CT respiratory motion synthesis using joint supervised and adversarial learning.

Area of Science:

• Medical Imaging
• Radiotherapy
• Deep Learning

Background:

• Four-dimensional computed tomography (4DCT) tracks organ and tumor motion for radiotherapy planning but increases complexity and radiation dose.
• Current 4DCT methods may not perfectly align with patient breathing during treatment, impacting accuracy.

Purpose of the Study:

• To develop a deep synthesis method for generating pseudo-respiratory CT phases from static images for motion-aware radiotherapy planning.
• To reduce radiation exposure and protocol complexity associated with conventional 4DCT.

Main Methods:

• A deep synthesis model generates patient-specific deformation vector fields (DVFs) conditioned on surface-based motion estimation.
• Supervised DVF training and an adversarial loss on both warped images and DVF magnitude ensure realism and respiratory correlation.
• The method avoids excessive smoothness often seen in unsupervised deep learning approaches.

Main Results:

• Generated pseudo-respiratory CT phases accurately capture organ and tumor motion, comparable to repeated 4DCT scans.
• Tumor center-of-mass distances (1.97 mm vs. 2.35 mm) and Dice similarity coefficients (0.63 vs. 0.71) show favorable results for synthetic vs. real 4DCT.
• Performance compares favorably against a state-of-the-art technique (RMSim).

Conclusions:

• The proposed deep image synthesis method effectively generates pseudo-respiratory CT phases from static images, addressing 4DCT limitations.
• This approach has the potential to reduce radiation exposure in radiotherapy planning while maintaining accurate motion representation.
• Further studies are needed to assess the dosimetric impact.