Cross-Modality Medical Image Segmentation via Enhanced Feature Alignment and Cross Pseudo Supervision Learning

Area of Science:

• Medical Imaging Analysis
• Computer Vision
• Machine Learning

Background:

• Traditional medical image segmentation models struggle with domain shift due to diverse imaging modalities.
• Unsupervised Domain Adaptation (UDA) methods, often using Generative Adversarial Networks (GANs), aim to address cross-modality analysis but face limitations with significant feature gaps.
• Existing UDA approaches assume feature alignment, which is often not true for modalities like MRI and CT, leading to training instability.

Purpose of the Study:

• To develop a novel unsupervised domain adaptation approach for medical image segmentation that effectively bridges domain discrepancies between modalities.
• To improve the stability and accuracy of cross-modality medical image segmentation.
• To enhance the learning efficiency of segmentation networks when dealing with heterogeneous medical image data.

Main Methods:

• Introduction of a novel approach with two key sub-networks: a cross-modality feature alignment sub-network and a cross pseudo supervised dual-stream segmentation sub-network.
• The feature alignment sub-network employs bidirectional alignment and a self-attention module for learning structurally consistent features.
• The segmentation sub-network utilizes an enhanced cross-pseudo-supervised loss, assessing inter-domain pseudo-distances to improve pseudo-label quality.

Main Results:

• Demonstrated notable advancements in segmentation precision across target domains for both abdomen and brain imaging tasks.
• Successfully bridged domain discrepancies, leading to more effective cross-modality image segmentation.
• Ensured a more stable training environment compared to traditional UDA methods.

Conclusions:

• The proposed novel approach effectively addresses the domain shift problem in medical image segmentation across different modalities.
• The combination of feature alignment and cross pseudo-supervised learning significantly improves segmentation performance and stability.
• This method offers a promising solution for robust cross-modality medical image analysis.