A Comprehensive Review of Lightweight and Attention-Driven Deep Learning Models for Automated Cataract Detection

Abstract

Cataract is the leading cause of reversible blindness globally, accounting for nearly 51% of all blindness cases according to the World Health Organization (WHO). Traditional diagnostic procedures such as slit-lamp examination and ophthalmoscopy require expert supervision and expensive imaging devices, limiting their accessibility in rural and low-resource regions. Artificial Intelligence (AI) and Deep Learning (DL) have emerged as transformative technologies that can automate cataract detection from ocular images, enabling early diagnosis through mobile and edge devices. This review provides a comprehensive synthesis of recent research on lightweight and attention-driven deep learning frameworks for cataract detection. It critically evaluates four cornerstone approaches: the Optimised Lightweight Deep Edge Intelligent Model (SDLM), CNN-based cataract severity detection, Global–Local Attention Augmented
Models (GLAAM and GLAAI), and Mobile Net-based transfer learning. We present an extensive comparative analysis covering datasets, architectures, accuracy, computational efficiency, and deployment feasibility. Furthermore, this review explores interpretability techniques such as Grad-CAM and attention visualization that enhance the transparency of AI systems. The paper concludes by identifying emerging research directions, challenges, and opportunities toward federated, explainable, and globally accessible cataract detection systems.