Interpretable prediction of occult lymph node metastasis in pancreatic ductal adenocarcinoma using a model fusing habitat radiomics and deep learning

Abstract

Purpose

To evaluate the value of integrating habitat radiomics features and deep learning features for predicting occult lymph node metastasis (OLNM) in pancreatic ductal adenocarcinoma (PDAC).

Methods

Data from 212 eligible PDAC patients across two institutions were analyzed. Cohorts were allocated as follows: training (n = 115), internal validation (n = 50), and external validation (n = 47). Habitat subregion partitioning of the tumor volume of interest (VOI) from portal venous phase computed tomography images was performed using a K-means clustering algorithm, and radiomics features were subsequently extracted. A 2.5D deep learning model based on ResNet18 was used to extract features from the whole VOI. After feature selection, models based on single-feature types and a fusion model integrating habitat radiomics features and deep learning features were developed. Model performance was assessed using receiver operating characteristic curves, decision curve analysis (DCA), and calibration curves. Model interpretability was evaluated via SHapley Additive exPlanations (SHAP).

Results

Relative to single-feature-based models, the fusion model achieved superior predictive performance with an area under the curve (AUC) of 0.832 (95% CI: 0.712–0.951) in the external validation cohort. DCA and calibration curves revealed that this model provided greater net clinical benefit compared with other models and demonstrated good calibration. SHAP analysis indicated that deep learning features were the top and third most important predictors.

Conclusion

The fusion model exhibited favorable predictive performance for preoperative OLNM diagnosis in PDAC and represents a promising auxiliary tool for personalized therapeutic decision-making.