Fusión interpretable de aprendizaje profundo en RM de superresolución para la predicción de invasión perineural en...

To develop an interpretable fusion deep learning model based on super-resolution (SR) MRI for predicting preoperative perineural invasion (PNI) in pancreatic ductal adenocarcinoma (PDAC) and to evaluate its role in guiding postoperative prognostic and therapeutic decision-making.

In this multicenter retrospective study, we enrolled 714 eligible patients, allocating 608 to a development/internal validation set and 106 from three external centers to an external validation set. The fusion clinical-radiomics-deep transfer learning (FCDR) was developed to predict PNI by integrating imaging signatures derived from deep learning and radiomics on SR-MRI with clinical risk factors and was optimized by selecting the best-performing among seven machine learning algorithms. The final model was subsequently validated for its incremental value in predicting postoperative prognosis and guiding adjuvant therapy.

The FCDR model achieved superior performance of PNI prediction with AUCs of 0.929, 0.886, and 0.832 across development, internal and external validation sets, significantly outperforming single clinical, DL, or radiomics models. The FCDR model-stratified high-risk group was associated with significantly worse postoperative OS and RFS (p < 0.05). Moreover, the high-risk PNI subgroup stratified by this model derived a significant OS and RFS benefit from adjuvant therapy. Model interpretability was affirmed by SHAP analysis.

The proposed interpretable fusion model serves as an effective tool for PNI evaluation, prognostic stratification, and tailoring of adjuvant therapy in PDAC, holding significant promise for personalized precision medicine.

The constructed fusion model offers a robust, non-invasive tool for identifying PNI in pancreatic ductal adenocarcinoma, showing significant potential to guide personalized treatment strategies and improve patient outcomes.