FlowDIP: Real-Time Phase-Contrast MRI Reconstruction with Flow-Conditional Deep Image Prior (RF_SA_447)

Segmented cardiac phase-contrast (PC) MRI relies on ECG synchronization and breath holding (BH), which fails in patients with arrhythmia or poor respiratory control. Real-time (RT) PC-MRI overcomes these limitations; however, compressed sensing (CS)-based reconstruction methods often underestimate the peak velocity in accelerated RT PC-MRI. To overcome this issue, we propose a deep image prior (DIP) [1]-based reconstruction method, called FlowDIP.

Methods:

Fig. 1 presents the outline of the Flow-conditional Deep Image Prior (FlowDIP) framework. Each reconstructed image is formed as a linear combination of R=60 spatial basis generated by a U-Net. The input to the U-Net is a static code vector z₀, initialized as a three-channel real-valued noise image. The dynamic and flow-encoding information is modeled by temporal coefficients ν_t^e generated by a Multilayer Perceptron (MLP). These coefficients are driven by a dynamic code vector z_t and flow-conditional index e. Here, z_t is a real-valued size-3 vector, while e assumes values of +1 or -1, representing flow-encoded and flow-compensated images, respectively. During the image reconstruction process, the parameters {θ,Φ,z₀,z_t} were optimized to ensure the reconstructed images are consistent with the acquired k-space y_t^e. The optimized z₀ and z_t  are shown in Fig. 1, where both cardiac and respiratory motions are visible in the dynamic code vectors.

Ten healthy volunteers were imaged using both BH segmented PC-MRI (GRAPPA rate 2, temporal resolution ~55.3 ms, spatial resolution ~2.7x2.1 mm², acquisition time 15 s, and ~20 frames) and free-breathing RT PC-MRI (Cartesian CAVA sampling [2], rate 15, temporal resolution ~54.8 ms, spatial resultion~2.7x2.6 mm², acquisition time 10 s, and ~182 frames) on a 1.5T scanner (MAGNETOM  Avanto, Siemens Healthcare, Erlangen, Germany).

RT data were reconstructed offline using CS [3] and FlowDIP. After reconstruction, background phase correction was applied using a second-order weighted regularized least squares fitting [4]. Net positive flow, peak velocity, and peak flow rate of the ascending aorta were quantified in suiteHEART (NeoSoft, Pewaukee, WI, USA). For RT acquisitions, the quantification results from all beats were averaged.

Results:

Fig. 2 shows representative images and flow curves from one of the volunteers. FlowDIP demonstrates sharper edges in the magnitude and phase images, as well as a higher peak flow rate compared to CS. Fig. 3 shows (a) scatter plots and (b) Bland-Altman plots for net positive flow, peak velocity, and peak flow rate. For net positive flow, both FlowDIP and CS show minimal underestimation. However, CS significantly underestimates both peak velocity and peak flow rate, whereas FlowDIP reduces the mean difference from -16.9 to -4.5 cm/s for peak velocity and from -54.8 to -7.3 ml/s for peak flow rate.

Conclusion:

We proposed a novel unsupervised RT PC MRI reconstruction method, called FlowDIP. Compared to CS, FlowDIP offers better agreement with BH reference for both peak velocity and peak flow rate.