Deep learning reconstruction in MRCP: Impact on IPMN characterization and common bile duct stone detection

Magnetic Resonance Cholangiopancreatography (MRCP) is the gold standard for evaluating ductal pathologies, particularly Intraductal Papillary Mucinous Neoplasms (IPMN) and common bile duct (CBD) stones. However, image quality can be limited by noise, artifacts, and acquisition constraints. Deep learning (DL) reconstruction algorithms may improve image quality and diagnostic confidence. This study aims to compare DL and non-DL reconstructions in MRCP regarding diagnostic confidence for IPMN evaluation and CBD stone detection.

This single-center retrospective study included 91 patients who underwent MRCP between March 2024 and January 2025 for IPMN assessment (51 patients) or suspected CBD stones (40 patients). Four sequence types were analyzed (2D and 3D, with and without DL reconstruction). Image quality was objectively assessed (SNR, CR, CNR). Subjective criteria (overall image quality, artifacts, noise, contrast) and diagnostic confidence scores for specific IPMN and CBD stone features were independently evaluated by two radiologists.

DL reconstruction significantly improved SNR, contrast, and CNR for the CBD, particularly in 2D sequences. For the pancreatic duct, improvement was more limited and variable. Subjectively, DL images received higher ratings, notably for contrast, with significant noise reduction and better overall quality. Diagnostic confidence was enhanced for IPMN typing, malignancy criteria, and CBD stone detection.

Deep learning reconstruction improves MRCP image quality and reader diagnostic confidence for evaluating IPMNs and CBD stones. While prospective studies with procedural standards are needed to evaluate direct impacts on diagnostic accuracy and patient outcomes, these findings support integrating DL technology into routine clinical protocols.