COBRA2026: a large-scale multicenter pelvic CBCT projection dataset

The COBRA2026 dataset is a large-scale, multicenter dataset of raw cone-beam computed tomography (CBCT) acquisitions for radiotherapy. It was created to support the development and reproducible benchmarking of conventional and deep learning-based methods for CBCT reconstruction and image correction. A dataset paper with a detailed description of the dataset is available on arXiv: 📝 arxiv.org/abs/2607.20037 📝 Each case includes measured CBCT projections acquired during routine image-guided radiotherapy, the corresponding acquisition geometry and metadata, a planning CT deformably registered to the daily CBCT anatomy, and simulated projections generated from the aligned CT using the recorded acquisition geometry. The COBRA2026 dataset contains 867 cases from six participating centers: LMU Klinikum Munich (Germany) University Hospital Cologne (Germany) UMC Utrecht (The Netherlands) Lyon (France) Medical University Graz (Austria) Medical University Vienna (Austria). The cases are divided into training, validation, and test sets as follows: Dataset split Center A Center B Center C Center D Center F Center G Total Training 120 107 120 120 92 133 692 Validation 9 8 9 9 7 10 52 Test 21 19 21 21 17 24 123 Total 150 134 150 150 116 167 867 Each case has a unique identifier containing the center and a three digit case ID (e.g. C231). Download links: Due to the large dataset size the data is hosted on a institutional repository and can be downloaded via the links below. The uploaded .zip file only contains overview .pngs for each case in the training set. ⚠️ Please note: No registration/account is required to download the files. Download links will only be functional after the release date of the respective datasets. Please do not request an account for the data repository. Set Download Link Size Release Date Training Download Training Set ~ 750 GB 13.07.2026 Validation Download Validation Set ~ 60 GB 01.03.2032 Test Download Test Set ~ 140 GB 01.03.2032 Files included in the COBRA2026 dataset Each case directory in the released COBRA2026 dataset contains the files listed below. ⚠️ Important: For the COBRA2026 challenge only specific files are available to algorithms on the grand-challenge.org platform. Please visit the challenge website for further details: cobra2026.grand-challenge.org/dataset/ File Description projections.mha Measured raw CBCT projection images geometry.xml Cone-beam acquisition geometry in RTK format projections_simulated.mha Cone-beam projections simulated from the deformed planning CT ct_original.mha Original planning CT without registration or air-cavity correction ct_def.mha Deformed planning CT with the full available field of view ct_def_masked.mha Deformed planning CT masked to the CBCT field of view cbct_clinical.mha Clinically reconstructed CBCT cbct_rtk.mha Reconstructed CBCT (using RTK) fov_cbct.mha CBCT field-of-view mask fov_cbct_nocouch.mha CBCT field-of-view mask with the patient couch removed metadata.yaml Acquisition and imaging parameters extracted from the CT, projection headers, and vendor files reconstruction.yaml Acquisition and reconstruction metadata extracted from the Elekta .INI files Scan.xml Acquisition, geometry, and reconstruction metadata for Varian scans Calibrations/ Acquisition-specific calibration and correction files for Varian scans overview_<ID>.png Multi-panel quality-control figure showing the principal images and representative measured and simulated projections Pre-processing pipeline The dataset was generated from exported raw data using an open-source pre-processing pipeline with the following major steps: Data conversion CBCT baseline reconstruction CT-CBCT deformable registration Projection simulation Postprocessing The preprocessing code is available in the COBRA2026 preprocessing repository. A detailed description of the pre-processing is available in the dataset paper License The COBRA2026 dataset is released under the Creative Commons Attribution–NonCommercial 4.0 International license. The data may be shared and adapted for non-commercial purposes provided that appropriate attribution is given.