Artificial intelligence based detection of infectious keratitis using slit-lamp images

Abstract

Purpose :
Slit-lamp examination, to detect ocular surface diseases including fungal keratitis (FK), bacterial keratitis (BK), acanthameoba keratitis (AK), and post-surgical infections, is intricate which requires high expertise. There is a dearth of skilled professionals in developing countries. Against this backdrop, we envisage to develop an artificial intelligence (AI) based system to enable early diagnosis, high throughput screening and remote eye-care. To this end, we attempted binary (healthy-diseased) classification based on slit-lamp images as a first step towards developing a comprehensive AI based disease detection tool.

Methods : Noting the prevalence, we considered only three diseases (FK, BK and AK) for initial AI model development. 928 subjects (364-healthy-both-eyes, 291-FK-single-eye, 96-BK-single-eye and 177-AK-single-eye) were recruited for the study. Slit-lamp (Haag-Streit BX 100 with Canon 40D camera) examination was performed under diffused illumination and 10X magnification. A total of 1292 images were acquired (Figure). Inspired from the ability of deep convolution neural networks (CNNs) to classify image data on par with humans, we trained and tested three CNN based architectures AlexNet, VGG-16 and VGG-19. Training and testing was performed with a split ratio of 80:20.

Results : Training accuracies of all three models (AlexNet, VGG-16 and VGG-19) were observed to be 100% while the testing accuracies were observed to be 100%, 99.1% and 99.6% respectively (Table). Further, the sensitivity (specificity) values for three models were observed to be 100% (100%) and greater than 99% (99%) during training and testing.

Conclusions : The proposed deep learning models have shown close to 100% accuracy, with AlexNet performing marginally better. Although the results look promising, it is wise to corroborate the performance with a larger dataset which is under progress.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Dataset used in the study, training and testing was performed with a split ratio of 80:20

 

Performance comparison of three CNN architectures under consideration