International Journal of Science, Engineering and TechnologyAuthors: Md. Alamgir Hossain, Paritosh Chandra Ray, Md. Mehedi Hasan, Abdul Al Mahin, Ferdowsi Khatun
Abstract: Automated analysis of retinal fundus images can support scalable screening, but multi-disease classification remains difficult because retinal abnormalities may exhibit overlapping visual characteristics and clinically important classes are often underrepresented. This paper investigates a four-class retinal image classification framework using a strictly filtered subset of the Retinal Fundus Multi-Disease Image Dataset (RFMiD). The considered classes are No_Disease_Risk, diabetic retinopathy, age-related macular degeneration, and macular hole. Images with multiple disease labels, unselected diseases, or inconsistent annotations were excluded, resulting in 855 training, 294 validation, and 305 test images. Training data were balanced to 401 entries per class through sampling with replacement and stochastic augmentation. A pretrained EfficientNet-B0 with frozen parameters was used to extract a 1,280-dimensional representation from each image. Support vector machine, Random Forest, K-nearest neighbors, XGBoost, and CatBoost classifiers were subsequently evaluated. On the official test subset, CatBoost attained the highest accuracy of 0.9016, weighted F1-score of 0.9022, and macro-ROC-AUC of 0.9842. SVM achieved the highest macro precision and macro F1-score, at 0.9075 and 0.8628, respectively, while XGBoost produced the highest macro recall of 0.8774. Grad-CAM visualizations were generated for selected test images to examine the retinal regions associated with the model predictions. The results show that fixed deep representations can support effective retinal image classification with conventional learners, although the small number of unique ARMD samples limits the statistical reliability and generalizability of the findings.
