Deep Learning for Three Types of Keratitis Classification based on Confocal Microscopy Images

Abstract

Accurate diagnosis of keratitis is important for the follow up treatment. The confocal microscope can scan different depth and layer of the cornea, therefore is an important tool for clinical diagnosis of keratitis. We collected, augmented and preprocessed the confocal microscopic images. In this paper, three kinds of infectious keratitis samples including viral keratitis, bacterial keratitis, and fungal keratitis were classified with ResNet (Residual Network). The results show that the recognition rate of three kinds of keratitis can reach 91.82%, and the accuracy rate of single keratitis could reach 99.09%. In addition, cross-validation was performed on each patient in the dataset. The classification accuracy rate reached 75.00%). This work extended the previous work of identifying fungal keratitis only to three categories and reach a good classification rate of keratitis.