Generating synthesized computed tomography from CBCT using a conditional generative adversarial network for head and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Cone-beam computed tomography (CBCT) suffers from imaging artifacts and Hounsfield unit (HU) inaccuracy.
• These limitations hinder its use in quantitative applications like radiotherapy planning.
• High-quality computed tomography (CT) images are crucial for accurate dose calculation and treatment delivery.

Purpose of the Study:

• To develop a method for synthesizing high-quality CT-like images from CBCT images.
• To address the limitations of CBCT, specifically imaging artifacts and HU inaccuracy.
• To enable quantitative applications of CBCT-derived images in radiotherapy.

Main Methods:

• A conditional generative adversarial network (cGAN) with a U-Net backbone and residual blocks was developed.
• The cGAN learned a mapping from CBCT images to planning CT images.
• 120 paired CBCT and CT scans from head and neck cancer patients were used for training, validation, and testing.

Main Results:

• The cGAN successfully synthesized CT-like images that were visually similar to planning CT images.
• Quantitative metrics (MAE, RMSE, SSIM, PSNR) showed the cGAN significantly outperformed U-Net and CycleGAN.
• The cGAN achieved MAE of 16.75 ± 11.07 HU, RMSE of 58.15 ± 28.64 HU, SSIM of 0.92 ± 0.04, and PSNR of 30.58 ± 3.86 dB.

Conclusions:

• The proposed cGAN method generates synthetic CT images with accurate Hounsfield units and preserved anatomical structures from CBCT.
• This technique effectively overcomes CBCT's inherent limitations.
• The synthesized CT images are suitable for quantitative applications in radiotherapy, enhancing treatment planning and delivery.