Pixel attention meets M-shaped networks: A cutting-edge AI solution for diabetic retinopathy classification and stroke risk prediction

Abstract

Visual impairment is a major global health concern, with Diabetic Retinopathy (DR) recognized as a leading cause of vision loss and blindness among individuals with diabetes. Despite its widespread occurrence, the early detection of DR poses significant challenges due to the subtlety of initial symptoms, the requirement for expert analysis of retinal images, and the drawbacks of traditional diagnostic techniques, which can be time-consuming, subjective, and susceptible to human error. DR impacts the retina and the ocular blood vessels, making timely and accurate diagnosis crucial to prevent irreversible damage. However, recent advancements in Artificial Intelligence (AI) have opened up new avenues in medical diagnostics, offering sophisticated approaches for the detection and classification of DR with impressive accuracy. This article introduces a Pixel Attention-based M−Shaped Architecture (PAM), an innovative AI-powered diagnostic tool designed to enhance the detection and classification of DR. The PAM system consists of two key components: 1) a Pixel Attention (PIAT) network that aids in the precise identification of abnormalities in retinal blood vessels, and 2) an M−shaped neural network architecture that delivers strong segmentation and classification capabilities. Trained and validated on three diverse benchmark datasets with over 35,000 fundus images spanning multiple DR severity levels, PAM exhibits exceptional adaptability across different levels of DR severity (mild, moderate, severe, and proliferative), achieving state-of-the-art results with an accuracy of 98.73%, precision of 98.82%, sensitivity of 98.67%, specificity of 98.72%, F1-score of 99.13%, and a Dice coefficient of 98.74%. Comparative studies demonstrate that PAM outperforms current methodologies, especially in terms of accuracy and F1-score. Its key advantage is its ability to offer healthcare professionals a scalable, efficient, and dependable resource for improving clinical decision-making. Moreover, by focusing on pixel-level vascular pathology, PAM can simultaneously yield insights into systemic microvascular health, specifically functioning as an early-warning tool for stroke risk. This dual-purpose utility positions PAM as a powerful asset for telemedicine, broad screening programs, and AI-enhanced healthcare systems, where a single retinal exam can address both ocular and cerebrovascular health. By enabling timely and accurate diagnosis of DR and proactively recognizing individuals at elevated risk for stroke, PAM can significantly reduce vision loss associated with diabetes and aid in stroke prevention. This approach effectively connects advancements in AI with the clinical requirements of ophthalmology and neurology. Current efforts involve implementing PAM on mobile devices and performing external clinical validations to confirm its applicability in real-world settings and its impact across multiple domains.