A Deep Learning-Based Approach for Retinal Image Classification Using EfficientNet-B0 and Channel Attention

Retinal fundus images contain visual information that can assist in the identification of several ocular conditions. This paper presents an image-classification pipeline based on EfficientNet-B0 with Squeeze-and-Excitation (SE) channel attention for the Ocular Disease Intelligent Recognition (ODIR) dataset. The pipeline applies Contrast Limited Adaptive Histogram Equalization (CLAHE) to the luminance channel in LAB colour space, extracts image features using an ImageNet-pretrained EfficientNet-B0 backbone, and applies channel-wise feature recalibration before multi-label prediction. Gradient-weighted Class Activation Mapping (Grad-CAM) is used to visualize image regions associated with model outputs. The original manuscript reports an overall ODIR accuracy of 98.5%. It also reports class-wise precision values from 0.91 to 0.96, recall values from 0.93 to 0.98, and F1-scores from 0.93 to 0.97 across the eight ODIR label categories. These values are reproduced from the supplied manuscript and require confirmation against the original experiment logs and the exact test-set protocol. The study is presented as an applied deep-learning implementation rather than a clinically validated diagnostic system. Further work is required to document the complete experimental configuration and assess performance on independent data.