Deep learning to predict extrapancreatic perineural invasion at CT images.

BACKGROUND

Extrapancreatic perineural invasion (EPNI) was an adverse prognostic factor in patients with pancreatic ductal adenocarcinoma (PDAC) and responsible for positive resection margin. This study aimed to develop an automatic model for segmenting the extrapancreatic nerve plexus and diagnosing EPNI.

METHODS

In this retrospective study, patients diagnosed with PDAC who underwent enhanced computer tomography scans between August 2018 and December 2022 were enrolled. These cases were divided into training sets with radiological EPNI labels, and validation sets with pathological EPNI labels. The extrapancreatic nerve plexus was segmented first via the nnUNet network and attention mechanism under the background of segmentation of PDAC and adjacent vessels. A 2D classifier was applied to diagnose EPNI based on the segmentation of the nerve plexus. The Dice similarity coefficients (DSCs), receiver operating characteristic (ROC) curve, and diagnostic accuracy were employed to evaluate the performance of the model.

RESULTS

A total of 332 consecutive patients were enrolled and classified into the training (n = 282) and validation (n = 50) sets. Patients diagnosed with EPNI accounted for 177 of the 332 patients (53.3%). On the one hand, the model showed modest DSCs in segmenting nerve plexus around celiac axis (CA), superior mesenteric artery (SMA), and common hepatic artery (CHA), which were 60, 68.2 and 35.7%, respectively. On the other hand, the model had a favorable performance in diagnosing EPNI; the accuracy and areas under the ROC curve were 0.797, 0.8 in training set and 0.72, 0.85 in the validation set.

CONCLUSIONS

The fully automatic deep learning model for segmenting the nerve plexus and diagnosing EPNI was a novel and promising tool. Further studies are required to improve the model performance.