629 Views
SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Yanjie Zhu, PhD
Research Fellow
Beth Israel Deaconess Medical Center and Harvard Medical School
Ahmed Fahmy, PhD, MSc
Research Fellow
Beth Israel Deaconess Medical Center and Harvard Medical School
Chong Duan, PhD
Research Fellow
Beth Israel Deaconess Medical Center and Harvard Medical School
Shiro Nakamori, MD
Research fellow
Beth Israel Deaconess Medical Center
Reza Nezafat, PhD
Associate Professor of Medicine
Department of Medicine (Cardiovascular Division) Beth Israel Deaconess Medical Center
Background: T1, T2, and ECV are often measured manually by an experienced observer in analyzing the myocardial tissue mapping data. This process is often time consuming and negatively impact reproducibility and standardization of measurement. Therefore, there is a need for automating data analysis for myocardial tissue mapping. Machine learning based technique has a potential to facilitate analyzing myocardial tissue mapping. However, it requires big datasets for training and evaluation. In addition, generalization of machine learning based automatic measurement for different tissue parameters is unknown. In this study, we sought to investigate the performance of deep fully convolutional neural network (FCN) based tissue mapping analysis which is trained using T1 mapping dataset for analyzing T2 and ECV measurements, a concept referred to as transfer learning.
Methods: Figure 1 shows the flowchart of the automated analysis platform. The FCN based on the U-Net architecture is used for myocardium segmentation. It was trained using 11550 T1-weighted images with different inversion times from native T1 mapping and transferred to T2 and ECV mapping. We prospectively acquired T2 maps in 401 patients (256 male; age 55±15 years) using slice-interleaved myocardial T2 mapping sequence [2] and ECV maps in 381 patients (250 male; age 55±15 years) using slice-interleaved myocardial T1 mapping sequence [3]. The reference values were manually analyzed by an experienced cardiologist (8 years CMR experience) using an in-house analysis tool (including image registration, curve fitting, and manual analysis) implemented in Matlab. The Pearson correlation coefficient (R) and Bland-Altman analysis were used to evaluate the performance of automatic measurements compared with the manual measurements on per-patient and per-slice basis.
Results: The automatic values showed a strong correlation with the manual values in per-patient (T2: 390 patients: R=0.891, slope=1.010; ECV: 319 patients: R=0.918, slope=0.992) and per-slice (T2: 1873 slices: R=0.825, slope=1.001; ECV: 1489 slices: R=0.859, slope=0.989) (Figure 2a). The automatic and manual values were in good agreement within 95% confidence interval of the measurements located between the limits-of-agreement lines in per-patient (T2: 0.6±5.9ms; ECV: -0.2±3.2%) and per-slice (T2: 0.2±9.4ms; ECV: -0.2±4.9%) analyses (Figure 2b). Table 1 shows the total patient, slice, and image numbers as well as the corresponding successful rates using the automated analysis for myocardial T2 and ECV mapping.
Conclusion:
The transfer learning-based automated analysis platform shows good performance for myocardial T2 and ECV mapping and has potential to replace the need for image registration and manual analysis in myocardial tissue mapping. Further studies are warranted to evaluate the performance of this technique for parametric maps acquired using different vendors, sequences and field strength.