A Federated Faster R‐CNN Framework With Contour Detection and Grad‐CAM for Interpretable Multi‐Class Retinal Disease Diagnosis in Fundus Images

Retinal abnormalities are increasing day by day vastly. Mostly the vision loss in children are leading challenges in the current era. Retinal diseases are increasing which has deepened the need for accurate and privacy preserving diagnostic systems which enable the fast and early diagnosis of retinal abnormalities and improve patient care. This study proposes a novel framework assimilating faster region‐based convolutional neural network (R‐CNN) for automatic contour detection and federated learning for privacy preserving decentralised training in the domain of retinal fundus image categorisation. The dataset, which included annotated images of four retinal fundus pathological images—choroidal neovascularisation, myopia, normal retina and tessellated fundus, which were further fragmented among two fictitious clients. Stochastic gradient descent (SGD) was used to train an adapted Faster R‐CNN model backbone over 10 epochs, resulting in a substantial drop in training loss from 19.79 to 3.08. Each category endured inference, and Matplotlib was used to show the bounding box localisation and class predictions. Gradient‐weighted class activation mapping (Grad‐CAM) was also used for explainability, which generates heatmaps that precisely identify problematic areas in individual client fundus images. The federated learning configuration meritoriously preserved data privacy while sustaining model performance during the process. Experimental results which were obtained, indicate that all classes have worthy prediction confidence, with tessellated fundus images presenting the highest success rate. The study shows impending for practical clinical applications in the diagnosis of ophthalmic diseases due to its efficient object‐level feature learning, scalable deployment, and better‐quality interpretability.