Matched clinical states, divergent courses

At one scan, these two glioblastomas were closely matched in volume, imaging features and RANO rating. Their subsequent observed courses diverged.

—

Enhancing tumorNecrotic coreEdema / FLAIRObserved anatomy per scan; scaled between scans

Total segmented lesion volume, both patients (cm³)

ABDots: scans, colored by expert rating (red progressive, green stable, blue response, white surgery)Dashed: enhancing tumor

Matched state

Observed courses

Divergence among matched scan pairs

Method and provenance

What the ratings do not include. RANO requires a stable or decreasing corticosteroid dose for a response to count. RANO 2.0 requires confirmation of objective response (after at least four weeks) and requires or recommends confirmation of progression in specified settings, particularly within 12 weeks of radiotherapy because of pseudoprogression. LUMIERE records neither steroid dose nor confirmation status. The ratings shown are the source dataset’s expert RANO labels and have not been re-adjudicated under RANO 2.0; read them as imaging assessments only.
Observed geometry. By default each patient's 3-D view shows that patient's own tumor compartments at the current scan and a brain surface from one of their scans, reconstructed from LUMIERE's atlas-space DeepBraTumIA segmentation and simplified for display. Between scans the most recent scan's shape is scaled to the estimated volumes. The Geometry control switches to equivalent-volume shapes.
Zoom and cut-away (pinch, the + and − buttons, or the mouse wheel after clicking the view; ◐ slices the tumor open) enlarge the current 3-D view to show how its compartments nest. They do not reveal detail finer than the MRI, whose resolution is on the millimetre scale.
Cells (✺ switches between cells and measurement dots) are illustrative renderings in the style of fluorescence microscopy, not imaged cells. Magenta cells are tumor cells: rounded in the tumor body, elongated where they infiltrate the surrounding edema, and shown dividing at a rate that rises when the enhancing tumor is growing. Grey, fragmenting cells are dying; they increase when the enhancing tumor shrinks or the necrotic core grows. Pale blue cells with branching processes are healthy brain cells around and among the infiltrating tumor cells. Infiltrating cells spread outward faster when edema is growing. The rates come from the displayed volumes (observed between scans, or calculated in the intervention analyses); the line under the legend states them. Zooming in brings the cells forward and fades the outlines.
Data: LUMIERE longitudinal glioblastoma dataset (Suter et al., Scientific Data 9:768, 2022): DeepBraTumIA compartment volumes, expert RANO ratings with written rationale, demographics, MGMT/IDH status and overall survival as provided.
Matching: a search over rated scans from patients represented each scan by radiomic features reduced to principal components, and paired scans with similar states whose next expert ratings went opposite ways. "Closer than X% of scan pairs" refers to that representation. It does not establish identical biology.
Compartment view shows illustrative cells, not imaged cells: the MRI cannot resolve individual cells. What is patient-specific is the pace: in each zone, the rates of division, death and migration follow that patient's measured rate of change in the matching compartment (enhancing volume for the rim, necrotic volume for the core boundary, edema/FLAIR volume for the infiltrating edge) at the current point in the course. Cell appearance and the zone layout follow published glioblastoma histology (pseudopalisading cells around necrosis, microvascular proliferation near the enhancing rim, tumor cells infiltrating the non-enhancing region); cell counts are not calibrated to a density.
Between scans the volumes follow shape-preserving interpolation; only the scan timepoints are measurements. In observed-anatomy mode, each patient uses that patient's reconstructed tumor geometry at the measured scans; between scans, the most recent observed shape is scaled to the interpolated compartment volumes. Equivalent-volume geometry remains available as an explicit comparison mode.
Not in these data: treatment details after surgery. Differences in treatment, timing since surgery, age or earlier history could explain the divergence; this demo shows that the matched snapshot alone did not predict it.
Expert ratings can be driven by findings that total volume does not capture, such as T2 progression or new non-target lesions. Rationale text is shown as provided.
Overall survival is the dataset's value in weeks; the panel dims when it is reached.
Automated segmentation volumes can fluctuate between scans for technical reasons.
Patient-071's first scan has no rating in the dataset; it is labelled "before surgery" because it precedes the same-week post-operative scan.