Deep learning algorithm shows accuracy in detecting glaucoma on fundus photographs
Automated deep learning analysis of fundus photographs showed high diagnostic accuracy in determining primary open-angle glaucoma, with increased ability to detect glaucoma earlier than human readers.
A deep learning (DL) algorithm was trained, validated and tested on the fundus stereophotographs of participants enrolled in the Ocular Hypertension Treatment Study (OHTS), a randomized clinical trial evaluating the safety and efficacy of IOP-lowering medications in preventing progression from ocular hypertension to primary open-angle glaucoma (POAG). Assessment of optic disc and visual field changes in the OHTS was performed by two reading centers and a masked committee of glaucoma specialists, “a demanding, laborious and complicated task,” according to the authors.
The OHTS data set consisted of fundus photographs from 1,636 participants, of which 1,147 were included in the training set, 167 in the validation set and 322 in the test set. The DL model detected conversion to POAG with high diagnostic accuracy, suggesting that artificial intelligence can offer a reliable tool to automate the determination of glaucoma for clinical trial management, simplifying the process of human interpretation and, possibly, making it more standardized, objective and accurate. Noticeably, DL analysis was associated with a higher false-positive rate in early photographs of eyes that later developed POAG as compared with eyes that did not develop POAG, suggesting that DL models may be able to detect glaucoma in some eyes earlier than human readers.
“These false-positives likely were true-positives detecting disease-related change on average more than 4 years earlier in eyes with ocular hypertension,” the authors wrote.
Integration of DL image analysis in clinical trials could improve the consistency and accuracy of endpoint assessment and significantly reduce the need for human resources and related costs.
“Moreover, given the performance of the DL analysis in comparison with expert human observation, this approach may be promising to provide decision support in clinical settings,” the authors wrote.