Deep learning automatic segmentation and radiomics model for diagnosing pancreatic solid neoplasms in MRI

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Pancreatic solid neoplasms present diagnostic challenges in MRI.
• Accurate segmentation is crucial for diagnosis and treatment planning.

Purpose of the Study:

• To develop and validate a deep learning tool for automatic segmentation of pancreatic solid neoplasms on MRI.
• To establish a radiomics model for diagnosing these neoplasms.

Main Methods:

• A 3D nnU-Net deep learning model was trained on plain MRI scans for automatic segmentation.
• A radiomics model was developed using segmentations for neoplasm diagnosis.
• Performance was evaluated using Dice Similarity Coefficient (DSC) and receiver operating characteristic (ROC) analysis.

Main Results:

• The deep learning model achieved high segmentation performance (mean DSC 0.82-0.91 training, 0.64-0.70 testing).
• Segmentation was reasonably efficient for lesions < 2 cm (DSC 0.51-0.92).
• The radiomics model showed high accuracy in differentiating pancreatic ductal adenocarcinomas (PDACs) (AUC 0.968 training, 0.790 testing).

Conclusions:

• Deep learning automatic segmentation is accurate for pancreatic neoplasms on MRI.
• The radiomics model effectively differentiates PDACs from other pancreatic neoplasms.