Enhancing medical image registration via appearance adjustment networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deformable image registration is crucial for medical image analysis but is challenged by appearance variations (texture, intensity, noise).
• Deep learning-based registration (DLR) methods offer speed but often lack adaptability to specific image variations.
• Existing DLRs struggle with significant appearance variations common in medical imaging, particularly brain MRI.

Purpose of the Study:

• To propose an Appearance Adjustment Network (AAN) that enhances the adaptability of DLRs to appearance variations.
• To develop an anatomy-constrained loss function for AAN to ensure anatomy-preserving transformations.
• To integrate AAN into existing DLRs for improved performance without significant computational overhead.

Main Methods:

• Developed an Appearance Adjustment Network (AAN) for DLRs to perform appearance transformations.
• Incorporated an anatomy-constrained loss function to guide AAN in generating anatomy-preserving transformations.
• Integrated and trained AAN cooperatively with three state-of-the-art DLRs (Voxelmorph, DifVM, LapIRN) in an unsupervised, end-to-end manner.

Main Results:

• The AAN consistently improved the performance of existing DLRs across three public 3D brain MRI datasets (IBSR18, Mindboggle101, LPBA40).
• The integrated AAN-DLR models demonstrated superior registration accuracy compared to state-of-the-art optimization-based registration methods (ORs).
• The AAN added only a fractional computational load to the existing DLRs, maintaining efficiency.

Conclusions:

• The proposed Appearance Adjustment Network (AAN) effectively addresses appearance variations in medical image registration.
• AAN enhances the adaptability and accuracy of deep learning-based registration methods.
• This approach offers a significant advancement for accurate and robust deformable image registration in clinical applications.