Transformer based Generative Adversarial Network for Liver Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automated liver segmentation from CT and MRI scans is crucial for clinical applications.
• Convolutional Neural Networks (CNNs) have been standard, but Transformers offer advantages in modeling long-range dependencies via attention mechanisms.

Purpose of the Study:

• To develop a novel hybrid segmentation approach combining Transformers and Generative Adversarial Networks (GANs).
• To leverage Transformer's self-attention for global information modeling and GANs for mask credibility assessment.
• To enhance the accuracy and reliability of automated liver segmentation.

Main Methods:

• A hybrid model integrating Transformer architecture with a GAN framework.
• Utilizing the Transformer's self-attention mechanism for high-dimensional feature aggregation.
• Employing a GAN discriminator to validate generated segmentation masks against expert annotations.

Main Results:

• The proposed hybrid model achieved a Dice coefficient of 0.9433, recall of 0.9515, and precision of 0.9376.
• The model demonstrated superior performance compared to other Transformer-based segmentation approaches.
• The method effectively extracts high-dimensional topological information for reliable biomedical image segmentation.

Conclusions:

• The hybrid Transformer-GAN approach offers a significant advancement in automated liver segmentation.
• This method provides more reliable segmentation results by incorporating adversarial training for mask validation.
• The findings support the use of this advanced AI technique for improved surgical and therapy planning.