Reproduce “Model Agnostic Preference Optimization for Medical Image Segmentation” on OpenResearch

About this paper

Yunseong Nam, Jiwon Jang, Dongkyu Won, Sang Hyun Park, Soopil Kim

Researchers at the Daegu Gyeongbuk Institute of Science and Technology (DGIST) introduced Model-Agnostic Preference Optimization (MAPO), a framework that leverages dropout during inference to generate diverse segmentation candidates for medical images. This approach consistently enhanced segmentation performance and boundary adherence across various deep learning architectures and datasets, improving stability and precision in fine-tuning.

Key insights

• Dropout, traditionally a regularization technique, can effectively generate diverse and meaningful segmentation hypotheses during inference, providing rich signals for preference optimization.
• Combining a Direct Preference Optimization (DPO) objective with traditional supervised losses in an online training loop stabilizes model optimization, particularly for unstable architectures, and enhances structural consistency.
• Leveraging relative preference signals derived from dropout-generated candidates, rather than solely relying on absolute pixel-level ground truth, improves boundary adherence and overall segmentation quality across various medical imaging tasks.

Summary by alphaXiv.