Summarize Paper logo SummarizePaper.com

AI summary · arXiv: 2610.41328v1 · eess.IV

With foundation-model features, aggregator complexity does not improve cross-hospital discrimination in gastric whole-slide image classification

Nheng Vanchhay, Ali Hussain, Khan Ziaullah, Moon-Il Joo, Md Ariful Islam Mozumder, Rashedul Islam Sumon, Won-Ae Lee, Min-Sun Cho,
and 9 more authorsHyeong-Chan Shin, Yujun Park, Baek Hwan Cho, Nam-Hoon Cho, Heejung An, Heung-Kook Choi, Haeyoun Kang, Hyunki Kim, Hee Cheol Kim
Original paper Submitted to arXiv on 30 Oct. 2026 License: CC BY 4.0 Also in: fr

How to read this summary? Illustrated guide

The essentials

On five-hospital gastric slide classification with frozen foundation-model features, hierarchical aggregation did not outperform simpler pooling; label-free onboarding and auditable evidence offered more value than architectural complexity.

Question

Does increasing multiple-instance learning aggregator complexity improve gastric whole-slide classification across unseen hospitals when strong foundation-model features are available? The study also examines label-free hospital alignment, calibration and attention faithfulness.

Method

The authors retrospectively assembled 5,036 gastric H&E slides, including 1,674 malignant and 3,362 normal slides, from five Korean hospitals scanned at 20× and 40×, and evaluated classifiers using leave-one-hospital-out validation. They compared hierarchical CHIME-MIL with attention, transformer and global max pooling aggregators using frozen UNI or GenBio-PathFM features, including recipe-matched controls for site mean-centring and 256- and 392-pixel fusion. Analyses included 10,000-replicate paired slide-level bootstraps, unlabelled onboarding experiments, calibration assessment and attention-patch removal tests.

Findings

  • Result Under matched alignment and two-scale fusion, global max pooling achieved AUC 0.8552 versus CHIME-MIL’s three-seed mean of 0.8374, while matched hierarchical aggregation also failed to exceed ABMIL.
    See the passage in the paper
    We therefore claim no discrimination advantage for the hierarchical architecture: with strong patch features, hierarchical complexity does not improve AUC over simpler aggregators on this binary task.
  • Result In the deterministic component ablation, multi-scale fusion was the only step with a statistically supported CHIME-MIL AUC gain; mean-centring and hospital-stratified validation were individually nonsignificant.
    See the passage in the paper
    Multi-scale fusion is the only step with a statistically supported gain on CHIME-MIL: +0.0092 (95% CI +0.0038 to +0.0149, P=0.001), reaching 0.8434, with positive per-fold gains everywhere (Supplementary Table S2).
  • Result Label-free alignment supported onboarding from ten unlabelled slides without retraining, and excluding held-out-hospital statistics reduced fused-model AUC by only 0.0019.
    See the passage in the paper
    Replacing the held-out-site mean with the average of the four training-site means, which uses no held-out information, costs only ΔAUC=−0.0019 on the fused model (0.8374 to 0.8355; Supplementary Table S4).
  • Result Despite low top-label calibration error, the model systematically underestimated malignancy probabilities.
    See the passage in the paper
    The headline model has a low top-label expected calibration error (ECE 0.053 ± 0.014; Brier score 0.163 ± 0.020), but its malignancy probabilities are systematically under-confident (mean positive-class reliability gap 0.108 ± 0.028).
  • Authors’ interpretation Attention-patch ablations supported the authors’ interpretation that highlighted patches carried predictive evidence, without establishing diagnostic utility for pathologists.
    See the passage in the paper
    The attended patches therefore carry the predictive evidence rather than merely looking plausible, which is the property a pathologist-facing audit trail requires (population statistics in Supplementary Note 6).

Limitations

  • The gastric benchmark does not establish generalisation to multiclass subtyping or populations outside the represented Korean centres.
  • The retrospective evaluation did not test whether attention evidence improves pathologists’ decisions in a reader study.
    See the passage in the paper
    All experiments are retrospective, and the decision-support framing rests on the premise, not yet tested in a reader study, that the surfaced attention is actionable for pathologists; the interpretability claim should be read as auditability of the surfaced evidence, not as validated diagnostic utility.
  • External CAMELYON16 validation tested architecture transfer rather than cross-hospital alignment.
    See the passage in the paper
    The external CAMELYON16 result supports architecture transfer only, because its single-source data disables the per-site alignment by construction.

Why it matters

The authors argue that strong frozen pathology features shift the emphasis away from increasingly complex aggregators toward feature-side improvements and rigorous recipe-matched comparisons. For deployment, they emphasise label-free onboarding, calibration checks and auditable slide evidence rather than claiming an accuracy advantage for hierarchical aggregation.

Written by AI from the paper. Passages are checked word for word against the text we analysed; the rest may contain errors.

AI-generated keywords: Multiple-instance learning; Gastric histopathology; Pathology foundation models; Cross-hospital generalisation; Label-free adaptation

Read further

  • Comprehensive Summary
  • Layman's Summary
  • Blog article
  • Original abstract
This paper examines whether increasingly complex multiple-instance learning (MIL) aggregators improve cross-hospital classification of gastric whole-slide images when strong, frozen pathology foundation-model features are already available. It also addresses practical deployment requirements: adaptation to an unfamiliar hospital without labelled examples, probability calibration, and identification of image regions supporting predictions. Its central finding challenges architectural complexity: under comparable training conditions, a hierarchical graph model does not outperform simpler attention pooling, while global max pooling performs better. The retrospective dataset comprises 5,036 haematoxylin-and-eosin-stained gastric slides from five Korean tertiary hospitals, including 1,674 malignant and 3,362 normal slides scanned at 20× or 40× magnification. Evaluation used leave-one-hospital-out validation, with models trained on four hospitals and tested on the fifth. Images were standardised to a physical resolution corresponding to 20× magnification, divided into patches, and encoded using frozen GenBio-PathFM features. UNI features provided an additional comparison. The proposed CHIME-MIL model combines patch-level gated attention, attention over a 4 × 4 spatial grid, and graph-based reasoning between regions. Predictions from separately trained models using 256-pixel and 392-pixel patches were averaged. A label-free alignment procedure subtracts each hospital’s mean feature vector. The authors also evaluated a region-ablation training loss and compared CHIME-MIL with conventional attention-based and transformer-based aggregators, as well as global max pooling. CHIME-MIL achieved a mean hospital-level area under the receiver operating characteristic curve (AUC) of 0.8374, with a standard deviation of 0.0041 across three training seeds. Although it outperformed six conventional GenBio-PathFM baselines on hospital-averaged AUC, those comparisons used different processing pipelines. Under matched single-scale conditions with alignment, CHIME-MIL achieved an AUC of 0.8274, compared with 0.8358 for attention-based MIL; the difference was not statistically significant. Global max pooling with the same alignment and two-scale fusion achieved an AUC of 0.8552, significantly exceeding the headline CHIME-MIL result by 0.0178 (95% confidence interval, 0.0092–0.0261; P = 0.0002). In the deterministic component analysis, multi-scale fusion was the only individual addition associated with a statistically supported improvement, increasing AUC by approximately 0.0092. The region-ablation loss did not improve discrimination. Mean-centring substantially reduced hospital-identifying information: the accuracy of a nearest-neighbour hospital classifier fell from 0.910 to 0.709. However, classification gains were small and inconsistent across aggregators and evaluation metrics. Replacing held-out-hospital statistics with training-hospital means reduced the fused model’s AUC by only 0.0019, from 0.8374 to 0.8355. In the single-scale onboarding experiment, ten unlabelled slides yielded a mean AUC within 0.001 of that obtained using all available slides, suggesting that low-cost adaptation without retraining may be feasible. Calibration and interpretability analyses revealed both strengths and limitations. Top-label expected calibration error was 0.053, but malignancy probabilities were systematically underconfident. Removing highly attended patches reduced predicted malignancy probability by 0.104 on average, compared with 0.0006 for random removal. This supports the relevance of these patches to model predictions but does not establish clinical usefulness. On a single-source CAMELYON16 subset, CHIME-MIL achieved an AUC of 0.9916, without a statistically significant difference from simpler attention models. This experiment assessed architectural transfer rather than cross-hospital alignment. The authors conclude that stronger feature representations, rather than more elaborate aggregators, are the more promising route to substantial improvements in discrimination for this task. Label-free onboarding, calibration assessment, and auditable evidence nevertheless remain important deployment considerations. Limitations include the retrospective design, a binary gastric classification task restricted to five Korean centres, unavailable sex and gender metadata, and the absence of a reader study demonstrating diagnostic benefit. External multi-institutional validation of alignment and broader clinical evaluation remain necessary.
Created on 02 Nov. 2026

Was this synthesis useful?

Score: 0

Why do we need votes?

Votes are used to determine whether we need to re-run our summarizing tools. If the count reaches -10, our tools can be restarted.

Ask questions about this paper to our AI assistant

How to chat with this paper? Illustrated guide

Compare with the PDF

You can also chat with multiple papers at once here.

Figures and plots

Ask about a figure by number or use its question button to request visual analysis. Up to two figures and four image panels per question. Values read from plots are approximate; check the original data.
AI assistant instructions?

Tools for this paper

Explain it differently 3 free rewrites left today
Save to a collection Group papers in collections to compare them, chat with them and export them. Create a free account Log in
Export BibTeX · RIS (Zotero, Mendeley, EndNote) · Obsidian (Markdown) · Notion (Markdown)

Advanced tools Premium

For researchers who follow a field closely: citation alerts, journal-club slides, version comparison and Zotero export. Discover Premium

Zotero
Citation alerts Citation alerts guide
Who works on it? The people and institutions who publish on its topic
Weekly letter Presented in "Computing beyond AI this week" of 2 November 2026. Computing beyond AI this week: Better algorithms, smaller machines, harder tests
Journal club Premium+ Slides to present it Propose for the journal club
Versions What changed between versions
Talk about it in public Premium+ Press kit: release, questions, hype check Press kit guide

Papers around this one

See them on a map

Ranked by how close their title and abstract are to this paper’s. Select some to read them together.

Closest papers

  1. Close Broader context
    Federated Model Aggregation via Self-Supervised Priors for Highly Imbalanced Medical Image Classification
    Marawan Elbatel, Hualiang Wang, Robert Martí et al. · 2023 · Computer Vision and Pattern Recognition · cited 11 times · AI summary available

    Federated aggregation addresses scanner- and population-related variation across medical-image clients, while this study tests unseen-hospital generalization for gastric slides.

    Abstract

    In the medical field, federated learning commonly deals with highly imbalanced datasets, including skin lesions and gastrointestinal images. Existing federated methods under highly imbalanced datasets primarily focus on optimizing a global model without incorporating the intra-class variations that can arise in medical imaging due to different populations, findings, and scanners. In this paper, we study the inter-client intra-class variations with publicly available self-supervised auxiliary networks. Specifically, we find that employing a shared auxiliary pre-trained model, like MoCo-V2, locally on every client yields consistent divergence measurements. Based on these findings, we derive a …

  2. Close Shared data, tool or method
    Virchow: A Million-Slide Digital Pathology Foundation Model
    Eugene Vorontsov, Alican Bozkurt, Adam Casson et al. · 2023 · Image and Video Processing · preprint, not peer reviewed · cited 164 times · AI summary available

    Virchow provides a self-supervised pathology foundation model for H&E slides, whereas this study evaluates classifiers on frozen pathology foundation-model features.

    Abstract

    The use of artificial intelligence to enable precision medicine and decision support systems through the analysis of pathology images has the potential to revolutionize the diagnosis and treatment of cancer. Such applications will depend on models' abilities to capture the diverse patterns observed in pathology images. To address this challenge, we present Virchow, a foundation model for computational pathology. Using self-supervised learning empowered by the DINOv2 algorithm, Virchow is a vision transformer model with 632 million parameters trained on 1.5 million hematoxylin and eosin stained whole slide images from diverse tissue and specimen types, which is orders of magnitude more data t…

  3. Close Shared data, tool or method
    PathoDuet: Foundation Models for Pathological Slide Analysis of H&E and IHC Stains
    Shengyi Hua, Fang Yan, Tianle Shen et al. · 2023 · Computer Vision and Pattern Recognition · preprint, not peer reviewed · cited 42 times · AI summary available

    PathoDuet develops pretrained pathology models for H&E and IHC slides, while this study uses frozen pathology foundation-model features for gastric classification.

    Abstract

    Large amounts of digitized histopathological data display a promising future for developing pathological foundation models via self-supervised learning methods. Foundation models pretrained with these methods serve as a good basis for downstream tasks. However, the gap between natural and histopathological images hinders the direct application of existing methods. In this work, we present PathoDuet, a series of pretrained models on histopathological images, and a new self-supervised learning framework in histopathology. The framework is featured by a newly-introduced pretext token and later task raisers to explicitly utilize certain relations between images, like multiple magnifications and …

  4. Close Another approach to the problem
    Towards better understanding and better generalization of few-shot classification in histology images with contrastive learning
    Jiawei Yang, Hanbo Chen, Jiangpeng Yan et al. · 2022 · Image and Video Processing · cited 28 times · AI summary available

    The few-shot study uses contrastive learning and latent augmentation for cross-domain histology classification, while this study evaluates cross-hospital gastric classification with frozen features.

    Abstract

    Few-shot learning is an established topic in natural images for years, but few work is attended to histology images, which is of high clinical value since well-labeled datasets and rare abnormal samples are expensive to collect. Here, we facilitate the study of few-shot learning in histology images by setting up three cross-domain tasks that simulate real clinics problems. To enable label-efficient learning and better generalizability, we propose to incorporate contrastive learning (CL) with latent augmentation (LA) to build a few-shot system. CL learns useful representations without manual labels, while LA transfers semantic variations of the base dataset in an unsupervised way. These two c…

  5. Related Broader context
    Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy
    Gioele Barabucci, Victor Shia, Eugene Chu et al. · 2024 · Artificial Intelligence · cited 7 times · AI summary available

    This study aggregates language-model diagnostic responses on clinical vignettes, whereas the gastric slide study compares multiple-instance learning aggregators for image classification.

    Abstract

    Background: Large language models (LLMs) such as OpenAI's GPT-4 or Google's PaLM 2 are proposed as viable diagnostic support tools or even spoken of as replacements for "curbside consults". However, even LLMs specifically trained on medical topics may lack sufficient diagnostic accuracy for real-life applications. Methods: Using collective intelligence methods and a dataset of 200 clinical vignettes of real-life cases, we assessed and compared the accuracy of differential diagnoses obtained by asking individual commercial LLMs (OpenAI GPT-4, Google PaLM 2, Cohere Command, Meta Llama 2) against the accuracy of differential diagnoses synthesized by aggregating responses from combinations of th…

  6. Related Broader context
    Is Dataset Quality Still a Concern in Diagnosis Using Large Foundation Model?
    Ziqin Lin, Heng Li, Zinan Li et al. · 2024 · Image and Video Processing · preprint, not peer reviewed · cited 1 time · AI summary available

    This study examines dataset quality and bias in fundus diagnosis using a foundation model, while the gastric study tests unseen-hospital discrimination with pathology features.

    Abstract

    Recent advancements in pre-trained large foundation models (LFM) have yielded significant breakthroughs across various domains, including natural language processing and computer vision. These models have been particularly impactful in the domain of medical diagnostic tasks. With abundant unlabeled data, an LFM has been developed for fundus images using the Vision Transformer (VIT) and a self-supervised learning framework. This LFM has shown promising performance in fundus disease diagnosis across multiple datasets. On the other hand, deep learning models have long been challenged by dataset quality issues, such as image quality and dataset bias. To investigate the influence of data quality …

  7. Related Broader context
    Customizing General-Purpose Foundation Models for Medical Report Generation
    Bang Yang, Asif Raza, Yuexian Zou et al. · 2023 · Computer Vision and Pattern Recognition · preprint, not peer reviewed · cited 15 times · AI summary available

    This study customizes general-purpose foundation models for medical report generation, whereas the gastric study uses pathology-model features for slide classification.

    Abstract

    Medical caption prediction which can be regarded as a task of medical report generation (MRG), requires the automatic generation of coherent and accurate captions for the given medical images. However, the scarcity of labelled medical image-report pairs presents great challenges in the development of deep and large-scale neural networks capable of harnessing the potential artificial general intelligence power like large language models (LLMs). In this work, we propose customizing off-the-shelf general-purpose large-scale pre-trained models, i.e., foundation models (FMs), in computer vision and natural language processing with a specific focus on medical report generation. Specifically, follo…

  8. Related Another approach to the problem
    StyPath: Style-Transfer Data Augmentation For Robust Histology Image Classification
    Pietro Antonio Cicalese, Aryan Mobiny, Pengyu Yuan et al. · 2020 · Computer Vision and Pattern Recognition · cited 19 times · AI summary available

    StyPath uses stain-style transfer augmentation to improve kidney histology classification robustness, while this study tests feature alignment and aggregation across hospitals.

    Abstract

    The classification of Antibody Mediated Rejection (AMR) in kidney transplant remains challenging even for experienced nephropathologists; this is partly because histological tissue stain analysis is often characterized by low inter-observer agreement and poor reproducibility. One of the implicated causes for inter-observer disagreement is the variability of tissue stain quality between (and within) pathology labs, coupled with the gradual fading of archival sections. Variations in stain colors and intensities can make tissue evaluation difficult for pathologists, ultimately affecting their ability to describe relevant morphological features. Being able to accurately predict the AMR status ba…

Go further

How are these papers found?

We compare the title and abstract of this paper with those of every paper summarized on this site, and with the papers Semantic Scholar recommends, those it cites and those citing it (for a much-cited paper, its most recent citations and, from OpenAlex, the most cited ones). Among the prior work and the follow-ups, the papers that use each other in depth, a method it uses and the much-cited follow-ups come a little higher. The levels (very close, close, related) come from the similarity of the texts, not from a judgment of quality; * marks a paper without an abstract, compared by its title. The sentence under a paper is written by an AI from the two abstracts: check it in the papers. A paper marked not relevant by three readers leaves the list.

Updated on 2 November 2026.

Disclaimer: The AI-based summarization tool and virtual assistant provided on this website may not always provide accurate and complete summaries or responses. We encourage you to carefully review and evaluate the generated content to ensure its quality and relevance to your needs.