Predicting Intraductal Papillary Mucinous Neoplasms grade of dysplasia using convolutional neural networks: results from a multicenter study

Aims Pancreatic cystic lesions (PCLs) are increasingly common and pose a higher risk of malignancy, particularly mucinous PCLs (M-PCLs). M-PCLs vary in risk based on lesion type and dysplasia level. Intraductal papillary mucinous neoplasms (IPMNs)—including side-branch (SB-IPMNs), main duct (MD-IPMNs), and mixed-type (MT-IPMNs)—are the most frequent M-PCLs. While many have low malignant potential, some progress to invasive carcinoma. Management focuses on dysplasia degree to determine surgery necessity (high-grade dysplasia/carcinoma, HGD/C) or surveillance (low-grade dysplasia, LGD). This stratification is done through cytologic analysis of cystic fluid after endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA) which entails significant sensitivity limitations. AI is increasingly aiding in identifying premalignant pancreatic lesions, helping to prevent misdiagnosis, over-diagnosis, and unnecessary surgeries. This study aimed to develop an AI algorithm for automatic detection and stratification of IPMNs into HGD/C and LGD lesions [1] [2].

Methods This multicenter study included EUS images from four centers centers, three in Spain (Hospital Universitario Puerta de Hierro Majadahonda, Hospital Universitario Ramón y Cajal and Hospital Universitario Marqués de Valdecilla) and one in Portugal (Centro Hospitalar Universitário São João). IPMNs were categorized as having HGD/C or LGD according to available histopathological results, either from PCL cytological analysis, EUS-guided through-the-needle biopsy or surgical specimen. A CNN was developed using the PyTorch model. Performance metrics, including sensitivity, specificity, accuracy, and area under the curve (AUC) were calculated.

Results A total of 15756 images were extracted from 63 EUS exams. From these images, 7303 showed IPMNs with LGD and 8453 with HGD/C. The model distinguished IPMNs with HGD/C from those with LGD with a sensitivity of 96.9%, a specificiity of 98.5%, and an overall accuracy of 97.7%. The AUC was 0.99.

Conclusions To our knowledge, our work is one of the first to assess the potential of AI to predict the grade of dysplasia using EUS images. An accurate stratificiation of patients with HGD/C or LGD is paramount as it enables to decrease the risk of undertreating patients with higher risk lesions, while sparing those with lesions with lower malignant potential to the morbidity of unnecessary surgical treatment. Subsequent development and validation of this model can provide endoscopists with real-time feedback, potentially mitigating the impact of the limitations and adverse events associated with the current diagnostic techniques.

Publikationsverlauf

Artikel online veröffentlicht:
27. März 2025

• References

• 1 Ohno E, Balduzzi A, Hijioka S, De Pastena M, Marchegiani G, Kato H, Takenaka M, Haba S, Salvia R.. Association of high-risk stigmata and worrisome features with advanced neoplasia in intraductal papillary mucinous neoplasms (IPMN): A systematic review. Pancreatology 2024; 24 (01): 48-61
• 2 Ohtsuka T, Fernandez-Del Castillo C, Furukawa T, Hijioka S, Jang JY, Lennon AM, Miyasaka Y, Ohno E, Salvia R, Wolfgang CL, Wood LD.. International evidence-based Kyoto guidelines for the management of intraductal papillary mucinous neoplasm of the pancreas. Pancreatology 2024; 24 (02): 255-270