Enhanced prediction of gene mutation and risk stratification in non-small-cell lung cancer through dual-pathway fusion of radiomics and pathomics

Purpose This study aimed to develop and validate a multimodal combined model that integrated radiomics, pathomics and clinical features to precisely predict EGFR status and risk stratification in NSCLC. Materials and methods We retrospectively analyzed 387 patients with NSCLC from two hospitals (the train cohort: n=193; the internal validation cohort: n=83; the external validation cohort: n=111). Radiomics models were developed using 3D CNN for the construction of deep learning radiomics (DLRadiomics). Weakly supervised learning and multi-instance learning were used to develop pathomics signature (Pathomics). We conducted an in-depth analysis of clinical features resulting in a clinical signature (Clinical). Finally, we integrated them into a comprehensive nomogram-Nomogram. The comparative analysis of all models was conducted through a comprehensive evaluation. The distribution of predictive features for Nomogram across different EGFR mutation subtypes was evaluated. The Kaplan-Meier curve was employed to assess the predictive capability of Nomogram in risk stratification among cases with survival outcomes. Results In comparison to Clinical, DLRadiomics and Pathomics models, Nomogram exhibits superior predictive performance (the train cohort: AUC=0.986, 95%CI=0.969-1.000; the internal validation cohort: AUC=0.796, 95%CI=0.659-0.932; the external test cohort: AUC=0.850, 95%CI=0.719-0.981). Nomogram could also be used to predict effectively EGFR mutation subtype (P<0.05). In the validation and test cohorts, the log rank test proved the effectiveness of Nomogram in predicting risk stratification (P<0.05). Conclusions We demonstrated that Nomogram which integrated radiomics, pathomics and clinical features, could be served as a noninvasive and reusable tool to precisely predict EGFR status and risk stratification in NSCLC.