Unsupervised multimodal deformable medical image registration based on feature perceptual contrast learning

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Multimodal medical image registration is crucial for clinical diagnosis but challenged by modality differences.
• Existing methods struggle with noise sensitivity or introduce inaccuracies in anatomical alignment.

Purpose of the Study:

• To develop a novel unsupervised approach for accurate multimodal medical image registration.
• To learn robust feature descriptors that bridge modality gaps while preserving anatomical details.

Main Methods:

• Proposed a Feature Perceptual Contrast Learning Network (FP-net) for unsupervised learning of modality-bridging descriptors.
• Utilized local sampling-based feature perceptual contrast learning and image reconstruction.
• Employed the trained FP-net to drive an unsupervised registration framework without ground-truth deformation fields.

Main Results:

• Achieved superior performance over state-of-the-art methods on BraTS 2021 and Learn2Reg 2021 datasets.
• Reached Dice Similarity Coefficients (DSC) of 76.3% (T2-T1) and 77.7% (T1-T1ce) for brain registration.
• Obtained a DSC of 50.1% for challenging abdominal CT-MR registration, significantly improving structural alignment.

Conclusions:

• The FP-net effectively bridges the modality gap, enabling standard U-Net architectures for state-of-the-art deformable registration.
• Provides a robust, accurate, and deployable unsupervised solution for clinical multimodal image analysis.