MCSE-U-Net: multi-convolution blocks and squeeze and excitation blocks for vessel segmentation

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate segmentation of retinal blood vessels from fundus images is critical for diagnosing various retinal vascular diseases.
• The complex vascular structure and ambiguous clinical criteria present significant challenges in precise blood vessel segmentation.

Purpose of the Study:

• To develop an advanced deep learning model for enhanced retinal vessel segmentation.
• To improve the accuracy and efficiency of automated retinal vessel segmentation in medical imaging.

Main Methods:

• Developed the MCSE-U-net model, integrating multi-convolution (MC) and squeeze-and-excitation (SE) blocks into a U-shaped network architecture.
• Employed YCbCr color space conversion for improved image visibility and utilized MC and SE blocks to enhance segmentation of both major and fine retinal vessels.

Main Results:

• The MCSE-U-net model achieved high performance on the DRIVE dataset, with a Dice coefficient of 0.8430, sensitivity of 0.8752, specificity of 0.9902, accuracy of 0.9725, and mIoU of 0.8473.
• Demonstrated significant improvements over the original U-net, with increases of 3.08% in Dice coefficient, 6.22% in sensitivity, 0.62% in specificity, 0.61% in accuracy, and 3.01% in mIoU.
• The model showed particular effectiveness in segmenting peripheral vascular anatomy.

Conclusions:

• The MCSE-U-net model offers superior performance for retinal vessel segmentation compared to existing technologies.
• The proposed architecture effectively addresses the challenges in segmenting complex retinal vascular structures.