A Deep Learning-Based Approach to Reduce Rescan and Recall Rates in Clinical MRI Examinations.

BACKGROUND AND PURPOSE: MR imaging rescans and recalls can create large hospital revenue loss. The purpose of this study was to develop a fast, automated method for assessing rescan need in motion-corrupted brain series. MATERIALS AND METHODS: A deep learning-based approach was developed, outputting a probability for a series to be clinically useful. Comparison of this per-series probability with a threshold, which can depend on scan indication and reading radiologist, determines whether a series needs to be rescanned. The deep learning classification performance was compared with that of 4 technologists and 5 radiologists in 49 test series with low and moderate motion artifacts. These series were assumed to be scanned for 2 scan indications: screening for multiple sclerosis and stroke. RESULTS: The image-quality rating was found to be scan indication- and reading radiologist-dependent. Of the 49 test datasets, technologists created a mean ratio of rescans/recalls of (4.7 ± 5.1)/(9.5 ± 6.8) for MS and (8.6 ± 7.7)/(1.6 ± 1.9) for stroke. With thresholds adapted for scan indication and reading radiologist, deep learning created a rescan/recall ratio of (7.3 ± 2.2)/(3.2 ± 2.5) for MS, and (3.6 ± 1.5)/(2.8 ± 1.6) for stroke. Due to the large variability in the technologists' assessments, it was only the decrease in the recall rate for MS, for which the deep learning algorithm was trained, that was statistically significant (P = .03). CONCLUSIONS: Fast, automated deep learning-based image-quality rating can decrease rescan and recall rates, while rendering them technologist-independent. It was estimated that decreasing rescans and recalls from the technologists' values to the values of deep learning could save hospitals $24,000/scanner/year.

Concomitant magnetic-field-induced artifacts in axial echo planar imaging.

When a linear magnetic field gradient is used, spatially higher-order magnetic fields are produced to satisfy the Maxwell equations. It has been observed that the higher-order magnetic field produced by the readout gradient causes axial echo planar images acquired with a horizontal solenoid magnet to shift along the phase-encoding direction and lose image intensities. Both the shift and intensity reduction become increasingly severe as the slice offset from the isocenter increases. These phenomena are quantitatively analyzed, and good correlation between experiments and theory has been established. The analysis also predicts a previously unreported Nyquist ghost on images with very large slice offsets. This ghost has been verified with computer simulations. Based on the analysis, several methods have been developed to eliminate the image shift, the intensity reduction, and the ghost. Selected methods have been implemented on a commercial scanner and proved effective in removing these image artifacts.

Concomitant gradient terms in phase contrast MR: analysis and correction.

Whenever a linear gradient is activated, concomitant magnetic fields with non-linear spatial dependence result. This is a consequence of Maxwell's equations, i.e., within the imaging volume the magnetic field must have zero divergence, and has negligible curl. The concomitant, or Maxwell field has been described in the MRI literature for over 10 years. In this paper, we theoretically and experimentally show the existence of two additional lowest-order terms in the concomitant field, which we call cross-terms. The concomitant gradient cross-terms only arise when the longitudinal gradient Gz is simultaneously active with a transverse gradient (Gx or Gy). The effect of all of the concomitant gradient terms on phase contrast imaging is examined in detail. Several methods for reducing or eliminating phase errors arising from the concomitant magnetic field are described. The feasibility of a joint pulse sequence-reconstruction method, which requires no increase in minimum TE, is demonstrated. Since the lowest-order terms of the concomitant field are proportional to G2/B0, the importance of concomitant gradient terms is expected to increase given the current interest in systems with stronger gradients and/or weaker main magnetic fields.

Effect of and correction for in-plane myocardial motion on estimates of coronary-volume flow rates.

The sensitivities of phase-difference (PD) and complex-difference (CD) processing strategies to in-plane motion were examined theoretically and experimentally. Errors in velocity and volume flow rate (VFR) estimates were attributed to (a) motion between different velocity encodings and, in the case of segmented k-space acquisition strategies, (b) motion over the segment duration. PD estimates were found to be insensitive to in-plane motion between velocity encodings, whereas CD VFR estimates were found to be sensitive to this motion. PD estimates, however, were affected by partial volume effects. A corrected CD (CD') scheme was developed that minimizes both partial-volume and in-plane motion effects. Segmented k-space acquisitions with sequential offset and sequential interleaved offset (or centric) phase-encoding schemes were studied. Images obtained using these techniques were found to include both blurring and replication artifacts. The amount of artifact generally increased with the number of views per segment (vps) and the in-plane velocity. PD, CD, and CD' VFR estimates were found to be degraded by these artifacts. The sequential offset phase-encoding scheme generally had acceptable VFR errors (at 4 vps. a CD' VFR error of 7.0%) when averaged over the physiologic range of myocardial motion (> 12 cm second-1); however, larger errors were observed outside this range. VFR estimates obtained using the sequential interleaved phase-encoding scheme at 4 vps were unacceptable. More accurate VFR measurements were obtained using a revised segmented PC strategy, which reversed the order in which the velocity and phase encodings were interleaved. The weighted average CD' VFR error obtained using the revised strategy was 24.5% (for 4 vps). Using displacement information obtained from the two velocity-encoded images, an estimate of the in-plane velocity was obtained and used to correct the acquired data. This decreased the VFR error (weighted average CD' error at 4 vps decreased from 24.5% to -6.3%); however, the implemented correction algorithm could potentially introduced other artifacts in the images.

Measurement of coronary blood flow and flow reserve using magnetic resonance imaging.

PURPOSE: It was the purpose of this study to demonstrate the feasibility of performing coronary artery flow and coronary flow reserve (CFR) measurements in normal human volunteers using a magnetic resonance (MR) phase contrast technique. MATERIALS AND METHODS: Coronary flow rate, flow velocity, peak flow and CFR were determined at rest and during pharmacologically induced hyperemia in 10 healthy volunteers. The flow measurements were obtained during a single breath-hold by using a fast, prospectively gated, segmented k-space gradient-echo phase contrast acquisition with view sharing (FASTCARD PC) that was modified to improve sampling of the diastolic flow. Data were processed using the standard phase difference (PD) processing techniques as well as a new complex difference (CD) flow measurement method intended to improve the accuracy of flow measurements in small vessels. RESULTS: Mean hyperemic flow velocity (40 +/- 16 cm/s) and blood flow (3.9 +/- 1.5 ml/s) rates differed significantly from resting velocity (13 +/- 6.6 cm/s) and flow (1.1 +/- 0.4 ml/s) measurements (p < 0.0001). PD methods consistently measured larger flow rates at rest (24% larger, p < 0.0005) and stress (29% larger, p < 0.0001). CFR, calculated as the ratio of the mean PD flows (4.7 +/- 2.8), was higher than CFR calculated as the ratio of mean CD flows (4.2 +/- 1.8); however, the differences did not reach statistical significance (p = 0.07). Flow measurements performed in adjacent slices of the same vessel correlated well (r = 0.88). CONCLUSIONS: Coronary flow and CFR measurements using the MR techniques are feasible and are similar to those reported in the literature for healthy volunteers.

Frequency response of multi-phase segmented k-space phase-contrast.

A theoretical analysis of the temporal frequency response of multi-phase segmented k-space phase-contrast was developed. This includes the effects of both segment duration and the number of cardiac phases that are reconstructed. An increase in the number of views per segment and the corresponding increase in segment duration results in an increased smoothing or low-pass filtering of the time-resolved flow waveform. Reconstruction of all intermediate cardiac phases makes the Nyquist sampling frequency independent of the number of views per segment. This analysis was verified experimentally using a multi-phase phase-contrast segmented k-space MR pulse sequence. This sequence reconstructs all intermediate cardiac phases and uses fractional segments at the end of the cardiac cycle if an entire segment does not fit. The use of fractional segments increases the portion of the cardiac cycle over which data are acquired.

Effects of through-plane myocardial motion on phase-difference and complex-difference measurements of absolute coronary artery flow.

We have previously reported on a complex-difference (CD) flow measurement technique that produces more accurate results than the phase-difference (PD) flow measurement technique due to the greater immunity of the former method to partial volume effects. We report here on some of the ways in which through-plane myocardial motion affects the accuracy of absolute coronary artery flow measurements obtained using the PD and CD techniques. We also discuss motion correction schemes that can be applied to the PD and CD processing methods to improve their accuracy. Computer simulations have been performed to assess the magnitude of the errors associated with these flow measurement techniques when they are applied to small vessels that are attached to a moving background. Laminar and plug flow, with and without complete background suppression, have been considered. Experiments with a moving vessel phantom have been conducted to test the performance of the PD and CD flow measurement techniques in circumstances similar to those simulated. The simulations and the experiments showed that, after corrections for through-plane motion are made, the CD method generally yields more accurate flow results than the PD method. As shown by the simulations, however, both methods yield compromised results due to subtle saturation effects that occur when the direction of myocardial motion is opposite the direction of blood flow. Unvalidated PD and CD measurements of coronary artery flow waveforms in human volunteers are presented to illustrate the magnitude of the proposed through-plane motion effects in vivo.

Respiratory blur in 3D coronary MR imaging.

3D MR imaging of coronary arteries has the potential to provide both high resolution and high signal-to-noise ratio, but it is very susceptible to respiratory artifacts, especially respiratory blurring. Resolution loss caused by respiratory blurring in 3D coronary imaging is analyzed theoretically and verified experimentally. Under normal respiration, the width for any Gaussian point spread function is increased to a new value that is at least several millimeters (about 3-4 mm). In vivo studies were performed to compare respiratory pseudo-gated 3D acquisition with breath-hold 2D acquisition. On average, the overall quality of a pseudo-gated 3D image is worse than that of the corresponding breath-hold 2D image (P = 0.005). In most cases, respiratory blur caused coronary arteries in pseudo-gated 3D data to have lower resolution than in breath-hold 2D data.

A complex-difference phase-contrast technique for measurement of volume flow rates.

Magnetic resonance (MR) phase-difference methods work well for measuring volumetric flow rates when the vessel diameter is large compared with the in-plane voxel dimensions. For small vessels (eg, coronary arteries), partial-volume effects introduce substantial errors in the measured volume flow rate. To correctly measure flow rates through a voxel, both the fraction of the voxel containing moving spins and the phase shift imparted to those spins must be known. The authors propose a flow measurement method that combines information obtained with both the complex-difference and phase-difference processing techniques and thereby provides the fractional volume occupied by the moving spins and the phase of those spins. The complex-difference flow map method proposed results in improved accuracy of MR phase-contrast flow measurements in the presence of partial-volume effects.

Coronary MRI with a respiratory feedback monitor: the 2D imaging case.

The inability to return the heart to the same position for all breath-holds during 2D coronary MR imaging can result in imaging different locations than desired. This can lead to problems such as (i) missing a whole vessel, or a part of it, (ii) misaligning segments of vessels imaged in different breath-holds, and (iii) degrading image quality when a single slice is acquired in multiple breath-holds. To reduce inconsistencies in the breath-hold level, we designed a respiratory feedback monitor (RFM) that uses a bellows to monitor the circumference of the subject's chest. When the circumference of the subject's chest is within preset limits, an audio signal alerts subjects to hold their breath at that position. Use of the RFM significantly reduces the problems caused by inconsistent breath-holds and the number of breath-holds for an examination in 2D coronary MR imaging.

MR angiography using velocity-selective preparation pulses and segmented gradient-echo acquisition.

We describe a cardiac-gated MR angiographic imaging method that employs velocity-selective preparation (VSP) pulses in conjunction with segmented gradient-echo acquisition and subtraction to produce images that, ideally, contain no signal from stationary tissues and display vessels with a signal intensity that is dependent on the velocity of the blood in the vessels. The novel features of this method are a) it acquires several phase-encoding values/application of a single VSP pulse, b) it uses subtraction to eliminate signal that is not sufficiently suppressed by the VSP pulses, and c) it uses VSP pulses that are synchronized with the cardiac cycle so it can be used to produce ghost-free images of pulsatile blood. An advantage of this sequence is that it detects a signal that, after preparation, is relatively unaffected by changes in blood velocity. This leads to a large signal-to-noise ratio for all the phase-encoding values, a reduction of ghosting artifacts, and the ability to visualize blood that is in motion for only a short time during the cardiac cycle. Because the signal is prepared during peak flow, venous signal can be suppressed by making the sequence sensitive to high velocities. An additional advantage of this sequence is that it permits sampling with a short TE because the velocity-encoding gradient can be applied in a preparatory interval. Signal loss that results from dephasing during the longer TE preparation interval can be reduced or eliminated by allowing the dephased spins to flow out of the region of complex flow, and perhaps out of the field-of-view, by introducing a delay between the finish of the VSP pulse and the beginning of data acquisition.

Effect of elastase and ventilation on elastic recoil of excised dog lungs.

The effect of porcine pancreatic elastase and mechanical ventilation on tissue elastic recoil was examined in excised dog lung lobes. Lobes incubated for one hour with an elastase-buffer mixture showed a significant (P less than 0.001) left shift of the liquid-filled pressure-volume curve at all pressures measured (0 to 12 cm H2O) when compared to lobes treated with buffer only. These results suggest that the contribution of elastin to the elastic properties of lung tissue is greatest at mid-lung volumes, but that it also contributes to delimiting maximal lung volume. Elastase and buff-treated lobes were inflated cyclically with humidified air to a pressure of 20 cm H2O 6 times per min during a 16-hour period. This mechanical ventilation caused no further decrease of tissue elastic recoil. Ventilation did cause an unexpected increase in the elastic recoil of liquid-filled lobes that was significant at pressures of 4 cm H2O (P less than 0.025) or more (P less than 0.001). Elastase and buffer-treated lobes showed an almost identical rightward shift of the pressure-volume curve after ventilation when compared to the respective nonventilated control lobes. This increased recoil cannot be attributed to altered surface tension.