Highly Accurate and Precise Automated Cup-to-Disc Ratio Quantification for Glaucoma Screening

• DOI: 10.1016/j.xops.2024.100540
• Publication Date: 2024-04-27T01:31:09Z
• Authors: Abadh K. Chaurasia; Connor J. Greatbatch; Xikun Han; Puya Gharahkhani; David A. Mackey; Stuart MacGregor; Jamie E. Craig; Alex W. Hewitt
• Keywords: 0301 basic medicine; 03 medical and health sciences; 550; 610; Original Article

ObjectiveAn enlarged cup-to-disc ratio (CDR) is a hallmark of glaucomatous optic neuropathy. Manual assessment the CDR may be less accurate and more time-consuming than automated methods. Here, we sought to develop validate deep learning–based algorithm automatically determine from fundus images.DesignAlgorithm development for estimating using data population-based observational study.ParticipantsA total 181 768 images United Kingdom Biobank (UKBB), Drishti_GS, EyePACS.MethodsFastAI PyTorch libraries were used train convolutional neural network–based model on UKBB. Models constructed image gradability (classification analysis) as well estimate (regression analysis). The best-performing was then validated use in glaucoma screening multiethnic dataset EyePACS Drishti_GS.Main Outcome MeasuresThe area under receiver operating characteristic curve coefficient determination.ResultsOur vgg19_batch normalization (bn) achieved an accuracy 97.13% validation set 16 045 images, with 99.26% precision 96.56%. Using regression analysis, our (trained vgg19_bn architecture) attained determination 0.8514 (95% confidence interval [CI]: 0.8459–0.8568), while mean squared error 0.0050 CI: 0.0048–0.0051) absolute 0.0551 0.0543–0.0559) 12 183 determining CDR. point converted into classification metrics tolerance 0.2 20 classes; 99.20%. (98 172 healthy, 3270 glaucoma) externally classification, accuracy, sensitivity, specificity 82.49%, 72.02%, 82.83%, respectively.ConclusionsOur models precise Although artificial intelligence–derived estimates achieve high threshold will vary depending other clinical parameters.Financial Disclosure(s)Proprietary or commercial disclosure found Footnotes Disclosures at end this article.