Prospective intraindividual comparison of a standard 2D TSE MRI protocol for ankle imaging and a deep learning-based 2D TSE MRI protocol with a scan time reduction of 48%

Purpose Magnetic resonance imaging (MRI) scan time remains a limited and valuable resource. This study evaluates the diagnostic performance of a deep learning (DL)-based accelerated TSE study protocol compared to a standard TSE study protocol in ankle MRI. Material and methods Between October 2020 and July 2021 forty-seven patients were enrolled in this study for an intraindividual comparison of a standard TSE study protocol and a DL TSE study protocol either on a 1.5 T or a 3 T scanner. Two radiologists evaluated the examinations regarding structural pathologies and image quality categories (5-point-Likert-scale; 1 = “non diagnostic”, 5 = “excellent”). Results Both readers showed almost perfect/perfect agreement of DL TSE with standard TSE in all analyzed structural pathologies (0.81–1.00) with a median “good” or “excellent” rating (4–5/5) in all image quality categories in both 1.5 T and 3 T MRI. The reduction of total acquisition time of DL TSE compared to standard TSE was 49% in 1.5 T and 48% in 3 T MRI to a total acquisition time of 5 min 41 s and 5 min 46 s. Conclusion In ankle MRI the new DL-based accelerated TSE study protocol delivers high agreement with standard TSE and high image quality, while reducing the acquisition time by 48%.


Introduction
The ankle is the second most injured body part in sports [1,2] and commonly evaluated by magnetic resonance imaging (MRI). Ongoing issue thereby is the acceleration of acquisition time. For musculoskeletal imaging, usually multiple two-dimensional (2D) multi-slice acquisitions of anisotropic voxels [3] are used. Previous studies tested three-dimensional (3D) sequences with isotropic voxels in ankle imaging with secondarily reconstructed slice orientations as a method for acceleration [4][5][6][7][8][9]. However, the positive correlation of image quality and acquisition time leads to an acquisition time of about 10 min for a 3D data set with small voxel sizes [10]. Furthermore, artifacts may corrupt the whole data set which might be seen only post-acquisition and require a repeated examination.
An established approach for accelerating the image acquisition is to acquire undersampled data and perform a parallel imaging (PI) reconstruction [4,5,11]. The approach is hereby limited by a decreased signal-to noise ratio (SNR) and an increased level of artifacts [12]. A further approach is compressed sensing (CS) [13][14][15] which combines an incoherent undersampling scheme with a sparsity enforcing reconstruction to recover the undersampled data. Limitations are residual blurring and an unrealistic image presentation.
Deep learning (DL) algorithms for image reconstruction from undersampled MRI data deliver a method to overcome these problems. They offer the potential of image reconstructions with higher SNR compared to conventional reconstruction techniques while at the same time delivering a realistic image presentation [16]. The gain in SNR can be spent on a reduction of the acquisition time, improved image quality or combinations thereof [17]. Recent studies have shown promising results in the knee [12].
The purpose of this study was to evaluate the diagnostic performance of DL-based PD-and T1-weighted turbo spin echo (TSE) sequences in a routine study protocol for ankle imaging at 1.5 T and 3 T MRI in a prospective setting.

Materials and methods
This study was approved by the institutional review board (Eberhard Karls University Tuebingen, project identification code: 055/2017BO2). The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013).

Study population
The study period of this prospective single center study was October 2020 till July 2021. All patients who underwent clinically indicated MRI of the ankle were included. Excluded were patients with divergent study protocols due to the specific clinical problems, e.g., follow-up care of tumor patients.

Hardware parameters of the MRI acquisition
The examinations were performed in random selection either on a 1.5 T scanner (MAGNETOM Aera or MAGNETOM Avanto fit , Siemens Healthcare, Erlangen, Germany) or on a 3 T scanner (MAGNETOM Skyra, MAGNETOM Prisma fit or MAGNETOM Vida, Siemens Healthcare, Erlangen, Germany). All patients were examined in supine position, feet first using a dedicated 16-channel foot-ankle coil (Foot/ Ankle 16, Siemens Healthineers, Erlangen, Germany).

Standard MRI protocol for ankle imaging
According to international guidelines (e.g., European Society of Musculoskeletal Radiology) our study protocol consisted of a coronal T1-weighted TSE sequence and coronal, sagittal and transversal PD-weighted TSE sequences with spectral fat suppression (see Figs. 1, 2). Total acquisition time of the standard study protocol was 10 min 59 s on a 1.5 T scanner and 11 min 0 s on a 3 T scanner. Further imaging parameters are displayed in Table 1.

DL-based MRI protocol for ankle imaging
As in PI a conventional equidistant undersampling pattern was used as it provides equal performance with DL-based reconstruction in contrast to incoherent sampling patterns known from CS. The basis for the prototypical reconstruction is a fixed iterative reconstruction scheme comprising a variational network [18][19][20]. Inputs to this DL-based reconstruction are k-space data, bias field correction and coil sensitivity maps. Multiple cascades made up from a data consistency step using a trainable Nesterov momentum and a following convolutional neural network regularization constitute the DL-based reconstruction. The parameters of the reconstruction where obtained through supervised training using volunteer data acquired on different 1.5 T and 3 T scanners (MAGNETOM, Siemens Healthcare, Erlangen, Germany). The reconstruction method is described in more detail in [16].  Table 1.

Evaluation of the image quality
The image quality was evaluated independently by two radiologists (six years of experience on MSK imaging including  fellowship training; five years of experience on MSK imaging) on a 5-point Likert scale (1 = "non-diagnostic image quality", 2 = "low image quality", 3 = "moderate image quality", 4 = "good image quality", 5 = "excellent image quality"). Criteria were image noise, sharpness, banding artifacts and diagnostic confidence.

Evaluation of the diagnostic accuracy
The readers rated the MRI blinded to patients' data, acquisition technique, magnetic field strength of the scanner and in random order. The relevant ligamentous structures were graded into four groups from intact to complete rupture, posttraumatic fiber alterations such as thickening were classified as 1. The relevant tendinous structures were also graded into four groups from intact to complete rupture with inflammatory or reactive changes such as pathological surrounding fluid or hyperintense signal of the tendon on PD-images were classified as 1. Cartilage defects and osteochondral lesions were graded as per International Cartilage Regeneration and Joint Preservation Society (ICRS) and Kramer's classification [21]. Other pathological findings were rated dichotomously. In addition, a consensus reading of the standard TSE examinations was carried out to determine the presence of pathologies in the study collective.

Statistical analysis
Statistical analysis was performed using the software packages JMP (Version 15.  Except for the extensor tendons there were both inconspicuous and pathological findings in all rated anatomical structures. The pathological findings in the study examinations, which are based on a consensus reading of the two readers on the well-established standard TSE imaging are listed in Table 2 (an exemplary lineup of structural pathologies is shown in Fig. 3).

Image quality of standard TSE and DL TSE ankle MRI
In PD-weighted images, the DL TSE sequence showed significantly less image noise (1.5 T and 3 T), while banding   artifacts were rated significantly worse (1.5 T and 3 T) and sharpness was rated slightly better for DL TSE without statistical significance (see Table 3 and exemplary patients' images in Figs. 4, 5, 6).
In the T1-weighted images banding artifacts were also rated worse for DL TSE on the 3 T scanners, but not on 1.5 T scanners. Other rated image quality criteria were not significantly different in T1-weighted images neither on 1.5 T nor on 3 T scanners.
In almost all patients, diagnostic confidence was rated excellent or good without any significant difference between standard TSE and DL TSE on 1.5 T scanners and 3 T scanners, respectively, as well as for PD-weighted or T1-weighted sequences (Table 3).

Diagnostic accuracy of standard TSE and DL TSE ankle MRI
The intermethod agreement on all rated anatomical structures was almost perfect/perfect for both readers, both grouped regarding field strength (1.5 T and 3 T) and in total of the whole study collective. Thereby all anatomical structures in the group "Others" showed perfect agreement. For more details see Table 4.

Discussion
The presented results show almost perfect/perfect agreement of DL TSE with standard TSE in all analyzed structure categories (ligamentous lesions, tendinous lesions, cartilage defects, OC, fluid sensitivity pathologies) with a median "good" or "excellent" rating in all image quality categories in both 1.5 T and 3 T MRI. The few discordant image interpretations referred to adjacent categories of the reading scheme and seemed to remain in the normal range of interindividual image interpretation (illustrated for example in the image of the ATFL in the middle row of Fig. 3).
The reduction of total acquisition time of DL TSE compared to standard TSE was 49% in 1.5 T and 48% in 3 T MRI. There are no comparable studies on DL-based sequences in ankle MRI published yet, however, studies on the knee support our results of a high agreement of DL TSE with standard TSE maintaining realistic image impression, meanwhile saving a substantial amount of acquisition time [16,17]. Though not in scope of the present study, the authors hypothesize that the results obtained here may be generalizable to other MSK imaging applications where similar contrasts are acquired i.e., hand-wrist or shoulder imaging. The existence of banding artifacts in DL TSE sequences in MSK imaging has been reported in the literature [16]. While further algorithmic improvement of the DL reconstruction is desirable in this regard, the diminution of the image quality in the presented study was little.
Another interesting approach to accelerate ankle MRI imaging is the acquisition of 3D data sets with isotropic voxels [4][5][6][7]. The benefits are obvious: the acquisition of a single volume instead of coronal, sagittal and transversal data sets is supposed to save time, and additionally delivers the possibility for the reconstruction of oblique reformats for special issues. Previous works indicate that a high-quality data set with small voxel sizes (0.5 mm 3 ) requires an acquisition time of about 10 min, what might be the reason why this approach has not been widely adopted in clinical practice. Moreover, simultaneous multi-slice excitation combined with PI has been reported an acceleration capability comparable to those of DL TSE combined with PI in the knee [23,24]. Further studies have to prove the feasibility of multi-slice excitation techniques in the ankle. Besides the direct comparison of simultaneous multi-slice excitation sequences and conventional sequences, another comparison using DL-based reconstruction techniques would be interesting. Furthermore, DL-based reconstructions of simultaneous multi-slice excitation sequences are promising [25].
Limitations of the presented study were the single center design and different MRI scanners from a single vendor, however, it in turn shows that the results may not be restricted to specific MR scanners and receive coils of this vendor. Suitable studies could substantiate the auspicious results on DL TSE in ankle MRI.
In conclusion, DL TSE is a promising new technique for 1.5 T and 3 T ankle MRI with high agreement compared to standard TSE and excellent image impression while saving 48% and 49% acquisition time.
Funding Open Access funding enabled and organized by Projekt DEAL. The authors have not disclosed any funding.

Conflict of interest
The authors of this manuscript declare relationships with Siemens Healthineers. The prototype DL reconstruction was provided by Siemens Healthcare, Erlangen, Germany. Full control of patient data was maintained by the authors.

Consent to participate Written informed consent was given by all participants.
Human and animal rights This study was approved by the institutional review board (Eberhard Karls University Tuebingen, project identification code: 055/2017BO2). The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). Animals were not involved.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.