Deep learning based infectious keratitis classification system using multi-type anterior segment images

Abstract

Keratitis is an eye disease with a high risk of vision loss if not treated properly. The incidence of keratitis continues to increase as increased use of contact lens and the population aging. Infectious keratitis such as bacterial and fungal keratitis and non-infectious keratitis are difficult to diagnose because they do not show clinically complete distinction. If keratitis patients do not receive the right treatment at the right time, they suffer from decreased vision and complications. However, standardized diagnostic methods for keratitis do not exist and depend entirely on the experience of an ophthalmologist. Therefore, in this paper, we propose a keratitis diagnosis prediction system to assist the doctor's diagnosis that outputs the diagnosis prediction result and the suspected lesion area by inputting the anterior segment image. The proposed network contains Lesion Guidance Module (LGM) for the network to learn knowledge pointing suspicious lesions which can be ground of ophthalmologists’ diagnosis and to robust from elements which can be a cause of misdiagnosis such as reflected light in anterior segment image. It also utilizes a novel module, slit lamp Mask Adjusting Module (MAM), and loss functions for learning various types of images efficiently to overcome the data shortage problem. Experimental result demonstrated that the proposed network improves keratitis diagnosis accuracy and predicts more accurate lesion areas.