Clinical implementation of deep learning-based synthetic CT for MRI-only volumetric modulated arc therapy in head and neck and pelvic cancer patients

Abstract

Rationale and objective

This study investigates the implementation of an MRI-only workflow in radiotherapy, focusing on synthetic CT (sCT) images generated through a deep learning-based commercial software to facilitate dose calculation and treatment delivery for volumetric modulated arc therapy (VMAT) in head and neck (H&N) and pelvic cancer patients.

Methods

This retrospective analysis included 33 patients (10 H&N, 9 prostate, 9 rectum, and 5 cervical cancer). All patients underwent CT and MRI for radiotherapy planning. sCT images were generated using deep learning-based MRI Planner™ software. Clinical treatment plans were initially optimized on CT and then recalculated on sCT with identical parameters. Image quality was evaluated using the dice similarity coefficient (DSC), mean error (ME), and mean absolute error (MAE). Dosimetric accuracy was assessed by comparing dose-volume histogram (DVH) differences and performing global gamma analysis (3%/2 mm to 1%/1 mm) between CT- and sCT-based plans. Treatment plan quality assurance (QA) conducted with an electronic portal imaging device (EPID) was compared between sCT and CT using global gamma analysis (3%/2 mm to 2%/2 mm). Patient set-up verification was evaluated by comparing CT-cone beam CT (CBCT) and sCT-CBCT.

Results

sCT images demonstrated high accuracy, with average body DSCs of 0.99 (SD = 0.00), ME of -8.94 HU (SD = 9.50), and MAE of 67.30 HU (SD = 6.94) for H&N, and DSCs of 1.00 (SD = 0.00), ME of -8.74 HU (SD = 6.26), and MAE of 33.64 HU (SD = 3.91) for the pelvis. Lower DSCs were observed in bone, with 0.87 (SD = 0.03) in H&N and 0.88 (SD = 0.02) in pelvis. Dose differences were within 2%, with average gamma pass rates of all criteria 92% for H&N and 96% for pelvis, across all criteria. QA plan evaluation revealed the gamma pass rate between CT and sCT of 0.05% for H&N and 0.12% for pelvis. Mean positioning differences between CT-CBCT and sCT-CBCT were 0.13 mm and 0.06 for the H&N and 0.17 mm and 0.13 for the pelvis in all directions.

Conclusion

Deep learning-based software successfully generated accurate sCT images for both H&N and pelvic cancer patients, supporting reliable dose calculation, treatment plan QA, and patient set-up verification. This enables the potential implementation of MRI-only workflows for VMAT in H&N and pelvic cancer treatment.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13014-025-02744-2.

Introduction

In conventional radiotherapy workflow, computed tomography (CT) is employed for dose calculations based on tissue electron density, while magnetic resonance imaging (MRI) is utilized for structure delineation, benefiting from its advanced soft tissue visualization where accurate visualization of tumors and organs-at-risk (OARs) is crucial such as the head and neck (H&N) and pelvis. However, combining CT and MRI through image registration can potentially introduce systematic uncertainties of 2–5 mm [1], primarily arising from imperfect registration caused by patient setup variations during image acquisition. This may result in reduced tumor local control and an elevated risk of treatment-related toxicities [2]. Particularly in highly conformal treatments such as volumetric modulated arc therapy (VMAT), where precise dose delivery is essential for achieving tumor coverage while preserving surrounding healthy tissues.

To address these challenges, an emerging MRI-only workflow offers a promising strategy to mitigate such uncertainties by eliminating the need for multi-modality image registration. Since MR does not provide electron density information, the conversion of MR data into Hounsfield unit (HU) to facilitate dose calculations, known as synthetic computed tomography (sCT) must be generated to replace conventional CT in the treatment planning process [3]. sCT is applicable throughout the radiotherapy process, from dose calculation to registration with cone beam computed tomography (CBCT) for set-up verification before radiation therapy delivery. Additionally, the MR-only workflow offers several advantages, including reduced radiation dose, simplified planning processes, decreased patient discomfort due to fewer scans, and lower planning-related costs [1, 4, 5]. Furthermore, it holds promising potential for MR-linear accelerator (MR-linac) treatment [1, 4].

Six methods have been explored for sCT generation: sequence-based, bulk density, atlas-based, patch-based, deep learning-based, and hybrid methods. Sequence-based methods estimate tissue-specific HUs from MRI signal variations across multiple sequences without requiring CT data [6]. Bulk density methods assign predefined electron density values to MRI-segmented tissues [4], offering consistent HU and minimizing artifacts and noise [7]. Nonetheless, their accuracy depends on accurate segmentation and specialized MR sequences, such as ultrashort echo time (UTE) or zero echo time (ZTE) [6] and may not be clinically acceptable for heterogeneous tissue [2, 8]. Atlas-based methods register one or multiple MR-CT image pairs, known as a CT atlas, to align with the input MRI, through their accuracy may be limited when patient anatomy differs significantly from the atlas [1, 7]. Patch-based methods aim to reduce registration errors by matching small MRI regions to a patch library derived from paired MRI-CT datasets to estimate local HU values [6]. Deep learning-based methods have been introduced. These models can learn complex anatomical features and their corresponding HU values using a single standard MRI sequence and offer fast sCT generation [9, 10]. However, model robustness depends on the diversity of the training data [10]. Lastly, hybrid methods combine two or more of these approaches to improve sCT accuracy and robustness, particularly beneficial in anatomically complex regions such as the H&N [6].

Among various methods, deep learning-based sCT generation has rapidly gained attention as a promising solution to enable MRI-only radiotherapy workflows. Numerous studies have investigated its application across various anatomical regions, including the brain [10–12], H&N [13–15], and pelvis [16, 17]. While promising results have been consistently demonstrated for the brain and pelvic regions, the implementation in more anatomically complex and variable sites such as the H&N remains challenging and requires further enhancement [18]. Despite growing interest and the potential to streamline radiotherapy workflows, the clinical adoption of sCT solutions has been limited, with most implementations relying on in-house models with limited generalizability across institutions.

To facilitate broader clinical translation, commercially available deep learning-based solutions, such as MRI Planner™ (Spectronic Medical AB, Helsingborg, Sweden), have been introduced. This software offers dedicated sCT generation models for the brain, H&N, and pelvic anatomies. The software leverages multiple high-resolution convolutional neural networks trained on large and multicenter MR-CT datasets. This design ensures robustness across various MR scanners and imaging parameters [19]. Several studies have explored the use of MRI Planner™ for generating sCT images in the brain [20], H&N [3, 9], and pelvis [19, 21], demonstrating promising technical and dosimetric performance.

However, sCT accuracy and clinical suitability can be highly site-dependent and influenced by factors such as anatomical complexity, imaging protocol variations, and differences in treatment planning system (TPS). Moreover, the lack of standardized guidelines for sCT commissioning, quality assurance (QA), and clinical evaluation methodologies complicates cross-study comparisons and hinders consistent and safe clinical implementation [22, 23]. Therefore, a comprehensive and site-specific evaluation that considers both technical and clinical factors, including MR scanner systems, acquisition protocols, TPS platforms, and patient setup workflows, is therefore essential before integrating sCT solutions into routine clinical practice.

Accordingly, this study aims to conduct a thorough clinical validation of deep learning-based sCT images generated by MRI Planner™ within a complete radiotherapy workflow. This includes assessing image and dosimetric accuracy, treatment plan quality assurance, and patient setup verification, benchmarked against conventional CT as the clinical gold standard. The outcomes of this study are expected to provide critical insights for advancing safe, efficient, and clinically robust MRI-only radiotherapy workflows for H&N and pelvic cancer treatments.

Materials and methods

Patient data and image acquisitions

A retrospective non-invasive study was conducted, including a cohort of 33 patients with 10 H&N and 23 pelvic cancers (9 with prostate, 9 with rectal, and 5 with cervical cancer). The study was received approval from the Regional Ethics Review Board (MURA2023/860). The inclusion criteria comprised patients with soft tissue tumors or soft tissue metastases in the H&N, prostate, rectal, and cervical regions who underwent both CT and MRI scans covering the patient’s body for VMAT planning. Patients with major implants that could interfere with MRI images were excluded. The study was designed to investigate sCT images without disrupting the clinical treatment workflow.

Each patient underwent head-first supine CT imaging using the GE scanner (Optima 580, GE Healthcare, Milwaukee, WI, USA) with individual support and immobilization. The scanning parameters were set to 120 kVp and 320 mA for the H&N and 250 mA for the pelvis. The slice thickness of H&N was 1.25 mm with a 1.25 mm interval, and the image resolution was 512 × 512 with a 1 × 1 mm² pixel size. For the pelvis, the slice thickness was 2.5 mm with a 2.5 mm interval, and the image resolution was 512 × 512 with a 1 × 1 mm². Subsequently, MR images were acquired using a 1.5 T MRI scanner (Magnetom Aera, Siemens, Erlangen, Germany) within the time interval from 1 h to 3 days. Specifically, the MRI Planner software requires particular MR sequences for sCT generation. The 3D T1 VIBE Dixon sequence was acquired for the H&N, and the 2D T2 turbo spin echo (TSE) sequence was acquired for the pelvis (refer to Table 1 for details). These sequences provide the necessary tissue contrast and geometric fidelity to enable accurate sCT prediction based on the software’s trained models [19]. In our study, these MR sequences were integrated into the standard clinical MRI protocol for radiotherapy planning, ensuring that the workflow remains compatible with routine clinical practice. The MR scan was conducted with a flat tabletop (CIVCO Radiotherapy, Iowa, USA), maintaining the same position and immobilizations as the CT scan and radiotherapy treatment sessions. To prevent body contour deformation in the pelvis, a coil bridge was employed. The large field of view (FOV) covered the patient’s outline and the entire volume in the craniocaudal direction for radiation dose calculation. All images were acquired with high acquisition bandwidth and with vendor-specific distortion correction to minimize MRI geometric distortion. The MRI system had been previously validated for geometric accuracy and distortion, as reported by Chaknam et al., [24] and the imaging sequence was deliberately selected to represent a worst-case scenario for residual distortion.

Table 1.

The specific scanning parameters of MRI for sCT generation in the H&N and pelvis

Parameters	H&N	Pelvis
Sequence type	T1 VIBE DIXON	T2 TSE
Acquisition type	3D	2D
Slice orientation	Transverse	Transverse
Slice thickness (mm)	2.0	2.5
Slice gap (mm)	0	0
Image resolution	448 × 448	512 × 512
Bandwidth (Hz/pixel)	795	200
FOV (mm)	500	448
Repetition time (ms)	8.0	12370.0
Echo time (ms)	2.4	97.0
3D Geometry correction	Yes	Yes
Scan time (min)	4.14	7.27

All included cases were evaluated at each step of MRI-only workflow against the conventional CT-based workflow, as illustrated in Fig. 1, which presents a side-by-side comprehensive comparison with associated evaluation metrics.

Fig. 1.

Schematic representation of sCT evaluation within the MRI-only workflow. The diagram illustrates key processing steps and quantitative metrics used to assess sCT accuracy, with conventional CT serving as the ground truth

Synthetic CT generation

The sCT images were automatically generated through the transfer function estimation (TFE) algorithm within MRI Planner™ (version 2.3, Spectronic Medical AB, Helsingborg, Sweden) integrated into the hospital network. The algorithm utilizes a specialized structure of 3D deep convolutional neural networks (CNNs) architecture with multiple residual connections and a large receptive field to ensure robust, context-aware image synthesis. Trained on database containing matched MR and CT images from multi centers, the network estimates spatially variant coefficients of a polynomial transfer function based on standardized MR inputs. These coefficients are then applied to the original MRI to generate the sCT while preserving image resolution. Following verification of the compatibility and consistency of user-defined acquisition protocols, the algorithm estimates the spatially variant coefficients of a polynomial transfer function for the incoming MRI and creates the corresponding sCT. Notably, the gas in the rectal cavities is automatically configured as water equivalent in the sCT due to these cavities not appearing in the same location during treatment for the pelvis. Additional information is provided by the vendor [19].

Geometric and image intensity accuracy

For image and dosimetric comparisons, sCT and CT images were aligned to ensure proper anatomical correspondence through the image registration process. The CT images were registered with the sCT using two-step processes, involving rigid image registration (RIR) followed by deformable image registration (DIR), using Velocity software (version 4.1.0, Varian Medical Systems, Palo Alto, CA). The deformed CT was then resampled to match the in-plane resolution and reconstruction matrix of the sCT. The registration accuracy was assessed through crossfading and visual inspection of bone and tissue interfaces before importing the images into the TPS.

The evaluation of the sCT involved assessing geometric accuracy and image intensity, focusing on the segmented structures and body contours. Specific thresholds were applied to account for varying bone intensity and tissue components across anatomical regions. For the H&N, the thresholds were set at > 250 HU for bone, -200 to 250 HU for soft tissue, and <-200 HU for air. In the pelvis, the thresholds for bone and soft tissue were adjusted to > 100 HU and < 100 HU, respectively. The body contours were automatically generated by the TPS. The geometric accuracy was evaluated employing the dice similarity coefficient (DSC), to determine alignment between segmented structures on the sCT and CT, in which specific equation is as follows:

1

where the terms and represent the structure volume of the sCT and CT, respectively. The DSC range is 0 to 1, and an acceptance criterion should be more than 0.8–0.9 according to AAPM TG 132 [25].

To assess image intensity, represented by the HU value, the mean error (ME) and mean absolute error (MAE) were calculated. To minimize uncertainty caused by streak artifacts from dental fillings, affected CT slices in the H&N were excluded from the ME and MAE evaluations. The ME and MAE equations are described as follows:

2

3

where the term N represents the number of voxels in the reference structure obtained from the CT.

Dosimetric evaluation

To assess the feasibility of sCT for radiotherapy planning, dosimetric differences between sCT- and CT-based plans were evaluated for each structure and the overall dose distribution. For each patient, the delineation of target volumes and OARs was performed on the CT with guidance from MRI. Subsequently, A 6 MV VMAT plan was optimized and calculated using an anisotropic analytical algorithm (AAA) within Eclipse TPS (version 16.1.0, Varian Medical System, Palo Alto, CA) with a grid size of 2.5 mm. To evaluate the dose difference between sCT and CT, the clinical structures and plan parameters from CT were transferred to the sCT images, excluding the automatically generated body structure by the TPS. The propagated contours underwent review and approval by a radiation oncologist. In order to minimize dose comparison uncertainty in each specific area, the streak artifact on the CT of the H&N region was replaced with the soft tissue HU, and the rectal gas on the CT of the pelvis was replaced with water equivalent HU. Then, the VMAT plans were recalculated with fixed beam geometry and monitor unit (MU), without re-optimization, to preserve the original plan parameters and control points. The default CT calibration curve in the TPS was used for conversion of HUs to relative electron density.

The dosimetric evaluation focused on the high dose planning target volume (PTV), which encompassed the margin from the cranial vertex to the clavicle for the H&N and from the iliac crest to the pubic arc for the pelvis. The percent dose differences of PTV and OARs were evaluated according to ICRU No.83 dose volume histogram (DVH) report [26]. For both H&N and pelvic, PTV D95%, D2%, and Dmean were assessed. In the H&N region, OARs evaluated included the spinal cord (D2%), brainstem (D2%), and the left and right parotid glands (Dmean). For the pelvic region, the maximum dose (Dmax) to the left and right femoral heads was analyzed. Additionally, D_mean and D_2% of the entire body were examined. The percent dose difference was calculated relative to the CT dose. It is important to note that the bladder and rectum were excluded from dose evaluation due to shape and volume variability, which may introduce confounding factors in assessing sCT performance.

Furthermore, the dose distribution was evaluated using global gamma analysis at the criteria of 3%/2 mm, 2%/2 mm, and 1%/1 mm, with a 10% low dose cut-off. The global gamma analysis was performed relative to the prescribed dose using MICE Toolkit (MICE Toolkit™, v.2022.4.9, Nonpi Medical, Umea, Sweden).

Quality assurance of treatment plan

The QA of the treatment plan is an essential step before treatment delivery, particularly for multileaf collimator (MLC) modulated beam intensity, which introduces the complexity of the treatment plan and delivery system. To identify discrepancies between the radiation dose calculated by the TPS and the actual dose delivered, the Varian aSi 1200 electronic portal imaging device (EPID) was employed to verify 27 compatible cases (9 H&N and 18 pelvic) from a total of 33 patients, as treatments were performed on different machine with distinct MLC model. EPID-based portal dosimetry was utilized due to its integration into the clinical workflow, providing clinically acceptable QA results while offering high spatial resolution. For sCT evaluation, it is essential to conduct verification under the same clinical workflow used in routine QA. Dark-field correction, flood-field correction, and absolute dose calibration were performed to mitigate the effects of long-term radiation exposure and ensure detector accuracy [27]. The gamma pass rate criteria recommended by AAPM TG-218 [28] at 3%/2 mm and stricter criteria at 2%/2 mm, with a 10% low-dose cut-off, were applied and compared between sCT and CT plans against measurement data. Finally, the percent differences in gamma pass rates between sCT and CT were evaluated.

Patient set-up verification

To accomplish the MRI-only workflow, sCT images served as the reference images for verifying patient set-up prior to treatment by registering with CBCT. This process was assessed using one fraction of CBCT images from 18 cases (10 H&N and 8 pelvic) out of 33 total cases, which were acquired on Varian TrueBeam and Edge systems. These CBCT images were retrospectively rigidly registered to both sCT and CT within the TPS. Auto-matching with a bone threshold of 200 to 1700 HU was performed in all six degrees of freedom. The resulting subtraction of CT-CBCT and sCT-CBCT was calculated and compared for translation and rotation in each degree of freedom.

Results

Image evaluation

The proposed method successfully generated sCT images from the corresponding MRI data, compared with the CT images in the H&N (Fig. 2) and pelvis (Fig. 3) in the axial, coronal, and sagittal views. The sCT images clearly illustrate the software’s capability to generate sCT images that preserve anatomical structures and image contrast among various tissue types, similar to CT images. Regarding the H&N, MRI signal void caused by dental fillings have a limited spatial extent on the sCT, whereas streak artifacts affect image quality over several centimeters away from the filling in the CT (Fig. 2).

Fig. 2.

Representative examples of sCT images for a H&N case. Panels (a1–a3) show the MRI images used for sCT generation. Panels (b1–b2) present the reference CT images. Panels (c1–c3) display the sCT images generated using a deep learning-based software. Panels (d1–d3) illustrate the HU difference maps between the sCT and CT, with the CT serving as the ground truth. The profiles along the red line in (e1) are plotted in (e2) to provide a quantitative comparison

Fig. 3.

Representative examples of sCT images for a pelvis case. Panels (a1–a3) show the MRI images used for sCT generation. Panels (b1–b2) present the reference CT images. Panels (c1–c3) display the sCT images generated using a deep learning-based software. Panels (d1–d3) illustrate the HU difference maps between the sCT and CT, with the CT serving as the ground truth. The profiles along the red line in (e1) are plotted in (e2) to provide a quantitative comparison

The quantitative analysis findings are summarized in Table 2, presenting calculations of DSC, ME, and MAE in comparison to CT. Regarding the geometric evaluation, the mean DSC values for the H&N were 0.99 0.00 for the body, 0.96 0.01 for soft tissue, 0.87 0.03 for bone, and 0.82 0.04 for air. For the pelvis, the mean DSC values were 1.00 0.00 for the body, 0.98 0.01 for soft tissue, and 0.88 0.02 for bone. For specific pelvic subregions, the body DSCs were 1.00 ± 0.00 for prostate, 0.99 ± 0.01 for rectum, and 1.00 ± 0.01 for cervix. The soft tissue DSCs were consistently 0.98 ± 0.00 to 0.98 ± 0.01 across all pelvic subgroups. The bone DSCs were 0.89 0.01, 0.88 0.02, and 0.88 0.01 for the prostate, rectum, and cervix, respectively. These results meet the clinical criteria of 0.8–0.9 followed by AAPM TG-132 [25], indicating a satisfactory geometric agreement between sCT and CT images.

Table 2.

Assessment of sCT in terms of geometric evaluation represented by the dice similarity coefficient (DSC) and image intensity evaluation represented by the mean error (ME) and mean absolute error (MAE) of the H&N, prostate, rectum, and cervix reported in mean 1 standard deviation (S.D.)

Tumor location	Segment	DSC	ME (HU)	MAE (HU)
H&N	Body	0.99 0.00	-8.94 9.50	67.30 6.94
	Soft tissue	0.96 0.01	7.75 7.34	29.91 3.30
	Bone	0.87 0.03	-121.08 34.45	203.57 26.79
	Air	0.82 0.04	118.65 41.65	211.20 34.51
Prostate	Body	1.00 0.00	-7.17 6.88	32.65 4.92
	Soft tissue	0.98 0.00	-0.93 6.42	23.99 5.61
	Bone	0.89 0.01	-61.20 17.65	104.31 13.64
Rectum	Body	0.99 0.01	-11.44 7.04	35.29 2.86
	Soft tissue	0.98 0.01	-2.19 4.10	24.55 1.97
	Bone	0.88 0.02	-86.27 33.43	121.58 16.76
Cervix	Body	1.00 0.01	-6.74 4.15	32.44 3.34
	Soft tissue	0.98 0.01	-0.13 3.35	23.79 1.62
	Bone	0.88 0.01	-62.56 16.03	112.82 9.59

In the evaluation of image intensity, notable differences in HU between sCT and CT images were observed in bone and air (Fig. 2 (d1-d3)). These differences were reflected in the mean ME and MAE values. The mean ME values indicated an underestimation of bone in both H&N and pelvis and an overestimation of air in the H&N across all patients. In terms of absolute HU differences, the mean MAEs were 67.30 6.94 HU for the body and 23.99 5.61 HU for soft tissue in the H&N and 33.64 3.91 HU for the body, and 24.17 3.51 HU for soft tissue in the pelvis. For specific pelvic subregions, the mean body MAE was 32.65 ± 4.92 HU in the prostate, 35.29 ± 2.86 HU in the rectum, and 32.44 ± 3.34 HU in the cervix. The corresponding soft tissue MAEs were 23.99 ± 5.61 HU in the prostate, 24.55 ± 1.97 HU in the rectum, and 23.79 ± 1.62 HU in the cervix. The highest mean MAEs were found in bone, with values of 203.57 26.79 HU for the H&N and 112.92 15.24 HU for the pelvis (104.31 13.64 HU for the prostate, 121.58 16.76 HU for rectum, and 112.82 9.59 HU for cervix). Additionally, the mean MAE for air in the H&N reached 211.20 34.51 HU.

Dosimetric evaluation

The percent dose differences of the DVH parameters for the H&N and pelvis are presented in Fig. 4 (a-d), in which overall dose differences demonstrated a high agreement between sCT and CT, with differences within 2% for both PTV and normal organs. Specifically, for the case of the H&N, the mean dose differences were 0.04 0.18% for the PTV and 0.76 0.55% for normal organs. In the pelvis, the mean dose differences were 0.25 0.20% in the PTV and 0.35 0.27% for normal organs. For individual pelvic sites, the PTV mean dose differences were 0.25 0.08% for the prostate, 0.34 0.34% for the rectum, and 0.38 0.15% for the cervix. Corresponding, the mean dose differences for normal organ were 0.31 0.22%, 0.39 0.36%, and 0.33 0.28% for the prostate, rectum, and cervix, respectively. It is important to note that assessment did not include the bladder and rectum due to their changes in shape and volume during image acquisition. These variations resulted in dose differences between sCT and CT that were unrelated to the quality of sCT.

Fig. 4.

Boxplots show the dosimetric evaluation of sCT compared to CT for H&N (a), prostate (b), rectum (c), and cervix (d). The percent dose difference was calculated as . DVH parameters for the PTV and normal organs were evaluated. It should be noted that the evaluation of the pelvis excluded the bladder and rectum due to its variation in shape and volume during image acquisition

In the comparison of dose distributions between sCT and CT using gamma analysis, Table 3 shows the mean gamma pass rates at the criteria of 3%/2 mm, 2%/2 mm, and 1%/1 mm, with a 10% low dose cut-off. The mean gamma pass rates exceeded 92% for the H&N and 96% for the pelvis for all criteria.

Table 3.

Gamma pass rates using a 10% low dose cut-off comparing CT and sCT

Gamma criteria	Gamma pass rate 1 S.D. (%)
	H&N	Prostate	Rectum	Cervix
3%/2 mm	96.34 1.31	99.80 0.27	98.57 0.73	98.85 0.38
2%/2 mm	95.06 1.40	99.67 0.38	98.04 0.97	98.24 0.52
1%/1 mm	92.09 2.02	99.23 0.65	96.29 1.48	96.91 0.86

Treatment plan QA evaluation

The comparisons of plan dose differences using gamma pass rate at the criteria of 3%/2 mm and 2%/2 mm with a 10% low dose cut-off are shown in Fig. 5 ((a) for the H&N and (b) for pelvis). The mean gamma pass rates reached 99% for both H&N and pelvis at the 3%/2 mm and 97% for H&N and 94% for pelvis at the 2%/2 mm. Minimal differences were observed between sCT and CT, with mean gamma pass rate differences of 0.01 0.04% for H&N and 0.05 0.08% for pelvis at 3%/ 2 mm and 0.05 0.06% for H&N and 0.12 0.16% for pelvis at 2%/2 mm.

Fig. 5.

Percent gamma pass rate differences between sCT and CT plans relative to measurement, calculated a , where GPR denotes gamma pass rate. Panel (a) presents the differences in the H&N at 3%/2 mm and 2%/2 mm criteria. Panel (b) presents the differences for the pelvis at 3%/2 mm and 2%/2 mm criteria

Patient set-up verification

The mean differences between image registrations for patient set-ups verification using CBCT and sCT compared to CT for each degree of freedom are presented in Table 4. Representative rigid registrations in the axial, coronal, and sagittal planes for CT-CBCT and sCT-CBCT of the H&N and pelvic regions are shown in Additional file 1: Fig. S1 and S2, respectively. In the case of H&N, the mean differences between sCT-CBCT and CT-CBCT were 0.13 mm in all translations and 0.06 in all rotations. For the pelvis, the mean differences between sCT-CBCT and CT-CBCT were 0.17 mm in all translations and 0.13 in all rotations. The absolute differences in translation and rotation values for all assessed cases were below 1.5 mm and 0.8°, respectively.

Table 4.

Differences in translation and rotation between CT-CBCT and sCT-CBCT registrations

Location	Direction	Translation (mm)		Rotation ()
		Mean 1 S.D.	(minimum, maximum)	Mean 1 S.D.	(minimum, maximum)
H&N	Right-left/pitch	0.13 0.40	(-0.7, 0.8)	0.01 0.40	(-0.4, 1.00)
	Posterior-anterior/yaw	-0.13 0.43	(-0.9, 0.40)	0.02 0.23	(-0.20, 0.60)
	Inferior-superior/roll	-0.06 0.77	(-1.50, 1.50)	-0.06 0.10	(-0.20, 0.10)
Pelvis	Right-left/pitch	0.11 0.25	(-0.20, 0.50)	0.02 0.10	(-0.10, 0.20)
	Posterior-anterior/yaw	-0.17 0.25	(-0.40, 0.40)	-0.13 0.28	(-0.80, 0.10)
	Inferior-superior/roll	-0.14 0.47	(-1.20, 0.20)	0.01 0.08	(-0.10, 0.20)

Discussion

To enable the implementation of sCT in the MRI-only workflow for VMAT, a reliable sCT solution is required to eliminate CT-MRI registration uncertainties and enhance spatial accuracy. sCT also streamlines the workflow by eliminating the need for separate CT scans, thereby reducing radiation exposure, patient time, and imaging-related costs, while leveraging MRI’s superior soft tissue contrast to support high-precision radiotherapy planning [29]. This study evaluated the performance of sCT generated by an approved deep learning-based software, focusing on image and dosimetric accuracy, treatment plan verification, and patient set-up for H&N and pelvic regions, compared to CT as the reference.

Geometric accuracy assessment showed that sCT images provided lower DSC values in bone and air regions due to CT and MRI modality differences, aligning with prior studies [3, 9, 10, 13, 17, 30]. Despite this, overall DSC values remained with clinical reference of 0.8–0.9 [25], consistent with previous studies on the H&N [9] and pelvis (prostate) [19]. These differences are potentially contributed by misregistration between image sets [17, 30] and anatomical differences between CT and MRI acquisitions. To address anatomical variations between CT and MRI, particularly in the pelvic region, this study applied DIR to account for differences in patient shape and organ position, aiming to provide a more accurate evaluation of sCT performance [31]. The DIR method, which accounts for internal organ motion and external positioning, achieved lower MAE and higher DSC compared to RIR, demonstrating improved dosimetric accuracy while minimizing uncertainties caused by organ and body misalignment between planning CT and MRI acquisitions in the H&N and pelvic regions. However, it exhibited limitations in areas with substantial anatomical variability, such as the bladder and rectum, where deformation accuracy was reduced due to differences in organ filling and shape. In comparison to the study by Fridström et al. [31], who used rigid registration for prostate sCT evaluation and reported MAEs of 71 HU for the body and 126 HU for the femoral heads, our approach incorporating both rigid and deformable registration achieved a substantially lower mean body MAE of 32.65 ± 4.92 HU in the prostate. This highlights the potential benefit of DIR in reducing intensity discrepancies and improving geometric alignment in MRI-only workflows. Using the same deformable registration method, our results are comparable to those by Palmér et al. [9] in the H&N (mean MAEs: 67 14 HU for body, 38 6 HU for soft tissue, 195 27 HU for bone, and 198 68 HU for air). For the pelvis, our findings also align with Cronholm et al. [19] for prostate (mean MAEs: 39.6 6 HU for body, 24.2 3 HU for soft tissue, and 95.0 12.8 HU for bone). Additional comparison with Boulanger et al. further supports that our body MAEs fall within the 65 to 131 HU for the H&N and 27 to 65 HU for the pelvis, both within the typical range reported for deep learning-based sCT generation methods [32]. For a more detailed MAE evaluation, dividing the CT intensity range into discrete intervals, such as 20 HU bins, could provide an alternative method to assess error variation across tissue types based on their HU values especially relevant to the H&N region [6].

Additionally, accurate differentiation between bone and air in MRI presents challenges, resulting in an indistinct representation of these voxels in sCT data [9, 17]. Bone-specific applications may be improved by using alternative sequences in sCT model training, such as UTE [33] or ZTE [34]. Notably, the H&N region, characterized by more prominent bone-air interfaces, presents more challenges for the sCT-generated algorithm, registration, and postures positional changes during image acquisitions (e.g., flexion or extension of the neck and jaw bone), exhibiting higher HU errors compared to pelvic region.

The dosimetric comparisons revealed mean dose differences between CT and sCT of within 1% for both H&N and pelvic regions, consistent with previous deep learning-based sCT publications [9, 14, 21, 35–37]. These differences are within the 2% tolerance recommended for MR-based dose calculations [1, 38]. Considering achievable dosimetric accuracy in external beam radiotherapy workflow, accounting for a combined uncertainty of 5% encompassing beam calibration, relative dosimetry, dose calculation, and dose delivery [39, 40], our observed mean dose differences (0.76% for the H&N and 0.35% for the pelvis) represent a minimal contribution to the overall error. These may indicate that megavoltage photon dose calculations are relatively insensitive to small HU variations [30]. Larger dose differences were particularly pronounced in the bladder and rectum due to variable organ filling and shape changes between image acquisitions [41], which even deformable registration could not fully compensate for. Therefore, the blader and rectum were excluded from the final dosimetric analysis.

Gamma analysis further validated the dose distribution and confirmed the dosimetric reliability of sCT, with global pass rates exceeding clinical criteria of 95% at 3%/2 mm [28]. Using the stricter 2%/2 mm criteria, gamma pass rates remained above 90% [42] for both H&N and pelvic regions. Our findings are in line with Singhrao et al. [3], who reported global and local gamma pass rates in the H&N of 92.7% (2%/2 mm). For the pelvis, our study aligns with the findings of Cronholm et al. [19], who reported mean gamma pass rates of 99.9% (2%/2 mm). Additionally, Fridström et al. [31] validated sCT for both VMAT photon and two field intensity modulated proton therapy (IMPT) in prostate cancer patients, achieving median global gamma pass rates of 99.4% (VMAT) and 95.8% (IMPT), and local gamma pass rates of 97.7% (VMAT) and 94.9% (IMPT), using 2%/2 mm. In comparison, our study demonstrated slightly better results than previously reported values under the stringent 1%/1 mm, compared to commercial bulk density methods studies by Gonzalez-Moya et al. [43] (73.2% for brain and 84.7% for prostate) and Farjam et al. [44] (90.6% for prostate). To facilitate clearer comparison, an overview of clinical evaluation criteria with corresponding results from this study for clinical readiness and key metrics from previous studies are presented in Additional file 1: Tables S1 and S2, respectively.

When benchmarked against other clinically implemented methods, such as the multi-planar cGAN approach by Hsu et al. [45] for MR-guided adaptive prostate radiotherapy, our results demonstrated comparable dosimetric accuracy. Hsu et al. reported body MAEs of 30.1 ± 4.2 HU, mean dose differences < 1%, and gamma pass rates of 99% (2%/2 mm–1%/1 mm), while our pelvic study showed body MAEs of 33.64 ± 3.91 HU, mean dose differences within 0.25% for the PTV and 0.35% for OARs, and gamma pass rates of 96–99% (3%/3 mm–1%/1 mm). These findings highlight the potential of commercial sCT solutions to achieve clinically acceptable accuracy in supporting both routine MRI-only workflows and adaptive radiotherapy applications, thereby facilitating broader clinical implementation.

To maintain confidence in patient safety, the pre-treatment QA plan was evaluated. The deviations between the treatment plan and the delivered plan were compared between sCT and CT plans. The average gamma pass rates achieved the clinical criteria, exceeding 95% for 3%/2 mm [28]. Under the more stringent 2%/2 mm, gamma pass rates remained high, achieving over 97% for head and neck (H&N) and 94% for pelvic cases. Negligible absolute differences observed between sCT and CT plans, with mean absolute differences 0.05% for H&N and 0.12% for pelvis across all criteria. However, the resulting gamma pass rates with EPID were slightly higher than those obtained with devices such as Delta4 or ArcCHECK, largely due to differences in detector resolution and geometry [27]. Further investigation using multiple QA platforms may help assess reproducibility across devices, particularly in anatomically complex or high-gradient regions.

Accurate patient positioning is crucial for MRI-only workflows, as sCT serves as the reference for patient position verification before beam delivery. Our findings showed negligible differences between sCT-CBCT and CT-CBCT registration (0.17 mm in translation and 0.13 for rotation), indicating sCT’s reliability for image-guided radiotherapy. These results were favorable compared to previous reports using DRR or CBCT rigid registration studies [23, 43, 46, 47], further supporting the clinical applicability of sCT in daily setup workflows. As an additional consideration, when sCT is intended for use in patient position verification such as with 2D kV images, evaluating these registration techniques is essential to ensure accurate patient alignment.

While previous studies [9, 19, 20] have focused primarily on aspects such as image and dosimetric accuracy, as well as patient setup verification, treatment plan QA has been excluded, despite its crucial role in ensuring that sCT can support safe and reliable dose delivery. A comprehensive evaluation of the sCT workflow should incorporate these additional steps to enable more robust validation and facilitate its clinical implementation.

Additionally, pre-assessment steps should be performed prior to image and dosimetric analyses as integral part of clinical sCT implementation for MRI-only workflows. These steps include verifying sCT DICOM data transfer. Followed by a thorough visual inspecting the sCT for anatomical abnormalities to ensure that the MR coil does not cause body distortions. Such distortions can be mitigated by careful positioning the coil or by using coil bridge for pelvic region to prevent deformation of patient’s body. In addition, motion artifacts that may affect sCT accuracy should be assessed, and sufficient sCT coverage of the patient’s body in both in-plane and through-plan directions (SI) must be ensured, with a margin of at least 5 cm, for accurate dose calculation [48]. Furthermore, MRI system QA focused on geometric distortion across large FOV is required [23], particularly in pelvic region, where the distortion tends to increase toward the periphery of FOV. Employing high acquisition bandwidth and vendor-specific distortion correction has been shown to minimize MRI geometric distortion, as evidenced by studies with negligible dosimetric impact [20, 49, 50]. Nevertheless, it is imperative to characterize potential MRI distortions that could affect dosimetric accuracy, which is unrelated to the performance of the sCT-generated algorithm [21]. The MRI system used in this study was previously validated for geometric accuracy as reported in [24]. The mean distortion caused by gradient nonlinearity and B0 inhomogeneity was less than 2 mm within a 42 × 42 cm² FOV at all band width varying from 130 to 840 Hz/pixel, falling within the American College of Radiology (ACR) guideline [51], with resulting dose errors within 2% for relevant target and organ structures.

Several commercial solutions for sCT generation are now available to support MRI-only radiotherapy workflows. The MR for Calculating ATtenuation (MRCAT) (Philips Healthcare) employs a bulk density approach using Dixon MRI sequences with fixed parameters, followed by AI-based HU assignment to segmented tissues. It has been validated for brain and pelvic applications [52–54]. Siemens Healthineers first developed bulk density-based sCT using Dixon sequences [43] and subsequently introduced a deep learning-based solution trained on a dual-network architecture consisting of CNNs and conditional generative adversarial networks (cGANs) [55]. Despite regulatory approval and encouraging results, site-specific validation remains essential, as sCT performance may vary depending on anatomical site, MRI system, and TPS configuration.

To improve the accuracy of deep learning-based sCT generation in the future, particular attention should be given to enhancing the bone–air interface. Addressing variability in MRI systems and acquisition protocols between training and clinical data is essential. Although manufacturers typically define software applicability limits, future strategies may include expanding training datasets through federated learning or fine-tuning pre-trained models to specific scanners and sequences using transfer learning [23]. Enhancing the generalizability and robustness requires training on larger and more diverse patient cohorts encompassing various anatomical sites and clinical scenarios to better address site-specific variations [7]. To facilitate the clinical adoption of the MRI-only workflow, this study applied the established clinical guidelines as evaluation criteria, ensuring that image quality and dosimetric accuracy meet the required standards for safe and effective treatment planning and delivery. However, the development of standardized consensus guidelines remains critical for regulating the commissioning and QA processes for MRI-only workflows, particularly in the absence of a gold-standard CT reference [7, 22, 23].

This study provided a comprehensive evaluation encompassing image and dosimetric accuracy, treatment plan QA, and patient set-up verification, ensuring robust validation of the sCT solution for clinical implementation. The results across the H&N and pelvic regions were pooled to demonstrate the generalizability of the MRI Planner software, while site-specific evaluations (prostate, rectum, and cervix) were also conducted and reported separately to capture anatomical and disease-related variations commonly encountered in routine clinical practice. This thorough approach encourages confidence in the safe and effective use of sCT in the MRI-only workflow, highlighting the importance of implementing a rigorous QA process in real-world clinical settings.

While the overall findings are promising, the limited number of cervical cancer cases may restrict the depth of subgroup evaluation. Future study should expand the patient cohort to include a boarder range of anatomical variations and disease stages. This would enhance the generalizability and clinical robustness of the deep learning model, supporting safe clinical implementation across various treatment scenarios. Additionally, further work is needed to evaluate the model’s applicability in high-precision techniques such as stereotactic radiosurgery (SRS) and stereotactic body radiotherapy (SBRT).

Conclusion

The implementation of commercially available deep learning-based software (MRI Planner™) demonstrates the capability to generate sCT images with geometric and intensity accuracy comparable to conventional CT. This advancement streamlines MRI-only radiotherapy by enabling accurate dose calculation, treatment plan quality assurance (QA), and patient set-up verification. These capabilities support the potential adoption of an MRI-only workflow for head and neck (H&N) and pelvic VMAT treatments, offering an effective alternative to traditional CT-based approaches. Furthermore, the findings highlight the importance of treatment site-specific and comprehensive evaluations across the entire clinical workflow to ensure the safe, reliable, and clinically robust implementation of deep learning-based sCT in MRI-only radiotherapy.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

CT

Computed Tomography

CBCT

Cone Beam Computed Tomography

CNNs

Convolutional Neural Networks

DSC

Dice Similarity Coefficient

DVH

Dose Volume Histogram

FOV

Field of View

H&N

Head and Neck

HU

Hounsfield Units

MRI

Magnetic Resonance Imaging

MAE

Mean Absolute Error

ME

Mean Error

MU

Monitor Unit

OAR

Organs at Risk

PTV

Planning Target Volume

QA

Quality Assurance

RT

Radiotherapy

sCT

Synthetic Computed Tomography

TPS

Treatment Planning System

VMAT

Volumetric Modulated Arc Therapy

Author contributions

P.E. Writing original draft, Visualization, Validation, Methodology, Investigation, Formal analysis, Data curation, and Conceptualization. C.P., S.S., L.W., C.J., T.S., P.C., S.A., S.K. Writing – Review & Editing, Methodology, Formal analysis and Supervision. C.P. and S.K. Data curation, Validation, Investigation, and Conceptualization. S.K. Project administration.

Funding

Open access funding provided by Mahidol University

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This study was conducted following the approval of the Institutional Review Board (IRB) of Ramathibodi Hospital, Bangkok, Thailand (Approval No. MURA2023/860), in accordance with the Declaration of Helsinki, which waived the requirement for written informed consent due to the retrospective nature of the study.

Consent for publication

Not applicable.

Generative AI and AI-assisted technologies in the writing process

During the preparation of this work the author(s) used ChatGPT in order to improve readability and language. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the publication.

Competing interests

The authors declare no competing interests.