ViT-AMD: A New Deep Learning Model for Age-Related Macular Degeneration Diagnosis From Fundus Images

Abstract

Age-related macular degeneration (AMD) diagnosis using fundus images is one of the critical missions of the eye-care screening program in many countries. Various proposed deep learning models have been studied for this research interest, which aim to achieve the mission and outperform human-based approaches. However, research efforts are still required for the improvement of model classification accuracy, sensitivity, and specificity values. In this study, we proposed the model named as ViT-AMD, which is based on the latest Vision Transformer (ViT) structure, to diagnosis a fundus image as normal, dry AMD, or wet AMD types. Unlike convolution neural network models, ViT consists of the attention map layers, which show more effective performance for image classification task. Our training process is based on the 5-fold cross-validation and transfer learning techniques using Chula-AMD dataset at the Department of Ophthalmology, the King Chulalongkorn Memorial Hospital, Bangkok. Furthermore, we also test the performance of trained model using an independent image datasets. The results showed that for the 3-classes AMD classification (normal vs. dry AMD vs. wet AMD) on the Chula-AMD dataset, the averaged accuracy, precision, sensitivity, and specificity of our trained model are about 93.40%, 92.15%, 91.27%, and 96.57%, respectively. For result testing on independent datasets, the averaged accuracy, precision, sensitivity, and specificity of trained model are about 74, 20%, 75.35%, 74.13%, and 87.07%, respectively. Compared with the results from the baseline CNN-based model (DenseNet201), the trained ViT-AMD model has outperformed significantly. In conclusion, the ViT-AMD model have proved their usefulness to assist the ophthalmologist to diagnosis the AMD disease.