A MACHINE LEARNING-BASED APPROACH FOR THE DETECTION AND SEGMENTATION OF RETINAL VESSELS IN OPHTHALMIC DISEASE ANALYSIS

Abstract

The research proposes an advanced deep learning infrastructure which develops two unique neural architectures named PLS-Net (Pool-less Semantic Network) and PLRS-Net (Pool-less Residual Semantic Network) using Convolutional Neural Networks (CNNs) while being motivated by U-Net’s encoder-decoder methodology. The research utilized CHASE_DB1 dataset to improve retinal blood vessels detection and segmentation through two deep learning models called PLS-Net and PLRS-Net. This database consists of 28 fundus images from pediatric patients which dual experts annotated for ground truth reliability. The framework implements a detailed preprocessing flow that includes pixel resizing to 256 × 256 pixels through bilinear interpolation followed by pixel normalization to [0, 1] range and the application of real-time data augmentation through multiple transforms (0–90° rotations and flips combined with ±20% brightness changes). This preprocessing method effectively quadruples each dataset size. The satisfactory outcomes from performance assessment demonstrated that PLRS-Net delivered industry-leading results exceeding U-Net (96.72% accuracy) and Attention U-Net (97.85% accuracy) benchmarks. Its performance showed 99.70% accuracy, a 0.997 sensitivity rate, 0.998 specificity rate and 0.997 Dice coefficient and 0.9972 Intersection over Union (IoU). The comparison against current leading methods shows our models win while two primary obstacles remain; boundaries of vessels near optic discs get incorrectly identified due to similar intensities and our system could overfit the limited dataset. These research results establish the framework as an essential tool for telemedicine because it provides pathways to increase diagnostic access in underserved regions which have an extremely low ratio of ophthalmologists to patients as demonstrated by the Pakistan scenario. This research advances the field of retinal vessel segmentation through the combination of lightweight pool-less designs with residual learning and creates groundwork for extensive AI-driven eye care solutions worldwide.