Revolutionizing Glaucoma Screening: An Innovative Low-Cost Protocol Empowered by Automated Machine Learning

Theme: Glaucoma
What: Glaucoma
Part of: Glaucoma IV: New Tech / Glaucome IV: Nouvelles technologies
When: 6/1/2024, 11:15 AM - 12:45 PM
Where: Room | Salle 801

Abstract

Purpose: There is an undeniable need for a low-cost glaucoma screening strategy. Artificial intelligence (AI) is revolutionizing large-scale screening by standardizing the process, enhancing its accuracy, and reducing costs. However, employing AI models on cloud systems can be expensive. Automated Machine Learning (AutoML) enables clinicians to construct their own deep learning (DL) models. AutoML Vertex Edge offers offline AI models, enabling screening without incurring model costs and the necessity of an internet connection, which is particularly advantageous for resource-limited regions. This study assesses the performance of low-cost AutoML models (Cloud and Edge) in glaucoma screening using fundus images and compares it to expert-designed DL models.

Study Design: Artificial intelligence diagnostic algorithm design and validation.

Methods: Ophthalmology trainees with no programming experience constructed AutoML model design using a dataset comprised of 101,442 labeled fundus images. We designed two binary models, a Cloud model and an Edge model, to distinguish between normal and glaucomatous eyes. Subsequently, we conducted external validation on two separate datasets.

Results: The AutoML models demonstrated high diagnostic capabilities, comparable or superior to bespoke models. The binary Cloud model had an area under the precision-recall curve (AuPRC) of 0.97, sensitivity of 88%, specificity of 92% and accuracy of 91% (compared to a sensitivity of 85% and specificity of 95% for the best expert models). The binary Edge model had an AuPRC of 0.97, sensitivity of 85%, specificity of 94% and accuracy of 91%. For external validation, our Cloud model showed a sensitivity, specificity, positive predictive value, and accuracy ranging from 75 to 83%, 97 to 99%, 73 to 94%, and 95 to 98% with the REFUGE database, and 92 to 100%, 94 to 96%, 94 to 96%, and 94 to 97% with the GAMMA database.

Conclusions: AutoML models created by ophthalmologists without programming experience were comparable or better than expert models. This innovative approach demonstrates how AutoML has the potential to revolutionize global glaucoma screening, facilitating accessible, precise, and low-cost screening protocols.

Presenter(s)

Presenting Author: Daniel Milad

Additional Author(s):

Fares Antaki, Department of Ophthalmology, Centre Hospitalier de l'Université de Montréal (CHUM), Montreal, Quebec, Canada

Andrew Farah, Department of Medicine, McGill University, Montreal, Canada

Jonathan El-Khoury, Department of Ophthalmology, University of Montreal, Montreal, Canada

Samir Touma, Department of Ophthalmology, University of Montreal, Montreal, Canada

Georges Durr, Department of Ophthalmology, Centre Hospitalier de l'Université de Montréal (CHUM), Montreal, Quebec, Canada

Taylor Nayman, Department of Ophthalmology, Hôpital Maisonneuve-Rosemont (CUO-HMR), Montreal, Quebec, Canada

Pearse Keane, Moorfields Eye Hospital, University of College London, London, United Kingdom

Renaud Duval, Department of Ophthalmology, Hôpital Maisonneuve-Rosemont (CUO-HMR), Montreal, Quebec, Canada

---