Abstract:

This article presents a comparative analysis of various training strategies for convolutional neural networks (CNN) aimed at the automatic classification of retinal diseases, with the goal of enhancing medical diagnosis. Using the RFMiD dataset, which contains 3,200 fundus images representing 46 different pathological conditions, multiple CNN architectures, including DenseNet201, RestNest50, RestNest51, and EfficientNetB4, are evaluated. Approaches such as transfer learning are explored, and the precision results obtained from each model are analyzed
in detail. Comprehensive analyses include the distribution of diseases within the dataset and the effectiveness of each strategy, highlighting that DenseNet201 achieves an outstanding accuracy
of 95.3 percentage. This study not only provides a comparative evaluation of the techniques employed but also contributes to the improvement of computer-aided diagnostic systems, suggesting future applications in the field of ophthalmology. Keywords: Convolutional Neural Networks, Retinal Diseases, Automatic Classification, Transfer Learning, Computer-Aided Diagnosis, RFMiD Dataset, DenseNet201, EfficientNetB4. Index Terms—Deep Learning, RFMiD, retinal diseases, classification, convolutional neural networks, transfer learning.