LFU-Net: LoRA-enhanced frequency-aware U-Net with recursive residual attention fusion for retinal segmentation

Abstract

Abstract Long-tail segmentation is a crucial challenge in computer vision, where most models prioritize common head classes over rare tail classes. This problem is particularly prominent in retinal vessel segmentation, as conventional approaches often struggle to overcome underrepresented faint vessels, noise-induced boundary ambiguity, and excessive parameters that prohibit portable deployment. To address these challenges, we introduce LFU-Net, a lightweight and clinically applicable method for long-tail retinal vessel segmentation. It integrates a three-component ensemble: a Frequency-Aware Encoder with a Multi-Branch Frequency Convolution block, which uses wavelet decomposition to suppress noise and retain details; Hierarchical frequency-token enhanced Low-Rank Adaptation, which efficiently enhances the representation of tail classes (faint vessels) with minimal parameters; and a Recursive Residual Attention Fusion module to ensure vascular topological continuity. Extensive experiments on four public benchmark datasets demonstrate that LFU-Net achieves competitive performance compared to recent relevant models. Its lightweight nature supports real-time inference on portable devices. Ablation studies confirm the improvement contribution of each core component, indicating its potential utility in early disease detection when clinical resources are limited.