A CT-based deep learning model to predict local recurrence-free survival in primary retroperitoneal sarcoma

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Deep learning (DL) and radiomics show promise in survival prediction.
• Predicting local recurrence-free survival (LRFS) in retroperitoneal sarcoma (RPS) using these methods is unexplored.
• This study aimed to develop a DL framework for LRFS prediction in RPS.

Purpose of the Study:

• To construct a DL framework using preoperative CT scans to predict LRFS in RPS patients.
• To evaluate the performance of the DL model against conventional radiomics and clinical models.
• To assess the utility of an integrated DL and clinical model for improved prediction.

Main Methods:

• Retrospective enrollment of 115 primary RPS patients, split into training (N=86) and validation (N=29) sets.
• Development of an end-to-end DL model for LRFS prediction using contrast-enhanced CT images.
• Comparison of DL-based score (DL-score) with handcrafted radiomics (Rad-score) and clinical models; integrated models (DLCM, RSCM) were also evaluated.

Main Results:

• The DL-score demonstrated superior performance over Rad-score and clinical models, with higher C-indices in both training and validation sets.
• The DL-score was an independent predictor of LRFS and effectively categorized patients into high- and low-risk groups.
• The integrated DLCM achieved the highest performance, showing strong calibration and clinical utility for risk classification.

Conclusions:

• A CT-based DL model can effectively predict LRFS preoperatively in RPS patients.
• The developed DL framework aids in risk stratification.
• This tool can guide individualized therapeutic strategies for RPS management.