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
AI-based Optic Disc and Cup Segmentation Can Provide Consistent Cup-Disc Ratio (CDR) Measurements and Precise Progression Tracking
Author Affiliations & Notes
  • Scott Kinder
    Ophthalmology, University of Colorado Anschutz Medical Campus School of Medicine, Aurora, Colorado, United States
  • Steve McNamara
    Ophthalmology, University of Colorado Anschutz Medical Campus School of Medicine, Aurora, Colorado, United States
  • Christopher Clark
    Ophthalmology, University of Colorado Anschutz Medical Campus School of Medicine, Aurora, Colorado, United States
  • Yoga Advaith Veturi
    Ophthalmology, University of Colorado Anschutz Medical Campus School of Medicine, Aurora, Colorado, United States
  • Galia Dietz
    Ophthalmology, University of Colorado Anschutz Medical Campus School of Medicine, Aurora, Colorado, United States
  • Talisa E De Carlo
    Ophthalmology, University of Colorado Anschutz Medical Campus School of Medicine, Aurora, Colorado, United States
  • Malik Kahook
    Ophthalmology, University of Colorado Anschutz Medical Campus School of Medicine, Aurora, Colorado, United States
  • Praveer Singh
    Ophthalmology, University of Colorado Anschutz Medical Campus School of Medicine, Aurora, Colorado, United States
  • Jayashree Kalpathy-Cramer
    Ophthalmology, University of Colorado Anschutz Medical Campus School of Medicine, Aurora, Colorado, United States
  • Footnotes
    Commercial Relationships   Scott Kinder None; Steve McNamara None; Christopher Clark None; Yoga Advaith Veturi None; Galia Dietz None; Talisa De Carlo None; Malik Kahook New World Medical, Spyglass Pharma, Code C (Consultant/Contractor), Spyglass Pharma, Code O (Owner), New World Medical, Alcon, SpyGlass Pharma, Code P (Patent); Praveer Singh None; Jayashree Kalpathy-Cramer Siloam, Code C (Consultant/Contractor), Genentech, Code F (Financial Support), Boston AI Lab, Code R (Recipient)
  • Footnotes
    Support  Unrestricted Research grant to the Department of Ophthalmology from RPB
Investigative Ophthalmology & Visual Science June 2024, Vol.65, 6189. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Scott Kinder, Steve McNamara, Christopher Clark, Yoga Advaith Veturi, Galia Dietz, Talisa E De Carlo, Malik Kahook, Praveer Singh, Jayashree Kalpathy-Cramer; AI-based Optic Disc and Cup Segmentation Can Provide Consistent Cup-Disc Ratio (CDR) Measurements and Precise Progression Tracking. Invest. Ophthalmol. Vis. Sci. 2024;65(7):6189.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : We evaluated the performance of Artificial Intelligence (AI)-predicted cup-to-disc ratio (CDR) on a large cohort of glaucoma patients, comparing inter-rater variability of AI-clinician vs clinicians alone, and tested whether an AI pipeline can provide more precise measurements to track CDR progression.

Methods : An AI-based optic disc + cup segmentation pipeline was trained using 1,919 fundus photos (1,434 full, 485 cropped fundus) from public datasets (Drishti-GS, RIGA, REFUGE-1, RIM-ONE DL) and preprocessed to match the field of view needed to train models in the inference pipeline. The inference pipeline handles full fundus and the left half of stereoscopic images by cropping (using YOLOv8), then performs segmentation of optic disc and cup (using MaskFormer w/ Swin backbone) to obtain the vertical CDR. The AI pipeline performance was evaluated on an independent, retrospective clinical dataset acquired over a 10-year period at a single institution. The clinical dataset consisted of 5,752 full fundus and 22,799 stereoscopic images (4,015 patients diagnosed with glaucoma), filtered down to 19,696 images by keeping only those above a YOLOv8 detection threshold. Each image had a CDR measurement gathered from electronic health record (EHR) on the same day of image capture. Longitudinal results were calculated comparing AI-predicted delta change vs EHR delta change.

Results : The Pearson correlation coefficient between AI-predicted and EHR-based CDR across all eyes is 0.758. The AI-prediction is within 0.2, 0.15, and 0.1 of the EHR-based CDR 88.3%, 77.8%, and 60.6% of the time, respectively. The mean, standard deviation and 95% confidence interval for fixed-time point differences between AI-predicted and EHR-based CDRs is 0.022±0.127 [95% CI: 0.020, 0.024], and for longitudinal differences between AI-predicted and EHR-based CDR delta change is -0.001±0.074 [95% CI: -0.002, 0.000]. Common failure modes of the AI pipeline included images with severe peripapillary atrophy and poor image quality.

Conclusions : AI-based CDR assessments can provide estimates with low inter-rater variability compared to humans even when trained on diverse, out-of-domain datasets. Longitudinal progression tracking from AI models allows for more precise measurement of CDR compared to the common notation of 0.05 or 0.1 intervals, potentially allowing for subtle changes to be identified sooner.

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.

×