Detection and characterization of pancreatic lesions with 3D CT data using an automated deep learning-based solution

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Oncology

Background:

• Pancreatic ductal adenocarcinoma (PDA) presents a significant challenge in cancer mortality due to late diagnosis.
• Current imaging techniques for early detection of pancreatic cancer are limited.
• Nonspecific symptoms of PDA further complicate early identification.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) algorithm for automated classification of pancreas lesions on contrast-enhanced CT scans.
• The algorithm aims to distinguish between normal, benign, and malignant lesions.
• To assist radiologists in the early detection of pancreatic cancer.

Main Methods:

• A dataset of 1,037 CT scans was utilized, split into training and validation sets.
• A DL model was developed using TotalSegmentator for segmentation and an nnUNet decoder for classification.
• Performance was evaluated using precision, recall, AUC, and a final score (FS) via cross-validation.

Main Results:

• The DL model achieved a consistent Final Score (FS) of 0.72 across internal and external validation datasets.
• Area Under the Curve (AUC) for overall classification was 0.84.
• Malignant lesions were detected with a high AUC of 0.97 in the external validation set.

Conclusions:

• Deep learning algorithms show feasibility for automated pancreas lesion classification on CT scans.
• The developed model demonstrates strong performance, especially in detecting malignant pancreatic lesions.
• Further validation in real-world clinical settings is necessary to assess practical applicability.