Free‐Breathing Dynamic, Regularized, Adaptive Cluster Optimization (DRACO) Cine Cardiac MRI in Atrial Fibrillation

ABSTRACT

Background

Existing strategies for free‐breathing cine cardiac MR, including real‐time imaging, struggle with complex arrhythmias such as atrial fibrillation (AF). Dynamic regularized adaptive clustering optimization (DRACO) has been proposed to generate high‐quality and quantifiable breath‐held cine images for patients in AF.

Purpose

To tailor DRACO for free‐breathing and evaluate its potential to simultaneously handle cardiac and respiratory motion.

Study Type

Prospective.

Subjects

10 sinus rhythm (10 males), 20 AF (18 males).

Field Strength/Sequence

3.0 T, balanced steady‐state free precession (bSSFP) with sorted golden‐step, ECG‐gated segmented cine, and real‐time.

Assessment

Patients were imaged using the golden‐step sequence under breath‐held and free‐breathing conditions, and the reference bSSFP cine sequence. Signal‐to‐noise ratio (SNR), contrast‐to‐noise ratio (CNR), edge sharpness, and image quality (4‐point Likert scale) were assessed.

Statistical Tests

Paired t‐test, Friedman test, regression analysis, Fleiss' Kappa. Significance level p < 0.05.

Results

The SNR and CNR values of DRACO images showed no significant differences between breath‐held and free‐breathing conditions in both sinus rhythm (SNR p = 0.53; CNR p = 0.87) and in AF (SNR p = 0.56; CNR p = 0.55). Image sharpness was not significantly different between breath‐held and free‐breathing in both sinus rhythm (p = 0.24) and AF (p = 0.79). In AF patients, breath‐held DRACO and free‐breathing DRACO had significantly higher image quality scores than real‐time cine. In sinus rhythm patients, there was no significant difference (p = 0.13) between breath‐held or free‐breathing DRACO and segmented cine. Despite significant within‐patient ejection fraction (EF) differences between breath‐held versus free‐breathing in 80% of sinus rhythm and 70% of AF patients, the average EF between breath‐held versus free‐breathing DRACO showed high correlation (sinus rhythm R ² = 0.71; AF R ² = 0.71), demonstrating DRACO's robustness to motion variation.

Data Conclusion

DRACO is robust to respiratory motion and AF‐related irregular cardiac motion. It enables high‐quality cine image generation in both regular and irregular cardiac motion, with or without breath holding.

Level of Evidence

1.

Technical Efficacy

Stage 2.

Plain Language Summary

Cine cardiac MRI is a technique that captures the motion of a beating heart and provides information about structure and function. Most existing methods can handle motion that occurs at repeatable intervals; however, they are less effective when the interval duration becomes irregular or unpredictable. Dynamic, Regularized, Adaptive Cluster Optimization (DRACO) was developed to capture irregular cardiac motion in the absence of respiratory motion. We demonstrate DRACO's capability for simultaneously capturing irregular cardiac motion affected by respiratory motion. We showed that DRACO can produce high‐quality cine images in cases of regular and irregular heart motion, with or without respiratory motion.

Abbreviations

AF

atrial fibrillation

BPM

beat per minute

bSSFP

balanced steady‐state free precession

CNR

contrast‐to‐noise ratio

DRACO

dynamic regularized adaptive cluster optimization

HIPAA

Health Insurance Portability and Accountability Act

IQR

interquartile range

MR

magnetic resonance

PCA

principal component analysis

ROI

regions of interest

SD

standard deviation

SI

signal intensity

SNR

Signal‐to‐noise ratio

1. Introduction

Cine cardiac magnetic resonance (MR) is the reference standard for quantification of cardiac function and morphology. Data acquisition is usually performed under ECG‐gating with breath holding, then binned into phases based on the ECG signal to generate multi‐phase cine images. However, breath holding is not always feasible in patients with limited breath holding capacity. Free‐breathing acquisition, therefore, is one solution that has the added benefit of enhancing diagnosis in conditions such as pericardial constriction by revealing pathologic interactions between the cardiac and respiratory systems [1, 2].

Existing methods to deal with respiratory motion during free breathing include diaphragmatic gating and respiratory self‐gating [3]. Diaphragmatic gating relies on the heart–diaphragm motion model and only samples data during end‐expiration, which is inferred from the position of the diaphragm. This will increase scan time, and the respiratory motion within a heartbeat is not always well resolved because the navigator is only acquired once or twice per heartbeat [4]. Self‐gating methods use acquired imaging data to resolve motion and include principal component analysis (PCA) motion binning [5, 6, 7, 8, 9], low‐resolution image‐based motion correction [10, 11, 12, 13], and k‐means clustering for motion binning [14]. For PCA‐based motion binning, frequency filters are used explicitly by directly limiting frequency ranges or implicitly by choosing the window size to associate principal components with either cardiac or respiratory motion [5, 6, 7, 8, 9]. Data binning is then performed based on the component signals prior to reconstruction. Image‐based motion correction uses navigator data to generate low‐resolution images to estimate motion displacement, which is then corrected during reconstruction of the high‐resolution images. For motion binning using k‐means, the similarity of the motion signal is used to generate respiratory bins with k‐means clustering, and the data are then divided into the corresponding bins.

In complex arrhythmias such as atrial fibrillation (AF), these self‐gating strategies may work less well under free breathing. Because AF has unpredictable cardiac motion frequencies, filters in PCA‐based strategies may fail to differentiate between types of motion. Low‐resolution image‐based approaches require additional image registration, and their inherent temporal resolution may be insufficient to capture abrupt cardiac motion. Although k‐means clustering has been used for motion binning in several free‐running frameworks, k‐means may generate abrupt transitions between motion states during complex arrhythmias and has not been demonstrated under free breathing [15].

Dynamic regularized adaptive cluster optimization (DRACO) CMR was recently introduced as an effective approach for generating high‐quality cine images that enable quantification of cardiac function during complex arrhythmias, including AF [16]. The DRACO approach uses adaptive clustering specifically tailored for complex arrhythmias and makes no prior assumptions about the motion type or pattern, except for the continuity of motion. Given the data‐driven nature of DRACO, we expect the framework to generalize to the combination of cardiac and respiratory motion and to resolve both types of motion simultaneously. In this work, we aim to validate the DRACO framework using free‐breathing data from patients in sinus rhythm and in patients with persistent AF. The objectives were to assess the capability of DRACO to simultaneously resolve irregular cardiac motion and respiration motion and to demonstrate its effectiveness at capturing beat‐to‐beat variation in the setting of free‐breathing image acquisition.

2. Materials and Methods

2.1. Data Acquisition

The study was approved by our local Institutional Review Board. All subjects provided written informed consent before participation, and their data were handled in compliance with the Health Insurance Portability and Accountability Act (HIPAA). A total of 10 patients with sinus rhythm and 20 patients with persistent AF were enrolled. A 3.0 T scanner (Skyra; Siemens Healthcare, Erlangen, Germany) with an 18‐channel body coil was used. Base, mid, and apical ventricular short‐axis slices were acquired using a free‐running sorted Cartesian golden‐step sequence [15] under breath‐held and free‐breathing conditions. For each slice, both the breath‐hold and free‐breathing acquisitions lasted for the same amount of time (~18 s). As a reference, the same slices were scanned under breath holding using a commercially available ECG‐gated segmented cine sequence for sinus rhythm and a real‐time sequence for AF. Representative sequence parameters are in Table S1.

Details of the golden step sequence with additional navigators can be found in [15]. Briefly, Figure 1A shows the golden step sequence using balanced steady‐state free precession (bSSFP) readouts [17]. The blue points represent acquired image data without ECG gating. Additional k‐space center lines represented by the red points were added after every four TR to obtain navigator data for resolving motion. The main scan parameters were the following: FOV = 380 × 285 mm², spatial resolution = 1.8 × 1.8 mm², slice thickness = 8 mm, flip angle = 45°–55°, TR/TE = 3.4/1.7 ms.

FIGURE 1.

DRACO pipeline for free‐breathing cardiac cine in atrial fibrillation. (A) Data acquisition. The sorted golden step sequence with navigator lines is used to acquire image data and navigator data. The navigator data with T time frames serve as inputs into a clustering‐based motion resolving algorithm. (B) Motion identification. The DRACO algorithm described by Equation (1) uses the similarity of navigator data, constrains the generated K clusters to be temporally continuous and cluster‐spatially continuous, and then generates probability‐like cluster matrix M with dimension K * T to represent motion. (C) Image reconstruction. The k‐space data X is used in the image reconstruction step to generate images of K clusters (motion states). In the Equation (2) reconstruction, the cluster matrix M is multiplied by the data fidelity term and the total variation term to consider the effect of motion. DRACO, dynamic regularized adaptive clustering optimization.

2.2. Data Processing and Image Reconstruction

The DRACO pipeline has been shown to be effective at handling cardiac motion typical of AF under breath holding conditions [16]. Figure 1 summarizes the DRACO framework. Briefly, the DRACO approach resolves motion through data‐driven clustering of the navigator lines to capture distinct motion states. It computes a K × T cluster matrix M for K motion state clusters and T time frames; its elements M _kt represent the probability that the data at the tth time frame belongs to the kth cluster. The cluster probability matrix M and the cluster centroid matrix C are estimated according to

$$M,C=\underset{M,C}{\arg }\min {\sum }_{t=1}^T{\sum }_{k=1}^K{M}_{\mathit{kt}}{‖y_t-c_k‖}^2+{\lambda }_M{R}_M\left(M\right)+{\lambda }_c{R}_c\left(C\right)$$ (1)

where y _t is the navigator data at time frame t; c _k is the centroid of cluster k; R _M is a regularization function to enforce temporal continuity of M; and R _C is a regularization function to enforce continuity across cluster orders in C. Next, DRACO reconstructs images for all clusters by solving

$$I_K=\underset{I_K}{\arg }\min {‖E\left(I_{K}M\right)-X‖}^2+\mathit{\lambda TV}\left(I_{K}M\right)$$ (2)

where E is the overall encoding operator incorporating sampling pattern, Fourier transform, and coil sensitivities, X is the k‐space data, and I _K is the collection of reconstructed images corresponding to the K clusters. The temporal total variation term TV(I _K M) ensures that the temporal continuity of the corresponding cine images is mapped to their chronological ordering. The final images are represented in chronological ordering as I _K M, where M assigns the images of the K clusters (motion states) to T time frames.

In the framework, the continuity of motion was assumed without specifying any specific pattern of motion. In the setting of AF with free‐breathing, both irregular cardiac and respiratory motion occur simultaneously. Respiratory motion can vary substantially, and the breath‐to‐breath variation has been described as both correlated and uncorrelated random variations as well as nonrandom variations (i.e., sometimes shallow and/or deep, or intermittently periodic and/or sporadic) [18]. The breath‐to‐breath variation can be likened to that of beat‐to‐beat variation in AF (i.e., its non‐periodic and irregularly irregular pattern). We hypothesize that our motion continuity assumption remains relevant under free‐breathing conditions, and the same pipeline could be used with or without respiratory variation because the irregular attributes of both types of motion could be generalized and handled by the DRACO approach with additional hyperparameter optimization.

In this study, both breath‐held and free‐breathing data acquired using the golden step sequence were processed using the same workflow as shown in Figure 1. The cluster‐based motion identification algorithm (Equation 1) used the navigator data as its input and output the cluster matrix M. To resolve motion, M was used after cluster‐based reconstruction (Equation 2) to generate images of K clusters (motion states). The only difference for free‐breathing was the values of the hyperparameter λ, which dictates the regularization strength of the temporal regularization in the cine reconstruction of Equation (2). This was motivated by the observation that the motion under breath holding was predominantly cardiac motion and warranted suppression of lower oscillations (or a smaller λ); whereas the motion during free‐breathing was a mixture of cardiac and respiratory motion. The λ _M and λ _C were tuned to make the scale of λ _M R _M (M) and λ _C R _C (C) one order lower than the scale of ${\sum }_{t=1}^T{\sum }_{k=1}^K{M}_{\mathit{kt}}{‖y_t-c_k‖}^2$. The M _kt value of the cluster matrix ranged from 0 to 1 and the input data y _t was normalized for each time frame. When the scales of λ _M and λ _C were varied and tested, the final results were not sensitive to the specific values. Based on these observations, the same λ _M and λ _C were used across all the experiments. Where appropriate, a larger λ was used for additional respiratory motion during free‐breathing.

In sinus rhythm datasets, λ was 0.0003 for breath‐held data and 0.0005 for free‐breathing data, and in the AF datasets, λ was 0.0005 for breath‐held data and 0.0008 for freebreathing data. K was set to be 25. Based on [16], 25 clusters were sufficient to describe the motion states in the 18‐s acquisition. Sample illustration of results from K = 25 and K = 35 can be found in Figure S1 of the Supplement Material.

2.3. Image Quality Assessment and Evaluation of In‐Slice Beat‐to‐Beat EF

The quality of images derived from DRACO were evaluated using signal‐to‐noise ratio (SNR), contrast‐to‐noise ratio (CNR), and edge sharpness. To test DRACO's robustness to respiratory motion in sinus rhythm and AF, these image quality metrics were compared between breath‐held and free‐breathing conditions. SNR and CNR were calculated based on regions of interest (ROI) manually drawn in the septum, ventricle blood pool, and air during end‐diastole and end‐systole. Mean signal intensity (SI) and standard deviation (SD) in these ROIs were used for computation. SNR was computed as: $\mathrm{SNR}=\frac{\text{mean}\left({\mathrm{SI}}_{\text{tissue}}\right)}{\mathrm{SD}\left({\mathrm{SI}}_{\mathrm{air}}\right)}$. Blood‐myocardium CNR was computed as: $\mathrm{CNR}=\frac{\text{mean}\left({\mathrm{SI}}_{\text{blood}}\right)‐\text{mean}\left({\mathrm{SI}}_{\text{myocardium}}\right)}{\mathrm{SD}\left({\mathrm{SI}}_{\mathrm{air}}\right)}$.

Edge sharpness was calculated along a straight line passing through the left ventricle, the interventricular septum, and the right ventricle of the same end‐systolic and end‐diastolic phase images. After plotting the SI profile of the straight line, the slope between the 25% and 75% of maximum intensity points of the left‐ventricle‐to‐septum edge and the right‐ventricle‐to‐septum edge in the profile was recorded. The average slope of the two edges at end‐systole and end‐diastole was calculated to represent edge sharpness. The slope was normalized by the maximum intensity of the SI profile to reduce the effect of intensity variation. The unit measure of slope is pixel⁻¹. A higher average slope reflects a sharper septal edge.

Three readers with 20 (J. Paul Finn), 12 (Kim‐Lien Nguyen), and 4 (Arutyun Pogosyan) years of experience in cardiac MRI independently scored the randomized images using a 4‐point Likert scale (Table S2). Because DRACO can be used to generate beat‐to‐beat pseudo real‐time cine images based on the temporal cluster matrix M, we analyzed in‐slice left ventricle EF to depict the beat‐to‐beat variation that can be used to render a time resolved EF. In the EF analysis, left ventricular area from each short axis slice was segmented and quantified by a reader with 5 (Zhengyang Ming) years of experience in cardiac MRI. The areas within the ROIs were temporally reordered based on cluster‐time associations to generate a time‐dependent curve of the left ventricular area (Figure 2). The peaks and valleys of this curve were determined by peak and valley detection, with peaks defined as diastole and valleys as systole. The EF of every heartbeat in each slice was calculated using values at peaks and valleys. Average EF values under breath held and free‐breathing were compared to assess the effect of respiration for each patient. The relationship between average EF under breath held and free breathing was analyzed by linear regression.

FIGURE 2.

Beat‐to‐beat in‐slice LV EF analysis based on DRACO pipeline. (A) LV segmentation is performed on images of every motion state (cluster). (B) The cluster temporal order is calculated by picking the clusters having the largest cluster membership values at each time frame of the motion‐resolved cluster matrix. (C) Permutation is performed on the LV areas of each cluster based on the cluster temporal order to generate a temporal curve of the LV areas. (D) Beat‐to‐beat EF is calculated by detecting peaks (diastolic LV area) and valleys (systolic LV area) in the temporal curve. EF, ejection fraction; DRACO, dynamic regularized adaptive clustering optimization; LV, left ventricle.

2.4. Statistical Analysis

Statistical analysis was performed in MATLAB version 2021a (MathWorks, Natick, MA, USA). Variables were summarized as mean ± SD or median and IQR. The Kolmogorov–Smirnov test was used to test normality of the data. A two‐tailed paired sample t‐test was performed for group comparisons of SNR, CNR, and edge sharpness. A two‐tailed unpaired t‐test was performed for comparisons between average EF for breath‐held and free‐breathing for each patient. The Friedman test was performed for multiple image quality score comparison (breath‐held DRACO, free‐breathing DRACO, and reference method) with subsequent post hoc group analysis using Tukey's test. Inter‐reader image scoring agreement between the three readers was calculated using Fleiss' Kappa. Fleiss' Kappa value was interpreted as agreement level: ≤ 0 no, 0.01–0.20 slight, 0.21–0.40 fair, 0.41–0.60 moderate, 0.61–0.80 substantial, and 0.81–1.00 almost perfect. p < 0.05 was considered significant.

3. Results

Data belonging to 10 subjects with sinus rhythm (62 ± 17 years, n = 10 males) and 20 subjects with AF (73 ± 10 years, n = 18 males) were collected. All AF patients had persistent AF (AF burden = 100%) and were in AF during the scans. The average heart rate of subjects in sinus rhythm was 64 ± 11 beats per minute (BPM), and the average heart rate of subjects in AF was 69 ± 15 BPM.

3.1. Illustrative Examples of Image Quality During Breath Holding and Free Breathing Acquisitions

Figure 3 provides examples of breath‐held and free‐breathing images from patients in sinus rhythm and AF. Time curves of 1D profile across the septum are shown for each slice under breath‐held and free‐breathing. The upper row is from the sinus rhythm patient and the lower row is from the AF patient. During breath holding, a periodic pattern is seen on the temporal 1D profile during sinus rhythm while an irregular pattern is shown for the AF. During free breathing, the time curve for sinus rhythm still has a relatively periodic pattern, but the time curve for AF has an irregular pattern. When comparing the time curve under breath holding versus free breathing for the AF patient, the in‐slice breath‐held EF and free‐breathing EF show differences.

FIGURE 3.

Temporal curves of 1D profile across the ventricular septum and select phases from DRACO breath‐held and free‐breathing acquisition. The images are from the mid ventricular slice of a patient in sinus rhythm (upper row, HR = 65 BPM) and a patient in atrial fibrillation (lower row, average HR = 42 BPM). A periodic motion pattern can be observed on the temporal curve in sinus rhythm; whereas, an irregular pattern is observed in atrial fibrillation. DRACO can generate high‐quality images for both patients regardless of the pattern of cardiac motion and the presence (or absence) of respiration motion. DRACO, dynamic regularized adaptive clustering optimization. BPM, beat per minute; HR, heart rate.

Figure 4 shows the performance of DRACO under breath holding and free breathing conditions across a range of heart rates for AF. Diastolic and systolic images of three patients with different heart rates are included. For heart rates ranging from 42 to 105 BPM, DRACO can generate high‐quality images for patients in AF under breath‐holding and free breathing conditions. All the images of the three patients received image scores ≥ 3. Multi‐phase videos of these images can be found in Video S1.

FIGURE 4.

Comparison of breath‐held and free‐breathing DRACO reconstruction in atrial fibrillation at several heart rates (average HR = 42,62,105 BPM). A mid left ventricular axis slice is shown in diastole and systole. All the images were scored ≥ 3 on a 4‐point Likert scale. The corresponding video is provided in Video S1. HR, heart rate, BPM, beat per minute.

3.2. Illustrations of Cluster Behavior During Respiratory Maneuvers

Figures 5 and 6 illustrate how cardiac motion and respiratory motion affect the cluster‐motion association behavior. Figure 5 belongs to a patient with sinus rhythm and Figure 6 belongs to a patient in AF. Both figures show the clustering matrix and the temporal self‐correlation map under breath hold and free breathing. In the setting of sinus rhythm with breath holding, periodic peaks are observed in the clustering matrix because of periodic cardiac motion (Figure 5). The corresponding self‐correlation map shows repetitive and highly correlated values because of periodic cardiac motion, as suggested by dense and parallel yellow lines. Under free breathing, the cluster matrix has peaks, but the periodic feature is less well preserved than that during breath holding, and the gaps between the peaks are wider. The correlation map displays more sparse but parallel yellow lines because the patient had a regular respiratory pattern, and the period of the respiratory motion is larger than the period of cardiac motion. Cine images can be found in Video S2.

FIGURE 5.

Cluster value matrix and its temporal correlation in sinus rhythm (HR = 65 BPM) under breath‐held and free‐breathing conditions. Under breath‐held, peaks in the cluster matrix show a periodic pattern and the temporal correlation shows dense and parallel diagonal lines. The gap width corresponds to the heartbeat period. Under free breathing, the peaks still have periodic patterns to a certain extent, but are less periodic than breath‐held conditions. Temporal correlation demonstrates more sparse diagonal lines because this patient had a periodic respiration pattern, and its period was longer than the heartbeat period. Videos of cine images and navigator signals can be found in Video S2. BPM, beat per minute; HR, heart rate.

FIGURE 6.

Cluster value matrix and its temporal correlation for atrial fibrillation (average HR = 80 BPM) under breath‐held and free‐breathing conditions. Under breath‐held conditions, no obvious periodic patterns can be observed in the cluster matrix, and the temporal correlation only shows a high correlation in adjacent time frames rather than a high correlation globally. Under free‐breathing, temporal correlation demonstrates only nearby correlation, and the high correlation was interrupted at 8 s by a long RR interval with minimal mechanical cardiac motion, which correlated with the simultaneously recorded ECG. Videos of cine images and navigator signals can be found in Video S3. BPM, beat per minute; ECG, electrocardiogram; HR, heart rate.

In the setting of AF, irregular cardiac motion leads to fewer detectable periodic peaks under breath holding (Figure 6). Consistently, there are few parallel lines in the correlation map. Under free breathing, there are rare sub‐diagonal lines between 8 and 10 s. Simultaneously recorded ECG showed a long cardiac cycle length (a long interval between two R waves on the ECG) during this time and the navigator data showed the heart had little motion during this beat. The corresponding cine videos are shown in Video S3.

3.3. Image Quality Comparison

Table 1 summarizes the results of qualitative and quantitative metrics. The SNR and CNR values of DRACO images acquired under breath holding and free breathing conditions were not significantly different between sinus rhythm (SNR 78.1 ± 23.5 vs. 79.8 ± 25.3, p = 0.53; CNR 82.2 ± 26.4 vs. 81.7 ± 25.0, p = 0.87) and AF (SNR 74.4 ± 28.7 vs. 73.3 ± 27.9, p = 0.56; CNR 88.4 ± 34.7 vs. 86.9 ± 34.4, p = 0.55). Image sharpness was also not significantly different between breath holding and free breathing conditions for both sinus rhythm (0.62 ± 0.32 vs. 0.57 ± 0.28 pixel⁻¹, p = 0.24) and AF (0.48 ± 0.22 vs. 0.47 ± 0.21 pixel⁻¹, p = 0.79). Figure 7 provides an illustrative comparison of SNR, CNR, image sharpness, and image quality scores.

TABLE 1.

Quantitative and qualitative comparisons among breath‐held and free‐breathing DRACO and reference methods.

Method	DRACO BH		DRACO FB		Reference
	Sinus Rhythm	Atrial Fibrillation	Sinus Rhythm	Atrial Fibrillation	Sinus rhythm (segmented cine)	Atrial fibrillation (real‐time cine)
SNR	78.1 ± 23.5	79.8 ± 25.3	74.4 ± 28.7	73.3 ± 27.9	NA	NA
CNR	82.2 ± 26.4	88.4 ± 34.7	81.7 ± 25.0	86.9 ± 34.4	NA	NA
Edge sharpness (pixel−1)	0.62 ± 0.32	0.48 ± 0.22	0.57 ± 0.28	0.47 ± 0.21	NA	NA
Image score: diastole	3.70 ± 0.51	3.49 ± 0.59	3.68 ± 0.57	3.40 ± 0.58	3.50 ± 0.69	3.02 ± 0.78 a
Image score: systole	3.41 ± 0.67	3.31 ± 0.55	3.22 ± 0.61	3.10 ± 0.54	3.25 ± 0.46	2.28 ± 0.63 a
Percent of images with quality scores ≥ 3: diastole (%)	97.8	95.6	95.3	94.6	95.7	72.2
Percent of images with quality scores ≥ 3: systole (%)	95.6	92.8	91.1	90.0	94.2	36.1

Abbreviations: AF, atrial fibrillation; BH, breath held, FB, free breath, SR, sinus rhythm.

^a Indicates statistical significance (p < 0.05) when compared to DRACO BH and FB.

FIGURE 7.

Image quality metrics based on SNR, CNR, image sharpness, and image quality scores during sinus rhythm and atrial fibrillation, with and without breath holding. The reference method was ECG‐gated segmented cine for sinus rhythm and real‐time cine for atrial fibrillation. The asterisk (*) denotes a statistically significant difference. For atrial fibrillation, the image quality scores of diastolic and systolic images were significantly higher for BH DRACO and FB DRACO relative to the reference method. AF, atrial fibrillation; BH, breath hold; CNR, contrast‐to‐noise ratio; DRACO, dynamic regularized adaptive clustering optimization; FB, free breathing; Ref: Reference; SNR, signal‐to‐noise ratio; SR, sinus rhythm.

For sinus rhythm, the Friedman test showed no significant difference (p = 0.13) among image scores for breath‐held DRACO, free breathing DRACO, and the reference segmented cine images. For AF, the Friedman test showed a significant difference among image scores for breath‐held DRACO, free‐breathing DRACO, and reference real‐time cine images. The corresponding post hoc test for AF data showed no significant difference between breath‐held and free breathing DRACO images (p = 0.64). However, there were significant differences between real‐time cine images and breath‐held DRACO images in both diastole and systole. There was moderate inter‐reader agreement for image quality of patients in sinus rhythm (Fleiss' Kappa = 0.41) and moderate agreement for images of patients in AF (k = 0.46).

For sinus rhythm, all three methods received high image quality scores on both diastolic and systolic images, and over 90% of images received scores ≥ 3. In AF, both breath held and free breathing DRACO had higher image quality scores than the reference real‐time cine method, especially on systolic images. Even though SNR, CNR, edge sharpness, and image quality score metrics for free‐breathing DRACO were slightly lower than those of breath‐held DRACO, no significant difference was found between any metric for the two methods. Based on the above statistical tests, DRACO reconstructions under breath‐held and free breathing conditions showed no significant differences in both qualitative and quantitative metrics.

3.4. In‐Slice Beat‐to‐Beat EF

Figure 8 shows the in‐slice left ventricle area temporal changes in sinus rhythm and AF during breath holding and free breathing conditions. There were significant within‐patient average EFs differences during breath holding and during free breathing in 80% of the 10 patients in sinus and in 70% of the 20 patients in AF. Figure 9 plots the average EF values derived from breath‐held and free‐breathing images. In both sinus rhythm and AF, the linear regression analysis shows good correlation between the average EF obtained with breath hold and free breathing DRACO (sinus rhythm, y = 0.78x + 3.03, R ² = 0.71; AF, y = 0.86x + 5.19, R ² = 0.71). In the patients with sinus rhythm (n = 10), the average EF of breath‐held DRACO and free‐breathing DRACO was 50.6% ± 12.4% versus 42.3% ± 11.5% (p < 0.05). In the 20 patients with AF, the EF values were 49.5% ± 11.5% versus 47.9% ± 11.8% (p = 0.30) for breath‐held versus free‐breathing DRACO, respectively.

FIGURE 8.

Temporal in‐slice LV area change in sinus rhythm (HR = 75 BPM) and atrial fibrillation (HR = 78 BPM), under breath‐held and free breathing conditions. The sinus rhythm patient shows periodic changes under both breath‐held and free‐breathing conditions, whereas for atrial fibrillation, irregular changes occur under both breath‐held and free‐breathing conditions. The effect of free‐breathing can be observed in both patients. BPM, beat per minute; HR, heart rate; LV, left ventricle.

FIGURE 9.

Average beat‐to‐beat ejection fraction (EF) under free‐breathing and breath‐held conditions for sinus rhythm and atrial fibrillation. In both sinus rhythm and atrial fibrillation, the average EF under free‐breathing shows a relatively high correlation with the average EF under breath‐held conditions. EF, ejection fraction.

4. Discussion

Current techniques [5, 6, 7, 8, 9, 10, 11, 12, 13, 14] to handle respiratory motion during free breathing acquisition of cine cardiac MR face challenges in the setting of AF, while few methods have been systematically tested in the setting of complex arrhythmias including AF. We optimized the DRACO approach and systematically demonstrated its ability for resolving simultaneous irregular cardiac and respiratory motion in the setting of AF. Moreover, our results support DRACO's effectiveness as a flexible framework for addressing simultaneous irregular cardiac and respiratory motion.

For AF, we noticed that the average real‐time cine image quality score for diastolic images was significantly higher than the average score of systolic images. We speculate that cardiac motion during the systolic phase could be more irregular during AF than that during the diastolic phase, which leads to more motion blurring and image artifacts on real‐time cine. In contrast, both breath‐held and free‐breathing DRACO performed well in AF datasets. The reliability of DRACO's performance during breath‐holding and free‐breathing acquisitions adds to its appeal by enabling the effect of respiration on image quality to be compensated through the tuning of λ in the cluster‐based reconstruction. The temporal features of AF can also be different based on heart rate. Regardless of breath‐holding or free‐breathing acquisitions, DRACO was able to capture the nuanced irregularity of the temporal features at different heart rate ranges. The cardiac motion of the patient in AF with a heart rate of 105 BPM appears closer to sinus rhythm.

The EF analysis based on DRACO demonstrates the possibility of analyzing beat‐to‐beat variation in AF and generating time‐resolved EF. Although respiration can affect the maximum and minimum left ventricle areas in both sinus rhythm and AF, the temporal curve of the left ventricular area for sinus rhythm has a higher prevalence of periodic patterns compared to AF, and this is consistently observed even during free breathing. The average EF under breath‐holding and free breathing showed a relatively high correlation in both sinus rhythm and AF despite significant within‐patient differences for breath holding and free breathing in both sinus rhythm and in AF. Through‐plane motion may be an important factor affecting in‐slice analysis, especially under free breathing. Earlier studies [19, 20, 21] comparing EF between breath‐holding and free breathing generally used ECG‐gated cine bSSFP under breath‐holding as the reference and applied parallel imaging [19] or compressed sensing [20, 21] to accelerate cine sequences under free breathing. The EF differences between breath‐held DRACO and free‐breathing DRACO were larger than the differences previously reported in references [19, 20, 21]. In the free‐breathing acquisition for references [19, 20, 21], few heartbeats and at most one respiration cycle were covered for each slice because of the acceleration. This strategy limits the effects of respiration and heartbeat variation on the EF results. Our EF analysis of DRACO included multiple heartbeats and multiple respiratory cycles in each slice. We speculate these differences represent the main reasons for the larger EF differences aside from through‐plane motion.

The option for beat‐to‐beat analysis also provides the possibility of quantifying a range of EF values that reflect the stroke volume of each heartbeat. A static, single‐beat EF is transformed by the DRACO method to capture beat‐to‐beat variation and thus becomes more reflective of the dynamic nature of cardiac function observed when patients are in AF. The dynamic representation of different heartbeats rather than a single EF value can be meaningful for patients who are in various arrhythmic states. Additional metrics such as standard deviation of the EF over a set duration may provide more information for management decisions.

Cardiac motion and respiration motion are generally considered differently and processed separately. In this work, cardiac motion and respiration motion can be resolved simultaneously by applying a data‐driven clustering DRACO algorithm that makes no assumption about the motion pattern other than continuity. DRACO's performance during the reconstruction of FB datasets affirms our confidence in DRACO's effectiveness for generalizing its ability to resolve motion. Future work includes testing DRACO's effectiveness for other types of motion, such as abdominal imaging, where both respiratory and peristalsis can be challenging, or speech imaging with lingual motion. In our current work, the image data were from the golden step sequence, and the motion signal was derived from the additional navigators. Future investigations related to the DRACO framework could be expanded to a broader family of sequences including motion encoding sequences.

A commercially available version of real‐time cine available at our institution was used as the reference for AF. A direct comparison with more advanced real‐time techniques with compressed sensing [22, 23, 24] was not performed. The faster acquisition in compressed sensing real‐time may reduce motion blurring from arrhythmias with customized k‐space sampling patterns and domain sparsity during reconstruction. Given the irregularity of arrhythmias such as AF, the acquisition must last long enough to cover multiple heartbeats rather than one single heartbeat and show the irregular motion of the arrhythmia. High temporal resolution can reduce motion blurring, but quantification would become cumbersome due to the large number of images generated by high temporal resolution for multiple heartbeats [25]. In contrast, the DRACO pipeline generates a limited number of representative images corresponding to each cluster with different motion states, and these cine images are based on the motion‐cluster association. Hence, DRACO is less time‐intensive and enables beat‐to‐beat EF as well as the overall EF to be efficiently calculated.

Other representative frameworks for cardiac and respiratory motion include MR multitasking [14], XD‐GRASP [7], 5D whole‐heart MRI [6], and nonrigid motion‐corrected 3D whole‐heart coronary vessel wall imaging [10]. MR multitasking generates respiratory bins and cardiac bins using two k‐means clustering algorithms, respectively, and produces images for every respiratory phase and every cardiac phase. In contrast, the DRACO pipeline performs a more tailored and arrhythmia‐targeted clustering algorithm to cope with cardiac motion and respiratory motion simultaneously. The reconstruction is based on adaptive clusters rather than rigid bins. XD‐GRASP and 5D whole‐heart MR separate cardiac signal and respiratory signal from motion signal using PCA and bandpass filters in specific frequency ranges. DRACO does not use bandpass filters to avoid neglecting subtle motion in AF and directly employs data similarity to find motion states. Motion‐corrected 3D whole‐heart vessel wall imaging reconstructs low‐spatial resolution iNAV images with high‐temporal resolution to estimate translational and non‐rigid motion, and then corrects the motion in the later reconstruction based on the iNAV information. In contrast, the DRACO pipeline relies on 1D projection of navigator data rather than 2D iNAV images, and resolves motion implicitly without using an explicit motion field matrix.

4.1. Limitations

First, because the DRACO pipeline is based on regularized reconstruction, the choice of the regularization hyperparameter λ affects the balance between robustness and its ability to resolve motion. Further work is needed to characterize the tradeoff between temporal continuity and motion representation. Self‐tuning of the hyperparameters is ideal for wider applications but can be time‐consuming. Second, our current 2D multislice approach does not account for slice synchronization. Although the sequential, multislice acquisition approach allows capture of a wider spectrum of motion states, and the DRACO approach yields robust motion handling of both cardiac and respiratory motion, future work could consider slice synchronization to simplify post‐processing logistics. Third, our cohort of 20 AF patients may be insufficient to capture the spectrum and complexity of AF and the mixture of irregular cardiac motion and respiratory motion. Although we studied a relatively wide range of heart rates, ranging from 42 to 105 BPM, the range of respiratory rate and respiratory patterns was limited to a typical outpatient population. In addition, our current technical development recruited consecutive patients who were male. We do not anticipate DRACO to be influenced by sex. However, arrhythmias and respiratory patterns may manifest differently in males and females, and DRACO is well‐poised to be used for investigations of sex‐specific differences in cardiorespiratory motion once it has been technically validated. Future work could consider adequately powered studies using the DRACO‐based framework to assess sex differences in clustering behavior. Fourth, to make a fair comparison between the breath holding and free breathing, and to limit the scan time in sick patients, free breathing data were acquired with the same time duration as breath held data. A longer acquisition time per slice would capture more motion states under free breathing; whereas, a whole short‐axis stack under free breathing acquisition is typically desired to fully leverage free breathing for enhanced image quality and comprehensive assessment of cardiac structure and function. Lastly, in the absence of a reference standard for AF cine imaging, the commercially available real‐time sequence on our system was used. In future work, more state‐of‐the‐art real‐time imaging methods could be compared with the DRACO method.

5. Conclusion

The proposed DRACO pipeline demonstrated robustness against free breathing respiratory motion and irregular cardiac motion. Compared with breath‐held acquisitions, the DRACO method produces images of comparable quality under free‐breathing conditions across a broad range of heart rates in AF. Upon further validation, the DRACO framework has the potential to serve as a solution for more generalized motion representation.

Funding

This work was supported in part by the National Institutes of Health (R01HL148182, R01HL127153) and the Veterans Health Administration (I01CX001901).

Supporting information

Table S1: Representative pulse sequence parameters for sorted Cartesian golden‐step bSSFP and reference segmented cine and real‐time at 3.0 T.

Table S2: Image quality scoring criteria.

Figure S1: Multi‐phase (randomly selected) DRACO image comparison of K = 25 (${\lambda }_{\mathrm{BH}}=0.0005,{\lambda }_{\mathrm{FB}}=0.0008)$ and K = 35 (${\lambda }_{\mathrm{BH}}=0.0008,{\lambda }_{\mathrm{FB}}=0.001)$ of a mid‐short‐axis slice belonging to a patient in atrial fibrillation. The image quality between K = 25 and K = 35 is similar under both BH and FB. BH, breath held, FB, free breathing.

Video S1: Cine images of several patients in atrial fibrillation across a spectrum of heart rates during breath‐holding and free‐breathing. In this video, multi‐phase videos of the three patients under BH and FB conditions are shown. Images were ordered based on the motion cluster matrix $M$ to show the motion in the whole 18‐s scan. BH, breath held, FB, free breathing; HR, heart rate.

Video S2: Cine images of a patient in sinus rhythm during breath‐holding and free‐breathing. In this video, the periodic respiratory motion can be easily observed under FB as demonstrated by the correlation map in Figure 5. Even though the period of the respiratory motion seems dominant in the cluster matrix and in the correlation map under FB, cardiac motion can be clearly seen in the cine image. BH, breath held, FB, free breathing.

Video S3: Cine images of a patient in atrial fibrillation during breath‐holding and free‐breathing. In this video, dynamic changes in cardiac motion along the time dimension can be observed under BH and FB. The respiratory motion under FB is irregular, and there is an obvious respiratory shift occurring at around 7 s of the video. A heartbeat with minimal contraction (long RR interval) can be detected at ~9 s of the video. These videos capture the dynamic irregular motion states, supporting the analysis shown in the correlation map and the recorded ECG. BH, breath held, FB, free breathing.