Developing a deep learning classifier for uveitis scoring on mouse OCT images

Abstract

Purpose : To develop and validate an automated approach to scoring OCT images of ocular inflammation in experimental animals using deep learning.

Methods : Ocular inflammation was generated using the primed mycobacterial uveitis (PMU) model. Anterior chamber and posterior chamber optical coherence tomography (OCT) images were obtained using the Envisu R2300. On each image, degree of inflammation was classified using a 6 step categorical system with scores ranging from 0-4+ by three masked graders (A). Training set images were labeled with the score assigned by at least 2 out of the 3 graders. Images were divided into a training, validation and test set in a 80:10:10 ratio and a Deep Learning classifier algorithm based on a modified ResNet50 but allowing higher resolution images, was developed. Accuracy of the classifier compared to the human score was determined for anterior and posterior chamber images separately. Agreement between the classifier and the human graders was determined using Cohen’s linear weighted Kappa. Comparison between human and classifier scores was performed using a confusion matrix on the test set.

Results : 1115 images (575 anterior chamber, 540 posterior chamber) were scored by human graders. Agreement between human graders was 0.87 and 0.83 on AC and PC images respectively. Agreement between the automated classifier and the human graders was 0.56 and 0.68 for AC and PC respectively. Absolute accuracy of the automated classifier was 56% for AC images and 57% for posterior images. The confusion matrix analysis (B) determined that accuracy was highest for scores of 0 which was also the most frequent score in the training set.

Conclusions : In this study, we found that human graders demonstrated strong agreement if one step differences inscore were accounted for with a weighted kappa. This is consistent with prior studies of inter-human agreement scoring in uveitis. Automated classification did not perform as well as human graders. Better performance and agreement between the automated and human graders was identified for posterior chamber images than on anterior chamber images. Performance was likely limited by the disproportionate number of score 0 images in the training set. Additional images at each score level will be required to improve classification accuracy.

This is a 2021 ARVO Annual Meeting abstract.

 

A. Representative OCT images for all scores
B. Confusion Matrix Analysis