LHR-Net: a lightweight high-resolution network for retinal vessel segmentation

Retinal vessel segmentation plays a critical role in the diagnosis of ophthalmic diseases. However, challenges arise due to low image contrast and the complex morphology of blood vessels, hindering automatic segmentation. To overcome these challenges, we propose lightweight high-resolution network (LHR-Net), a LHR-Net built upon HR-Net. LHR-Net includes a high-resolution primary path, two low-resolution branches, and a multi-scale feature extraction branch. Before-activation residual blocks (BRBs) are incorporated to improve the network's feature extraction capability. The high-resolution main path is further enhanced with a parallel channel attention mechanism to capture rich semantic and spatial information, improving microvessel prediction. Additionally, atrous convolution with varying dilation rates is used to capture multi-scale features of the blood vessels. The performance of LHR-Net was evaluated using the DRIVE, STARE, CHASE_DB1 and HRF fundus datasets, achieving accuracy of 96.96%, 97.40%, 97.38% and 96.78%, sensitivity of 84.70%, 84.60%, 85.06% and 80.52%, and area under curve of 98.69%, 98.76%, 98.86% and 98.20%, respectively. Compared to other state-of-the-art methods, our network delivers superior segmentation performance and is more lightweight, with only 1.0 M parameters.