Mask-UnMask Regions (MUMR) Framework for Classifying AMD Grades Using Inter-Regional Interaction Analysis

Early diagnosis and effective treatment of age-related macular degeneration (AMD), a leading cause of vision impairment, are critically dependent on accurate grading. This paper presents a novel framework, named Mask-UnMask Regions (MUMR), designed to differentiate between normal retina, intermediate AMD, geographic atrophy (GA), and wet AMD using standardized retinal fundus images with an input resolution of 1024 \times 1024 pixels. The framework initiates with the downscaling of images to a quarter of their original size via a Preserving High-Frequency Information (PHFI) module, which retains key details essential for further analysis. Additionally, we developed a simple, lightweight, yet efficient ResNet-like network for feature extraction and introduced a Region Interaction (RI) module. This module incorporates Adaptive Mask and UnMask Sub-Modules, identifying significant regions while reconstructing less relevant areas using a direction-constrained self-attention mechanism to ensure the learning of global structural cues critical for AMD grade classification. The proposed method was evaluated on a dataset comprising 864 retinal fundus images. Our model consistently outperforms state-of-the-art approaches, achieving mean accuracy, mean F1-score, and mean Cohen's Kappa of 92.55%, 92.59%, and 89.97%, respectively. In the binary classification task of distinguishing between Non-AMD and AMD cases, the proposed approach also surpasses competing models, achieving mean accuracy, mean F1-score, and mean Cohen's Kappa of 97.11%, 97.03%, and 94.06%, respectively. Furthermore, statistical analysis of these metrics confirms that the improvements are statistically significant, demonstrating the robustness and improved performance of our proposed framework in AMD grading.