Attention-Guided Deep Neural Network With Multi-Scale Feature Fusion for Liver Vessel Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Manual liver vessel segmentation from CT images is laborious and time-consuming.
• Existing deep learning methods struggle with the complex variations and structures of liver vessels.
• The UNet architecture, while common, has limitations in feature utilization for liver vessel segmentation.

Purpose of the Study:

• To develop a novel deep neural network, LVSNet, for accurate liver vessel segmentation.
• To address challenges posed by intricate liver vessel anatomy and variations.
• To improve automated segmentation for clinical diagnosis and surgical planning.

Main Methods:

• Proposed LVSNet with an Attention-Guided Concatenation (AGC) module to select relevant contextual features.
• Introduced a multi-scale fusion block with hierarchical residual-like connections to link vessel fragments.
• Created a new dataset of 40 thin-slice CT volumes with annotated liver vessels.
• Developed an automatic stratification method to differentiate major and minor liver vessels.

Main Results:

• LVSNet demonstrated superior performance in liver vessel segmentation compared to previous methods.
• The AGC module effectively captured detailed and complementary information.
• The multi-scale fusion block successfully integrated local vessel fragments.
• Ablation studies confirmed the effectiveness of the proposed modules and dataset.

Conclusions:

• LVSNet offers a significant advancement in automated liver vessel segmentation.
• The proposed methods enhance the accuracy and detail of segmentation for clinical applications.
• The new dataset and stratification method facilitate research on minor liver vessels.