Reproduce “COME: Dual Structure-Semantic Learning with Collaborative MoE for Universal Lesion Detection Across Heterogeneous Ultrasound Datasets” 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
Lingyu Chen, Yawen Zeng, Yue Wang, Peng Wan, Guo-chen Ning, Hongen Liao, Daoqiang Zhang, Fang Chen
COME achieves robust lesion detection across highly diverse ultrasound datasets by integrating dual shared experts (structural and semantic) with source-specific experts within a Mixture-of-Experts framework. This approach yields a mean Average Precision of 0.6587 in inter-organ integration, outperforming previous methods by 5.14% to 7.95%.
Key insights
- Integrating universal structural and semantic prior knowledge with source-specific expert specialization effectively addresses extreme data heterogeneity in medical imaging tasks like lesion detection.
- Novel clustering strategies (Fine2Coarse and Multi-Step) are crucial for mitigating multi-dataset noise and attribute entanglement, enabling better specialization of experts within an MoE framework.
- Tailored loss functions that enforce expert-source mapping and balance expert load are essential for achieving stable, specialized learning and preventing under-utilization of experts in heterogeneous Mixture-of-Experts models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.