Paired cycle-GAN-based image correction for quantitative cone-beam computed tomography.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Cone-beam computed tomography (CBCT) is crucial for image-guided radiation therapy, but its clinical utility is limited by severe image artifacts.
• Developing methods to improve CBCT image quality is essential for advancing radiation therapy techniques.

Purpose of the Study:

• To propose a deep learning-based method for generating high-quality corrected CBCT (CCBCT) images.
• To enhance the clinical potential of CBCT in image-guided radiation therapy.

Main Methods:

• A residual block concept was integrated into a cycle-consistent adversarial network (cycle-GAN) framework, termed res-cycle GAN.
• The model learns a mapping between CBCT and planning CT images, enabling end-to-end CBCT-to-CT transformations.
• The algorithm was evaluated on patient data from brain and pelvis scans, using metrics like MAE, PSNR, NCC, and SNU.

Main Results:

• The proposed method significantly improved image quality metrics compared to standard CBCT images in both brain and pelvis datasets.
• Quantitative analysis showed substantial reductions in mean absolute error (MAE) and spatial non-uniformity (SNU), and improvements in peak signal-to-noise ratio (PSNR) and normalized cross-correlation (NCC).
• The res-cycle GAN method outperformed conventional scatter correction and another machine learning-based CBCT correction technique in reducing noise and artifacts.

Conclusions:

• A novel deep learning-based method effectively generates high-quality corrected CBCT images.
• The developed technique significantly enhances onboard CBCT image quality, making it comparable to planning CT.
• This advancement holds promise for enabling quantitative adaptive radiation therapy with further clinical evaluation.