Explainable Deep Learning for Automated Histopathological Diagnosis of Gastrointestinal Malignancies: A Multicenter External Validation Study
Advanced Journal of Biomedicine & Medicine · 30.06.2026 · 14 (2) · 119-149
DOI: 10.18081/ajbm.2026.2.119 ↗Kurz gefasst
Fachartikel in der Zeitschrift Advanced Journal of Biomedicine & Medicine, online erschienen am 30.06.2026. Verfasst von Min-Seo Park und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Background
Histopathological examination remains the gold standard for diagnosing gastrointestinal (GI) malignancies; however, increasing diagnostic workloads, growing cancer incidence, and interobserver variability continue to challenge pathology services worldwide. Recent advances in artificial intelligence (AI), particularly deep learning, have demonstrated considerable potential to enhance diagnostic accuracy and efficiency in digital pathology. Nevertheless, the clinical applicability of many existing models remains limited because of inadequate external validation, restricted multicenter datasets, and insufficient interpretability. This study aimed to develop and externally validate an explainable deep learning framework for the automated histopathological diagnosis of gastrointestinal malignancies using multicenter whole-slide images obtained from tertiary hospitals in South Korea.
Methods
A retrospective multicenter diagnostic study was conducted using digitized hematoxylin and eosin–stained whole-slide images collected from four tertiary referral hospitals in South Korea. After quality control, 5,028 patients with 7,588 whole-slide images were included. The development cohort consisted of 3,954 patients, while an independent external cohort included 1,074 patients. Whole-slide images underwent standardized preprocessing, stain normalization, tissue segmentation, and patch extraction before training an attention-based deep learning architecture using a convolutional neural network integrated with multiple-instance learning. Model performance was evaluated by the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, accuracy, positive predictive value, negative predictive value, F1 score, calibration analysis, and external validation. Explainability was assessed using attention heatmaps and Gradient-weighted Class Activation Mapping (Grad-CAM). A reader study involving gastrointestinal pathologists, general pathologists, and pathology trainees evaluated the clinical impact of AI-assisted diagnosis.
Results
The proposed model demonstrated excellent diagnostic performance across both internal and external validation cohorts. In the independent external cohort, the model achieved an AUROC of 0.972 (95% CI: 0.963–0.980), with a sensitivity of 93.8%, specificity of 91.5%, overall accuracy of 92.8%, positive predictive value of 94.5%, negative predictive value of 90.3%, and an F1 score of 0.941 for distinguishing malignant from non-malignant gastrointestinal tissue. Patient-level AUROC reached 0.978, confirming robust generalizability across institutions and scanner platforms. Multiclass classification achieved an overall external accuracy of 89.3%, with the highest performance observed for esophageal squamous cell carcinoma, colorectal adenocarcinoma, and gastric adenocarcinoma. Explainability analyses demonstrated that attention maps consistently localized diagnostically relevant histological structures, including infiltrative glands, nuclear pleomorphism, desmoplastic stroma, and tumor necrosis. In the reader study, AI assistance increased overall diagnostic accuracy from 88.6% to 94.1% (P < 0.001) while reducing median slide interpretation time by 34.3%. The greatest improvement was observed among general pathologists and pathology trainees.
Conclusion
An explainable deep learning framework demonstrated high diagnostic accuracy, excellent external generalizability, and meaningful clinical interpretability for the automated histopathological diagnosis of gastrointestinal malignancies. Integration of AI into routine digital pathology significantly improved diagnostic performance and workflow efficiency while maintaining expert pathological oversight. These findings support the implementation of explainable deep learning as an effective clinical decision-support tool in gastrointestinal pathology and highlight its potential contribution to precision oncology. Prospective multinational validation studies are warranted to confirm clinical utility and facilitate widespread implementation in routine pathological practice. Keywords: Artificial intelligence; Deep learning; Digital pathology; Whole-slide imaging; Gastrointestinal malignancies; Histopathology; Computational pathology
Zitieren
APA
Park, M., & Kim, J. (2026). Explainable Deep Learning for Automated Histopathological Diagnosis of Gastrointestinal Malignancies: A Multicenter External Validation Study. Advanced Journal of Biomedicine & Medicine, 14(2), 119-149. https://doi.org/10.18081/ajbm.2026.2.119
Vancouver
Park M, Kim J. Explainable Deep Learning for Automated Histopathological Diagnosis of Gastrointestinal Malignancies: A Multicenter External Validation Study. Advanced Journal of Biomedicine & Medicine. 2026;14(2):119-149. doi:10.18081/ajbm.2026.2.119
BibTeX
@article{Park2026,
author = {Park, Min-Seo and Kim, Ji-Hoon},
title = {Explainable Deep Learning for Automated Histopathological Diagnosis of Gastrointestinal Malignancies: A Multicenter External Validation Study},
journal = {Advanced Journal of Biomedicine & Medicine},
year = {2026},
volume = {14},
number = {2},
pages = {119--149},
doi = {10.18081/ajbm.2026.2.119}
}
EUVIMED