[Application of Novel Down-sampling Method in Retinal Vessel Segmentation].

Area of Science:

• Medical image analysis
• Deep learning for computer vision
• Ophthalmology imaging

Context:

• Accurate retinal blood vessel segmentation is crucial for diagnosing and monitoring eye diseases.
• Current deep learning models, like U-Net, often use max pooling, which can lead to information loss during down-sampling.
• There is a need for improved down-sampling techniques to preserve finer details in medical images.

Purpose:

• To propose and evaluate a novel, lightweight, and generalizable down-sampling module called Pixel Fusion-pooling (PF-pooling).
• To enhance the performance of convolutional neural networks for medical image segmentation, specifically for retinal blood vessels.
• To demonstrate the effectiveness of PF-pooling in fusing adjacent pixel information and reducing information loss.

Summary:

• A new down-sampling method, Pixel Fusion-pooling (PF-pooling), is introduced to improve medical image segmentation.
• PF-pooling effectively fuses adjacent pixel information, overcoming limitations of traditional max pooling.
• Experimental results on DRIVE and STARE datasets show significant improvements in U-Net model performance, including a 1.98% increase in F1-score on the STARE dataset.

Impact:

• The proposed PF-pooling module offers a lightweight and universally applicable solution for enhancing deep learning models in medical image segmentation.
• Demonstrated performance improvements in U-Net, Dense-UNet, and Res-UNet models highlight the generalizability and effectiveness of PF-pooling.
• This advancement can lead to more accurate and reliable diagnosis and monitoring of eye conditions through improved retinal image analysis.