Related Experiment Video
Updated: Jun 20, 2025
07:13
Author Spotlight: An Efficient and Robust Software for Automated Fusion of Multiple Preclinical Imaging Modalities
Published on: October 27, 2023
1.1K
Spatial-aware contrastive learning for cross-domain medical image registration
Chenchu Rong1, Zhiru Li1, Rui Li2
1School of Electronic Science and Engineering, Nanjing University, Nanjing, China.
Medical Physics
|July 20, 2024
Summary
This study introduces a spatial-aware contrastive learning method for cross-domain medical image registration, significantly improving accuracy between CT and MRI scans. The approach optimizes feature representation, enhancing diagnostic precision and treatment planning for better patient outcomes.
Area of Science:
- Medical Imaging Analysis
- Machine Learning in Healthcare
Background:
- Precise medical image analysis is vital for patient care.
- Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) offer complementary imaging strengths.
- Cross-domain registration faces challenges due to modality-specific differences.
Purpose of the Study:
- To develop a spatial-aware contrastive learning approach for CT-MRI cross-domain registration.
- To optimize feature space representation using reconstruction and contrastive losses.
- To enhance structural similarity learning and registration accuracy across imaging domains.
Main Methods:
- Extracting equivalent feature spaces from CT and MRI images for cross-domain matching.
- Utilizing an autoencoder-like structure with custom reconstruction and contrastive losses.
- Implementing region masks to balance spatial correlation and distinctiveness in feature representation.
Main Results:
- Achieved high accuracy with Dice Similarity Coefficient (DSC) of 85.68%, Target Registration Error (TRE) of 1.92 mm, and Mean Hausdorff Distance (MHD) of 1.26 mm.
- Demonstrated significant reduction in registration time to 2.67 seconds on GPU.
- Validated effectiveness and adaptability across various medical imaging scenarios.
Conclusions:
- The proposed spatial-aware contrastive learning method offers a novel solution for cross-domain medical image registration.
- The approach significantly enhances registration accuracy and stability between CT and MRI.
- Results show substantial application value for precise diagnosis and personalized treatment planning.
