XAI-CAD: An Explainable CAD Framework for the Classification of Diabetic Retinopathy

Diabetic retinopathy(DR) is a major vision-threatening consequence of diabetes. It can be identified by computer-aided diagnosis (CAD) automatically or manually. The manual inspection of the retinal images is time-consuming. CAD methods can identify and evaluate DR based on the morphological changes that occur in the retina due to long-term diabetes. Deep learning(DL) techniques have recently proliferated and are producing cutting-edge results when used to analyze medical images. Despite its effectiveness, the DL-based CAD system has a fatal flaw in clinical applications. Without a plausible rationale, medical professionals cannot trust a DL diagnosis. Given that a DL-based CAD system can demonstrate high prediction accuracy when applied to substantial test data, we may be inclined to believe it due to its black-box nature. Even if the prognosis is accurate, believing without a sound foundation becomes challenging. Hence, this paper aims to deploy the model-driven explainability of the CAD system (XAI-CAD) for the early detection and classification of DR. In the proposed framework, state-of-the-art DL based Convolution neural network (CNN) models along with transfer learning models like DenseNet121, ResNet-50, MobileNetV2, Inception-ResNetV2, Xception, and Inception-v3 have been evaluated, and compared on publicly available fundus image datasets. The Xception model has been proven as the best model in terms of accuracy and sensitivity of 92.72% and 98.7% respectively. Gradient-based interpretation has been performed for a visual explanation of the model's decision. Thus the resulting XAI-CAD classification becomes more informative and reliability also increases.