A Longitudinal Glioblastoma MRI Dataset with Anatomical Landmark Pairs for DIR Validation

This dataset contains longitudinal, multi-modal MRI data from glioblastoma patients treated at Duke. Matching anatomical landmark points are labeled on blood vessel bifurcations of both the pre-operative and post-recurrence scans for each patient. These anatomical landmark pairs can be used for deformable image registration (DIR) validation and benchmarking. Automatically segmented tumor volumes, demographic data, and relative treatment dates are also available for each patient. The full description of the dataset and its acquisition will be described in a future publication. Dataset Organization The dataset includes longitudinal, skull-stripped MRI images for 61 patients, resampled to an isotropic voxelsize of 1mm, and co-registered to the Montreal Neurological Institute brain atlas for FOV standardization. Each patient has 3 to 18 scanning sessions (total of 503 across the full dataset), each of which have a T1, T2, FLAIR, ADC, DWI, and T1-contrast enhanced (T1-c) MRI image. The images are saved in NIfTI format. Landmark data is included for the pre-operative and post-recurrence scans for each patient. Recurrence was defined either by the volumetric change of the tumor or clinical progression noted in the patient's chart. (The full details of the patient selection criteria will be published in a future dataset paper). The preoperative and recurrence scans are clearly identified by the "preoperative" and "recurrence" suffixes for the scans for each patient. The landmark positions are saved as csv files in each patient folder, named by the scan they correspond with. Landmark positions are recorded as positive voxel indices, set to start at [0,0,0]. They are organized by their distance to the tumor: "near" are landmarks <3cm from enhancing tumor boundary, while "far" are those >3cm from tumor boundary. So, for example, the files "patient17_scan1_landmarks_far.csv", and "patient17_scan4_landmarks_far.csv", saved in the "Patient17" folder, lists the position of the same matching positions >3cm from the enhancing tumor boundary in Scan 1 and Scan 4 of Patient 17. The dataset also includes tumor segmentation masks for each scanning session generated with a nnUNet segmentation software described in "Development and Evaluation of Automated Artificial Intelligence–Based Brain Tumor Response Assessment in Patients with Glioblastoma" (Zhang et al, 2025). In these segmentations, the labels are as follows: 1: Necrotic Tumor Core 2: Nonenhancing T2-FLAIR signal abnormality 3: Enhancing Tumor 4: Resection Cavity Additional Patient Data Additional data for each patient including scan dates, patient demographics, treatment information, and more are tabulated in the csv files published alongside the dataset in Zenodo. Treatment info includes relative dates of diagnosis, chemoradiation, radiation, major surgeries, and death. Demographic info includes patient ages, race/ethnicity, smoking status, IDH mutation, and BT-RADS score of scans for which it is available. To ensure anonymity, all dates are set relative to the first scan session available for each patient. Additional Guidance for using the dataset can be found on our Github at https://github.com/deshanyang/GBM-DIR-QA