Does a deep learning-assisted reconstruction algorithm improve the precision and reproducibility of accelerated single-breath-hold pancreatic PDFF mapping compared to conventional compressed sensing in adults with obesity?
Accelerated chemical shift encoding-based MRI with deep learning-assisted reconstruction provides precise and reproducible single-breath-hold pancreatic fat quantification.
AbstractPurpose With the rising prevalence of obesity and metabolic syndrome, there is an increasing need for noninvasive quantification of pancreatic fat as a marker of metabolic risk. Chemical shift encoding (CSE)-based water–fat separation enables pancreatic proton density fat fraction (PDFF) mapping. This study evaluates techniques for accelerating high-resolution, single-breath-hold PDFF mapping using sparse sampling with compressed sensing with sensitivity encoding (C-SENSE) and a deep learning (DL)-assisted reconstruction algorithm, focusing on reproducibility, precision, and clinical applicability. Methods 104 abdominal MRI datasets were obtained from 71 adults (58 % female; age 18–65 years; body mass index (BMI) 30.0–39.9 kg/m2; without diabetes) enrolled in a lifestyle intervention trial. Imaging was performed at 3 T (Ingenia Elition X, Philips) using two six-echo gradient-echo acquisitions (2 × 2 × 3 mm3, identical TR/TE/echo spacing). Acceleration factors of R=6 (16.9 s) and R=10 (10.3 s) were reconstructed using vendor compressed sensing (C-SENSE6, C-SENSE10); the DL-assisted reconstruction (C-SENSE AI10) was applied only to R=10 to evaluate denoising of higher-acceleration data. PDFF maps were analyzed using three regional regions of interest (ROIs) (head, body, tail) and whole-pancreas segmentation. Results A Mean pancreatic PDFF measured with C-SENSE6 was 15.0 10.9 – 23.0 % at baseline (V1) and 8.2 7.1 – 11.4 % after one year (V3). Across all reconstructions, PDFF ranged 3.5 – 47.6 %. Strong linearity was observed between C-SENSE10 and C-SENSE AI10 compared with C-SENSE6 (R2 ≥ 0.99). Whole-pancreas analysis showed high reproducibility (intraclass correlation coefficient = 0.87 – 1.00 across methods). The DL-assisted reconstruction reduced map noise compared with conventional C-SENSE10 without affecting PDFF accuracy. Conclusion Accelerated CSE-based pancreatic PDFF mapping enables precise, reproducible, and clinically feasible single-breath-hold fat quantification. The approach provides a robust tool for evaluating pancreatic steatosis in obesity and metabolic disease research.
Rupp et al. (Fri,) studied this question.