Multimodal radiomics and pathomics model for predicting postoperative local recurrence in T3–4 non-small cell lung cancer

Abstract Background Patients with resected pT3–4 non-small cell lung cancer (NSCLC) remain at substantial risk of local recurrence, highlighting the need for individualized postoperative risk stratification. We investigated whether integrating computed tomography radiomics with weakly supervised histopathology could improve risk stratification for postoperative local recurrence. Methods This two-center retrospective study included consecutive patients with pT3–4N0–2M0 NSCLC who underwent R0 resection at a development center between January 2015 and December 2021 and a separate external center between January 2018 and December 2021. Preoperative contrast-enhanced CT, postoperative hematoxylin and eosin (H&E) whole-slide images, and clinicopathological variables were analyzed. The primary outcome was time from surgery to first documented local recurrence; distant-first recurrence and death before local recurrence were competing events. Penalized Fine–Gray models for clinicopathological, radiomics, and CNN-derived pathomics features, together with score-level fusion, were developed and internally evaluated using repeated nested five-fold cross-validation in the development cohort. The upstream pathology CNN was trained separately using a fixed 7:3 patient-level split and was not retrained within the cross-validation folds. The final models were locked before external-center evaluation. Performance at 2 years was assessed using time-dependent area under the curve (AUC) accounting for competing risks, Brier score, calibration, and paired differences between models. Results The development and external-center cohorts included 135 and 31 patients, with 30 and 5 primary local-recurrence events, respectively. The 2-year AUCs of the clinicopathological, intratumoral radiomics, peritumoral radiomics, Path-DL, and fusion models were 0.492 (95% confidence interval [CI], 0.317–0.658), 0.689 (0.555–0.816), 0.694 (0.555–0.818), 0.804 (0.665–0.856), and 0.827 (0.689–0.936) in repeated internal validation and 0.310 (0.062–0.604), 0.605 (0.256–0.905), 0.616 (0.296–0.854), 0.619 (0.337–0.902) and 0.683 (0.411–0.917) in external-center evaluation. For the multimodal model versus Path-DL, the paired difference in AUC was 0.023 (95% CI, − 0.097 to 0.062) internally and 0.064 (− 0.270 to 0.413) externally; corresponding differences in Brier score were 0.001 (− 0.002 to 0.005) and 0.001 (− 0.003 to 0.004). Conclusions Multimodal integration of CT radiomics and weakly supervised histopathology showed potential for estimating 2-year local-recurrence risk after R0 resection of pT3–4 NSCLC while accounting for competing events.