Generating synthesized computed tomography (CT) from cone-beam computed tomography (CBCT) using CycleGAN for adaptive...

Area of Science:

• Medical Physics
• Radiology
• Artificial Intelligence

Background:

• Adaptive radiation therapy (ART) requires accurate patient imaging throughout treatment, as anatomy can change significantly.
• Cone-beam computed tomography (CBCT) is used for positioning and re-planning in ART, but suffers from noise, artifacts, and inaccurate Hounsfield Unit (HU) values, compromising dose calculation accuracy.
• Synthesized CT (sCT) images derived from CBCT can potentially improve accuracy for treatment planning.

Purpose of the Study:

• To develop and evaluate a CycleGAN framework for synthesizing CT images from CBCT images.
• To assess the accuracy of synthesized CT (sCT) images for radiation dose calculation in adaptive radiation therapy.
• To compare the performance of CycleGAN against other unsupervised learning methods and existing image registration techniques.

Main Methods:

• A cycle-consistent generative adversarial network (CycleGAN) was developed for unsupervised image-to-image translation from CBCT to sCT.
• The model was trained and evaluated using unpaired CT and CBCT images, particularly for head-and-neck (H&N) cancer patients.
• Quantitative metrics including Mean Absolute Error (MAE) and 3D gamma index analysis were used to assess image and dose accuracy. Comparisons were made with Deep Convolutional Generative Adversarial Networks (DCGAN), Progressive Growing of GANs (PGGAN), and Deformable Image Registration (DIR).

Main Results:

• CycleGAN significantly reduced the MAE from 69.29 HU to 29.85 HU, producing sCT images visually and quantitatively similar to real CT images.
• Dose distributions calculated on sCT images showed higher accuracy than those on CBCT, with the gamma index pass rate increasing from 86.92% to 96.26% (1 mm/1% criteria) compared to the deformed planning CT (dpCT).
• CycleGAN outperformed DCGAN and PGGAN in unsupervised learning and demonstrated superior anatomical accuracy compared to DIR in phantom studies (Structural Similarity Index increased from 0.91 to 0.93).

Conclusions:

• CycleGAN is an effective unsupervised method for synthesizing accurate CT images from CBCT, crucial for improving adaptive radiation therapy.
• The generated sCT images enhance the precision of radiation dose calculations, leading to more reliable treatment planning.
• This approach offers a promising solution for overcoming the limitations of CBCT in ART, improving patient outcomes.