Dataset Card for PREDICT-GBM

Dataset Description

The PREDICT-GBM dataset originates from an analysis of real patient data aimed at enhancing glioma treatment planning. It includes derived data for 253 patients from the mpMRI imaging database at Klinikum Rechts Der Isar, the Río Hortega University Hospital Glioblastoma dataset, and the LUMIERE dataset.

Files in the PREDICT-GBM dataset are:

• <modality>_bet_normalized.nii.gz: The skull-stripped, normalized, atlas-aligned modality in pre-op space.
• <modality>_warped_longitudinal.nii.gz: The skull-stripped, normalized, follow-up modality registered to pre-op space.
• tumor_seg.nii.gz: The segmentation of the pre-operative tumor. Labels are: 1=necrotic, 2=edema, 3=enhancing tumor
• recurrence_preop.nii.gz: The segmentation of the recurrence from the follow-up exam, registered to pre-op space. Labels are: 1=necrotic, 2=edema, 3=enhancing tumor, 4=resection cavity
• <tissue>_pbmap.nii.gz: Probability map of a tissue class (white matter, gray matter, cerebrospinal fluid).
• <model_id>_pred.nii.gz: The prediction from a model.
• <model_id>_pred.nii.gz: The radiotherapy plan derived from the model prediction, iso-volumetric to the standard plan.

The TUM-GBM dataset contains 142 patients with modalities from preop, postop, and follow-up MRI exams. Note that post-op is not available for all patients and some modalities are missing for individual cases.

We note that all niftis are of isotopic 1 mm³ resolution. It is recommended to overlay files with an identity affine as axis conventions may differ between NIfTIs.

Note that sbtc was renamed to LOTI for the PREDICT-GBM publication.

Intended Use

The dataset is intended for use in glioblastoma growth modeling, particularly for evaluating growth models using the recurrence segmentations. A pipeline is provided at https://github.com/BrainLesion/PredictGBM.

Citation

Soon.