A Two-stage Algorithm for Automatic Diagnosis of Keratitis

Abstract

Keratitis, as a corneal disease, is an important factor resulting in corneal blindness. Therefore, the key to prevent corneal blindness is how to diagnose keratitis with high accuracy. Hence, in this paper, a two-stage deep learning algorithm for automatic diagnosis of keratitis is proposed based on automatic localization of ocular surface. Firstly, the target detection algorithm, Single Shot MultiBox Detector (SSD) is used to automatically locate the conjunctival and corneal regions on the slit-lamp images, and remove the eyelid and surrounding noise area. Then, DenseNetl21 is applied to automatically classify the located conjunctival and corneal regions, so as to realize the diagnosis of keratitis, other corneal abnormalities, and normal cornea. Based on the dataset of 6,567 slit-lamp images, the accuracy of the proposed method improves by 0.8% compared to categorizing raw images directly, and also improves by 2.1% and 2% respectively on external test sets from two other hospitals. This study provides an effective artificial intelligence algorithm for assisting doctors in the early diagnosis of keratitis, and has great significance for the treatment of keratitis.