Interpretable deep learning for multicenter gastric cancer T staging from CT images

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Preoperative T staging of gastric cancer is crucial for treatment planning.
• Conventional contrast-enhanced CT interpretation for gastric cancer staging lacks consistent reliability and is subjective.

Purpose of the Study:

• To develop and validate an automated, interpretable deep learning framework (GTRNet) for preoperative T staging of gastric cancer using routine CT images.
• To compare the performance of GTRNet against radiologists in classifying tumor stages.

Main Methods:

• A retrospective multicenter study involving 1792 patients was conducted.
• GTRNet, an end-to-end deep learning model, was trained on CT images without manual segmentation or annotation.
• The model's performance was evaluated on internal and two independent external cohorts.

Main Results:

• GTRNet demonstrated high discrimination (AUC 0.86-0.95) and accuracy (81-85%) in both internal and external validation cohorts.
• The deep learning model outperformed human radiologists in T staging accuracy.
• Explainable AI (Grad-CAM) highlighted attention to the gastric wall and serosa, enhancing interpretability.

Conclusions:

• The developed GTRNet framework offers an automated and interpretable solution for CT-based gastric cancer T staging.
• This pipeline has the potential to standardize preoperative staging and improve the selection of neoadjuvant therapy.
• A nomogram combining GTRNet's rad-score with clinical factors showed improved clinical utility over conventional methods.