AnyMetric: Universal Metric Optimization for Medical Image Segmentation With Preference Learning

Medical image segmentation is crucial for accurate disease diagnosis and treatment. To assess segmentation performance, various evaluation metrics have been developed. Existing methods typically design surrogate loss functions tailored to specific metrics for model optimization. However, it is challenging and laborious to customize loss functions for all non-differentiable and complex evaluation metrics. In this paper, we propose Preference-Guided Segmentation Network (PGSNet), a novel framework that enables the unified optimization of arbitrary evaluation metrics, eliminating the need for handcrafted surrogate loss functions. To this end, PGSNet designs a preference-based optimization strategy, which optimizes the segmentation model by directly learning human-defined evaluation preferences rather than relying on loss functions. Specifically, PGSNet first generates diverse segmentation candidates by leveraging multi-scale representations that capture different levels of medical information. These candidates are then assessed and ranked using a reward function derived from target evaluation metrics, reflecting segmentation preference according to task metrics. Finally, segmentation preference optimization is developed to leverage these preference signals to guide model training, ensuring that the probability of generating superior segmentation results over that of inferior ones. Instead of designing specific surrogate loss functions, PGSNet offers a generalizable and flexible optimization framework capable of handling various metrics. Extensive experiments on several challenging public datasets demonstrate that PGSNet achieves superior segmentation results compared to existing methods.