When the Scout Goes Blind: Drift-Aware Pruning for Gigapixel Pathology Under Hospital Shift

Fast whole-slide pathology pipelines save compute with a "scout." A cheap network looks at a low-resolution overview of a gigapixel slide and decides which regions deserve expensive high-resolution analysis. The rest, often 85 to 90 percent of the tissue, is never examined at full resolution. This works well when test slides come from the same hospital as the training slides. Deep networks, however, are known to amplify small statistical mismatches layer by layer. When a slide arrives from a new hospital with a different scanner or stain, the scout's internal features drift, and the drift grows with depth. In an ordinary classifier this costs a few points of accuracy. In a scout cascade it does something worse: a small lesion wrongly scored as background is simply never looked at again, and no later stage can recover it.

This note proposes to measure that failure directly and to fix it with a small, testable mechanism. The scout reads its own internal statistics on each incoming slide, turns their mismatch with the training hospitals into a label-free drift score, and enlarges its high-resolution budget on slides where it is likely to be unreliable. The claim to test is specific. At the same average compute, a drift-conditioned budget recovers more held-out-hospital lesions than any fixed budget, and more than simply repairing the scout's normalization statistics at test time.

Core Claim

Cumulative feature degradation, meaning the super-additive growth of statistical mismatch through network depth, has been documented only in single-stage classifiers. Three papers establish it. Delving into the Estimation Shift of Batch Normalization in a Network measures how far estimated normalization statistics sit from ideal ones and finds that the gap grows much faster in deeper layers. Reducing Divergence in Batch Normalization for Domain Adaptation shows the same growth under source-to-target domain shift on 224-pixel natural images. Testing the "Benign Running Statistics" Hypothesis shows that swapping statistics across 20 layers of a ResNet is nearly three times more damaging than the sum of single-layer swaps. These papers motivate the proposal.

Scout cascades are a different setting. AHDMIL distills a high-resolution teacher's attention into a low-resolution pre-screening network, and SEW uses a thumbnail "Focus Predictor" to pick the top-K regions. Both convert a continuous feature error into a hard, irreversible pruning decision. Neither evaluates across hospitals, and SEW's authors name the Focus Predictor's tendency to miss subtle lesions as open work.

Our hypothesis has three parts. First, under cross-center shift, the scout's recall of lesion regions falls by more than its in-distribution error would predict. Second, that loss tracks a measurable, label-free drift in the scout's layer statistics. Third, using the drift to set the per-slide budget beats both fixed budgets and statistics repair on the recall-versus-compute trade-off.

Research Program

Measurement study. We build scout cascades in the style of AHDMIL and SEW, using the public hierarchical attention-guided MIL code BearCleverProud/HAG-MIL as a cascade scaffold. MIL (multiple instance learning) here means learning a slide-level label from a bag of unlabeled patches. Where the AHDMIL scout is not publicly available, we reimplement it from the paper.

We train on CAMELYON17 in a leave-one-center-out protocol, holding out each of the five hospitals in turn. The central quantity is cascade lesion recall: the fraction of annotated metastatic patches that survive pruning. We report it separately for slides containing only micrometastases, since those are the cases that lack low-magnification hallmarks.

To transfer the degradation analysis into this setting, we compute per-layer statistic gaps between the held-out center and the training centers. We then repeat the layer-wise substitution ablation of Testing the "Benign Running Statistics" Hypothesis inside the scout: swap source statistics for target statistics one layer at a time and then jointly, and check whether recall loss is super-additive, as accuracy loss was in that paper.

Mechanism. For a slide $s$, let $\mu_l(s)$ and $\sigma^2_l(s)$ be the channel means and variances of scout layer $l$, computed over that slide's own thumbnail tiles. Let $\mu_l$ and $\sigma^2_l$ be the pooled training statistics. The drift score and the per-slide budget are:

$$ D(s) = \sum_{l=1}^{L} w_l \left( \frac{\lVert \mu_l(s) - \mu_l \rVert_2}{\lVert \mu_l \rVert_2} + \frac{\lVert \sigma^2_l(s) - \sigma^2_l \rVert_2}{\lVert \sigma^2_l \rVert_2} \right), \qquad k(s) = \min\big(N(s),\; k_0 \cdot g(D(s))\big) $$

Here $L$ is the number of normalization layers, $w_l$ are layer weights, $N(s)$ is the slide's tissue-patch count, $k_0$ is the base number of high-resolution patches, and $g$ is a monotone map.

The weights $w_l$ and the map $g$ are fitted without target labels. Inside the training centers we hold out one center at a time, simulate an unseen hospital, and regress the observed recall loss on per-layer drift. We then choose $g$ so that the average budget over a slide set equals a target compute level. We include the variance term separately because the variance mismatch was the more destructive and less explained component in the substitution study, which recorded 27.2% accuracy for variance substitution alone versus 76.7% for mean substitution alone. The only extra cost is one pass of the scout, which already runs on every slide.

Interaction with statistics repair. An obvious fix is to recompute the scout's normalization statistics on each target slide. The cancellation identity in Testing the "Benign Running Statistics" Hypothesis warns against this in one case. When statistics are dominated by one class, re-centering subtracts exactly the signal the scoring head expects, which dropped CIFAR-10 accuracy from 91.1% to 3.5%. For slides with large tumor burden, per-slide statistics are tumor-skewed and could suppress tumor scores. We therefore test per-slide repair as a competitor, both plain and with a trimmed version that excludes the highest-scoring tiles. We also test whether drift-aware budgeting adds anything on top of it.

Baselines

Full tiling (no scout). Every tissue patch goes to a frozen encoder with Maxsoft pooling, following Navigating the MIL Trade-Off. Maxsoft uses max pooling in the forward pass and log-sum-exp gradients in the backward pass, trained with independent per-patch augmentation. This is the accuracy ceiling and the compute worst case. It is the strongest existing method on this benchmark: that paper reports a CAMELYON17 WSI-level AUC of 1.000 and a patch-level AUC of 0.948 with Prov-GigaPath. It also reports that a DINO domain encoder matched it on several metrics at 1,000× less compute.

Because slide AUC is already saturated, "beating" this baseline cannot mean higher AUC. It means matching its slide-level AUC and micrometastasis sensitivity within confidence intervals, under our leave-one-center-out split, while processing a fraction of the patches. We rerun it on our split rather than quoting it, since its split may differ.

Uniform random subsampling. We draw 512 patches per slide, as in eWSI, which reports 98.5% AUC on CAMELYON16 with an in-domain encoder but notes calibration bias from sparse sampling. This is the naive "do nothing clever" reference.

Fixed-budget scouts. These are AHDMIL-style and SEW-style scouts at a swept set of fixed budgets. For AHDMIL, the reported results are a 1.2× to 2.1× speedup and a 4.8% AUC gain over TransMIL on CAMELYON16, where TransMIL reports 93.09%. No reported result was found on Lacuna for AHDMIL or SEW on CAMELYON17.

Scout-hardening variants at a fixed budget. These are the scout with Macenko stain normalization, with per-patch augmentation, with a group-normalization circuit breaker in deeper blocks as in RBNBlock, and with per-slide statistics repair.

Fairness. All methods share the same high-resolution encoder, aggregator, training centers, and hyperparameter search budget. Compute is reported as high-resolution patches encoded per slide and as wall-clock time on one GPU.

Ablations isolating the key ingredient. The first replaces $D(s)$ with a random score that has the same budget distribution. The second replaces it with the scout's own output entropy, which tests whether internal statistics carry information beyond confidence. The third uses mean-only and variance-only drift.

The proposal wins if its recall-versus-compute curve on held-out centers lies above the frontier traced by every fixed-budget and hardened scout. The target is at least a 25% reduction in missed micrometastatic slides at matched average compute. It must also close most of the gap to full tiling at no more than half its patch count.

Why not just…?

Why not just give the scout a bigger fixed budget? This is the most serious threat, because a larger fixed budget will recover lesions. The question is whether spending compute uniformly is wasteful. If drift concentrates on a subset of slides, such as one scanner, conditioning the budget should dominate. If drift is uniform within a held-out center, a fixed per-center budget is equivalent, and the matched-compute frontier comparison will show that plainly. In that case the contribution shrinks to a label-free way of choosing the budget per hospital, which is still useful but smaller.

Why not just recompute the scout's normalization statistics on each target slide? This is cheap and may repair most of the drift. It is a real competitor and is run as a baseline. Two things argue it will not suffice. The cancellation identity predicts harm on tumor-dominated slides. And Delving into the Estimation Shift found that removing batch dependence entirely hurt accuracy relative to hybrid designs. If repaired scouts show no residual recall loss, we will report that drift-aware budgeting is unnecessary and that repair is the answer.

Why not just stain-normalize, or train the scout with per-patch augmentation? Both act at the input. The results on internal compounding in Reducing Divergence in Batch Normalization and the failure of uniform slide-level augmentation in Navigating the MIL Trade-Off suggest that residual drift survives. We include both as hardened scouts and measure the drift that remains after them.

Why not just drop the scout and use a compact encoder on everything? Maxsoft with a DINO encoder is already cheap. If full tiling with that encoder runs within the same wall-clock time as a scout cascade, the problem vanishes, and we will report the break-even point.

Decisive Checks

The first check can kill the project. If held-out-center lesion recall of fixed-budget scouts is not measurably below their in-distribution recall, or if the loss does not correlate with $D(s)$ across slides (Spearman correlation below roughly 0.3), the premise fails.

The second check asks whether the layer-wise substitution ablation shows super-additive recall loss inside the scout. That would extend the degradation result from single-stage accuracy to cascade recall.

The third check is the matched-compute frontier comparison against fixed budgets and statistics repair. It is judged on micrometastasis-only held-out slides with bootstrap confidence intervals over slides and all five center rotations. Results are secondarily confirmed on TCGA-IDH, a set of glioma slides from eight institutions, under a leave-one-institution-out split.