CBCT-based synthetic CT generation using generative adversarial networks with disentangled representation

Area of Science:

• Medical Imaging
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Cone-beam computed tomography (CBCT) is crucial for image-guided radiotherapy (IGRT).
• Poor CBCT image quality has limited its clinical use.
• This study addresses CBCT image quality limitations.

Purpose of the Study:

• To develop a deep learning approach for CBCT-to-CT image translation.
• To generate synthetic CT (sCT) images with improved quality from CBCT.
• To preserve anatomical structures from CBCT in the generated sCT images.

Main Methods:

• A novel synthetic CT generative adversarial network (sCTGAN) was developed.
• Disentangled representation was used for CBCT-to-CT translation.
• The network was trained on 40 patients' CBCT and CT data and tested on 12 patients' data.

Main Results:

• sCTGAN achieved superior quantitative metrics (PSNR, SSIM, MAE, RMSE) compared to deformed planning CT (dpCT).
• sCTGAN's RMSE was 60.53 HU, significantly lower than original CBCT (112.13 HU).
• sCTGAN outperformed three state-of-the-art CycleGAN-based methods in RMSE (60.53 HU vs. 64.93-72.40 HU).

Conclusions:

• The sCTGAN network generates high-quality sCT images closer to the ground truth (dpCT).
• This method offers an effective solution for improving CBCT image quality.
• The generated sCT has broad applications in IGRT and adaptive radiotherapy.