Image-based cataract diagnosis

Abstract

Cataracts, the clouding of the eye's lens, poses a global health challenge as a leading cause of visual impairment. There is a need for improved cataract screening and diagnosis as traditional diagnosis methods are limited in access, involving expensive equipment, and requiring great expertise.

Handheld devices such as slit lamp cameras promise greater portability and accessibility, but they often suffer from poorer image quality. Therefore, this project leverages image augmentation techniques, transfer learning from pretrained convolutional neural networks (CNNs), and combining patient metadata, to propose the Image+metadata model. The model achieved an accuracy of 0.960, F1-score of 0.959, a sensitivity of 0.960 and specificity of 0.960, making it comparable to the results from other studies while using a smaller dataset (n=187) and relatively lower quality images. To further validate the effectiveness of the model, saliency maps were generated to explain the predictions made by the model.

This approach holds the potential to vastly increase the accessibility of handheld cataract screening, offering an accurate, cost-effective, and approachable solution for cataract diagnosis and intervention, ultimately reducing the incidence of cataract-related visual impairment.