MRI-Derived Synthetic CT: Methods & Applications

• MRI-derived synthetic CT (sCT) is a method that converts MRI data into CT-equivalent images using image regression and translation techniques.
• It employs various modeling approaches such as CNNs, GANs, and transformer-based diffusion, addressing challenges in registration, intensity normalization, and spatial standardization.
• sCT is pivotal for MRI-only radiotherapy, PET/MR attenuation correction, focused ultrasound planning, and pediatric imaging, reducing reliance on separate CT scans.

MRI-derived synthetic CT (sCT) denotes CT-equivalent images generated from MRI, typically with voxel values in Hounsfield Units, so that MRI can support tasks that ordinarily require CT-calibrated attenuation or electron-density information. In radiotherapy, the principal motivation is MRI-only treatment planning: MRI provides superior soft-tissue contrast, whereas CT is needed for dose calculation and related attenuation modeling; sCT attempts to unify these roles in a single modality space and thereby reduce dependence on separate CT acquisition and MRI–CT registration (Fan et al., 2024, Rogowski et al., 13 May 2026). Beyond external-beam radiotherapy, MRI-derived sCT has also been used for PET/MR attenuation correction, transcranial focused ultrasound planning, and pediatric cranial bone-and-suture analysis (Gholamiankhah et al., 2021, Liu et al., 2022, Iyer et al., 29 Dec 2025).

1. Clinical role and problem formulation

CT remains fundamental in radiotherapy because its Hounsfield Units are directly related to electron density and x-ray attenuation, which are required for dose calculation and, in some workflows, digitally reconstructed radiographs for patient positioning (Fu et al., 2018). MRI, by contrast, is preferred for target and organ-at-risk delineation because of its superior soft-tissue contrast, but its signal intensities are not directly related to attenuation coefficients or electron density (Fan et al., 2024, Gholamiankhah et al., 2021). MRI-derived sCT is therefore a cross-modality image-regression or image-translation problem in which MRI is mapped to a CT-like volume suitable for downstream quantitative use.

The clinical rationale is strongest where MRI is already central to the workflow. In MRI-only radiotherapy, sCT can eliminate separate planning CT, reduce radiation exposure, reduce cost, and circumvent MRI–CT registration errors (Fan et al., 2024). In male pelvic radiotherapy, CT–MR registration uncertainties were described as being in the $2$–$5$ mm range for various sites, motivating direct MR-only workflows (Fu et al., 2018). In MR-guided adaptive radiotherapy on MR-Linac systems, the same-day or on-table MRI provides the anatomy of the day, while sCT provides the CT-like density map required for dose recalculation and optimization (Asher et al., 2023). In low-field abdominal MR-only liver radiotherapy, the objective is similar: derive electron density from MRI without separate planning CT, even in the presence of abdominal variability and low-field image limitations (Fu et al., 2019).

The concept extends beyond radiotherapy. In transcranial focused ultrasound, CT is ordinarily required because skull HU are converted into acoustic properties such as density, speed of sound, and attenuation; MRI-derived sCT offers a CT-free route to skull modeling and phase correction (Liu et al., 2022, Gao et al., 11 Jul 2025). In pediatric cranial imaging, MRI-derived sCT has been used not only to emulate CT appearance but also to recover cranial bone segmentation and suture probability maps from routine T1-weighted MRI, addressing the problem that MRI poorly depicts bone and sutures while CT entails ionizing radiation (Iyer et al., 29 Dec 2025).

2. Data curation, alignment, and preprocessing

MRI-derived sCT is fundamentally dependent on paired or at least geometrically corresponding MRI–CT data. Several studies used retrospectively paired MRI and CT acquired in separate sessions and then aligned by rigid, affine, or deformable registration. In the male pelvis study, CT was registered to bias-corrected MRI by rigid registration, affine registration, and multi-resolution B-spline deformable registration, and the resulting deformed CT was resampled to the MR resolution (Fu et al., 2018). In brain MRI-to-CT synthesis, MR images were rigidly and non-rigidly registered to CT using Elastix, then resampled to the CT grid (Gholamiankhah et al., 2021). In challenge-derived datasets such as SynthRAD2023 and SynthRAD2025, pre-registration by the organizers was frequently assumed, but later work showed that the quality of this alignment can materially affect both training and evaluation (Fan et al., 2024, Boussot et al., 24 Oct 2025).

Intensity preprocessing is likewise central because MRI has no standardized physical intensity scale. Representative strategies include N4 bias-field correction, histogram-based normalization, z-score normalization, percentile clipping, and modality-specific scaling (Fu et al., 2018, Fu et al., 2019, Raggio et al., 2024). In one SwinUNETR-based brain-and-neck pipeline, MRI intensities were divided by $1000$, whereas CT was shifted by subtracting the per-volume minimum and then scaled by $1/2000$ (Fan et al., 2024). In the SynthRAD2025 flow-matching submission, MRI was z-score normalized per volume and clipped to $[-3,3]$, while CT was clipped to $[-1024,3071]$ HU and divided by $1000$ (Hadzic et al., 6 Oct 2025). Body or patient masks are commonly used to suppress background and to restrict losses and metrics to anatomically relevant voxels (Fu et al., 2018, Fan et al., 2024).

Spatial standardization varies by method. Some models use full volumes resampled to isotropic grids, such as $128 \times 128 \times 128$ at $1 \times 1 \times 1\ \text{mm}^3$ in the flow-matching approach (Hadzic et al., 6 Oct 2025), whereas others use patch-based training and inference because of GPU memory constraints. The choice between whole-volume and patch-based processing has implications for computational efficiency, boundary artifacts, and preservation of fine structure (Fan et al., 2024, Pan et al., 2023).

3. Methodological families

MRI-derived sCT methods now span deterministic CNN regression, adversarial image translation, transformer-based models, diffusion and flow-based generative models, federated learning, and task-specific hybrids.

A common baseline is voxel-wise regression with CNN encoder–decoder architectures. In the male pelvis, 2D and 3D CNNs were trained from scratch to map T1-weighted MRI directly to CT HU values with an L1 loss inside the body mask (Fu et al., 2018): $$\text{loss} = \frac{1}{N}\sum_{i=1}^{N} | \text{sCT}_i - \text{CT}_i |.$$ The 3D model exploited inter-slice context and improved MAE, bone DSC, and bone precision relative to the 2D model (Fu et al., 2018). A related line uses residual CNNs rather than adversarial training; in brain MRI-to-CT synthesis, a deep ResNet with dilated convolutions outperformed a GAN in whole-head MAE, PSNR, and SSIM (Gholamiankhah et al., 2021).

Adversarial methods remain prominent. Conditional GANs and CycleGANs have been used in pelvis, abdomen, brain, and MR-Linac settings (1802.06468, Fu et al., 2019, Asher et al., 2023). In the low-field abdominal liver study, the cGAN objective combined adversarial loss with an L1 penalty: $5$0 with $5$1 (Fu et al., 2019). Unpaired and weakly paired settings have also been explored with CycleGAN variants and DualGAN-based systems, including coordinate convolution and perceptual losses for difficult bone–air disambiguation in head MRI-to-CT translation (Prokopenko et al., 2019, Crespi et al., 2024).

More recent work emphasizes volumetric transformers and generative transport models. A 3D transformer-based denoising diffusion model conditioned on MRI used a Swin-VNet denoiser to iteratively transform Gaussian noise into CT, reporting strong brain and prostate results (Pan et al., 2023). Flow Matching has also been applied in a fully 3D conditional setting, learning a velocity field $5$2 and integrating the ODE

$5$3

to synthesize CT from MRI or CBCT (Hadzic et al., 6 Oct 2025). In pediatric cranial imaging, a MAISI-based variational autoencoder with adversarial and perceptual losses was coupled to an atlas-guided segmentation VAE, so that MRI yielded not only sCT but also seven cranial bone labels and suture probability heatmaps (Iyer et al., 29 Dec 2025).

Patch-based inference introduces a distinct methodological issue: stitching artifacts at patch boundaries. A SwinUNETR-based study addressed this with a 3D subvolume merging rule

$5$4

applied in overlapping subvolumes along all three spatial dimensions (Fan et al., 2024). This post-inference design reduced visible seams and improved MAE, illustrating that sCT quality can depend not only on network architecture but also on inference-time reconstruction.

4. Evaluation paradigms and reported performance

Reported performance is heterogeneous because anatomies, MR sequences, alignment quality, and endpoints differ. Typical image-space metrics include MAE in HU, PSNR, SSIM or MS-SSIM, and NCC; structure-aware studies additionally report Dice and HD95; radiotherapy-focused studies add gamma analysis and DVH comparisons; application-specific studies may instead evaluate acoustic metrics or segmentation equivalence (Fu et al., 2019, Liu et al., 2022, Rogowski et al., 13 May 2026). Representative results are summarized below.

Study	Setting	Key reported results
(Fu et al., 2018)	Male pelvis, 3D CNN from T1-weighted MRI	Whole-body MAE $5$5 HU; bone DSC $5$6
(Pan et al., 2023)	Brain and prostate, 3D transformer-based DDPM	Brain MAE 43.317 HU, PSNR 27.046 dB, SSIM 0.965, NCC 0.983; prostate MAE 59.953 HU
(Asher et al., 2023)	MR-Linac multi-site Cycle-GAN	MAE $5$7 HU; SSIM $5$8; PTV D95 median absolute difference 0.46 Gy
(Fu et al., 2019)	0.35T abdominal MR-only liver RT, cGAN	MAE $5$9 HU; gamma passing rates $1000$0 at 2%/2 mm
(Fan et al., 2024)	Brain and neck, SwinUNETR with 3D subvolume merging	MAE reduced from 52.65 HU to 47.75 HU
(Raggio et al., 2024)	Federated multicentre brain MRI-to-CT, unseen centre	Median MAE 102.0 HU; SSIM 0.89; PSNR 26.58
(Rogowski et al., 13 May 2026)	SynthRAD2025 Task 1 challenge report	Top MRI-to-CT MAE $1000$1 HU; PSNR $1000$2 dB; Dice 0.79; photon $1000$3

These values are not directly interchangeable benchmarks. The male pelvic CNN used standard T1-weighted MRI in 20 patients (Fu et al., 2018), whereas the diffusion model reported separate brain and prostate cohorts (Pan et al., 2023), the MR-Linac study evaluated deformed CT references near the PTV (Asher et al., 2023), and the SynthRAD2025 challenge aggregated five European centers across head and neck, thorax, and abdomen (Rogowski et al., 13 May 2026). A plausible implication is that anatomical site, input protocol, and reference alignment contribute as much to apparent performance as network class.

The challenge-scale picture is important. SynthRAD2025 reported that top MRI-to-CT performance reached MAE $1000$4 HU, PSNR $1000$5 dB, MS-SSIM $1000$6, Dice 0.79, photon $1000$7, and proton $1000$8 (Rogowski et al., 13 May 2026). The same report found strong image–segmentation correlations but only moderate dose correlations, confirming that image quality is insufficient as a dosimetric surrogate (Rogowski et al., 13 May 2026).

5. Downstream applications

The most established application is MRI-only radiotherapy. In general pelvis MR-only radiotherapy, a cGAN trained on prostate cancer patients generated sCT in 5.6 s on GPU and 21 s on CPU, and the average target dose increase on sCT was at most 0.3%, supporting feasibility across prostate, rectal, and cervical cancer (1802.06468). In low-field abdominal MR-only liver radiotherapy, cGAN-generated sCT achieved average gamma passing rates higher than 95% using a 2%, 2 mm criterion and 99% using a 3%, 3 mm criterion, while average differences in the mean dose and DVH metrics were within $1000$9 for the planning target volume and within $1/2000$0 for evaluated organs (Fu et al., 2019). On a 0.35T MR-Linac, Cycle-GAN sCTs produced PTV D85, D90, and D95 median absolute differences of 0.45, 0.47, and 0.46 Gy, with median OAR change above 33 Gy of 0.01 Gy (Asher et al., 2023).

The same principle applies in adaptive and repeated-imaging settings. MRI-derived sCT eliminates repeated CT scans and can be generated from on-table MRI, which is especially relevant for MR-guided adaptive radiotherapy (Asher et al., 2023, Hadzic et al., 6 Oct 2025). This suggests that deployment constraints are not only about final HU accuracy but also about runtime, patching strategy, and robustness to anatomical changes.

Outside radiotherapy, skull-focused applications are prominent. In transcranial focused ultrasound, a 3D patch-based cGAN from T1-weighted MRI yielded skull density ratio, skull thickness, and number of active elements with Pearson’s correlation coefficients of 0.94, 0.92, and 0.98 relative to real CT, and the distance between peak focal locations was less than 1.3 mm for all simulated cases (Liu et al., 2022). A later fully CT-free framework combined MRI-derived sCT with k-Wave, hybrid angular spectrum, and Rayleigh-Sommerfeld methods, reporting sub-millimeter targeting deviation, focal shape consistency with FWHM $1/2000$1–$1/2000$2 mm, and $1/2000$3 normalized pressure error compared to CT-based reference (Gao et al., 11 Jul 2025).

Pediatric cranial imaging represents a distinct endpoint: not dose, but bone and suture depiction. A dual-stage VAE framework generated pediatric cranial sCT from T1-weighted MRI and achieved SSIM $1/2000$4, PSNR $1/2000$5, mean Dice $1/2000$6 across seven bones plus sutures, and suture Dice $1/2000$7, with equivalence between CT- and sCT-based segmentation in all regions except the parietal bones under the reported TOST margins (Iyer et al., 29 Dec 2025).

6. Limitations, controversies, and future directions

A recurring limitation is that supervised MRI-to-CT synthesis inherits any registration error present in the training targets. This issue is explicit in the SynthRAD2025 registration study: IMPACT-based registration produced more accurate and anatomically consistent local alignments than mutual-information-based Elastix, improving local sCT synthesis, but public validation still favored Elastix-trained models because the evaluation pipeline itself was aligned with Elastix (Boussot et al., 24 Oct 2025). This directly supports the proposition that registration errors can propagate into supervised learning and potentially inflate performance metrics at the expense of anatomical fidelity (Boussot et al., 24 Oct 2025).

Another persistent problem is domain shift across MRI contrasts, scanners, and institutions. In brain MRI-to-CT synthesis, a baseline GAN trained without FLAIR performed worst on FLAIR with MAE $1/2000$8 HU; domain randomisation improved FLAIR performance to $1/2000$9 HU, but remained worse than training with FLAIR itself, which yielded $[-3,3]$0 HU (Nijskens et al., 2023). In multicentre federated brain MRI-to-CT synthesis, a cross-silo model achieved a median MAE of 102.0 HU on an unseen external centre, illustrating both the difficulty of external generalisation and the utility of federated collaboration under privacy constraints (Raggio et al., 2024).

Computational trade-offs remain unresolved. Patch-based models can suffer stitching artifacts unless overlap and merging are carefully designed (Fan et al., 2024). Whole-volume 3D generative models preserve volumetric context but often pay a resolution penalty: the fully 3D flow-matching model reported no artifacts but noticeably blurry outputs because training was constrained to $[-3,3]$1 volumes, and MRI-to-sCT validation performance reached MAE $[-3,3]$2 HU with Dice $[-3,3]$3 and HD95 $[-3,3]$4 mm (Hadzic et al., 6 Oct 2025). This suggests that full-volume 3D synthesis, patch-based high-resolution synthesis, and latent-space compression remain competing design choices rather than a settled progression.

A common simplification is to treat low MAE or high SSIM as sufficient evidence of clinical readiness. The challenge report explicitly argues otherwise: strong image–segmentation correlations but only moderate dose correlations showed that image quality is insufficient as a dosimetric surrogate, and residual errors at tissue interfaces propagated along beam paths, affecting proton dose more than photon dose (Rogowski et al., 13 May 2026). Thoracic and abdominal cases were more variable than head-and-neck, and MRI-to-CT remained distinctly harder than CBCT-to-CT (Rogowski et al., 13 May 2026).

Current future directions therefore cluster around four themes. First, better geometric supervision: registration methods such as IMPACT, or alternatives that reduce reliance on imperfect paired alignment (Boussot et al., 24 Oct 2025). Second, stronger robustness: domain randomisation, multi-sequence conditioning, and federated learning for multicentre generalisation (Nijskens et al., 2023, Raggio et al., 2024). Third, higher-fidelity generative modeling: transformer, diffusion, and flow-based methods that preserve global structure while recovering fine detail (Pan et al., 2023, Hadzic et al., 6 Oct 2025). Fourth, broader validation: not just image-space similarity, but structure-specific accuracy, dosimetric endpoints, and task-specific criteria such as acoustic focusing or cranial suture equivalence (Liu et al., 2022, Iyer et al., 29 Dec 2025).

MRI-derived sCT has thus evolved from a narrow image-translation problem into a modality-bridging infrastructure for MRI-only radiotherapy, adaptive replanning, CT-free focused ultrasound, and radiation-sparing pediatric assessment. The field’s current trajectory suggests that future progress will depend less on any single network family than on the coordinated treatment of registration, contrast generalisation, inference geometry, and endpoint-specific validation.