EICSeg: Universal Medical Image Segmentation via Explicit In-Context Learning.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning models for medical image segmentation face challenges in generalizing to new tasks, requiring extensive retraining.
• In-context learning (ICL) offers a promising alternative by using example prompts but often relies on implicit modeling.
• Existing ICL methods lack end-to-end integration and efficient adaptation to diverse medical imaging scenarios.

Purpose of the Study:

• To develop an end-to-end in-context learning framework for universal medical image segmentation.
• To improve the generalization capabilities of segmentation models to unseen tasks, anatomies, and modalities.
• To reduce the reliance on extensive human effort and computational resources for model adaptation.

Main Methods:

• Introduced Explicit In-Context Learning (E-ICL) by redefining ICL as an explicit retrieval process using vision foundation models (VFMs).
• Proposed EICSeg, an end-to-end framework integrating complementary VFMs with a lightweight SD-Adapter for enhanced segmentation.
• Developed a scalable self-prompt training strategy and an adaptive prompt selection mechanism for efficient training and inference.

Main Results:

• EICSeg achieved a 74.0% average Dice score on nine unseen datasets, outperforming existing methods by 4.5% in few-shot generalization.
• Demonstrated strong performance with a single prompt (60.1% Dice score), comparable to interactive models.
• Showcased automatic segmentation without manual prompt engineering, significantly reducing the need for labeled data.

Conclusions:

• EICSeg offers a powerful and efficient solution for universal medical image segmentation through explicit in-context learning.
• The framework exhibits remarkable few-shot generalization, adaptability across diverse datasets, and minimal data requirements.
• EICSeg represents a significant advancement in enabling automated and accurate medical image analysis with reduced human intervention.