Amsterdam IMAging and Clinical GliOma Dataset; IMAGO Authors: Ivar J H G Wamelink, Alle Meije Wink, Niels Verburg, Roelant S Eijgelaar, Emmanouil Koltsakis, Marcus Cakmak, Maarten Balder, Henk J M M Mutsaerts, Philip de Witt Hamer, Mathilde Kouwenhoven, Pieter Wesseling, Frederik Barkhof, Vera C Keil Published: 2026 ABSTRACT AbstractThe Amsterdam University Medical Centers IMAging in GliOma (IMAGO) dataset contains 1,700 adult-type diffuse glioma patients with structural MRI, histological, molecular, and survival information. Its first release includes 500 patients with preoperative structural MRI (T1w pre-contrast, T1w post-contrast, T2w, and T2-FLAIR). Postoperative MRI with the same sequences as the preoperative MRI are included for patients scanned within 14 days of surgery (381 patients). Pre- and postoperative apparent diffusion coefficient (ADC; derived from diffusion-weighted MRI; 497/500 and 377/381 respectively) and relative cerebral blood volume (rCBV; derived from dynamic susceptibility contrast MRI; 269/500 and 24/381 respectively) images are provided. All sequences are registered to MNI standard space and skull stripped. Histological and molecular data include IDH (500/500) and 1p/19q codeletion status (170/500), and WHO CNS5 (2021) tumor grade (500/500). Preoperative and postoperative nnUnet-based tumor region segmentations with human quality check are available for a subgroup of cases. The dataset is usable to create synthetic images, tumor segmentations, and to predict histological and molecular features in adult-type diffuse glioma.Subject terms: CNS cancer, Brain imaging, Brain, CNS cancer Background & Summary Research involving patients with adult-type diffuse glioma has expanded significantly, driven by a convergence of recent technological and methodological advancements. One key contributor has been the evolution of magnetic resonance imaging (MRI), which has enabled a more detailed and multi-parametric characterization of brain tumors, including structural, functional, and metabolic information1–4. The increasing availability of high-resolution imaging data has, in turn, facilitated the application of artificial intelligence (AI) and machine learning approaches to neuro-oncology4,5. AI-driven techniques have shown considerable promise across a wide range of neuro-oncological applications, including automated tumor segmentation, which reduces observer variability and could enhance workflow efficiency6–9. Furthermore, predictive modeling using radiomic and deep learning features extracted from MRI has enabled more accurate survival prognostication and treatment response assessment9. Recent advances also include the development of synthetic imaging methods, such as generating contrast enhanced T1w imaging10,11. Collectively, these innovations are reshaping the landscape of diffuse glioma research, facilitating more personalized, data-driven approaches to diagnosis, prognosis, and therapeutic planning. A critical factor enabling these advancements is the increasing availability of accessible external MRI datasets. Notable examples include the Multimodal Brain Tumor Segmentation (BraTS)12–14 challenge and The Cancer Genome Atlas Glioblastoma Multiforme (TCGA-GBM)15 dataset. More recently, these resources have been complemented by additional datasets such as the Rotterdam Erasmus Glioma Database (EGD)16 and the University of California San Francisco Preoperative Diffuse Glioma MRI (UCSF-PDGM)17 collections. Despite the growth of open-source imaging data, these datasets typically include a limited set of clinical variables and are often restricted to conventional structural MRI sequences, such as T1-weighted (T1w, pre- and post-contrast), T2-weighted (T2w), and T2-FLAIR imaging. This constraint highlights the ongoing need for more comprehensive and multimodal data resources to support continued progress in both imaging research and AI-based approaches in glioma studies. The IMAging in GliOma (IMAGO) dataset18 contains MRI and clinical data of 1,700 subjects with cerebral adult-type diffuse gliomas scanned between 2010 and 2024. This first release of the IMAGO dataset18 consists of 500 patients and contains four different data types: 1. Preoperative and postoperative structural MRI and perfusion and diffusion MRI where available; 2. Histological and molecular labels specifying the WHO CNS5 (2021) tumor types; 3. Clinical parameters; 4. nnUnet-based segmentations of edema/non-enhancing lesion, necrosis, and enhancing tissue. The inclusion of perfusion and diffusion MRI is important as these sequences provide critical information on tumor vascularity and cellularity, respectively. Perfusion MRI helps assess tumor angiogenesis and blood flow, while diffusion MRI informs on tumor cellular density and potential treatment response. Together, they enhance the characterization of glioma biology and are therefore a valuable addition to the dataset. The inclusion of follow-up imaging, an increased number of pre-operative scans, and more extensive longitudinal clinical data has the potential to substantially enhance the performance of predictive AI models and to support clinical decision-making in patients with adult-type diffuse cerebral gliomas. Furthermore, the availability of such data can contribute to improved patient management by enabling more accurate modeling of tumor progression and prediction of growth dynamics over time. In addition, incorporating data from multiple centers can increase the robustness and generalizability of these segmentation and predictive models, making them more trustworthy and easier to implement across diverse clinical settings19,20. Finally, the use of external validation datasets is one of the criteria for the METRICS and CLEAR tools21,22. Methods Data from patients with an adult-type diffuse glioma treated at the Amsterdam University Medical Center location VUmc, The Netherlands, between 2010 and 2024 were retrospectively collected. Patients were included if they were 16 years or older, if isocitrate dehydrogenase (IDH) status was known, and in the case of IDH-mutated patients if 1p/19q codeletion status was known. This study was approved by the institutional review board (VUmc_2021-0385) and all patients gave written informed consent. Additionally, all four structural sequences - pre-contrast T1w, post-contrast T1w (CET1), T2w, and T2-FLAIR; preferably 3D - needed to be available for the preoperative MRI. Subsequently, the surgery date was used to identify the first postoperative MRI. Postoperative MRI scans were only included if the patient was scanned within 14 days of the surgery to avoid confounders from, e.g. early progression. All four structural sequences also had to be available for the postoperative MRI. MRI Structural preoperative MRI were retrospectively collected from the hospital picture archiving and communication system (PACS) (pre-contrast T1w, CET1, T2w, and T2-FLAIR) for all patients. The MRI sequences were automatically labeled using an in-house Digital Imaging and Communications In Medicine (DICOM) labeling tool CINDERELLA23, which uses a minimal number of sequence parameters for classification. T2w images were scanned as 2D multislice. Apparent diffusion coefficient (ADC) maps are provided when available. Relative cerebral blood volume (rCBV) maps are provided when available and are superseded by motion-corrected rCBV maps if available. All rCBV maps were obtained using a 50% Dotarem contrast agent preload followed by the 50% dose dynamic scanning with a 4 mL per second injection rate using the autoinjector to the antecubital vein. The ADC and rCBV maps were post-processed with scanner-integrated software and are, therefore, not post-processed with the same software. Units may differ between images. rCBV maps from before 2022 were leakage corrected, while the presence of leakage correction before 2022 was inconsistent. An overview of all available images can be found in Fig. 1. An overview of all available MRI sequences. Postoperative MRI scans are included only if performed within 14 days after surgery. Images were converted from DICOM to Neuroimaging Informatics technology Initiative (NIfTI) using dcm2niix version v1.0.20230411. All sequences were registered to the CET1 images. ADC maps were registered to the CET1 images via the T2w images. The post-contrast T1w scan was registered to the Montreal Neurological Institute (MNI) 152 atlas, version ICBM 2009a nonlinear, which has a 1 mm isotropic resolution and a size of 197 × 233 x 189 voxels, using Ants version 0.4.2 in Python version 3.9.1824. After registration to the MNI152 atlas, the CET1 images were skull stripped using HD-BET version 1.0. The computed CET1 to MNI152 atlas and brain mask were subsequently applied to all other sequences. Technical details on scanner and sequence settings, such as MRI vendor, scanner model, repetition time, and inversion time (among others) are provided per sequence. Clinical data Clinical information was retrospectively obtained from the electronic patient system. The postoperative MRI scan is only provided if the MRI was performed within 14 days of surgery (arbitrary cut-off to avoid the inclusion of cases showing potentially early progression and to ensure comparability among patients). The number of days the early-postoperative MRI were scanned are additionally marked to identify cases performed within the 72 h window. Tumor enhancement was obtained from the preoperative radiologic reports. All patients had undergone surgery (biopsy or resection) in the period 2010–2021 as part of their routine clinical care and were diagnosed with an adult-type diffuse glioma. Between 2014 and 2016, the genetic characteristics of the tumors were mostly assessed using immunohistochemistry for the IDH R132H mutant protein, in a subset of cases (especially those histologically showing oligodendroglial characteristics) supplemented with loss of heterozygosity (LOH) analysis for the chromosome 1p/19q codeletion status. During this phase, certain IDH-mutant gliomas may not have been detected as such due to limitations inherent to this diagnostic approach. However, approximately 90% of the IDH-mutant gliomas carry the IDH1 R132H mutation and can thus be identified as such by immunohistochemistry. Therefore, the vast majority of cases diagnosed between 2014 and 2016 can be expected to be accurately classified25. Between 2016 and 2019, a dedicated glioma DNA next-generation sequencing (NGS) panel was mostly used to screen for IDH1 and IDH2 hotspot mutations as well as for 1p/19q codeletion. Since 2019, methylome profiling has been applied to analyze these markers26. For patients who underwent surgery between 2010 and 2014, the IDH mutation and 1p/19q codeletion status of their gliomas were determined from tissue obtained at recurrence and analyzed at a later time. For other patients, retrospective assessment was conducted using immunohistochemistry for the IDH1 mutant protein, and in some cases, this was supplemented with loss of heterozygosity (LOH) analysis for 1p/19q. Tumor segmentations All images were automatically segmented using the PICTURE pre- and postoperative segmentation models6,27–30. The preoperative model provides the following labels: (0) background, (1) edema/non-enhancing, (2) necrosis, and (3) enhancing tissue; the postoperative model solely provides the enhancing tumor residue. The postoperative model was only applied if the postoperative MRI scan occurred within three days of the surgery, as instructed by the software developers. Both models used all four structural MRI sequences. The segmentations were visually checked for acceptance by a neuroradiologist (6 years of experience). Data Record IMAGO has been made accessible through EUCAIM on the Health-RI XNAT platform and can be downloaded free-of-charge from the Health-RI website by academic researchers. Access is provided through the EUCAIM platform. By default, the dataset may only be used for internal research/analysis and there is no external interaction with participants. Patients can be downloaded separately, by session i.e. pre- and postoperative, by scan type, and by data format (NIFTI and JSON). The NIFTI files have the study patient ID, scan session and scan type in the name, which allows the downloader to re-create a complete BIDS-compliant directory tree. MRI For all 500 patients, the images are provided as NIfTI files with the following names “D_T1.nii.gz” (T1w precontrast MRI), “D_CET1.nii.gz” (T1w postcontrast MRI), “D_T2.nii.gz” (T2w MRI), “D_FLAIR.nii.gz” (T2-FLAIR MRI), “ADC.nii.gz” (DWI MRI), and “CBV.nii.gz” (DSC MRI) with ‘D’ being the scanning dimension (2D or 3D). The scans were acquired on 23 different scanners from 4 vendors: General Electric (GE), Boston, Massachusetts, USA (DISCOVERY MR750 (3 T), Signa HDxt (1.5 T), and OPTIMA MR360 (1.5 T)), Philips Healthcare, Best, The Netherlands (Achieva (1.5 T and 3 T), Ingenia (1.5 T and 3 T), Ingenia Elition X (3 T), Ingenia Elition S (3 T), Achieva Stream (1.5 T and 3 T), and Intera (1.5 T)), Siemens Healthineers, Erlangen, Germany (MAGNETOM Avanto (1.5 T), Vida (3 T), Sonata (1.5 T), Avanto fit (1.5 T), Aera (1.5 T), Sola (1.5 T), Skyra (3 T), Espree (1.5 T), Verio (3 T), ESSENZA (1.5 T), Symphony (1.5 T)), and Toshiba Medical Systems Corporation, Otawara, Japan (Titan3T (3 T)). Postoperative MRI scans are available for 381/500 patients. Axial 2D ADC and rCBV maps are available for 497/500 and 269/500 patients preoperatively and 377/381 and 24/381 postoperatively. Technical details on scanner and sequence settings, such as MRI vendor, scanner model, repetition time, and inversion time (among others) are provided in an accompanying .json file per sequence. Clinical data Per patient, the following clinical information is provided: sex (male/female), date of preoperative MRI (in days after birth), year of diagnosis, age at preoperative MRI (age at radiology detection), surgery type (biopsy (B), partial resection (P), gross total resection (G); obtained from the first postoperative radiological report), tumor type (glioblastoma IDH-wildtype, astrocytoma IDH-mutant, and oligodendroglioma IDH-mutant and 1p/19q-codeleted), CNS WHO (2021) grade (2, 3, 4), surgery date (days after preoperative MRI), the number of days the postoperative MRI was obtained after surgery, preoperative Karnosfky score (KPS) (250/500 patients), survival (number of days between date of birth and last clinical follow-up), OS (overall survival in days between initial diagnosis and last clinical follow-up), enhancement (1: enhancement; 0: no enhancement), IDH state (wt-IDH and IDH-mutation; 500/500) and 1p/19q state (codeleted, not codeleted, and unknown) (170/500 patients). Clinical data is provided in a CSV file entitled “Clinical_data.csv”. Dataset demographics can be found in Table 1. Table 1Dataset demographics grouped by tumor type.Tumor typeIDH-mutant 1p/19q codeleted oligodendrogliomaIDH-mutant astrocytomaIDH wild-type glioblastomaTotal6072368Sex Male3839226 Female2233142Age at diagnosis (years)46 (±14.2)39 (±11.7)59 (±12.3)Surgery type Biopsy4795 Partial4554232 Gross total111141WHO grade Grade 239410 Grade 321190 Grade 4-12368Number of days preoperative MRI was before surgery24.2 (±32.9)14.3 (±20.6)8.2 (±11.8)Number of days postoperative MRI was after surgery1.6 (±0,9)2.1 (2.15)1.8 (±1.6)IDH state60723681p/19q codeletion state607223 Dataset demographics grouped by tumor type. Tumor segmentations Segmentations are provided for 249 preoperative cases and 190 postoperative cases. Technical Validation All patients were diagnosed with adult-type diffuse gliomas according to the WHO CNS5 (2021) criteria. Only patients with an availability of all four structural MRI sequences (T1w precontrast, CET1, T2w, and T2-FLAIR) were included. MR images were skull stripped for anonymization. Cases were assigned a random number; to ensure anonymity, no key-file was stored. All automatically generated segmentations were verified by an experienced neuroradiologist and not included if incorrect. The data has been directly downloaded from the Health-RI website and assessed to ensure all images were in the same atlas and skull stripped. Bar plots (Fig. 2) and patient survival curves (Fig. 3) were generated, demonstrating survival outcomes that were comparable to or exceeded those reported in previous studies31,32. and showed distributions as reported in other papers. Bar graph of number of patients per surgery type and tumor type. Pane 2c shows the number of patients by age group and tumor type (bins of 10 years). Kaplan-Meier survival curves by IDH mutation, sex, and tumor type. Usage Notes Release 1.0 of the dataset is available through the EUCAIM platform and can be obtained after completing the registration steps. Users first register with EUCAIM (EUCAIM Catalogue) and request access to the IMAGO dataset (EUCAIM Catalogue) by 1. Adding the dataset; 2. Requesting the dataset; 3. Sending the request to the negotiator. Users must then sign a DUA that prohibits patient re-identification and unauthorized onward sharing. After this, they can request an account for the data host by completing the Health-RI request form on the EUCAIM helpdesk page and then download the data18. Alternatively, users may request access by sending an email with the subject line “Request IMAGO database access”, including the following information: 1. First name; 2. Last name; 3. Organization; 4. GDPR or NON GDPR Country; 5. E-mail; 6. Project title; 7. Project description to servicedesk@health-ri.nl. MRI data can be visualized using any platform that supports the NIfTI format, such as ITK-SNAP, while clinical data can be opened with any CSV viewer, such as Excel. Ethics statement The dataset was developed in accordance with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. Patients provided written informed consent to use their data and to publish the data after anonymization. The dataset building process and online publication was supported and approved by the hospitals’ data privacy officer and research data management team. Acknowledgements This work was supported by the HANARTH fonds call 2022 (GLIOCARE: Glioma Imaging Omitting Contrast through Artificial intelligence and Risk Evaluation) and the EUCAIM 2024 open call. Author contributions The first draft of the manuscript was written by I.J.H.G.W. and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. Conceptualization: I.J.H.G.W., F.B., V.C.K.; Data curation: I.J.H.G.W., E.K., M.C., M.B., P.W.; Data postprocessing: I.J.H.G.W., A.M.W. Data availability The IMAGO dataset is available through the ECUAIM platform (EUCAIM Catalogue) after requesting access: EUCAIM. Once data access is granted the user can download the dataset from the data host: Health-RI XNAT. Direct access to the Health-RI XNAT can be obtained by sending an email with the required information (see usage notes) to servicedesk@health-ri.nl. Code availability DICOMs were converted to NIFTI using dcm2niix version v1.0.2023041133. Image registration was done with the Advanced Normalization Tools (ANTs) version 0.4.234. For skull stripping HD-BET version 1.0 was used35. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The IMAGO dataset is available through the ECUAIM platform (EUCAIM Catalogue) after requesting access: EUCAIM. Once data access is granted the user can download the dataset from the data host: Health-RI XNAT. Direct access to the Health-RI XNAT can be obtained by sending an email with the required information (see usage notes) to servicedesk@health-ri.nl. DICOMs were converted to NIFTI using dcm2niix version v1.0.2023041133. Image registration was done with the Advanced Normalization Tools (ANTs) version 0.4.234. For skull stripping HD-BET version 1.0 was used35.