Liver tumor segmentation method combining multi-axis attention and conditional generative adversarial networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Manual segmentation of liver and tumors in CT images is inefficient and inaccurate.
• Deep learning offers automatic segmentation but faces challenges like class imbalance and poor feature fusion.
• Existing methods struggle with fuzzy boundaries, irregular shapes, and small lesions in liver tumor segmentation.

Purpose of the Study:

• To develop an advanced deep learning model for accurate and efficient automatic liver tumor segmentation.
• To address limitations of existing methods, including class imbalance and inadequate feature representation.
• To improve the perception of local details and global contexts in abdominal CT image segmentation.

Main Methods:

• Proposed a Multi-Axis Attention Conditional Generative Adversarial Network (MA-cGAN).
• Introduced a Multi-Axis attention mechanism (MA) for projecting 3D CT images and fusing 2D features from different axes.
• Integrated MA into a U-shaped network generator and combined it with a discriminator for enhanced segmentation stability and accuracy.

Main Results:

• MA-cGAN demonstrated superior performance on the LiTS dataset compared to state-of-the-art models.
• Achieved improvements in Dice coefficient, Hausdorff distance, and average surface distance metrics.
• Generated segmented liver and tumor models with clearer edges, fewer false positives, and higher fidelity to true labels.

Conclusions:

• MA-cGAN effectively addresses class imbalance and enhances feature fusion for liver tumor segmentation.
• The model improves segmentation accuracy and detail perception, crucial for medical adjuvant therapy.
• The proposed approach offers a significant advancement in automated medical image analysis for liver cancer detection and treatment planning.