Deep Learning to Classify Intraductal Papillary Mucinous Neoplasms Using Magnetic Resonance Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Intraductal papillary mucinous neoplasia (IPMN) is a premalignant lesion of the pancreas.
• Accurate identification of dysplasia and malignancy in IPMN is crucial for patient management.
• Current radiographic criteria have limitations in precisely classifying IPMN risk.

Purpose of the Study:

• To evaluate a novel deep learning (DL) protocol for identifying neoplasia in IPMN.
• To compare the diagnostic performance of the DL protocol against established radiographic criteria.
• To assess the potential of DL frameworks as aids for radiologists in IPMN risk stratification.

Main Methods:

• A computer-aided framework utilizing convolutional neural networks was developed to classify IPMN.
• The DL protocol was applied to magnetic resonance images of the pancreas.
• Sensitivity and specificity were calculated by comparing DL classifications with histopathological findings from resected lesions or controls.

Main Results:

• The DL protocol demonstrated a sensitivity of 92% and specificity of 52% for detecting any dysplasia.
• For high-grade dysplasia or adenocarcinoma, the DL protocol achieved 75% sensitivity and 78% specificity.
• The area under the receiver operating characteristic curve for the DL protocol (0.78) was comparable to American Gastroenterology Association (0.76) and Fukuoka (0.77) guidelines.

Conclusions:

• The deep learning protocol exhibits diagnostic accuracy comparable to existing radiographic criteria for IPMN.
• Computer-aided DL frameworks show promise as assistive tools for radiologists in identifying high-risk IPMN.
• Implementation of DL could enhance the detection and management of pancreatic neoplasia.