Deep learning-based thoracic CBCT correction with histogram matching.

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Kilovoltage cone-beam computed tomography (CBCT) is crucial for daily image-guided radiation therapy (IGRT), particularly for stereotactic body radiation therapy (SBRT).
• Clinical use of CBCT is limited by poor soft tissue contrast, image artifacts, and unstable Hounsfield unit (HU) values, impacting setup accuracy.

Purpose of the Study:

• To develop a novel deep learning-based method for generating synthetic CTs (sCT) from thoracic CBCT images.
• To enhance image quality and accuracy for IGRT applications, especially for SBRT.

Main Methods:

• A deep learning model, HM-Cycle-GAN, integrating histogram matching (HM) into a Cycle-GAN framework, was trained to map thoracic CBCTs to paired planning CTs.
• Perceptual supervision and an informative maximizing loss were employed to minimize blurring and evaluate histogram matching.
• Deformable image registration was used for pre-processing to reduce anatomy mismatch effects during training and assessment.

Main Results:

• The HM-Cycle-GAN method achieved average evaluation metrics of 66.2 HU (MAE), 30.3 dB (PSNR), and 0.95 (NCC) across all CBCT fractions.
• The proposed method demonstrated superior image quality, with reduced noise and artifact severity compared to the standard Cycle-GAN.
• Planning CTs served as the ground truth for evaluating the derived sCTs.

Conclusions:

• The developed HM-Cycle-GAN method effectively generates high-quality synthetic CTs from thoracic CBCTs.
• This technique can improve the accuracy of IGRT by providing better soft tissue contrast and reduced artifacts.
• Corrected CBCTs hold potential for improving online adaptive radiotherapy through enhanced contouring accuracy and dose calculation.