A deep learning pipeline for liver macromolecular proton fraction quantification without subject-specific B1 acquisition using spin-lock MRI.
Authors
Affiliations (5)
Affiliations (5)
- Department of Imaging and Interventional Radiology, The Chinese University of Hong Kong, Hong Kong, China.
- CUHK Lab of AI in Radiology (CLAIR), The Chinese University of Hong Kong, Hong Kong, China.
- Department of Medicine and Therapeutics, State Key Laboratory of Digestive Disease, The Chinese University of Hong Kong, Hong Kong, China.
- Illuminatio Medical Technology Limited, Hong Kong, China.
- Philips Healthcare, Hong Kong, China.
Abstract
Macromolecular proton fraction (MPF) is a promising noninvasive biomarker for staging liver fibrosis. However, current post-processing requires B1-inhomogeneity acquisition and manual region of interest (ROI) selection, introducing subjectivity and variability. In this study, we propose a deep learning pipeline that enables liver MPF quantification without subject-specific B1 acquisition by leveraging an atlas-derived B1 map constructed from measured B1 data. This retrospective study used data collected at one institution from April 2019 to October 2019. The pipeline contains three models: a segmentation network for obtaining liver masks, a registration network for aligning the atlas B1 map with liver masks, and a quantification network for MPF quantification. An uncertainty-guided strategy is proposed to automatically select ROIs for assessing liver MPF. The accuracy of MPF quantification was evaluated using mean absolute error (MAE), structural similarity index measure (SSIM), and peak signal-to-noise ratio (PSNR). Reproducibility of liver MPF measurement was assessed by comparing manually selected ROIs from two experts with automated ROIs, employing intraclass correlation coefficient (ICC) and Bland-Altman analysis. The study included 44 patients (mean age, 59.4±9.7 years; 20 male patients, 24 female patients). MAE for MPF quantification within the whole liver and ROI are 0.49%±0.31% and 0.45%±0.26%, respectively. ICC of liver MPF assessments are 0.931 [95% confidence interval (CI): 0.887, 0.963] between expert 1 analyst and automated analysis, 0.949 (95% CI: 0.924, 0.971) between the expert 2 analyst and automated analysis, and 0.938 (95% CI: 0.903, 0.959) between expert 1 and expert 2 analyst. Mean bias [95% limits of agreement (LOA)] were 0.042% (-0.514%, 0.598%), -0.029% (-0.445%, 0.388%), and 0.033% (-0.497%, 0.563%) for expert 1 <i>vs</i>. automated analysis, expert 2 <i>vs</i>. automated analysis, and expert 1 <i>vs</i>. expert 2, respectively. The proposed deep learning pipeline enables liver MPF quantification without subject-specific B1 acquisition by employing an atlas-based B1 substitution strategy, while maintaining high reproducibility in a fully automated manner.