Dense-MoE vs Lite-MoE: A Gating-Weight-Aware Pruning Framework for Unpaired Multimodal Breast Cancer Diagnosis

Abstract

This study proposes a unique Mixture-of-Experts (MoE)-based deep learning framework for the effective use of unpaired multimodal images in breast cancer diagnosis. Mammography (MG), ultrasonography (US), and magnetic resonance imaging (MRI) data are modeled as independent expert networks based on DenseNet-121, and a Gating Network mechanism is integrated to dynamically determine the contribution of each modality in the decision-making process. Thus, the model offers a flexible and clinically relevant structure without requiring all modalities to be available in every patient. The unique contribution of the study, the Gating-Weight-Aware Selection strategy, performs pruning by considering not only the individual Area Under the Curve (AUC) performance of the experts but also their usage rate within the model. Experimental results show that the Dense-MoE model achieves an AUC of 0.9661 and 89% accuracy. The pruned Lite-MoE model, despite reducing the number of parameters by approximately 33%, largely maintains its performance with an AUC of 0.9529 and 88% accuracy. Clinical evaluations of the proposed model were examined using Gradient-weighted Class Activation Mapping (GradCAM) heatmaps. The results showed that the model has the potential to be a flexible, scalable, and high-performance clinical decision support system in scenarios where data is incomplete.