TY - GEN
T1 - Conditional Iterative α-(de)Blending Model for CBCT-to-sCT Synthesis
T2 - International Workshop on Simulation and Synthesis in Medical Imaging
AU - Cao, Jiaming
AU - Sargeant, Chelsea A. H.
AU - McWilliam, Alan
AU - Osorio, Eliana Vasquez
PY - 2025/9/21
Y1 - 2025/9/21
N2 - Cone-beam CT (CBCT) is widely used in adaptive radiotherapy (ART) but often suffers from image artifacts and poor soft tissue contrast, limiting its wider application in ART workflows including segmentation and dose calculation. In this work, we propose a conditional Iterative α-(de)Blending (cIADB) for CBCT image quality improvement. cIADB employs a deterministic blending-deblending mechanism that reduces sampling randomness, enabling more stable and efficient image generation compared to conventional conditional denoising diffusion probabilistic model (cDDPM), which relies on stochastic sampling. We comprehensively evaluate the proposed method on head-and-neck CBCTs across different training approaches and anatomical planes. Quantitative results demonstrate that cIADB achieves better performance compared to cDDPM in terms of PSNR, SSIM, and SSE, while qualitative assessments further confirm improved denoising effect and structural fidelity. Moreover, the lightweight inference process of cIADB facilitates its potential integration into ART workflows. Our study highlights the promise of deterministic IADB model as a robust solution for clinical CBCT enhancement.
AB - Cone-beam CT (CBCT) is widely used in adaptive radiotherapy (ART) but often suffers from image artifacts and poor soft tissue contrast, limiting its wider application in ART workflows including segmentation and dose calculation. In this work, we propose a conditional Iterative α-(de)Blending (cIADB) for CBCT image quality improvement. cIADB employs a deterministic blending-deblending mechanism that reduces sampling randomness, enabling more stable and efficient image generation compared to conventional conditional denoising diffusion probabilistic model (cDDPM), which relies on stochastic sampling. We comprehensively evaluate the proposed method on head-and-neck CBCTs across different training approaches and anatomical planes. Quantitative results demonstrate that cIADB achieves better performance compared to cDDPM in terms of PSNR, SSIM, and SSE, while qualitative assessments further confirm improved denoising effect and structural fidelity. Moreover, the lightweight inference process of cIADB facilitates its potential integration into ART workflows. Our study highlights the promise of deterministic IADB model as a robust solution for clinical CBCT enhancement.
U2 - 10.1007/978-3-032-05573-6_15
DO - 10.1007/978-3-032-05573-6_15
M3 - Conference contribution
SN - 9783032055729
T3 - Lecture Notes in Computer Science
SP - 149
EP - 158
BT - Simulation and Synthesis in Medical Imaging
A2 - Fernandez, Virginia
A2 - Wiesner, David
A2 - Zuo, Lianrui
A2 - Casamitjana, Adrià
A2 - Remedios, Samuel W.
PB - Springer Cham
CY - Cham
Y2 - 23 September 2025 through 23 September 2025
ER -