Introduction
Gastric cancer remains a leading cause of cancer-related mortality worldwide. Early detection through the identification of precursor lesions, specifically gastric dysplasia, is critical for improving patient outcomes. However, histopathological diagnosis of these lesions is challenging due to significant interobserver variability among pathologists, particularly in distinguishing benign regenerative changes from low-grade dysplasia (LGD).
Furthermore, the clinical deployment of artificial intelligence (AI) in pathology is often hindered by "domain shift"—a phenomenon where models trained on data from one institution perform poorly on data from another due to variations in staining, sectioning, and scanning protocols. This study introduces DLDA, a deep learning framework designed to provide accurate, generalizable diagnosis of gastric dysplasia across different hospital settings.
Methodology
The researchers developed a two-stage diagnostic framework utilizing a ResNet-50 backbone. The architecture is designed to handle both local tissue features and global slide context while addressing the challenge of data variability.
Two-Stage Framework
- Patch-level Module with Domain Adaptation (DA): To overcome domain shift, the model incorporates an unsupervised DA strategy. When applied to a new "target" hospital, the model processes unlabeled slides and uses high-confidence predictions as "pseudo-labels" to refine its parameters. This allows the CNN to adapt to the specific visual characteristics of the new dataset without requiring manual re-labeling by pathologists.
- WSI-level Module: The refined patch-level model generates a probability map for the entire whole-slide image (WSI). A second model then analyzes this map to produce a final diagnosis categorized as Negative for Dysplasia (NFD), Low-Grade Dysplasia (LGD), or High-Grade Dysplasia/Intramucosal Neoplasia (HGD/IMN).
Experimental Results
The study evaluated the model using 897 WSIs from two different hospitals and conducted an observer study involving 12 pathologists of varying experience levels.
Performance and Generalizability
On the internal validation set, the model achieved a macro-averaged Area Under the Curve (AUC) of 0.97. Crucially, the unsupervised DA strategy significantly improved performance on external data, raising the macro-averaged AUC from 0.67 to 0.82.
In the observer study, the DLDA model demonstrated superior sensitivity and Negative Predictive Value (NPV) compared to junior and senior pathologists for NFD and HGD cases. For instance, the model achieved 95% sensitivity for NFD, outperforming even specialist pathologists (88.7%). This high NPV suggests the tool is particularly effective as a screening layer to prevent overdiagnosis and reduce pathologist workload.
Visual Interpretability
To assist clinical decision-making, the model generates color-coded heatmaps that highlight suspicious regions: green for NFD, yellow for LGD, and red for HGD. These maps allow pathologists to quickly navigate to areas of interest on the slide.
Discussion and Limitations
While the DLDA model shows remarkable performance in identifying high-grade lesions and negative cases, diagnosing Low-Grade Dysplasia remains a challenge for both the AI and human experts. LGD often lacks distinct architectural features at high magnification, requiring broader context that patch-based models may struggle to capture fully.
Future work will focus on improving LGD accuracy and expanding the model's scope to include rarer gastric cancer subtypes. Additionally, integrating multi-modal data, such as endoscopic images and immunohistochemistry, could further enhance diagnostic precision. While validated across two Chinese hospitals, broader international validation is necessary to ensure the model's robustness across different global diagnostic standards.
Summary updated