Toward Accurate Cardiac MRI Segmentation With Variational Autoencoder-Based Unsupervised Domain Adaptation.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Cardiovascular Imaging

Background:

• Accurate myocardial segmentation in Late Gadolinium Enhancement (LGE) cardiac magnetic resonance (CMR) images is vital for diagnosing and treating myocardial infarction (MI).
• Segmentation annotations for LGE-CMR images are scarce, hindering supervised learning approaches.
• Domain shift between modalities (e.g., LGE vs. balanced-Steady State Free Precession (bSSFP)) significantly degrades model performance when transferring knowledge from annotated to unannotated datasets.

Purpose of the Study:

• To develop an unsupervised domain adaptation framework for myocardial structure segmentation in LGE-CMR images.
• To address the domain shift problem by aligning feature distributions between different CMR image modalities.
• To improve the accuracy of myocardial segmentation without relying on LGE-specific annotations.

Main Methods:

• Proposed a Variational autoencoder based feature Alignment Module Combining Explicit and Implicit features (VAMCEI) for end-to-end feature alignment.
• Utilized Kullback-Leibler (KL) divergence to measure global distribution distance between domains.
• Implemented a prototype contrastive loss to align same-category prototypes and separate different-category prototypes across domains.
• Incorporated a domain discriminator to refine feature distributions for improved segmentation.
• Combined VAMCEI with CycleGAN to create a multi-stage unsupervised domain adaptation (UDA) framework.

Main Results:

• The proposed VAMCEI module effectively aligns feature distributions between LGE and bSSFP CMR images.
• The multi-stage UDA framework significantly mitigates performance degradation caused by domain shift.
• Experimental results on MSCMRSeg 2019, MyoPS 2020, and MM-WHS 2017 datasets demonstrate superior performance compared to state-of-the-art methods.

Conclusions:

• The developed unsupervised domain adaptation framework offers a robust solution for myocardial segmentation in LGE-CMR images.
• The VAMCEI module and its integration into a multi-stage framework effectively handle domain shift challenges.
• This approach advances the field of automated cardiac image analysis, particularly in scenarios with limited annotated data.