CIM-VTP: Correlation-Guided Image Modeling with Visual-Textual Task Prompt for Universal Medical Image Registration.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Universal medical image registration aims for a single model across diverse tasks.
• Existing deep learning methods struggle with generalizable features and dynamic adaptation for cross-task registration.
• Fixed model architectures limit flexibility and zero-shot performance on unseen tasks.

Purpose of the Study:

• To propose a novel two-stage universal registration framework, CIM-VTP, addressing limitations in current deep learning methods.
• To enhance generalizable feature representation and dynamic task adaptation for universal medical image registration.
• To improve zero-shot performance on unseen registration tasks.

Main Methods:

• Developed a two-stage framework: CIM-VTP.
• Stage 1: Correlation-guided Image Modeling (CIM) pretraining for spatial correspondence and universal representations.
• Stage 2: Registration task classifier and Visual-Textual Task Prompt (VTP) modules for adaptive decoder adjustment using task similarity scores.

Main Results:

• CIM-VTP demonstrated superior performance in universal medical image registration.
• The framework achieved strong results across six diverse registration tasks.
• The approach effectively captures spatial correspondence and adapts to different input domains.

Conclusions:

• CIM-VTP offers a promising solution for universal medical image registration.
• The proposed pretraining and adaptive prompting strategies enhance generalization and flexibility.
• The framework advances the state-of-the-art in handling various registration tasks with a single model.