A Generalized and Interpretable Multi-Label Multi-Disease Screening System for Ocular Anterior Segment Disease Detection

Abstract

Objective

To develop and validate a multi-label, multi-disease, well-generalized, and interpretable screening system applied to the detection of common ocular anterior segment diseases based on ocular surface slit-lamp images.

Design

A multicenter artificial intelligence diagnostic study.

Participants

A total of 1990 patients were randomly selected from 2 medical centers: the Second Affiliated Hospital of Zhejiang University and the Affiliated People’s Hospital of Ningbo University, between November 2016 and March 2022.

Methods

The data set was retrospectively collected from 2 clinical centers and composed of 5132 anonymized slit-lamp images of 13 ocular anterior segment diseases. The screening system was trained and validated in the internal data set composing randomly selected phenotypes and was tested in both internal and external data sets with less trained or new phenotypes included. The performance of the model was further compared with ophthalmologists.

Main Outcome Measures

The diagnostic accuracy, precision, recall, sensitivity, specificity, F1 score, Matthews correlation coefficient, confusion matrix, and area under the receiver operating characteristics curve.

Results

The multi-label multi-disease detection ability of the screening system was evaluated in 3 stepwise levels and reached the average accuracy of 0.969 and 0.923 in binary image-level anomaly detection, 0.940 and 0.827 in the 4-class region-level anomaly detection, and 0.972 and 0.911 in the 13-class lesion-level anomaly detection in the internal and external test data sets, respectively, showing comparable performance with the ophthalmologists. Furthermore, the screening system presented the average accuracy of 0.950 and 0.852 in internal and external test data sets in images of phenotypes that were less trained or untrained.

Conclusions

Our screening system showed excellent multi-label and multi-disease detection ability and generalization ability in identifying ocular anterior segment disease, regardless of the limited phenotypes in the training data set. Thus, the screening system is anticipated to offer easily available primary medical information for patients and assist ophthalmologists in clinical practice.

Financial Disclosure(s)

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.