A Liver Segmentation Method Based on the Fusion of VNet and WGAN

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate liver segmentation is crucial for diagnosing and treating liver diseases.
• Existing 2D Convolutional Neural Network (CNN) methods struggle with interlayer information, limiting segmentation accuracy.
• Integrating high-level and low-level features effectively in 2D networks remains a challenge.

Purpose of the Study:

• To develop an improved liver segmentation method using a 2.5D CNN architecture.
• To enhance segmentation accuracy by effectively utilizing cross-sectional and interlayer information.
• To optimize feature fusion and address limitations of existing loss functions.

Main Methods:

• Designed and implemented VNet_WGAN, a 2.5D CNN incorporating three adjacent liver image slices as input.
• Employed serial convolution kernels to integrate cross-sectional and interlayer spatial information.
• Introduced a chain residual pooling module for multilevel feature fusion and optimized skip connections.
• Utilized a boundary loss function to improve marginal pixel accuracy, complementing Dice loss.

Main Results:

• Achieved Dice coefficients of 92% on the LiTS dataset and 90% on the CHAOS dataset.
• Demonstrated superior performance compared to existing 2D segmentation networks.
• Showcased reduced computational consumption while maintaining high segmentation accuracy.

Conclusions:

• The VNet_WGAN method significantly improves liver segmentation accuracy by effectively leveraging 2.5D information.
• The approach offers a practical and efficient solution for liver segmentation in clinical settings.
• Provides a valuable reference for clinicians and researchers in medical image analysis.