FreNet Uses Visual Priors for Medical Lesion Segmentation

Summary

Medical lesion segmentation remains difficult because complex backgrounds can interfere with image features and lesions can have highly varied shapes. The paper introduces FreNet, a feature-reconfiguration framework that applies pixel-level reconfiguration before encoding and feature-level reconfiguration during encoding. Its Implicit Prior Neural Network models a continuous spatial field and uses a visual prior from SAM to suppress background responses before the image enters the encoder. A Dual-domain Feature Reconfiguration module progressively updates backbone features during encoding. Within it, a Frequency Decoupling Module separates features in the frequency domain to improve foreground-background discrimination, while a Spatial Localization Module relocates features to improve spatial stability after decoupling. Experiments across nine medical image segmentation benchmarks and three imaging modalities report that FreNet outperforms the compared state-of-the-art methods. On the challenging ETIS dataset, it improves Dice by 5.0 percentage points over the state-of-the-art method and by 7.2 percentage points over SAM.