Related Experiment Video
Updated: Aug 5, 2026
Quantitation of Protein Expression and Co-localization Using Multiplexed Immuno-histochemical Staining and Multispectral Imaging
Published on: April 8, 2016
Integration of pathomics and AutoML for precise classification of intestinal metaplasia subtypes in standard H&E
Xinyu Fu1,2, Tianming Guo2, Min Zhang2
1Department of Gastroenterology, The First Affiliated Hospital of Dalian Medical University, Dalian, China.
Objectives:
Gastric intestinal metaplasia (GIM) is a key precancerous lesion, with incomplete intestinal metaplasia (IIM) conferring a higher risk of malignant transformation. While experienced pathologists can preliminarily differentiate complete intestinal metaplasia (CIM) from IIM on H&E-stained slides, manual assessment is limited by substantial inter-observer variability and the potential for missed detection of microlesions in routine practice. Alcian blue/hematoxylin (AB/HID) special staining remains the diagnostic gold standard; however, its procedural complexity and limited accessibility impede early standardized management of GIM.
Methods:
A retrospective study included 324 GIM patients (193 CIM, 131 IIM including mixed types; 113 pure IIM for sensitivity analysis) after excluding 14 poor-quality samples. Prov-GigaPath extracted 768-dimensional features from H&E whole-slide images, refined via Spearman correlation and LASSO regression. Models were built using H2O AutoML and AutoGluon, with performance evaluated by AUC.
Results:
In the initial cohort containing mixed-type lesions, the optimal H2O-based model achieved a cross-validated AUC of 0.835 and an external validation AUC of 0.850, outperforming the AutoGluon model (0.816). After excluding all mixed-type lesions in a sensitivity analysis, the AutoGluon model showed superior performance, with a cross-validated AUC of 0.907 and an external validation AUC of 0.886. Core discriminative features were identified via SHAP interpretability analysis. Human-machine comparison further demonstrated that the optimal model outperformed two pathologists with five years of experience in gastrointestinal pathological diagnosis.
Conclusion:
The H&E-based pathomics model reduces reliance on specialized staining, offering high accuracy, cost-effectiveness, and broad applicability. It facilitates precise management of gastric precancerous lesions and contributes to the prevention of early gastric cancer.