Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Eliminating Registration Bias in Synthetic CT Generation: A Physics-Based Simulation Framework

Lukas Zimmermann, Michael Rauter, Maximilian Schmid, Dietmar Georg, Barbara Knäusl

TL;DR

This work tackles registration bias in supervised CBCT-to-CT synthesis by introducing a physics-based CBCT simulation that yields geometrically aligned training pairs and evaluating models with geometric alignment against input CBCT. The authors develop a three-stage pipeline (motion simulation, forward projection, reconstruction) implemented in CUDA/CuPy to generate large-scale, geometrically coherent data, enabling training without imperfect ground-truth alignment. They demonstrate that models trained on simulated data achieve superior geometric fidelity (e.g., $NMI$ improvements from $0.22$ to $0.31$) and are more aligned with clinical perception, as reflected by observer preferences, even when intensity metrics are lower and across different registration methods. The findings suggest that incorporating geometric alignment into evaluation and prioritizing spatial fidelity can improve clinical relevance for adaptive radiotherapy, with the framework offering scalable, site-specific data generation and potential applicability to segmentation and registration tasks.$

Abstract

Supervised synthetic CT generation from CBCT requires registered training pairs, yet perfect registration between separately acquired scans remains unattainable. This registration bias propagates into trained models and corrupts standard evaluation metrics. This may suggest that superior benchmark performance indicates better reproduction of registration artifacts rather than anatomical fidelity. We propose physics-based CBCT simulation to provide geometrically aligned training pairs by construction, combined with evaluation using geometric alignment metrics against input CBCT rather than biased ground truth. On two independent pelvic datasets, models trained on synthetic data achieved superior geometric alignment (Normalized Mutual Information: 0.31 vs 0.22) despite lower conventional intensity scores. Intensity metrics showed inverted correlations with clinical assessment for deformably registered data, while Normalized Mutual Information consistently predicted observer preference across registration methodologies (rho = 0.31, p < 0.001). Clinical observers preferred synthetic-trained outputs in 87% of cases, demonstrating that geometric fidelity, not intensity agreement with biased ground truth, aligns with clinical requirements.

Eliminating Registration Bias in Synthetic CT Generation: A Physics-Based Simulation Framework

TL;DR

This work tackles registration bias in supervised CBCT-to-CT synthesis by introducing a physics-based CBCT simulation that yields geometrically aligned training pairs and evaluating models with geometric alignment against input CBCT. The authors develop a three-stage pipeline (motion simulation, forward projection, reconstruction) implemented in CUDA/CuPy to generate large-scale, geometrically coherent data, enabling training without imperfect ground-truth alignment. They demonstrate that models trained on simulated data achieve superior geometric fidelity (e.g., improvements from to ) and are more aligned with clinical perception, as reflected by observer preferences, even when intensity metrics are lower and across different registration methods. The findings suggest that incorporating geometric alignment into evaluation and prioritizing spatial fidelity can improve clinical relevance for adaptive radiotherapy, with the framework offering scalable, site-specific data generation and potential applicability to segmentation and registration tasks.$

Abstract

Supervised synthetic CT generation from CBCT requires registered training pairs, yet perfect registration between separately acquired scans remains unattainable. This registration bias propagates into trained models and corrupts standard evaluation metrics. This may suggest that superior benchmark performance indicates better reproduction of registration artifacts rather than anatomical fidelity. We propose physics-based CBCT simulation to provide geometrically aligned training pairs by construction, combined with evaluation using geometric alignment metrics against input CBCT rather than biased ground truth. On two independent pelvic datasets, models trained on synthetic data achieved superior geometric alignment (Normalized Mutual Information: 0.31 vs 0.22) despite lower conventional intensity scores. Intensity metrics showed inverted correlations with clinical assessment for deformably registered data, while Normalized Mutual Information consistently predicted observer preference across registration methodologies (rho = 0.31, p < 0.001). Clinical observers preferred synthetic-trained outputs in 87% of cases, demonstrating that geometric fidelity, not intensity agreement with biased ground truth, aligns with clinical requirements.
Paper Structure (26 sections, 11 equations, 4 figures, 7 tables)

This paper contains 26 sections, 11 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Previous work
  3. Materials and Methods
  4. Physics simulation framework
  5. Respiratory Motion Simulation
  6. Motion Field Derivation
  7. Motion-adapted Volume Generation
  8. Dynamic X-ray Projection
  9. Contrast Media Removal
  10. Scanner Geometry Configuration
  11. Motion-Adapted Projection Generation
  12. Physics-Based Imaging Model
  13. Scatter Radiation Model
  14. Noise Model and Projection Finalization
  15. CBCT Reconstruction
  16. ...and 11 more sections

Figures (4)

  • Figure 1: Simulated generation workflow. The fan-beam is deformed using a respiratory motion field to simulate breathing motion (c). Motion-adapted volumes are forward projected to create simulated X-ray projections (a), which are then reconstructed with simulated artifacts (b) to generate the final simulated .
  • Figure 2: Motion vector field generation (x and y component of 3D motion vector visualized as 2D vector overlay on a central axial slice of test patient volume): the sparse motion vector field (\ref{['fig:motion_field:MF_strong_border_gradients']}) encodes strong gradient vectors in border voxels, the dense motion vector field (\ref{['fig:motion_field:MF_propagated_gradients1']}) has a motion vector estimate for every voxel position, the updated motion vector field (\ref{['fig:motion_field:MF_propagated_gradients2']}) attenuates motion vectors that deviate from the desired main motion direction, and the final motion vector field (\ref{['fig:motion_field:MF_final']}) only keeps motion vectors in the extended proximity of the reference structure and eliminating motion vectors at voxels representing bone or voxels behind bones.
  • Figure 3: Qualitative comparison of geometric alignment between Msimulated (second row) and Mreal (third row) visualized using local heatmaps for SynthRAD test case 2PA020. Higher values (blue regions) indicate stronger local alignment with the source . Msimulated demonstrates superior alignment throughout the volume, particularly in anatomically complex regions.
  • Figure 4: Relationship between quantitative metrics and clinical visual assessment scores. Green and red markers show samples from SynthRAD and Clinical datasets, respectively. Circles indicate Mreal and crosses show Msimulated samples. Note the inverted - relationship for Clinical data (red markers) compared to SynthRAD (green markers).