U-Net Optimization for Hyperreflective Foci Segmentation in Retinal OCT.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Hyperreflective foci (HRF) are key optical coherence tomography (OCT) biomarkers linked to retinal disease progression.
• Accurate HRF identification is challenging due to small size, class imbalance, and imaging artifacts.
• Optimal U-Net model configurations for HRF segmentation require further investigation.

Purpose of the Study:

• To evaluate optimal U-Net-based model settings for hyperreflective foci (HRF) segmentation.
• To compare standard U-Net and attention U-Net performance under various preprocessing conditions.

Main Methods:

• Evaluated standard U-Net and attention U-Net models on 435 OCT B-scans with annotated HRF masks.
• Applied preprocessing techniques including contrast-limited adaptive histogram equalization (CLAHE).
• Utilized focal Tversky loss for standard U-Net and soft dice loss for attention U-Net.

Main Results:

• Standard U-Net achieved a dice score of 0.5207, AUC of 0.8411, and recall of 0.6439 on raw OCT images.
• The standard U-Net with CLAHE and focal Tversky loss improved recall by 19.4% compared to attention U-Net.
• This configuration demonstrated increased sensitivity and a 23% relative reduction in false negatives.

Conclusions:

• The standard U-Net model combined with CLAHE and focal Tversky loss is the optimal configuration for HRF segmentation.
• This approach effectively handles class imbalance and enhances sensitivity for HRF detection.
• The standard U-Net provides a robust framework for automated HRF segmentation, suitable for further clinical validation.