A multimodal radiomics, deep learning, and pathomics signature for predicting the prognosis in central conventional chondrosarcoma.

OBJECTIVE: This study aims to develop and validate a multimodal feature-integrated signature (MS) by combining radiomics, deep learning (DL), and pathomics features to predict prognosis in central conventional chondrosarcoma (CS) patients. METHODS: In this multicenter retrospective study, 166 central conventional CS patients who underwent surgery were enrolled from two institutions and divided into a training cohort (n = 117) and a validation cohort (n = 49). Using preoperative non-enhanced computed tomography (CT) images and postoperative whole slide imaging (WSI) of hematoxylin and eosin (H&E)-stained pathological sections, we extracted radiomics, DL, and pathomics features. These features were used to construct and validate 4 prognostic prediction signatures: the radiomics signature (RS), the DL signature (DS), the pathomics signature (PS), and the MS. Predictive performance was assessed via Harrell concordance index (C-index) and hazard ratio (HR). The SHapley Additive exPlanations (SHAP) analysis was applied to interpret the signature prediction process. RESULTS: The RS, DS and PS demonstrated predictive ability, with C-indices of 0.735 (95% confidence interval‌ [CI]: 0.587-0.884), 0.767 (95% CI: 0.613-0.921) and 0.711 (95% CI: 0.555-0.867) in the validation cohort, respectively. The MS achieved the highest C-index among the signatures, with a C-index of 0.821 (95% CI: 0.660-0.982). Progression-free survival (PFS) significantly differed between high-risk and low-risk patients stratified by MS (log-rank test, P < 0.001). SHAP analysis quantified the contribution of individual features to the signature's predictions. CONCLUSION: The MS showed promising prognostic value for prognostic prediction in central conventional CS. It may provide useful support for postoperative risk stratification.