Interpretable Deep Learning for Gastric Cancer T Staging
Original Title: Interpretable deep learning for multicenter gastric cancer T staging from CT images
Journal: NPJ digital medicine
DOI: 10.1038/s41746-025-02002-5
Overview
Gastric cancer remains a significant global health challenge, requiring precise preoperative T staging to determine the appropriate therapeutic strategy, such as neoadjuvant chemotherapy or direct surgical intervention. Standard contrast-enhanced computed tomography is the primary tool for this evaluation, yet its accuracy often ranges between 65% and 75% due to subjective interpretation and the difficulty of identifying subtle serosal invasion. This study introduces GTRNet, an automated deep-learning framework designed to classify gastric cancer into four T stages from routine portal venous phase images. Developed using a retrospective multicenter dataset of 1,792 patients, the system utilizes a modified ResNet-152 backbone to analyze the largest axial tumor cross-section. In internal testing, the model achieved an accuracy of 89.9% and an area under the curve (AUC) of 0.97. External validation across two independent cohorts demonstrated consistent performance, with accuracies between 87% and 94% and AUC values ranging from 0.91 to 0.95. Compared to expert radiologists, who achieved independent accuracies of 55.3% to 59.7%, GTRNet showed superior discrimination and consistency.
Novelty
The framework distinguishes itself by implementing an end-to-end pipeline that eliminates the need for time-consuming manual tumor segmentation or annotation, which are common bottlenecks in clinical AI applications. While previous research often focused on binary classifications like early versus advanced stages, GTRNet provides a complete four-category T-staging output. The architecture incorporates parallel max-pooling and center-cropping streams to capture both local tumor details and broader contextual information of the gastric wall. Furthermore, the researchers developed a comprehensive nomogram by integrating a deep-learning-derived Rad-score with clinical variables, including tumor size, differentiation status, and Lauren classification. This multimodal approach significantly improved model fit and clinical utility. To address the opaque nature of neural networks, Gradient-weighted Class Activation Mapping was utilized to visualize model attention. These heatmaps showed a high degree of spatial overlap with expert-annotated regions, specifically targeting the mucosa in T1 lesions and the organ interface in T4 cases, with Dice similarity coefficients ranging from 0.56 to 0.63.
Potential Clinical / Research Applications
This technology has direct implications for refining neoadjuvant therapy selection. By accurately identifying T3 and T4 cases, the system can ensure that patients who require preoperative chemotherapy receive it, while sparing T1 and T2 patients from unnecessary toxicity. Decision curve analysis indicated a higher net benefit for the AI model compared to endoscopic ultrasound, showing lower over-treatment (2.09% vs. 12.97%) and under-treatment (2.51% vs. 17.57%) rates. In research settings, the automated nature of GTRNet allows for the rapid processing of large-scale imaging datasets in retrospective studies or clinical trials. Additionally, the interpretable heatmaps can serve as an educational resource for junior radiologists, helping them recognize the subtle radiological signs of serosal invasion and transmural spread. The framework could eventually be expanded into a unified system covering the entire TNM staging protocol, offering a more comprehensive auxiliary diagnostic tool for gastric cancer management.