Predicting Choroidal Nevus Transformation to Melanoma Using Machine Learning

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning
• Oncology

Background:

• Choroidal nevi are common melanocytic tumors of the eye.
• Predicting transformation to melanoma is crucial for patient management.
• Multimodal imaging offers detailed characterization of ocular lesions.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting choroidal nevus to melanoma transformation.
• To assess model performance using multimodal imaging data.
• To identify key imaging features predictive of nevus transformation.

Main Methods:

• Retrospective multicenter study involving patients with choroidal nevus.
• Utilized multimodal imaging: fundus photography, autofluorescence, OCT, and ultrasonography.
• Developed and optimized XGBoost, LGBM, Random Forest, and Extra Tree models, validated externally.

Main Results:

• The Simple AI Nevus Transformation System (SAINTS) XGBoost model achieved high performance (pooled AUROC 0.864 in test, 0.931 in validation).
• Key predictive features included tumor thickness, basal diameter, shape, distance to optic nerve, and subretinal fluid.
• A model using nevi with ≥5 years follow-up showed improved AUPRC.

Conclusions:

• ML models can accurately and generally predict choroidal nevus to melanoma transformation.
• Multimodal imaging is valuable for risk stratification.
• SAINTS model provides a robust tool for clinical decision-making.