Datasets:

AIOmarRehan
/

medtrace-rhuh-gbm-derived

Delivery artefacts for the MEDTRACE longitudinal brain-MRI workstation: 28 glioblastoma patients, 3 timepoints each, prepared so a browser can render them directly.

This repository is the Space's data layer. The Space is a Static Space with no server, so the browser range-fetches these files straight from here.

Research artefacts. Not a medical device. Not for diagnosis, treatment planning, or any clinical decision. Not clinically validated.

Attribution

Derived from the RHUH-GBM collection, used and redistributed under CC BY 4.0.

Cepeda, S., García-García, S., Arrese, I., Herrero, F., Escudero, T., Zamora, T., & Sarabia, R. (2023). The Río Hortega University Hospital Glioblastoma dataset: a comprehensive collection of preoperative, early postoperative and recurrence MRI scans (RHUH-GBM) [Dataset]. The Cancer Imaging Archive. https://doi.org/10.7937/4545-c905

Cepeda, S., et al. (2023). Data in Brief, 50, 109617. https://doi.org/10.1016/j.dib.2023.109617

Only the CC BY 4.0 assets are used: the brain-extracted NIfTI images, the expert-corrected segmentations and the clinical CSV. The DICOM assets are under the NIH Controlled Data Access Policy and are not included, redistributed, or required.

What is here

Directory	Files	Naming	Contents
`images/`	336	`{series_id}.nii.gz`	4 sequences x 3 timepoints x 28 patients, windowed to uint8
`masks/`	336	`{series_id}.nii.gz`	Tumour mask reoriented into that series' own voxel grid
`volumes/`	84	`{segmentation_id}.nii.gz`	Skull-stripped atlas volume per study, uint8, for volume rendering
`brain/`	84	`{segmentation_id}.ply`	Brain surface mesh per study, binary PLY, for orientation
`meshes/`	120	`{observation_id}.ply`	Lesion surface mesh per observation, binary PLY

index.csv maps every identifier to its subject, timepoint and sequence. Start there if you are browsing by hand.

Files are named by identifier rather than by subject and sequence on purpose. The Space's URL builders are synchronous and receive only an identifier, so a readable tree would require fetching an id-to-path map before any volume could be requested. Naming by identifier removes that ordering dependency, and index.csv keeps the tree readable without duplicating 830 MB.

Source imaging is 1 mm isotropic, co-registered to the SRI24 atlas and brain-extracted upstream with SynthStrip, then intensity-normalised with CaPTk.

Label mapping

1 necrosis, 2 peritumoral signal alteration, 3 enhancing tumour.

Enhancing tumour was verified by measurement, not taken from documentation: label 3 is the brightest compartment on T1ce in 65 of 65 studies where it appears alongside another compartment, with a minimum margin of 0.244 standard deviations above the brain mean and a median of 1.514.

The necrosis and oedema order is adopted from the source documentation and corroborated, not independently proven. No single intensity feature reproduces it reliably: brightest-on-FLAIR selects the documented peritumoral label in 48 of 52 three-compartment studies, and darkest-on-T1ce selects the documented necrosis label in only 39 of 52. Label 2 does appear in 84 of 84 studies, and that ubiquity is consistent with it being the oedema compartment.

This distinction matters and is stated rather than smoothed over. MEDTRACE measures only the enhancing compartment, which is the one that was proven.

Two heuristics were tried and rejected, recorded here so they are not retried:

Largest compartment is oedema. False. RHUH-0001's necrotic core is 52.1% of its tumour.
Darkest on T1ce is necrosis. Unreliable, 39 of 52.

Preparation

Sequence volumes are windowed to uint8 with the window fitted to brain voxels at the 1st and 99th percentiles, background forced to exactly 0 so a transfer function can discard it.

The background is detected as the volume minimum rather than assumed to be zero. CaPTk normalises after skull stripping, so air is z-scored too: measured here it is a single constant near -6.8 filling 81% of the volume, and 43.5% of the remaining brain voxels are negative. Thresholding at zero discards nearly half the brain, and the rendered head comes out eaten away.

Windowing reduces the imaging from 1797 MB to 522 MB with no change to the stored source data.

Notes on completeness

28 of the 40 RHUH-GBM patients are included, selected as the first 28 grid-consistent subjects by sorted identifier so the set is reproducible.

RHUH-0028 is excluded: its preoperative study is on a 230x230x138 grid while its other two are 240x240x155. Lesions cannot be matched across differing voxel grids without registering first, so it cannot contribute a comparable timeline. RHUH-0029 takes its place.

Every included patient has at least two timepoints with a usable series, so a comparison is always possible.