Abstract

Info

Title:: Automatic diagnosis method for keratitis based on regional image patches from slit-lamp images

Author(s):: JIANG Jiewei1; DING Ke1; FENG Yangyang1; XIN Yu1; GONG Jiamin1; LI Zhongwen2; 1. School of Electronic Engineering, Xian University of Posts and Telecommunications, Xian 710121, China 2. Ningbo Eye Hospital, Wenzhou Medical University, Ningbo 315000, China

Keywords:: keratitis automatic diagnosis slit-lamp image patch feature fusion convolutional neural network cost-sensitive

Abstract:: Abstract: A method that integrates the features of image patches from corneal lesions and conjunctival congestion-like complications is proposed to address the limitations of manual keratitis diagnosis (i.e., time-consuming, laborious, high subjectivity) and the generally low accuracy of automatic keratitis diagnosis based on original slit-lamp images. Specifically, samples are acquired from the corneal and conjunctival regions. A cost-sensitive convolutional neural network is then used to extract and concatenate the high-level features of these image patches. After dimensionality reduction through principal component analysis, the processed features are input into the fully connected layer for classification. Trained and evaluated on 6 414 slit-lamp images collected from Ningbo Eye Hospital, the proposed method achieves accuracies of 97.8%, 98.6%, and 97.0% for keratitis, normal cornea, and other abnormal corneas, respectively. This method effectively integrates relevant features and provides a feasible solution for high-accuracy keratitis diagnosis.