RetMobileNet: A New Deep Learning Approach for Multi-Class Eye Disease Identification

Image processing plays a vital role in disease identification, particularly in analyzing retinal fundus images to detect various eye conditions, such as Diabetic Retinopathy, Drusen, and Central Retinal Vein Occlusion.This study aims to develop a CNN for multi-class eye disease identification using retinal fundus images.The MuReD dataset combines 1855 images across 20 classes from ARIA, STARE, and RFMiD datasets.Eye specialists validated 1122 images, representing 12 disease classes.All images were standardized to 224×224 pixels.Data was split 80%:10%:10% for training, validation, and testing, respectively, using K-Fold Cross-validation (k=10).Seven CNN models (VGG19, InceptionV4, Resnet50, MobileNetV2, MobileNetV1, MobileNetV3 Large, and MobileNetV3 Small) were tested, with MobileNetV1 and MobileNetV2 showing the most promise.We innovatively enhanced the architectures of MobileNetV1 and MobileNetV2 through proper CLAHE configuration, optimization of training parameters (Batch size, learning rate, optimizer), and layer modification (skip connection, Dense layer, convolution layer, batch normalization, and dropout), culminating in a novel approach named Retina MobileNet (RetMobileNet).RetMobileNet consistently outperformed other models, achieving an average accuracy, recall, precision, and F1-score of over 90%.Our developed model, RetMobileNet, presents a significant advancement in the accurate identification of eye diseases from fundus images, with improved efficiency.