Do Habitat Radiomics Signatures See Tumor Geography, or Only Its Histogram?

Tumor "habitat" analysis is a popular way to predict cancer survival from scans. It splits a tumor's voxels (3D pixels) into a few intensity clusters, such as a dense poorly perfused region or a necrotic region. It then feeds each cluster's features, together with clinical variables, into a survival model. The recipe works: in ovarian cancer and lung cancer it beats whole-tumor features on external cohorts. But the clusters are formed from voxel intensities alone. That means two tumors with the same mix of habitats get nearly the same summary even if their spatial layout differs completely. In one tumor the hypoxic habitat might be a single core wrapped in enhancing rim. In another it might be scattered foci. Population genetics met this exact problem decades ago. The site frequency spectrum, a histogram of allele frequencies, cannot tell a population bottleneck from background selection, because both skew the histogram the same way. The field learned to answer such questions by building a null model: preserve the summary, scramble the structure the summary ignores, and see what changes.

This note proposes that audit for prognostic imaging. The central tool is a composition-preserving spatial shuffle. It rearranges a tumor's habitats in 3D while keeping every habitat's volume fraction and internal texture fixed. Running it asks two questions. First, whether current habitat and deep-learning survival models use spatial arrangement at all. Second, whether arrangement carries survival information beyond composition. A second null, re-slicing thin-slice scans to clinical slice thickness, tests a related worry: acquisition resolution may shift habitat features in the same direction that biology does, the way demography mimics selection. If arrangement matters and current summaries miss it, the fix is a small set of 3D arrangement descriptors built on a habitat adjacency graph. If arrangement does not matter, the field learns that habitat composition is sufficient, which is also worth knowing.

Core Claim

Habitat radiomics is an inverse problem with the same kind of identifiability failure that population geneticists have documented for compressed genomic summaries. One study showed that site frequency spectrum inference confuses gradual population declines with sharp crashes, because distinct histories produce the same summary (A previously reported bottleneck in human ancestry 900 kya is likely a statistical artifact). Others showed that background selection and population expansion both produce an excess of rare variants (Transition densities and sample frequency spectra of diffusion processes with selection and variable population size; On the importance of skewed offspring distributions and background selection in virus population genetics). These papers motivate the proposal: a summary that throws away ordering, whether linkage order in genomes or spatial layout in tumors, cannot separate generators that differ only in that ordering.

The specific claim has three parts. Intensity-clustered habitat pipelines are nearly invariant to mesoscale 3D rearrangement. At least one clinically meaningful survival signal in glioblastoma lives in that arrangement. And through-plane resolution loss, which is literally volumetric context loss, is a confounder that moves habitat features along the same axis as prognosis.

I deliberately do not claim that imaging can detect clonal selective sweeps. The research file sketches a mapping in which necrosis stands in for a bottleneck and a habitat stands in for a sweep. Without spatially resolved sequencing, that mapping is itself unidentifiable, which is the lesson of the first direction. Population genetics supplies the method here, a null model that preserves the summary. It does not supply a biological analogy.

Research Program

The shuffle operator. The first component is the composition-preserving shuffle. For each tumor, the habitat label map is partitioned into supervoxels, which are compact contiguous 3D chunks of a few hundred voxels, each inside a single habitat. The supervoxels are then permuted in position within the tumor mask. This keeps each habitat's volume fraction and within-chunk texture intact and destroys the adjacency structure between habitats. A graded version swaps only a fraction $\rho$ of supervoxels, which gives a dose-response curve rather than a single test.

The quantity of interest for a trained survival model $f$ is

$$\Delta(\rho) = C(f, X) - C(f, \pi_\rho(X)),$$

where $C$ is Harrell's concordance index (the probability that the model ranks a random comparable patient pair in the correct survival order), $X$ are the real volumes, and $\pi_\rho$ applies the shuffle at fraction $\rho$. A model whose $\Delta$ stays near zero is using composition only. This mirrors the test in Your other Left! Vision-Language Models Fail to Identify Relative Positions in Medical Images. That paper flipped CT slices and found GPT-4o dropped from 75% to 50% accuracy, which showed the model relied on memorized anatomy rather than image geometry. We ask the analogous question of prognostic models.

A caveat is needed. If $\Delta \approx 0$ for a model, that tells us about the model, not about biology. To test biology, we retrain models on shuffled training data. If models trained on shuffled tumors match models trained on real ones, arrangement carries no learnable survival signal at this cohort size.

Arrangement descriptors. The second component is a small, pre-registered set of descriptors that a composition summary cannot express, computed from a 3D graph whose nodes are connected habitat components. The descriptors are the number of components per habitat, meaning fragmentation. Next is the contact area between each pair of habitats; necrosis–enhancing contact is the prime candidate. Then comes the distance-to-necrosis gradient of each habitat, and the enclosure of the necrotic core by enhancing tissue. The glioblastoma sex-dimorphism study motivates the choice of these descriptors (Sexually dimorphic radiogenomic models identify distinct imaging and biological pathways that are prognostic of overall survival in glioblastoma). It extracted 2,850 texture features from three subcompartments and found that male risk tracked enhancing and edema textures while female risk tracked the necrotic core. That points to interfaces between compartments as places where signal may live. We keep the descriptor set to roughly 20 features. The ovarian habitat study extracted 27,888 features from about 500 patients (mpMRI-based habitat analysis for predicting prognoses in patients with high-grade serous ovarian cancer: a multicenter study), and that feature-to-patient ratio is the false-discovery risk we want to avoid.

The resolution null. The third component is the analogue of demography mimicking selection. Thin-slice volumes are re-slice-averaged to 3, 5, and 7 mm through-plane resolution, and then the full pipeline is re-run: habitat clustering, features, and the frozen survival model. We measure two things. One is how far the risk scores move. The other is whether they move along the model's high-risk direction. A systematic shift would mean thick-slice centers look higher- or lower-risk for technical reasons. A pooled multi-center model would absorb that as biology. The pancreas segmentation work in the file showed that slice-wise processing fragments structures across $z$ (Deep convolutional networks for pancreas segmentation in CT imaging). We expect thick slices to fragment habitats in the same way, which inflates exactly the fragmentation descriptor we propose. That is why this null has to run alongside the descriptors, not after them.

Baselines

The naive reference point is a Cox proportional hazards model on clinical and molecular covariates alone: age, extent of resection where available, IDH status, and MGMT methylation from UPenn-GBM (The University of Pennsylvania glioblastoma (UPenn-GBM) cohort: advanced MRI, clinical, genomics, & radiomics). Any imaging model must beat it. Otherwise the imaging, arrangement included, adds nothing.

The second baseline is whole-tumor and subcompartment handcrafted radiomics. These are the 145 standardized CaPTk features per compartment released with UPenn-GBM, fed to LASSO-Cox with the same covariates.

The third is a reimplementation of intensity K-means habitat radiomics in the style of the ovarian and lung studies. The lung study (Habitat radiomics analysis for progression free survival and immune-related adverse reaction prediction in non-small cell lung cancer treated by immunotherapy) reported an external progression-free survival AUC of 0.808, against 0.626 for whole-tumor features on toxicity. The ovarian study reported an external overall-survival C-index of 0.745 for its combined model. We rerun this baseline on the glioblastoma MRI sequences with the number of clusters chosen by internal cross-validation. It is the direct target of the arrangement hypothesis.

The strongest existing method is deep radiomics. The first variant follows A Deep Learning-Based Radiomics Model for Prediction of Survival in Glioblastoma Multiforme: 2D ImageNet-pretrained CNN features, multi-step selection, and LASSO-Cox with age and performance score. That paper reports a validation C-index of 0.710 for the deep signature and 0.739 for the combined nomogram, on a TCGA/TCIA discovery cohort with a Sun Yat-Sen validation cohort rather than UPenn-GBM. The research file lists no reported survival C-index on UPenn-GBM itself, so we establish that number ourselves. The second variant is an end-to-end 3D model with a linear-cost state-space backbone, as in Swin-UMamba: Mamba-based UNet with ImageNet-based pretraining, which reported 0.7760 Dice on AbdomenMRI. We fine-tune it with a Cox loss. This is the most serious competitor, because a native 3D network could learn arrangement implicitly.

For the BraTS 2017 survival task we also report against Brain Tumor Segmentation and Radiomics Survival Prediction: Contribution to the BRATS 2017 Challenge, which reached 52.6% three-class accuracy and a Spearman correlation of 0.496.

Fairness rules apply to every comparison. All models use the same patient splits, the same covariates, and the same number of hyperparameter trials, 50 per model, chosen by nested cross-validation. Deep models get equal GPU hours.

The key ablation compares the habitat model with and without arrangement descriptors. We also compare the arrangement model trained on real versus shuffled volumes, which isolates whether arrangement is the ingredient that helps. Beating the strongest baseline means two things. The arrangement-augmented model must exceed the 3D deep model's C-index on held-out data, with a bootstrap confidence interval on the difference that excludes zero. And its gain must shrink toward zero when it is retrained on shuffled data.

Why not just…?

Why not just train a 3D network end to end? It already sees the full volume. This is the real threat. With a few hundred patients, a 3D network may learn arrangement, or it may learn only volume and texture. Nobody currently knows which, and $\Delta(\rho)$ measures it directly. If the 3D model shows large $\Delta$ and matches the descriptor model, we report that the descriptors add interpretability but no accuracy. The contribution then narrows to the audit and the resolution confound, and we would say so plainly.

Why not just rely on GLCM and other texture features, which are already spatial? Co-occurrence textures measure voxel neighborhoods a few voxels across. The shuffle keeps those neighborhoods intact inside each supervoxel by construction, so texture-based models should show near-zero $\Delta$. If they do not, our supervoxels are too small and the test is not measuring mesoscale arrangement. This is a built-in check on the operator.

Why not just add slice thickness or center as a Cox covariate, or harmonize features across centers? Additive adjustment is the clinical analogue of adding ancestry principal components. The sex-dimorphism study found that pooled models with sex treated as a mere adjustment failed on sex-specific cohorts, and the drift and eigenvector-GWAS literature shows that linear correction misses nonlinear distortions (EigenGWAS: finding loci under selection through genome-wide association studies of eigenvectors in structured populations). We test the adjustment directly. We train with a thickness covariate and then check whether risk still drifts under controlled re-slicing within the same patient. A covariate cannot remove a within-patient shift it was never shown.

Why not just use spatially regularized clustering, so habitats are contiguous from the start? Smoother habitats reduce fragmentation noise, but the downstream summary is still habitat composition, so the identifiability gap remains. We include a Markov-random-field-smoothed clustering as an ablation to separate "cleaner habitats" from "arrangement information."

Proposed Experiments

The primary cohort is UPenn-GBM: 630 patients with expert subcompartment segmentations, structural MRI for everyone, and perfusion MRI for 79.6%. We train on it with five-fold nested cross-validation. External tests use the BraTS 2017 survival cohort and, if accessible, the TCGA/TCIA glioblastoma data used by the deep-radiomics paper. The metrics are C-index, time-dependent AUC at 12 and 24 months, and calibration slope.

Three results are decisive. The first is the $\Delta(\rho)$ curve for every baseline. A flat curve for K-means habitats together with a rising curve for the 3D model would show that intensity habitats are blind to layout while volumetric models are not. The second is the retrain-on-shuffled test for the descriptor model. If performance does not drop, the arrangement hypothesis is false at this cohort size, and we report that as a negative result. The third is the resolution null. We report the correlation between the risk shift caused by thick-slice re-slicing and the model's own risk axis, both overall and stratified by sex and IDH status, because the dimorphism study suggests the necrotic-core signal is sex-specific. A strong positive correlation would mean multi-center habitat biomarkers are partly measuring slice thickness.

All three outcomes are informative whichever way they fall. Pre-registering the roughly 20 descriptors before any survival model is fit keeps the study from becoming another high-dimensional fishing trip.