PURPOSE

Microsatellite instability (MSI) plays a crucial role in determining the therapeutic outcomes of gastroesophageal junction (GEJ) adenocarcinoma. This study aimed to develop a deep learning model based on H&E-stained pathological specimens to accurately identify MSI-H in GEJ adenocarcinomas patients.

METHODS

A total of 416 H&E-stained slides of 212 GEJ adenocarcinoma patients were collected to establish an artificial intelligence (AI) model using digital pathology (DP) for of MSI-H prediction. Simple Vit and ResNet18 Neural networks were trained and tested on models developed from patch-level images. A whole-slide image (WSI)-level AI model was constructed by integrating deep learning- generated pathological features with six machine learning algorithms.

RESULTS

The MLP model showed demonstrated the highest performance in predicting MSI-H in the test cohort, achieving an AUC of 93.3%, a sensitivity of 0.841, and a specificity of 0.952. Similarly, Decision Curve Analysis (DCA) revealed that WSI-level H&E-stained slides offered significant clinical MSI-H prediction in GEJ adenocarcinoma patients.

CONCLUSION

The AI model based on digital pathology exhibits great potential for predicting MSI-H in GEJ adenocarcinoma, suggesting promising clinical applications.