A GLOBAL-LOCAL FEATURE LEARNING FOR RETINAL VESSEL SEsGMENTATION USING AN IMPROVED U-NET

Abstract

The diagnosis of ophthalmic diseases such as glaucoma, diabetic retinopathy, hypertension, and cardiovascular disorders can be recognized through retinal vessel segmentation from fundus images. The low contrast, noise, blurred boundaries, and substantial variations in vessel thickness make the vessel segmentation a challenging task for semantic segmentation of retinal vessels. To mitigate these limitations, an enhanced U-Net variant is proposed. The framework integrates an initial Inception block, a sequential Global-Local (GL) feature engineering module utilizing zooming and shrinking operations, parallel 1×1 and 3×3 skip paths coupled through element-wise addition, and a terminal Atrous Spatial Pyramid Pooling (ASPP) stage. Internal representations across these stages were systematically profiled using Mutual Information (MI) metrics. Quantitative evaluations demonstrate the network's strong performance across standard benchmarks, recording an mIoU of 80.87%, a Mean BF-score of 91.89%, a center-line Dice (clDice) of 80.83%, and an 84.6% vessel classification accuracy alongside a reduced 15.4% false-negative rate. Ablation trials validate the individual architectural contributions: introducing the Inception layer increases IoU from 50.84% to 59.59%, whereas employing element-wise addition over concatenation in the GL module raises clDice to 81.04% (versus 80.72%). Additionally, MI profiling confirms that average pooling and multi-rate ASPP convolutions retain vital spatial entropy and micro-vascular geometry far more effectively than standard operations.