Magnetic Resonance Imaging, Cine

ECV imaging was performed between Sept 2009 and May 2011, 952 patients referred for CMR assessment of known or suspected heart disease prospectively underwent ECV imaging. Hematocrit was measured from a venous blood sample drawn just prior to the CMR study. This study was approved by the local Institutional Review Boards of the National Heart, Lung, and Blood Institute and Suburban Hospital, and all subjects gave written informed consent to participate. A total of 1680 ECV maps were generated by automated off-line processing as described above (728 studies imaged matching short and long axis views both pre- and post-contrast, and the 224 remaining studies only acquired a single slice). Scoring of ECV map quality was performed on 600 maps (338 subjects) acquired consecutively, and statistics for automatic blood measurements were made based on the complete set. Additionally, cine MRI of cardiac function and phase sensitive inversion recovery (PSIR) late Gadolinium enhancement imaging [22 (link)] was performed on all subjects.

The Henry Ford Health Institutional Review Board reviewed and approved the study protocol (#IRB13570). Eighty-four patients had ^Tc99mPYP scan done in the selected time period. We reviewed and collected patients’ demographics including age, sex and race; medical histories including history of heart failure, atrial fibrillation, cerebrovascular accidents, carpal tunnel syndrome, spinal stenosis and peripheral neuropathy; admissions for heart failure exacerbation and arrhythmia, biomarkers (Troponin and B-Type natriuretic peptide); electrocardiogram and echocardiography results including ejection fraction, diastolic dysfunction grade, diastolic septal (IVSd) and posterior wall thickness (PWd), diagnostic testing including serum protein electrophoresis, urine protein electrophoresis, free light chains, cardiac magnetic resonance imaging, the ^Tc99m pyrophosphate (PYP) scan planar uptake ratio and tissue biopsies. Standard cine MRI images, delayed gadolinium enhancements and look-locker sequence were used in the imaging analysis. Biopsy specimens were analyzed by the Department of Pathology using Congo red staining and mass spectrometry.

The experimental datasets used in this study are 2D multi-slice cardiac cine MRI sequences collected following standard CMR protocols^{87 (link)}, where each dataset showing a single cardiac cycle divided into 25 cardiac phases. For all the datasets, the first time frame corresponds to diastole. Thus, each dataset starts at diastole, then continues to systole, and finally comes back to diastole. The in-plane resolution of each dataset is 1.77 × 1.77 mm ${}^2$ with slice thickness of 8 mm and no gap between slices. The information for each dataset about the number of slices covering from the base to the apex of the ventricles, the slice locations, and the image size of each slice is given in Table 10.Table 10

Description of the 2D multi-slice cardiac cine MRI datasets used in the experimental validation.

Dataset	Image Size (pixels)	Number of Slices	Slice locations
			Basal	Mid	Apical
1	204 × 243	8	{1,2,3}	{4,5,6}	{7,8}
2	188 × 192	9	{1,2,3}	{4,5,6}	{7,8,9}
3	211 × 297	9	{1,2,3}	{4,5,6}	{7,8,9}
4	284 × 294	8	{1,2,3}	{4,5,6}	{7,8}
5	243 × 268	9	{1,2,3}	{4,5,6}	{7,8,9}
6	331 × 322	8	{1,2,3}	{4,5,6}	{7,8}
7	293 × 328	9	{1,2,3}	{4,5,6}	{7,8,9}
8	284 × 291	8	{1,2,3}	{4,5,6}	{7,8}

For each dataset, the image size of a single slice, number of slices in that dataset, and the slice locations relative to basal, mid-ventricular, and apical sections are provided.

The analysis of the cine MRI datasets were done offline via using the freely available software, Segment version 3.3 R9405b^{88 (link)}. The LV and RV segmentations were performed by using the automatic segmentation algorithm in the software^{89 (link)}. The ground truth points were extracted from the data via the feature tracking algorithm in the strain analysis module of the software^{90 (link)}. Figure 10 shows the images and corresponding segmentations obtained via the Segment software for the basal, mid-ventricular, and apical slices.
For the real MRI datasets, 5th order one-step estimators were used in the motion update step. The particle filter was initialized with 1000 particles per time step. A single MRI slice was used per time step for the tracking algorithm.
Two sets of experiments were performed for the real cardiac MRI datasets. The first experiment was same as the numerical phantom. Separate trials were performed for the basal, mid-cavity, and apical slice planes. In each trial, once the 2D image slice plane was determined at the first time step, the same image was used in the remaining time steps.Figure 10

Shows the segmentations obtained via Segment software respectively for the (a) basal (b) mid-ventricular (c) apical slices.

In the second experiment, the algorithm was evaluated with changing slice planes at each time step; i.e., the slice plane was allowed to be different than the one in the previous time step. Once the initial 2D image slice plane was selected, it was not necessarily needed to be the same one in the remaining time steps. In a typical application, this varying single image slice could be either manually selected from a stack of slices by a clinician or automatically by an active sensing algorithm^{91 (link)}. Here, the set of slice planes are selected randomly between the basal and mid-ventricular slices (Fig. 1). At each time step, the slice plane variable $v_p$ ; indicating which slice to be selected from a given stack of slices, is generated randomly from a discrete uniform distribution; $v_v\sim \mathcal{U}(v_b,v_m)$ , where $v_b=2$ is the second basal slice and $v_m=7$ is the first apical slice.