Synopsis

Keywords: Flow, Cardiovascular

Quantification of blood flow using 2D or 4D phase-contrast (PC) MRI is routinely being used to evaluate blood flow in cardiovascular disease. However, PC has only a modest temporal resolution compared to echocardiography. We sought to develop a deformation-encoding transformer (DENT) model for cardiac frame interpolation and evaluate its potential in increasing temporal resolution. DENT was trained using a large multi-center (centers = 3), multi-vendor (vendors = 3) and multi-field strength (1.5T, 3T) cine MRI dataset (patients = 3178). The model was successfully applied to 2D/4D-PC MRI without modifications, enabling a 2-fold gain in temporal resolution.

Background

Quantification of blood flow using 2D or 4D phase-contrast (PC) MRI is routinely being used to evaluate blood flow in cardiovascular disease. Cardiac 2D or 4D phase contrast (PC) flow imaging allows evaluation of blood flow and hemodynamics within cardiovascular system. In ECG-segmented 2D or 4D PC imaging, there is a trade-off between scan time and spatiotemporal resolution. Insufficient temporal resolution leads to imprecise peak velocity measurements. We sought to develop a Deformation ENcoding Transformer (DENT) for high-frame-rate 2D/4D PC MRI.

Methods

DENT worked in the image domain and generated forward-backwards mappings describing the underlying deformation of multiple cardiac phases. Such deformation, as well as a blending mask, was used to synthesize images with high framerate. Input was T = 4 cine images of size W × H collected with temporal resolution ∆t at t - ∆t, t, t + ∆t and t + 2∆t. Output was interpolated image at t + 0.5∆t (Fig. 1a). An embedding layer extracted F = 32 features per pixel from inputs. Downsampling layers reduced image dimensionality by half while doubling feature dimensionality. Inputs to transformer layers were split into windows. Multiscale attention was used to learn a spatiotemporal correspondence within each window (Fig. 1b). For spatial attention, inputs were split into N = W•H•T / M2 windows of size M2 × F along spatial dimension (M = 8). For temporal attention, inputs were split into N = W•H windows of size T × F (Fig. 1c). Decoder layers upsampled the dimensionality of encoder outputs by 2 while reducing feature dimensionality by half. These were passed to the motion synthesizer block. Additional convolutional layers in the block upsampled the input. The final outputs of the model were deformation , scaling components , and blending mask (Fig. 1d). A new cardiac frame was synthesized from each input image using deformation and scaling components via bilinear interpolation. The blending mask was used to combine synthesized images onto a single frame at t + 0.5∆t. Multi-center (centers = 3), multi-vendor (GE, Philips, Siemens), multi-field strength (1.5T, 3T) scans from 3178 patients (2139 male, 54 ± 16 years) undergoing clinical MRI for various cardiac indications were used for training. Cine images were collected using a breath-hold ECG-gated segmented SSFP at 1.5T (n = 1831) and 3T (n = 1347) in short axis and 2-, 3- and 4-chamber. A sample was as a cine slice with T ≥ 7 frames; an epoch one optimization loop across all training samples. First, a center-frame was randomly selected and used as the ground-truth. Second, 4 frames adjacent to were selected as inputs: either or , which had a 2∆t and 4∆t temporal resolution, accordingly (Fig. 2). Each cine frame was normalized by min-max prior to processing. Ground-truth and input images were randomly cropped to 256 × 256. Training used a batch size = 10 for 100 epochs (~200 hours) using AdaMax. Learning rate was 2 × 10-2 and gradually decayed to 1 × 10-6. 2D or 4D PC MRI imaging dataset from patients who were imaged using Siemens 3T Vida system were extracted to demonstrate feasibility of DENT-enabled high-frame-rate flow. Breath-hold ECG-segmented 2D PC scans were acquired with the following imaging parameters: spatial resolution = 1.9 × 1.9 mm2, temporal resolution = 37 ms; GRAPPA rate = 3, number of phase-encode line per segment = 4. Navigated free-breathing ECG-segmented 4D PC scans were acquired (acquisition time = 4.5 min) with the following imaging parameters: TE/TR = 2.3/4.58 ms, spatial resolution = 2.5 × 2.5 × 2.5 mm3, acquired temporal resolution = 37 ms, reconstructed number of phases = 25 ms, vendor-provided WIP compressed sensing rate 7.2, number of phase-encode line per segment = 2. Encodings , and were derived from the anatomical images of flow datasets (Fig. 3). These encodings were applied to both the anatomical and flow images to synthesize a new cardiac frame. For 4D flow, the encodings were applied to flow images in x, y and z directions separately. 4D flow were processed and visualized in Medis Suite (4.0.50.2).

Results

trained DENT model enabled both anatomical and flow images with high-frame-rate, as demonstrated for 2D PC (Fig. 4). A two-fold gain was achieved in 4D flow datasets with simulated and reconstructed temporal resolutions of 54 and 27 ms, accordingly. Our approached resulted in 4D flow with temporal resolution as low as 14 ms (Fig. 5). DENT was trained in large cine-based dataset. Translational feasibility for PC was demonstrated. However, additional (i.e., phantom) studies should be performed to assess the contribution of each encoding component. For instance, in cases where deformation does not play a major role (e.g., center of vessels), the synthesized flow values are essentially a linear combination of the four adjacent frames. The coefficients of that linear combination are given by the blending mask. Nevertheless, our study showed that transformer-based frame interpolation has the potential to increase the temporal resolution of PC imaging

Conclusion

We trained a transformer-based frame interpolator using a large multi-center, multi-vendor and multi-field strength dataset and demonstrate its potential in increasing the temporal resolution of 2D/4D PC MRI.

Acknowledgements

No acknowledgement found.

References

No reference found.