Deep learning to predict extrapancreatic perineural invasion at CT images

Area of Science:

• Medical Imaging
• Oncology
• Artificial Intelligence

Background:

• Extrapancreatic perineural invasion (EPNI) is a significant adverse prognostic factor in pancreatic ductal adenocarcinoma (PDAC).
• EPNI is associated with positive resection margins, complicating surgical outcomes.
• Accurate identification of EPNI is crucial for patient management and treatment planning.

Purpose of the Study:

• To develop an automated deep learning model for segmenting the extrapancreatic nerve plexus.
• To create a model for diagnosing EPNI based on segmented nerve structures.
• To evaluate the model's performance in a retrospective cohort of PDAC patients.

Main Methods:

• Retrospective analysis of enhanced CT scans from 332 PDAC patients (August 2018-December 2022).
• Utilized nnUNet and attention mechanisms for extrapancreatic nerve plexus segmentation.
• Employed a 2D classifier for EPNI diagnosis, validated using Dice Similarity Coefficients (DSCs) and ROC curves.

Main Results:

• The model achieved modest DSCs for nerve plexus segmentation around major arteries (e.g., SMA 68.2%).
• The EPNI diagnostic model demonstrated favorable performance with an accuracy of 0.797 and AUC of 0.8 in the training set.
• Validation set results showed an accuracy of 0.72 and AUC of 0.85 for EPNI diagnosis.

Conclusions:

• A fully automatic deep learning model for nerve plexus segmentation and EPNI diagnosis is a novel and promising tool.
• The developed model shows potential for improving the identification of EPNI in PDAC.
• Further research is needed to enhance model performance and clinical applicability.