A core needle biopsy removes a sliver of tissue, often less than 1% of a tumor. Pathologists then treat that sliver as if it spoke for the whole lesion. Tumors vary from region to region, so the sliver is sometimes wrong. Receptor status, proliferation, immune checkpoints and resistant cell nests can all differ across a tumor. Current machine learning for pathology does not help with this. It either classifies the sample it is given or assumes the whole slide is available. It never asks whether the sample in hand is likely to represent the rest of the tumor.
This note proposes to learn exactly that signal: a representativeness score computed from the biopsy alone. The score comes from how a continuous-time quantum walk behaves on the core's tissue graph. A tissue graph is a network whose nodes are tissue regions and whose edges link neighbors. A continuous-time quantum walk (CTQW) is a wave-like analogue of a random walk. It spreads probability over the graph through a reversible, oscillating evolution rather than a diffusion that settles into equilibrium.
The bet is that a mixed, representative piece of tumor and a homogeneous or margin-dominated piece leave different interference fingerprints. This matters most on the small, cut-off graphs that biopsies produce. Those graphs are also exactly the size at which exact CTQW computation is cheap.
Core Claim
Given a whole-slide graph and virtual cores sampled from it, we can label each core by whether a classifier's prediction on that core disagrees with the slide-level (or patient-level) label. The claim is that time-averaged CTQW statistics of the core graph predict this disagreement better than three alternatives:
- the classifier's own confidence;
- distribution statistics of the core's features;
- the same pipeline with the quantum walk replaced by classical heat diffusion on the identical graph.
The final comparison is the one that matters. If heat diffusion does as well, the quantum ingredient is unnecessary. We will report that outcome as the finding, and the reusable contribution becomes the discordance benchmark itself.
The walk is defined by a Hamiltonian, the matrix that plays the role of the graph in the walk's dynamics. We build one that carries tissue attributes, not just connectivity:
$$H = W + \mathrm{diag}(\lambda\, p), \qquad M = \lim_{T\to\infty}\frac{1}{T}\int_0^T \left|e^{-iHt}\right|^{\circ 2} dt .$$
Here $W_{uv} = \exp(-\lVert x_u - x_v\rVert^2/\sigma^2)$ weights the edge between adjacent tissue nodes $u,v$ by the similarity of their morphology embeddings $x$. The vector $p$ holds each node's predicted tumor probability and acts as an on-site potential scaled by $\lambda$. The symbol $\circ 2$ denotes elementwise squared modulus, and $M$ is the average mixing matrix: the long-run fraction of time a walker started at one node spends at another.
The on-site potential is our answer to a problem the quantum-walk graph convolution work leaves open (Learning Graph Convolution Networks Based on Quantum Vertex Information Propagation): how to let node attributes shape the walk without discarding its spectral structure. From $M$ we read off two summaries that feed a small classifier:
- the distribution of per-node quantum Shannon entropies, as in QBMK;
- an "escape profile", the time-averaged mass that leaves tumor-labeled nodes.
Positioning Against Prior Work
The closest bridge is Hybrid Quantum-Classical GNNs for Tumor Classification. That work attaches a variational quantum classifier to embeddings from a classical tissue-graph network (HACT-NET) on BRACS. Its quantum part is a readout head only, it uses no walk dynamics, and it never considers sampling. We use no quantum hardware and make no claim about quantum speed or classification accuracy. The walk is a classically simulated structural probe, aimed at a different target: sample representativeness.
The methods we reuse come from graph learning on standard benchmarks:
- QSGCNN replaces the adjacency matrix with the average mixing matrix and argues that interference suppresses back-and-forth "tottering" (Learning Graph Convolution Networks…).
- QBMK builds entropy signatures from the average mixing matrix that mix local and global context (QBMK).
- Quantum positional encodings separate strongly regular graphs that classical encodings cannot (Quantum Positional Encodings for GNNs).
All three were tested on molecules and social graphs, not tissue.
That last result also works as a caution. Strongly-regular-graph expressivity has little obvious relevance to noisy tissue graphs with rich node features. Its reported ZINC gain is small (MAE 0.058 vs 0.060). We therefore do not rest the proposal on expressivity arguments.
The most relevant motivation is On Quantum Random Walks in Biomolecular Networks. On gene networks, its CTQW escaped bottlenecks that trapped random walk with restart, reaching AP@20 ≈ 0.30 against 0.14. The same paper also constrains us: the gap narrowed at larger retrieval depths. Any quantum advantage may therefore be confined to intermediate ranges, which is why our design sweeps the Hamiltonian scale.
On the pathology side, the evidence that single samples mislead is clinical:
- regional receptor variation in colorectal cancer (IGF-IR/EGFR/HER-2 coexpression);
- sparse resistant stem-cell clusters missed by needle biopsy (cancer stem cell factors in rectal cancer);
- Ki-67 sampling variability (Ki-67 radiomics).
Two lines of computational work have tried to route around this. Whole-lesion imaging such as PD-L1 from PET/CT avoids the core entirely. Heterogeneity-aware aggregation such as quantile MIL pools the patches within whatever sample it receives. Neither estimates whether the physical sample is trustworthy. We target that estimate because molecular assays still run on cores.
Research Program
The work proceeds in three stages.
Stage 1: build a virtual-biopsy discordance benchmark. From BRACS regions of interest and TCGA-COAD whole slides, we cut simulated needle cores. These are elongated strips of realistic width and length, placed uniformly or biased toward margins. We build superpixel tissue graphs for each core, following the HACT-NET style used in the hybrid paper. This keeps graphs to a few hundred nodes, so exact eigendecomposition, which scales as $O(n^3)$ in node count, takes milliseconds per core.
The Carolina Breast Cancer Study tissue microarray set supplies real cores: 5,970 cores from 1,713 patients, with several cores per patient. There, discordance is measured directly as disagreement between core-level and patient-level ER status or subtype. No simulation is needed.
Stage 2: fix a core classifier and compute discordance labels. We use both quantile MIL and a tissue-graph network as core classifiers. Each is trained once, and each yields discordance labels. We then fit the representativeness predictor on held-out slides and patients.
Stage 3: analyze what the walk detects. We relate the entropy and escape features to measured structure: the tumor–stroma interface fraction, the number of distinct morphology clusters, and cut-boundary effects.
Baselines
Naive reference. Trust every core, with coverage fixed at 100%. As a second naive signal, rank cores by tissue area or tumor-node count. Size is the first thing a pathologist checks, so a proposal that cannot beat core size has not earned its complexity.
Classifier self-uncertainty. This is the strongest practical competitor. We use predictive entropy from a deep ensemble of five core classifiers of each type, drawn from the same backbones and training data. Disagreement usually shows up as low confidence, and this signal costs nothing extra to obtain.
Feature-distribution statistics. A classifier receives the core's morphology-feature histogram, quantile summaries, and interface fraction. This tests whether topology adds anything beyond "what is in the bag."
Classical spectral probes. On the identical attributed graph, we compute:
- heat-kernel signatures from $e^{-Ht}$;
- random-walk-with-restart profiles;
- Laplacian eigenvector encodings fed to a graph network trained directly to predict discordance.
The heat-kernel variant is the key ablation. It keeps $H$, the readout and the tuning grid, and changes only unitary evolution into diffusion. It therefore isolates interference.
Fairness. All representativeness predictors receive identical splits, grouped by patient or slide. Each gets the same hyperparameter budget of 50 trials, selected on validation discordance AUROC, and runs on one GPU.
Anchors for the core classifiers. These keep the classifiers themselves at a credible level:
| Classifier / setting | Reported result | Source |
|---|---|---|
| Quantile MIL, CBCS ER status | 85.6% accuracy (68.6% with single-patch training) | quantile MIL |
| HACT-NET, BRACS | weighted F1 0.70 | hybrid paper |
| Deformable Attention Graph, TCGA-COAD full slides | 87.31% accuracy | Deformable Attention Graph |
The Deformable Attention Graph serves as a full-slide ceiling, not a core-level method.
No reported result was found on Lacuna for predicting biopsy-to-slide discordance. Beating a baseline therefore means higher discordance AUROC over five seeds with non-overlapping confidence intervals. It also means better selective accuracy at 80% coverage, the accuracy on cores the method chooses to trust, than the ensemble-entropy baseline.
Why not just…?
Why not just use the classifier's confidence? It may be enough, and it is our main threat. Confidence, however, reflects what the model sees. A homogeneous core of confidently ER-positive tumor from a regionally mixed tumor produces a confident, wrong answer. The regional receptor variation documented in IGF-IR/EGFR/HER-2 coexpression predicts exactly this failure. We will report discordance AUROC separately for high-confidence cores. If the walk features add nothing there, the idea fails.
Why not just summarize the patch distribution, as quantile MIL does? Quantile MIL shows that distribution tails matter within a sample. The same paper's drop from 85.6% to 68.6% accuracy without context shows that arrangement matters too. The feature-statistics baseline tests directly whether arrangement carries representativeness information beyond composition.
Why not just use heat diffusion or random walk with restart? Diffusion is cheaper and standard. The only evidence for interference helping on biological graphs is the retrieval gain in On Quantum Random Walks in Biomolecular Networks, and that gain faded at longer range. The matched heat-kernel ablation settles this question. We commit to reporting it whichever way it falls.
Why not just image the whole lesion? Whole-lesion approaches work for some targets, for example PD-L1 from PET/CT (AUC ≥ 0.82) and CT radiomics survival signatures (C-index 0.65–0.69). They do not produce the tissue that genomic and receptor assays require, and they vary with scanner and acquisition settings. Our score is complementary: it flags when a physical core should be repeated.
Decisive Checks
The first test is on real CBCS cores: does walk-based representativeness predict core–patient ER discordance better than ensemble entropy? This matters most for the subset of cores the classifier is confident about.
The second test compares the unitary walk against heat diffusion at matched $H$. We sweep the time scale and the potential strength $\lambda$ to locate any range where interference helps, echoing the range dependence seen on gene networks.
The third test concerns transfer. Scores trained on margin-biased virtual cores in BRACS should rank discordance on TCGA-COAD without retraining, which checks that the signal is not tied to one tissue type.
The fourth test is selective prediction. When cores with low representativeness scores are withheld, patient-level accuracy should recover toward the multi-core quantile-MIL figure of 85.6%.
A clear positive result on the first two tests justifies the quantum walk. A negative result on the second test still leaves a benchmark and a simpler classical representativeness score.
References
- Learning Graph Convolutional Networks Based on Quantum Vertex Information Propagation
- QBMK: Quantum-based Matching Kernels for Un-attributed Graphs
- Hybrid Quantum-Classical Graph Neural Networks for Tumor Classification in Digital Pathology
- Quantum Positional Encodings for Graph Neural Networks
- On Quantum Random Walks in Biomolecular Networks
- Coexpression of the IGF-IR, EGFR and HER-2 is common in colorectal cancer patients
- Cancer stem cell-related factors are associated with the efficacy of pre-operative chemoradiotherapy for locally advanced rectal cancer
- Prediction of Ki-67 expression in invasive breast cancer with dual-modality radiomics
- Non-invasive measurement of PD-L1 status and prediction of immunotherapy response using deep learning of PET/CT images
- Multiple Instance Learning for Heterogeneous Images: Training a CNN for Histopathology
- Deformable Attention Graph Representation Learning for Histopathology Whole Slide Image Analysis
- Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach