Explainable Classification of Macular Degeneration Using Deep Learning

Age-related macular degeneration is one of the leading causes of vision loss in individuals. This paper proposes a fundus image-based analysis and automatic grading of AMD using deep learning techniques. We utilize data points from three different datasets and perform augmentation and data sampling to equalise the data distribution. We apply deep learning by proposing EfficientNet-B3 for training a three-class classification model which categorizes the fundus images as having no AMD, mild AMD or severe AMD. The results obtained are evaluated on various metrics and we obtain 93.6% accuracy. We also analyse the trained model by visualising the information learned by it using class activation map algorithms. The accuracy of the proposed method outperforms the majority of the existing methods. The confidence in the proposed model is further increased by the CAM algorithms which indicate that the model predicts the output based on the presence of retinal lesions.