The Pulse Sequence as Confounder Label: Paired Black-Blood Scans for Auditing and Removing Blood-Signal Shortcuts

Image classifiers in medicine often get high accuracy for the wrong reasons. They latch onto a strong signal that happens to correlate with the diagnosis, such as a scanner tag or a patient's sex, instead of the pathology itself. In vascular and cancer imaging, the most obvious candidate for such a shortcut is flowing blood. With contrast or flow enhancement, the lumen (the channel the blood flows through) appears bright. Its brightness depends on contrast timing, flow speed and scanner, and it sits right next to the vessel wall, where the disease actually is. MRI and CT already have a physical way to remove that signal. "Black-blood" acquisitions null or subtract the luminal blood and leave the wall and surrounding tissue visible. Many vessel-wall protocols acquire bright-blood and black-blood images of the same patient in the same session.

This proposal treats that pairing as a measuring instrument for interpretability. For each patient we have the same anatomy imaged with and without luminal blood signal. From the pair we can compute, without any human labeling, how much luminal blood intensity each lesion carries. That quantity then serves two purposes. It audits whether a bright-blood classifier depends on blood intensity. It also supplies the confounder variable that existing shortcut-removal methods need but cannot obtain for unlabeled image artifacts. The two-sentence version: the black-blood sequence physically labels the blood-signal confounder, so we can measure and penalize a model's reliance on it. We can do this without asking anyone to annotate the confounder, and without throwing away the lumen's shape, which is real diagnostic information.

Core Claim

Dependence-minimization methods divide a model's internal features into task features and confounder features, then push the two toward statistical independence. Müller et al. benchmark these methods and show that the mutual-information version (MINE, a neural estimator of how much one variable tells you about another) keeps an AUROC of 87.7 on an inverted CheXpert split, where an unconstrained network collapses to 45.8. That paper also states the method's main limitation: every confounder must carry a ground-truth label during training. Unlabeled image-level signals cannot be penalized.

Our claim is that, in vascular imaging, a co-acquired black-blood image turns luminal blood intensity into a measured, per-patient continuous variable. Feeding this variable into MINE-style independence training should give bright-blood classifiers two properties. First, they should survive distribution shifts in contrast timing and flow. Second, their attributions should agree better with radiologists' wall-centered reasoning. They should achieve both without losing the lumen-geometry information that blunt lumen masking would discard.

The key distinction is between lumen geometry, such as stenosis and abutment angle, which a model should use, and lumen intensity, which it should not. Physical suppression lets us separate the two.

Relation to Prior Work

The motivating evidence comes from three clinical studies, none of which involves learned diagnosis.

  • Black-blood CT for pancreatic cancer. Lin et al. subtract arterial-phase from delayed-phase CT to suppress luminal enhancement, so radiologists can tell tumor abutment (≤180°) from encasement (>180°) in pancreatic cancer. Readers reach an AUC of 0.942. The deep network in that work is only a denoiser.
  • Black-blood MRI for glioma. Finck et al. show that black-blood T1 MRI raises tumor contrast-to-noise from 32.5 to 43.4 and reveals lesions invisible on bright-blood scans.
  • Automated vessel-wall processing. Zhang et al. deploy an automated vessel-wall pipeline across 1,981 patients and both bright- and black-blood sequences. They report degraded performance in small distal arteries but offer no explanation of the failures.

Together these studies establish that luminal signal is strong, variable and adjacent to the decisive anatomy. That is the profile of a shortcut. None of them asks what a classifier does with it.

The method comes from two sources. The first is Müller et al., whose independence penalty we reuse. We change its input: a physically derived confounder measurement replaces a human-annotated label. The second is the polar unwrapping of the tracklet-refinement vessel-wall work. That paper converts each artery cross-section into a strip indexed by radius and angle, which lets us define lumen and wall regions and compute per-angle measurements. We use it as a measurement coordinate system rather than as an interpretability device.

Three papers constrain the design.

  • Off-manifold counterfactuals. Laugel et al. show that up to 37% of post-hoc counterfactual explanations sit in regions unsupported by data. This is why we avoid synthetic "remove the blood" edits and instead use acquired counterfactuals, which are real scans by construction.
  • Generative reconstruction. SPIRiT-Diffusion and the review by Rueckert and Schnabel warn that generative reconstruction can produce non-anatomical detail. We therefore do not synthesize black-blood images from bright-blood ones in the main method, and use synthesis only as a clearly separated ablation.
  • Attribution alignment. Haddada et al. measure how well SHAP rankings agree with expert rankings using Kendall's τ, a rank correlation between −1 and 1. They show that models with similar accuracy can range from τ = 1.0 to τ = −0.39. We adopt their metric for regional attributions.

Yuksekgonul et al. build concept bottleneck models, which predict from human-named intermediate concepts, and raise confounded NIH-age accuracy from 9.4% to 50.7%. Their approach is the strongest competing route to robustness, and we treat it as a baseline rather than a component.

Research Program

Confounder measurement. For each paired study, we register the black-blood image to the bright-blood image within a local vessel neighborhood. The lumen boundary is taken from the black-blood image, where it is a sharp dark-to-wall edge. The confounder vector $b$ for a lesion consists of the luminal intensity statistics on the bright-blood image within that boundary, per polar angle sector. These statistics are normalized to adjacent muscle or parenchyma so that they capture enhancement and flow rather than global gain.

Lumen geometry, meaning area, minimum diameter and the angular extent of tumor contact, is deliberately kept out of $b$. It stays available to the model.

Registration error is itself a nuisance. We record the residual registration error with each pair and test whether results hold when pairs above an error threshold are excluded. This matters because Lin et al. note that subtraction is sensitive to motion.

Audit. For any trained bright-blood classifier $f$, we compute three reliance measures:

  • Lumen attribution share: the fraction of attribution mass that falls inside the lumen mask. Attributions are computed with Integrated Gradients and with occlusion.
  • Predictive leakage: how well $b$ can be decoded from the penultimate features.
  • Conditional sensitivity: how much $f$'s output varies with $b$ among lesions matched on wall measurements.

Training. Let $z$ be the image representation and $y$ the label. We minimize

$$\mathcal{L} = \mathcal{L}_{\text{task}}(f(x_{\text{bright}}), y) + \lambda\, \hat I(z; b),$$

where $\hat I$ is the MINE estimate of mutual information between the representation and the physically measured luminal intensity, and $\lambda$ sets the trade-off. Black-blood images are needed only at training time. At test time the model takes bright-blood images alone, which matches sites or protocols where black-blood imaging is unavailable.

Tasks. We use two tasks:

  • Intracranial and carotid plaque presence or high-risk plaque. The data are the multi-sequence head-and-neck vessel-wall cohort of Zhang et al., which contains black-blood T1WI, T1WI-CE and TOF-MRA. Labels come from the lesion measurements that cohort already records, plus a new plaque-grade annotation, which is labor we must budget for.
  • Pancreatic-cancer vessel invasion. The data are the CT cohort of Lin et al., with the arterial phase as the bright image and the subtraction as the black-blood image. It has surgical and pathological ground truth on 175 vessels, 30 of them invaded. This is far too small to train on, so it serves as an external audit set for models trained on the vessel-wall data and adapted with few labels.

CARE-II provides additional carotid pairs where both contrasts exist.

Baselines

Naive reference. The do-nothing reference is a standard convolutional network trained by ordinary empirical risk minimization on bright-blood images only, with the same backbone, augmentation, data and hyperparameter-search budget as the proposed model. It shows how much shortcut reliance exists by default. In the closest analogous published setting, this kind of model drops to an AUROC of 45.8 on an inverted split (Müller et al.).

Black-blood-only model. This is the same architecture trained and tested on black-blood images. It is an upper reference for wall-centered reasoning, but not a fair deployment competitor, since it needs the black-blood sequence at test time. We report it so readers can see how much of its robustness the bright-blood regularized model recovers.

Lumen masking. The bright-blood image's lumen, segmented by a vessel-wall segmentation network of the polar CNN type, is replaced by the local wall-adjacent tissue mean before training. This is the cheapest strong competitor and needs no pairing at inference or training beyond the segmenter.

Proxy-label MINE. This is MINE as in Müller et al. with a human-recordable proxy confounder: contrast phase, scanner vendor and site. It is the strongest existing method for this question, because it is exactly the published technique given the labels a hospital could realistically provide. It isolates whether the physical measurement $b$ adds anything beyond metadata.

The strongest number we can cite for this family is the inverted-split AUROC of 87.7 on CheXpert from that paper. That figure is for a different task. We found no reported result on Lacuna for any learned classifier on the vessel-wall or pancreatic-cancer cohorts. For pancreatic invasion, the only reported figure is radiologists' AUC of 0.942 reading black-blood CT (Lin et al.), which we treat as a human ceiling rather than a model baseline.

Concept bottleneck. This baseline predicts the label from measured wall concepts: thickness, eccentricity, stenosis ratio and contact angle, following Yuksekgonul et al.. The concepts are computed from segmentations rather than from language-model vocabularies.

Ablations. Two ablations isolate the key ingredient. The first replaces $b$ with a shuffled $b$ (same marginal distribution, no patient pairing). The second replaces acquired black-blood images with generatively synthesized ones, to test whether the physical acquisition, rather than any lumen-dark image, is what matters.

Criteria for beating the baselines. All models get the same search budget and five seeds. The proposal beats a baseline if it does three things:

  • It improves inverted-split AUROC by a margin whose 95% bootstrap interval excludes zero.
  • It does not lose more than two AUROC points on the uninverted split.
  • It achieves lower lumen attribution share and higher Kendall τ with radiologist region rankings.

Why not just…?

Why not just mask the lumen? This is the real threat. If masking matches the proposal on robustness, the contribution shrinks to the audit protocol, and we will report it that way. We expect masking to fall short in two measurable ways. It deletes lumen geometry, so stenosis- and contact-angle-dependent labels, such as abutment versus encasement in Lin et al., should degrade on the uninverted split. It also depends on segmentation, which Zhang et al. report degrades in distal arteries and complex vasculopathies. We stratify every result by vessel caliber to test whether masking's robustness collapses exactly there.

Why not just train on black-blood images? Where black-blood imaging is always available, this may be the right answer, and the black-blood-only reference tells us so. However, many protocols and sites only acquire bright-blood or contrast-enhanced images at test time. Black-blood imaging also has its own failure mode in slow or incompletely suppressed flow, which can mimic wall thickening. The proposal uses black-blood images only during training and reports the gap to the black-blood-only model explicitly.

Why not just inspect Grad-CAM or saliency maps? Saliency on a lumen-adjacent wall is spatially ambiguous at the resolution of these maps. Laugel et al. also show that post-hoc explanations often rest on unsupported regions. We test this directly. If lumen attribution share from the naive model fails to predict its inverted-split degradation across seeds and tasks, while the paired conditional-sensitivity measure does, that shows that visual saliency inspection is not a sufficient audit.

Why not just use a concept bottleneck? Concept models are interpretable by construction, but they cap performance at whatever the concepts capture. They also inherit segmentation errors in every concept. We run them as a baseline. If they match the proposal's robustness at equal accuracy, we will recommend them, and the paired measurement will still serve as their audit.

Proposed Experiments and Decisive Checks

Planted-shortcut test. We construct an inverted split in the style of Müller et al.. Within the vessel-wall cohort, we subsample training lesions so that the label correlates with high normalized luminal enhancement (for instance, early-phase acquisitions among positives), then test on a split with the correlation reversed. The ground truth about which signal is spurious is known by construction. This makes it the decisive test of both the audit, which should flag the naive model, and the training penalty, which should restore performance.

Natural-shift test. We train on some centers and test on others, using the center differences in contrast timing and vendor that Zhang et al. span. This shows whether the effect matters without planting anything.

External audit. On the pancreatic cohort of Lin et al., we measure each model's vessel-invasion AUC against pathology alongside its conditional sensitivity to luminal enhancement. The goal is to test whether lower reliance accompanies better invasion staging. With 30 positive vessels, we report intervals, not point claims.

Alignment test. Two radiologists rank the regions that drive their reads: lumen, inner wall, outer wall and perivascular tissue. We compute Kendall τ between those rankings and each model's region-aggregated attributions, following Haddada et al., and compare against inter-reader τ as the ceiling.

Outcomes. The study succeeds if three things hold. The physically measured confounder must beat the proxy-label MINE baseline on inverted and cross-center AUROC. The shuffled-pairing ablation must erase that gain. And lumen masking must lose on geometry-dependent labels or in small vessels. If the shuffled ablation performs as well, the benefit came from regularization, not from the physics. If masking matches everywhere, the method's value reduces to diagnosis rather than cure. Either outcome would be reported as the finding.