CT image generation from ultrashort echo time MRI for dosimetric assessment in lung cancer radiotherapy

Abstract

Background and purpose

To facilitate magnetic resonance imaging (MRI)-only radiotherapy workflows and overcome the lack of electron density information in MRI for accurate radiotherapy dose calculation, we proposed a cycle-consistent generative adversarial network-based framework, Fast Fourier Transform Spatial-Channel Net (FFTSC-Net). This study aimed to synthesize high-quality synthetic computed tomography (sCT) from lung ultrashort echo time-MRI (UTE-MRI) and assess its potential clinical dosimetric viability.

Materials and methods

UTE-MRI and CT data from 60 patients with lung cancer were retrospectively collected and spatially aligned through deformable registration. FFTSC-Net was developed by integrating spatial and channel reconstruction convolution, squeeze-and-excitation, and residual fast Fourier transform modules, alongside a dual contrastive loss function. The synthesis quality was evaluated using structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and mean absolute error (MAE). The clinical feasibility was assessed through Hounsfield unit (HU) consistency analysis in organs at risk (OARs) and planning target volume (PTV), dose volume histogram parameters, and gamma passing rates.

Results

FFTSC-Net achieved an SSIM of 0.80, PSNR of 20.07 dB, and MAE (HU) of 59.31. No statistically significant differences were observed in HU distributions for lungs, spinal cord, and PTV compared with planning CT (p > 0.05). Dosimetric analysis revealed no significant differences across all PTV and OAR indices (p > 0.05). Gamma passing rates reached 95.39% (2%/2 mm) and 98.08% (3%/3 mm).

Conclusion

FFTSC-Net demonstrates high image fidelity and dosimetric accuracy, supporting its potential for MRI-only radiotherapy planning in patients with lung cancer.

Graphical abstract

Highlights

• •FFTSC-Net produces synthetic computed tomography from lung ultrashort echo time magnetic resonance imaging for dosimetric validation.
• •FFTSC-Net outperforms competing networks with a structural similarity index of 0.80, peak signal-to-noise ratio of 20.07 dB, and mean absolute error of 59.31 HU.
• •The generated synthetic computed tomography presents stable HU values in the planning target volume and organs at risk.
• •Mean relative dose differences between synthetic computed tomography and planning computed tomography were 0.01%–0.59%; standard deviations were 0.39%–3.44%.
• •Gamma passing rates all exceed 95% under three standard evaluation criteria.

1. Introduction

Magnetic resonance imaging (MRI)-guided radiotherapy has emerged as a promising approach for the treatment of thoracic malignancies [1]. In particular, MRI-linear accelerator (MRI-Linac) systems offer superior soft-tissue visualization and real-time motion tracking, making online adaptive radiotherapy feasible for lung cancer patients [2], [3]. Despite these advantages, the implementation of an MRI-only radiotherapy workflow remains challenging. The main obstacle is that treatment planning systems (TPS) require electron density (ED) information from computed tomography (CT) for dose calculation, whereas MRI primarily provides information related to proton density and tissue relaxation properties (T1 and T2) [4]. Consequently, MRI cannot be directly used for radiotherapy dose calculation.

To bridge this gap, synthesizing pseudo-CT or synthetic CT (sCT) directly from MRI has become important for MRI-only radiotherapy workflows. For thoracic imaging, ultrashort echo time (UTE) MRI is particularly advantageous because it can capture rapidly decaying signals from the lung parenchyma with non-Cartesian k-space trajectories [5], [6], [7], [8], [9]. Notably, UTE itself cannot suppress motion artifacts; central k-space sampling provides self-navigator signals to sort k-space data into respiratory bins, and motion-compensated reconstruction further reduces respiratory artifacts [10], [11]. However, mapping UTE-MRI (proton density) to CT (ED) is non-linear, particularly across heterogeneous regions encompassing low-density lungs, dense ribs, and mobile cardiac borders. Deep learning methods, including U-Net [12] and generative adversarial networks (GANs) [13], have shown promising performance in sCT generation [14], [15], [16], but their application to thoracic imaging often results in blurred anatomical boundaries and inaccurate Hounsfield Unit (HU) estimation [17]. Furthermore, state-of-the-art supervised models (for example, conditional GANs) rely on paired datasets [18]. Nevertheless, cycle-consistent GANs (CycleGANs) [19] offer a compelling solution by enabling image-to-image translation on unpaired data. However, standard CycleGANs struggle to accurately characterize complex lung textures and high-frequency structural boundaries under unpaired constraints. To address these limitations, we developed an enhanced cycle-consistent framework, termed FFTSC-Net, for thoracic UTE-MRI-to-CT synthesis, incorporating spatial and channel reconstruction convolution (SCConv) [20], squeeze-and-excitation (SE) attention [21], residual fast Fourier transform (FFT) modules [22], and a dual contrastive loss [23] to improve anatomical feature representation, preserve high-frequency structural boundaries, and enhance voxel-wise HU fidelity.

Therefore, this study aimed (1) to synthesize high-fidelity sCT images from UTE-MRI to support MRI-only thoracic radiotherapy and (2) to evaluate their clinical feasibility through dosimetric consistency validation for lung cancer treatment planning.

2. Materials and methods

2.1. Data acquisition

This study was approved by the Ethics Committee of Hubei Cancer Hospital (No. LLHBCH2026YN-050). Written informed consent was obtained from all participants. A total of 60 patients (32 males and 28 females) with clinically confirmed lung cancer were enrolled in this study (mean age ± standard deviation (SD): 61 ± 10 years). Tumors were predominantly located in the right lung (33 cases), whereas 27 cases involved the left lung. All patients underwent planning CT (pCT) scanning using a GE Revolution EVO CT scanner (General Electric, Waukesha, WI, USA) with a tube voltage of 120 kVp. The acquired CT images had a matrix size of 512 × 512 × 55 and a slice thickness of 5 mm. On the same day, lung UTE-MRI data were obtained using the United Imaging uMR 790 3.0 T MRI system (United Imaging Healthcare, Shanghai, China) with the UTE sequence.

The UTE-MRI protocol used a respiratory-gated 3D radial sequence with a TR/TE of 2/0.16 ms and a slice thickness of 1.15 mm. Data acquisition lasted 4–5 min depending on individual respiratory patterns. Respiratory motion was reduced using a navigator sequence that monitored vertical diaphragmatic displacement. Images were reconstructed by re-gridding radial k-space data onto a Cartesian grid using a Kaiser-Bessel kernel, followed by three-dimensional FFT [24]. This non-iterative regridding pipeline carries low computational overhead, enabling on-scanner online reconstruction without high-memory GPU demands. Detailed UTE-MRI acquisition parameters are provided in the Supplementary Material A.

2.2. Data preprocessing

Intensity-based 3D registration was performed using 3D Slicer software to achieve precise spatial alignment [25], [26]. With the UTE-MRI volume serving as the fixed (reference) image and the pCT as the moving image, a two-step registration pipeline was employed. Initially, global rigid and affine transformations were applied using the BRAINSFit module. A non-linear deformable registration was performed using the Elastix module, based on B-spline transformations, to account for local anatomical variations. Both modules are powered by the underlying Insight Toolkit framework [27], [28]. The optimized deformation field was used to resample and interpolate the moving CT image (5 mm slice thickness) into the high-resolution UTE image space (1.15 mm slice thickness), to improve voxel-wise anatomical correspondence for model evaluation [29].

After registration, the 60 samples were randomly divided into training, validation, and testing sets at a ratio of 6:2:2. All competing models were trained on this identical dataset split and processed with the same preprocessing pipeline. External contour masks were generated to remove non-anatomical structures, such as the scanning bed and clothing, and to constrain the synthesis region. Normalization was applied to scale the grayscale values of UTE and CT images to a numerical range suitable for the deep learning model, thereby optimizing network training efficiency.

2.3. Network architecture

FFTSC-Net was developed based on a CycleGAN framework, with its overall architecture shown in Fig. 1. The proposed model builds upon the U-Net architecture [30], with several modifications introduced for UTE-MRI-to-CT synthesis (Fig. S1). An SCConv module was integrated into the first encoder layer to enhance low-level feature extraction and reduce feature redundancy. Transformer encoders combined with SE attention were used to improve feature recalibration during cross-modality translation. Furthermore, an FFT module was embedded in the skip connections to strengthen spatial–frequency feature representation, and DC loss was introduced to constrain the synthesized sCT toward the pCT domain. These components were designed to improve anatomical reconstruction and HU fidelity, particularly in high-frequency structures such as bone.

Fig. 1.

Schematic diagram of FFTSC-Net structure.

2.4. Comparison between different models

To validate the proposed approach, FFTSC-Net was compared with three representative image-synthesis models using the same test set: Cytran [31], Pix2Pix [32], and Utrans [33]. Cytran contains two generators and discriminators and uses cycle-consistency loss for unsupervised cross-modality translation to generate high-quality images [31]. Pix2Pix, a conditional GAN model, produces images conditioned on input images to realize cross-modality mapping [32]. Constructed based on the U-Net architecture, Utrans integrates generative adversarial learning for cross-modality image generation. By leveraging an encoder–decoder structure with skip connections, the model preserves fine structural details and maintains semantic consistency in high-resolution synthesized images [33].

2.5. Training parameter setting

The experiments were conducted on a workstation equipped with an Intel (R) Xeon (R) Silver 4210R CPU @ 2.40 GHz, 64.0 GB of RAM, and an NVIDIA GeForce RTX 3090 GPU (24 GB VRAM), running on the CentOS operating system. The model was developed using Python 3.7.16 and PyTorch 1.13.0. During the training phase, the initial learning rates for the generator and discriminator were set to 0.0002 and 0.0001, respectively. To balance the adversarial training, the generator was updated twice for every single update of the discriminator. A linear decay strategy was applied over a total of 150 epochs: The learning rates remained constant for the first 80 epochs and decayed linearly to zero over the remaining 70 epochs. The batch size was fixed at 8. Network training time was 45 h.

2.6. Image analysis

To quantitatively evaluate the accuracy and fidelity of the synthesized images, three metrics were employed: Peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and mean absolute error (MAE). The definitions and mathematical formulations for these metrics are as follows:

$$\mathit{SSIM}\left(x,y\right)=\frac{\left(2{\mu }_x{\mu }_y+C_1\right)\left(2{\sigma }_{\mathit{xy}}+C_2\right)}{\left({\mu }_x^2+{\mu }_y^2+C_1\right)\left({\sigma }_x^2+{\sigma }_y^2+C_2\right)}$$ (1)

where $\mu$represents the mean,${\sigma }^2$denotes the variance, ${\sigma }_{\mathit{xy}}$signifies the covariance, and$C_1$ and $C_2$ are constants that maintain stability.

$$\mathit{PSNR}=10·{\mathit{log}}_{10}\left(\frac{\mathit{MA}{X}_I^2}{\mathit{MSE}}\right)$$ (2)

Here, $\mathit{MA}{X}_I$ represents the maximum possible pixel value of an image.

$$\mathit{MAE}=\frac{1}{N}\sum _{i=1}^N\left|x_i-y_i\right|$$ (3)

where $N$ is the total number of pixels, and $x_i$ and $y_i$ represent the pixel values at position $i$ in the generated image and the standard image, respectively.

2.7. Treatment planning parameters

All treatment plans were designed and optimized using the Monaco TPS (Version 6.0.11, Elekta AB, Stockholm, Sweden). For each patient, intensity-modulated radiation therapy plans were developed using 6 MV photon beams from an Elekta linear accelerator equipped with an Agility 160-leaf multileaf collimator. Dose calculation was performed using the Monte Carlo algorithm with a 0.3 cm isotropic grid spacing and a statistical uncertainty of 1.0% per calculation. The prescribed dose was 60 Gy delivered in 30 fractions (2 Gy per fraction).

2.8. Dose calculation

The dosimetric impact of sCT was assessed through a plan-transfer study. Anatomical structures, including the planning target volume (PTV) and organs at risk (OARs: Heart, spinal cord, and bilateral lungs), were delineated on the pCT and mapped to the sCT via the registration. A clinical treatment plan was optimized and calculated on the pCT to satisfy standard dosimetric objectives for lung cancer. The optimized beam arrangements, fluences, and monitor units from the pCT plan were transferred to the corresponding sCT images. The dose was recalculated on the sCT to evaluate the consistency of dose distribution.

2.9. Dosimetric evaluation

Dosimetric agreement between sCT- and pCT-based plans was evaluated using dose volume histogram (DVH) parameters and gamma analysis. For target volume evaluation, PTV D₉₅, D₉₈, and D_{2 cm}³ were calculated to evaluate dose coverage within the target, where D_x denotes the dose covering x% of the PTV, and D_{2 cm}³ is defined as the minimum dose received by the most irradiated 2 cm³ of the target volume. Lung dose was evaluated using V₂₀, V₁₀ and V₅, with dose constraints of V₂₀ < 25%, V₁₀ < 40%, and V₅ < 50%. Heart dose was evaluated using V₃₀ and mean dose (D_mean), with constraints of V₃₀ < 40% and D_mean < 20 Gy. Spinal cord dose was assessed using the maximal dose (D_max), which was less than 45 Gy. In these metrics, V_x denotes the percentage volume of OARs receiving a dose greater than x Gy. The dose constraints were based on the Quantitative Analysis of Normal Tissue Effects in the Clinic (QUANTEC) guidelines [34]. 2D gamma analysis was performed to quantify global dose distribution consistency, using criteria of 2%/2 mm, 3%/2 mm, and 3%/3 mm with a 10% dose threshold.

2.10. Statistical analysis

In this study, all data were processed using SPSS 22.0 software. Continuous variables are expressed as mean ± SD. A paired t-test was performed to evaluate differences in HU values of OARs and the PTV between pCT and sCT, and to assess differences in DVH metrics between sCT and pCT-based plans. All statistical analyses were conducted using two-sided tests, and p < 0.05 indicated statistically significant differences between the compared groups.

3. Results

Table 1 summarizes the quantitative evaluation results of all sCT images in the test set. FFTSC-Net achieved an SSIM of 0.80, PSNR of 20.07 dB, and MAE of 59.31 HU. Compared with Cytran, Pix2Pix, and Utrans on the same test set, FFTSC-Net showed the best performance across all three metrics.

Table 1.

SSIM, PSNR, and MAE (HU) of images generated by different models.

Model	SSIM	PSNR	MAE(HU)
Cytran	0.72	16.58	110.65
Pix2Pix	0.73	14.86	120.75
Utrans	0.75	16.57	95.98
FFTSC-Net	0.80	20.07	59.31

Abbreviations: SSIM, Structural Similarity Index; PSNR, Peak Signal-to-Noise Ratio; MAE (HU), Mean Absolute Error (Hounsfield unit).

Fig. 2 shows representative sCT images generated by different models and their corresponding MAE (HU) maps. Compared with the other methods, FFTSC-Net better preserved anatomical structures, including the spine and lesions. The MAE maps of the comparison networks showed broad red/yellow regions around complex anatomical boundaries (Figs. 2C, D, and E), whereas the MAE map of FFTSC-Net showed fewer red/yellow regions and a more uniform low-error background (Fig. 2F), indicating improved consistency with the pCT.

Fig. 2.

Transverse slices of UTE-MRI, pCT, and sCT obtained from different networks, (A)pCT; (B)UTE-MRI; (C)Cytran; (D)Pix2Pix; (E)Utrans; (F)FFTSC-Net.

Table 2 summarizes the mean HU values of all OARs and the PTV in the test set. No significant differences were observed between sCT and pCT in the bilateral lungs, spinal cord, or PTV (all p > 0.05), indicating comparable tissue density representation in these regions. In contrast, the heart showed a significant HU difference between sCT and pCT (p < 0.05), suggesting relatively reduced synthesis accuracy in this region.

Table 2.

The HU difference between pCT and sCT was evaluated for different OARs and the PTV in the 12 test cases.

Pt#	R Lung		L Lung		Spinal Cord		Heart		PTV
	pCT	sCT	pCT	sCT	pCT	sCT	pCT	sCT	pCT	sCT
Pt1	−726.31	−735.43	−748.26	−745.09	28.84	17.61	38.17	34.45	−101	−81
Pt2	−698.89	−715.64	−693.01	−745.05	24.88	5.85	38.26	28.94	−353	−343
Pt3	−760.89	−763.99	−753.04	−768.68	30.38	3.74	36.81	37.65	−394	−399
Pt4	−761.03	−695.92	−813.85	−743.41	25.05	27.71	32.34	29.38	−261	−194
Pt5	−777.69	−724.24	−730.43	−727.28	19.66	15.29	36.47	25.85	−336	−328
Pt6	−574.20	−610.01	−696.76	−729.24	24.37	11.96	34.84	27.74	−79	−77
Pt7	−725.84	−739.31	−753.30	−767.39	16.77	20.11	36.95	30.61	−513	−535
Pt8	−831.28	−826.92	−848.47	−811.06	17.39	29.85	38.33	21.33	−411	−409
Pt9	−687.23	−686.74	−799.72	−760.90	20.22	9.96	34.97	34.80	−141	−117
Pt10	−818.63	−752.66	−788.72	−700.95	25.54	28.96	32.11	30.80	−315	−299
Pt11	−873.56	−862.69	−839.16	−803.86	25.34	34.15	33.71	21.96	−203	−201
Pt12	−824.88	−746.56	−802.89	−702.96	36.28	10.31	34.97	32.64	−401	−350
p value	0.45		0.33		0.08		<0.05		0.06

Abbreviations: R Lung, (Right Lung); L Lung, (Left Lung); PTV, Planning Target Volume.

Table 3 summarizes the dosimetric differences between sCT and pCT-based dose calculations. Across all treatment plans, the mean relative differences for lung V₂₀, V₁₀, and V₅ were − 0.07% ± 0.51%, −0.45% ± 0.68%, and 0.43% ± 1.07%, respectively. For the heart, the mean differences in V₃₀ and D_mean were 0.27% ± 3.44% and 0.01% ± 0.50%, respectively, while the mean difference in spinal cord D_max was −0.59% ± 1.84%. For PTV dose evaluation, the mean relative differences in D₉₅, D₉₈, and $D_{2{\mathrm{cm}}^3}$were − 0.20% ± 0.39%, −0.05% ± 0.42%, and 0.02% ± 0.69%, respectively. No significant differences were observed between sCT and pCT across all dosimetric parameters (all p > 0.05). Fig. S2 demonstrates the 2D dose distribution cross-sections at the target level for two randomly selected patients. The isodose distributions on the sCT are similar to those on the pCT. Fig. 3 displays DVH plots for two patients, where the DVH curves for the PTV and OARs are similar for sCT and pCT.

Table 3.

Comparison of dosimetric parameters between pCT and sCT for different OARs and the PTV in the 12 test cases.

Parameter	pCT	sCT	Relative Error ± SD	p value
Lung V20(%)	13.41% ± 5.57%	13.41% ± 5.56%	−0.07% ± 0.51%	0.973
Lung V10 (%)	20.98% ± 8.18%	20.87% ± 8.10%	0.45% ± 0.68%	0.065
Lung V5 (%)	33.42% ± 12.80%	33.29% ± 12.71%	0.43% ± 1.07%	0.231
Spinal Cord Dmax(Gy)	20.12 ± 8.97 Gy	20.19 ± 8.93Gy	−0.59% ± 1.84%	0.460
Heart V30 (%)	3.33% ± 4.19%	3.31% ± 4.16%	0.27% ± 3.44%	0.278
Heart Dmean (Gy)	6.68 ± 5.55 Gy	6.67 ± 5.53Gy	0.01% ± 0.50%	0.340
PTV D95 (Gy)	61.10 ± 0.18Gy	61.22 ± 0.21Gy	−0.20% ± 0.39%	0.107
PTV D98 (Gy)	60.77 ± 0.21Gy	60.80 ± 0.28Gy	−0.05% ± 0.42%	0.701
PTV$D_{2{\mathrm{cm}}^3}$(Gy)	64.13 ± 0.30Gy	64.12 ± 0.43Gy	0.02% ± 0.69%	0.918

Abbreviations: V_x, the percentage of an organ or tissue volume receiving ≥ x Gy; D_max, the maximum dose received by a given organ or tissue; D_mean, the mean dose received by a given organ or tissue; D_x, the minimum dose received by at least x% of the target volume; $D_{2{\mathrm{cm}}^3}$, the minimum dose received by the most irradiated 2 cm³ of the target volume; SD, standard deviation.

Fig. 3.

The DVH curves of the OARs and PTV for two cases, comparing sCT (dash lines) and pCT (solid lines).

Gamma index analysis was performed across the entire test set to evaluate dosimetric agreement. Under the criteria of 2%/2 mm, 3%/2 mm, and 3%/3 mm, the average gamma passing rate reached 95.39%, 97.30%, and 98.08%, respectively. Notably, all passing rates exceeded the clinical acceptance threshold of 95%, supporting the high dosimetric accuracy of the synthetic sCT for radiotherapy planning purposes. Fig. S3 shows representative gamma index maps under the three criteria, indicating good spatial agreement between sCT and pCT-based dose distributions.

4. Discussion

FFTSC-Net was used to generate sCT images from lung cancer UTE-MRI, followed by dosimetric consistency validation. Compared with the evaluated networks, FFTSC-Net achieved superior image synthesis performance, with an SSIM of 0.80, a PSNR of 20.07 dB, and an MAE (HU) of 59.31. The generated sCT images showed good HU consistency in the OARs and PTV. Mean relative dosimetric differences were within 0.6%, with SDs ranging from 0.39% to 3.44%, and mean gamma passing rates exceed 95% under all three criteria.

A baseline CycleGAN-based U-Net framework [30] was used as the initial benchmark. However, this baseline showed limited modality translation, with residual UTE-specific grayscale patterns and insufficient reconstruction of anatomical structures such as muscle and spinal cord. To address these limitations, several targeted modifications were introduced. SCConv was incorporated into the generator encoder to reduce feature redundancy and improve low-level feature extraction in complex pulmonary regions [20], [35]. SE attention and Transformer encoders were used to enhance feature recalibration during cross-modality translation [36]. In addition, FFT modules were embedded in the skip connections to strengthen spatial–frequency feature representation, particularly for high-frequency structures such as bone and chest wall boundaries [37]. DC loss was further introduced to improve voxel-wise HU consistency by constraining the synthesized images toward the pCT domain [38]. These modifications contributed to improved anatomical reconstruction and image synthesis performance compared with the baseline model.

Previous studies have explored deep learning-based MRI-to-CT synthesis [14], but relatively few have focused on lung sCT generation. Kang et al. [39] employed a basic UNet and CycleGAN for thoracic CT synthesis and reported MAE values of 134 HU and 63.7 HU, respectively, both higher than that achieved by FFTSC-Net in the present study. Bombini et al. [17] achieved a lower whole-body MAE of 35.9 HU using an adaptive Fourier neural operator, whereas FFTSC-Net showed higher synthesis accuracy in the spinal cord region. Although Kang et al. reported superior performance for the spinal cord [39], their study included only six lung cases, limiting the precision and generalizability of the results. Training and inference time are important indicators of model efficiency and practical deployability, particularly for translation-oriented clinical applications. Longuefosse et al. [40] applied a GAN-based approach, whereas Lenkowicz et al. [41] employed a conditional GAN for lung CT synthesis. The former reported an SSIM of 0.92, and the latter achieved a whole-body MAE of 56.8 HU; however, their corresponding training times were substantially prolonged (∼400 h and 70 h, respectively). In comparison, FFTSC-Net achieved competitive quantitative performance with a training time of 45 h and an inference time of 5 min 10 s, suggesting its potential for practical clinical application.

Dosimetric validation further supported the feasibility of FFTSC-Net-generated sCT for dose calculation. HU consistency of different OARs and the PTV was evaluated as an indirect verification of whether the image-level MAE (HU) translates into dose calculation accuracy. For the bilateral lungs, spinal cord, and PTV, no statistically significant differences in mean HU values were observed between sCT and pCT (all p > 0.05), indicating good HU reconstruction consistency of the generated sCT in critical anatomical structures. While Wang et al. [42] reported a statistically significant HU difference in the spinal cord (p = 0.007), no such discrepancy was identified in our work. Moreover, DVH analysis demonstrated that none of the nine evaluated DVH metrics exhibited statistically significant differences between sCT and pCT-based dose calculations (all p > 0.05) (Table 3). Furthermore, several dosimetric trends were consistent with those reported in previous studies [17], [39], [41], supporting the dosimetric consistency of the proposed approach. A gamma passing rate exceeding 95% or even 98% under the 3%/2 mm criterion suggests high spatial dose agreement between sCT and pCT dose distributions. Under the 2%/2 mm criterion, the gamma passing rates achieved in this study compare favorably with those reported by Duetschler et al. [43]. Under the 3%/3 mm criterion, our results were consistent with those reported in related studies [41], [42], supporting the potential utility of the generated sCT for dose calculation.

Several limitations should be acknowledged. (1) The relatively small sample size may limit the generalizability of the model, and larger multi-institutional studies are needed for further validation. (2) Significant HU discrepancies were observed in the heart (p < 0.05), which may be related to cardiac motion because the UTE-MRI protocol was not ECG-gated. Future work may incorporate cardiac-resolved imaging or region-aware constraints to improve synthesis accuracy in mobile regions. (3) The effects of MRI acquisition parameters, such as flip angle and repetition time, were not systematically evaluated. A more T1-weighted regime may improve tissue contrast and potentially enhance the generative model's feature learning, which will be explored in future work. (4) Deformable registration uncertainty and the spatial resolution mismatch between CT and UTE-MRI may have affected quantitative evaluation, and future studies using higher-resolution matched datasets or super-resolution-assisted approaches may help reduce these errors.

CRediT authorship contribution statement

Kaiyi Zheng: Writing – review & editing, Writing – original draft, Data curation. Xudong Xue: Writing – review & editing, Methodology, Funding acquisition. Qi Zhang: Writing – review & editing, Methodology. Fengqin Zhou: Data curation. Jia Xu: Data curation. Xingyu Wang: Data curation. Xiaopeng Song: Software, Methodology. Wei Mao: Software, Methodology. Zilong Yuan: Writing – review & editing, Supervision, Methodology, Funding acquisition. Yulin Liu: Writing – review & editing, Supervision, Resources, Funding acquisition.

Ethics declaration

Written informed consent to take part in the study and to publish the article has been obtained from all participants or their legal representatives. The privacy rights of participants have been observed.

This study was performed in compliance with relevant laws, regulatory frameworks and guidelines where the research took place.

This study was approved by the Hubei Cancer Hospital.

(Approval No. No. LLHBCH2026YN-050)

Funding

This study was supported by Talent Project of Hubei Cancer Hospital (No. 2025HBCHLHRC001); Hubei Province “Chutian Talents Program” medical and health personnel project (No. CTYC001); Beijing Huikang Charity Foundation (No. HKRA2025040009); Scientific Research Projects of Hubei Cancer Hospital (No. 2025HBCHYN23); Project of the Joint Fund for Innovation and Development of the Natural Science Foundation of Hubei Province of China (Grant No. 2026AFC1478); the National Key Clinical Specialty Construction Discipline Project (No. HBCHBCCC02); the Talent Project of Hubei Cancer Hospital (No. 2025HBCHQHRC017).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

Supplementary material