Multi-label Deep Learning Framework for Early Detection of Macular Hole Diseases

Macular holes were first identified in the context of trauma and have since been linked to a large number of ocular disorders. In this article, four different convolutional neural network (CNN) models have been selected based on their abilities to extract useful information. We trained these models on 3200 fundus pictures from 46 separate fundus cameras for multi-label classification tasks. The suggested models in our work have predominantly shown a higher accuracy for the dataset under consideration. The comparison among different models relies upon certain evaluation metrics conducted for 15 epochs (early stopping) with 60 iterations each. The proposed method for detecting macular hole disease attained an accuracy of 92% using VGG19, 92% with VGG16, 92% with EfficientNetB0, and 93% with ResNet152 on the Retinal Fundus Multipath Imaging Dataset (RFMiD). Out of the four models, ResNet152 dominates over the rest for the RFMiD dataset. This method lays the groundwork for developing a diagnostic support system, thereby assisting ophthalmologists in identifying and assessing retinal diseases.