Developing an Ultrasound-Electrocardiography Deep Learning Framework for Predictive Cardiac Gating under Arrhythmia and Variable Heart Rate Conditions

Introduction:: Cardiovascular disease (CVD) is the leading cause of death in the world, contributing to approximately 17.9 million deaths yearly [1]. The current gold standard for evaluating CVD, catheter coronary angiography, is invasive and expensive. Alternatively, cardiac computed tomography angiography (CTA) is a faster, less invasive diagnostic imaging approach for CVD. However, using conventional methods for gating CTA acquisition, patients with arrhythmias have an increased rate of non-diagnostic scans due to motion artifacts and are also subjected to higher radiation dose than patients without arrhythmias. Ultrasound (US) has been used in previous studies to provide the ground truth cardiac quiescence, i.e., the time point of minimum motion in the cardiac cycle [2]. Until recently, US gating was not feasible due to the lack of CT-compatible US transducers, as commercial US transducers produce prominent streak artifacts disrupting the CT scan [3]. We present a novel hybrid gating approach using both US and electrocardiography (ECG). As a first step, we employ a US-ECG deep-learning framework evaluated in a variety of cardiac motion cases, including arrhythmias, using a programmable commercial dynamic heart motion phantom (Simutec DHP-01) controlled by software (Visual Show Automation, VSA). A volume rendering of acquired 2D ultrasound images of the phantom cross-section is shown in Fig. 1a.


Materials and Methods:: Thirteen different experiments were executed, with motion ranging in complexity from simple combinations of linear and torsion motion to more complex, physiologically accurate, cardiac-like motion. A trimodal data acquisition system asynchronously acquired echocardiography using a Phillips P4-1 phased array transducer, artificial ECG from a passive ECG adapter in the phantom, and optical images from a camera. For all experiments, a ground truth quiescent period (QP) was determined by 4 separate methods: 1) Horn-Schunck optical flow method, 2) particle image velocimetry, 3) a correlation-based phase-to-phase deviation measure on the B-mode echocardiography images, and 4) template tracking from the optical images of the phantom (Fig. 1b). In addition, the interventricular septum (IVS), a key indicator of cardiac quiescence, was manually segmented. The deep learning model, consisting of 2 branches, was used to predict the frames corresponding to the phantom’s cardiac QPs on a beat-by-beat basis using US and ECG data. In the first branch, US frames are fed into a 3D Convolutional Neural Network (CNN), and in the second, 3 features extracted from the passive ECG signals are fed into an artificial neural network (ANN) model. The predicted QP outputs of the two branches were then fused to predict the common QP frames. In addition, partial cardiac cycle segments (i.e. the first 30%, 60%, or 80% of US and ECG data from each cycle) were used as input to test the model’s ability to predict the QPs within the same cardiac cycle before the occurrence of cardiac quiescence.


Results, Conclusions, and Discussions:: Frame-based validation across the ground truth measures—namely the optical and the VSA modalities—reveals a 95.6% accuracy for identifying the center of the phantom QP window for the US-based modality, and an 88.3% accuracy for identifying the QP window length. Velocities across different sections of the phantom are compared (Fig. 1c). Of the 520 cardiac cycles collected for each modality, 416 cycles were used for training and 104 cycles for validation and testing of cardiac quiescence prediction. The deep learning model predictions based on incomplete phantom datasets (30%, 60% and 80% of input data) achieve prediction accuracies of 68.8%, 86.1% and 91.3% for the whole phantom, while the IVS dataset achieves accuracies of 62.9%, 78.5% and 88.2%, respectively. The IVS and whole phantom datasets achieve a similar degree of accuracy under the constraints of partial data, indicating that imaging the IVS may be sufficient for QP prediction. Furthermore, using the IVS dataset for QP prediction decreases the neural network’s computational complexity. In summary, deep learning was utilized to predict QPs using a combination of echocardiography frames and passive ECG signals. 13 different types of cardiac motion were validated on the phantom, the results of which were used to further validate the QPs predicted by the US-ECG gating framework.


Acknowledgements (Optional): :

References (Optional): : [1] World Health Organization cardiovascular disease statistics, https://www.who.int/health-topics/cardiovascular-diseases#tab=tab_1, Accessed:06-07-2023.



[2] J. Yao, S. Tridandapani, C. A. Wick, and P. Bhatti, “ Seismocardiography-Based Cardiac Computed Tomography Gating Using Patient-Specific Template Identification and Detection,” IEEE Journal of Translational Engineering in Health and Medicine, 5, pp. 1-14, 2017.



[3] S. S. Rojas, S. Tridandapani, and B. D. Lindsey, “A thin transducer with integrated acoustic metamaterial for cardiac ct imaging and gating,” IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 69, no. 3, pp.1064–1076, 2021.