Automated CAD System for Intermediate Uveitis Grading Using Optical Coherence Tomography Images

Abstract

Intermediate uveitis is a major cause of vitritis and can be considered a leading cause of blindness. Clinical records show that accurate detection and hence grading of vitritis will result in a great reduction of blindness rate. This paper proposes an automatic vitritis grading computer aided diagnostic (CAD) system using optical coherence tomography images (OCT), which consists of two stages. The first is a U-net convolutional neural network (U-CNN), which is used to segment the vitreous. The vitreous is very difficult to segment directly from the original OCT due to the high similarity in visual appearance with background tissues. Instead, the U-CNN is based on processing of an input proposed fused image (FI) that integrates the original image, a distance map, and an adaptive appearance map. To assess the vitritis severity, the second stage utilizes the cumulative distribution function of the vitreous intensity as a discriminatory feature for a two-level machine learning classifier with 4 classes (grades 0 - 3). System performance is evaluated on a 200 images dataset. Segmentation stage performance is evidenced by both Dice similarity coefficient of 98.8% and Hausdorff distance of 0.3 μm. Second stage performance is evidenced by the classifier accuracy of 90.5% for the first level and 81% for the second level. These results support using the proposed CAD as an aid to early diagnosis of uveitis. © 2022 IEEE.

Keywords Convolutional neural networks, Deep Learning, U-net, Vitreous inflammation grading, Vitreous segmentation