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

PURPOSE: Pancreatic ductal adenocarcinoma (PDA) is a leading cause of cancer-related deaths, with early diagnosis hampered by nonspecific symptoms and limitations of existing imaging techniques. This study aimed to develop a deep learning (DL) algorithm to automatically classify pancreas lesions on contrast-enhanced CT scans as normal, benign, or malignant, to assist radiologists in detecting early-stage pancreatic cancer.

MATERIALS AND

METHODS

A dataset of 1,037 portal-phase CT scans was compiled from 18 institutions. The dataset was divided into a training set (N = 732) and a test set (N = 305), which was further divided into an internal validation test set (N = 139) and an external validation test set (N = 166). After segmentation using the TotalSegmentator algorithm, the pancreas was isolated from each CT scan. A DL model combining TotalSegmentator's pre-trained encoder and nnUNet decoder layers was developed to classify pancreas lesions. Ten-fold cross-validation was applied, and model performance was assessed using precision, recall, area under the curve (AUC) and a final score (FS) representing a weighted average of the previous three metrics. The final prediction combined the outputs of ten models.

RESULTS

Across all validation datasets (139 and 166 patients, respectively, in the internal and external dataset), precision and recall were 0.57 and 0.63, respectively, while AUC was 0.84. In the external validation dataset, malignant lesions were detected with an AUC of 0.97. The model achieved an FS of 0.72 in both internal and external validation datasets, indicating consistent performance across datasets.

CONCLUSION

This study demonstrated the feasibility of using a DL algorithm for automated pancreas lesion classification in CT scans. The model showed strong performance, particularly in detecting malignant lesions. Further research is needed to assess the model's clinical applicability and performance in real-world settings.