Accurate and Explainable Cataract Detection Using Eye Images Taken by Hand-held Slit-lamp Cameras

Cataract is a leading cause of visual impairment in the elderly. With a greying population globally, there is a pressing need to improve the accessibility of cataract screening. Hand-held Slit-lamp Cameras (HSCs) are often preferred in community eye screening due to their great portability and accessibility. However, the image quality from HSCs is generally inferior to that from conventional bulky fundus cameras. In this paper, we extend the pre-trained ResNet-18 neural network to analyze a limited number of HSC images (n=187) for cataract detection. Model accuracy is improved through augmenting training data samples and complementing the visual features with patients’ vision measurements. Explainability (of focal model areas for decision making) is attained via extracting the saliency maps using the Grad-CAM method. Our model achieves a high accuracy of 0.96, on par with state-of-the-art results reported in the literature. Our approach demonstrates the potential of large-scale community-based cataract detection using HSCs and our highly accurate and explainable AI-assisted model.