Synthetic CT generation from CBCT images via unsupervised deep learning

Area of Science:

• Medical Physics
• Radiotherapy
• Medical Imaging

Background:

• Adaptive radiation therapy (ART) requires accurate anatomical data to adjust treatment plans.
• Cone-beam computed tomography (CBCT) is used for patient positioning but has inaccurate Hounsfield Units (HU), limiting its use in dose calculations.
• Current adaptive re-planning methods using CBCT are affected by artifacts and noise, impacting accuracy.

Purpose of the Study:

• To develop an unsupervised style-transfer method for generating synthetic CT (sCT) images from CBCT and planning CT (pCT).
• To create sCT images that retain CBCT's anatomical structure while achieving accurate HU values similar to pCT.
• To enable advanced applications in adaptive radiation therapy by improving image quality for dose calculation and treatment planning.

Main Methods:

• An unsupervised style-transfer deep learning model was proposed, utilizing CBCT for anatomical information and pCT for HU accuracy.
• The model was trained using a loss function that minimized contextual loss (CBCT structure) and style loss (pCT image quality).
• The model was trained and validated on 114 patient cases and tested on 29 independent cases.

Main Results:

• The generated sCT images demonstrated superior quality compared to CBCT, with higher structure-similarity index (0.9723 vs 0.9182) and peak-signal-to-noise ratio (33.68 vs 29.67).
• Mean absolute error in HU was significantly lower for sCT (28.52) than for CBCT (49.90).
• Quantitative comparisons confirmed the model's effectiveness in synthesizing CT-quality images from CBCT and pCT.

Conclusions:

• The proposed unsupervised style-transfer approach effectively generates synthetic CT images with accurate HU values from CBCT and pCT.
• This method addresses the limitations of CBCT in adaptive radiation therapy by providing high-quality images suitable for dose calculation.
• The developed model shows promise for enabling advanced adaptive treatment planning applications, improving patient care.