Spatially Covariant Image Registration With Text Prompts

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Medical images possess structured anatomy and inhomogeneous contrasts, posing challenges for analysis.
• Anatomical priors in neural networks can improve medical imaging utility, especially in resource-limited clinical settings.
• While effective for segmentation, anatomical priors have shown limited progress in deformable image registration.

Purpose of the Study:

• To introduce textSCF, a novel method for deformable image registration that leverages anatomical priors and textual prompts.
• To enhance computational efficiency and registration accuracy in medical imaging tasks.
• To address the modest progress in deformable image registration by integrating novel techniques.

Main Methods:

• textSCF integrates spatially covariant filters with textual anatomical prompts encoded by visual-language models.
• An implicit function correlates text embeddings of anatomical regions to filter weights for optimization.
• The method was evaluated on intersubject brain MRI and abdominal CT registration tasks.

Main Results:

• textSCF demonstrated superior performance, outperforming state-of-the-art models in the MICCAI Learn2Reg 2021 challenge.
• In abdominal CT registrations, textSCF improved the Dice score by 11.3% compared to the second-best model.
• A smaller textSCF variant achieved similar accuracy with significant reductions in network parameters (89.13%) and computational operations (98.34%).

Conclusions:

• textSCF offers improved computational efficiency and registration accuracy by capturing contextual interplay between anatomical regions.
• The method exhibits impressive interregional transferability and preserves structural discontinuities during registration.
• textSCF represents a significant advancement in deformable image registration, particularly for brain MRI and abdominal CT.