When a medical imaging model goes live, its behavior changes, and the people running it usually cannot tell why. In a prospective mammography triage deployment, a threshold set on historical data to flag 14% of exams flagged 29% once it ran on live patients (Triaging mammography with artificial intelligence). That jump has two possible causes. The screened population may really contain more suspicious findings, which is benign. Or the live data pipeline may be feeding the model images that differ from the training images (different compression, pixel spacing, or vendor post-processing), which is dangerous. A study that took the trouble to separate these causes for a hospital infection model found that the pipeline mismatch, which it called infrastructure shift, cost 0.016 AUROC. Passage of time actually added 0.005 (Mind the Performance Gap). That attribution needed a third cohort pulled months later, and it needed outcome labels. For breast cancer, outcome labels arrive weeks to months after the scan, because biopsy confirmation alone took 39 to 56 days on average in the triage study.
We propose using the internal structure of a foveated vision model as a label-free instrument for that attribution. A foveated model imitates the eye: it looks at a coarse, low-resolution view of the whole image (the "periphery"), chooses a few spots to look at closely, and examines those spots at full resolution (the "fovea"). The two views have different sensitivities. Pipeline faults such as lossy re-encoding, wrong pixel spacing, or altered sharpening mostly damage fine detail. That means they change what the fovea sees while leaving the periphery's coarse picture largely intact. A real shift in the population changes both views together. So the agreement between what the periphery expected to find at a fixated spot and what the fovea then found there should drop under pipeline faults and stay stable under real population change. That agreement can be computed on every exam, immediately, with no labels.
Core Claim
We claim that, in a peripheral–foveal mammography model, the disagreement between the two streams at the fixated locations is a specific, early, label-free signal of infrastructure shift. We further claim it separates infrastructure shift from changes in disease prevalence and case mix better than three alternatives: monitoring the output score distribution, monitoring feature drift, or comparing a standard model's predictions on low- and full-resolution copies of the whole image.
The second claim matters because of an obvious competitor. One could take any uniform-grid model and compare its predictions on a downsampled copy and a full-resolution copy of the whole image. Our hypothesis is that whole-image comparison dilutes the signal. Most of a mammogram is fat and stroma where fine-scale corruption barely changes the prediction. A foveated model measures disagreement only where it chose to look closely, which is where high-frequency detail such as roughly 10×10-pixel microcalcifications decides the answer (Mammo-Clustering). If this hypothesis fails, the contribution shrinks to "cross-resolution consistency is a useful monitor." We design the experiments so that this outcome is visible.
A secondary claim concerns compute. Foveated models with confidence-based early stopping, such as FoveaTer, keep taking glimpses until the predicted class is confident enough. Under shift, raw confidence drifts in the way Mind the Performance Gap documented for calibration. That drift can cause either stopping too early or near-unbounded looping. We propose gating the stopping rule on cross-stream agreement rather than on the model's own confidence, so that the glimpse budget stays bounded and suspicious exams are routed to a human instead of absorbing more compute.
Research Program
The model builds on what is already known to work. Its periphery is a downsampled whole-breast view across the four standard views. It selects fixations from a peripheral saliency map. Its fovea encodes native-resolution crops. This mirrors the global, feature-local, and patch-local levels of Mammo-Clustering and the patch-to-breast design of Deep Neural Networks Improve Radiologists' Performance in Breast Cancer Screening. The one architectural change is that the periphery must also predict, for each fixation $i$, a malignancy probability $p_i^{\text{per}}$ for that location before the fovea sees it. The fovea then returns its own estimate $p_i^{\text{fov}}$.
We train the periphery's prediction to match the fovea's estimate on clean training data. The per-exam consistency score is
$$c(x) = \frac{1}{K}\sum_{i=1}^{K} \mathrm{KL}\big(p_i^{\text{fov}} \,\|\, p_i^{\text{per}}\big),$$
where $x$ is the exam, $K$ is the number of fixations taken, and KL is the Kullback–Leibler divergence between the two Bernoulli predictions. On a stream of exams, we monitor the running distribution of $c(x)$ alongside the distribution of final risk scores. The two signals fall into four diagnostic quadrants. If both are stable, nothing has changed. If risk scores shift while $c$ stays stable, we read that as population change. If $c$ rises, whatever the risk scores do, we read that as infrastructure shift. If both rise, we read that as mixed causes.
For the fovea, we start from a pretrained backbone adapted with low-rank updates, following FOVI. FOVI retained over 96% of full-resolution ImageNet accuracy with 1/16 of the pixels. It also found that moderate foveation beats uniform subsampling at matched pixel budgets, which is the property our hypothesis about undiluted signal relies on.
The program has three stages. The first trains the model on Vindr-Mammo and checks that it is competitive with the published results; the monitor is only worth having if the model is clinically credible. The second is a controlled shift study, in which we apply known infrastructure faults and known prevalence changes to held-out data and measure detection and attribution. The third is a silent prospective deployment at a partner screening site. The model scores live DICOM traffic without influencing care, and we reproduce the three-cohort design of Mind the Performance Gap: the training cohort, the live-pipeline cohort, and the same months re-extracted later from the research archive. The re-extracted cohort gives a ground-truth estimate of how much shift came from the pipeline, against which the label-free score can be checked after outcome labels mature. Stage three depends on a clinical partner and ethics approval. Stages one and two stand on their own as a publishable contribution.
Baselines
The naive reference point is a fixed threshold with no monitoring. This is the configuration that produced the 14%-to-29% alert surge in Triaging mammography with artificial intelligence. Any monitor must at least detect a shift of that size.
The first real competitor is monitoring the output score and the alert rate. It is what deployed systems actually watch, and it costs nothing. It detects that something changed but, by construction, not why. It wins if it attributes shifts as well as $c(x)$. We run it on the same exam streams with the same false-alarm calibration.
The second competitor is feature-drift monitoring: a distance between the distribution of penultimate-layer features on live data and on training data, computed for a uniform model and separately for each of our two streams. Running it on each stream separately is our main ablation. It tells us whether the useful information lies in the paired disagreement or merely in having a fine-detail feature space to watch. I found no reported result on Lacuna for feature-drift shift detection on mammography.
The third and most dangerous competitor is uniform two-resolution consistency. We take a strong uniform model and measure the disagreement between its whole-image predictions at low and full resolution, giving it the same total compute as our model. If it matches us, foveation is not what makes the method work.
For diagnostic accuracy, the strongest existing method on our public data is Mammo-Clustering, which reports AUC 0.828 ± 0.020 on Vindr-Mammo and 0.805 ± 0.020 on CBIS-DDSM (Mammo-Clustering). Our model should come within that error band. Being slightly below it would be acceptable. Being clearly below it would mean we are monitoring a model nobody would deploy.
Every monitor gets the same calibration data, the same false-alarm target, and the same hyperparameter search budget.
Why not just…?
Why not just watch the alert rate and recalibrate the threshold? Recalibrating the threshold restores the alert volume. But if the cause was a pipeline fault, recalibrating hides real sensitivity loss instead of fixing it. Getting the labels to tell the two cases apart takes weeks to months (Triaging mammography with artificial intelligence), and decomposing the cause took a retrospective re-extraction in Mind the Performance Gap. The controlled shift study tests directly whether score monitoring can tell a prevalence change from a JPEG2000 re-encoding that moves scores by the same amount. Our hypothesis is that it cannot.
Why not just check DICOM headers and pipeline metadata? This is cheap and should be done anyway. But many harmful faults leave the headers untouched, for example vendor processing changes or silent lossy compression in an intermediate image router. We include faults that preserve the headers specifically, and we report header checks as a complementary baseline rather than claiming to replace them.
Why not just compare any model's low-resolution and full-resolution predictions? This is the real threat, because it captures the same physical intuition without a foveated architecture. Our answer is an empirical bet: whole-image disagreement is averaged over mostly uninformative tissue, while disagreement at fixations focuses on the regions where corruption changes the diagnosis. The experiments use matched compute and report detection delay for both methods. If uniform two-resolution consistency matches us, we will report that cross-resolution consistency, not foveation, is the finding.
Proposed Experiments
The controlled shift study uses held-out Vindr-Mammo exams and transfers to CBIS-DDSM, whose digitized film scans are a natural infrastructure contrast. We inject infrastructure faults at graded severity: lossy re-encoding, pixel-spacing mislabeling that changes the physical crop size, intensity-windowing and sharpening changes, and dropped views. Separately, we inject population changes by resampling to raise BI-RADS-positive prevalence, shifting breast-density mix, and changing age distribution. Each fault is calibrated so that it moves the alert rate by the same amount as some prevalence change. This matching is what makes attribution non-trivial.
We measure three things. The first is detection delay: how many exams pass before an alarm, with the false-alarm rate fixed on clean streams. The second is attribution accuracy: whether each shift is assigned to the correct quadrant. The third is how well the change in $c$ predicts the change in AUC across fault severities. We also report the distribution of the glimpse budget and sensitivity under the confidence-gated and agreement-gated stopping rules, both on clean data and under shift.
There are three decisive checks. First, under faults matched for alert rate, $c(x)$ must separate infrastructure faults from prevalence changes where score monitoring cannot. Second, $c(x)$ must beat uniform two-resolution consistency at equal compute, most clearly for the faults that degrade microcalcification detection; if it does not, the foveation claim is withdrawn. Third, under agreement gating, the glimpse count must stay bounded while sensitivity holds, whereas confidence gating either stops too early or runs away.
In the silent prospective stage, success means that the share of the AUROC gap our label-free score attributes to infrastructure, estimated in real time, agrees in sign and rough size with the share recovered from the later re-extracted cohort once labels mature. We would also log pipeline feed failures, which ran 10 out of 356 days in Mind the Performance Gap, and check whether $c$ rises before each one.
References
- Triaging mammography with artificial intelligence: an implementation study
- Mind the Performance Gap: Examining Dataset Shift During Prospective Validation
- Mammo-Clustering: Context Clustering based Multi-view Tri Level Information Fusion for Lesion Location and Classification in Mammography
- FoveaTer: Foveated Transformer for Image Classification
- Deep Neural Networks Improve Radiologists’ Performance in Breast Cancer Screening
- FOVI: A biologically-inspired foveated interface for deep vision models