An inline deep learning based free-breathing ECG-free cine for exercise cardiovascular magnetic resonance

Stress cardiac imaging is the first-line provocative test for the diagnosis and prognosis of coronary artery disease (CAD). During the past decade, stress cardiovascular magnetic resonance (CMR) imaging has become widely used clinically, and is increasingly recognized as an important testing tool in CAD [1‐5]. CMR has many well established benefits over other imaging modalities such as fewer limitations from body habitus and no ionizing radiation. CMR cine imaging is also the non-invasive gold standard for assessment of left-ventricular (LV) anatomy and function and may be utilized for detection of stress-induced LV regional wall motion abnormalities [6, 7]. Nevertheless, current clinical CMR requires pharmacological stimulation to induce stress, which does not elicit the same cardiovascular response as physical stress [8‐10] and lacks important prognostic data [11, 12].

Coalescing technical innovations in CMR over the past 25 years are enabling CMR imaging during exercise (Ex-CMR) as an alternative to pharmacologic stress, including the advent of highly accelerated real-time cine sequences [13]. Ex-CMR was initially demonstrated by Weiss et al. using isometric handgrip exercise [14]. Later studies using a CMR-compatible bicycle ergometer required breath-holding to evaluate cardiac function during exercise stress [15, 16], which was problematic as breath-holds during exercise are difficult to perform. Lurz et al. later proposed a radial k-t SENSE sequence to assess biventricular volume and function during free-breathing continuous exercise [17]. However, an electrocardiogram (ECG) signal was needed to retrospectively reconstruct images, and ECG gating issues during maximal exercise intensity compromised image quality [18, 19]. Thus, a number of clinical studies since then have relied on the ungated real-time sequence proposed by La Gerche et al. using Cartesian k-space undersampling [20]. More recently, Haji-valizadeh et al. reported a free-breathing ECG-free real-time radial cine sequence with higher spatiotemporal resolution [21] using the golden-angle radial sparse parallel (GRASP) reconstruction framework [22]. Nevertheless, without specialized techniques such as parallel imaging and compressed sensing, highly accelerated cine images suffer from significant aliasing artifacts that can degrade LV assessment or render images non-diagnostic. From a clinical perspective, a practical yet crucial limitation of these specialized strategies is that reconstruction times can be long and not truly real-time. Alternatively, faster image reconstruction and de-aliasing of Cartesian [23‐25] and radial [26‐28] data is feasible with deep learning (DL). However, most methods have been evaluated on retrospectively undersampled or breath-hold images. Further, DL is yet to be evaluated for Ex-CMR, which is potentially more difficult since real-time images acquired during stress may have increased artifacts compared to rest.

In this study, we sought to develop an inline (i.e., on-scanner) DL-based approach for seamless, low-latency image acquisition and reconstruction of real-time ECG-free radial cine with an acceleration rate of 12 for Ex-CMR. In contrast to previous methods that use magnitude images for learning, our model was trained using complex-valued images simulated from raw k-space data acquired at rest. This approach preserves phase information in the images. We hypothesized that a DL model trained for image de-aliasing with phase-preserved, at-rest images is robust to the significant respiratory and cardiac motion that occurs during exercise stress. We evaluated and compared our method to GRASP on healthy subjects and patients with suspected CAD referred to Ex-CMR.

A DL radial acceleration with parallel reconstruction (DRAPR) model based on a three-dimensional U-Net was implemented to suppress streaking artifacts in highly accelerated real-time radial cine images. The model was trained using real-time radial images synthesized from conventional ECG-gated segmented Cartesian k-space data from 503 patients at rest. An inline pipeline was then implemented to enable fast, parallel reconstruction during a prospective Ex-CMR study.

Resting ECG in all patients showed sinus rhythm. Standard cine in one patient showed poor image quality due to off-resonance artifacts. This patient was excluded from the analysis of LV function since proper contouring was not possible, but was included in the analyses of image quality. In addition, one of the patients was unable to complete the Ex-CMR protocol due to a leg cramp during supine cycle exercise. The baseline and post-exercise heart rates were 72 ± 15 and 124 ± 19 bpm, which represent an average increase of 53 ± 20 bpm.

Reconstruction of real-time radial cine images with GRASP and DRAPR was successful in all cases. Using GRASP to reconstruct cine data whose frame matrix size was 288 × 288 required 1.2 s per frame for pre-processing and 6.1 s per frame for application of compressed sensing to the coil-combine complex images. With DRAPR, pre-processing times of 10 ms per frame were achieved, and application of the U-Net model to the coil-combine complex images was achieved in 6.6 ms per frame. Thus, for instance, an acquisition over 4 heart beats (i.e., ~ 100 frames) and 12 slices would require 2.4 h with GRASP and 20 s with DRAPR.

Measures of LV function across all subjects based on breath-hold ECG-gated segmented cine and free-breathing ECG-free real-time radial cine reconstructed with GRASP and DRAPR were 114 ± 15 mL, 118 ± 17 mL, and 117 ± 17 mL for LVEDV; 43 ± 9 mL, 49 ± 10 mL, 50 ± 10 mL for LVESV; and 62 ± 7%, 58 ± 10%, 57 ± 7% for LVEF. Comparison of measures based on ECG-gated and GRASP cine showed differences in LVEDV (4.4 mL [− 14.1, 22.8], P = 0.138) that were not significantly different from zero. However, differences in LVESV (6.4 mL [− 3.5, 16.3], P < 0.001) and LVEF (− 4.3% [− 14.3, 5.6], P = 0.003) were significant. Similarly, comparison of measures based on ECG-gated and DRAPR showed differences in LVEDV (3.0 mL [− 11.7, 17.8], P = 0.320) that were not significantly different from zero. Differences in LVESV (7.0 mL [− 1.3, 15.3], P < 0.001) and LVEF (− 5.0% [− 11.1, 1.0], P < 0.001) were significant. There were no significant differences between GRASP- and DRAPR-based measures of LV function (Fig. 4).

The mean artifact level score in ECG-gated segmented cine images was 3.9 ± 0.3. Only 3.0% of the scores were graded as severe and the remaining 97.0% were graded as minimal. All NUFFT-reconstructed real-time cine images were deemed as having a non-diagnostic level of artifacts, while none remained non-diagnostic after application of GRASP or DRAPR. The mean score in real-time cine images at rest reconstructed with GRASP was 3.7 ± 0.6, which was comparable (P = 0.3) to ECG-gated images. However, 9.1% of scores were graded as severe. The mean scores of images reconstructed with DRAPR was 3.3 ± 0.7, which reflects a significantly higher artifacts relative to ECG-gated (P < 0.001) and GRASP (P = 0.02) images. In addition, 10.6% of scores were graded as severe (Fig. 5). Various subjects representing a range of scores are shown in Fig. 6 and Additional file 1: Video 1, i.e., from a mean score of 3 (reader A, B, C = 2, 4, 3) to a mean score of 4 (reader A, B, C = 4, 4, 4) with DRAPR. The artifact level of images reconstructed with GRASP (3.5 ± 0.8) and DRAPR (3.1 ± 0.6) during post-exercise stress was significantly different (P = 0.002). Relative to scores of images at rest, the percentage of scores graded as severe increased to 17.9% and 15.4%, accordingly. A range from a mean artifact score of 2.3 with DRAPR (reader A, B, C = 2, 3, 2) to 3.7 (reader A, B, C = 3, 4, 4) in subjects post-exercise is shown in Fig. 7 and Additional file 2: Video 2.

Representative real-time cine images reconstructed with DRAPR in patients with suspected CAD are shown in Fig. 8 and Additional file 3: Video 3. Ex-CMR showed absence of regional wall motion abnormalities in seven of eight patients, which was consistent with findings from other modalities. Specifically, exercise tolerance test (ETT) was negative in two patients. Stress echocardiography tests were also negative in three patients, and nuclear stress showed a moderate perfusion defect in one patient. No other imaging tests were performed in any of these six patients. In one patient, stress echocardiography was positive but subsequent coronary angiography was negative for significant CAD. In the remaining patient whose Ex-CMR showed a regional wall motion abnormality, stress echocardiography showed hypokinesis of the LV basal inferoseptum, and coronary angiography revealed a heavily calcified right coronary artery with subtotal and chronic total occlusions (Fig. 9 and Additional file 4: Video 4).

In this study, we assessed the feasibility of inline real-time cine with DL-based radial acceleration for Ex-CMR. We demonstrated that (1) while our training dataset was based on synthetic real-time images simulated from images acquired during breath-hold and at rest, DRAPR substantially suppressed artifacts in free-breathing real-time cine images acquired during post-exercise stress; (2) the proposed inline implementation for seamless, low-latency reconstruction is a feasible approach for evaluation of cardiac function in Ex-CMR.

Ex-CMR with real-time radial cine has the potential to detect regional wall motion abnormalities in patients with suspected CAD. Fractional flow reserve (FFR) is the gold standard invasive diagnostic test to assess functional significance of coronary artery stenosis [34]. However, studies suggest that CMR can provide highly sensitive and specific CAD assessment compared to invasive FFR, and may be non-inferior at predicting adverse events [4, 35, 36]. Compared to pharmacological stress, Ex-CMR may provide additional exercise-capacity and hemodynamic prognostic information and elicit a more powerful stress response. Thus, exercise stress is preferable for cardiac stress testing when feasible. Despite its clinical importance, the clinical usefulness of Ex-CMR has been limited by lack of adequate temporal resolution and image quality due to elevated heart rates, exaggerated breathing, and gross patient motion during imaging. In recent years, the advent of highly accelerated sequences have enabled free-breathing ECG-free cine imaging [13]. In this feasibility study, we implemented a prototype cine sequence with DL-based radial acceleration and spatiotemporal resolution of 37.7 ms and 2 × 2 mm², respectively. Our inline implementation was able to quickly reconstruct and suppress streaking artifacts in 16.6 ms per frame, and evaluation of LV function in all patients with suspected CAD using both rest and post-exercise cine images was in agreement with other stress imaging modalities.

Isometric handgrip was the first exercise method proposed for Ex-CMR [14]. Supine ergometry was later proposed by Mohiaddin et al. as an alternative exercise method that does not require patient repositioning and engages large muscle groups [37]. Previous Ex-CMR studies have reported exercise can be performed inside the scanner bore using a cycle ergometer [20]. However, we found that many patients are not able to adequately cycle inside the scanner due to limitations of the bore size. With our approach, there was only a small delay from peak exercise intensity to stress imaging. In addition, with the exception of one patient who experienced a leg cramp during supine cycling, all healthy subjects and patients were able to complete the Ex-CMR protocol.

The underestimation of LVEF in real-time cine relative to ECG-gated was due to greater volume overestimation at end-systole compared to end-diastole. Physiologic and parameter variations between ECG-gated and real-time imaging could lead to systematic differences in LV parameters. Further, our three-dimensional U-Net implementation applies a temporal filtering that could result in reduced temporal resolution and may explain the LVEF underestimation. In addition, it might also make visual detection of regional wall motion abnormalities more difficult, especially during post-exercise or dobutamine stress. Alternatively, a recurrent two-dimensional U-Net could avoid temporal smoothing. Nevertheless, these volume differences in volume were smaller than those reported in a previous validation study of a similar real-time radial cine sequence reconstructed with GRASP [21]. In that study, comparison to ECG-gated quantification of LV parameters showed mean differences in LVEDV and LVESV of 15.2 mL and 7.9 mL, respectively.

In this study, we also compared DRAPR- to GRASP-based dealiasing. Although the artifact level scores were higher with GRASP, the difference between the two was small and may not necessarily impact clinical interpretation or quantification. Indeed, we found no significant differences between GRASP- and DRAPR-based measures of LV function, and both methods received a similar amount of severe artifact scores. Further, DRAPR was able to de-aliased images with a 1000-fold reduction in computational time compared to GRASP. In Ex-CMR studies, cine images may be acquired over multiple heart beats and multiple slices. DRAPR enables inline reconstruction of these images in the order of seconds instead of hours, which is important to prevent delays in the clinical workflow.

We did not observe any dependence on heart rate in the artifact scores in our study. However, subjects with high body mass index showed increased artifact levels. The lowest artifact level scores at rest represent reduced suppression of streaking artifacts arising from high-intensity signals and field inhomogeneities in the peripheral field-of-view. Since the peripheral FOV normally contains a higher amount of artifacts, better artifact suppression could also be achieved by using the entire image matrix during training. However, such implementation would require larger GPU memory. Alternatively, a patch-level training approach could perform better than or similar to an image-level approach with reduced memory demand.

Low latency enables quick evaluation of images, and also prevents computational bottlenecks that may occur in imaging protocols with finite computing resources. Our integration enabled us to perform NUFFT reconstruction and coil combination in 10 ms per frame, which was almost one quarter of the temporal resolution. This approach could be used to speed up previously proposed DL methods. For example, in the model proposed by Shen et al. for real-time radial cine, these pre-processing steps were done in ~ 0.3 s per frame [27]. Better artifact suppression might be possible if additional GPUs are available, or in protocols where some latency can be sacrificed. For instance, we concatenated real and imaginary components to enable DL processing with real-valued convolutional kernels. Alternatively, complex-valued convolutions may be used, but would require a two-fold increase in the number of parameters and a four-fold increase in the number of operations. Also, the artifact level in gridded images could be reduced by increasing the number of neighbors and point precision in the NUFFT, or by including prior smoothing and increasing the number of iterations during coil sensitivity estimation. Nevertheless, we showed that streaking artifacts were substantially suppressed using our fast inline model in various Ex-CMR studies in both healthy subjects and patients with suspected CAD.

Combination of highly accelerated radial sequences with DL enables fast acquisition and reconstruction of real-time cine images with suppressed streaking artifacts. Our study demonstrated the feasibility of low-latency, inline implementation of such methodology for clinical Ex- CMR.