Pathology slides are gigapixel images, so the strongest classifiers do not read them at full resolution. They first look at a cheap, low-magnification version, score each square tile, and zoom only into the few tiles that look most suspicious. Anything the first pass does not pick is never examined at high resolution. This proposal argues that the zoom step should be designed as a budgeted coverage problem with feedback, rather than as a fixed ranking of tile scores.
Two predictable failure modes make the current ranking design fragile, and both have direct analogues in wireless network engineering. First, a small tumor deposit that a tile edge cuts in half shows each tile only a fraction of the lesion, so both halves can score below the cutoff. Sensor-network theory calls this a boundary effect: a node at an edge or corner loses half to three quarters of its neighborhood. Second, benign tissue that looks like cancer (a "morphological mimic," such as lymphoid follicles in a lymph node) can soak up the whole zoom budget, crowding out the real lesion. This resembles how many transmitters packed into one area interfere with a weak distant signal. We propose to measure how often each failure happens. We then repair it with three pieces: boundary-corrected coarse scores, a per-region cap on how much zoom budget one cluster of look-alike tissue can take, and a recovery loop that lets high-resolution evidence pull in neighboring tiles the first pass skipped.
Core Claim
In hierarchical multiple instance learning (MIL), a weakly supervised setup where a slide label is predicted from a bag of unlabeled tiles, many slide-level false negatives on small lesions come from selection misses rather than from classification errors at high resolution. These misses cluster on two kinds of lesion: those split by tile boundaries, and those that compete with spatially concentrated mimics. Treating zoom selection as "maximize the chance that every lesion is examined, under a fixed budget" should reduce these misses at equal compute. It should also make the classifier stable when the tile grid is shifted, which current methods do not test.
A useful theoretical lens comes from Parametric estimation for Gaussian fields indexed by graphs. That paper corrects boundary bias in spatial estimators with a taper, but only for sets whose boundary is negligible relative to their volume. A micrometastasis that covers two or three coarse tiles is the opposite case: it is almost entirely boundary. This is why we expect boundary truncation to matter most for small lesions and to wash out for large ones. The prediction is testable by stratifying results by lesion size.
Relation to Prior Work
The closest work is Diagnose Like a Pathologist: Transformer-Enabled Hierarchical Attention-Guided Multiple Instance Learning for Whole Slide Image Classification (HAG-MIL, code at BearCleverProud/HAG-MIL). It builds a multi-resolution quadtree and zooms into the top-$k$ tiles by attention at each level. The second closest is Towards Hierarchical Regional Transformer-based Multiple Instance Learning, which pools fixed 2×2 and 4×4 regions and, at inference, re-runs the model only on high-attention clusters found by K-means. Both use hard, one-shot gates on a fixed square grid. Neither reports how often true lesions are pruned, how predictions change when the grid moves, or how mimics compete for the selection budget. We keep their backbones and change only the selection mechanism, so any gain is attributable to selection.
Prior work also shows that the cascade problem is real and general. A Part-Based Deep Neural Network Cascade Model for Human Parsing documents that a wrong coarse crop is unrecoverable downstream. Content-Based Retrieval from Medical Image Databases shows the same irreversibility when a root-level disease classification is wrong.
The same human-parsing paper also constrains our design. Its "refinement paradox" finding is that smoothing with conditional random fields hurt accuracy because it erased thin, local features. We therefore avoid any spatial smoothing of tile scores. Our boundary correction operates by pooling evidence across neighboring tiles before selection, not by smoothing outputs after it.
Two lines of work show how to move away from single-path commitment. Joint Object and Part Segmentation lets coarse and fine predictions update each other. Chu et al. keep a beam of candidate decompositions. Our recovery loop is a cheaper, spatially targeted form of the first idea: fine evidence reopens only the immediate neighbors of a positive tile, rather than updating everything.
The method borrows ideas from wireless networks. From Results on Finite Wireless Sensor Networks: Connectivity and Coverage we take the practice of treating interior, edge, and corner positions separately rather than assuming every location sees a full neighborhood. That paper shows asymptotic, edge-free formulas can be badly wrong in finite settings. From Interference Reduction in Virtual Cell Optimization we take the idea of a distance threshold that defines which units interfere, together with a budget constraint across those units. The same paper warns that making the threshold too large wastes capacity, with sum-rate falling past roughly 140 m in their setting. That warning tells us to sweep the cap rather than fix it. When Pilots Should Not Be Reused Across Interfering Cells in Massive MIMO motivates partitioning a shared resource hierarchically by geometry, but its proofs assume an idealized infinite lattice, which is exactly the assumption we drop.
Research Program
The work has three components, each built as a drop-in replacement for the top-$k$ step in HAG-MIL, plus one diagnostic.
Boundary-corrected coarse scores. Each coarse tile's score is recomputed from a small set of offset sub-grids, at half-tile shifts. We record whether the tile's evidence peaks at its border. A tile whose suspicious content touches an edge is credited with the pooled score of the corresponding neighbor pair. Coarse-level compute is a small fraction of total cost, so this is cheap. It deliberately absorbs the obvious "just use shifted grids" alternative into the method rather than competing with it.
Mimic-aware budget allocation. Coarse tiles are grouped into spatial clusters, joining tiles within distance $d$ whose coarse features are similar. Selection then maximizes total score subject to two limits: the global zoom budget $k$, and a cap of $m$ zooms per cluster. A sprawling field of look-alike follicles therefore cannot consume the whole budget. The leftover budget goes to the best tiles in other clusters. We sweep both $d$ and $m$.
Recovery loop. After the first round of high-resolution processing, a tile whose fine-level evidence is positive near its border triggers zoom on the adjacent unselected tiles. This draws on a small reserved slice of the budget. The loop gives downstream evidence a way to correct an upstream gate, which neither closest paper has.
Diagnostic. We define zoom recall as the fraction of annotated lesions that have at least one zoomed tile. We also define grid-shift flip rate as the fraction of slides whose predicted label changes when the tile grid is translated by half a tile. Both are measured for every method, stratified by lesion size and by whether a mimic cluster is present.
Baselines
The naive reference point is plain HAG-MIL run from the released code with its default top-$k$ selection. On Camelyon16 its authors report an area under the ROC curve (AUC) of 0.946 and an F1 of 0.874. They report AUCs of 0.982 on TCGA-RCC and 0.921 on TCGA-NSCLC (HAG-MIL). We re-run it ourselves with identical feature extractors, splits, and seeds. All comparisons use our re-run numbers, with the published ones as a sanity check. This is the strongest existing method in the file and the one our changes target directly.
The second existing baseline is Hierarchical Regional MIL with high-attention inference. Its authors report a Camelyon16 AUC of 0.914, against 0.764 for their plain baseline, and 0.891 on TCGA-THCA (Towards Hierarchical Regional Transformer-based Multiple Instance Learning). It matters because its K-means filter is a different hard gate, so it tests whether our selection changes help beyond one architecture. We will apply our selection module to it as well.
The upper reference is exhaustive high-resolution processing with no pruning, using the same aggregator. It shows how much of the gap is due to selection at all. We found no reported result on Lacuna for an unpruned variant of either method, so we will measure it ourselves along with its compute cost.
The compute-matched alternatives are the following. First, simply raising $k$ until FLOPs equal our method's total cost. Second, a beam over the top candidate paths in the spirit of Chu et al.. Third, overlapping or multi-offset grids alone. Ablations remove one component at a time: boundary correction, cluster cap, and recovery loop.
Every variant receives the same hyperparameter search budget, measured in trials, on the same validation folds. Beating the baselines means two things at equal or lower FLOPs. First, a statistically significant gain over our re-run HAG-MIL on zoom recall for lesions in the smallest size tertile, together with a lower grid-shift flip rate. Second, no loss in overall slide AUC. Given HAG-MIL's 0.946 AUC on Camelyon16, we do not expect large movement in aggregate AUC there. The decisive evidence is on the small-lesion and mimic-present strata.
Why not just…?
Why not just use overlapping tiles or average over shifted grids? This is the most serious threat to the boundary component, and we do not argue around it. Multi-offset coarse scoring is literally part of our method. The experiment that matters is whether overlap alone fixes zoom recall. If grid shifting alone matches the full method on split lesions but not on mimic-present slides, then the boundary contribution reduces to "use overlap," and the paper's claim rests on the budget cap and recovery loop. We will report it that way.
Why not just raise $k$, or skip pruning entirely? With enough budget, selection misses vanish, which is why the unpruned model is our upper reference. The question is compute. The latency cost of a cascade is already large, as in the human-parsing cascade at 4.98 s versus 0.29 s per image (A Part-Based Deep Neural Network Cascade Model for Human Parsing). We will plot zoom recall against FLOPs for top-$k$ and for our selection. Our method is worth having only if its curve dominates in the low-budget region. If larger $k$ catches up at modest cost, the result is a calibration recommendation for $k$, not a new method.
Why not just train harder against mimics, for example with hard-negative mining? Better features could lower mimic scores and make the budget cap unnecessary. Disease-related pretraining did help a mimic problem in prostate MRI, raising T2-weighted AUC from 0.75 to 0.86 (Classification of Prostate Transitional Zone Cancer and Hyperplasia). We will train a hard-negative-mined HAG-MIL and apply our cap on top of it. If the cap adds nothing once features improve, we will say so. We expect some residual benefit, because coarse resolution limits how separable mimics can be at the stage where selection happens.
Why not just smooth the attention map before selecting? The refinement-paradox evidence from the human-parsing cascade shows that smoothing erases small, sharp features, which are exactly what micrometastases are. We include Gaussian smoothing of coarse scores as a control. We predict it raises zoom recall for large lesions and lowers it for small ones.
Proposed Experiments
The primary benchmark is Camelyon16 (397 slides). Its lesion annotations allow zoom recall and size stratification. Lymphoid tissue there supplies natural mimic clusters. TCGA-RCC and TCGA-NSCLC test whether the method holds up on subtyping, where "lesions" are diffuse and boundary effects should matter less. That is a useful negative control. TCGA-THCA tests the regional MIL backbone. The in-house IMGC set used by HAG-MIL is unavailable, so we do not use it.
The first decisive check is whether the problem exists at all. If plain HAG-MIL already examines nearly all annotated lesions, and its false negatives are mostly high-resolution misclassifications, the proposal's premise fails. We will report that as the main result. The second check is whether miss rate rises for lesions intersecting tile edges and corners compared with interior lesions, and whether the gap shrinks with lesion size as the amenability argument predicts. The third is whether each component's gain appears in its intended stratum: boundary correction on split lesions, the cluster cap on mimic-present slides, and the recovery loop on lesions adjacent to selected tiles. Each must hold at matched FLOPs, with grid-shift flip rate reported throughout.
References
- Parametric estimation for Gaussian fields indexed by graphs
- Diagnose Like a Pathologist: Transformer-Enabled Hierarchical Attention-Guided Multiple Instance Learning for Whole Slide Image Classification
- Towards Hierarchical Regional Transformer-based Multiple Instance Learning
- A Part-Based Deep Neural Network Cascade Model for Human Parsing
- Content-Based Retrieval from Medical Image Databases: A Synergy of Human Interaction, Machine Learning and Computer Vision
- Joint Object and Part Segmentation Using Deep Learned Potentials
- BeamAggR: Beam Aggregation Reasoning over Multi-source Knowledge for Multi-hop Question Answering
- Results on finite wireless sensor networks: Connectivity and coverage
- Interference Reduction in Virtual Cell Optimization
- When pilots should not be reused across interfering cells in massive MIMO
- Classification of Prostate Transitional Zone Cancer and Hyperplasia Using Deep Transfer Learning From Disease-Related Images