Introduction

SIMBA was proposed as a fast data-driven approach for the reconstruction of static motion-consistent 3D whole-heart images from untriggered and ungated free-running acquisitions¹. This approach relies on the clustering of reference datapoints, i.e. superior-inferior (SI) projections, regularly acquired during the scan. The clustering is based on physical similarities in the signals, irrespective of their timepoint of acquisition. The original implementation (SIMBAc) uses the whole extent of the SI projections of the four central elements of the body array coil, located at the center of the chest, as an input to target the heart and liver². However, in some instances SIMBAc fails at isolating precise motion states, leading to suboptimal image quality. We therefore propose a novel approach (SIMBAp) that consists of an automated subject-specific pixels selection to extract the signal components that can best promote motion consistency for SIMBA. This new algorithm was tested in a series of ferumoxytol-enhanced scans obtained from 15 patients with congenital heart disease.

Methods

Pixels selection: First, all the SI projections were concatenated in time, for each coil, obtaining a 3D matrix $$$S$$$ of size ($$$N_{pixels} \text{x} N_{segments} \text{x} N_{coils}$$$). For each pixel $$$p$$$ for each coil $$$c$$$, the power spectral density $$$s(p,c)$$$ of each timeseries $$$t(p,c)$$$ is computed. $$$s(p,c)$$$ was then split into two parts, one encompassing only low-frequency components $$$s_{low}(p,c)$$$, i.e. containing relevant physiological information, and a high-frequency one $$$s_{high}(p,c)$$$, containing mainly components from trajectory imperfections. Considering only $$$s_{low}(p,c)$$$, the low-frequency energy of each pixel is calculated as $$$E(p,c) = \sum s_{low}(p,c,f)^2$$$. Finally, the pixel selection was set to include pixels with an energy value higher than an empirically selected 15% of the maximal energy at low frequencies (Figure 2). At this point, the final matrix $$$\hat{S}$$$ has a size of ($$$M \text{x} N_{segments}$$$), with $$$M$$$ being the total number of selected pixels. Finally, $$$\hat{S}$$$ was low-pass filtered to minimize the effect of non-physiological components.
SIMBA: The main steps of SIMBA were implemented as in the original publication¹ as illustrated in Figure 1.
Data acquisition: Fifteen congenital heart disease patients (17$$$\pm$$$10 years; 56$$$\pm$$$26 kg; 13 males) were scanned on a 1.5T (MAGNETOM Sola, Siemens Healthcare, Erlangen, Germany), after injection of ferumoxytol (2mg/kg). All datasets were acquired using a prototype free-running GRE sequence¹ with a 3D radial phyllotaxis trajectory³, and the ECG recorded.
Data analysis: The low-dimensional space (Figure 1) was characterized in terms of compactness of the largest cluster, inversely related to the mean distance of each point from the centroid, the number of segments in the largest cluster, and the trajectory uniformity¹. Analysis of the cardiac data selection for SIMBAc and SIMBAp was performed by calculating the percentage of diastolic data, taking the QT interval from the ECG as the duration of systole⁴. The respiratory data selection was assessed using a self-gating respiratory signal extraction⁵.
Finally, the image quality for the two algorithms was compared in terms of blood-myocardium sharpness⁶, the visible length and sharpness of the right coronary artery (RCA), and those of the combined left main (LM) + left anterior descending coronary artery (LAD), determined using the Soap-Bubble software tool⁷.

Results

Pixels selection: SIMBAp always resulted in a smaller input matrix, with an average of 83$$$\pm$$$42 pixels vs 764 obtained with SIMBAc. In only 5 of 15 cases the selected pixels were all located in the four coil elements chosen by SIMBAc.
Physiological data selection: For 7 of 15 cases, SIMBAp resulted in a more accurate selection of end-expiration, and in 5 of these 7 cases in a more precise selection of diastolic data: SIMBAp 86.8$$$\pm$$$13.8% vs SIMBAc 67.7$$$\pm$$$32.5% (Figure 3A). This greatly improved the image quality (Figure 3B), resulting in a significantly higher blood-myocardium sharpness (P<0.05). Moreover, these instances were the ones for which the pixels selection differed for up to 75% from the standard body coil selection of SIMBAc.
Data analysis: The compactness of the largest cluster was significantly higher for SIMBAp compared to SIMBAc, while both trajectory uniformity and size of the largest cluster were not different (Figure 4A). A comparison of the mean blood-myocardium sharpness, visible length and sharpness of RCA and LM+LAD for all 15 patients didn’t show any statistically significant differences between the two methods (Figure 4B). Figure 5 shows an example reconstruction with both methods.

Discussion

In this work we developed and tested an alternative approach to the original fixed coil selection for the creation of the input data matrix for SIMBA. We demonstrated the advantages of an automated subject-specific selection driven by the frequency spectrum of single pixels from the SI projections of individual coil elements. This new approach improved the physiological data selection, resulting in sharper images where the original approach was incapable of adequately capturing all physiological components. Using SIMBAp we observed some instances where a good suppression of respiratory motion was not paralleled by an equally good suppression of cardiac motion. The algorithm’s steps should be optimized to improve the results in these cases as well. We plan to further improve this technique, also considering different dimensionality reduction and clustering methods. Moreover, we would like to test this pixel selection for the self-gating signal extraction as part of the 5D free-running framework.

Acknowledgements

No acknowledgement found.