Deep learning for identifying corneal diseases from ocular surface slit-lamp photographs

Hao Gu,Youwen Guo,Xiaoqing Liu,Xingtao Zhou,Anji Wei,Lei Gu,Jianjiang Xu,Yi Lu,Shirong Xie,Zhengqiang Ye,Jiaxu Hong

doi:10.1038/s41598-020-75027-3

Abstract

To demonstrate the identification of corneal diseases using a novel deep learning algorithm. A novel hierarchical deep learning network, which is composed of a family of multi-task multi-label learning classifiers representing different levels of eye diseases derived from a predefined hierarchical eye disease taxonomy was designed. Next, we proposed a multi-level eye disease-guided loss function to learn the fine-grained variability of eye diseases features. The proposed algorithm was trained end-to-end directly using 5,325 ocular surface images from a retrospective dataset. Finally, the algorithm’s performance was tested against 10 ophthalmologists in a prospective clinic-based dataset with 510 outpatients newly enrolled with diseases of infectious keratitis, non-infectious keratitis, corneal dystrophy or degeneration, and corneal neoplasm. The area under the ROC curve of the algorithm for each corneal disease type was over 0.910 and in general it had sensitivity and specificity similar to or better than the average values of all ophthalmologists. Confusion matrices revealed similarities in misclassification between human experts and the algorithm. In addition, our algorithm outperformed over all four previous reported methods in identified corneal diseases. The proposed algorithm may be useful for computer-assisted corneal disease diagnosis.

Highlights

To demonstrate the identification of corneal diseases using a novel deep learning algorithm
Recent advances on deep learning algorithms, in particular convolutional neural networks (CNN), have made it possible to learn the most predictive features of disease directly from medical images when given a large dataset of labeled e xamples[5,6]
We examined the internal features via t-distributed Stochastic Neighbor Embedding (t-SNE)[22] where point clouds with different colors represent different disease categories

Summary

Introduction

To demonstrate the identification of corneal diseases using a novel deep learning algorithm. The assessment of the ocular surface, primarily the cornea and conjunctiva, by ocular slit-lamp examination is the foundation of corneal disease diagnosis This is highly dependent on the grader’s clinical experience, which is time-consuming and may have interobserver variation on the same patient. Gulshan et al.[10] demonstrated the detection of diabetic retinopathy through fine-tuning a pre-trained Inception-v38 network on retinal fundus images. Long et al.[17] developed a technique for diagnosis of congenital cataracts with acceptable diagnostic accuracy Their method was trained based on images covering the pupil area only. Unlike Long’s17 and Williams’s w ork[18], to cover a wider spectrum of ocular surface diseases, we utilized the whole ocular surface image and were not limited to the pupil This makes our algorithm capable of detecting corneal diseases related to the peripheral cornea and limbus

Methods

Results

Conclusion