A non-invasive artificial intelligence model to predict the severity of intrapancreatic perineural invasion.

Introduction: Intrapancreatic perineural invasion (IPNI) was a poor prognostic factor in pancreatic ductal adenocarcinoma (PDAC), but few studies quantified the severity of IPNI. Additionally, there was an absence of non-invasive tool to predicted IPNI.

Methods: Patients who underwent radical pancreatectomy betweem 2018 and 2022 were enrolled. Severity of IPNI was assessed via binary classification, numeration, and neural invasion (NI) severity score. Cox regression analysis was performed to clarify the relationship between IPNI and prognosis. Clinical data, artificial intelligence (AI) , and radiomics were used to develop predictive models, evaluated by receiver operating characteristics (ROC) curve and decision curve analysis (DCA). Kaplan-Meier analysis confirmed the model's clinical utility.

Results: A total of 192 patients were enrolled and rate of IPNI was 93.8%. Cox regression analysis and ROC curve indicated NI severity score>10.5 was responsible for poor prognosis (p<0.05). Further analysis indicated the AI model had better performance in predicting NI severity score>10.5 in both training set (clinical model AUC=0.66, AI model AUC=0.85, radiomics model AUC=0.81) and validation set (clinical model AUC=0.74, AI model AUC=0.84, radiomics model AUC=0.73). According to the predictive results of AI model, patient with NI severity score>10.5 had decreased recurrence free survival(RFS) and overall survival(OS) in training set (RFS 12months vs. 24 months, p=0.002; OS 24 months vs. 35 months, p=0.024) and validation set (RFS 8 months vs. 16 months, p=0.03; OS 25 months vs. 36 months, p=0.04).

Conclusions: NI severity score >10.5 predicted poor PDAC prognosis, and AI was an useful tool to predict it.