Deep learning-based classification of infectious keratitis on slit-lamp images

Zijun Zhang; Haoyu Wang; Shigeng Wang; Zhenyu Wei; Yang Zhang; Zhiqun Wang; Kexin Chen; Zhonghong Ou; Qingfeng Liang

doi:10.1177/20406223221136071

Deep learning-based classification of infectious keratitis on slit-lamp images

Ther Adv Chronic Dis. 2022 Nov 14:13:20406223221136071. doi: 10.1177/20406223221136071. eCollection 2022.

Authors

Zijun Zhang¹, Haoyu Wang², Shigeng Wang², Zhenyu Wei¹, Yang Zhang¹, Zhiqun Wang¹, Kexin Chen¹, Zhonghong Ou², Qingfeng Liang³

Affiliations

¹ Beijing Institute of Ophthalmology, Beijing Tongren Eye Center and Beijing Key Laboratory of Ophthalmology and Visual Sciences, Beijing Tongren Hospital, Capital Medical University, Beijing, China.
² Beijing University of Posts and Telecommunications, Beijing, China.
³ Beijing Institute of Ophthalmology, Beijing Tongren Eye Center and Beijing Key Laboratory of Ophthalmology and Visual Sciences, Beijing Tongren Hospital, Capital Medical University, Beijing 100005, China.

Abstract

Background: Infectious keratitis (IK) is an ocular emergency caused by a variety of microorganisms, including bacteria, fungi, viruses, and parasites. Culture-based methods were the gold standard for diagnosing IK, but difficult biopsy, delaying report, and low positive rate limited their clinical application.

Objectives: This study aims to construct a deep-learning-based auxiliary diagnostic model for early IK diagnosis.

Design: A retrospective study.

Methods: IK patients with pathological diagnosis were enrolled and their slit-lamp photos were collected. Image augmentation, normalization, and histogram equalization were applied, and five image classification networks were implemented and compared. Model blending technique was used to combine the advantages of single model. The performance of combined model was validated by 10-fold cross-validation, receiver operating characteristic curves (ROC), confusion matrix, Gradient-wright class activation mapping (Grad-CAM) visualization, and t-distributed Stochastic Neighbor Embedding (t-SNE). Three experienced cornea specialists were invited and competed with the combined model on making clinical decisions.

Results: Overall, 4830 slit-lamp images were collected from patients diagnosed with IK between June 2010 and May 2021, including 1490 (30.8%) bacterial keratitis (BK), 1670 (34.6%) fungal keratitis (FK), 600 (12.4%) herpes simplex keratitis (HSK), and 1070 (22.2%) Acanthamoeba keratitis (AK). KeratitisNet, the combination of ResNext101_32x16d and DenseNet169, reached the highest accuracy 77.08%. The accuracy of KeratitisNet for diagnosing BK, FK, AK, and HSK was 70.27%, 77.71%, 83.81%, and 79.31%, and AUC was 0.86, 0.91, 0.96, and 0.98, respectively. KeratitisNet was mainly confused in distinguishing BK and FK. There were 20% of BK cases mispredicted into FK and 16% of FK cases mispredicted into BK. In diagnosing each type of IK, the accuracy of model was significantly higher than that of human ophthalmologists (p < 0.001).

Conclusion: KeratitisNet demonstrates a good performance on clinical IK diagnosis and classification. Deep learning could provide an auxiliary diagnostic method to help clinicians suspect IK using different corneal manifestations.

Keywords: auxiliary diagnosis; deep learning; infectious keratitis; slit-lamp photos.