Deformable registration of chest CT images using a 3D convolutional neural network based on unsupervised learning

Area of Science:

• Medical Image Analysis
• Deep Learning in Radiology
• Computational Anatomy

Background:

• Deformable registration of 3D chest CT images is crucial for medical analysis.
• Respiratory motion causes nonlinear deformation and large displacements in lung tissues, posing significant challenges.
• Existing methods struggle with accuracy and speed due to these complexities.

Purpose of the Study:

• To develop an efficient and accurate unsupervised learning-based method for 3D chest CT image registration.
• To address the challenges of nonlinear deformation and large displacements in lung tissues.
• To improve the precision and speed of deformable image registration.

Main Methods:

• An end-to-end unsupervised learning framework was proposed, optimizing a U-Net architecture with inception modules.
• Jacobian regularization was incorporated into the loss function to prevent voxel folding and ensure a smooth displacement field.
• Data augmentation using 3D thin plate spline (TPS) transforms was employed to mitigate overfitting with limited datasets.

Main Results:

• The proposed method achieved a competitive target registration error of 2.09 mm and an optimal Dice score of 0.987.
• The technique demonstrated minimal voxel folding, outperforming VoxelMorph, ANTs, and Elastix.
• The registration process was significantly faster than traditional methods.

Conclusions:

• The developed method is efficient and robust for 3D chest CT image registration.
• The approach shows strong potential for clinical applications requiring accurate lung image alignment.
• The findings highlight the efficacy of unsupervised learning with architectural enhancements for complex medical image registration tasks.