LDVoxelMorph: A precise loss function and cascaded architecture for unsupervised diffeomorphic large displacement...

Area of Science:

• Medical imaging
• Computer vision
• Computational anatomy

Background:

• Traditional non-rigid medical image registration methods often struggle to balance registration accuracy with the diffeomorphic property.
• The diffeomorphic property is essential for ensuring the credibility and reliability of registration outcomes in medical image analysis.

Purpose of the Study:

• To improve the diffeomorphic property and registration accuracy in 3D medical image registration.
• To address the limitations of invariant smoothness regularization parameters in existing learning-based methods.

Main Methods:

• Proposed LDVoxelMorph, a diffeomorphic cascaded network utilizing a compressed loss (CL).
• Employed deep supervision with cascade-variant smoothness regularization parameters across constituent U-Nets.
• Developed CL as a penalty for the velocity field to prevent deformation field overlap.

Main Results:

• Achieved high Dice scores: 0.892 ± 0.040 (SLIVER), 0.848 ± 0.044 (LiTS), and 0.689 ± 0.014 (LPBA).
• Demonstrated minimal overlapping voxels (325, 159, 0) in deformation fields across datasets.
• Ablation studies confirmed CL's superior effectiveness in enhancing the diffeomorphic property.

Conclusions:

• LDVoxelMorph achieves superior registration accuracy and maintains the diffeomorphic property, even with large deformations.
• The proposed method offers a more credible and accurate solution for 3D medical image registration tasks.