Reproduce “MammoMix: Leveraging Mixture of Experts for Robust Mammogram Breast Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Dinh Tan Nguyen, Hoang Quan Dang, Chen Zhang, Sai Ho Ling

MammoMix introduces a Mixture-of-Experts (MoE) framework, enhanced with confidence calibration (MoCAE), for robust breast lesion detection across diverse mammography datasets. This system maintains high detection accuracy and improves the identification of small and challenging lesions by specializing expert models and fusing their calibrated predictions, effectively addressing data heterogeneity.

Key insights

  • An MoE architecture with specialized experts significantly improves robustness and generalization capabilities for breast lesion detection across heterogeneous mammography data by effectively handling diverse data distributions.
  • Confidence calibration of individual expert predictions is crucial for reliable ensemble performance, ensuring that expert contributions are accurately weighted based on their true predictive accuracy.
  • The fusion of calibrated expert predictions through techniques like Soft NMS and Score Voting can consolidate fragmented detections and yield more precise and complete lesion localizations than individual experts.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.