Investigative Ophthalmology & Visual Science Cover Image for Volume 65, Issue 7
June 2024
Volume 65, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2024
Glaucoma Suspects Detection by Combined Machine Learning-Based Risk Score Generation and Feature Optimization
Author Affiliations & Notes
  • Alauddin Bhuiyan
    iHealthscreen Inc., New York, United States
  • James C Tsai
    New York Eye and Ear Infirmary of Mount Sinai, New York, New York, United States
  • Tak Yee Tania Tai
    New York Eye and Ear Infirmary of Mount Sinai, New York, New York, United States
  • Roland Theodore Smith
    New York Eye and Ear Infirmary of Mount Sinai, New York, New York, United States
  • Arun Govindaiah
    iHealthscreen Inc., New York, United States
  • Footnotes
    Commercial Relationships   Alauddin Bhuiyan None; James Tsai None; Tak Tai None; Roland Smith None; Arun Govindaiah None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2024, Vol.65, 1620. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Alauddin Bhuiyan, James C Tsai, Tak Yee Tania Tai, Roland Theodore Smith, Arun Govindaiah; Glaucoma Suspects Detection by Combined Machine Learning-Based Risk Score Generation and Feature Optimization. Invest. Ophthalmol. Vis. Sci. 2024;65(7):1620.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : Glaucoma is one of the leading causes of blindness globally which often progresses unnoticed until significant vision loss occurs. Early detection is crucial for effective management. This study introduces a novel machine-learning algorithm for early detection of glaucoma suspects with the utilization of the retinal cup-disc ratio (CDR), disc hemorrhages, and peripapillary atrophy.

Methods : Our approach utilizes a divide-and-conquer strategy, breaking down the screening task into three subproblems: CDR estimation, disc hemorrhage detection, and peripapillary atrophy identification. Each sub-model is enhanced through domain-specific pre-training on a massive multi-label fundus image dataset, including images from diverse sources. We employed EfficientNet B4 architecture pre-trained on ImageNet for disc hemorrhage detection, with automated AI cropping to focus on the optic disc area. Data augmentation techniques were applied to counter the small dataset size. For external validation, we used the UK biobank dataset.

Results : The disc hemorrhage detection achieved an accuracy of 93.13% (95% CI: 91.35% to 94.64%), sensitivity of 71.53% (95% CI: 63.42% to 78.73%), and specificity of 96.87% (95% CI: 95.45% to 97.95%). The peripapillary atrophy model exhibited a sensitivity of 93.18% (95% CI: 81.34% to 98.57%) and specificity of 97.67% (95% CI: 87.71% to 99.94%). For CDR classification, the model categorized images into three classes with a weighted kappa of 0.785. The final logistic model tree classifier combines these individual models' outputs, providing a comprehensive risk score for glaucoma suspects. The combined model detected glaucoma with an accuracy of 94.3% (95% CI: 89.05% to 97.50%), a sensitivity of 95.0% (95% CI: 83.08% to 99.39%), and a specificity of 94.0% (95% CI: 87.40% to 97.77%)

Conclusions : This study presents an innovative, data-driven approach for glaucoma suspect screening by leveraging advanced machine learning techniques and color fundus imaging. The high accuracy and robustness of the individual models demonstrate the potential of this system in early glaucoma detection, which could significantly impact public health by enabling timely interventions and prevention of glaucoma.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×