Fusion of Deep Learning Radiomics and Clinical Features for Individualized Prediction of Neoadjuvant Chemotherapy Response in Osteosarcoma

Background: Stage IIB conventional osteosarcoma exhibits highly heterogeneous responses to neoadjuvant chemotherapy, necessitating reliable, noninvasive pre-treatment evaluation systems. Methods: This retrospective study of 198 patients utilized pre-treatment T2-weighted magnetic resonance imaging (MRI) and baseline clinical data. Twelve classical convolutional neural network architectures were benchmarked to identify the optimal imaging backbone. This selected radiomics score was then integrated with core clinical variables using logistic regression to construct a multimodal model. Gradient-weighted Class Activation Mapping++ (Grad-CAM++) and SHapley Additive exPlanations (SHAP) algorithms provided image- and feature-level interpretability. Results: DenseNet121 achieved the highest mean cross-validated area under the curve (AUC; 0.775) and was selected, with AUCs of 0.797 (95% CI: 0.716–0.877) and 0.701 (95% CI: 0.513–0.877) in the training and test cohorts. Adding the clinical variables improved discrimination, yielding a test-cohort AUC of 0.827 (95% CI: 0.661–0.992) with significant reclassification gain over the radiomics-only model (NRI = 1.28, IDI = 0.177; DeLong p = 0.011). Grad-CAM++ activation concentrated on the solid medullary and extraosseous soft-tissue components, and SHAP ranked the radiomics score as the strongest contributor. Conclusions: Integrating deep learning radiomics with key clinical features provides a noninvasive tool for the individualized prediction of neoadjuvant chemosensitivity in osteosarcoma.