Exploring the Acceleration Limits of Deep Learning Variational Network–based Two-dimensional Brain MRI

Abstract

Purpose

To explore the limits of deep learning–based brain MRI reconstruction and identify useful acceleration ranges for general-purpose imaging and potential screening.

Materials and Methods

In this retrospective study conducted from 2019 through 2021, a model was trained for reconstruction on 5847 brain MR images. Performance was evaluated across a wide range of accelerations (up to 100-fold along a single phase-encoded direction for two-dimensional [2D] sections) on the fastMRI test set collected at New York University, consisting of 558 image volumes. In a sample of 69 volumes, reconstructions were classified by radiologists for identification of two clinical thresholds: (a) general-purpose diagnostic imaging and (b) potential use in a screening protocol. A Monte Carlo procedure was developed to estimate reconstruction error with only undersampled data. The model was evaluated on both in-domain and out-of-domain data. The 95% CIs were calculated using the percentile bootstrap method.

Results

Radiologists rated 100% of 69 volumes as having sufficient image quality for general-purpose imaging at up to 4× acceleration and 65 of 69 volumes (94%) as having sufficient image quality for screening at up to 14× acceleration. The Monte Carlo procedure estimated ground truth peak signal-to-noise ratio and mean squared error with coefficients of determination greater than 0.5 at 2× to 20× acceleration levels. Out-of-distribution experiments demonstrated the model’s ability to produce images substantially distinct from the training set, even at 100× acceleration.

Conclusion

For 2D brain images using deep learning–based reconstruction, maximum acceleration for potential screening was three to four times higher than that for diagnostic general-purpose imaging.

Keywords: MRI Reconstruction, High Acceleration, Deep Learning, Screening, Out of Distribution

Supplemental material is available for this article.

© RSNA, 2022

Summary

Evaluation of a two-dimensional brain MRI reconstruction neural network suggested that maximum possible acceleration was three to four times higher than that for diagnostic general-purpose imaging when the purpose is restricted to screening for major disease or structural abnormality.

Key Points

• ■ In a sample of two-dimensional T1-weighted fast spin-echo, T2-weighted, and T2 fluid-attenuated inversion recovery brain MRI scans from 69 patients, 69 scans (100%) were deemed acceptable for general-purpose imaging at up to 4× acceleration; if the purpose of imaging was narrowed to screening, 65 of 69 scans (94%) were deemed acceptable at up to 14× acceleration.

• ■ The Monte Carlo procedure was used to estimate ground truth metrics of peak signal-to-noise ratio and mean squared error with coefficients of determination greater than 0.5 at 2× to 20× acceleration levels using only undersampled data.

• ■ In an experiment on dataset diversity and out-of-distribution samples, the trained model provided reconstructions distinct from the training data, suggesting that the models act as complex projectors rather than as memorizers of the training set, even at acceleration factors up to 100×.

Introduction

Over the past several years, there has been a growing interest in using deep learning–based reconstruction to accelerate MRI acquisition (1–4). Reducing acquisition time provides many potential benefits: maximizing patient comfort, increasing access, improving scan quality in children and other individuals for whom holding still for a long time is difficult, minimizing sedation and anesthesia (5), reducing motion-related artifacts, and providing overall higher-quality diagnostic images. Many approaches have been developed over the past few decades, including novel acquisition strategies (6,7), parallel imaging (8–10), and compressed sensing (11). Commercial prototypes and products using deep learning for reconstruction are currently under development and are available in the market across the industry. To date, investigations using deep neural network (DNN)–based image reconstruction have applied advanced models and substantial computational power to achieve modest gains in acceleration (eg, moving from twofold accelerations commonly afforded by parallel imaging to fourfold accelerations with DNNs).

Here, we studied the limits of deep learning–based reconstruction by examining reconstructions across a broad range of accelerations using both quantitative metrics and qualitative expert assessments. Specifically, we enlarged and adapted the End-to-End Variational Network (12) to reconstruct standard clinical two-dimensional (2D) spin-echo brain MR images (fast spin-echo T1-weighted, fast spin-echo T2-weighted, and T2 fluid-attenuated inversion recovery [FLAIR] images) over a range of undersampling rates at up to 1/100 of fully sampled levels. We restricted our investigation to accelerations along a single phase-encoding direction, for which image quality is known to degrade rapidly with conventional techniques even at modest acceleration levels. Our goal was to explore the properties of deep learning–based reconstructions throughout the regimen; identify potential use cases, such as short-scan diagnosis screening beyond general imaging; and expose points of fallibility.

Materials and Methods

Dataset

This was a retrospective study performed with local institutional review board approval. Written informed consent was waived owing to the minimal risk and retrospective nature of the study and the use of de-identified data. The data used in this study are the brain subset of the fastMRI dataset collected at New York University in 2018; data were de-identified by New York University (13,14). In the current study, analysis was performed on the test set, whereas previous analysis was performed on the challenge set (14). Following standard machine learning practice, models were not applied on the test set until all model development was complete. The dataset includes raw k-space data from 6405 fully sampled clinical brain MR images (including normal and abnormal clinical studies) obtained using standard 2D imaging sequences on multiple imagers at multiple imaging sites within the one contributing academic center. Imagers include 3-T (Magnetom Prisma, Magnetom Skyra, Biograph mMR, Magnetom Tim Trio; Siemens Healthcare) and 1.5-T (Magnetom Avanto, Magnetom Aera; Siemens Healthcare) units. Axial 2D fast spin-echo T1-weighted (both noncontrast and contrast-enhanced), 2D fast spin-echo T2-weighted, and 2D T2 FLAIR images were used. Imaging parameters varied depending on the imager but followed standard clinical protocols in all cases. Imaging parameters (repetition time msec/echo time msec) for T1-weighted, T2-weighted, and T2 FLAIR imaging were 250/2.43, 6000/113, and 9000/81, respectively, for 3-T MRI, and 461/9.4, 5120/102, and 9000/86, respectively, for 1.5-T MRI. Section thickness was 3–5 mm, and field of view range was 200–220 mm, depending on patient head size. We used the standard fastMRI brain data splits, which include a training set (4469 volumes), validation set (1378 volumes), and test set (558 volumes), with no patient overlap between the splits. We tuned model parameters on the validation set. After standard machine learning practice, we used a fully sampled version of the test set for our evaluations.

Model Architecture and Training

To improve performance, we introduced several changes to the End-to-End Variational Network (12), enlarging it from 29.5 million to 231 million parameters. Architectural modifications included the following: (a) using all k-space data instead of only the densely sampled center to estimate coil sensitivities; (b) using a DenseNet-inspired (15) skipped connection to the image reconstructed from zero-filled k-space for all layers in the cascade; (c) increasing the number of cascades to 18 and increasing the size of each cascade U-Net to have 20 channels with five pooling layers; and (d) using a residual U-Net (16) rather than a standard U-Net in each layer. We call this model a zero-filled skipped connection network (ZSNET). A block diagram of the model is shown in Figure E1 (supplement). We trained for accelerations at 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 24, 32, 48, 64, and 100 times. The training procedure is described in Appendix E1 (supplement).

Quantitative and Qualitative Evaluation

We evaluated the models by calculating the structural similarity index (SSIM) (17) and peak signal-to-noise ratio (PSNR) between the images reconstructed by fine-tuned models and the ground truth images in the fastMRI test set. We used the volume variants of both metrics as implemented in the fastMRI GitHub repository. To further evaluate whether the model was merely memorizing training data at high accelerations, we performed a separate analysis in which we ran our model on out-of-distribution data, including a sample from the validation set of the fastMRI knee data (18) and a sample of pure noise.

Radiologist Evaluation

For radiologist evaluation, we reviewed a subgroup of 69 volumes (29 T2-weighted, 20 T2 FLAIR, and 20 T1-weighted examinations) from the test set with a balance of normal and abnormal cases for each contrast type. Volumes were classified as normal if either no abnormality was present or if white matter changes consistent with mild microvascular change were present on the imaging volume. Abnormal volumes were selected from the larger pool to include a broad variety of abnormal conditions, including acute and chronic ischemia, hemorrhage, intra- and extra-axial masses, extra-axial collections, hydrocephalus, postoperative change, white matter disease more severe than mild microvascular change, leptomeningeal disease, vascular malformation, and mass effect. Abnormal cases included both low- and high-contrast lesions. Two neuroradiologists with 20 (Y.W.L.) and 5 (E.L.) years of experience independently reviewed ground truth reconstruction and deep learning reconstructions at all accelerations. In deep learning reconstructions, dithering (ie, addition of low levels of noise) was applied prior to presentation; this has been shown to be helpful for radiologist evaluation (19).

Radiologists used a modified simultaneous double stimulus presentation method previously described for the evaluation of a continuum of image examples for threshold determination of image quality (20) to categorize reconstructions into one of three levels separated by two thresholds. Level 1 was the highest acceleration level with acceptable clinical diagnostic quality, defined as image quality undifferentiable from state-of-the-art routine clinical diagnostic imaging currently obtained at an academic tertiary care center based on subspecialist neuroradiology review. Features include preserved tissue contrast appropriate to each pulse sequence; accuracy of anatomic detail of both intra- and extracranial features, including faithful rendering of any abnormalities present; preserved gray-white differentiation; and lack of any major artifacts, such as the introduction of any hallucinatory features or noticeable blurring. Level 2 was the highest acceleration acceptable if the images were to be used as a screening method to identify major intracranial disease or structural abnormality, such as mass, herniation, or hydrocephalus. Level 2 screening would need to avoid gross pseudonormalization of any abnormalities present, establish sufficient concern about any abnormalities to refer patients appropriately for further imaging, and avoid the introduction of any problematic hallucinations or oversimplification of the gyral pattern. Level 3 included all other reconstructions. Unless noted otherwise, disagreements between radiologists concerning threshold accelerations were resolved by choosing the more conservative acceleration.

Monte Carlo Metric Estimation

We developed a Monte Carlo procedure to estimate reconstruction error using only undersampled data to address the limitations of quantitative image quality assessment without ground truth available in clinical scenarios. At a high level, our procedure assumes that the real-valued DNN reconstruction is the ground truth image and simulates the forward sampling and reconstruction process to estimate the variation between estimates produced by the DNN model and the actual ground truth. This forward simulation is repeated several times with different sampling patterns and noise realizations. Using the multiple simulations, it is possible to compute estimates of ground truth reconstruction statistics. Because the simulation is nonideal, we estimate correction factors for the metrics estimation using hold-out validation data. The procedure is described in detail in a diagram (Fig E2 [supplement]).

Statistical Analysis

For the quantitative evaluation, we calculated mean metrics for each acceleration and contrast. We also calculated 95% CIs on the mean using the percentile bootstrap method (21) with 10 000 bootstrap iterations. We chose to use the percentile bootstrap owing to the nonnormal distribution of SSIM scores and intuitive relationship between the percentile bootstrap and patient selection. Processing was performed using a custom script in Python (version 3.8; Python Software Foundation; https://www.python.org/). For the radiologist evaluation, intraclass correlation coefficient and mean absolute deviation of scores between readers were calculated to measure interreader variability. For the radiologist evaluation, statistical analysis was performed using R software (version 3.6.0; The R Project for Statistical Computing).

Results

Qualitative Image Overview

Reconstructed images decayed in quality as the acceleration rate increased, as shown in Figure 1, which includes examples of a normal case and two abnormal cases. The degree of image quality degradation was dependent on image content. For the normal case, T2-weighted image quality is relatively stable throughout the acceleration regimen, with blurring gradually introduced at very high accelerations (Fig 1A). Similar behavior is observed in one of the abnormal cases, with a large amount of vasogenic edema observed on the T2-weighted images (Fig 1B); however, there is some pseudonormalization of mass effect at very high acceleration levels. Pseudonormalization indicates an effect of the model that makes an abnormal brain appear normal. In the abnormal case with a complex ring-enhancing mass (Fig 1C), the abnormality appears to be gradually erased with increasingly accelerated T1-weighted reconstructions, becoming more pseudonormalized by 100-fold acceleration.

Figure 1:

Sample reconstructed brain MR images at various acceleration levels, showing (A) normal and (B, C) abnormal examples. In each panel, deep neural network (DNN)–based reconstructions are shown above corresponding baseline reconstructions using simple zero-filling and Fourier inverse transformation, and an unaccelerated ground truth (GT) image is shown at the far left. (A) Reconstruction quality is steady in axial T2-weighted images throughout the acceleration range, with gradual blurring introduced as the acceleration rate reaches 100. (B) Gross features of vasogenic edema are also relatively well maintained in axial T2-weighted images across accelerations, with some loss of mass effect visualization at higher accelerations. (C) For some cases, abnormal features can be lost as acceleration increases, as in the axial T1-weighted images with a complex peripherally enhancing mass that begins to disappear at 24×.

We observed that reconstructions could exhibit complex interactions with image features depending on contrast and acceleration rate (Fig 2). Other examples are shown in Figure E6 (supplement). A T2 FLAIR example in Figure 2A shows blurring and, at high acceleration levels, a simplification of the gyral and sulcal contours of the brain, giving this normal brain a pseudolissencephalic appearance. A T2-weighted example in Figure 2B also shows smoothing, in this case in a patient with vasogenic edema from multiple masses, with masses replaced by hallucinatory prominent sulci along the medial margin of the superior frontal gyrus at high acceleration. In another T2-weighted example (Fig 2C), multiple faint white matter lesions are blurred into horizontal bands at 20× and are completely lost by 100×. All cases with the described problematic features were excluded by radiologists from both level 1 and level 2.

Figure 2:

Axial brain MR images of feature alteration across increasing acceleration. (A) A normal T2 fluid-attenuated inversion recovery image with increasing blurring at higher accelerations, with a final simplification of the sulcal and gyral pattern creating a pseudolissencephalic appearance of the brain at 100× acceleration. (B) A T2-weighted image with two T2 hyperintense lesions in the left superior frontal gyrus that become incorporated into left frontal paramedian sulci by 100×. (C) A T2-weighted image with many faint hyperintense lesions in the corona radiata white matter bilaterally that are smoothed at higher acceleration levels so that they are imperceptible by 20× acceleration. GT = ground truth.

Quantitative Evaluation

When fine-tuned for standard equispaced sampling patterns, ZSNET achieved 0.9595 SSIM at 4× and 0.9433 SSIM at 8× on the fastMRI public leaderboard (versus 0.9591 and 0.9426 for the original End-to-End Variational Network [12]), placing it in second place for both accelerations at the time of upload.

Figure 3 shows quantitative metrics calculated on the test set broken down by contrast with 95% CIs. Image quality decayed monotonically with increasing acceleration, although the pace of image quality decay shifted between 18× and 20× acceleration, which corresponds to the point at which no outer k-space lines remained in the sampled data. The 95% CIs were concentrated tightly around the mean, with the largest spread occurring with T2 FLAIR images, likely owing to the smaller number of T2 FLAIR cases.

Figure 3:

Average (A) structural similarity index (SSIM) and (B) peak signal-to-noise ratio (PSNR) across accelerations as calculated on the test set. The 95% CIs of the mean were calculated using the percentile bootstrap method. Metrics decay monotonically with increasing acceleration in two distinct regimens. The transition between standard acceleration and superresolution regimens can be observed around 18×–20× acceleration. Mean SSIM and PSNR exhibit tight CIs, but variation can be larger on a case-by-case basis. Tabular data for SSIM and PSNR results are shown in Tables E2 and E3 (supplement). FLAIR = fluid-attenuated inversion recovery.

Radiologist Evaluation

An overview of images at different threshold accelerations is shown in Figure 4, with an extended version shown as Figure E4 (supplement). Threshold levels are defined as the maximum acceleration rate before an image is no longer classified at the given level, with level 1 being appropriate for general purpose diagnostic imaging and level 2 having screening potential. The images exhibited great diversity in their qualitative properties depending on the contrast and pathologic feature of concern. Some abnormalities, such as the occipital lobe mass in Figure 4D, were observable even at reconstructions up to 64× acceleration. In other cases, deep learning–based reconstruction with undersampled data seemed to improve image quality over ground truth at modest acceleration rates, as illustrated in Figure 4B, in which motion artifacts present in the ground truth image are reduced at 6× acceleration.

Figure 4:

Examples of axial reconstructed brain MR images at threshold levels, that is, the upper-bound accelerations at which images were felt to be satisfactory for general-purpose diagnostic imaging (level 1) and potential screening (level 2) as determined by consensus review of two neuroimaging experts. (A) A T1-weighted postcontrast image with a cavitary area in the right frontal subcortical white matter that is reasonably rendered at 4× acceleration and would certainly be detected as abnormal at 18× acceleration. (B) A T2 fluid-attenuated inversion recovery (FLAIR) image showing mild ventriculomegaly and nonspecific patchy periventricular white matter T2-weighted hyperintensities with motion artifact present in the ground truth (GT). Motion artifact is reduced at 6× with arguably improved overall clinical image quality, but at even higher accelerations, unacceptable blurring occurs. (C) A T2 FLAIR image showing a large right frontal convexity meningioma with mass effect effacing the right lateral ventricle and midline shift. At 6× acceleration, the heterogeneity and overall signal of the mass is well observed, as are the cerebrospinal fluid cleft and displaced surface veins, characteristic of an extra-axial mass, and the midline shift and mass effect. (D) A T2-weighted image with a complex cystic occipital lobe mass is accurately rendered at accelerations up to 10× and clearly present at accelerations up to 64×, which would trigger further investigation in any potential screening scenario.

Table 1 shows the percentage of MR images classified to each clinical level across accelerations (level 2 inclusive of level 1). Radiologists found all images through 4× acceleration to be of sufficient quality for general-purpose diagnostic imaging. Above 4× acceleration, radiologist scores of the images decreased; no images at 14× acceleration or above for any contrast were deemed sufficient for general diagnostic imaging. Radiologists did, however, rate all images through 6× acceleration sufficient for potential screening purposes, and 65 of 69 volumes (94%) were deemed sufficient for screening up to 14× acceleration. Some images were rated sufficient for screening at accelerations up to 100×. Interreader variability measurements showed an intraclass correlation coefficient of 0.875 and mean absolute difference of 0.7 for level 1 and 5.4 for level 2.

Table 1:

Study Results for Classifying MR Images Appropriate for General-Purpose Imaging (Level 1) or Potential Screening (Level 2)

Table 2 shows average threshold accelerations for general-purpose diagnostic quality images (level 1), as well as potential screening use (level 2) for normal and abnormal cases considered in the study; in cases of disagreement between radiologists, an average score was calculated. Maximum accelerations were highest for T2-weighted images for both level 1 and level 2.

Table 2:

Threshold Accelerations Appropriate for Diagnostic Imaging (Level 1) and Potential Screening (Level 2) Obtained from Radiologist Evaluation, Separated into Normal and Abnormal Studies

Quantitative metrics were cross tabulated with the threshold levels for levels 1 and 2 (taking the minimum from disagreeing radiologists) and are shown in Figure E3 (supplement). Metrics trended higher with more stringent diagnostic thresholds (ie, SSIM was higher for level 1 T2-weighted scans than for level 2 T2-weighted scans), but there was substantial overlap between groups at many accelerations.

Monte Carlo Estimation

Results of the Monte Carlo procedure to estimate degree of reconstruction error are shown in Figure 5. Monte Carlo estimates correlated with ground truth metrics computed from fully sampled data. Overall, R² for both PSNR and mean squared error (MSE) were above 0.5 across all accelerations. Monte Carlo estimates were most accurate for MSE at accelerations above 2× with R² of 0.8 up to 10× acceleration. Estimates were least accurate for SSIM, but even SSIM estimates had R² values above 0.2 throughout the acceleration range.

Figure 5:

Results of Monte Carlo reconstruction error estimation showing identity lines for (A) structural similarity index (SSIM), (B) peak signal-to-noise ratio (PSNR), and (C) mean squared error (MSE). Each dot represents an individual volume from the test set, color-coded based on acceleration level (A [up to 18]). (D) The Monte Carlo procedure provided relatively high-quality estimates of ground truth PSNR and MSE using only undersampled data, with R² above 0.5 for the entire range. Monte Carlo estimates were most accurate for MSE for the majority of accelerations. Monte Carlo estimates of SSIM decayed with increasing acceleration, exhibiting wide tails at high accelerations. In principle, such an approach could be used to evaluate the quality of reconstructed images; however, results shown in Figure E3 (supplement) suggest that further work is necessary to relate quantitative metrics to clinical use thresholds.

Handling Anatomic Diversity and Out-of-Distribution Testing

Despite the extremely high accelerations, ZSNET was able to retain coarse morphologic features unique to an individual across a variety of brains. Figure 6 includes a sample of images from three different individuals representing a spectrum of different brain morphology. An extended version of Figure 6 is shown as Figure E5 (supplement). Major individual variations are generally preserved, even out to 100× acceleration.

Figure 6:

MRI reconstructions from applying the zero-filled skipped connection network (ZSNET) to in-distribution and out-of-distribution cases. Rows 1-3: Axial T2-weighted images from brain samples of the fastMRI dataset. Even out to 100× acceleration, ZSNET retains the gross morphologic features unique to each individual brain, accurately capturing variations in anatomy, including volume loss and asymmetry of the calvarium. Row 4: Coronal T1-weighted images from knee samples of the fastMRI dataset. When applied out-of-distribution on knees, ZSNET gives recognizable reconstructions at initial low accelerations. At high acceleration levels, ZSNET reconstructs some gross coarse features, but with increasing alteration to texture. Bottom image: When applied to pure noise, ZSNET removes the noise.

Out-of-distribution results are also shown in Figure 6. Although the model was trained on only brain data, it was able to reconstruct knees at low accelerations, retaining at least the overall morphology of the original input image. However, higher accelerations resulted in altered texture. In terms of quantitative performance, running ZSNET on the fastMRI knee test set, ZSNET showed an expected decrease in performance compared with the End-to-End Variational Network trained on appropriate knee MR images, with SSIM of 0.8942 at 4× acceleration and 0.8436 at 8× acceleration (compared with 0.9302 and 0.8920, respectively, for the End-to-End Variational Network [12]).

Discussion

We examined the capabilities of a deep learning–based MR image reconstruction model, ZSNET, at acceleration levels from 2× to 100×, and evaluated image quality at two levels: (a) acceptable image quality for diagnostic imaging and (b) potential screening applications for fast MRI. Our examination found that 65 of 69 cases (94%) were acceptable at 14× acceleration to identify gross pathologic features and flag them in a clinical setting for return and diagnostic imaging. The expert evaluation portion of the study was enriched to include over 50% abnormal cases.

Our study design was motivated by insights obtained from the 2019 (22) and 2020 (14) fastMRI reconstruction challenges. These challenges found that at high acceleration rates, deep learning compressed sensing reconstruction methods could reconstruct images that were visually indistinguishable from unaccelerated data acquisitions and achieved high ratings of quantitative metrics, such as root MSE to the ground truth or SSIM, but suffered from missing pathologic features, as illustrated by the pseudonormalization effect in this study. Pseudonormalization was more severe for subtle abnormalities. This has become a known concern in the broader field of deep learning compressed sensing reconstruction.

In our study, acceleration was surprisingly advantageous in some scenarios in which the ground truth, fully sampled reconstruction, was afflicted by motion artifacts. In such cases, modest acceleration rates up to approximately 6× improved image quality from a clinical perspective. At extreme acceleration rates, however, the pseudonormalization property of the model began to take effect, as observed in abnormal cases. This led to radiologists down-rating cases at these levels. For diagnostic quality–level imaging, our investigation confirms previous studies that images from 2D acquisitions are acceptable up to approximately 4× acceleration (14,19,22), and moving beyond this acceleration level likely requires advances in 2D reconstruction performance or alternative acquisition strategies. In practice, this translates to reduced scan time for axial T2-weighted, axial T2 FLAIR, and axial T1-weighted brain imaging performed with a 3-T MRI machine from 2 minutes 47 seconds, 4 minutes 31 seconds, and 1 minute 21 seconds with a fully sampled k-space to 42 seconds, 1 minute 8 seconds, and 20 seconds with 4× acceleration, respectively.

Our study shows that even for extreme acceleration rates, our model was able to reconstruct the coarse morphologic features of ground truth images. This could raise concern that some form of memorization is taking place at these accelerations. However, the preservation of anatomic features unique to each individual brain as well as the comparatively faithful reconstruction of knee and noise images in Figure 6 suggest an alternative perspective: that variational networks are complex projectors of the input data, corroborating other studies on transfer learning (23) and studies claiming that neural networks are extrapolators (24). One form of this projection appears to be pseudonormalization. Pseudonormalization has also been observed in other studies on self-supervised learning and applied to the task of anomaly detection (25). Further understanding of this phenomenon is a topic for future research.

From a practical perspective, screening 2D fast spin-echo acquisitions through the brain at 16× acceleration could be accomplished in 8–22 seconds per sequence; 2D fast spin-echo fully sampled acquisition times for brain imaging vary from approximately 2 to 6 minutes depending on contrast, magnet strength, and platform. This suggests that a three-to-four series protocol could be obtained for screening in approximately 1 minute, within the timescale of CT while providing the multiple soft-tissue contrasts of MRI, a potentially valuable contribution. Ultimately, extreme acceleration such as 100× may not add clinical value as gains in time become incremental; however, rapid imaging approaches do have the potential to facilitate new modes for dynamic imaging. Here, we isolated the effect of undersampling in the context of a deep learning–based reconstruction model; thus, the study uniquely examined one approach to a faster MRI protocol. Other methods explored for screening-type studies include abbreviated protocols (26,27), which typically remove a subset of pulse sequences. The approach described here could be applied to cases for which it is desirable to preserve image contrasts. Alternatively, our approach could be combined with others for further acceleration. To facilitate reproducibility, our open-source code is available in GitHub (https://github.com/facebookresearch/fastMRI).

Our study had some limitations. The dataset was from one institution and one vendor with no external testing and included limited sequences (ie, T2 weighted, T2 FLAIR, and T1 weighted). The data came from limited section thickness (3–5 mm). Additionally, our sampling pattern was retrospective; therefore, we were unable to measure potential alterations in spin-echo sequence contrast found in prospectively obtained images. However, it is important to note that although undersampling was performed retrospectively, it was performed on actual raw, multireceiver channel k-space data provided in the fastMRI dataset, which was obtained directly from MRI scanners. Therefore, in contrast to retrospective studies that simulate k-space data from already reconstructed images, the image data that are considered ground truth in the fastMRI dataset (used in our study) are not precompressed or filtered (28). Another limitation was that despite the higher screening threshold, we found that metrics were generally not discriminative between potential screening and general-purpose imaging.

In conclusion, we investigated DNN-based reconstructions across a broad range of accelerations, including an extremely high acceleration regimen. We identified a gap between acceleration thresholds acceptable for general-purpose diagnostic imaging and acceleration thresholds considered acceptable for potential screening. In a potential screening or fastMRI protocol setting, we found an additional three- to fourfold acceleration increase over DNN-based acceleration suitable for diagnostic-quality images. We presented a procedure used to estimate reconstruction error using only subsampled data and found that it produced reasonable estimates of PSNR and MSE. Further work is necessary to develop metrics that capture all features of interest to radiologists. Recent progress has been made using pixelwise error bars (29) and neural networks (30–32) that may have the potential to fulfill this task, and our Monte Carlo procedure could be complementary to such efforts. Last, we observed that throughout the high-acceleration regimen, the variational network under consideration did not output samples from the training distribution but rather applied complex projections on the input data.