Assessing the impact of image quality on deep learning classification of infectious keratitis

Abstract

Purpose : To investigate features affecting the quality of external corneal photographs, and their impact on a convolutional neural network (CNN) trained to classify bacterial and fungal corneal ulcers

Methods : A CNN was trained and tested using photographs of culture- and stain-proven bacterial and fungal corneal ulcers from subjects presenting to Aravind Eye Hospital in Madurai, India between 1/1/21 and 12/31/21. 897 images obtained by handheld cameras were labeled according to 5 quality parameters: gaze direction, eyelid position, exposure, focus, and light reflection. A MobileNet deep CNN was trained on a separate set of corneal ulcer images to perform multilabel classification of bacterial and fungal keratitis. The relative diagnostic performances of the model on this image set with regard to presence, absence or a combination of varying quality parameters were compared. Gradient class activation (Grad-CAM) heatmaps were generated to qualitatively assess which image regions were most influential on CNN predictions. CNN performance was evaluated using area under the receiver operating characteristic curves (AUROC) and area under the precision recall curves (AUPRC). Logistic loss was calculated to measure individual prediction accuracies.

Results : The CNN attained an AUROC of 0.83 (95% CI, 0.78-0.87) and AUPROC of 0.83 (95% CI, 0.80-0.87). Cases with either significant light reflections or eyelid obscuration were associated with statistically greater CNN performance than those without. No other quality parameter significantly influenced CNN performance. In a subset of cases correctly classified by the CNN, 78% of Grad-CAM heatmaps indicated the cornea and associated infiltrate as having the highest diagnostic relevance. However, the remaining 22% revealed a more diffuse region of “interest” to the ocular surface and eyelids, excluding the cornea entirely in 50% of these cases.

Conclusions : The CNN demonstrated expert-level diagnostic performance, even in cases of suboptimal image quality, implicating handheld photography as a viable means of capturing corneal images for deep learning analysis. Future studies will investigate use of smartphone cameras and image sets with greater variance in image quality to further explore the influence of these parameters on model performance.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.