SynthRAD2025 Grand Challenge dataset: generating synthetic CTs for radiotherapy

• The listed keywords are image synthesis, artificial intelligence, CT, MR, CBCT, and deep learning.Source: Potential applications · p. 2

• CBCT artifacts limit dose-calculation accuracy, while MRI lacks the electron-density information required by dose algorithms.MRI also offers soft-tissue contrast and functional imaging without ionizing radiation; CBCT is commonly used for daily pretreatment alignment.Source: §1 · p. 3
• MRI-only workflows are of interest because MRI provides soft-tissue contrast without ionizing radiation but cannot directly provide electron density for dose calculations.MR-guided proton therapy remains under research and development, with clinical availability described as a possibility.Source: §1 · p. 3
• Synthetic CT generation from CBCT and MRI is being explored to support accurate dose calculations and address the shortage of public paired datasets.sCTs have also been used to facilitate organ-at-risk and target-volume auto-segmentation.Source: §1 · p. 3
• SynthRAD2025 adds paired CBCT-to-CT and MRI-to-CT data for head-and-neck, thoracic, and abdominal cancer patients from five European university medical centers.The dataset is intended to support development, validation, and benchmarking of sCT algorithms.Source: §1 · p. 4

• Task 1 benchmarks MRI-to-CT conversion for MR-only and MR-guided photon and proton radiotherapy, while Task 2 targets CBCT-to-CT translation for daily adaptive workflows.Source: §2.1 · p. 4
• The dataset covers head-and-neck, thoracic, and abdominal cancer patients treated at five European university medical centers.The centers are UMC Groningen, UMC Utrecht, Radboud UMC, LMU University Hospital Munich, and University Hospital of Cologne.Source: §2.1 · p. 4
• The 2362 cases comprise 890 MRI-CT pairs for Task 1 and 1472 CBCT-CT pairs for Task 2.Eligibility required external-beam radiotherapy at a participating center and available imaging from a specified anatomical region; no further age, sex, or tumor-characteristic restrictions were imposed.Source: §2.1 · p. 4
• The planned training, validation, and test split is 65/10/25%, with slight variation possible by center, task, and anatomy; initially, only training data are publicly released.Table 1 organizes case counts by set, task, anatomical site, and center.Source: §2.1 · p. 4, Table 1 · p. 5

2.2 Task 1

• Task 1 includes MRI acquired for treatment planning at centers A, C, and D and low-field MR-Linac imaging at center B, paired with planning CTs.MRI sequences include T1-weighted gradient echo and balanced steady-state free-precession.Source: §2.2 · p. 6

2.2.1 Head-and-Neck (1HN)

• The head-and-neck subset contains 340 MRI-CT pairs from centers A, C, and D; center C's diagnostic scans have a small field of view and different immobilization, making them challenging for sCT generation.Similar immobilization and table tops at centers A and C provide similar patient positioning on MRI and CT, improving pre-processed data quality.Source: §2.2.1 · p. 6

2.2.2 Thorax (1TH)

• Only two of five institutes had suitable thoracic MRIs, yielding 280 images with equal contributions from centers A and B.Acquisition parameters for thoracic MRI and CT are listed in Table 3.Source: §2.2.2 · p. 7, Table 3 · p. 8

2.2.3 Abdomen (1AB)

• The abdominal subset contains 270 MRI-CT pairs from centers A, B, and C, including 95 cases from A and 30 from C; center B contributed 140 cases.Abdominal MRI and CT acquisition parameters are listed in Table 4.Source: §2.2.3 · p. 9, Table 4 · p. 9

2.3 Task 2

• 1496 CBCT-CT pairs were available across all five centers and all anatomical regions, acquired on three treatment-machine/CBCT-system types.The section describes this coverage as representing many clinically used CBCT scanners and acquisition protocols.Source: §2.3 · p. 10

2.3.1 Head-and-Neck (2HN)

• Head-and-neck CBCTs came from Elekta linacs at centers A–D, a Varian linac at center E, and an IBA proton therapy machine at center D.Table 5 lists CBCT and CT acquisition parameters.Source: §2.3.1 · p. 10, Table 5 · p. 11

2.3.2 Thorax (2TH)

• Thoracic CBCTs came from Elekta linacs at centers A–D, an IBA proton cyclotron at center D, and a Varian linac at center E.Table 6 lists detailed acquisition parameters.Source: §2.3.2 · p. 12, Table 6 · p. 12

2.3.3 Abdomen (2AB)

• Abdominal CBCTs were predominantly acquired on Elekta and Varian linacs, with only a minimal number of proton-therapy patients at center D.Table 7 presents CBCT and corresponding CT acquisition parameters.Source: §2.3.3 · p. 13, Table 7 · p. 13

2.4 Pre-processing

• The workflow aimed to harmonize image parameters, anonymize images, reduce file size, generate patient outlines, and prepare data for synthetic CT evaluation.Preprocessing code is publicly available.Source: §2.4 · p. 14

2.4.1 Rigid registration

• MRIs and CBCTs were rigidly registered to their planning CTs using Elastix, with task- and anatomy-specific parameter files.The final parameter files are included in the public repository.Source: §2.4.1 · p. 15

2.4.2 Defacing

• Visible facial structures were automatically defaced using TotalSegmentator-derived skull and brain masks to define the facial region, which was overwritten with background intensity.The algorithm was reported robust to positioning and orientation changes and required no manual corrections; Figure 2 illustrates the mask points and overwritten region.Source: §2.4.2 · p. 15, Figure 2 · p. 15

2.4.3 Resampling

• All MRI, CBCT, and CT images were resampled to a consistent voxel spacing of 1 × 1 × 3 mm.Source: §2.4.3 · p. 15

2.4.4 Outline Segmentation

• Patient-outline masks were generated by histogram thresholding and morphological erosion and dilation, with thresholds manually tuned for some patients.The final dilation includes surrounding air so models must reconstruct the patient outline; automated masks can be inaccurate and have variable margins, and manual correction was infeasible at this dataset scale.Source: §2.4.4 · p. 16

2.4.5 Cropping; 2.4.6 File conversion

• Images were cropped to a bounding box extending 10 pixels beyond the patient-outline mask, then compressed as .mha files with INT16 pixel data.Source: §2.4.5 · p. 16, §2.4.6 · p. 16

2.4.6 Deforming CT to MRI/CBCT

• Deformed CTs match the input MRI or CBCT anatomy to reduce anatomical differences for image-similarity and dose-accuracy evaluation.To avoid bias toward paired-training approaches, deformed CTs are withheld from training and released only with validation and testing data.Source: §2.4.6 · p. 16

2.5 Data validation

• All datasets received institutional visual checks for facial-feature removal, and overview images showed central axial, sagittal, and coronal slices with image overlays for registration review.Because checks were limited to three planes per image and patient, the overviews are unsuitable for image-intensity quantification.Source: §2.5 · p. 16
• Visual review found variable arm positions, occasional metal artifacts, and some thoracic-abdominal region overlap across centers.Automatic body-mask thresholding could include couch structures or exclude lung regions, while dilation margins varied between patients and datasets.Source: §2.5 · p. 17
• Limited-field-of-view MRIs in center C’s 1HN subset made rigid and deformable registration particularly challenging, and these cases may challenge sCT generation.Some center B 1AB and 1TH outline masks were cropped inferiorly-superiorly because of MRI intensity variation and edge artifacts; cropping could leave artifacts or remove regular slices.Source: §2.5 · p. 17

3.1 Data structure and file formats

• Training folders are split into MRI Task 1 and CBCT Task 2, then by head-and-neck, thorax, and abdomen, with a unique alphanumeric case code for each patient.Figure 4 depicts the task-and-anatomy folder split.Source: §3.1 · p. 17, Figure 4 · p. 18
• Each patient folder contains the input image, corresponding CT, and patient-outline mask; validation and test folders will also include deformed CTs.Patient overviews and acquisition-parameter spreadsheets are provided in each region’s overview directory.Source: §3.1 · p. 17
• Data from centers A, B, C, and E is provided under a CC-BY-NC 4.0 license.Source: §3.1 · p. 18
• Center D data is usable only while the challenge is active; afterward, downloads are deactivated and the data must be deleted.Participants request access through the SynthRAD2025 challenge website.Source: §3.1 · p. 18
• Table 8 summarizes included files, release dates, and dataset download links for training, validation, and test sets.Source: Table 8 · p. 18, §3.1 · p. 18

3.2 Usage notes

• Compressed .mha images can be read, written, and modified with ITK; SimpleITK provides simplified interfaces for several programming languages.The preprocessing scripts include SimpleITK examples, and GUI viewers include 3DSlicer, ITK-SNAP, and vv.Source: §3.2 · pp. 18–19

4 Discussion

• SynthRAD2025 is presented as a large-scale, multi-center, publicly available resource for developing, validating, and benchmarking MRI-to-CT and CBCT-to-CT algorithms.The discussion frames the dataset as addressing a critical gap in sCT research.Source: §4 · p. 19

4.1 Implications for adaptive radiotherapy

• The dataset may improve dose-calculation accuracy in MR-guided and CBCT-guided adaptive radiotherapy, with particular relevance to proton therapy.It covers head-and-neck, thorax, and abdomen across institutions, supporting development and benchmarking for photon and proton radiotherapy.Source: §4.1 · p. 19

4.2 Strengths of the dataset

• Five-center, almost 2,400-case coverage and standardized quality-controlled preprocessing make SynthRAD2025 a substantial, publicly available sCT resource.It separates MRI-to-CT and CBCT-to-CT tasks and releases preprocessing code and parameter files.Source: §4.2 · p. 19

4.3 Limitations and challenges

• Protocol heterogeneity, incomplete patient characteristics, and possible preprocessing effects limit homogeneous comparisons and whole-dataset subgroup analyses.Training data omit deformed CTs, requiring users of registration-dependent models to perform and validate deformable registration themselves.Source: §4.3 · p. 20

4.4 Future Directions

• Future work should test computational efficiency, robustness, and clinical feasibility in real-time workflows, while considering additional regions, populations, and modalities.The discussion specifically identifies extremities, pediatric patients, ultrasound, and PET as possible expansions.Source: §4.4 · p. 20

5 Conclusion

• SynthRAD2025 offers a comprehensive, publicly available dataset for synthetic CT generation in radiotherapy.The conclusion presents it as a resource for the radiotherapy research community.Source: §5 · p. 20
• Dataset users should account for its limitations and aim to develop robust, generalizable, and clinically feasible algorithms.Source: §5 · p. 20
• Releasing the validation and test sets after the challenge will further support users in validating and refining their approaches.Source: §5 · p. 20