Development and validation of an interpretable CT-based scoring model for gastric cancer aggressiveness

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Accurate prediction of adverse histopathological status (AHS) in gastric cancer (GC) is critical for patient prognosis.
• Current methods may lack precision in identifying high-risk GC cases preoperatively.

Purpose of the Study:

• To develop and validate a CT-based machine learning model for predicting AHS in GC.
• To establish an interpretable imaging score (I-score) for prognostic stratification and treatment planning.

Main Methods:

• A dual-center retrospective study included 1164 GC patients.
• Semantic CT features (cN, LD, TT, serosal status) and radiomic features were analyzed.
• XGBoost model was trained and validated for AHS prediction, leading to the I-score development.

Main Results:

• The XGBoost model demonstrated high predictive performance for AHS (AUCs ranging from 0.764 to 0.848).
• The I-score successfully stratified patients into low-risk and high-risk groups.
• High-risk patients exhibited significantly poorer 1000-day overall survival (55.7% vs. 73.8%).
• The I-score was confirmed as an independent prognostic factor.

Conclusions:

• A simplified CT-based machine learning model using semantic features accurately predicts AHS in GC.
• The developed I-score facilitates effective preoperative risk stratification and supports individualized treatment planning.