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

• DOI: 10.1088/1361-6560/ad1cfc
• Publication Date: 2024-01-10T22:24:22Z
• Authors: Branimir Rusanov; Ghulam Mubashar H...; Mark Reynolds; Mahsheed Sabet; Pejman Rowshanfarzad; Nicholas Bucknell; Suki Gill; Joshua Dass; Martin Ebert
• Keywords: Hounsfield scale; Concordance correlation coefficient

Abstract Objective . Clinical implementation of synthetic CT (sCT) from cone-beam (CBCT) for adaptive radiotherapy necessitates a high degree anatomical integrity, Hounsfield unit (HU) accuracy, and image quality. To achieve these goals, vision-transformer anatomically sensitive loss functions are described. Better quantification quality is achieved using the alignment-invariant Fréchet inception distance (FID), uncertainty estimation sCT risk prediction implemented in scalable plug-and-play manner. Approach Baseline U-Net, generative adversarial network (GAN), CycleGAN models were trained to identify shortcomings each approach. The proposed CycleGAN-Best model was empirically optimized based on large ablation study evaluated classical metrics, FID, gamma index, segmentation analysis. Two methods, Monte-Carlo Dropout (MCD) test-time augmentation (TTA), introduced epistemic aleatoric uncertainty. Main results FID correlated blind observer scores with Correlation Coefficient −0.83, validating metric as an accurate quantifier perceived mean absolute error (MAE) 42.11 ± 5.99 25.00 1.97 HU, compared 63.42 15.45 31.80 HU CycleGAN-Baseline, 144.32 20.91 68.00 5.06 CBCT, respectively. Gamma 1%/1 mm pass rates 98.66 0.54% CycleGAN-Best, 86.72 2.55% CBCT. TTA MCD-based maps well spatially poor synthesis outputs. Significance Anatomical accuracy by suppressing CycleGAN-related artefacts. better discriminated quality, where alignment-based metrics such MAE erroneously suggest poorer outputs perform better. Uncertainty shown correlate has clinical relevancy toward assessment assurance. accompanying evaluation tools necessary additions clinically robust generation models.