Objective:

To develop a multi-task deep learning framework that simultaneously predicts microsatellite instability (MSI) and tumor mutation burden (TMB) using routine histopathological images and clinical data, thereby enhancing patient selection for immunotherapy.

Key Findings:

• The model achieved AUC values of 0.828 for MSI and 0.836 for TMB on the internal TCGA test set, indicating strong predictive performance.
• Performance on the external validation set yielded AUCs of 0.78 for MSI and 0.74 for TMB, indicating a moderate decrease due to domain shifts, which may affect clinical applicability.
• Attention heatmaps provided insights into the spatial concordance of predictive regions for MSI and TMB, suggesting areas for further biological investigation.

Interpretation:

The study demonstrates the feasibility and accuracy of a unified, multi-task deep learning framework for predicting key immunotherapy biomarkers in gastric cancer using routine histopathological images.

Limitations:

• External validation highlighted challenges in generalizability across different scanners, which may limit the model's application in diverse clinical settings.
• Performance decreased on the external validation set compared to the internal test set, raising concerns about the model's robustness in real-world scenarios.

Conclusion:

The framework represents a significant innovation with potential to lower barriers to precision oncology in clinical practice, serving as a cost-effective preliminary screening tool for MSI and TMB in gastric cancer, and warrants further research to enhance its generalizability.