A Retinal Vessel Segmentation Method Based on the Sharpness-Aware Minimization Model.

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Accurate retinal vessel segmentation is vital for detecting eye conditions like diabetic retinopathy and glaucoma.
• Existing models may struggle with generalization, leading to suboptimal diagnostic performance.
• Novel optimization techniques are needed to improve the robustness of segmentation models.

Purpose of the Study:

• To evaluate the impact of Sharpness-Aware Minimization (SAM) on the generalization performance of the RF-UNet model for retinal vessel segmentation.
• To quantify the improvements in accuracy, loss reduction, and other key metrics using SAM.

Main Methods:

• Experiments were conducted using the Digital Retinal Images for Vessel Extraction (DRIVE) dataset, a standard benchmark.
• The RF-UNet model was trained with and without SAM to compare performance.
• Key performance indicators including training/validation loss, accuracy, sensitivity, specificity, AUC, and F1 score were analyzed.

Main Results:

• SAM-trained RF-UNet demonstrated substantially lower training loss (0.094225 vs. 0.45709) and validation loss (0.08053 vs. 0.40266) compared to the non-SAM model.
• Training accuracy increased from 0.90169 to 0.96225, and validation accuracy improved from 0.93999 to 0.96821 with SAM.
• Significant improvements were observed in sensitivity, specificity, AUC, and F1 score, indicating enhanced generalization.

Conclusions:

• Sharpness-Aware Minimization effectively reduces overfitting and enhances the generalization capabilities of RF-UNet for retinal vessel segmentation.
• The findings support the principle that SAM promotes learning flatter minima, leading to more robust models.
• This approach holds promise for improving other medical imaging tasks and warrants further investigation on diverse datasets and in clinical settings.