TotalSynth: Robust Whole-Body Synthetic CT from MRI and CBCT

Valentin Boussot; Cedric Hemon; Anais Barateau; Caroline Lafond; Jean-Claude Nunes; Jean-Louis Dillenseger

cs.CVSep 12, 2026

TotalSynth: Robust Whole-Body Synthetic CT from MRI and CBCT

Authors: Valentin Boussot, Cedric Hemon, Anais Barateau, Caroline Lafond, Jean-Claude Nunes, Jean-Louis Dillenseger

Organizations: Univ. Rennes, CLCC Eugène Marquis, INSERM, LTSI - UMR 1099 F-35000 Rennes, France

Abstract

Purpose: To develop and evaluate TotalSynth, a reusable pretrained framework for whole-body synthetic CT (sCT) generation from MRI and cone-beam CT (CBCT) images. Materials and Methods: In this retrospective technical study, the dataset was assembled between 2020 and 2026 from SynthRAD challenge data, four prostate cohorts, and BIC-MAC. After registration quality control, 1450 of 1800 public challenge pairs were retained; 350 were excluded for insufficient registration quality or major source/CT mismatch. The corpus also included 84 additional prostate MRI/CT and 60 external BIC-MAC MRI/CT cases. Three 5-fold model families were evaluated with image-domain, anatomy-aware, registration-based, and uncertainty metrics. Age and sex were not consistently available across public datasets. Results: The released MRI-to-CT model achieved an overall MAE of 67.49 HU, SSIM of 0.920, and PSNR of 29.28 dB. The CBCT-to-CT model achieved an overall MAE of 53.55 HU, SSIM of 0.939, and PSNR of 32.09 dB. The unified model maintained similar performance on MRI inputs (MAE, 67.68 HU) and CBCT inputs (MAE, 54.22 HU). On external BIC-MAC data, MRI-to-CT MAE was 100.91 HU without fine-tuning and 62.21 HU after fine-tuning. Conclusion: TotalSynth provides reusable MRI- and CBCT-based CT synthesis models with broad anatomical coverage, while external evaluation highlights the need for local validation and optional fine-tuning under domain shift.

Figures & tables

Figure 1: TotalSynth data coverage by source modality and anatomical region. MRI = magnetic resonance imaging; CT = computed tomography; CBCT = cone-beam CT; BIC-MAC = Big Cross-Modal Attenuation Correction. The bar chart summarizes registered SynthRAD2023/2025 cases counted from the public registration repositories, 60 BIC-MAC whole-body MRI cases counted for each anatomical region, and 84 additional prostate MRI/CT cases. Representative aligned CT slices are shown with segmentation-mask overlays to illustrate anatomical coverage and registration-derived supervision.

Dataset	Modality pair	Region	Source cases	Retained cases	Excluded cases	Use
SynthRAD2023	MRI+CBCT	Brain, Pelvis	720	618	102	Training/evaluation
SynthRAD2025	MRI+CBCT	Head & Neck, Thorax, Abdomen	1080	832	248	Training/evaluation
BIC-MAC	MRI(Dixon)/CT	Whole-body	75	60	15	Fine-tuning/OoD
D1	MRI/CT	Prostate	24	24	0	Training/evaluation
D2	MRI/CT	Prostate	18	18	0	Training/evaluation
GoldAtlas (D3)	MRI/CT	Prostate	19	17	2	Training/evaluation

Table 1: Summary of datasets used in this study after eligibility screening and registration quality control. MRI = magnetic resonance imaging; CT = computed tomography; CBCT = cone-beam CT; OoD = out-of-distribution.

Figure 2: Overview of the TotalSynth training and inference pipeline. MRI = magnetic resonance imaging; CT = computed tomography; CBCT = cone-beam CT. MRI or CBCT images are preprocessed and used as input to a 2.5D synthesis model that predicts a CT-like image. During training, reference CT images are spatially aligned to the input images and used as supervised targets. During inference, predictions are generated using test-time augmentation and 5-fold model ensembling, then restored to the original image geometry.

Figure 3: Representative CT synthesis examples across source modalities and anatomical regions. MRI = magnetic resonance imaging; CT = computed tomography; CBCT = cone-beam CT; sCT = synthetic CT; HU = Hounsfield unit. For each region, the source MRI or CBCT image, generated sCT, registered CT target, and absolute HU error map are shown. These examples were generated with a single ensemble member and no test-time augmentation for visual illustration; quantitative tables report the released model evaluation.

Model and setting	Evaluation domain	MAE (HU)	SSIM	PSNR (dB)
MRI-to-CT	Primary MRI/CT	$67.49\,\pm\,15.97$	$0.920\,\pm\,0.025$	$29.28\,\pm\,2.10$
CBCT-to-CT	Primary CBCT/CT	$53.55\,\pm\,17.79$	$0.939\,\pm\,0.026$	$32.09\,\pm\,2.93$
Unified model	MRI inputs	$67.68\,\pm\,16.42$	$0.920\,\pm\,0.025$	$29.22\,\pm\,2.18$
Unified model	CBCT inputs	$54.22\,\pm\,17.54$	$0.938\,\pm\,0.025$	$31.89\,\pm\,2.87$
MRI-to-CT, direct	BIC-MAC	$100.91\,\pm\,18.67$	$0.968\,\pm\,0.007$	$26.69\,\pm\,1.40$
MRI-to-CT, fine-tuned	BIC-MAC	$62.21\,\pm\,6.19$	$0.982\,\pm\,0.002$	$30.45\,\pm\,0.92$

Table 2: Overall performance of released TotalSynth models and BIC-MAC fine-tuning. MRI = magnetic resonance imaging; CT = computed tomography; CBCT = cone-beam CT; BIC-MAC = Big Cross-Modal Attenuation Correction; MAE = mean absolute error; HU = Hounsfield unit; SSIM = structural similarity index; PSNR = peak signal-to-noise ratio. Values report mean

\pm

standard deviation. Detailed regional results are provided in the Appendix.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Region	MAE	PSNR	SSIM	Dice	SAM	Reg	Uncertainty
AB ( $n=22$ )	$60.10\,\pm\,9.42$	$30.01\,\pm\,1.47$	$0.908\,\pm\,0.022$	$0.598\,\pm\,0.070$	$27.405\,\pm\,4.156$	$0.511\,\pm\,0.117$	$1543.81\,\pm\,580.52$
HN ( $n=21$ )	$80.89\,\pm\,13.12$	$28.44\,\pm\,1.35$	$0.925\,\pm\,0.030$	$0.645\,\pm\,0.106$	$13.412\,\pm\,2.108$	$0.435\,\pm\,0.066$	$1308.83\,\pm\,512.26$
TH ( $n=23$ )	$58.45\,\pm\,11.20$	$31.21\,\pm\,1.86$	$0.940\,\pm\,0.019$	$0.645\,\pm\,0.051$	$21.638\,\pm\,4.232$	$0.466\,\pm\,0.089$	$751.61\,\pm\,262.12$
brain ( $n=27$ )	$79.71\,\pm\,8.64$	$27.44\,\pm\,0.92$	$0.915\,\pm\,0.015$	$0.751\,\pm\,0.096$	$13.209\,\pm\,2.612$	$1.199\,\pm\,0.138$	$3402.08\,\pm\,633.51$
pelvis ( $n=23$ )	$48.16\,\pm\,8.30$	$31.28\,\pm\,1.42$	$0.901\,\pm\,0.025$	$0.756\,\pm\,0.108$	$25.949\,\pm\,3.617$	$0.574\,\pm\,0.094$	$2348.75\,\pm\,648.73$
CSIRO ( $n=24$ )	$75.99\,\pm\,10.13$	$27.67\,\pm\,1.02$	$0.929\,\pm\,0.016$	$0.774\,\pm\,0.035$	$30.898\,\pm\,3.947$	$0.434\,\pm\,0.054$	$1272.28\,\pm\,153.55$

Appendix

Table 3: Evaluation of the released MRI-to-CT model by anatomical region. Values report mean

\pm

standard deviation. Case counts refer to the image-domain and Dice evaluation.

Region	MAE	PSNR	SSIM	Dice	SAM	Reg	Uncertainty
AB ( $n=32$ )	$57.18\,\pm\,14.96$	$31.12\,\pm\,2.42$	$0.926\,\pm\,0.023$	$0.673\,\pm\,0.086$	$17.706\,\pm\,4.475$	$1.049\,\pm\,0.016$	$730.95\,\pm\,198.42$
HN ( $n=37$ )	$67.53\,\pm\,15.80$	$29.97\,\pm\,2.19$	$0.946\,\pm\,0.021$	$0.717\,\pm\,0.075$	$8.853\,\pm\,1.528$	$1.052\,\pm\,0.036$	$661.14\,\pm\,231.48$
TH ( $n=34$ )	$57.27\,\pm\,15.33$	$32.03\,\pm\,2.69$	$0.928\,\pm\,0.019$	$0.724\,\pm\,0.075$	$16.241\,\pm\,4.455$	$1.045\,\pm\,0.021$	$793.57\,\pm\,252.09$
brain ( $n=26$ )	$39.42\,\pm\,6.13$	$33.96\,\pm\,1.56$	$0.970\,\pm\,0.007$	$0.905\,\pm\,0.071$	$7.219\,\pm\,1.024$	$1.013\,\pm\,0.002$	$879.35\,\pm\,256.72$
pelvis ( $n=21$ )	$34.84\,\pm\,8.32$	$35.12\,\pm\,2.49$	$0.924\,\pm\,0.023$	$0.835\,\pm\,0.078$	$20.066\,\pm\,4.641$	$1.023\,\pm\,0.005$	$949.44\,\pm\,381.04$
overall ( $n=150$ )	$53.55\,\pm\,17.79$	$32.09\,\pm\,2.93$	$0.939\,\pm\,0.026$	$0.758\,\pm\,0.113$	$13.703\,\pm\,6.015$	$1.039\,\pm\,0.027$	$784.24\,\pm\,278.43$

Appendix

Table 4: Evaluation of the released CBCT-to-CT model by anatomical region. Values report mean

\pm

standard deviation.

Region	MAE	PSNR	SSIM	Dice	SAM	Reg	Uncertainty
MRI-to-CT/AB	$59.64\,\pm\,9.56$	$30.04\,\pm\,1.48$	$0.909\,\pm\,0.023$	$0.591\,\pm\,0.066$	$27.362\,\pm\,4.014$	$0.526\,\pm\,0.140$	$1451.52\,\pm\,614.79$
MRI-to-CT/HN	$80.34\,\pm\,12.50$	$28.39\,\pm\,1.34$	$0.926\,\pm\,0.029$	$0.641\,\pm\,0.106$	$13.378\,\pm\,2.122$	$0.423\,\pm\,0.080$	$1181.19\,\pm\,467.90$
MRI-to-CT/TH	$58.00\,\pm\,11.26$	$31.19\,\pm\,1.88$	$0.940\,\pm\,0.019$	$0.635\,\pm\,0.054$	$21.851\,\pm\,4.555$	$0.436\,\pm\,0.087$	$673.25\,\pm\,226.95$
MRI-to-CT/brain	$80.76\,\pm\,9.21$	$27.26\,\pm\,0.99$	$0.913\,\pm\,0.016$	$0.763\,\pm\,0.108$	$13.355\,\pm\,2.870$	$1.185\,\pm\,0.132$	$2897.51\,\pm\,530.13$
MRI-to-CT/pelvis	$47.43\,\pm\,8.03$	$31.35\,\pm\,1.49$	$0.902\,\pm\,0.025$	$0.771\,\pm\,0.102$	$27.045\,\pm\,3.892$	$0.579\,\pm\,0.096$	$2171.19\,\pm\,487.25$
MRI-to-CT/CSIRO	$76.98\,\pm\,10.54$	$27.55\,\pm\,1.05$	$0.928\,\pm\,0.016$	$0.775\,\pm\,0.038$	$31.038\,\pm\,4.174$	$0.274\,\pm\,0.060$	$1303.97\,\pm\,165.31$

Appendix

Table 5: Evaluation of the released unified MRI/CBCT-to-CT model by input domain and anatomical region. Values report mean

\pm

standard deviation.

Setting	Contrast	MAE	PSNR	SSIM	Dice	SAM	Reg	Uncertainty
Direct	IN	$85.27\,\pm\,6.15$	$27.89\,\pm\,0.61$	$0.973\,\pm\,0.003$	$0.568\,\pm\,0.046$	$18.991\,\pm\,1.608$	$0.180\,\pm\,0.028$	$917.41\,\pm\,115.51$
Direct	OUT	$115.50\,\pm\,14.04$	$25.56\,\pm\,0.91$	$0.962\,\pm\,0.004$	$0.350\,\pm\,0.070$	$20.804\,\pm\,1.503$	$0.204\,\pm\,0.027$	$2132.74\,\pm\,263.72$
Direct	Overall	$100.91\,\pm\,18.67$	$26.69\,\pm\,1.40$	$0.968\,\pm\,0.007$	$0.455\,\pm\,0.124$	$19.929\,\pm\,1.799$	$0.193\,\pm\,0.030$	$1546.03\,\pm\,641.28$
Fine-tuned	IN	$60.00\,\pm\,5.43$	$30.73\,\pm\,0.87$	$0.983\,\pm\,0.002$	$0.662\,\pm\,0.025$	$13.113\,\pm\,1.069$	$0.198\,\pm\,0.028$	$272.78\,\pm\,33.12$
Fine-tuned	OUT	$64.43\,\pm\,6.12$	$30.16\,\pm\,0.87$	$0.981\,\pm\,0.002$	$0.648\,\pm\,0.026$	$13.354\,\pm\,1.063$	$0.209\,\pm\,0.029$	$332.65\,\pm\,40.43$
Fine-tuned	Overall	$62.21\,\pm\,6.19$	$30.45\,\pm\,0.92$	$0.982\,\pm\,0.002$	$0.655\,\pm\,0.026$	$13.234\,\pm\,1.073$	$0.203\,\pm\,0.029$	$302.72\,\pm\,47.56$

Appendix

Table 6: BIC-MAC external evaluation before and after fine-tuning. Values report mean

\pm

standard deviation.

Explore similar work

Generating synthetic computed tomography for radiotherapy: SynthRAD2025 challenge report
May 13, 2026Viktor Rogowski, Maarten L. Terpstra, Niklas Wahl +30Cone-Beam CT Medical Image Synthesis
Radiation Physics, Department of Hematology, Oncology, and Radiation Physics, Skåne University Hospital, Lund, Sweden · Medical Radiation Physics, Department of Clinical Sciences Lund, Lund University, Lund, Sweden · Radiotherapy Department, University Medical Center Utrecht, Utrecht, The Netherlands +20
A Proof-of-Concept Study of Multitask Learning for Cranial Synthetic CT Generation Across Heterogeneous MRI Field Strengths
Apr 30, 2026Zhuoyao Xin, Yiren Zhang, Christopher Wu +6Domain Generalization Multi-Task Learning
F.M. Kirby Research Center for Brain Imaging, Kennedy Krieger Institute, Baltimore, Maryland, 21205, United States. · Neurosection, Division of MR Research, Russell H. Morgan Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, Baltimore, Maryland, 21287, United States. · Department of Biomedical Engineering, Johns Hopkins University, Baltimore, Maryland, 21218, United States. +6
MRI-to-CT synthesis using drifting models
Mar 30, 2026Qing Lyu, Jianxu Wang, Jeremy Hudson +2Medical Image Synthesis Drifting Models
Department of Radiology and Biomedical Imaging, Yale School of Medicine, New Haven, CT · Department of Biomedical Engineering, Rensselaer Polytechnic Institute, Troy, NY

May 13, 2026J/K move · Enter open · S save

Generating synthetic computed tomography for radiotherapy: SynthRAD2025 challenge report

Viktor Rogowski, Maarten L. Terpstra, Niklas Wahl, Florian Kamp+29

Radiation Physics, Department of Hematology, Oncology, and Radiation Physics, Skåne University Hospital, Lund, Sweden · Medical Radiation Physics, Department of Clinical Sciences Lund, Lund University, Lund, Sweden · Radiotherapy Department, University Medical Center Utrecht, Utrecht, The Netherlands+20

Cone-Beam CT Medical Image Synthesis Medical Image Benchmarks

Open paper AskPDF arXiv

Radiation therapy (RT) requires precise dose delivery over multiple fractions, with CT fundamental for treatment planning due to its electron density information. Repeated CT acquisitions impose radiation exposure and logistical burdens, MRI lacks electron density, and cone-beam CT (CBCT) requires correction for dose calculation. Synthetic CT (sCT) generation addresses these by converting MRI or CBCT into CT-equivalent images with accurate Hounsfield Unit (HU) values, enabling MRI-only RT and CBCT-based adaptive workflows. Building on SynthRAD2023, SynthRAD2025 benchmarked sCT methods on 2,362 patients from five European centers across head and neck, thorax, and abdomen. Two tasks: MRI-to-CT (890 cases) and CBCT-to-CT (1,472 cases), evaluated via image similarity (MAE, PSNR, MS-SSIM), segmentation (Dice, HD95), and dosimetric metrics from photon and proton plans. With 803 participants and 12/13 valid submissions, Task 1 top performance reached MAE

64.8\pm21.3

HU, PSNR

\sim

30 dB, MS-SSIM

\sim

0.936, Dice 0.79, photon

γ_{2\%/2\text{mm}}>98\%

, proton

γ\approx85\%

. Task 2 improved: MAE

48.3\pm13.4

HU, PSNR 32.6 dB, MS-SSIM 0.968, Dice 0.86, photon

γ>99\%

, proton

γ\approx89\%

. Strong image--segmentation correlations (

ρ=0.78

0.79

) but moderate dose correlations confirmed image quality is insufficient as a dosimetric surrogate. Head-and-neck cases were most consistent; thoracic and abdominal cases showed greater variability. Residual errors at tissue interfaces propagate along beam paths, affecting proton dose more than photon. SynthRAD2025 demonstrates that deep learning yields clinically relevant sCTs, especially for CBCT-to-CT, while identifying persistent MRI-to-CT challenges and underscoring dose-based evaluation as essential for clinical validation.

Abstract

Explore similar work

Generating synthetic computed tomography for radiotherapy: SynthRAD2025 challenge report

A Proof-of-Concept Study of Multitask Learning for Cranial Synthetic CT Generation Across Heterogeneous MRI Field Strengths

MRI-to-CT synthesis using drifting models