Deep learning-based approach for corneal ulcer screening

Abstract

Corneal ulcer is a common corneal symptom that, upon infection, can lead to destruction of the corneal tissues, resulting in corneal blindness. To ease the corneal ulcer screening process, this paper introduces a deep transfer learning architecture based on various backbone networks to help identify two severity levels of the symptom: early stage and advanced stage. The total of 15 well-known deep convolutional neural networks are used as the base model. The proposed transfer learning-based architectures are trained, validated, and tested on 426, 143, and 143 fluorescein staining slit-lamp images from the public SUSTech-SYSU dataset. The experimental results show that ResNet50 is the best model achieving the best accuracy, sensitivity, F1 score, and Cohen’s kappa of 95.10%, 94.37%, 95.04%, and 0.9021, respectively, on the blind test set of the cropped corneal images. This model is further evaluated on an external dataset and its prediction is also explained using Integrated Gradients to provide an insight into its generalization performance.