Interpretable Deep Learning for Accurate Multicenter Gastric Cancer T Staging

Overview

GTRNet, an interpretable deep learning model, accurately classifies gastric cancer T stages (T1–T4) from routine CT scans without manual segmentation. In a multicenter study of 1792 patients, it outperformed expert radiologists and demonstrated robust external validation, with AUCs ranging from 0.86 to 0.95 and accuracies between 81% and 85%. Integration of GTRNet outputs with clinical factors into a nomogram further enhanced preoperative risk stratification.

Background

Gastric cancer remains a leading cause of cancer mortality worldwide, with accurate preoperative T staging critical for guiding treatment strategies. Conventional contrast-enhanced CT interpretation is limited by subjectivity and moderate accuracy, especially in distinguishing T2 from T3 tumors and detecting serosal invasion. Endoscopic ultrasound offers improved layer visualization but is operator-dependent and less reliable in advanced or proximal tumors. Deep learning approaches, particularly convolutional neural networks, have shown promise in medical imaging but often require manual tumor segmentation, limiting scalability. GTRNet addresses these challenges by providing a fully automated, interpretable framework for four-class T staging using a single CT slice.

Data Highlights

Key Findings

• GTRNet achieved high discrimination with AUCs between 0.86 and 0.95 across internal and external cohorts.
• Accuracy of T staging ranged from 81% to 85%, outperforming expert gastrointestinal radiologists in a comparative reader study.
• The model requires only a single axial CT slice of the largest tumor cross-section, eliminating the need for manual segmentation or annotation.
• Grad-CAM heatmaps provided interpretable visualizations, localizing model attention to the gastric wall and serosa, enhancing clinical trust.
• Combining the deep learning rad-score with tumor size, differentiation, and Lauren subtype in a nomogram improved calibration and net clinical benefit over conventional methods.

Clinical Implications

GTRNet offers a standardized, objective tool for preoperative gastric cancer T staging that can reduce inter-observer variability and improve diagnostic accuracy. Its interpretability via saliency maps facilitates clinician acceptance and aids in understanding tumor invasion patterns. Integration into clinical workflows may support more precise therapeutic stratification, including neoadjuvant therapy selection, potentially improving patient outcomes.

Conclusion

This multicenter study demonstrates that GTRNet is a robust, interpretable deep learning framework that enhances the accuracy and consistency of gastric cancer T staging from routine CT imaging. Its automated approach and integration with clinical factors provide a promising avenue for improving preoperative decision-making.

References

1. GTRNet Study Authors 2024 -- Interpretable Deep Learning Approaches for T Staging of Gastric Cancer Using CT Imaging Across Multiple Centers