Simpler pooling fares better in a five-hospital gastric pathology AI study

A five-hospital study compares ways to combine evidence for gastric pathology classification.

Researchers testing AI on gastric tissue slides found that a simpler way of combining image features performed better across hospitals than a more elaborate model. The October 30 preprint draws on 5,036 stained slides from five hospitals in South Korea. In each round of validation, the researchers held out a hospital and tested on its slides. That design asks a harder question than whether a model works on fresh examples from a site that contributed training data. [Study preprint](https://arxiv.org/abs/2610.41328)

What changed in a fair comparison

Whole-slide images are too large to treat like ordinary photographs. A common approach divides a slide into patches, extracts features from each, then combines the features into a slide-level result. The authors compared several aggregation methods using features from a frozen pathology foundation model. Their graph method first looked stronger than some alternatives. Those alternatives, however, had been tested without the same alignment and two-scale input recipe.

With that recipe matched, global max pooling reached a mean area under the receiver operating characteristic curve of 0.855, against 0.837 for the graph-based method. AUC measures how well the classifier ranks positive and negative examples across thresholds. It is one part of a clinical evaluation. A hospital still needs to know how a chosen threshold affects false alarms and missed cases, and whether the model’s probabilities reflect observed outcomes. The paper reported some underconfidence in malignancy estimates. [Study preprint](https://arxiv.org/abs/2610.41328)

Why the hospital boundary matters

The same kind of slide can look different across hospitals because of factors such as scanning and preparation. The study tested a site-alignment step that subtracts an estimate of each hospital’s average feature vector. The authors examined estimating it with ten unlabeled slides from a new site and reported that accuracy benefits were small and inconsistent. The result supports testing adaptation methods as separate components, with comparison groups that receive the same inputs.

For teams assessing pathology tools, the key contribution is the matched comparison across held-out hospitals. It shows how a change credited to an elaborate architecture can come from another part of the pipeline. The preprint offers evidence on this dataset and validation design. Clinical performance in a particular workflow would require additional evaluation, including how staff use the result and how the system behaves at the thresholds they choose. [Study preprint](https://arxiv.org/abs/2610.41328)