LEARNING MRI CONTRAST-AGNOSTIC REGISTRATION.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Classical image registration methods are accurate but computationally intensive, requiring optimization for each image pair.
• Current learning-based methods are fast but limited to training data's image contrasts and content.
• Existing approaches struggle with generalization across diverse magnetic resonance imaging (MRI) contrasts.

Purpose of the Study:

• To develop a learning-based image registration strategy independent of acquired imaging data and MRI contrast.
• To enhance the generalization capabilities of registration networks across various unseen contrasts.
• To eliminate the need for real imaging data during the training of registration models.

Main Methods:

• A generative strategy was employed, synthesizing a wide range of images from segmentations during network training.
• Networks were exposed to diverse, synthesized image contrasts to promote generalization.
• Training utilized arbitrary shapes from noise distributions and anatomical label maps to synthesize images.

Main Results:

• Networks trained with the proposed framework generalized effectively to unseen MRI contrasts.
• The method surpassed classical state-of-the-art brain registration accuracy by up to 12.4 Dice points.
• Synthesizing images from anatomical label maps significantly boosted performance.

Conclusions:

• The generative strategy enables robust image registration without reliance on acquired imaging data or specific MRI contrasts.
• This approach offers a powerful alternative to traditional and existing learning-based registration methods.
• The framework demonstrates the potential for data-efficient and highly generalizable medical image registration.