CCMNet: Cross-scale correlation-aware mapping network for 3D lung CT image registration.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• The lung's high elasticity and complex structure lead to substantial shape variability, posing challenges for accurate image registration.
• Traditional U-Net architectures struggle with large deformations due to limited receptive fields and weakened long-range dependencies.

Purpose of the Study:

• To develop a novel multilevel registration framework to enhance voxel correspondence for improved large deformation estimation in lung CT images.
• To address the limitations of existing methods in handling complex lung deformations and long-range dependencies.

Main Methods:

• A convolutional neural network (CNN) with a two-stream registration structure for robust feature extraction.
• A cross-scale mapping attention (CSMA) mechanism to establish frequent connections between network layers, maintaining image pair correlation.
• Leveraging multi-scale context information to map relationships between low- and high-resolution feature maps.

Main Results:

• Achieved TRE of 1.56 ± 1.60 on the DIRLAB dataset.
• Obtained NCC of 99.72% and SSIM of 91.42% on the POPI dataset.
• Demonstrated effective handling of large deformation issues and mitigation of long-range dependence.

Conclusions:

• The proposed multilevel registration framework significantly improves the accuracy and robustness of lung CT image registration.
• The CSMA mechanism effectively utilizes multi-scale information to overcome challenges associated with large deformations.
• This approach offers a promising solution for precise lung image analysis and clinical applications.