Comprehensive evaluation of a deep learning-based synthetic CT model for MR-only radiotherapy across multiple anatomical sites Authors: Xiao Liang, Chanwoong Lee, Andrew Godley, Junjie Wu, Mu-Han Lin, Trung Truong, Weiguo Lu, Viktor Iakovenko, Steve Jiang, Jie Deng Published: 2026 ABSTRACT Graphical abstractKeywords: MR-only radiotherapy planning, synthetic CT, Image similarity, Dose calculation accuracy, Patient positioning accuracy, Multi-site evaluation Introduction Magnetic resonance imaging (MRI) has become increasingly important in radiation therapy planning because of its superior soft-tissue contrast, which enables more accurate delineation of tumors and organs-at-risk (OARs) [1]. However, computed tomography (CT) remains essential in conventional workflows, as it provides electron density information required for dose calculation and serves as the reference for patient positioning through registration with cone-beam CT (CBCT) or digitally reconstructed radiographs (DRRs). Recent advances in deep learning have accelerated the development of MRI-only radiotherapy planning (MROP), where synthetic CT (sCT) images are generated directly from MRI [2]. This approach eliminates the need for CT simulation, reduces radiation exposure, and avoids MR–CT registration uncertainties. MROP has been extensively studied in individual anatomical sites such as the brain [3], [4], pelvis [5], [6], and head and neck (H&N) [7], with a particularly strong focus on prostate radiotherapy. Despite these advances, several challenges limit broader clinical adoption. Most existing studies have focused on site-specific models [8], [9], [10], [11], and relatively few address multi-site generalizability. Furthermore, validation efforts have primarily emphasized image similarity and dose/volume metric accuracy, while patient positioning accuracy—a critical component for image-guided radiation therapy (IGRT)—remains largely underexplored [12]. Various methods for sCT generation have been proposed, including segmentation-based [13], [14], [15], [16], [17], atlas-based [17], [18], deformation-based [19], [20], and deep learning [9], [21], [22] approaches. Deep learning methods, particularly generative adversarial networks (GANs), have demonstrated strong performance across several anatomical regions. However, thoracic applications remain challenging due to respiratory motion and low MRI signal in lung tissues, and whole-body or multi-site models are still limited. In prior work [21], we developed a weakly supervised deep learning framework using mDixon MRI to generate sCT images without paired MR–CT training data. While this method improved bone representation, it was limited to prostate cases and did not evaluate positioning accuracy. In this study, we extended this approach to a unified, all-sites model applicable across multiple anatomical regions, including H&N, thorax, abdomen, pelvis, and spine. We performed a comprehensive evaluation encompassing image similarity, dose calculation accuracy, and positioning accuracy. The primary contributions of this work were: 1.Development of a unified sCT model for multiple anatomical sites using standardized MRI inputs. Development of a unified sCT model for multiple anatomical sites using standardized MRI inputs. 2.Comprehensive multi-dimensional validation, including positioning accuracy to support clinical implementation of MROP. Comprehensive multi-dimensional validation, including positioning accuracy to support clinical implementation of MROP. Materials and methods Patient cohort and image acquisition This retrospective study included 164 patients treated with external beam radiotherapy between 2022 and 2025 across five anatomical regions: H&N, thorax, abdomen, pelvis, and spine (Table 1). This retrospective study was approved by the Institutional Review Board of The University of Texas Southwestern Medical Center (IRB No. 082013–008). Patients with same-day CT and MRI were prioritized for training to minimize anatomical variation, while testing datasets included both same-day and different-day acquisitions along with treatment CBCT images. Table 1Summary of training and testing cohorts across five anatomical sites. Patient tumor types include: a. malignant neoplasm of tonsil; b. Secondary and malignant neoplasm of lymph nodes of head, face and neck; c. malignant neoplasm of ethmoidal sinus; d. malignant neoplasm of tonsillar fossa; e. malignant neoplasm of maxillary sinus; f. malignant neoplasm of nasal cavity; g. malignant neoplasm of upper-outer quadrant of right female breast; h. malignant neoplasm of overlapping sites of accessory sinuses; i. malignant neoplasm of bronchus or lung; j. malignant neoplasm of connective and soft-tissue of thorax; k. secondary and malignant neoplasm of intrathoracic lymph nodes; l. localized enlarged lymph nodes; m. malignant neoplasm of kidney; n. malignant neoplasm of pancreas; o. secondary malignant neoplasm of liver and intrahepatic bile duct; p. secondary malignant neoplasm of kidney and renal pelvis; q. malignant neoplasm of overlapping sites of cervix uteri; r. malignant neoplasm of endometrium; s. malignant neoplasm of labium majus; t. secondary malignant neoplasm of bone; u. malignant neoplasm of prostate;H&NThoraxAbdomenPelvisSpineTraining number of patients301425460Testing number of patients101091010Sex (males/females)35/58/1622/1220/366/4Tumor typea, b, c, d, e, f, g, hi, j, k, lm, n, o, pq, r, s, t, utType of treatmentsIMRTIMRT/SBRTIMRT/SBRTIMRT/SBRTSBRTImage guidanceCBCTMRICBCTCBCTCBCT Summary of training and testing cohorts across five anatomical sites. Patient tumor types include: a. malignant neoplasm of tonsil; b. Secondary and malignant neoplasm of lymph nodes of head, face and neck; c. malignant neoplasm of ethmoidal sinus; d. malignant neoplasm of tonsillar fossa; e. malignant neoplasm of maxillary sinus; f. malignant neoplasm of nasal cavity; g. malignant neoplasm of upper-outer quadrant of right female breast; h. malignant neoplasm of overlapping sites of accessory sinuses; i. malignant neoplasm of bronchus or lung; j. malignant neoplasm of connective and soft-tissue of thorax; k. secondary and malignant neoplasm of intrathoracic lymph nodes; l. localized enlarged lymph nodes; m. malignant neoplasm of kidney; n. malignant neoplasm of pancreas; o. secondary malignant neoplasm of liver and intrahepatic bile duct; p. secondary malignant neoplasm of kidney and renal pelvis; q. malignant neoplasm of overlapping sites of cervix uteri; r. malignant neoplasm of endometrium; s. malignant neoplasm of labium majus; t. secondary malignant neoplasm of bone; u. malignant neoplasm of prostate; CT simulations were performed on a Brilliance Big Bore scanner (Philips, Netherlands) with 1–3 mm slice thickness, in-plane resolution of 0.9 × 0.9 to 1.4 × 1.4 mm2, and tube voltage of 120 kVp. Thoracic and abdominal patients underwent 4DCT or breath-hold scans, while other sites used free-breathing acquisitions. MRI scans were acquired on an Ingenia Ambition 1.5 T system (Philips, Netherlands) using a T1-weighted 3D mDixon Vane sequence before contrast injection. The mDixon sequence typically reconstructs water-only image, fat-only image, and in-phase (IP) images, with the IP image combining water and fat signals. Both water and IP images were used as model inputs. Sequence parameters for each body site are listed in Table 2. Table 2Imaging parameters for 3D T1-weighted mDixon Vane acquisition. FOV: Field-of-View; S/I: Superior/Inferior; SENSE: SENSitivity Encoding; TR/TE: repetition time/ echo time; TFE: Turbo field echo.ParameterBrainH&NThorax/Abdomen/PelvisFOV (mm2)280 × 280550 × 550550 × 550S/I coverage (mm)230300400In-plane pixel size (mm2)1.2 × 1.21.3 × 1.31.75 × 1.75Slice thickness/gap (mm)2/-13/0(2–3)/0SENSE factor (slice direction)1.52.2NoPartial Fourier factor0.80.80.6–0.7Radial percentage (%)180180120TR/TE1/TE2 (ms)6.1/1.8/3.55.9/1.8/3.55.6/1.8/3.5Water-fat shift (pixels)/BW(Hz)0.3/7200.3/7200.3/720TFE factor8551154Flip angle (°)101010 Imaging parameters for 3D T1-weighted mDixon Vane acquisition. FOV: Field-of-View; S/I: Superior/Inferior; SENSE: SENSitivity Encoding; TR/TE: repetition time/ echo time; TFE: Turbo field echo. For the testing cohort, CBCT images were obtained using on-board imagers from TrueBeam, VitalBeam, Halcyon (Varian, USA), and Versa (Elekta AB, Sweden), with a slice thickness of 2 mm, an in-plane resolution of 0.5 × 0.5 to 2 × 2 mm2, and tube voltages ranging from 100 to 125 kVp. All-sites sCT deep learning model The sCT images were generated using an all-sites deep learning model adapted from our previously published multichannel CycleGAN framework for MROP [21]. In the prior work, the model was developed and evaluated using prostate data only. In the current study, the framework was extended to support multiple anatomical regions across the body using a unified training strategy. The model employed a modified multichannel CycleGAN architecture consisting of three generators and three discriminators. The primary generator converted paired Dixon MR images into sCT images, while the remaining generators performed inverse image translation from CT to the corresponding MR domains. Similar to our prior work, U-Net-based generators and PatchGAN discriminators were used for image translation and adversarial learning. The model was trained using a composite objective function including adversarial loss, cycle-consistency loss, identity loss, and anatomy-specific regularization terms. In our previous study, Dixon fat and in-phase (IP) images were used as model inputs. However, the generated sCT occasionally demonstrated artificial voids within the bladder region, which were attributed to the relatively low signal intensity surrounding the bladder on fat images. To mitigate this limitation, the current study replaced fat images with Dixon water images while retaining the IP images as multichannel inputs. The water images provided improved soft-tissue signal representation around the bladder and reduced the appearance of artificial air pockets. In addition, a bladder-specific loss term was incorporated into the generator objective function to further improve anatomical consistency in the bladder region. Additional details regarding the network architecture and preprocessing pipeline are provided in Supplementary Material A. Image similarity Rigid and deformable image registration (DIR) between sCT and planning CT were performed using Velocity (Varian, USA). Masks were generated from the planning CT to focus on various regions: 1. Bone mask, 2. Soft-tissue mask, 3. Whole-body mask. Bone masks (MBone) were generated from the planning CT images using a global threshold of HU > 200. The bone threshold of 200 HU was selected based on meta-analyses and large cohorts study [23]. The body mask (MBody) was created from the planning CT images using a threshold of HU > -300 to exclude background air, followed by cavity filling to include internal air cavities and ensure complete body coverage. The continuity and validity of the resulting body contours were visually verified. Soft-tissue masks (Msofttissue) were obtained by subtracting MBone from MBody. Image similarity was evaluated by computing the mean absolute error (MAE) and mean error (ME) of HU values within MBody, Msofttissue, and MBone. The corresponding equations are shown below: where: Nx denotes the number of voxels in the mask, CTi is the intensity of the CT at voxel i, and sCTdef(i) is the intensity of the deformed sCT at voxel i. Bland–Altman analysis was performed to quantitatively assess the agreement between CT and sCT by calculating the ME (bias) and the 95% limits of agreement (bias ± 1.96 standard deviation). Dose calculation accuracy Clinical treatment plans were recalculated on sCT images and compared with CT-based plans. Dose calculations were performed using Monaco (Monte Carlo) for the thorax site and Eclipse (Acuros External Beam, v16.1.0) for all the other sites. A resolution of 1.5 mm was used for stereotactic body radiation therapy (SBRT) and 2.5 mm for intensity-modulated radiation therapy (IMRT), with heterogeneity correction enabled (dose-to-medium). Dose evaluation included gamma analysis (2 mm/2%, 10% threshold, global normalization) and dose-volume histogram (DVH) metrics for targets and OARs. For the planning target volumes (PTVs), we evaluated DVH metrics including maximum dose (Dmax), mean dose (Dmean), the percentage volume receiving 100% of the prescription dose (V100%(%)), and the dose received by 95% of the volume (D95%). For organs-at-risk (OARs), Dmax and Dmean were assessed. Positioning accuracy Positioning accuracy was assessed by comparing registration results from CBCT-to-CT versus CBCT-to-sCT. Registration procedures were site-specific: for H&N, thorax, abdomen, and pelvis, we performed automatic global registration followed by manual refinement using seeds or soft-tissue around the target; for spine, global registration was similarly followed by regional bony alignment adjustments. Because MRI and planning CT images were acquired in different coordinate systems, the generated sCTs (inheriting MRI coordinates) were not directly aligned with the CT. Therefore, each sCT was first rigidly registered to the planning CT, and then the CBCT was registered to the coordinate corrected sCT. For each case, both translational and rotational registration parameters (Δx, Δy, Δz, pitch, roll, yaw) were extracted from the registration outputs. The translational error and rotation error were defined as the difference in estimated shifts and rotations between the two registration methods: where T represented the 3D translation vector obtained from rigid registration, while R represented the rotation obtained from the rigid registration. We performed a Bland–Altman analysis to assess positional accuracy. Specifically, the mean error and the 95% limits of agreement were calculated for the evaluated positional differences. These results are shown in Table 5 to provide a quantitative assessment of agreement. Table 5Comparison of positioning accuracy using sCT versus CT as the reference image for CBCT-based registration. Bland–Altman analysis is used to evaluate translational and rotational differences. Results are reported as mean error and 95% limits of agreement.Translation (mm)Rotation (degree)xyzPitchYawRollH&N0.07[-1.30, 1.44]0.17[-1.25, 1.58]−0.38[-1.36, 0.60]−0.06[-1.61, 1.48]−0.16[-1.35, 1.03]−0.21[-1.47, 1.04]Abdomen−0.55[-4.75, 1.60]−0.20[-4.05, 1.76]−0.50[-4.91, 1.75]0.12[-4.38, 2.41]−0.53[-3.62, 1.05]−0.56[-3.37, 0.87]Pelvis−0.19[-3.19, 2.81]1.23[-2.92, 5.39]0.63[-2.77, 4.03]−0.18[-1.91, 1.56]−0.24[-1.14, 0.66]0.11[-1.04, 1.27]Spine0.31[-1.28, 1.90]−0.02[-1.14, 1.11]−0.37[-1.97, 1.23]0.12[-0.97, 1.22]0.04[-0.47, 0.56]0.01[-0.88, 0.89] Comparison of positioning accuracy using sCT versus CT as the reference image for CBCT-based registration. Bland–Altman analysis is used to evaluate translational and rotational differences. Results are reported as mean error and 95% limits of agreement. All registrations were performed in Velocity using rigid alignment only. Two observers (a medical physicist and a radiation therapist) performed and verified all registrations. The initial image registration was performed by a medical physicist. Each registration was then independently reviewed by a radiation therapist. If any disagreement was identified, the case was discussed jointly and the registration was refined until consensus was reached. Therefore, all final registrations used for analysis were consensus-approved by both observers. Because discrepancies were resolved through discussion rather than recorded as independent final registrations, a formal quantitative interobserver agreement metric was not calculated. This also reflects how differences are resolved during treatment. Thorax cases that have both CT and MRI simulation were all treated on MR-LINAC systems without CBCT; thus, thorax cases were excluded from the positioning analysis. Results Image similarity accuracy Representative axial-view sCT images across all anatomical sites in axial view are shown in Fig. 1. The full sets of axial, sagittal, and coronal views for each case are shown in Figs. S1-S5. HU differences between planning CT and deformed sCT in whole-body, bone, and soft-tissue regions are summarized in Table 3. Bland–Altman analysis was performed to assess agreement between CT and sCT, and the corresponding plots are included in Fig. S6, showing the mean bias and 95% limits of agreement across evaluated regions. Representative axial views of sCT images across all anatomical sites. Table 3Differences between planning CT and deformed sCT images evaluated using MAE and ME of HU values within MBody, Msofttissue, and MBone regions. Results are reported as mean (standard deviation) in the unit of HU across the test patients.Whole-BodyBoneSoft-tissueMAEMEMAEMEMAEMEH&N111(59)20(23)249(30)119(43)36(31)−17(36)Thorax83(20)37(31)155(32)94(31)74(21)30(34)Abdomen96(24)31(41)209(38)164(45)80(30)15(41)Pelvis59(13)27(21)191(38)53(62)43(16)17(23)Spine89(9)47(15)228(30)137(53)62(15)24(20) Differences between planning CT and deformed sCT images evaluated using MAE and ME of HU values within MBody, Msofttissue, and MBone regions. Results are reported as mean (standard deviation) in the unit of HU across the test patients. Across all anatomical sites, the deformed sCT images demonstrated agreement with planning CT, with whole-body MAE ranging from 59 HU in the pelvis to 111 HU in the H&N, and ME ranging from 20 to 47 HU. In bone regions, larger discrepancies were observed, with MAE values ranging from 155 HU (thorax) to 249 HU (H&N), and ME values ranging from 53 to 164. For soft-tissue, the agreement was improved compared to bone, with MAE values ranging from 36 HU (H&N) to 80 HU (abdomen), and ME values ranging from −17 to 30 HU. Overall, the results indicated that sCT provided better HU agreement in soft-tissue than in bone, with the largest discrepancies observed in high-density structures and in the H&N region (Table 3). Dose differences Dose differences between CT- and sCT-based plans for PTV and OARs across all sites are summarized in Tables S1–S5. Strong agreement was observed between CT- and sCT-based dose calculations. PTV dose differences were generally less than 1 Gy across all sites. Gamma passing rates exceeded 95% for most regions, including H&N (99.6%), abdomen (97.5%), pelvis (98.6%), and spine (95.5%) (Table 4). The thorax showed notably lower gamma passing rates (86.8% ± 7.8%), despite relatively modest PTV dose differences (Table 4). Table 43D Gamma-index passing rates comparing dose distributions calculated on planning CT and sCT. The gamma criteria are 2 mm/2% with 10% dose threshold and global normalization. Results are reported as mean (standard deviation) [range].Gamma passing rate (%)H&N99.6 (0.3) [99.0, 100.0]Thorax86.8 (7.8) [76.7, 97.5]Abdomen97.5 (3.1) [90.7, 100.0]Pelvis98.6 (1.8) [94.4, 100.0]Spine95.5 (3.1) [88.1, 98.4] 3D Gamma-index passing rates comparing dose distributions calculated on planning CT and sCT. The gamma criteria are 2 mm/2% with 10% dose threshold and global normalization. Results are reported as mean (standard deviation) [range]. For the abdomen, one outlier case showed reduced gamma agreement due to bowel gas differences and surgical clips visible on CT but absent on MRI (Fig. S7). For spine cases, one outlier was driven by dose differences near lung interfaces (Fig. S7). Higher gamma values were observed in low-density regions such as lungs. However, PTV dose differences remained small (e.g., Dmean 0.42 Gy, Dmax 0.21 Gy, V100% 0.3%, and D95% 0.6 Gy), whereas lung regions exhibited larger deviations (Dmax difference up to 1.11 Gy), consistent with HU mismatch. The gamma passing rate improved from 88.1% to 99.4% when lung HU in the sCT was overridden to a typical value of −750 HU. These findings indicate that the observed dose differences were primarily driven by HU mismatches in low-density lung regions. Positioning accuracy Table 5 summarizes the positioning accuracy of sCT-based alignment relative to planning CT. The observed differences reflect a combination of registration uncertainty and true anatomical variation between imaging sessions. Positioning analysis showed strong agreement between sCT- and CT-based registration in rigid anatomical regions. In H&N and spine, translational differences were within ± 2 mm. Greater variability was observed in the abdomen and pelvis, where translational limits extended up to ± 5 mm. These differences were primarily driven by anatomical changes such as organ motion and filling between when the CT and MR were acquired, rather than registration error. Discussion This study presented a comprehensive evaluation of a deep learning-based sCT model across multiple anatomical sites. The results demonstrated that the model achieved clinically acceptable performance in terms of image quality, dose/volume metric accuracy, and positioning consistency. A limitation of this study is the reliance on deformably registered CT as the reference for voxel-wise comparison. DIR introduces uncertainty, particularly in anatomically variable regions such as the lung and abdomen, where perfect spatial correspondence cannot be achieved. As a result, the reported HU differences reflect a combination of sCT generation error and residual registration error. Future work could further quantify DIR-related uncertainty (e.g., using landmark-based validation) to better isolate model-specific performance. Compared with published sCT studies [18], the observed HU errors fall within reported ranges, particularly for whole-body and soft-tissue regions. As expected, bone regions show larger discrepancies due to their complex structure. From a clinical perspective, dose/volume metric accuracy is the most important criterion. Established guidelines consider dose differences within 1–2% or ≤ 1–2 Gy to be acceptable [30–33]. Gamma analysis criteria of 3%/3 mm or 2%/2 mm with ≥ 95% passing rates are widely accepted benchmarks for clinical dose distribution validation [34–36]. Our results meet these thresholds across most sites, supporting the clinical viability of the model. The thorax remains the most challenging region due to respiratory motion, susceptibility artifacts, and difficulty visualizing lung structures on MRI. Reduced sCT accuracy in the lung region was primarily attributed to the model assigning voxel values closer to − 1000 HU rather than the typical lung density range of −800 to −500 HU [24]. Additional training data and advanced MRI sequences may improve performance in this area. Positioning accuracy analysis highlights an important distinction between rigid and deformable anatomical regions. In the H&N and spine, alignment is highly reproducible due to stable bony anatomy. In contrast, abdominal and pelvic variability is largely driven by physiological changes rather than technical limitations. This increased variability is unlikely to be explained by registration uncertainty alone and is more consistent with true anatomical variation, including changes in bladder filling, bowel gas, and soft-tissue deformation between CT, MRI, and CBCT acquisitions. MRI-related artifacts also contributed to registration uncertainty. Because sCT was derived from MRI, artifacts in the source images propagate into the sCT and could introduce misalignments greater than 1 mm. Fig. S8 illustrates a ring artifact near the prostate that degraded visualization on both MRI and sCT, likely caused by the combination of high Sensitivity Encoding (SENSE) factors (2–3) and the 3D Vane gradient-echo sequence. We have recently mitigated this artifact by disabling SENSE and increasing radial oversampling while maintaining short scan times. Rigorous patient-specific MRI QA is essential when MRI is used for sCT generation, and any imaging defects must be corrected before simulation. Patient-induced artifacts—such as susceptibility from bowel/rectal gas, metallic implants, or large body size—should also be assessed before proceeding. Site-dependent alignment strategies also contributed to the observed differences. H&N and spine alignment relied primarily on stable bony anatomy, whereas abdominal and pelvic alignment required refinement based on soft-tissue landmarks (e.g., bladder, bowel, rectum, liver, and kidneys) when visible on CBCT. Because these soft-tissue structures can change between CT (used for planning) and MRI (used for sCT generation), anatomical variation introduced greater registration variability. Fig. S9 shows a cervix case in which differences in bladder filling altered cervix position and shape across CT, CBCT, and sCT/MRI. Even with only 1 h between CT and MRI acquisitions, intra-day anatomical variation affected inter-modality alignment. These errors can be removed by an MR-only workflow. Because registrations were performed using a consensus workflow, interobserver variability was not independently quantified; therefore, the reported positioning differences reflect a combination of residual registration uncertainty and true anatomical variation. Anatomically stable regions (e.g., head & neck, spine) were more likely driven by registration uncertainty, whereas larger variability observed in the abdomen and pelvis is predominantly attributable to true anatomical changes such as organ filling and motion. Future work will include a formal assessment of interobserver variability using independent registrations and quantitative metrics (e.g., repeat registration analysis), which would allow more rigorous separation of observer-related uncertainty from anatomical variation. Several practical considerations remain for clinical implementation. MRI field-of-view limitations may lead to truncation effects, although these are often clinically insignificant depending on treatment geometry. Implants, particularly non-metallic ones, remain challenging due to their variable appearance on MRI and limited representation in training data. Generalizability across MRI vendors and protocols is also an important consideration. Although the model was trained primarily on mDixon images from a single vendor, variations in acquisition parameters and noise across platforms may influence sCT accuracy. It remains unclear whether full retraining is necessary or whether lightweight fine-tuning with a small site-specific dataset is sufficient. Given the consistent physics underlying mDixon imaging, we anticipate that transfer learning or modest fine-tuning will effectively harmonize data across platforms. Future studies should systematically evaluate cross-platform performance to confirm the model’s generalizability. Overall, this work demonstrates the feasibility of a unified all-sites sCT model for MRI-only radiotherapy. By integrating image similarity, dose/volume metric accuracy, and positioning evaluation, this study provides a comprehensive foundation for clinical adoption. Future work will focus on improving robustness in challenging regions, expanding multi-institutional validation, and refining quality assurance strategies. CRediT authorship contribution statement Xiao Liang: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Resources, Project administration, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Chanwoong Lee: Visualization, Formal analysis, Data curation. Andrew Godley: Writing – review & editing, Visualization, Validation, Supervision, Resources, Methodology, Investigation, Formal analysis, Conceptualization. Junjie Wu: Validation, Resources, Investigation, Data curation. Mu-Han Lin: Supervision, Resources, Conceptualization. Trung Truong: Validation, Investigation. Weiguo Lu: Resources, Data curation, Conceptualization. Viktor Iakovenko: Resources, Investigation, Data curation, Conceptualization. Steve Jiang: Supervision, Resources, Methodology, Investigation, Conceptualization. Jie Deng: Writing – review & editing, Validation, Supervision, Resources, Project administration, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. 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. Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.phro.2026.101009. Supplementary data The following are the Supplementary data to this article: Supplementary Data 1 Additional figures demonstrating sCT generation across anatomical sites, Bland–Altman and dosimetric analyses, CBCT-to-sCT alignment case studies, MRI artifact and anatomical variation examples, and details of the deep learning model architecture, preprocessing pipeline, and training strategy. References Associated Data Supplementary Materials Supplementary Data 1 Additional figures demonstrating sCT generation across anatomical sites, Bland–Altman and dosimetric analyses, CBCT-to-sCT alignment case studies, MRI artifact and anatomical variation examples, and details of the deep learning model architecture, preprocessing pipeline, and training strategy.