Tumor hypoxia means pockets of tissue starved of oxygen. It is one of the main reasons radiotherapy and other oxygen-dependent treatments fail, and it pushes cancer cells toward more aggressive states. Clinicians would like to know where inside a tumor the hypoxic regions sit, for example to boost radiation dose there. The trouble is that nobody can hand-draw hypoxia on a routine MRI. A radiologist can outline dead tissue or the contrast-enhancing rim, but oxygen level is not visible to the eye, and the direct measurements (needle electrodes, special PET tracers, biopsies) are invasive, rare, and spatially patchy. So the annotation bottleneck here is not only about cost. The pixel labels we would want cannot be produced at any price.
What does exist, for hundreds of patients, is a single number per tumor: a hypoxia score computed from bulk RNA sequencing of the resected tissue. We propose to treat that number as a weak label on the average hypoxia of the tumor. We then train a network that assigns a hypoxia probability to every voxel, constrained so that the voxel values, averaged over the tumor, reproduce each patient's bulk score. The same voxel-level function is shared across all patients. Because of that sharing, the cross-patient variation in bulk scores can pin down which local imaging patterns carry hypoxia signal. The deliverable is a spatial hypoxia map learned without a single hand-drawn hypoxia label. We will also test honestly whether such a map is recoverable at realistic cohort sizes, and whether it tells us anything that standard anatomical sub-regions do not.
Core Claim
A bulk molecular hypoxia score, used as an aggregate constraint over a shared per-voxel model, is enough weak supervision to learn a spatial hypoxia map from routine multi-contrast brain MRI. That map should predict held-out bulk hypoxia and survival better than the standard necrotic, enhancing, and edema sub-regions do.
The claim can fail in two separate ways. The map might not be identifiable at around 100–250 patients. Or it might be identifiable but add nothing beyond the necrosis fraction. The experiments below are designed to tell these two failures apart.
What prior work establishes and leaves open
The motivating result is Beig et al.. They computed a Hypoxia Enrichment Score (HES) from 21 hypoxia-associated genes on 115 TCGA glioblastoma cases. They then showed that texture features inside radiologist-drawn necrotic, enhancing, and edema regions correlate with HES and separate short-term from long-term survivors. Their pipeline, however, uses hypoxia only as a target to correlate against. It depends on full manual contouring and produces no map of where hypoxia is. We reuse their molecular score and cohort, but we use the score as supervision for a spatial model.
Bakas et al. released expert-corrected sub-region labels for 243 TCGA glioma scans. They seeded these with the automated GLISTRboost tool, and the automated output already agreed closely with the corrected labels (median Jaccard 0.96 whole tumor, 0.87 tumor core). Together with the 2,040-patient BraTS 2021 benchmark, this tells us something that constrains the design. For glioma, the anatomical annotation bottleneck is largely solved: a BraTS-trained segmenter can supply the tumor outline and the necrotic, enhancing, and edema regions automatically. We therefore do not claim to automate anatomical contouring. We use it as infrastructure, and our claim concerns a quantity that anatomical labels do not contain.
On the hypoxia-imaging side, Zhang et al.'s habitat imaging grouped breast tumor voxels with unsupervised fuzzy clustering over perfusion and diffusion maps. They called the low-vascularity, high-cellularity cluster the putative hypoxic core and linked its volume fraction to HER2 status and chemotherapy response. That work supplies a competing method: clustering inside a mask. It also exposes the open problem we target. Its clusters carry no semantic supervision, and the label "hypoxic" rests on a physiological assumption that was never checked against molecular data.
The label-efficient segmentation literature supplies both method and cautions. WESEG trained pan-cancer tumor segmentation on whole-slide pathology images from noisy slide-level tumor-percentage estimates. This is the closest methodological precedent: an aggregate proportion label supervising a dense map, reaching 0.932 AUC out of distribution. We transfer that idea from a visible class (tumor cells, which a pathologist can check) to an invisible one (hypoxia) and from pathology slides to MRI, which makes the identifiability question central in a way it was not for WESEG.
Two results constrain the design. Andermatt et al.'s PathoGAN and CAM-style methods such as HistoSegNet show that image-level labels tend to yield one monolithic class, or "peaky" maps that highlight only the most discriminative core. Self-training pipelines such as Zhang et al. on lung cysts degrade after a few rounds of pseudo-labeling (75.87% then 74.94% Dice). We therefore avoid recursive pseudo-labeling entirely and fit directly to the aggregate constraint.
Research Program
Each TCGA patient $i$ comes with four MRI contrasts, a bulk HES value $y_i$, and survival data. A pretrained BraTS segmenter produces a tumor-core region $T_i$, meaning the enhancing plus necrotic tissue, roughly the tissue that gets resected and sequenced. A 3D network $h_\theta$ maps the local multi-contrast appearance around voxel $x$ to a hypoxia probability $h_\theta(x) \in [0,1]$.
Training fits the following model:
$$ y_i = \alpha + \beta \cdot \frac{1}{|T_i|}\sum_{x \in T_i} h_\theta(x) + \varepsilon_i , $$
Here $\alpha$ and $\beta$ are learned scalars that calibrate the gene-expression scale, and $\varepsilon_i$ is noise. Because HES is a relative enrichment score, we fit with a pairwise ranking loss on the predicted tumor-mean, plus a small regression term.
Three design choices matter:
- Small receptive field. The network sees only a limited patch, so it cannot read off patient identity or overall tumor size. This forces it to learn a local function that is shared across patients.
- Smoothness term. A total-variation penalty discourages salt-and-pepper maps.
- Initialization from BraTS features. The encoder starts from BraTS 2021 self-supervised features, so the few hundred RNA-labeled patients only need to learn the hypoxia head.
We train on the TCGA-GBM and TCGA-LGG cases that have both the Bakas et al. imaging and RNA-seq. Including low-grade tumors widens the range of HES values, which should help identifiability.
The analysis has three parts, run in this order:
- Identifiability on synthetic data. We plant a known hypoxia field in BraTS 2021 volumes, generate synthetic bulk scores, and ask whether the per-voxel function is recoverable at the real cohort size.
- Held-out prediction on real TCGA data. We test whether the learned map predicts HES and survival in patients it has not seen.
- Biological checks on the learned maps. Does the map concentrate near, but not inside, necrosis, as expected of tissue at the edge of oxygen diffusion? Or does it simply copy the necrosis label? Does it vary with HES among patients whose necrosis fractions are similar?
Baselines
The naive reference point has two parts. The first predicts HES and survival from age, performance score, and total tumor-core volume alone. The second is a label-permutation control: we train our full model with HES shuffled across patients. Any spatial structure that survives shuffling is a prior of the architecture, not hypoxia. Beating the naive reference means a significantly higher held-out Spearman correlation with HES, and a higher survival concordance index (C-index, the fraction of patient pairs the model ranks correctly by survival time), under identical folds.
The strongest existing method is the Beig et al. pipeline, which uses texture features from expert sub-regions, mutual-information feature selection, and a random forest. They report a C-index of 0.83 for short- versus long-term survivor discrimination when combined with age and performance score, on an 85-case training and 30-case validation split. For predicting HES itself as a continuous target, no reported result was found on Lacuna. We will run this pipeline in two forms, each with the same cross-validation folds and the same hyperparameter search budget as our method:
- Beig with automatic sub-regions, replacing expert contours with a BraTS-trained segmenter. This is the realistic competitor once anatomical annotation is automated.
- Beig with expert sub-regions, using the corrected Bakas labels. This serves as an upper reference.
We also reproduce the survival evaluation on a split of the same sizes. Matching 0.83 without any manual contour is the minimum bar. The claim of added value requires beating the automatic-sub-region pipeline on held-out HES correlation.
Habitat clustering follows Zhang et al., applied to the four structural contrasts inside the automatic tumor mask. The clustering sees no HES. We pick the cluster whose volume fraction best correlates with HES using training folds only, and then evaluate that fraction on held-out patients. This isolates whether supervision helps beyond physiologically motivated clustering.
Global attention regression trains the same encoder to predict HES from the whole tumor through attention pooling, and reads the attention weights as a map. Two ablations isolate the key ingredient. The first replaces the mean-of-voxel constraint with this attention pooling, testing whether the aggregation constraint is what makes the map interpretable as a hypoxia fraction. The second enlarges the receptive field to the full tumor, testing whether sharing a local function across patients is what drives recovery.
Why not just…?
Why not just automate the sub-regions with a BraTS model and feed them into Beig-style radiomics? This is the most serious threat. If HES is mostly explained by how much necrosis a tumor has, it will perform as well as our method. Beig et al. already found signal inside predefined anatomical regions. It is therefore plausible that the necrotic fraction carries most of it. We run exactly this pipeline as a baseline. The decisive comparison is held-out HES correlation after controlling for necrosis fraction, and in particular whether our map explains HES variation among patients with matched necrosis fractions. If it does not, we will report that anatomical sub-regions suffice and the spatial map is redundant.
Why not just cluster habitats without labels? Clustering needs no RNA, and Zhang et al. found clinically meaningful habitat fractions this way. But nothing in the clustering objective ties any cluster to hypoxia, and the "hypoxic" cluster was named by assumption. The clustering baseline, given the advantage of choosing its best cluster on training folds, measures directly how much the molecular supervision adds.
Why not just regress HES from the whole image and look at the attention? Attention-pooled regression uses the same labels, and it may predict the scalar HES equally well. Attention weights, however, are not constrained to mean anything as a fraction of tissue. CAM-like maps are known to become peaky on discriminative cores, as HistoSegNet and Gao et al. illustrate. The synthetic experiment settles this directly: both methods are scored on how well they recover a planted voxel map, not only on scalar prediction.
Why not just have experts annotate hypoxia? For necrosis, experts can annotate, and Can et al. show that cheap scribbles would lose only a few Dice points. Hypoxia, though, is not visible on structural MRI. Expert annotations of the visible sub-regions already disagree at 74–85% inter-rater Dice (Menze et al.). No expert effort produces hypoxia labels from these scans. That is exactly why a molecular weak label is needed.
Proposed Experiments
Synthetic identifiability study. On BraTS 2021 volumes, we define a ground-truth hypoxia field as a fixed nonlinear function of local appearance and distance to necrosis. We compute bulk scores as noisy tumor-core means, with noise levels chosen so the score-to-mean correlation spans a range of plausible values. We then train each method at cohort sizes from 50 to 2,000 and measure voxel-level recovery: correlation with the planted field, and Dice at a fixed threshold. The key output is a recovery-versus-cohort-size curve. If recovery at roughly 200 patients is poor, the real-data spatial claim is not supported, and we say so. The study also tests a failure case: a planted field that is simply a copy of the necrosis label. In that case the method should recover necrosis, which shows why the real-data necrosis-controlled comparison is needed.
Real-data prediction. On TCGA glioma, we use nested cross-validation for held-out HES Spearman correlation, both overall and within strata of necrosis fraction. We also report survival C-index on a reproduction of the 85/30 split sizes, against all baselines above, with bootstrap confidence intervals. With cohorts this small, we pre-register the primary endpoint: held-out HES correlation controlling for necrosis fraction.
Map sanity checks. We run the label-permutation control. We measure the spatial relation of the learned map to necrotic and enhancing tissue. We test stability across folds by measuring voxel-wise agreement of maps trained on disjoint patient subsets. Hypoxia maps that change completely between folds would undercut any clinical use, regardless of their predictive score.
The decisive result. The claim is supported only if three things hold together:
- The synthetic study shows recovery at the real cohort size.
- The permutation control shows no spurious structure.
- The learned map beats the automatic-sub-region Beig pipeline on necrosis-controlled held-out HES correlation.
Spatial validation against a hypoxia PET tracer or tissue staining would be the natural next step. No such paired dataset was found on Lacuna, so this proposal does not depend on one.
References
- Radiogenomic analysis of hypoxia pathway is predictive of overall survival in Glioblastoma
- Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features
- The RSNA-ASNR-MICCAI BraTS 2021 Benchmark on Brain Tumor Segmentation and Radiogenomic Classification
- Study on heterogeneity of vascularity and cellularity via multiparametric MRI habitat imaging in breast cancer
- Weakly supervised pan-cancer segmentation tool
- Pathology Segmentation Using Distributional Differences to Images of Healthy Origin
- HistoSegNet: Semantic Segmentation of Histological Tissue Type in Whole Slide Images
- Self-learning to detect and segment cysts in lung CT images without manual annotation
- Multimodal Weakly Supervised Segmentation for Histopathology
- Learning to Segment Medical Images with Scribble-Supervision Alone
- The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS)