Clinical Implementation of Deep Learning MR reconstruction for TSE Sequences: Reduction of Acquisition Time and Maintenance


 Background: The application of Deep Learning (DL) in MR image reconstruction is increasingly gaining attention due to its potential of increasing image quality and reducing acquisition time. However, the technology hasn’t been yet implemented in clinical routine. The aim of this study was therefore to describe the implementation of this novel DL image reconstruction for turbo spin echo (TSE) sequences in clinical workflow including a thorough explanation of the required steps and an evaluation of the obtainable image quality compared to conventional TSE.Methods: DL image reconstruction using a variational network was clinically implemented to enable acquisition of accelerated TSE sequences. After internal review board’s approval and informed consent, 30 examinations for knee, shoulder, and lumbar spine in 15 volunteers at 3 T were included in this prospective study. Conventional TSE sequences (TSE) and TSE with deep learning reconstruction (TSEDL) were compared regarding overall image quality, noise, sharpness, and subjective signal-to-noise-ratio (SNR), as well diagnostic confidence and image impression. Comparative analyses were conducted to assess the differences between the sequences. A survey on technologists’ acceptance was performed for DL image reconstruction. Results: DL image reconstruction was successfully implemented in a clinical workflow and TSEDL allowed a remarkable time saving of more than 50%. Overall image quality, diagnostic confidence and image impression for TSEDL were rated as excellent (median 4, IQR 4-4) and comparable to TSE (image quality: p=0.059, diagnostic confidence: p=0.157, image impression: p=0.102). Noise, sharpness, artifacts, and subjective SNR for TSEDL reached significantly superior levels to TSE (noise: p<0.001, sharpness: p=0.001, artifacts: p=0.014, subjective SNR: p<0.001). Technologists reported high levels of acceptance for DL image reconstruction. Required time for the reconstruction process was rated moderate and longer than standard sequences (median 2, IQR 2-3). Required time and effort for the implementation in daily workflow was rated as low effort (median 4, IQR 3-4). General applicability of DL reconstruction as well as acceptance of DL sequences in clinical routine were rated excellent (median 4, IQR 3-4). Conclusion: DL image reconstruction for TSE sequences can be implemented in clinical workflow and enables a remarkable time saving (>50%) in image acquisition while maintaining excellent image quality.Trial registration: Your clinical trial is officially registered at the German DRKS with the registration number: DRKS00023278.


Introduction
Magnetic Resonance Imaging (MRI) has become a modality of choice for the diagnosis of several diseases and is currently indispensable in healthcare. One big disadvantage of MRI is the long examination duration, which are not tolerated by a substantial proportion of patients and, on the other hand, come along with other downsides such as decreased image quality due to motion artifacts, increased costs and reduced patient throughput (1).
The acquisition time of MRI is primarily determined by the achievable sampling rate for a given contrast and image quality which in turn is determined by the number of samples needed for the image reconstruction for a given size and resolution. Over the past decades, different acceleration strategies have been proposed and established such as parallel imaging (PI) and Compressed Sensing (CS). These techniques acquire reduced k-space data with an array of receiver coils and afterwards reconstruct images from the acquired, undersampled data (2,3). Artifacts due to residual aliasing, or stair casing and blurring can impair achievable image quality of those acceleration strategies.
Recently, a new acceleration strategy gained attention: Deep Learning (DL) reconstruction, as discussed here and detailed below, may solve non-linear and ill-posed reconstruction problems e ciently (4)(5)(6).
Instead of ad hoc regularization that enforces sparsity, the regularization is trained on representative images. This procedure also allows the regularization to generalize for different sampling patterns, acceleration factors, and artifact behavior. Most prominently, the DL networks are trained in a supervised manner, i.e. representative fully sampled data with known results for the given application -so called ground truth data -are available which allow retrospective subsampling and training of the architectures. Testing is performed on a separate set of samples (not seen in training) with subsampled datasets.
Realistic performance assessment should be performed on prospectively subsampled datasets. While training can be computationally intensive and takes rather long, it can be performed o ine. The trained architecture can then be used in testing to reconstruct an aliasing and noise-free image within a few seconds and with greatly reduced computational demand. DL reconstruction has been recently shown to potentially accelerate image acquisition in knee MRI (7), however, the majority of existing literature on this topic focusses on o ine solutions simulating accelerated image acquisition. The evaluation of the clinical value of DL reconstruction requires an implementation in clinical settings and a prospective acquisition of accelerated data.
Therefore, the aim of this report is to describe the implementation of DL image reconstruction in clinical work ow including a thorough explanation of the required steps in patient care and an evaluation of the obtainable image quality for accelerated DL-based TSE sequences (TSE DL ) in comparison to conventional TSE sequences (TSE).

Materials And Methods
Institutional review board approval was obtained for this prospective, monocentric study. All study procedures were conducted in accordance with the ethical standards as laid down in the 1964 Declaration of Helsinki and its later amendments.
Acceleration strategies for DL accelerated MRI With DL reconstruction providing higher signal-to-noise ratio (SNR) and allowing for higher acceleration with conventional sampling patterns, the data acquisition can be tailored.
For TSE acquisitions modifying the original acquisition's image contrast is usually not a desirable goal for tailored acquisitions. Rather, DL reconstruction in MRI can be used to improve on a combination of image resolution, acquisition time, and SNR while maintaining the original contrast.
In contrast to that, some applications like the Half-Fourier single-shot turbo spin echo (HASTE) sequence provide fast and robust acquisitions at the cost of image contrast as compared to the TSE sequence. Here, the improved DL reconstruction enables acquisition protocols providing improved image contrast.
Speci cally, with a higher acceleration factor the duration of the echo train can be shortened, and therefore, the effect of T2 decay can be reduced. As an additional bene t, the speci c absorption rate (SAR) is reduced with the number of required refocusing pulses. This allows for further sequence optimizations in the form of larger gaps between consecutively acquired slices and reduced repetition times.
Besides the data acquisition for the actual image data, calibration data for the coil-sensitivity estimation needs to be acquired. For the TSE sequence, these data which cover the center of k-space (typically about 16 phase-encoding lines) are acquired as part of the imaging scan.
A conventional under-sampling pattern as known from parallel imaging is used. As shown in earlier works (8,9), these provide the same performance when reconstructed with DL-based methods as incoherent sampling patterns favored by CS. They have the important advantages of being clinically established and are highly exible regarding adaptations of resolution and signal evolution during sampling. Furthermore, their artifact behavior with regard to motion, reduced eld-of-view, and aliasing is well-known and potentially even improved by the DL-based reconstruction. Also, as a part of conventional under-sampling patterns, a fraction of the k-space's periphery is often not acquired in order to reduce acquisition time.
This effectively reduces the resolution in the phase-encoding direction and is referred to as phase resolution. It describes the fraction of acquired data in the phase-encoding direction in percent neglecting the regular parallel imaging type of under-sampling. For illustration, an exemplary sampling pattern of acceleration factor 2 and a phase resolution < 100% is shown in Figure 1. A calibration region around the k-space center is fully sampled and used for the estimation of coil-sensitivity maps.

Deep Learning Image Reconstruction
For all discussed applications, the prototype image reconstruction comprises a xed iterative reconstruction scheme or variational network (9,10). The xed unrolled algorithm for accelerated MR image reconstruction consists of multiple cascades, each made up from a data consistency using a trainable Nesterov Momentum followed by a CNN-based regularization. The regularization model's architecture is based on a novel hierarchical design of an iterative network that repeatedly decreases and increases the resolution of the feature maps, allowing for a more memory-e cient model than conventional CNNs. In addition to the input under-sampled k-space data, coil-sensitivity maps are also provided, which are estimated from the calibration data as a pre-processing step. Also, a bias-eld is extracted from a separate adjustment acquisition for image homogenization. The architecture of the reconstruction network is illustrated in Figure 2. During the training phase, the bias-eld is inserted into the image reconstruction as a nal correction step.
For the image reconstruction k-space data, bias-eld correction and coil-sensitivity maps are inserted into the variational network. Compared to previous works and the previously described cascades, the variational network also utilizes two additional types of cascades, namely, pre-and post-cascades. Like regular cascades, pre-cascades employ trainable extrapolation; however, no regularization is applied, allowing the network to focus on parallel imaging. Such design is motivated by the empirical nding that initial steps in the variational network focus on the signal recovery of missing data near the k-space center. This approach supports acquisitions without integrated calibration and exible k-space sampling.
Finally, post-cascades employing non-trainable extrapolation are also utilized for further guarantees on the data consistency, which minimizes the risk of hallucination when adversarial training is applied. The network is rst trained to minimize the combined L1 and a multi-scale version of the structural similarity (SSIM) content losses between network prediction and ground truth images. A semi-supervised re nement is applied in a subsequent training step where an adversarial loss based on Wasserstein Generative Adversarial Networks (WGAN) is also added (11).
The reconstruction was trained on volunteer acquisitions using conventional TSE protocols. About 10,000 slices were acquired on volunteers using clinical 1.5T and 3T scanners (MAGNETOM scanners, Siemens Healthcare, Erlangen, Germany). Fully sampled acquisitions with high resolution were performed in head, pelvis, and knee using representative contrasts for the respective body regions. The training data therefore included a wide range of image contrasts, orientations, body regions, and resolutions. For both sequence types, the input to the reconstruction network was retrospectively down-sampled to an acceleration factor of 4. The training was implemented in PyTorch and performed on a GPU cluster NVIDIA Tesla V100 (32GB of memory) GPU.

Implementation of DL image reconstruction in clinical work ow
For deployment in the scanner reconstruction pipeline, the obtained network was converted to a C++ implemented inference framework. For the CPU-only reconstruction on a clinical MRI scanner, inference needed about 2 seconds per slice for the used protocol settings. The reconstruction was triggered after the end of the acquisition, which resulted in a perceived reconstruction time of 2-3 minutes including additional pre-and post-processing. GPU-based reconstruction brings the duration down to the order of 10 seconds for a complete dataset but was not available in the local setting of this study.

Diagnostic image evaluation
Accelerated TSE DL sequences were prospectively acquired along with standard TSE sequences at clinical 3T MRI scanner (MAGNETOM, Siemens Healthcare, Erlangen, Germany), and a nal sample of 30 examinations in 15 volunteers were included in this analysis, see Table 1.

Implementation of DL image reconstruction in clinical work ow
The aim was to implement DL image reconstruction for TSE sequences in clinical work ow. All TSE DL were successfully implemented in clinical work ow and were successfully acquired in all for all body parts. Fat suppression could be applied successfully for the implemented TSE DL . TSE DL allowed a remarkable time saving of more than 50% for each sequence, for instance T1-weighted TSE in sagittal orientation for lumbar spine required an acquisition time of 2:45 minutes compared to T1-weighted TSE DL with an acquisition time of 58 seconds.   An overview of all results is displayed in Table 3.

Questionnaire for technologists
Required time for the reconstruction process was rated moderate and slower than standard sequences (median 2, IQR 2-3). Required time and effort for the implementation in daily work ow was rated as low effort (median 4, IQR 3-4). Technical performance as well as acceptance of DL sequences in daily routine were rated excellent (median 4, IQR 3-4).

Discussion
The aim of this report was to describe the implementation of DL image reconstruction in clinical work ow and to evaluate the obtainable image quality for TSE sequences. DL image reconstruction could be easily implemented in clinical routine in our institution with high acceptance among technologists.
TSE DL provided excellent overall image quality and diagnostic con dence and is comparable to conventional TSE. Moreover, concerning noise, sharpness, artifacts, subjective SNR TSE DL was rated signi cantly higher compared to TSE. DL-based reconstructions produce images that can exhibit even lower noise levels than a corresponding fully sampled conventional acquisition and may therefore look arti cial to experienced radiologists.
Nonetheless, image impression was rated as excellent for both sequences.
Another aim of the study was to accelerate TSE sequence acquisition by the incorporation of DL image reconstruction. DL allowed for an acquisition time reduction of ≥50% while maintaining excellent image quality and diagnostic con dence. DL seems therefore to allow for higher accelerations than prior accelerations techniques.
Prior to DL image reconstruction, high acceleration levels beyond the Nyquist-Shannon sampling limit could be obtained by CS. In fact, if images can be sparsely represented in some transform domain, then a random and incoherent sub-Nyquist sampling together with an appropriate non-linear iterative image reconstruction allows aliasing-free recovery from incompletely sampled k-space data (2). CS employs iterative reconstruction algorithms that use a priori xed sparsity-promoting transformation. Furthermore, the a priori assumption on sparsity with application-speci c regularization weighting can, if not chosen appropriately, result in residual aliasing (under-regularized) or stair casing and blurring (over-regularized).
Depending on the imaging application and sequence, a sampling trajectory that follows a desired random distribution can be challenging to implement without introducing other artifacts, e.g. by eddy currents due to strong switching gradients. Overall, these factors can impair achievable image quality and/or limit the achievable acceleration. The implementation of DL can overcome this drawback, and enables the acceleration of MR acquisition without impairing image quality.
Current Deep Learning-based image reconstruction uses super-vised learning techniques with convolutional neural networks (CNNs) (8,12). DL networks have been proposed that operate on parallel imaging (PI)-accelerated acquisitions (9, 10, 13) and on CS-accelerated acquisitions (14). Proposed methods primarily differ in the way the DL network is applied in the image reconstruction and how data consistency is enforced between reconstructed images and acquired data: i) the network acts as trainable denoiser without explicit data consistency inside the architecture during training, but handling it in outer optimization schemes (plug-and-play denoisers), ii) physics-based reconstructions that incorporate data consistency during training. Networks furthermore differ in the chosen architectures (VN, UNet, cascaded CNN, etc.), raw k-space or noisy/aliased image input, the input dimensionality (2D, 3D, 2D + time, 3D + time, etc.), single-or multi-parametric input, complex-or real-valued processing of the complex-valued data, and single-coil (coil-combined) or multi-coil processing (9,10,(14)(15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26) This report has limitations. First, we included a small amount of image data. This impacts the generalizability of our ndings. However, this report primarily aims to describe the implementation of DLreconstruction in clinical routine rather than to systematically and comprehensively evaluate the resulting image quality. For image quality assessment, body region focused clinical studies are still required. A further limitation is the fact that DL-reconstruction wasn't applied on 3D MR sequences yet. DL-based reconstruction algorithms for 3D sequences are still being developed by our team.
To conclude, DL image reconstruction can be implemented in clinical work ow and enables accelerated image acquisition allowing a remarkable time saving of more than 50% while maintaining excellent image quality for TSE sequences.

Declarations
Ethics approval and consent to participate Institutional review board approval was obtained for this prospective, monocentric study and informed consent obtained from all participants. All study procedures were conducted in accordance with the ethical standards as laid down in the 1964 Declaration of Helsinki and its later amendments.

Consent for publication
Written informed consent was obtained from all subjects in this study.

Availability of data and materials
The datasets generated and analyzed during the current study are not publicly available to ensure data privacy protection but are available from the corresponding author on reasonable request.