VCU-Net: a vascular convolutional network with feature splicing for cerebrovascular image segmentation

Area of Science:

• Biomedical image processing
• Medical imaging analysis
• Neuroscience imaging

Background:

• Cerebrovascular image segmentation is vital in biomedical image processing.
• Traditional convolutional kernels struggle with the complex morphology of cerebral blood vessels, leading to feature loss.

Purpose of the Study:

• To propose a novel vascular convolutional U-Net (VCU-Net) for improved cerebrovascular image segmentation.
• To address the limitations of traditional methods in perceiving elongated blood vessel structures.

Main Methods:

• Introduced a vascular convolution method to adaptively extract features of blood vessels with varying morphologies and orientations.
• Implemented a new feature splicing technique in the encoding stage to enrich feature information.

Main Results:

• The VCU-Net achieved Dice Similarity Coefficient (DSC) of 53.57% and Intersection over Union (IOU) of 69.74%.
• Demonstrated a performance improvement of 2.11% (DSC) and 2.01% (IOU) over the best-performing GVC-Net.
• Showcased superior segmentation of complex cerebrovascular structures, particularly elongated vessels, with enhanced integrity and continuity.

Conclusions:

• The VCU-Net effectively addresses the challenges in segmenting elongated cerebral blood vessels.
• The proposed vascular convolution and feature splicing methods significantly improve segmentation accuracy and structural preservation.