MRI is increasingly being used to image the fetus as an adjunct to ultrasound, however motion remains a key limiting factor in studies of the fetal heart and great vessels [1]. The challenges are numerous when imaging a small, rapidly beating heart within the context of the maternal torso that is subject to various quasi-periodic (maternal respiration, fetal respiratory movement, fetal cardiac pulsation), and spontaneous, episodic movements (gross fetal movement). Electrocardiogam (ECG) gating of the fetal heart is unreliable, though segmented cine acquisitions have been achieved using self-consistency in reconstruction to infer a gating signal [2] or MR-compatible Doppler ultrasound-based triggering [3], but non-cardiac sources of motion may still corrupt the data.

An alternative approach is to use rapid dynamic (‘real-time’) imaging with high temporal resolution to obtain serial ‘snapshots’ of the fetal heart and surrounding anatomy that can be motion-corrected and reassembled to combine several cardiac cycles into a single cine. In this work, real-time imaging with k-t SENSE [4] was employed, taking advantage of the fact that the highly-dynamic fetal heart occupies only a small fraction of the field-of-view. Motion-correction processes were used to identify consistent real-time image frames and align them in space and time as a single cardiac cycle. Finally, a super-resolution reconstruction [5] was performed on the densely-sampled, but relatively low SNR and large temporal duration, data to improve image quality.

Imaging was performed in five singleton pregnancies (gestational age: 28-35 weeks) at 1.5 T (Ingenia, Philips) using torso and spine receiver coil arrays in short and long axis orientations of the fetal heart during maternal free-breathing. Real-time imaging employed a 2D balanced steady-state free precession sequence using a k-t under-sampling factor of 8 to achieve a temporal resolution of 68-80 ms per real-time frame (TR/TE: 4.2/2.1 ms; FA: 60°; FOV: 400 × 300 mm; voxels: 1.8-2.0 × 2.0-2.3 x 6.0 mm). A spatiotemporally optimal sampling pattern [6] was used to minimise the overlap of aliases in x-f space (Fig.1). Low SAR (<2W/kg) and low acoustic noise settings were employed throughout, necessarily limiting scanner performance.

Real-time images were reconstructed offline using auto-calibrated sensitivity maps derived from eigenvalue decomposition [7] of temporally-averaged under-sampled data and noise-decorrelation pre-processing. Minimal regularization was applied in order to preserve image features and temporal fidelity, at the cost of increased noise. Data collected during episodes of gross fetal movement were removed prior to k-t SENSE reconstruction.

A user-specified target frame and fetal heart region-of-interest (ROI) were used for motion-correction. The fetal heart rate was identified in x-f space based on the conspicuous peaks in the fetal heart ROI at fundamental and harmonic frequencies (Fig.1b), allowing for reordering of real-time image frames into a cardiac cine series. Spatial displacement of the fetal heart was corrected by employing self-gating, using per-frame sum-of-squared-difference with the target frame calculated across the full maternal field-of-view, to limit through-plane displacement of the fetal heart due to maternal respiration, and rigid registration to correct for in-plane motion.

Following temporal reordering and motion correction to align successive frames, a final cine sequence with uniformly-sampled image frames representing a single cardiac cycle was generated using super-resolution reconstruction with edge-preserving regularization [5].

Motion correction was validated on adult and neonatal cardiac real-time data against breath-held, ECG-gated cine acquisitions.

Highly-accelerated real-time imaging was able to capture the motion of the fetal heart as well as the surrounding anatomy. The full pipeline successfully achieved cine sequences in all subjects. Rapid imaging enabled retrospective correction of motion, namely: identification of the fetal heart rate (measured range: 128-158 bpm); self-gating to exclude through-plane displacement due to maternal respiration (Fig.2c); and motion correction to stabilise the in-plane location. Super-resolution reconstructed cardiac cine images (Fig.3, Fig.4) showed improved representation of anatomical features. Each reconstruction step was tested individually and found to contribute to successful outcome when combined.

Highly-accelerated real-time imaging was possible using an interleaved k-t sampling pattern due to the compact x-f representation of the small fetal heart within a large field-of-view. These high temporal resolution real-time images make possible direct assessment of and correction for motion and outliers. Higher acceleration-rate data has already been acquired and reconstructed, but its properties are still being evaluated. Image compounding was able to recover image SNR lost in the highly-accelerated real-time image acquisition and temporal super-resolution reconstruction further increased the visibility of dynamic anatomical features, allowing finer details to be visualised. The proposed method shows promise as a framework for comprehensive fetal cardiac MRI in the future.