A New Algorithm for Incidental Pancreatic Cyst Detection

ObjectivesTo develop an accurate, state-of-the-art algorithm for the incidental detection of pancreatic cystic lesions (PCLs) on computerized tomography (CT) and magnetic resonance imaging (MRI) scans. MethodsA SwinT-Unet-based architecture was developed for the incidental detection of PCLs. The algorithm was trained and validated on a robust dataset of retrospective CT and MRI studies collected from HT Medica centers located in eight different cities using scanners fabricated by four different manufacturers. ResultsOur algorithm was able to detect 91.6% of the confirmed PCLs in the initial dataset with 91.6% sensitivity and 92.3% specificity, while 91.7% of the healthy controls were also correctly identified. Furthermore, our tool was remarkably capable of classifying these PCLs as mucinous or non-mucinous, determining their location within the pancreas with an accuracy of 88.5%, and identifying the presence of calcifications or scars within the PCLs with an accuracy of 96%. ConclusionsBy integrating radiological data and state-of-the-art artificial intelligence techniques, we have developed an efficient tool for the incidental identification and initial characterization of PCLs, which present a substantial prevalence within the global population. Our algorithm facilitates early diagnosis of pancreatic abnormalities, which could have a profound impact on patient management and prognosis, particularly in the case of PCLs with malignant potential.