To develop an artificial intelligence construct that can automatically identify glaucoma from fundus photographs

We included fundus photographs from 267 normal eyes and 160 eyes with glaucoma in the study. We developed an artificial intelligence (AI) construct to distinguish normal eyes from eyes with glaucoma using fundus photographs. After training and testing the AI model, all 427 fundus photographs were used as input to the proposed model (based on a deep pre-trained architecture) consisting of region of interest on fundus photographs to identify clinically relevant glaucoma signatures. We then assessed different regions of the fundus photographs to identify potentially novel signatures of glaucoma. Those regions were more important in distinguishing fundus photographs of normal eyes from eyes with glaucoma. The clinical diagnostic ability of the AI model was evaluated by machine learning accuracy metrics and the identified glaucoma imaging signatures were validated by a glaucoma specialist to assure clinical relevance.

The accuracy of the method in discriminating normal eyes from eyes with glaucoma was 90%. Among fundus photographs that had been classified to glaucoma group, we observed that the AI model had identified significant features mostly in the superior/inferior peripapillary regions and within the optic nerve head.

We developed an AI model that was able to detect glaucoma from widely-used fundus photographs with high accuracy. This AI model also identified clinically relevant glaucoma signatures from fundus photographs. This approach could be useful in glaucoma research, clinical practice, and primary care settings as an assistive tool for screening glaucoma without the need for glaucoma clinicians. An independent dataset with larger number of fundus photographs is required to validate our findings.