High Performance Breast Cancer Diagnosis From Mammograms Using Mixture of Experts With EfficientNet Features (MoEffNet)

Abstract

As breast cancer is a leading cause of death for women globally, there is a critical need for better diagnostic tools. To address this challenge, we propose MoEffNet, a cutting-edge framework that offers high-performance breast cancer diagnosis. MoEffNet is characterised by its innovative hybrid integration of EfficientNet and Mixture of Experts (MoEs), using two powerful techniques developed to enhance accuracy and efficiency. EfficientNet, known for its robust feature extraction capabilities, utilises compound scaling and depth-wise separable convolutions to capture image features across multiple levels of abstraction. This is combined with MoEs framework, which employs specialised expert networks to analyse distinct aspects of mammograms. MoEffNet analyses features at various levels: low-level for basic patterns, mid-level for detailed analyses, and high-level for complex contents. Features extracted from various EfficientNet model stages are assigned to specialised experts to optimise diagnostic precision. A dynamic gating mechanism (EffiGate) is introduced to ensure that the most relevant experts contribute to each diagnostic decision, by dynamically adjusting their influence based on input data characteristics. This approach ensures that the most effective experts are utilised for each case, resulting in superior accuracy. The scalability of MoEffNet is highlighted by its ability to adapt to various computational constraints and accuracy requirements, using EfficientNet’s architecture, which ranges from B0 to B7 models. We have validated MoEffNet’s effectiveness on three mammographic datasets (MIAS, CBIS-DDSM, and INbreast) achieving outstanding results (AUC >0.99 across all datasets), outperforming existing methods. Particularly, EfficientNet B1 and B2 models with three or four experts achieved the highest accuracy, demonstrating MoEffNet’s potential as a robust diagnostic tool for early breast cancer detection. Through its innovative hybrid model, robust feature extraction, dynamic gating, and specialised expert networks, MoEffNet sets a new benchmark in automated mammogram analysis, offering a powerful tool for more accurate and reliable breast cancer diagnosis.

Keywords

• Breast cancer diagnosis
• computer-aided diagnosis
• deep learning
• machine learning
• multi-view analysis
• mammography