uniGradICON: A Foundation Model for Medical Image Registration

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Conventional medical image registration optimizes transformation parameters, offering generic applicability but lacking speed.
• Deep learning registration networks provide speed and accuracy but are task-specific, limiting their general use.

Purpose of the Study:

• To introduce uniGradICON, a foundation model for medical image registration.
• To achieve performance across multiple datasets and enable zero-shot registration for new tasks.
• To combine the speed of deep learning with the generic applicability of conventional methods.

Main Methods:

• Developed uniGradICON, a novel deep learning model for medical image registration.
• Trained and evaluated the model on twelve diverse public datasets.
• Demonstrated zero-shot capabilities for unseen registration tasks, acquisitions, regions, and modalities.

Main Results:

• UniGradICON achieved high performance across multiple datasets, overcoming limitations of current learning-based methods.
• The model demonstrated effective zero-shot capabilities for new registration tasks.
• UniGradICON provides a strong initialization for fine-tuning on out-of-distribution tasks.

Conclusions:

• UniGradICON represents a significant step towards a generic foundation model for medical image registration.
• The model successfully integrates the speed and accuracy of deep learning with the broad applicability of conventional techniques.
• The availability of code and weights facilitates further research and development in medical image registration.