Eliminating Registration Bias in Synthetic CT Generation using a physics-based simulation framework for pelvic anatomy

Why clinicians should know about it

• Picked for Medical Physics (paper of the day, 29 August 2026): Physics‑based synthetic CT reduces registration bias for CBCT

Abstract

Supervised synthetic computed tomography (sCT) generation from cone-beam CT (CBCT) requires spatially registered training pairs, yet perfect registration between separately acquired scans is unattainable. This registration bias propagates into trained models and corrupts intensity-based evaluation, so higher benchmark scores may reward reproduction of registration artifacts over anatomical fidelity. We propose physics-based CBCT simulation for geometrically aligned training pairs by construction, with bias-robust geometric metrics.

Approach:A framework simulated pelvic CBCT from fan-beam CT, modeling respiratory motion, X-ray scatter, and noise to yield aligned simulated-CBCT/CT pairs. On a clinical gynecological dataset (deformable registration) and the SynthRAD2023 pelvic dataset (rigid registration), sCT models trained on simulated data were compared against models trained on real pairs, a finetuned variant, and CycleGAN and RegGAN baselines. Evaluation combined intensity metrics (MAE, PSNR, SSIM) with geometric alignment metrics (normalized mutual information, NMI; correlation coefficient, CC) against input CBCT. Downstream segmentation of bladder, rectum and bowel bag was assessed in two modes: an sCT cascade applying a CT-trained model to sCT outputs, and direct segmentation by a model trained on simulated CBCT, plus a physics ablation and five-observer quality assessment.

Main results:Simulation-trained models achieved higher geometric alignment than real-trained models (cross-dataset NMI 0.31 vs 0.22) despite lower intensity scores. Intensity metrics correlated inversely with observer ratings under deformable registration, whereas NMI consistently predicted clinical preference (clinical ρ = 0.29, SynthRAD ρ = 0.31). Observers preferred simulation-trained outputs in 87% of cases. In the sCT cascade, simulation-trained models improved segmentation (DSC 0.91/0.86/0.54 vs 0.84/0.77/0.04), while direct simulation-trained CBCT segmentation reached 0.92/0.87/0.83, exceeding a phantom-based baseline on the bowel bag.

Significance:Physics-based simulation eliminates registration bias at its source, and downstream segmentation provides a task-based measure of sCT conversion quality that intensity metrics miss. Geometric fidelity, not intensity agreement with biased ground truth, aligns with the spatial-accuracy requirements of adaptive radiotherapy.

Abstract as published, via PubMed.