HME-NET: A HYBRID MULTIMODAL ENSEMBLE FRAMEWORK FOR EXPLAINABLE AND EFFICIENT BONE CANCER DIAGNOSIS

Early and precise diagnosis of bone cancer plays a critical role in improving patient outcomes and guiding treatment strategies. Although traditional radiology gives good results, the way experts interpret the images can cause delays or errors in diagnosis. For this reason, we introduce a new integrative framework called Hybrid-Multimodal Ensemble (HME) which mixes MRI, CT and PET images with both radiomicsand deep learning to detect bone cancer automatically. Handcrafted radiomic descriptors are extracted using PyRadiomics and these descriptors are fused with hierarchical semantic features extracted by EfficientNet-B3 and Swin Transformer models. All of these features are brought together by an Explainable Feature Fusion Module (EFFM), with SHAP and LIME providing interpretability, diminishing the number of features and using an attention-based process. Classification of the final feature vector is done by an ensemble of LightGBM, Logistic Regression and a two-layer MLP. Real-time use is possible through quantization and the system allows training to happen on various devices through federated simulation. Evaluations conducted on this curated, mixed dataset found that the hybrid model (HME) scored better, with 98.1% accuracy, 97.4% precision, 98.3% recall, 97.85% F1-score and 0.987 AUC. In addition, looking at deployment metrics shows how prepared the model is to be used in clinical settings. Based on these results, the proposed HME framework ensures excellent diagnoses while also offering good understanding, privacy and ability to grow with more data, making it suitable for clinical use in bone cancer care.