Respiratory deformation registration in 4D-CT/cone beam CT using deep learning.

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence

Background:

• Accurate deformable image registration is crucial for lung cancer radiotherapy.
• Current methods for 4D-CT/4D-CBCT registration can be time-consuming and user-dependent.
• Developing automated and efficient registration techniques is essential for improving treatment planning and delivery.

Purpose of the Study:

• To evaluate the feasibility of a supervised convolutional neural network (CNN) for registering lung 4D-CT/4D-CBCT deformable vector fields.
• To assess the performance of a CNN-based method for 4D dose accumulation, contour propagation, motion modeling, and target verification.

Main Methods:

• A CNN-based deep learning model was developed to directly register deformation fields between phases of 4D-CT/4D-CBCT.
• The network was trained using patch pairs from lung images and validated against VelocityAI.
• A half mean squared error loss function guided the training, utilizing data from nine patients.

Main Results:

• The CNN-based registration achieved comparable accuracy to VelocityAI, with main anatomic features matching well.
• Cross-correlation coefficients above 0.9 were observed in the diaphragm region for intra-patient cases.
• The CNN deformation field successfully registered images, demonstrating good performance in key anatomical areas.

Conclusions:

• Patch-based deep learning methods provide accurate deformable registration comparable to VelocityAI.
• The developed CNN method is fully automatic, faster, and less user-dependent than VelocityAI.
• This deep learning approach shows significant promise for clinical applications in radiotherapy.