A Cascaded Segmentation-Classification Deep Learning Framework for Preoperative Prediction of Occult Peritoneal...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Advanced gastric cancer (GC) poses challenges in predicting occult peritoneal metastasis (OPM) and early peritoneal recurrence (PR).
• Accurate preoperative risk stratification is crucial for personalized treatment strategies in GC patients.

Purpose of the Study:

• To develop and validate a cascaded deep learning (DL) framework for predicting OPM and early PR in advanced GC.
• To integrate tumor segmentation and risk stratification for enhanced preoperative prediction accuracy.

Main Methods:

• A multicenter study involving 765 advanced GC patients.
• Development of a two-stage DL framework: V-Net tumor segmentation followed by DL-based risk classification.
• Integration of clinicopathological predictors with DL probabilities for a combined predictive model.

Main Results:

• The combined model achieved high AUCs for OPM prediction (0.938 Train, 0.916 Test1) and early PR prediction (0.820-0.825 Test2/Test3).
• Demonstrated significant improvements over clinical and DL-only models for OPM prediction.
• Identified Borrmann type, CA125 levels, and CT-N+ stage as independent OPM predictors.

Conclusions:

• The CT-based cascaded DL framework provides reliable preoperative risk stratification for OPM and early PR in advanced GC.
• This approach can potentially refine treatment indications and personalize therapeutic pathways for GC patients.