Differentiating Choroidal Melanomas and Nevi Using a Self-Supervised Deep Learning Model Applied to Clinical...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Posterior uveal melanoma (UM) is a rare but serious ocular malignancy.
• Accurate differentiation between UM and benign nevi is crucial for appropriate patient management.
• Existing diagnostic methods can be limited, necessitating advanced analytical tools.

Purpose of the Study:

• To evaluate the performance of RETFound, a self-supervised deep learning model, in distinguishing between uveal melanoma and nevi.
• To assess the model's utility in classifying healthy eyes alongside UM and nevi.
• To validate the model's accuracy on a large, single-center dataset.

Main Methods:

• A case-control study utilizing ultrawidefield fundoscopy images (color and autofluorescence) from 4255 patients.
• Analysis of 18,510 UM, 8,671 nevi, and 1,192 healthy eye images after quality exclusion.
• Fine-tuning the RETFound deep learning model for binary (UM vs. nevi) and tertiary (UM vs. nevi vs. healthy) classification tasks.

Main Results:

• The model achieved an AUROC of 0.90 and accuracy of 0.83 for binary classification (UM vs. nevi).
• For tertiary classification, the model demonstrated a mean accuracy of 0.82 and an AUROC of 0.92.
• These results indicate strong performance in differentiating ocular pathologies.

Conclusions:

• Self-supervised deep learning models like RETFound are feasible for accurate differentiation of UM and nevi.
• The model shows high accuracy in a large, imbalanced dataset from a single clinical center.
• Further validation on external cohorts is planned to assess broader clinical applicability.