BoneXpertAI: An Explainable Multimodal Hybrid Expert Framework for Bone Cancer Detection, Segmentation and 5-Year Survival Risk Stratification

The diagnosis and prognosis of bone cancer is a challenging problem due to heterogeneous imaging appearances, lack of annotated cohorts, missing modalities, and interpretable decision support. The paper introduces BoneXpertAI, a multimodal hybrid expert system which integrates attention-based CNN and Transformer image encoders with radiological images with an MLP model of clinical and genomic data. Attention-based fusion module is trained on modality contribution and domain informed rules offer explicit clinical constraints and SHAP-based attribution offer explanation on a feature level. The strategy of a federated ensemble is tested under a simulation of a multi-site environment whereby raw data are local and model parameters are shared with a coordinator. The framework aids tumor detection/localization, segmentation, and 5-year survival-risk classifications. During the described internal analysis, BoneXpertAI demonstrated superiority to the chosen baseline settings. Nonetheless, it is a proof-of-concept small-group study, and the manuscript does not purport to be clinical superiority or deployment-ready. Clinical translation requires external validation, prospective analysis, stringent time to event analysis, and validation on larger independent curated cohorts.