Classification of Retinal Disorders from OCT Images using Attention based CNN

Efficient automated decision support systems for the detection of retinal disorders are crucial in ophthalmology. Optical Coherence Tomography (OCT), a widely used imaging modality, allows visualization and measurement of retinal layer thickness, aiding in the early detection of disorders such as age-related macular degeneration (AMD), diabetic macular edema (DME), and other abnormalities. Despite advancements, existing methodologies often lack generalization and primarily focus on entire retinal images, disregarding the central retinal region where most abnormalities manifest. This study proposes a deep learning model integrating attention mechanisms with a convolutional neural network (CNN) and auto-encoder for OCT image classification into four categories: Choroidal Neovascularization (CNV), DME, Drusen, and Normal. The attention mechanism emphasizes relevant features, while the auto-encoder detects anomalies effectively. Optimized using random search, the model achieves a remarkable accuracy of 97.8%, with precision, recall, and F2-scores of 98.4%, 98.3%, and 98.3%, respectively, demonstrating significant improvement over existing approaches. This model offers enhanced accuracy and efficiency for retinal disorder classification, promising improved diagnostic and treatment planning in clinical applications.