MedZeroSeg: Zero-shot medical image segmentation via vision foundation models

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Computer Vision

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning.
• Current methods often require extensive labeled datasets, limiting their applicability.
• Leveraging large vision foundation models offers a potential solution for data-efficient segmentation.

Purpose of the Study:

• To introduce MedZeroSeg, a novel framework for zero-shot medical image segmentation.
• To reduce the dependency on large annotated datasets in medical image analysis.
• To enhance segmentation robustness and adaptability across diverse medical imaging modalities.

Main Methods:

• Utilized vision foundation models: Contrastive Language-Image Pre-training (CLIP) and Segment Anything Model (SAM).
• Developed a Dual-Path Feature Extraction Module for integrated local and global perception.
• Introduced a Context-Enhanced Hard-Negative Contrast Loss for improved contrastive learning.

Main Results:

• Achieved superior zero-shot and weakly supervised segmentation performance on cardiac MRI, abdominal CT, and chest X-ray datasets.
• Demonstrated strong generalization capabilities across different medical imaging modalities without task-specific fine-tuning.
• Showcased minimal data dependency, significantly reducing the need for annotated data.

Conclusions:

• MedZeroSeg offers a significant advancement in medical image segmentation by leveraging foundation models.
• The framework demonstrates high adaptability and minimal data requirements, addressing key challenges in the field.
• Opens promising avenues for applying advanced AI models in healthcare applications.