Retinal Disease Identification From Oct Images Using Dictionary Learning and Yolov8 Publisher Pubmed



Hashemi SS ; Asadi M ; Rabbani H
Source: Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS Published:2025

Abstract

Automated identification of retinal diseases is critical for early and accurate diagnosis and treatment. This study presents a novel methodology that integrates representative dictionary learning with a deep neural network based on YOLOv8, to identify retinal diseases from OCT images. Applied to 2 OCT image datasets with 3 classes (NORMAL, DME, CNV for the first dataset and NORMAL, DME, AMD for the second dataset), the algorithm comprises four key steps. Firstly, OCT images are preprocessed using normalization and resizing methods. Then, minibatch online dictionary learning is employed to discover optimal atoms representing fundamental data features for each class. Subsequently, the learned atoms of dictionaries transform the dataset into lower-dimensional sparse encoded images, and finally, YOLOv8n classification model is trained on the transformed data through fine-tuning technique. With accuracies of 0.97 and 0.96 on both test datasets, the results indicate superior outcomes compared to most contemporary methods and significantly reduced computational costs.Clinical Relevance - This study presents an efficient method for automated retinal disease identification from OCT images, serving as an intelligent assistant for ophthalmologists to support early detection of macular diseases like DME and AMD. Its lightweight design enables deployment in clinics and telemedicine settings without requiring continuous internet access, ensuring reliable diagnosis even in resource-limited environments. © 2025 IEEE.