UAE: Universal Anatomical Embedding on multi-modality medical images.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Accurate identification of anatomical structures is vital for medical image analysis.
• Exemplar-based methods for landmark detection use self-supervised learning but face challenges with appearance variations and cross-modality matching.

Purpose of the Study:

• To develop a Unified framework for learning Anatomical Embeddings (UAE) that addresses limitations in current medical image analysis methods.
• To enable robust landmark detection and registration across different imaging modalities and conditions.

Main Methods:

• Proposed UAE framework incorporating semantic embedding learning with prototypical contrastive loss.
• Implemented a fixed-point-based matching strategy for improved voxel correspondence.
• Utilized an iterative approach for cross-modality embedding learning.

Main Results:

• UAE demonstrated superior performance in intra- and inter-modality tasks, including one-shot landmark detection and lesion tracking.
• Achieved state-of-the-art results in CT-MRI affine/rigid registration.
• Showcased robustness and versatility in diverse medical image analysis applications.

Conclusions:

• The proposed UAE framework offers a significant advancement in learning anatomical embeddings for medical image analysis.
• UAE provides a robust and versatile solution for landmark detection and image registration tasks.
• The framework's ability to handle appearance variations and cross-modality matching enhances its applicability in clinical settings.