StAC-DA: Structure aware cross-modality domain adaptation framework with image and feature-level adaptation for...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel at medical image segmentation but struggle with differing data distributions across modalities.
• This limitation hinders the integration of diverse imaging data, despite its clinical benefits.

Purpose of the Study:

• To present an unsupervised framework, Structure Aware Cross-modality Domain Adaptation (StAC-DA), for robust medical image segmentation across different imaging modalities.
• To address the performance degradation of CNNs when source and target datasets have different probability distributions.

Main Methods:

• StAC-DA employs a two-step approach: image-level translation using a CycleGAN-based model with structure preservation, followed by feature-level alignment.
• A U-Net with deep supervision is trained adversarially using transformed source and target domain images for segmentation.

Main Results:

• The framework was evaluated on cardiac substructure segmentation, demonstrating superior performance compared to existing unsupervised domain adaptation methods.
• StAC-DA achieved top rankings in segmenting the ascending aorta for both Magnetic Resonance Imaging (MRI) to Computed Tomography (CT) and CT to MRI adaptations.

Conclusions:

• StAC-DA effectively overcomes challenges posed by differing data distributions in medical imaging datasets.
• The framework shows significant potential for enhancing the accuracy of medical image segmentation across various imaging modalities.