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
Machine Learning Automated Segmentation for Early Detection of Uveal Melanoma
Author Affiliations & Notes
  • Michael Heiferman
    Ophthalmology, University of Illinois Chicago, Chicago, Illinois, United States
  • Jiechao Ma
    Bioengineering, University of Illinois Chicago, Chicago, Illinois, United States
  • Sabrina Iddir
    Ophthalmology, University of Illinois Chicago, Chicago, Illinois, United States
  • Sanjay Ganesh
    Ophthalmology, University of Illinois Chicago, Chicago, Illinois, United States
  • Darvin Yi
    Ophthalmology, University of Illinois Chicago, Chicago, Illinois, United States
  • Footnotes
    Commercial Relationships   Michael Heiferman None; Jiechao Ma None; Sabrina Iddir None; Sanjay Ganesh None; Darvin Yi None
  • Footnotes
    Support  Illinois Society for the Prevention of Blindness Research Grant, Research to Prevent Blindness Departmental Grant
Investigative Ophthalmology & Visual Science June 2024, Vol.65, 4278. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Michael Heiferman, Jiechao Ma, Sabrina Iddir, Sanjay Ganesh, Darvin Yi; Machine Learning Automated Segmentation for Early Detection of Uveal Melanoma. Invest. Ophthalmol. Vis. Sci. 2024;65(7):4278.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Uveal melanoma (UM) is the most common primary intraocular malignancy in adults and has a poor prognosis. Current screening and triaging methods for melanocytic choroidal tumors face inherent limitations, particularly in regions with limited access to ocular oncologists. One solution may lie in the implementation of enhanced screening techniques that are accessible to all providers with varying levels of expertise. Currently, there are no automated clinical tools for UM identification using ophthalmic imaging. Machine learning (ML) provides a promising method for automating tumor identification. In this study, we develop and evaluate an ML model designed for lesion segmentation in ultra-widefield fundus photography.

Methods : A retrospective chart review was conducted of patients diagnosed clinically with UM, choroidal nevi, or congenital hypertrophy of the retinal pigmented epithelium (CHRPE) at a single tertiary academic medical center (University of Illinois Chicago). Patients were included who had a single ultra-widefield fundus photo of adequate quality to visualize the lesion of interest as confirmed by a single ocular oncologist (MJH). All images were used to develop and train an ML algorithm for lesion segmentation.

Results : A total of 295 patients with choroidal nevi, 166 patients with UM, and 105 patients with CHRPE met the inclusion criteria. The Dice coefficient is a calculated measure of similarity between segmented pixels and the ground truth. Values closer to 1 indicate higher similarity. Of the images with lesions that were successfully detected, the ML segmentation yielded Dice coefficients of 0.87, 0.77, and 0.80 for UM, choroidal nevi, and CHRPE respectively. The multiclass Dice coefficients for UM, choroidal nevi, and CHRPE were 0.87, 0.78, and 0.81 respectively. Sensitivity/specificity for UM, choroidal nevi, and CHRPE were 82/100, 76/100, and 78/100 respectively.

Conclusions : Our study demonstrates that the ML algorithm’s notable performance indicates its potential clinical utility in screening choroidal tumors. As such, automated lesion segmentation provides a potential screening method that may be accessible to providers globally, irrespective of the availability of ocular oncologists. More methods of evaluation are necessary to enhance the model for lesion classification and improve diagnostic accuracy.

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.

×