Steady-state free precession (SSFP) cine imaging is the method of choice for evaluating ventricular size and function because of its short repetition time, high signal-to-noise ratio, and excellent blood-to-myocardium contrast [1]. The standard clinical approach uses a retrospective ECG-gated 2D segmented Cartesian SSFP sequence that acquires 1-2 slices per breath-hold. The technique is, however, vulnerable to misregistration error due to variations in breath-holding position [2] and requires operator-dependent planning. Recent advances in acceleration techniques have led to single breath-hold whole-heart 3D cine acquisitions [3]. However, the limited duration of a single breath-hold requires compromises in spatiotemporal resolution and anatomic coverage. To address this, we combined the recently developed respiratory motion compensation algorithm (Cine-Nav) with a Cartesian variable density Poisson disc undersampling pattern and compressed sensing (CS) reconstruction [4].The schematic diagram of Cine-Nav for the free-breathing, whole-heart 3D cine SSFP acquisition is shown in Fig. 1. Two pencil-beam navigators (NAVs) are performed during each beat to monitor the respiratory motion of the right hemidiaphragm (RHD). If the RHD position of NAV2 is within the acceptance window and greater than NAV1, the next beat is used for data acquisition. NAV1 is then performed at the conclusion of the beat, and, if it is within the acceptance window, the acquired data is accepted for image reconstruction. To reduce the scan time while maintaining a high spatiotemporal resolution, a new variable density Poisson disc undersampling pattern is generated for each cardiac phase. The sampling rate decreases exponentially, from fully sampled at the k-space center to 16-18% sampled at the k-space periphery (Fig. 2). After the acquisition, the data were retrospectively binned into 20 cardiac phases and a 4D k-space dataset was reconstructed. A non-linear CS reconstruction was performed using BART [4, 5]. The data were first coil-compressed to 12 virtual channels using geometric coil compression [6]. The fully sampled central 2% of k-space at each cardiac phase was averaged and used to estimate ESPIRiT coil sensitivities [7]. An IFFT was performed in the fully sampled readout direction, and the data were reconstructed slice-by-slice and regularized with spatial wavelets and temporal finite-differences. The reconstruction was then interpolated to 30 time points. To assess the utility of the proposed sequence, 10 patients (4 males; age 19±9 years) with informed consent underwent the proposed free-breathing 3D cine acquisition on a 1.5 T MR scanner (Philips Ingenia). The imaging parameters were FOV ≈240(SI)×240(AP)×150(RL) mm, isotropic spatial resolution 1.8-2.0 mm³; α/TE/TR 60°/1.52/3.0 ms, acquired heart phases 20, bandwidth ≈1.78 kHz, acceptance window of 10 mm and a tracking factor of 0.6, a 28-element phased-array coil, and a CS reduction factor of ≈6. The 3D cine datasets were analyzed using Arterys software to measure left and right ventricular end-diastolic (EDV), end-systolic (ESV), stroke volume (SV), and ejection fraction (EF) (Fig. 3). These measurements were then compared with those from a conventional multiple breath-hold 2D slice cine acquisition with the following parameters: FOV ≈270(SI)×270(RL)×107(AP) mm, spatial resolution 1.8×1.8×8 mm; slice gap 1 mm, α/TE/TR 60°/1.4/2.8 ms, acquired heart phases 20 and interpolated to 30, bandwidth ≈1.08 kHz, and SENSE with reduction factor of ≈2. A paired two-tailed Student’s t-test was used for the statistical analysis and a p-value ≤0.05 was considered statistically significant.Fig. 4 shows representative 3D cine images acquired from 2 patients. EDV, ESV, SV, and EF by 3D cine acquisition were not significantly different compared to the conventional 2D cine breath-hold technique (Table 1). Mean scan time for the 3D cine acquisition was 7.5±0.6 minutes. The mean image reconstruction time was 107 minutes on a standard r3.4xlarge Amazon EC2 instance. We developed a free-breathing 3D cine whole-heart acquisition accelerated with variable density Poisson disc undersampling that yielded similar measurements of ventricular volumes compared to a conventional 2D breath-hold approach. Future work will further optimize and expand this approach so as to reduce scan time, reconstruction time, and operator dependence.