Multiscale U-Net with Spatial Positional Attention for Retinal Vessel Segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate retinal vessel segmentation is vital for diagnosing eye conditions.
• Challenges include scale variation and low contrast, hindering precise boundary identification.
• Deep learning models excel at capturing distinguishing features for retinal vessels.

Purpose of the Study:

• To propose an improved U-Net algorithm for accurate retinal vessel segmentation.
• To enhance the identification of vessel boundaries and improve blood vessel-background differentiation.
• To achieve superior performance compared to existing retinal vessel segmentation methods.

Main Methods:

• Replaced traditional CNNs with global convolutional networks and boundary refinement in the encoder.
• Incorporated improved position and channel attention modules in skip connections.
• Utilized multiscale input, dense feature pyramid cascades, and convolutional long/short memory networks with dilated convolutions in the decoder.

Main Results:

• Achieved high accuracy rates of 96.99% on the DRIVE dataset and 97.51% on the CHASE_DB1 dataset.
• Demonstrated superior average performance over existing retinal vessel segmentation algorithms.
• The enhanced U-Net model effectively addresses challenges in scale variation and low contrast.

Conclusions:

• The proposed improved U-Net algorithm significantly enhances retinal vessel segmentation accuracy.
• The modifications effectively improve boundary detection and vessel-background separation.
• This advanced deep learning approach offers a promising tool for early eye disease detection and diagnosis.