Beyond Static Fusion: A Mixture-of-Experts Framework for Multimodal Breast Cancer Classification
Faseela Chakkalakkal Abdullakutty, Younes Akbari, Somaya Ali Al-Maadeed, Rafif Al-Saady, Hadi Mohamad Abu Rasheed · 2025
Accurate breast cancer diagnosis requires integrating heterogeneous data sources such as histopathology images and structured clinical records. However, traditional deep learning models often struggle to effectively combine modalities due to static fusion strategies that ignore case-specific modality relevance. This paper introduces an adaptive Mixture-of-Experts (MoE) framework for multi-modal breast cancer classification. The architecture comprises two specialized experts: a convolutional neural network for image feature extraction and a multi-layer perceptron for clinical data processing. A learnable gating network dynamically computes instance-specific weights to combine expert outputs, enabling context-aware fusion tailored to individual patient profiles. Evaluated on a real-world dataset with histopathology and electronic medical records, the proposed model achieves state-of-the-art classification accuracy. Experimental results demonstrate the effectiveness of adaptive modality weighting in enhancing diagnostic precision and robustness for clinical decision support.