RA-SIFA: Unsupervised domain adaptation multi-modality cardiac segmentation network combining parallel attention...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Convolutional neural networks (CNNs) are effective for cardiac image segmentation.
• Diversity in medical imaging equipment causes domain shift, challenging segmentation accuracy.
• Domain shift hinders the performance of CNNs across different imaging modalities.

Purpose of the Study:

• To investigate and test an unsupervised domain adaptation network, RA-SIFA, for multi-modality cardiac image segmentation.
• To address the domain shift problem in cardiac image segmentation.
• To improve the accuracy and robustness of cardiac image segmentation across different imaging devices.

Main Methods:

• RA-SIFA integrates a parallel attention module (PAM) in the generator for image alignment.
• A residual attention unit (RAU) in the shared encoder enhances feature alignment.
• The network employs unsupervised domain adaptation (UDA) combining image and feature alignment.

Main Results:

• RA-SIFA improved Dice scores by 8.4% (CT) and 3.2% (MR) compared to SIFA.
• Average symmetric surface distance (ASD) was reduced by 3.4 mm (CT) and 0.8 mm (MR).
• The model demonstrated enhanced accuracy in whole-heart segmentation on the MM-WHS2017 dataset.

Conclusions:

• The proposed RA-SIFA network effectively mitigates domain shift in cardiac image segmentation.
• RA-SIFA significantly improves segmentation accuracy for both CT and MR cardiac images.
• The study highlights the potential of attention-based UDA for robust medical image analysis.