A deformable patch-based transformer for 3D medical image registration.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image registration is crucial for clinical applications but faces challenges with complex physiological structures.
• Current registration algorithms require further development for enhanced accuracy and speed.

Purpose of the Study:

• To design a 3D medical image registration algorithm capable of high accuracy and speed for complex structures.
• To address limitations in existing medical image registration techniques.

Main Methods:

• Developed DIT-IVNet, an unsupervised learning algorithm for 3D medical image registration.
• Employed a hybrid convolution and transformer network architecture.
• Introduced a 3D_Depatch module for adaptive patch embedding and inception blocks for multi-scale feature learning.

Main Results:

• DIT-IVNet demonstrated superior performance across Dice score, Negative Jacobian determinant, Hausdorff distance, and Structural Similarity metrics compared to state-of-the-art methods.
• The model achieved the highest Dice score in generalization experiments, indicating robust generalizability.

Conclusions:

• The proposed unsupervised registration network, DIT-IVNet, effectively performs deformable medical image registration.
• The network architecture surpasses current state-of-the-art methods for brain dataset registration.