MSGAN: Multi-Stage Generative Adversarial Networks for Cross-Modality Domain Adaptation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep neural networks often perform poorly on target domain medical images due to domain shifts.
• Cross-modality medical image segmentation (e.g., CT to MRI) presents significant challenges due to differing image characteristics.

Purpose of the Study:

• To introduce an unsupervised domain adaptation approach, Multi-Stage GAN (MSGAN), for robust medical image segmentation.
• To address the domain shift problem in cross-modality medical imaging tasks, specifically for CT and MRI segmentation.

Main Methods:

• Developed a Multi-Stage GAN (MSGAN) employing a parallel multi-stage strategy to preserve information across resolutions.
• Utilized style layers to map learned style codes from Gaussian noise to input features for synthesizing diverse image styles.
• Implemented a strategy to transfer rough styles from low-resolution to detailed textures on high-resolution feature maps.

Main Results:

• Validated MSGAN on public datasets for cross-modality medical image segmentation (CT and MRI).
• Demonstrated the effectiveness of the proposed method in mitigating performance degradation caused by domain shifts.
• Achieved successful translation and segmentation of cross-modality medical images.

Conclusions:

• MSGAN offers a promising solution for unsupervised domain adaptation in medical image segmentation.
• The technique facilitates the translation of cross-modality images (MRI and CT) for improved segmentation.
• This approach can significantly reduce performance drops when applying deep learning models in cross-domain scenarios.