Source-linked AI summary
Seeing Beyond the Lesion: Disease Recognition from Reactive CNS Tissue
Jan Schnorrenberg, Jan Ernsting, Enrico Küllenberg, Tim Hahn, Benjamin Risse, Christian Thomas
TL;DR
Sampling error can leave intracranial biopsies with reactive, non-lesional tissue that does not reveal the underlying disease. The study benchmarks four frozen pathology foundation-model encoders in an attention-based MIL framework, audits slide-size confounding, and finds that finer disease distinctions remain above chance while encoder performance is statistically indistinguishable. The results support provenance-only shortcut checks and indicate that weak disease signal can be recovered from reactive tissue, although external validity is limited by the single-site cohort.
Problem
Reactive-only brain biopsies can leave the underlying disease undiagnosed because conventional histopathology reveals tissue response without identifying its cause.
Method
The study compares four frozen pathology foundation-model patch encoders within a shared attention-based MIL framework for classifying reactive, non-lesional tissue.
Results
After slide-size confounding was identified for coarse classification, disease labels remained above chance on three finer tasks and all four encoders performed without statistically significant pairwise differences.
Takeaways & Limitations
Provenance-only baselines are a necessary control for computational-pathology benchmarks, and weakly supervised models can recover disease signal from tissue regarded as non-diagnostic once that confound is removed.
Takeaways & Limitations
The single-site, modest cohort—especially the small metastatic-origin classes—does not support claims about cross-site generalization.
Abstract
from arXiv · showhide
Sampling error yields exclusively reactive, non-lesional brain parenchyma in a significant proportion of intracranial biopsies, leaving the underlying disease undiagnosed. We benchmark four pathology foundation models (UNI2-h, Virchow2, Prov-GigaPath, H-optimus-0) as frozen patch encoders within a shared attention-based multiple-instance learning framework using 245 whole-slide images from 186 patients with confirmed downstream diagnoses. We first show that coarse disease-category prediction can be reproduced largely from slide size alone. After restricting classification to three finer diagnostic distinctions within common tissue categories, this confound no longer explains performance, yet disease labels remain predictable above chance under permutation testing (p $\le 10^{-4}$ throughout). Surprisingly, performance is statistically indistinguishable across all foundation-model encoders, suggesting that recovering these weak morphological signatures is not limited by current patch representations. Signed instance-contribution maps and expert review further test whether predictive evidence localizes to reactive parenchyma rather than sampling-induced bias like blood introduced during tissue sampling. These results position acquisition-shortcut auditing via a provenance-only baseline as a necessary control in computational-pathology benchmarks, and show, once that confound is removed, that weakly supervised models still recover disease signal from tissue conventionally regarded as non-diagnostic.
1. Introduction
Reactive-only brain biopsies can conceal the disease process because conventional pathology sees gliosis rather than the underlying lesion. This study tests whether weakly supervised models can infer disease from reactive tissue and whether foundation-model representations matter after auditing acquisition shortcuts.
- Clinical gap: Up to 14% of intracranial biopsies yield only reactive, non-lesional parenchyma that conventional histopathology cannot link to the underlying disease.Reactive gliosis includes activated astrocytes and microglial proliferation, with context-dependent cellular states.
- Clinical gap: Prior weakly supervised CNS tumor models target overt lesions, where neoplastic morphology provides the dominant class-defining signal.Reactive-only specimens instead contain subtle, spatially diffuse disease-related information embedded within tissue responses.
- Study aim: Foundation models differ in pretraining data, objectives, architectures, and receptive fields, but their ability to capture subtle reactive CNS morphology remains unknown.Performance on tumor-centered benchmarks cannot be assumed to generalize to non-lesional brain tissue.
- Study aim: The study benchmarks multiple pathology foundation models as shared-framework patch encoders for disease classification from reactive, non-lesional tissue.Tasks include three-class category prediction, primary CNS tumor distinctions, and metastatic tissue-of-origin inference.
- Study aim: A provenance-only slide-size baseline exposes an acquisition shortcut on the coarse task, while permutation testing supports recoverable disease signal on three finer tasks.After correction, the four encoders are statistically indistinguishable, indicating representation choice is not the current bottleneck.
2. Related work
Computational pathology uses weakly supervised MIL and foundation-model patch encoders to extract slide-level information from gigapixel WSIs. Existing CNS applications focus on morphologically evident tumors, leaving reactive, non-lesional tissue as a distinct diagnostic setting.
- Weakly supervised computational pathology: Multiple instance learning aggregates patch features into slide-level representations when WSI annotations are available only at slide or case level.Attention-based variants can expose regions that drive a prediction.
- Pathology foundation models: Pathology foundation models provide transferable patch representations learned through self-supervised pretraining on large, diverse histopathology datasets.The benchmarked models differ in pretraining data, objectives, architectures, and receptive fields.
- Deep learning for CNS tumor histology: DEPLOY and Hetairos demonstrate strong weakly supervised CNS tumor classification from H&E WSIs with morphologically evident lesions.Hetairos combines Prov-GigaPath embeddings with transformer-based MIL to resolve many methylation-defined subtypes.
- Deep learning for CNS tumor histology: Reactive, non-lesional tissue differs from tumor-centered benchmarks because it lacks overt neoplastic morphology as a dominant class-defining signal.The study therefore evaluates whether disease-specific information can be recovered from subtle tissue responses.
- Dataset structure: The curated dataset contains 186 patients and 245 slides organized across Category, Subtype, Glioma, and Origin classification levels.Labels span tumor, inflammation, control, glioma-versus-metastasis, three glioma entities, and four metastatic origins.
3. Methods
The study builds patient-stratified weakly supervised WSI classifiers from reactive tissue, comparing four frozen foundation-model encoders within one gated-attention MIL pipeline. It audits provenance and interpretability using slide-size baselines, signed patch contributions, and expert review.
- Data and preprocessing: The dataset comprises 245 H&E WSIs from an institutional neuropathology archive, with research use approved by the local ethics committee.The cohort contains anonymized biopsy specimens.
- Data and preprocessing: Reactive-only ground truth was created by removing primary lesions from slides containing adjacent reactive tissue or using manually edited reactive regions linked to definitive diagnoses.These strategies produced systematically different tissue amounts, making slide size correlated with sampling provenance.
- Data and preprocessing: WSIs were converted to BigTIFF and down-sampled from native 40× to 20× resolution to match foundation-model training resolution.The pixel resolution changed from 0.233 µm/pixel to 0.5 µm/pixel.
- Encoders and MIL: UNI2-h, Virchow2, Prov-GigaPath, and H-optimus-0 were used as frozen patch encoders without end-to-end fine-tuning.Three encoders produce 1536-dimensional embeddings, while Virchow2 produces 2560-dimensional embeddings.
- Encoders and MIL: Each slide is treated as a bag of tissue patches whose embeddings are projected to 256 dimensions, pooled by gated attention, and classified with a shared linear head.The plain single-branch aggregator avoids spatial or relational inductive bias, so comparisons primarily reflect encoder feature spaces.
- Interpretability: Signed patch contributions decompose each class logit exactly into per-patch terms that distinguish evidence for versus against a class.Heatmaps combine attention salience with class-aligned projection and localize where reactive tissue drives predictions.
- Evaluation: Patient-stratified Monte Carlo cross-validation assigns all slides from each patient to the same split to prevent train-test leakage.Training uses class-balanced cross-entropy, while significance is assessed with conditional label permutations and corrected pairwise encoder tests.
- Interpretability: A neuropathologist reviewed the five highest-contribution tiles per slide and recorded whether each was predominantly sampling-introduced blood.The category task was used because it contained the most slides and therefore offered the greatest per-tile comparison power.
4. Results
Slide size substantially confounded coarse category prediction, whereas finer diagnostic tasks remained above chance and attention maps localized predictive evidence to reactive tissue. Across tasks, encoders showed no corrected-significant performance differences, while some differed markedly in attention to sampling-introduced blood.
- Slide-size confounding: 0.782 balanced accuracy (AUC 0.877) from slide size alone nearly matched the best encoder on coarse category prediction, exposing a major acquisition confound.The size-only baseline was near chance for metastatic origin and glioma typing, and only marginally above chance for tumor subtyping.
- Classification performance: BACC 0.798–0.792 and AUC 0.921–0.912 were achieved on broad tissue-category classification, while glioma-versus-metastasis subtyping reached BACC 0.807–0.819 and AUC 0.906–0.921.These coarse tasks were evaluated against chance levels of 0.33 and 0.50, respectively.
- Classification performance: BACC 0.437–0.581 for glioma typing and 0.371–0.416 for metastatic origin remained above chance, despite performance declining with finer diagnostic granularity.Metastatic origin was the hardest task and showed among the largest fold-to-fold variance.
- Attention localization: Attention maps localized diagnostic evidence to reactive, non-lesional tissue, with encoders broadly agreeing on predictions in representative correctly classified cases.For a consensus-correct lung metastasis, all four encoders localized to similar regions; in a glioblastoma case, they attended to different image areas.
- Sampling-artifact analysis: 27.0% of Virchow2’s top-five tiles contained blood, compared with 15.7% for H-optimus-0, 4.4% for Prov-GigaPath, and 4.0% for UNI2-h.The nearly sevenfold spread was significant by the Friedman test (χ2(3) = 151.6, p < 0.001), although blood attention did not differ between correctly and incorrectly classified cases.
- Encoder comparison: p ≤ 1 × 10^-4 occurred for every encoder-task-metric combination, but no pairwise encoder difference survived Holm correction across 48 comparisons.The smallest corrected p-value was pHolm = 0.25; the authors caution that 11–20 folds may be underpowered to resolve differences of the observed magnitude.
5. Discussion
The discussion separates a slide-size acquisition shortcut affecting coarse classification from above-chance disease signal recovered on finer tasks. It also shows that encoder choice does not determine slide-level performance, while attention behavior and external validity remain important qualifications.
- The size confound: A slide-size baseline nearly matches the best encoder on coarse tissue-category classification, so that task cannot establish learned morphological response.The baseline is within roughly two balanced-accuracy points of the best encoder.
- The size confound: The size-only baseline is at or near chance on glioma typing and metastatic-origin classification, whereas encoders classify both tasks above chance under permutation testing.This localizes the acquisition confound to the coarse task rather than the finer distinctions.
- The size confound: A provenance-only baseline is a cheap general control, but it cannot distinguish tile count from staining, scanning, or other provenance-linked covariates.A provenance-stratified evaluation is identified as the more complete control.
- Equivalent accuracy, divergent attention: Top-ranked attention tiles capturing sampling-introduced blood range from 4.0% for uni2-h to 27.0% for virchow2, despite statistically indistinguishable slide-level performance.Artifact rates did not differ between correct and incorrect cases, but the spread cautions against interpreting a single encoder’s attention map as definitive evidence.
- External validity: The single-institution, modest cohort supports feasibility rather than validated cross-site diagnostic performance.The finest-task confidence intervals are wide, and persistence across scanners, protocols, and centers remains to be established.
- Representation choice: No pairwise encoder difference is significant after correction across four tasks and both metrics, suggesting representation choice is not the bottleneck for this reactive-tissue signal.The authors instead point to signal subtlety, cohort scale, and shared 256-pixel tokenization as possible constraints.