Unsupervised Cross-Modality MR Image Segmentation via Prompt-Driven Foundation Model.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Pixel-level medical image annotation is costly and time-consuming, particularly for multi-modal data like MRI.
• Unsupervised domain adaptation methods for cross-modality segmentation often struggle with source-target domain discrepancies.

Purpose of the Study:

• To develop a robust cross-modality segmentation scheme using foundation models that bypasses domain discrepancies.
• To enable accurate segmentation in target imaging modalities using annotations from a single source modality without labels or registration.

Main Methods:

• Utilized a Segment Anything Model (SAM)-based approach, employing segmentation results from one modality as pseudo-labels and prompts for another.
• Introduced consistency-based prompt tuning and hybrid representation learning to handle unregistered data and noisy labels.
• Leveraged spatial consistency across multiple modalities to mitigate domain shift issues.

Main Results:

• Demonstrated significant performance improvements in cross-modality segmentation tasks.
• Validated the method on internal and public datasets for liver lesion and liver segmentation.
• Achieved state-of-the-art results compared to existing approaches.

Conclusions:

• The proposed foundation model-based scheme offers an efficient and effective solution for cross-modality medical image segmentation.
• This approach reduces the reliance on extensive expert annotations and complex registration processes.
• The method shows great potential for improving segmentation accuracy and applicability across diverse medical imaging scenarios.