MSCA++-UNet: Image Segmentation Model with Coordinate and Channel Attention Mechanisms

Abstract

Retinal vessel segmentation in fundus images is critical for ophthalmic and systemic disease diagnosis, yet it is challenged by intricate vessel structures, low image contrast, and fine detail loss in feature extraction. Existing U-Net-based deep learning methods suffer from insufficient channel feature utilization, incomplete single-dimension attention enhancement, and poor fine vessel continuity, while lightweight models trade accuracy for efficiency, limiting clinical application. This paper proposes MSCA++-UNet,a novel segmentation model integrating a Multi-Scale Coordinate and Channel Attention Plus Plus (MSCA++) dual module into the U-Net backbone. The MSCA++ module realizes parallel computation and early fusion of multi-scale coordinate and adaptive channel attention, capturing spatial positional dependencies and dynamic channel correlations synergistically; combined with the lightweight AttentionDoubleConv block and optimized encoder-decoder fusion, it enhances edge preservation and feature discrimination with high computational efficiency. Extensive experiments on the DRIVE dataset (comparative, ablation, quantitative/qualitative analyses) show that MSCA++-UNet achieves a Dice coefficient of 0.787, mIoU of 79.7% and loss of 0.3624, outperforming baseline and single-attention variants in key metrics. It improves fine vessel detection accuracy and continuity, suppresses background noise, and balances performance (2.92 M parameters, 7.48 G FLOPs) with efficiency for edge device deployment. This work verifies the dual attention fusion strategy’s effectiveness, and MSCA++-UNet provides a high-precision, clinically feasible solution for fundus image analysis, supporting computer-aided diagnosis and large-scale retinal screening.