Deep learning-based respiratory-triggered magnetic resonance cholangiopancreatography: comparison with conventional respiratory-triggered and breath-hold magnetic resonance cholangiopancreatography.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, Chung-Ang University Hospital, Seoul, Republic of Korea.
- College of Medicine, Chung-Ang University, Seoul, Republic of Korea.
- Department of Radiology, Chung-Ang University Hospital, Seoul, Republic of Korea. [email protected].
- Research & Clinical Translation, Magnetic Resonance, Siemens Healthineers AG, Erlangen, Germany.
- Department of MR Engineering, Siemens Healthineers AG, Seoul, Republic of Korea.
Abstract
To evaluate whether deep learning-based respiratory-triggered (DL) 3D magnetic resonance cholangiopancreatography (MRCP) improves acquisition efficiency while providing similar qualitative image assessment results to conventional respiratory-triggered (CON) and breath-hold compressed sensing (BH-CS) MRCP. This retrospective study included 116 patients who underwent DL, CON, and BH-CS MRCP. Relative contrast (RC) of the common bile duct (CBD) was quantitatively analyzed. Two reviewers qualitatively assessed ductal visibility (10 segments), background suppression, motion artifacts, overall image quality, and conspicuity of pancreatic cystic lesions. Acquisition efficiency of respiratory-triggered sequences was compared based on the number of effective respiratory cycles. Interobserver agreement was substantial to almost perfect (κ = 0.73-0.94). RC was significantly higher with DL (0.87 ± 0.07) and BH-CS (0.91 ± 0.08) compared with CON (0.83 ± 0.08; p < 0.01). Visibility of the CBD and common hepatic duct did not differ significantly among sequences, whereas DL and CON showed significantly higher visibility of peripheral intrahepatic and pancreatic duct segments than BH-CS (all p < 0.01). Overall image quality scores were significantly higher for DL and CON than for BH-CS for both readers (mean scores, 3.4-3.5 vs. 2.8). Pancreatic cystic lesions ≥ 1 cm were detected in 14.7% of examinations with no significant difference in detection rates among sequences. For lesion conspicuity, reader 1 found no significant between-sequence differences, whereas reader 2 reported higher conspicuity scores for BH-CS than for DL (adjusted P = 0.01). DL required fewer effective respiratory cycles than CON (19 vs. 60 cycles), corresponding to an approximately 68% reduction in respiratory-triggered acquisition burden. DL-based respiratory-triggered 3D MRCP demonstrated improved acquisition efficiency with fewer effective respiratory cycles than CON. No significant differences were detected between DL and CON in qualitative image assessment, while both respiratory-triggered techniques showed higher peripheral duct visibility than BH-CS.