Evaluation of a cycle-generative adversarial network-based cone-beam CT to synthetic CT conversion algorithm for...

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Adaptive radiation therapy (ART) enhances image-guided radiation therapy.
• Synthetic CT (sCT) generation from cone-beam CT (CBCT) is crucial for ART implementation.
• Cycle-generative adversarial networks (cycle-GANs) show promise for medical image synthesis.

Purpose of the Study:

• To evaluate a cycle-GAN-based CBCT-to-sCT conversion algorithm.
• Assess image quality, segmentation accuracy, and dosimetric accuracy across head and neck (H&N), thoracic, and pelvic regions.
• Determine the clinical feasibility of CBCT-derived sCT for ART.

Main Methods:

• Trained three body site-specific cycle-GAN models using paired CT and CBCT datasets.
• Generated sCT from first-fraction CBCT for 15 patients per region.
• Analyzed sCT image quality (ME, MAE), segmentation accuracy (Dice, surface distance), and dosimetric accuracy (dose-volume parameters, 3D-gamma).

Main Results:

• Mean errors (ME) and mean absolute errors (MAE) for sCT were within acceptable ranges across regions.
• High Dice similarity coefficients (up to 94.9% for lungs) and acceptable surface distances (e.g., 3.5 mm for brainstem) were observed.
• Mean dosimetric differences for target volumes were <1.7%, with >97.8% 3D-gamma pass rates.

Conclusions:

• The cycle-GAN method generates sCT images with quality comparable to planning CT (pCT).
• Clinically acceptable dosimetric deviations were achieved, validating the approach.
• This fulfills a key requirement for the clinical integration of CBCT-based ART.