Disease registries count only the cases that someone confirmed. Epidemiologists already correct these counts for the earlier steps in the funnel: who got sick, who showed symptoms, who sought care, and who was tested. The last step is usually treated as a fixed, perfect gate. A pathologist or a model looks at tissue and says yes or no. Increasingly that look happens remotely. A gigapixel slide image, roughly 100,000 by 100,000 pixels, is compressed and sent over a shared mobile or edge network to an expert or a model elsewhere.
We argue that this turns network congestion into a hidden form of ascertainment bias, meaning distortion in which true cases get counted. When the link is busy, slides are compressed harder, small diagnostic features such as isolated tumor cells or dividing nuclei are blurred, and the chance of confirming a true case drops. Because that chance now rises and falls with network load, it varies by site and by hour.
The proposal has two parts. First, measure how confirmation sensitivity depends on the bitrate a slide receives, for both human readers and models. Second, build a slicing policy that allocates bandwidth so that a correctable, low-variance incidence estimate survives at the registry. Network slicing means carving one physical network into isolated virtual networks with guaranteed resources. The goal of the policy is not to maximize per-slide accuracy or the number of tasks the network admits.
The core claim is a design principle. For surveillance, a known and stable sensitivity is worth more than a high but fluctuating one, because the first can be divided out and the second cannot. If the claim holds, slicing for medical imaging should be judged by the bias and variance of the downstream incidence estimate, not by per-task accuracy agreements.
What the literature gives us and where it stops
The surveillance work supplies both the correction machinery and the motivating failure modes.
Two papers back-calculate true incidence by dividing observed counts by assumed funnel probabilities. On the origin and timing of Zika virus introduction in Brazil scales reported Zika cases by an assumed subclinical fraction. Estimating incidence of sexually transmissible infections in Australia builds a Monte Carlo probability tree conditioned on testing volume. The Australian study also shows the damage a non-stationary reporting step can do: a 2006–2007 Medicare billing-code change invalidated two years of screening series. A congestion-dependent confirmation step is the same kind of hazard, but it changes on a timescale of minutes rather than years. Both papers treat test performance as a static parameter, and that is the assumption we relax.
Accurately Estimating Unreported Infections using Information Theory (MDLInfer) addresses a different problem. Compartmental epidemic models fit reported curves equally well under very different parameter values, and MDLInfer resolves this by choosing the parameters that compress the observed data most. It reports 1.2× to 7.21× lower error than standard calibration against CDC serology. It too assumes the case labels it fits are correct. In our design it is a downstream consumer whose robustness to label noise we test; it is not a competitor.
The pathology work shows that the diagnostic gate is noisy and that models can make it more consistent. Clinical utility of receptor status prediction in breast cancer used a model as a second reader on H&E slides. It reclassified 4.45% of ER-negative and 5.98% of PR-negative clinical diagnoses as positive and found 31 misdiagnoses. This gives an empirical scale for the false-negative rate of human confirmation.
Multiple instance learning (MIL) treats a slide as a bag of thousands of small patches and learns a slide-level label from them. Navigating the MIL Trade-Off shows that a cheap pooling rule called Maxsoft, paired with a domain-specific encoder, reaches WSI-level AUC 1.000 on CAMELYON17 using about 1,000× less computation than a large foundation model. Attention-based MIL with Mixed Supervision shows that slide-level accuracy can be high while localization stays poor: mean Dice is only 0.215 to 0.405 on Camelyon16. That gap matters here. It suggests the model's decision rests on a few small regions, which are exactly what heavy compression destroys.
Automatic and explainable grading of meningiomas shows that a mitotic-count feature alone predicts grade in 32 of 43 cases. Counting tiny, sparse mitotic figures is a natural stress test for compression.
On the network side, the closest prior work is SEM-O-RAN. It slices radio, GPU, and memory resources in an open radio access network (O-RAN, a standard that splits base stations into open software components) using each vision class's tolerance to compression. On the Colosseum emulator it admits on average 18% more deep-learning tasks than the Sl-EDGE baseline, and up to 169% more. It evaluates only natural scenes, and its objective is per-class accuracy agreements on each task. We change both the data, to gigapixel slides read by MIL models and humans, and the objective, to the statistical quality of an aggregate estimate several steps downstream. The O-RAN tutorial survey supplies the control timescales we design around. The Near-Real-Time controller runs at 10–1000 ms and the Non-Real-Time controller at one second or slower; our policy belongs in the slower loop. SLICES is a candidate open testbed for reproducing the work.
Core Claim
The confirmation sensitivity at site $k$ depends on the bitrate $b_k$ its slides receive, through a curve $s_k(b)$. Its confirmed count is roughly $C_k \approx I_k T_k s_k(b_k)$, where $I_k$ is true incidence among tested people and $T_k$ is testing volume.
If $b_k$ is logged with every slide and the curve $s_k$ is known, the corrected estimate $\hat I_k = C_k / (T_k\, s_k(b_k))$ is unbiased. Its variance still grows as sensitivity falls. Under a binomial approximation, $\mathrm{Var}(\hat I_k) \approx I_k\,(1 - I_k s_k)/(T_k s_k)$, plus a term from uncertainty in the estimated curve, which grows steeply where the curve is steep.
The allocation problem is therefore to minimize total estimator variance:
$$\min_{b_1,\dots,b_K} \sum_k w_k \,\mathrm{Var}(\hat I_k) \quad \text{subject to} \quad \sum_k b_k \le B,$$
where $w_k$ weights each site's public-health importance and $B$ is the total bandwidth. This looks like Neyman allocation from survey sampling, which assigns more effort to the noisiest strata. The solution sends bandwidth to sites where sensitivity is low and the curve is steep. These are often low-volume, under-served sites, not the sites with the most traffic. The prediction is that a variance-aware policy will differ systematically from both accuracy-agreement slicing and traffic-proportional slicing.
Research Program
The program has three linked pieces.
The first is measurement. We build compression-to-sensitivity curves for MIL models on CAMELYON16, CAMELYON17, and TCGA-Lung, and for mitotic counting on the meningioma cohort. We vary JPEG2000 and HEVC bitrates and also test region-of-interest compression guided by attention maps. A small reader study adds human curves, with several pathologists reading the same Camelyon slides at three bitrates. This yields inter-observer confusion matrices as a function of bitrate, separating disagreement caused by compression from disagreement that exists at full quality. No reported result was found on Lacuna for diagnostic accuracy of whole-slide classifiers as a function of transmission bitrate; this curve is the first deliverable.
The second is a surveillance simulator. It extends the Australian probability tree with a diagnosis stage whose sensitivity is drawn from the measured curves and is driven by network-load traces. Load comes either from Colosseum or from a trace-driven emulator configured like the SEM-O-RAN setup.
The third is the slicing policy. It runs in the Non-Real-Time controller and re-solves the allocation above each control period. It uses the measured curves and current per-site testing volumes, which registries already hold.
Baselines
The naive reference point is a shared network with static priority and uniform JPEG2000 compression, with no bitrate logging. The registry applies the standard tree correction with a fixed sensitivity. This is what passive surveillance would do today if telepathology simply spread, and it is the scenario in which we expect bias that tracks congestion.
The strongest statistical competitor is log-and-correct. Allocation is traffic-proportional, the bitrate of every slide is recorded, and the registry divides by the measured $s_k(b_k)$. This is the honest threat: it removes bias for free, and our method must beat it on variance and on interval coverage, not on bias.
The strongest network competitor is SEM-O-RAN-style slicing re-targeted to pathology. Each site gets an accuracy agreement on the slide-level AUC from the measured curves, and admitted tasks are maximized, with the same log-and-correct estimator downstream. For reference, SEM-O-RAN reports an 18% average and up to 169% gain in admitted tasks over Sl-EDGE on Colosseum.
The strongest diagnostic model is the reader whose curves we measure, which is shared by every arm. We use Maxsoft with a domain encoder, the setting reported at WSI AUC 1.000 on CAMELYON17. We also use CLAM, whose Camelyon16 slide AUC is reported as 0.895, or 0.924 with 10% patch annotations, in the mixed-supervision study.
The ablation that isolates our key ingredient keeps the full policy but sets the curve-uncertainty term to zero. This tests whether the gain comes from modeling where sensitivity is steep or merely from moving bandwidth around.
All arms share the same slides, load traces, total bandwidth budget, model weights, and simulator seeds. Policy hyperparameters get the same search budget in every arm. Beating log-and-correct means at least a 20% reduction in mean squared error of site-level incidence at equal bandwidth, with 95% interval coverage within two points of nominal. Beating SEM-O-RAN-style slicing means a lower incidence error even if it admits more slides.
Why not just…?
Why not just log the bitrate and correct after the fact? This is the baseline we take most seriously, and it may be enough when bandwidth is plentiful. Our claim is only that it leaves variance on the table under congestion. When sensitivity at a low-volume site falls on the steep part of its curve, the corrected estimate becomes very noisy. The simulator directly measures the regime where this happens. If log-and-correct matches our policy across realistic load traces, we report that the network layer needs only metadata, not new slicing. That would still be a useful finding, because current registries log neither bitrate nor reader.
Why not just run the model at the scanner site and send only the label? Maxsoft's low compute cost makes on-site inference plausible. If the model were the final reader, the transmission problem would largely disappear. Registries, however, record human-confirmed diagnoses, and remote expert review is the point at which compression meets inter-observer variability. The receptor-status study uses its model as a second reader, not a replacement. We therefore include a hybrid arm: on-site inference, with only patches the model is uncertain about sent to a remote reader. Our allocation then governs those uncertain patches. If this hybrid alone eliminates the variance gap, it becomes the recommended design.
Why not just use per-task accuracy slicing as SEM-O-RAN does? Accuracy agreements protect each slide individually. They spend bandwidth where traffic is heavy and are blind to how the errors add up across a region. We predict they over-serve high-volume sites whose estimates are already precise. The comparison above tests exactly this, and a null result would mean the two objectives coincide in practice.
Why not just fix sensitivity with stricter consensus rubrics? Consensus panels reduce disagreement at full image quality but cannot restore features that compression removed. The reader study separates these two sources of error. The meningioma study also suggests that manual counting of sparse events is itself error-prone.
Proposed Experiments and Decisive Checks
The first check is whether bandwidth-induced bias exists at all. On the measured curves, does slide-level sensitivity fall by more than the human false-negative scale of a few percent, roughly 4–6% in the receptor-status study, at bitrates a congested cell would realistically impose? If the curves stay flat down to a 50:1 compression ratio, the problem is not real and we stop.
The second check is the variance comparison. In the simulator we run 10,000 Monte Carlo iterations, matching the Australian tree. Each run uses ground-truth incidence and evaluates incidence mean squared error, bias, and coverage across all arms under load traces with congestion spikes.
The third check is downstream robustness. We feed each arm's registry series to MDLInfer and to a standard least-squares compartmental fit. We then ask whether variance-aware allocation keeps the inferred unreported-infection totals stable when congestion shifts mid-epidemic. This tests whether a fluctuating confirmation step creates the same kind of trend artifact that the billing-code change created in the Australian chlamydia series.
The fourth check is a feasibility run of the policy on Colosseum or SLICES. It verifies that re-solving every control period fits within the one-second-or-slower Non-Real-Time timescale. Because the policy needs no device-side feedback, it also avoids the missing feedback channel that SEM-O-RAN identifies in current O-RAN specifications.
References
- On the origin and timing of Zika virus introduction in Brazil
- Estimating incidence of sexually transmissible infections in Australia
- Accurately Estimating Unreported Infections using Information Theory
- Clinical utility of receptor status prediction in breast cancer and misdiagnosis identification using deep learning on hematoxylin and eosin-stained slides
- Navigating the MIL Trade-Off: Flexible Pooling for Whole Slide Image Classification
- Attention-based Multiple Instance Learning with Mixed Supervision on the Camelyon16 Dataset
- Automatic and explainable grading of meningiomas from histopathology images
- SEM-O-RAN: Semantic and Flexible O-RAN Slicing for NextG Edge-Assisted Mobile Systems
- A Comprehensive Tutorial and Survey of O-RAN: Exploring Slicing-Aware Architecture, Deployment Options, Use Cases, and Challenges
- SLICES, a scientific instrument for the networking community