SPW-TransUNet: three-dimensional computed tomography-cone beam computed tomography image registration with spatial...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Current Transformer-based medical image registration methods struggle with intensity differences and computational efficiency in 3D CT and CBCT.
• These limitations impede precise alignment for diagnosis and treatment planning.

Purpose of the Study:

• To develop a novel method enhancing feature interaction and computational efficiency for 3D medical image registration.
• To overcome limitations of existing methods in handling 3D CT and CBCT image alignment.

Main Methods:

• Introduced SPW-TransUNet integrating resizable spatial perpendicular window (SPW) attention and a self-learning mapping control (SLMC) mechanism.
• Utilized a mini convolutional neural network (CNN) within the UNet framework for adaptive feature vector transformation.
• Evaluated on CT-CBCT and inter-CT registration tasks using metrics like DICE, SSIM, TRE, and negative Jacobian percentage.

Main Results:

• Achieved lowest TRE (2.16 mm) and minimal negative Jacobian (0.126) for CT-CBCT lung image registration.
• Recorded highest SSIM (86.87%) and DICE (88.28%) for CT-CBCT lung images.
• Demonstrated strong performance on liver CT registration with peak SSIM (76.92%) and DICE (85.77%).

Conclusions:

• SPW-TransUNet significantly improves feature interaction and computational efficiency in 3D medical image registration.
• The method provides an effective solution for patient and target localization in image-guided radiation therapy.