FReSCO: Flow Reconstruction and Segmentation for low‐latency Cardiac Output monitoring using deep artifact suppression and segmentation

Figures

• 

  FIGURE 1. The “FReSCO” (Flow Reconstruction and Segmentation for low latency Cardiac Output monitoring) framework for low‐latency monitoring of aortic flow. A, Real‐time golden‐angle variable density spiral data are forwarded to an external computer using Gadgetron. The pipeline is initialized using the 10 first frames (eg, module imports, trajectory, density compensation, coil estimation, Maxwell correction), and the proposed reconstruction and flow monitoring are performed during scanning. Flow maps and segmentations are sent back to the scanner. B, Illustration of the proposed real‐time monitoring interface at the end of a 3‐min exercise scan. Top left: segmented magnitude images; top middle: phase images; top right: flow curves; bottom left: beat‐to‐beat heartrate; bottom middle: stroke volume; bottom right: cardiac output. Abbreviation: nuFFT, nonuniform fast Fourier transform

• 

  FIGURE 2. A, Overview of the FReSCO framework at inference. Deep artifact suppression is performed on the buffered gridded flow‐encoded and compensated data. Segmentation is performed on the combined magnitude contrast limited adaptive histogram equalized (CLAHE) images. B, Sliding window reconstruction timings (at scale). Each acquired frame is gridded independently but deep artifact–suppressed and segmented as a block of 24 frames

• 

  FIGURE 3. Heartrate (HR), stroke volume (SV), and cardiac output (CO) curves obtained from two subjects with different responses to exercise. In white, the median filtered curves. Blue arrows depict the 10 s areas used for comparison to CS and reference at rest (4–14 s) and CS only at peak exercise (107–117 s)

• 

  FIGURE 4. Reference (repeated for three cycles), compressed sensing (CS), and FReSCO images at rest, and CS and FReSCO images during exercise in a representative subject. Magnitude x‐y, magnitude x‐t, phase x‐y, and phase x‐t images are shown. Automatic segmentations (computed from the proposed deep learning [DL] images) are overlaid on top of the magnitude images. Corresponding video in Supporting Information Video S3

• 

  FIGURE 5. Correlation and Bland Altman plots comparing cardiac output from proposed FReSCO to reference at rest (A), compressed sensing (CS) to reference at rest (B), FReSCO to CS at rest (C), and FReSCO to CS at peak exercise (D). The values reported for CS and FReSCO are the average cardiac output obtained over 4 to 14 s (rest) and 107 to 117 s (exercise) of the real‐time acquisition