Transformer CycleGAN with uncertainty estimation for CBCT based synthetic CT in adaptive radiotherapy

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Clinical implementation of synthetic CT (sCT) from cone-beam CT (CBCT) requires high anatomical integrity, Hounsfield unit (HU) accuracy, and image quality.
• Existing methods face challenges in achieving the necessary precision for adaptive radiotherapy.

Purpose of the Study:

• To develop and evaluate an advanced CycleGAN model (CycleGAN-Best) for generating high-fidelity sCT from CBCT.
• To implement robust image quality quantification using Fréchet Inception Distance (FID) and uncertainty estimation for risk assessment.

Main Methods:

• A vision-transformer and anatomically sensitive loss functions were employed within a CycleGAN framework.
• Empirical optimization via ablation studies and evaluation using FID, gamma index, and segmentation analysis.
• Incorporation of Monte-Carlo Dropout (MCD) and test-time augmentation (TTA) for uncertainty estimation.

Main Results:

• CycleGAN-Best achieved superior image quality with FID of 42.11 ± 5.99 and MAE of 25.00 ± 1.97 HU, outperforming baseline models and CBCT.
• Gamma 1%/1 mm pass rates reached 98.66 ± 0.54% for CycleGAN-Best, significantly higher than CBCT's 86.72 ± 2.55%.
• FID demonstrated strong correlation with perceived image quality (r = -0.83), while uncertainty maps correlated with synthesis errors.

Conclusions:

• The CycleGAN-Best model effectively suppresses artefacts, ensuring anatomical accuracy and improved HU accuracy for sCT generation.
• FID is a more reliable metric for image quality assessment than alignment-based metrics like MAE.
• Uncertainty estimation provides clinically relevant insights for model risk assessment and quality assurance in adaptive radiotherapy.