A coarse-to-fine deep learning framework for optic disc segmentation in fundus images

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computer Vision for Medical Diagnosis

Background:

• Accurate optic disc (OD) segmentation is crucial for diagnosing retinal diseases.
• Existing methods may lack the precision required for early disease detection.
• Deep learning offers potential for automated and accurate medical image analysis.

Purpose of the Study:

• To develop and evaluate a coarse-to-fine deep learning framework for precise OD segmentation.
• To enhance the accuracy of automated optic disc identification in color fundus images.
• To improve early detection and quantitative diagnosis of retinal conditions.

Main Methods:

• Proposed a coarse-to-fine deep learning framework utilizing a U-net convolutional neural network (CNN).
• Trained the network on both color fundus images and grayscale vessel density maps.
• Employed an overlap strategy to identify disc candidate regions for refined U-net segmentation.

Main Results:

• Achieved an average intersection over union (IoU) of 89.1% and a dice similarity coefficient (DSC) of 93.9%.
• Demonstrated superior performance compared to the sole U-net model (IoU 87.4%, DSC 92.5%).
• Validated performance on a dataset of 2,978 images across multiple public and collected datasets.

Conclusions:

• The developed coarse-to-fine deep learning framework provides reliable and high-performance automated OD segmentation.
• This approach shows significant potential for clinical applications in diagnosing optic nerve diseases.
• The method offers a robust solution for quantitative analysis in ophthalmology.