Inline automatic quality control of 2D phase‐contrast flow MRI for subject‐specific scan time adaptation

To develop an inline automatic quality control to achieve consistent diagnostic image quality with subject‐specific scan time, and to demonstrate this method for 2D phase‐contrast flow MRI to reach a predetermined SNR.


INTRODUCTION
Conventional MRI uses a fixed acquisition duration to provide acceptable image quality for most patients.Indeed, MRI protocols are optimized to provide diagnostic images as fast as possible while maintaining sufficient information for evaluating the targeted structures.However, image quality can vary between patients.In cardiac imaging, image quality is worse in patients with larger body habitus, patients who move, patients with irregular breathing, or when coils are improperly positioned.In other patients, sufficient image quality is achieved quickly, and the default scan time is unnecessarily long.Insufficient image quality can result in longer exam time due to scan repetition, and in the worst case, the patient may need to return for a repeat exam.Rather than waiting until the end of the scan to identify poor quality, we hypothesized that suboptimal images could be detected and prevented by intermittent automated inline evaluation during the scan.
In recent years, automatic quality control tools have been designed to detect nondiagnostic images.For example, methods have been presented that automatically detect artifacts, 1,2 identify missing apical and basal slices on cine cardiac MR, 3 detect cardiac motion, 4 detect interslice motion along with heart coverage and image contrast estimation, 5 and identify segmentation failures. 6,7owever, these have been primarily applied offline after the exam, or online after a completed acquisition, but never during the acquisition.Meanwhile, low-latency and real-time processing applications have emerged.Jaubert et al. 8 implemented low-latency inline flow measurements monitoring.Frueh et al. 9 developed real-time landmark detection and tracking.Huttinga et al. 10 realized a real-time low-latency 3D nonrigid motion field thanks to an offline preparation from a previous acquisition.Furthermore, subject-specific adaptive acquisitions have been proposed.Contijoch et al. 11 developed a closed-loop system to choose optimal sampling for segmented cine radial acquisitions within the sequence controller.Breutigam et al. 12 developed an automatic feedback for adjusting the postlabeling delay in arterial spin labeling.Vidya Shankar et al. 13 developed a method using automatic slice tracking to follow a catheter for MR-guided cardiac catheterization.A closed-loop image-based assessment of scan quality will bridge modern image-reconstruction techniques and novel computer vision programs with the clinical workflows to maximize the effectiveness of these tools.
We proposed an inline automatic quality control based on a generalizable closed-loop feedback framework between image reconstruction and data acquisition to efficiently achieve consistent diagnostic image quality based on a predetermined metric.SNR is directly related to the confidence of flow measurements 14 ; therefore, we applied this framework for cardiac flow measurements with 2D phase-contrast MRI using an SNR threshold as a stop criterion for achieving accurate flow measurement across subjects with subject-specific imaging duration.

Theory
We implemented closed-loop feedback between the image reconstruction and data acquisition.Two-way communication between the image acquisition software and the Gadgetron 15 reconstruction software is handled using "FIRE" research framework 16 (Siemens Healthineers, Erlangen, Germany), which supports streaming of MRI raw data. 17Periodically during the acquisition, the quality of the image is automatically evaluated, and a feedback message is sent to the sequence controller (i.e., a small data packet that reports the quality metric).If the quality requirement is met, the acquisition will automatically stop itself; if not, the acquisition will continue.

Automatic SNR-driven inline quality control workflow
Figure 1 provides a schematic diagram of our workflow.We applied the closed-loop feedback framework for 2D pseudo-golden spiral phase-contrast flow acquisition that will automatically stop when the target SNR is achieved.
Every 20 s, SNR maps are estimated using the pseudo-replica method, 18 which is a Monte Carlo approach that emulates the gold-standard repeated-image signal-to-noise measurement (scan, rescan).Random Gaussian noise based on measured noise statistics is added to k-space before image reconstruction, generating a pseudo replica image.The process is repeated N times (N = 100, here) with different synthetic noise, resulting in a stack of independent pseudo-replica images.Then, the pixel-wise SNR is calculated using the ratio of the image over the standard deviation of the image replicas.
To minimize computation time, data from the entire cardiac cycle are used for interim SNR analysis instead of cardiac-resolved SNR maps.This time-averaged SNR was correlated with cardiac-resolved SNR to validate this approach.The SNR of the targeted tissue, either ascending aorta (AAo) or main pulmonary artery (MPA),

F I G U R E 1
Schematic diagram demonstrating the inline automatic quality control based on assessment of SNR and feedback messaging.Every 20 s, SNR maps are generated by Gadgetron, and the SNR feedback is extracted automatically in the targeted tissue and sent using the FIRE framework to the sequence controller.As soon as sufficient SNR is reached in the targeted tissue, a stop message is sent to the acquisition, and cardiac cycle-resolved magnitude and phase images are reconstructed with their segmentations.
was then extracted by automatic segmentation using a nnUNet 19 and sent to the sequence controller.The operator selected an image-reconstruction pipeline from a drop-down menu, to ensure that the appropriate segmentation network for AAo or MPA was used inline.
The stopping criterion was defined as a minimum SNR threshold (i.e., after the target SNR is achieved, the sequence will finish executing the current loop and stop).If the target SNR never reaches the threshold, the acquisition will stop at the maximum number of prescribed averages.At the end of the scan, a higher quality image reconstruction was performed inline using Gadgetron.The acquired data are retrospectively self-gated to 25 cardiac frames and reconstructed using Temporal (T)-conjugate-gradient (CG) SENSE with spatial and temporal constraints ( s = 0.1,  t = 1), and images are returned to the scanner host along with the segmentation of the target tissue for each frame using a nnUNet.The T-CG-SENSE reconstruction used to generate the final images following automatic stop will have inherently higher SNR than the nonuniform fast Fourier transform reconstruction used for the rapid inline quality assessment.

Healthy volunteer and patient imaging
Institutional review board approval and written informed consent from all study participants was obtained (ClinicalTrails.govidentifier NCT03331380).Ten healthy volunteers (body mass index [BMI] = 25.4 ± 2.5, age = 30 ± 8 years, male/female = 4/6) were imaged on a 0.55T MRI scanner (MAGNETOM Free.Max; Siemens Healthineers) with a prototype gradient coil.We used the vendor body array and the spine coil array, totaling up to 21 channels.A free-breathing, gradient-echo, single-slice, pseudo-golden-angle spiral flow sequence (TE/TR = 2.0/10.5 ms, flip angle [FA] = 25 • , [1.7 mm] 2 resolution, 8-mm slice thickness, through-plane v enc = 200 cm/s, FOV = [384 mm] 2 ) was modified to listen for and process the feedback messages.Two scans were performed in each subject: (1) full acquisition time and (2) a SNR-driven automatic stop acquisition with a maximum scan time of 4 min 50 s (AAo) or 6 min 10 s (MPA).Interim pseudo-replica SNR estimation, automatic segmentation, and image reconstructions were performed using Gadgetron on a computer equipped with four GPUs (NVIDIA A100-SXM, 80 GB) and 128 CPUs cores (2× AMD EPYC 7H12 64-core processors).
To test the clinical robustness of the inline automatic quality control, 1 patient (BMI = 29.3,age = 73 years old, female) with a prosthetic aortic valve was also recruited.

Image quality assessment: Automatic segmentation
We chose the nnUNet 19 framework to automatically segment the ascending aorta and the main pulmonary artery in our workflow.The nnUNet has already shown good performance for cardiac MR segmentation challenge. 20To train and evaluate neural network models, a mono-centric database with 138 patients was defined, divided into 128 patients for training and validation and 10 for testing, all of whom underwent a cardiac MRI exam at 1.5 T (MAGNETOM Aera; Siemens Healthineers) and/or 0.55 T (prototype MAGNETOM Aera or MAGNETOM Free.Max; Siemens Healthineers).For all enrolled subjects, 2D Cartesian phase-contrast MR images were acquired with the following parameters for 0.55 T (TE/TR = 4.3/14.1 ms, FA = 30 • , [1.56-2.2mm] 2 resolution, 6-mm slice thickness, through-plane v enc = 200 cm/s, FOV = 270 × 360 mm 2 , three averages) and at 1.5 T (TE/TR = 2.7/10.0ms, FA = 20 • , [1.40-1.56mm] 2 resolution, 6-mm slice thickness, through-plane v enc = 200 cm/s, FOV = 270 × 360 mm 2 , three averages).As a result, 454 acquisitions were obtained, as some patients had more than one acquisition per targeted tissue, divided in 59 (AAo = 30, MPA = 29) at 1.5 T and 395 (AAo = 200, MPA = 195) at 0.55 T. For reference, AAo and MPA were segmented for all 25 retrogated cardiac phases using an automatic tool provided by a commercial software (suiteHEART version 5.1.0;NeoSoft) and revised by experts.
The default 2D nnUNet networks were trained for AAo and MPA independently with 5-fold cross-validation, stratified by field strength and subject independence.The models were trained over 1000 epochs with a batch size of 106 using a stochastic gradient descent with Nesterov momentum (mu = 0.99) and an initial learning rate of 0.01.Data augmentation was performed on the fly.All images were resampled to (1.875 mm) 2 in-plane resolution with an interpolation of order 3.

Image quality assessment: Stopping criterion
The stopping criterion was defined by a target SNR threshold.A retrospective analysis of the SNR feedback messages sent every 20 s during the full acquisition duration was conducted in 10 healthy volunteers to determine an optimal SNR stopping threshold.The optimal stopping threshold was chosen to produce an accurate measurement of cardiac output, defined as less than 5% absolute relative error compared with the full acquisition time, which was chosen to be long enough to provide precise measurement (4 min 50 s [AAo] or 6 min 10 s [MPA]).This clinical criterion has been chosen empirically to represent the concept of a diagnostic measurement criterion, but it is equivalent to previously reported intersite variability of 5% in cardiac exams. 21This stopping threshold was then applied inline for 6 volunteers.

Statistical analysis and evaluations metrics
For the automatic segmentation, the nnUNet network was evaluated on a test data set including 10 healthy volunteers using conventional segmentation metrics (Dice similarity coefficient, Haussdorff distance, absolute relative surface error).Flow analysis was performed using MATLAB R2021a (The MathWorks; Natick, MA, USA).Statistical analysis has been conducted using R (version 4.3.0).

Feasibility of inline quality control
For the automatic segmentation, 2D nnUNet provided accurate segmentation of the AAo and MPA with mean Dice similarity coefficient of 0.95 ± 0.02.Segmentation cross-validation results are provided in Tables S1 and S2.
The automatic segmentation required 1.08 ± 0.09 s, and SNR map computation (100 pseudo-replicas) required 12.98 ± 5.49 s.The inline control computation time grows throughout the scan as more data are collected, starting at 4 s and ending at 20 s for the full acquisition.The SNR map reconstruction time increased based on the density compensation and nonuniform fast Fourier transform operations that scale with the number of spiral shots.SNR estimation using the pseudo replica method requires a trade-off between computer processing time and SNR measurement precision according to the number of replicas (Figure S1).The total latency of the computation was always compatible within the 20 s assessment interval, which is crucial to avoid a growing lag between acquisition and image-quality assessment that would result in incorrect stopping time and protocol inefficiency.The final images reconstruction using T-CG-SENSE was done inline using Gadgetron, such that final images return to the scanner after the acquisition, and required 1 min.Our reconstruction, SNR estimation, and segmentation implementations are available open-source using Gadgetron (https://github.com/NHLBI-MR/SNR-driven-flow).

SNR-based stopping criterion: Retrospective analysis
Retrospective analysis demonstrated that by choosing an SNR threshold of 175 for AAo and 140 for MPA (Figure 2A,B), we ensured sufficient image quality to maintain accurate quantitative cardiac output measurements with an error less than 5% relative to the full duration measurement (4 min 50 s for AAo and 6 min 10 s for MPA).These SNR threshold values are higher than typical cardiac-resolved 2D phase contrast because the intermittent SNR assessment is made on the time-averaged data (i.e., unbinned data, not cardiac-resolved).The time-averaged SNR correlated well (R 2 = 0.99) with the mean cardiac-resolved SNR (Figure S2).As expected, the time-averaged and cardiac-resolved SNR values are proportional by a factor corresponding to the square root of the number cardiac frames.

F I G U R E 2
Retrospective analysis of absolute error in cardiac output (CO) for ascending aorta (AAo; A) and main pulmonary artery (MPA; B) was performed to determine the SNR threshold used to stop the acquisition and then applied retrospectively across 10 healthy volunteers for the AAo (C) and MPA (D).The CO error is calculated relative to the data from the full scan duration (4 min 50 s for AAo and 6 min 10 s for MPA).SNR was calculated from the unbinned images, equivalent to the sum of all 25 cardiac frames.Open markers indicate an error less than 5%, and the region highlighted in yellow shows unstable cardiac output measurements.In (C) and (D), the dotted horizontal lines represent the selected SNR threshold, and the dashed vertical lines represent the "stop" message for each healthy volunteer (HV).For 1 subject (HV 2), the optimal SNR threshold would have never been reached in the AAo acquisition.
By applying these optimal SNR thresholds retrospectively (Figure 2C,B), acquisition would have automatically stopped at 2 min 41 s ± 62 s and 2 min 39 s ± 63 s, saving 41% ± 23% and 57% ± 18% of scan time for AAo and MPA, respectively.Compared to the full acquisition, the retrospective stopped acquisition had a CO % error of 1.3% ± 1.6%/1.4% ± 1.1% with a maximum of 5.0%/3.3%for AAo/MPA.For 1 healthy volunteer of BMI = 28.8, the full acquisition (4 min 50 s) in the AAo did not reach the target SNR (max SNR = 164).
Figure 4 compares MPA flow imaging and measurements between a fixed 2-min acquisition and SNR-driven automatic stop acquisition, both compared with the reference full acquisition (6 min 10 s) in 2 different healthy subjects.In one subject, compared with the reference, the fixed 2-min acquisition provided inaccurate flow measurements, and 5 min was the required acquisition time determined by the SNR-driven automatic stop method.In the other subject, 2 min was sufficient, and, in fact, the SNR-driven automatic stop occurred at 1 min 40 s.This illustrates the added value of SNR-driven acquisition to reduce protocol inefficiency.

F I G U R E 3
Example SNR-driven quality control applied inline.The final SNR maps are shown along with the magnitude and phase images after automatic stop, and the resulting flow curves are compared with the full acquisition time; the resulting error in cardiac output (CO) is provided.Results are displayed for the ascending aorta (A,B) and the main pulmonary artery (C,D) of the different healthy volunteers.The contour of the automatic segmentation is displayed in red.

F I G U R E 4
Comparison of phase images and flow measurements between a fixed 2-min acquisition and SNR-driven automatic stop acquisition (SNR threshold = 140 for main pulmonary artery) compared with the reference full acquisition (6 min 10 s).(A) The SNR-driven scan stopped the acquisition at 5 min for a female healthy volunteer (HV) of body mass index (BMI) = 28.8(A), whereas the acquisition stopped at 1 min 40 s for a male HV of BMI = 23.8(B).This illustrates the value of an SNR-driven stop criterion.
As illustrated in Figure 5, the SNR-driven automatic stop was also deployed inline in 1 patient with a prosthetic aortic valve and visible metallic artifacts caused by sternal wires.The acquisition stopped at 2 min with a SNR = 178, and it generated diagnostic flow measurements with a CO % error of 3.7% compared with the full acquisition.The automatic segmentation of the aorta for the inline quality control was robust to an artifact induced by a metallic implant.

DISCUSSION
This study aimed to develop a framework for inline automatic quality control based on a predetermined image-quality metric.We provided an illustrative example of an adaptive subject-specific MRI acquisition time for 2D phase-contrast MR flow measurements in the heart with an SNR-driven stop criterion.We demonstrated a generalizable framework for intermittent closed-loop communication between the image reconstruction software and the data-acquisition software, sending messages about image quality inline to the sequence controller in this application.For pseudo-golden-angle spiral 2D phase-contrast flow, the standard deviation of automatic stop times (±67 s for AAo, ±80 s for MPA) revealed the value of subject-specific acquisition time for consistent image quality.It resulted in saving approximately 50% of acquisition time while ensuring a diagnostic measurement with average error in quantitative measurements less than 2.1%/6.3%for AAo/MPA, which is in the order of variation for scan-rescan (3%) or intersite variability (5%) found in the literature. 21e chose 2D phase-contrast flow measurement as an example application because these scans can be time-consuming for specific clinical indications.For example, during MRI-guided invasive catheterization procedures, flow is measured repeatedly while the patient is instrumented.Similarly, several flow measurements may be required in pediatric and adult patients with congenital heart disease.Therefore, optimal patient-specific scan durations of 2D phase-contrast flow measurements may improve the efficiency of these exams.We used SNR as the image-quality metric for 2D phase-contrast MRI.SNR is directly linked to velocity noise ratio 22 ; therefore, the standard deviation of the velocity (σ v ) is given by the following formula: This formula enables calculation of the confidence interval of the flow measurement 14 and therefore is an Inline automatic quality control of the ascending aorta of a patient, displaying the SNR map and magnitude and phase images after early stop and flow measurements compared with the full acquisition time.The automatic segmentation for the quality control performed well even in the presence of an artifact induced by a sternal wire (blue arrow).CO, cardiac output.appropriate metric for this application.We chose a pseudo-golden-angle spiral sequence to provide flexibility to have fine control over the automatic stopping time, compared with Cartesian imaging, in which a full image average requires almost 1 min.The 2D spiral phase-contrast measurements have previously been validated against Cartesian sequences. 23The SNR thresholds used here were specifically optimized for our sequence and application and should be adapted for each application.
Inline quality control imposed the need for low-latency SNR calculation.For speed purposes, only the SNR of the time-averaged data was reconstructed in our study, as it is correlated (R 2 = 0.99) with the mean cardiac frame-resolved SNR.The pseudo replica method used to estimate SNR is computationally intensive for non-Cartesian imaging and required 12.98 ± 5.49 s; automatic segmentation required 1.08 ± 0.09 s.The required computation time restricted the maximum frequency of intermittent image-quality assessment (fixed at 20 s for our study).Instead, another possible implementation would be to calculate SNR once and extrapolate the predicted SNR versus scan time.
Our implementation of SNR calculation is available open-source using Gadgetron (https://github.com/NHLBI-MR/SNR-driven-flow).We used FIRE, which is a proprietary Siemens package for messaging between the reconstruction and acquisition, but this could be similarly achieved using the Gadgetron streaming capabilities, custom implementation, and/or vendor-provided software.This technology could benefit low-latency offline automatic image-quality control present in the literature [2][3][4][5][6][7]24,25 by enabling rapid inline implementation.
The main limitation affecting our method is that inline quality control is computationally intensive due to the use of the pseudo replica method for SNR map estimation, therefore restraining the deployment of this method in clinical settings with limited resources.It should also be noted that the choice of the clinical stopping criterion (CO % error ≤ 5%) was only intended as a proof of concept to illustrate the concept of an automatic stop when a certain level of diagnostic certainty was achieved.Defining such criterion from widespread application would require a larger cohort with diversity of patient profiles, which is beyond the scope of this study.In addition, the use of a temporal constraint in the image reconstruction may influence the flow measurement, especially when SNR is insufficient.
The concept of automatic scan termination based on a predetermined image-quality metric is widely applicable.Indeed, the inline quality control has been designed for a single-slice acquisition and could be extended to a multislice approach, in which each slice could have its own stopping time (based on a single SNR threshold) to ensure consistent quality across the whole volume.Moreover, the quality control could be extended with multiple targets and image-quality metrics, thereby increasing the complexity to design the stopping criterion.To alleviate this issue, the image quality could be based on a singular quality score or a binary classification (i.e., continue or stopping scanning).However, those approaches will likely require a large database with expert annotations to train a classification network.For other applications, different image-quality metrics may be relevant, such as contrast-to-noise ratio or sharpness metric.We applied this method on a contemporary 0.55T system, in which, given the intrinsically lower SNR, the ability to ensure consistent image quality is desirable.However, this approach is also valuable at other field strengths.Moreover, the closed-loop feedback workflow could also be extended to detect artifacts in real time and correct them by automatically adjusting sequence parameters.For instance, a feedback module that detects velocity aliasing and corrects it in real time may enable subject-specific optimal v enc for our 2D phase contrast application.

CONCLUSIONS
We demonstrated a framework for automatic real-time quality control for subject-specific acquisition timing adaptation and applied it using SNR-driven imaging on phase-contrast MRI as a proof of concept.We observed a wide distribution of automatic stopping times across the population, which revealed the value of subject-specific acquisition time for consistent image quality.