Early detection of pancreatic ductal adenocarcinoma (PDAC) is vital to improve survival. This study introduces a deep learning tool, pancreatic cancer detection with artificial intelligence (PANDA), capable of accurately detecting and classifying pancreatic lesions via non-contrast computed tomography (CT). The promising results suggest that PANDA could emerge as a valuable tool for large-scale pancreatic cancer screening.
Pancreatic ductal adenocarcinoma (PDAC) is the deadliest solid malignancy worldwide, causing approximately 466,000 deaths per year. Regrettably, this disease is also typically detected late and at an inoperable stage. Although early or incidental detection is associated with prolonged survival, screening asymptomatic individuals for PDAC using a single test remains unfeasible due to the low prevalence and potential harm of false positives.
Non-contrast computed tomography (CT), routinely performed for clinical indications, offers the potential for large-scale screening. However, identification of PDAC using non-contrast CT has long been considered impossible. This study developed a deep learning approach, pancreatic cancer detection with artificial intelligence (PANDA), that can detect and classify pancreatic lesions with high accuracy via non-contrast CT.
In this multicentre retrospective study, patient datasets from ten institutions were retrospectively collected. The study comprised five cohorts: an internal training cohort, on which the AI models were built; an internal test cohort, on which the model performance and reader study were assessed; an external multicentre test cohort, on which the generalisation across multiple centres was assessed; a chest non-contrast CT test cohort, on which the generalisation to chest CT scans was assessed; and a real-world clinical evaluation cohort, on which critical questions about the clinical translation were assessed. PANDA, the AI model, was developed in three stages for pancreas localisation, lesion detection, and differential diagnosis, including PDAC, pancreatic neuroendocrine tumour (PNET), solid pseudopapillary tumour (SPT), intraductal papillary mucinous neoplasm (IPMN), mucinous cystic neoplasm (MCN), chronic pancreatitis, serous cystic neoplasm (SCN) and “other”.
PANDA was trained on a robust dataset comprising 3,208 patients from a single centre. In a multicentre validation involving 6,239 patients across 10 centres, PANDA achieved an area under the receiver operating characteristic curve (AUC) of 0.986–0.996 for lesion detection. When compared to the mean radiologist performance, PANDA exhibited a remarkable 34.1% improvement in sensitivity and a 6.3% enhancement in specificity for PDAC identification. Additionally, in a real-world multi-scenario validation with 20,530 consecutive patients, PANDA achieved a noteworthy sensitivity of 92.9% and a specificity of 99.9% for lesion detection. Importantly, when utilised with non-contrast CT, PANDA demonstrated non-inferiority to radiology reports based on contrast-enhanced CT in differentiating common pancreatic lesion subtypes.
These compelling results suggest that PANDA could emerge as a valuable tool for large-scale pancreatic cancer screening.
Reference
Top
