Corroboration, Not Loss: Telling Rare Tumors From Missed Biopsies in Meta-Learning

When a model trained across many small clinical prediction problems keeps failing on one of them, there are three very different explanations. The labels may be wrong. The problem may be a rare but real biological pattern the model has not learned yet. Or it may be a single patient whose outcome no one could predict. Methods that decide how much each training problem counts all read the size of the error as the signal. Some turn toward the hardest problems, some discard them, and some learn a weighting from loss plus internal consistency. In oncology the first explanation has a particular shape. A needle biopsy samples a tiny track through a spatially varied tumor. When it misses a receptor-positive or checkpoint-positive region, it produces a consistent false negative for that patient. This is documented for PD-L1 by Mu et al. and for growth-factor receptors by Cunningham et al.. That error looks exactly like a hard, coherent task.

We propose to separate these cases with the test clinicians already apply to case reports: replication. A high-error training problem is allowed to pull the model toward it only if other high-error problems, drawn from independent sites or strata, push the model's parameters in the same direction. A rare real phenotype recurs and so corroborates itself. A biopsy miss or a one-off anecdote does not. Uncorroborated problems are neither amplified nor thrown away. They are held at baseline weight and flagged. This turns the "proof of existence versus proof of pattern" limit of single case reports, stated explicitly by Hideshima et al. and Wang et al., into an operational rule for weighting training tasks.

Core Claim

Meta-learning trains a model over many small "tasks" so that it adapts quickly to a new one. In the clinical setting here, a task is a group of patients from one stratum, such as a cancer type crossed with a collection site, split into a small training set (support) and a held-out set (query). Existing non-uniform weighting cannot distinguish a task whose labels are consistently wrong from a task that is genuinely hard.

Consider the Adaptive Task Scheduler (ATS). It detects noise through disagreement between support-set and query-set gradients, meaning the directions in which each set would move the model's parameters. When a biopsy miss flips labels coherently across a stratum, support and query are corrupted together and their gradients agree. The noise is therefore invisible from inside the task. The same flaw affects the other two families. TR-MAML optimizes the worst-case task and will chase the corrupted one. HeTRoM truncates the top-loss tasks by rank and will discard rare true phenotypes along with the noise. The research file identifies this tension (hardness versus noise) as unexplored for clinical task pools.

The fix we propose uses evidence from outside the task. Let $H$ be the set of tasks in a meta-batch whose query loss lies above quantile $q$. Let $g_i$ be the last-layer meta-gradient of task $i$, and let $s(i)$ be its collection site. We define the corroboration of task $i$ as

$$ c_i = \frac{1}{|H|-1}\sum_{j \in H,\ s(j)\neq s(i)} \mathbf{1}\left[\cos(g_i, g_j) > \tau\right], $$

the fraction of other high-loss tasks from different sites whose gradients point the same way. The threshold $\tau$ is fixed on a validation pool. Task weights are then

$$ w_i \propto \begin{cases} \exp(\beta \ell_i) & \text{if } i \notin H \text{ or } c_i \ge \kappa, \\ w_0 & \text{otherwise,} \end{cases} $$

where $\ell_i$ is the query loss, $\beta$ sets how strongly hard tasks are up-weighted, $w_0$ is the uniform weight, and $\kappa$ is the minimum corroboration needed to count as a pattern.

The cross-site requirement matters. Biopsy protocols are site-specific, so misses can be correlated within a site. Only corroboration across independent sites is taken as evidence that a failure reflects a real pattern. Tasks held at $w_0$ go into an "anecdote register" and are revisited as the meta-batch composition changes.

Relation to Prior Work

The three task-weighting papers supply the method components and define what must be beaten. ATS learns sampling probabilities with a separate scheduler network from loss and support–query gradient similarity. We keep a gradient-similarity signal but compute it between tasks rather than within one, and we need no second network. HeTRoM keeps tasks ranked between $m+1$ and $k$ by loss. We replace its fixed rank cut with a corroboration test, so a rare phenotype ranked in the top $m$ can survive. TR-MAML is the pure "upweight hardest" position. Our rule reduces to a softened version of it when every hard task is corroborated. Task affinity via bipartite matching measures relatedness between tasks to choose auxiliary data for a target. We use relatedness among failing tasks for a different question: is a failure real?

The clinical papers motivate the noise model and supply clean references. Zhang et al. show that radiologist-scored LIDC-IDRI labels are biased toward calling nodules malignant. They fix this with a one-time relabeling against 180 pathology-proven SCH-LND nodules. That approach requires a matched clean anchor set, which our method does not, and we use their setup as a real-noise evaluation. Deep attention multiple-instance learning and quantile aggregation over tissue cores handle spatial heterogeneity within a patient by pooling over regions. These pooling methods are complementary. They improve the per-patient representation but say nothing about which patients' labels to trust.

The case reports serve as the epistemic motivation. Our method does not ingest case-report text. It gives an $N=1$ failure a defined status: bounded weight and an explicit flag, instead of silent absorption or deletion.

Research Program

The work has three parts.

First, we build a benchmark with known ground truth that separates the three kinds of failure. We start from the TCGA Meta-Dataset (174 clinical tasks, 33 cancer types, RNA-Seq features, 50-train/100-test regime). We then define episodes by cancer type crossed with tissue source site, and plant three failure types.

The first type is structured biopsy misses: in selected strata, a fraction of positives is flipped to negative. The flips are concentrated in patients with high expression heterogeneity, and support and query are flipped together. The second type is rare real subgroups: a small feature–outcome rule is planted consistently in a handful of strata across different sites. The third type is singletons: a unique rule is planted in one stratum only.

Second, we implement corroboration weighting on top of first-order MAML using last-layer gradients. We measure its cost. HeTRoM reports 1.36 s per iteration against MAML's 1.22 s, and pairwise cosines over $|H|$ tasks should add comparable overhead.

Third, we move to real label noise. On the Carolina Breast Cancer Study tissue microarrays (5,970 cores, 1,713 patients, as used by the quantile MIL paper), each patient's multiple cores let us create "single-biopsy" labels from one core and compare them against consensus. Access to this dataset must be confirmed. On LIDC-IDRI, radiologist scores serve as the noisy label and SCH-LND as the pathology-proven test set, following Zhang et al..

An optional second ingredient is an imaging-derived credibility prior. It multiplies $w_0$ for uncorroborated tasks by an estimated miss probability computed from whole-lesion habitat heterogeneity, using K-means habitats as in Huang et al.. It is ablated separately and is not part of the core claim.

Baselines

The naive reference is uniform-sampling MAML together with pooled empirical-risk training of an MLP. On the TCGA Meta-Dataset, the benchmark paper reports average accuracy of 71.19% (±14.62%) for an MLP, 68.78% for logistic regression and 66.18% for a majority classifier, all without planted noise. We rerun these on our corrupted variant with identical splits.

The strongest competitors are ATS and HeTRoM. On miniImageNet, ATS reports a 13% accuracy gain over uniform sampling at 60% task noise. HeTRoM reports 70.45% on clean 5-way 5-shot against MAML's 63.11%, and about 68% at 60% noise. There is no reported result found on Lacuna for either method on the TCGA Meta-Dataset, LIDC-IDRI or tissue-microarray data. We therefore run both ourselves with the same backbone, the same meta-batch size and the same number of hyperparameter trials per method. HeTRoM's thresholds are set by its own log-normal warm-up (sg-HeTRoM), and ATS's scheduler receives the same wall-clock budget. TR-MAML is included as the worst-case-focused extreme.

For the real-noise LIDC setting, the static relabeling of Zhang et al. is the domain baseline, with a reported sensitivity of 0.67 and specificity of 0.66 against 0.97/0.23 on raw labels. It has the unfair advantage of using SCH-LND anchors, which our method does not.

Two ablations isolate the key ingredient. The first replaces cross-task corroboration with ATS-style within-task support–query similarity while keeping everything else fixed. The second allows corroboration from the same site, which tests whether site independence matters.

Beating the field means meeting three conditions on the planted benchmark. Clean-task accuracy must be no worse than HeTRoM. Accuracy on rare-subgroup query sets must be significantly higher than HeTRoM, which should discard them. The weight placed on biopsy-miss strata must be significantly lower than under TR-MAML and ATS. On LIDC, beating the domain baseline means matching or exceeding Zhang et al.'s specificity at comparable sensitivity without using anchors.

Why not just…?

Why not just use ATS, which already detects noisy tasks by gradient disagreement? ATS's noise signal compares support and query gradients inside one task. A biopsy miss corrupts both halves of a stratum coherently, so the two gradients agree and the task looks clean and hard. That is the precise failure mode this proposal targets. The within-task ablation tests it directly. If ATS's disagreement signal already flags our structured-miss strata, the core claim is wrong, and we will report that.

Why not just truncate the highest-loss tasks as HeTRoM does? Truncation removes coherent noise but also removes every rare real phenotype that happens to be hard. Clinically, these are exactly the exceptional responders that case reports keep surfacing. On the planted benchmark this shows up as low accuracy on rare-subgroup queries. This is a real threat on one axis: if rare subgroups are so small that they never enter the top-$m$, HeTRoM loses nothing. We therefore vary subgroup size and report where the two methods cross over.

Why not just fix labels offline with a clean anchor set, or pool multiple cores per patient? Anchor relabeling works when a matched, pathology-proven set exists, as SCH-LND does for solid nodules. The TCGA-style 174-task setting has no such set for most phenotypes. Pooling more cores is the true fix when more tissue exists. On the tissue-microarray data we report the multi-core consensus model as an upper bound, which shows how much of that gap corroboration weighting closes when only one core is available.

Why not just use the imaging heterogeneity prior alone? This alternative is plausible and could explain most of the gain on LIDC, since heterogeneity is what causes biopsy misses. The ablation that uses the prior without corroboration settles it. If the prior alone matches the full method, the contribution narrows to a credibility prior and we will say so. The TCGA benchmark, which has no imaging, still tests corroboration by itself.

Decisive Checks

The first decisive experiment is a synthetic replication on miniImageNet with two noise types at matched rates: within-task random label noise and task-level consistent label permutation. We predict that ATS handles the first but not the second, and that corroboration weighting handles both. If ATS handles both, the premise fails before any clinical data is touched.

The second is the planted TCGA benchmark. The primary metrics are clean-task accuracy, rare-subgroup query accuracy, total weight mass placed on miss strata, and precision of the anecdote register in identifying planted singletons. All are reported over at least five seeds with paired tests.

The third is external validity. We report LIDC-IDRI-trained, SCH-LND-tested sensitivity and specificity. On tissue microarrays, we report single-core-trained, consensus-tested accuracy for ER status against the multi-core upper bound.

A failure mode we will report if it appears: correlated misses across sites, for example from a shared biopsy protocol, could fake corroboration. We test this by planting cross-site correlated misses and measuring how much weight leaks to them.