Generative Adversarial Training with Dual-Attention for Vascular Segmentation and Topological Analysis

Area of Science:

• Medical Imaging
• Computational Biology
• Oncology

Background:

• Vascular topology is critical for developing effective liver cancer chemotherapy models.
• Accurate segmentation of liver vessels is essential for topological analysis.
• Existing methods may lack the precision required for complex vascular structures.

Purpose of the Study:

• To propose and validate a novel pipeline for segmenting liver vessels and extracting their skeletons.
• To enable detailed topological analysis for improved liver cancer chemotherapy models.
• To enhance the accuracy of vessel segmentation and skeletonization in liver imaging.

Main Methods:

• A dual-attention U-Net model was employed, trained using a generative adversarial network (GAN) framework.
• The pipeline segments liver vessels and extracts their skeletons.
• Extracted skeletons are classified by length and bifurcation order for topological analysis.

Main Results:

• The proposed pipeline demonstrated consistent superiority in segmentation accuracy compared to existing methods.
• Topology correctness was significantly improved, validating the pipeline's effectiveness.
• Experiments were conducted on 40 samples with meticulously annotated ground truth labels.

Conclusions:

• The developed pipeline offers a robust and accurate solution for liver vessel segmentation and topological analysis.
• This method provides a valuable tool for building more precise liver cancer chemotherapy models.
• The findings highlight the potential of deep learning, specifically GAN-trained U-Nets, in medical image analysis for oncology research.