Explainable XceptionNet-Based Keratoconus Detection Using Infrared Images and Edge Computing

Abstract

Keratoconus is a progressive eye disease affecting the cornea, potentially leading to vision impairment. Early detection is crucial to prevent long-term damage. Artificial intelligence (AI) with edge computing can significantly aid in the early diagnosis of keratoconus by reducing the computation time and increasing diagnostic accuracy. In this letter, we utilize a customized explainable Xception network for keratoconus detection. For this clinical infrared (IR) images are obtained from hospital under defined protocol. The IR images are preprocessed and pupil regions are extracted. The extracted regions are applied to XceptionNet, and integrated gradient-weighted class activation maps (GradCAM) are obtained. Experiment are performed to evaluate the performance of the network. Results show that the proposed approach is able to detect keratoconus images. The proposed approach yielded the average accuracy of 85%. GradCAM outcomes show that the trained model detected the regions similar to clinical interpretation. We assess the model's performance on an edge device, demonstrating its potential for real-time, on-site diagnostics. This application not only improves classification accuracy but also enhances interpretability, thereby assisting ophthalmologists in making informed decisions with greater confidence. Thus, the proposed architecture could be useful in clinical condition.