Resting-state functional MRI (fMRI) measures slow, spontaneous fluctuations in blood oxygenation while a person lies still in the scanner. Two brain regions whose fluctuations rise and fall together are said to be "functionally connected." A recent glioma study found that the connectivity between a recurrent tumor and one large-scale brain network, the dorsal attention network, predicts how long patients survive. That connectivity is computed by averaging the signal over every voxel inside a drawn tumor outline. Expert outlines of brain tumors routinely disagree at the edges, and the edges are exactly where tumor tissue meets healthy cortex, which is itself strongly coupled to that network. So the survival biomarker may reflect tumor biology. It may also reflect how much healthy cortex a particular outline happened to swallow.
This proposal turns inter-observer disagreement from a nuisance into a measuring tool. Several experts outline each tumor. We then split the tumor region into a core that every rater includes and a contested band that only some raters include. Connectivity and its link to survival are computed separately in the core, the contested band, and geometric control shells of matched width. If the prognostic signal lives in the core, the biomarker is robust and tumor-specific. If it lives in the band, and a rater-free geometric ring of the same width reproduces it, the biomarker is mostly boundary leakage. If the band carries signal that the geometric ring does not, human disagreement is marking something real, plausibly infiltrating tumor. Each of these outcomes changes how such biomarkers should be measured.
Core Claim
The closest existing work, Functional connectivity between tumor region and resting-state networks as imaging biomarker for overall survival in recurrent gliomas, took one PET-thresholded tumor volume in 82 patients. It used dual regression, which projects group-level network maps onto each subject's data, to estimate tumor-to-network connectivity against seven canonical networks. Higher connectivity to the dorsal attention network went with longer survival (HR 0.88, P = 0.007; multivariate HR 0.90, P = 0.033; median survival 31.9 vs 10.6 months). That study motivates this proposal. It also leaves the question open, because it never varies the mask.
Our claim has two parts. First, the reliability of a tumor-to-network biomarker cannot be read off segmentation overlap alone, because connectivity is a nonlinear function of which voxels are averaged. Second, the spatial pattern of human disagreement is a more informative perturbation than arbitrary erosion or dilation, because it concentrates on the anatomically ambiguous margin where tumor and cortex mix. The deliverable is a disagreement-stratified connectivity analysis and a clear verdict on whether the published prognostic effect survives it.
Research Program
For each patient we collect $K \geq 4$ independent tumor outlines on structural MRI, drawn by neuroradiologists and radiation oncologists of differing seniority. We also collect the PET biological tumor volume at several plausible uptake thresholds, since thresholding is itself a rater-like choice. From these outlines we form a voxelwise agreement map $p(v) \in [0,1]$, the fraction of raters who include voxel $v$. The consensus core is $\{p(v)=1\}$ and the contested band is $\{0<p(v)<1\}$. The healthy ring is the region just outside the union of all outlines, taken at the same width as the band.
We also build agreement-weighted and geometric control regions. The agreement-weighted tumor signal is
$$\bar{x}_t = \frac{\sum_v p(v)\, x_t(v)}{\sum_v p(v)},$$
where $x_t(v)$ is the preprocessed signal of voxel $v$ at time $t$. This is the continuous alternative to a single thresholded mask. For each contested band we then build a geometric twin: a shell obtained by eroding and dilating the consensus mask to match the band's volume and mean distance from the core, with no rater input. The contrast between band and twin is the ingredient that isolates what human disagreement adds beyond geometry.
Each region's time series then enters the Friedrich pipeline unchanged: dual regression against the seven Yeo networks, followed by univariate and multivariate Cox survival models with the same clinical covariates (tumor type, grade, IDH status, MGMT methylation, tumor volume). We report three things per region. The first is the between-rater intraclass correlation of each tumor-to-network connectivity value. The second is the spread of hazard ratios across the $K$ single-rater masks. The third is the survival concordance index, which measures how well the model ranks patients by survival time.
Two controls address anatomy rather than raters. The first is a mirrored-mask null. We reflect each patient's union mask into the contralateral, tumor-free hemisphere and repeat the analysis. Survival association in healthy mirrored tissue would indicate that location and network proximity, not tumor tissue, carry the effect. The second tests automated disagreement. We train a deep ensemble on the multi-rater brain-tumor task of QUBIQ, which found that simple ensembles beat more elaborate probabilistic models at reproducing rater disagreement. We then ask whether the ensemble's disagreement band reproduces the human-band findings. If it does, the method becomes deployable without a rater panel.
The broader multi-rater methods supply tools and framing rather than direct competition. The Latent Doctor Model models annotator style as a latent variable for pathology classification. Multi-rater Prism alternates between a calibrated mask and per-rater confidence maps and does best on high-disagreement structures such as brain tumors. Diagnosis-First calibration chooses the label fusion that maximizes downstream classification AUC. None of them touches physiological time series.
We deliberately do not optimize the mask for survival prediction, which would be the Diagnosis-First approach transplanted here. With 82-patient-scale cohorts, tuning a segmentation to maximize a survival fit is a recipe for overfitting, and it would answer the wrong question: it asks "which mask predicts best," not "where does the signal come from." Jensen et al.'s review reinforces the need for caution. It found that static whole-brain connectivity markers often fail to replicate across cohorts, so robustness has to be shown before utility is claimed.
Baselines
The naive reference point is a Cox model with clinical covariates only, including tumor volume. Any connectivity measure must add concordance over it to matter. No reported concordance index was found on Lacuna for this setting, so we compute it ourselves on the same cohort and folds.
The strongest existing method is the single PET-mask pipeline of Friedrich et al., with its reported dorsal attention network result of HR 0.88 (P = 0.007) univariate and HR 0.90 (P = 0.033) multivariate. We rerun it exactly on our cohort before anything else. A replication failure is itself a finding, and every other arm is compared against the replicated number rather than the published one.
Consensus fusion by majority vote and by STAPLE, an algorithm that jointly estimates a consensus mask and each rater's reliability, represents what a careful lab would do today. We run it with the same pipeline. Beating it means showing that its hazard ratio sits at an unrepresentative point in the across-rater distribution, or that core-only or weighted connectivity gives higher between-rater reliability at equal or better concordance.
A Multi-rater Prism calibrated mask serves as the strongest learned fusion, trained on QUBIQ brain tumor data under the same compute budget as our ensemble. The file gives no numeric Dice for it, and QUBIQ top teams reached average Dice of 0.80–0.85.
The key ablation is the geometric twin shell. All comparisons use identical preprocessing, identical folds, and bootstrap confidence intervals, with no survival-driven tuning in any arm. "Beating" the baselines here means an interpretable, reproducible decomposition: core-versus-band hazard ratios whose confidence intervals separate, or a band-versus-twin difference that exceeds the bootstrap spread.
Why not just…?
Why not just fuse the raters with majority vote or STAPLE and move on? Fusion produces one mask and therefore one hazard ratio. It hides whether that number is stable across plausible outlines. Diagnosis-First calibration showed that majority-vote labels are suboptimal for downstream classification, gaining up to 2.73% AUC on glaucoma. Our concern is different, though: we care about attribution, not accuracy. The fusion arm settles this empirically. If the fused-mask hazard ratio sits in the middle of a narrow across-rater distribution, fusion is adequate, and we will say so.
Why not just erode or dilate the mask geometrically, with no raters at all? This is the most serious threat, and we built it into the design as the twin-shell ablation. If rater disagreement bands are essentially uniform rings, the twin reproduces every result and the expensive rater panel is unnecessary. The proposal then reduces to a cheaper, still useful shell-sensitivity protocol. The rater panel earns its cost only if band and twin diverge in connectivity or in survival association.
Why not just replace humans with one deterministic automatic segmenter? A single network removes disagreement by fiat. It does not remove ambiguity: it inherits its training annotators' conventions, as the brain tumor segmentation survey notes about dependence on manual training labels. It also reports nothing about which boundary voxels are uncertain. The ensemble-disagreement arm is the honest version of this alternative. If it matches the human band, automation wins on cost.
Why not bypass tumor masks with whole-brain voxelwise or independent-component analysis? Mass effect makes registration to a standard template unreliable. More fundamentally, the scientific question is about the tumor's own signal, which requires deciding which voxels are tumor. The mirrored-mask null covers the part of this critique that matters: whether the effect is really about brain location rather than tumor tissue.
Proposed Experiments
The first check is a replication of the Friedrich result on a cohort with rs-fMRI, FET-PET or contrast MRI, and survival follow-up, extended with our multi-rater outlines. The second is the core/band/twin decomposition for all seven networks, with the dorsal attention network pre-registered as the primary test and the visual network in glioblastoma as secondary. The third is the mirrored-mask null. The fourth is the ensemble-versus-human band comparison.
Three outcomes would be decisive. The biomarker is robust if the dorsal attention network hazard ratio from the consensus core alone stays below 1 with a confidence interval excluding 1, the between-rater intraclass correlation of that connectivity is high, and the mirrored-mask null is flat. The biomarker is a leakage artifact if the effect appears only when the band is included, the twin shell reproduces it, or the mirrored mask shows it. Disagreement carries biology if the band predicts survival beyond both the core and its geometric twin.
A secondary analysis asks how much the across-rater hazard-ratio spread shrinks under agreement weighting. The answer tells future studies how many raters they need.
References
- Functional connectivity between tumor region and resting-state networks as imaging biomarker for overall survival in recurrent gliomas diagnosed by O-(2-[18F]fluoroethyl)-<scp>l</scp>-tyrosine PET
- QUBIQ: Uncertainty Quantification for Biomedical Image Segmentation Challenge
- The Latent Doctor Model for Modeling Inter-Observer Variability
- Multi-rater Prism: Learning self-calibrated medical image segmentation from multiple raters
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle
- Aberrant cortical–subcortical-cerebellar connectivity in resting-state fMRI as an imaging marker of schizophrenia and psychosis: a systematic review of data-driven whole-brain functional connectivity analyses
- A survey of MRI-based brain tumor segmentation methods