduces an innovative approach integrating multiscale spatial features derived from the peritumoral microenvironment. This integration aims to capture critical tumor–stroma interactions[11] and significantly enhances the model’s prediction performance. We observed that moderate peritumoral regions contain critical prognostic information, whereas excessively expanded areas may introduce substantial normal tissue or extraneous confounding factors. Such over-expansion leads to signal dilution and ultimately impairs model performance. Therefore, we systematically evaluated and combined peritumoral features across varying distances, identifying an optimal spatial combination strategy for predicting nonresponders. This finding underscores the critical role of spatial heterogeneity within the peritumoral microenvironment in predictive modeling, and provides significant guidance for future research. While this study demonstrates promising potential, several limitations remain. First, as a retrospective study, inherent risks such as information selection bias are unavoidable. Future large-scale studies should validate the consistency of identified feature importances across diverse cohorts and assess the generalizability of SHAP-derived biomarkers. Second, manual segmentation may vary subjectively. Although we minimized this variability through independent contouring by multiple senior imaging experts and consensus discussions, developing and applying highly reproducible deep learning-based automated segmentation algorithms will be crucial for reducing manual intervention in future research. Furthermore, uniformly extending the peritumoral region by a fixed distance may overlook the spatial heterogeneity of the microenvironment across different T-stage tumors. Future studies should explore personalized definitions of the peritumoral area to more accurately capture information relevant to treatment response. Finally, while SHAP analysis provided an initial interpretation of the model’s predictions, the underlying biological mechanisms of these radiomic features require further elucidation. Future multiomics integration studies will be essential to decipher the biological significance of imaging features and build causal models.