Enhancing Choroidal Nevus Position Identification through CNN-Based Segmentation of Eye Fundus Images.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diagnosing choroidal nevus from color fundus images presents challenges for clinicians lacking specialized experience.
• Machine learning (ML) offers a promising solution for accurate and efficient analysis of ocular abnormalities.

Purpose of the Study:

• To develop and compare convolutional neural network (CNN) segmentation models for identifying key areas in fundus images.
• To enhance the accuracy of choroidal nevus detection through improved image analysis.

Main Methods:

• Utilized fundus images from the Alberta Ocular Brachytherapy Program, including healthy and nevus-affected eyes.
• Developed and compared four CNN models: U-net, Residual U-net, Attention U-net, and an Ensemble model.
• Employed ground truth masks provided by an ocular oncologist for model training and external validation by experts.

Main Results:

• The CNN models achieved varying levels of segmentation accuracy.
• The Ensemble model demonstrated the highest performance with a Dice Coefficient score of 87.7%.
• Other models achieved scores of 85.02% (U-net), 85.66% (Residual U-net), and 86.89% (Attention U-net).

Conclusions:

• CNN-based segmentation models are effective for analyzing choroidal nevi in fundus images.
• The developed ensemble model shows significant potential for a clinical decision support system.
• Further validation and integration into clinical workflows can improve diagnostic accuracy for choroidal nevus.