Deep learning‐accelerated image reconstruction in MRI of the orbit to shorten acquisition time and enhance image quality

This study explores the use of deep learning (DL) techniques in MRI of the orbit to enhance imaging. Standard protocols, although detailed, have lengthy acquisition times. We investigate DL‐based methods for T2‐weighted and T1‐weighted, fat‐saturated, contrast‐enhanced turbo spin echo (TSE) sequences, aiming to improve image quality, reduce acquisition time, minimize artifacts, and enhance diagnostic confidence in orbital imaging.

Even the slightest eye movements during these long acquisition times can cause motion artifacts, further compromising image quality.As a result, critical anatomical details may be obscured, undermining the reliability of diagnostic images.Given these challenges, there is an urgent and compelling need to develop and implement advanced imaging techniques that significantly improve time efficiency and reduce artifacts in orbital MRI.Reducing acquisition time would reduce patient burden and promote better compliance, ultimately leading to more accurate imaging results.In addition, reducing the occurrence of artifacts is critical to improving the interpretability and diagnostic reliability of orbital MRI.Artifacts not only impede accurate anatomical imaging but can also lead to misinterpretation and influence patient care decisions. 2tion artifacts commonly occur and can lead to T2 hyperintensity of the optic nerves, which resembles optic neuritis. 3These artifacts can be recognized by the blurring of nerve contours and ghosting.To address this problem, repeating the scan with the patient's gaze fixed can potentially eliminate these artifacts and provide clearer images, with the drawback that motion artifacts may reappear even if the scan is repeated.
However, there have been significant recent advances in accelerated MRI that would require the gaze to be fixed for significantly shorter time periods.Techniques such as parallel acquisition (PAT) and compressed sensing (CS) are commonly used to reduce examination time and optimize imaging procedures.[6] To address these concerns, the use of artificial intelligence (AI) in medical imaging has generated great interest and revolutionized the field.AI has led to remarkable improvements in image quality and integration of neural network functions. 7,8New deep learning (DL)-based methods have been developed to overcome the limitations of conventional acceleration techniques.These approaches aim to increase efficiency and precision and reduce scan time while maintaining image quality. 9,10udies have demonstrated the effectiveness of these AI-based methods in improving the quality of T2-weighted fluid-attenuated inversion recovery sequences 11 and reducing acquisition time in various imaging scenarios, such as non-neuroradiologic imaging. 12[15] By integrating DL reconstruction techniques into medical imaging, there is potential to increase diagnostic accuracy, improve diagnostic efficiency, and enhance the effectiveness of therapeutic interventions.

Study design
The retrospective monocentric study received approval from the local review board, and informed consent was waived using the code 118/2023B02.The study followed the principles outlined in the Declaration of Helsinki guidelines and recommendations.Fifty patients who had undergone orbital MRI for the evaluation of orbital diseases between March 2023 and July 2023 were included in the study.Exclusion criteria were nonconditional MRI implants, claustrophobia, being under the age of 18, as well as having incomplete DL-based MRI datasets or no examination at a 3.0-Tesla scanner (Figure 1).If an orbital pathology was observed, then patients were grouped according to their respective pathology.and standard orbital sequences, which consisted of axial T2-weighted images (slice thickness, 2 mm), coronal T2-weighted images (slice thickness, 3 mm), axial T1-weighted contrast-enhanced images with spectral fat saturation (slice thickness, 2 mm), and coronal T1-weighted contrast-enhanced images with spectral fat saturation (slice thickness, 3 mm).These four authorized DL-MRI sequences were also routinely examined in our clinical practice using the DL algorithm with the same slice thickness and orientation.All sequences were examined in a defined and consistent order, starting with the conventional standard sequence, followed by the corresponding AI sequence.These datasets, consisting of conventional and DL-based TSE images, were retrospectively analyzed and compared.The acceleration factor for all DL images was set to four phase-encoding steps, and the other acquisition parameters were the same for all correlated sequences, as shown in Table 1.

MRI acquisition parameters
Notably, the acquisition time for standard axial orbital T2-weighted images was 2:47 minutes, while accelerated imaging only took 42 seconds, diminishing the scan time by 75%.Similar results were found in the other sequences measured.The added time saving for accelerated imaging was approximately 69% compared to standard imaging (Table 1).
In this study, an unrolled variational network 9 was employed for DL-based image reconstruction, which has previously demonstrated potential in reducing acquisition time in various applications. 16,17The network was trained using over 10,000 slices obtained from volunteer acquisitions on different clinical 1.5-and 3-Tesla scanners (MAGNE-TOM scanners; Siemens Healthcare).After training, the network was integrated into the scanner's reconstruction pipeline by a Siemens engineer, aiming for potential application in clinical practice.

Image reconstruction
The algorithm used is as described earlier by Herrmann et al. 18 The image reconstruction prototype utilizes either a fixed iterative reconstruction scheme or a variational network approach.For input, the prototype accepts undersampled k-space data and coil sensitivity maps, while a separate acquisition is used to extract a bias field for image homogenization.Lastly, the sharpness of images was rated from 1 (severely blurred edges) to 4 (no blurring).

Image evaluation
If the standard imaging dataset had a significant lesion such as an orbital tumor, inflammatory lesion, or infectious lesion, its delineation was evaluated compared with accelerated imaging.This resulted in an excellent consensus of the evaluating neuroradiologists with regard to the image findings, such as tumor or inflammation (Table 2).

Statistical analysis
The smallest sample size was determined before the start of the study by using the software R (version 4.3.1,Vienna, Austria, https://www.rproject.org).The power was set to at least .8and the alpha to .05.The smallest sample size of 14 was calculated based on the data collected in the previous pilot project.All other statistical analyses were conducted using IBM's SPSS Statistics (Version 28.0.0.0;IBM Corp., Armonk, NY, USA, https://www.ibm.com/de-de/products/spss-statistics).Continuous variables were presented using the mean and standard deviation (SD), while ordinal scaled variables were presented using the median and interquartile range (IQR).The Wilcoxon signed-rank test was used for paired data of ordinal structure and nonnormally distributed parametric variables, with p-values adjusted using the Bonferroni procedure.Intra-and interreader variability was assessed using Cohen's kappa.The significance level for all tests was set at .05.

Patient characteristics
This retrospective study included 50 consecutive patients who underwent orbital MRI.The patients' mean age was 56 ± 13 (SD) years, with 32 male and 18 female patients between 28 and 72 years old.The subgroup built consisted of patients with orbital tumors (n = 14) as well as patients with inflammatory (n = 5) or infectious lesions (n = 2).Further patient characteristics are given in Table 2.
The sequence protocol was used as part of routine MRI for the evaluation of orbital diseases.Examples of imaging examinations are displayed in Figures 2-5.

Image quality analysis
Cohen's kappa was applied to evaluate the agreement of image quality parameters between the two readers.The obtained values were .78and .77for standard axial and coronal T2-weighted imaging, respectively.For accelerated axial and coronal T2-weighted imaging, the values were .85 and .86,respectively.
In terms of fat-saturated contrast-enhanced T1-weighted imaging, the Cohen's kappa values were .77and .80 for standard axial and coronal imaging, respectively, and .79 and .82for accelerated axial and coronal imaging, respectively.
The subsequent section will detail the results obtained by the more experienced reader, Reader 2. The detailed evaluation of the two readers is displayed in Table 3.
The impact and extent of image noise were rated significantly less in accelerated imaging than in standard imaging in all planes: in axial T2-weighted imaging, the median was 4 (IQR 3.5-4) for T2 DL and 3 (IQR 2-4) for T2 S (p < .001); in coronal T2-weighted imaging, the median was 4 (IQR 3-4) for T2 DL and 3 (IQR 3-4) for T2 S (p < .05).The image sharpness was also rated significantly better in accelerated imaging than in standard imaging planes (all p < .001).
Overall image quality was rated higher in axial T2 DL (median of 4 The diagnostic confidence was evaluated to be higher in accelerated than in standard imaging, with a median of 4 (IQR 4-4) for axial T2 DL and a median of 4 (IQR 2-4) for axial T2 S (p < .05).In coronal T2-weighted imaging, a median of 4 (IQR 4-4) for T2 DL versus 4 (IQR 3-4) for T2 S (p = .017)was obtained, whereas in contrast-enhanced T1-weighted imaging, a median of 3.All results are included in Table 3.In 47 cases (94%), both readers chose accelerated imaging as their preference.In the other three cases, conventional images were preferred because the DL images showed more artifacts.The investigation revealed that nonremovable piercings in the face and nose area, permanent makeup, or artificial eyelashes were present in these cases.A possible causal relationship is obvious but ultimately cannot be distinguished with certainty from a coincidence.A relevant MRI scan finding was found in 21 of 50 patients, and it was found on all standard and accelerated images (100%).

DISCUSSION
The results of our study demonstrate a noteworthy 69% reduction in acquisition time with the TSE DL acquisition approach.Interestingly, this reduction did not compromise the quality of images or the level of diagnostic confidence; on the contrary, both aspects showed improvement.This is significant because the use of DL-based reconstruction techniques can lead to "instabilities" during the image reconstruction process.These instabilities can result in certain small pathological findings being "masked" or artifacts being introduced. 22 the field of DL imaging, certain artifacts have been noted in the literature, such as banding artifacts typically associated with Cartesian DL reconstruction, especially in areas of the reconstructed image with low signal-to-noise ratios. 13These artifacts appear as stripe patterns aligned with the phase encoding direction. 23However, in our study sample, we found no evidence of differences in artifacts, image quality, or diagnostic confidence between standard and accelerated imaging.[26][27][28][29] DL incorporation allows for a more significant level of subsampling compared to postprocessing techniques. 29,30These developments offer the potential to address the long-standing limitations in MRI capacity effectively.Moreover, the shortened acquisition time in medical imaging provides enhanced comfort, especially for elderly or critically ill patients who may have difficulty remaining completely motionless during MRI exams.Additionally, the reduced measurement time allows for a larger number of patients to undergo examinations, bringing about not only economic advantages but also serving the needs of the high volume of examination requests.
Unlike conventional acceleration techniques such as CS, 31 DL-based acceleration preserves image quality and resolution by incorporating physical modeling through coil sensitivity into the variable neural network architecture. 9,32Previous studies have demonstrated the accurate reconstruction of pixel-wise T2 maps from highly accelerated k-space data using DL reconstruction networks. 33cent advancements in DL-accelerated T2-weighted TSE sequences have been successfully applied in various indications of MRI, resulting in notable improvements in image quality, reduced noise, fewer artifacts, and a reduction in scan time by over 60%. 16,17ditionally, a study involving healthy volunteers demonstrated the feasibility of DL reconstructions in various non-neuroradiologic applications, including the shoulder, and spine. 18The results showed significantly improved image quality with enhanced edge sharpness and reduced noise.
The study acknowledges limitations.First, the study included a small sample size, but the necessary number of patients was previously determined by power analysis.In addition, a subgroup analysis of the influence of artifacts on the detection rate of small pathologies in the images was not feasible because of the limited sample size.Nevertheless, this is the first study evaluating this DL technique for orbital imaging.Second, the scope of the study is limited by the fact that it was conducted using only one scanner from a single manufacturer in a monocentric design, which compromises the generalizability of the results.Third, the image quality analysis relied on the ratings from two However, further research is needed to investigate the quality of DLgenerated images and optimize their practical application in clinical routine.This study aims to assess the technical feasibility and to evaluate the efficacy of novel DL-based turbo spin echo (TSE) sequences, such as thin-sliced T2-weighted and fat-saturated contrast-enhanced T1weighted sequences in orbital imaging.The evaluation will focus on several aspects, including acquisition time, image quality, resistance F I G U R E 1 Flow diagram of study inclusion and exclusion.DL, deep learning; n, number.to artifacts, and diagnostic confidence.The results of this research could provide valuable insight into the feasibility and practicality of implementing DL-based reconstruction methods.

A 3 -
Tesla clinical MRI scanner (MAGNETOM Vida fit; Siemens Healthcare) equipped with a 20-channel head coil was used for all examinations.The acquisition protocol included standard brain sequences TA B L E 1 Comparison of MRI acquisition parameters and acquisition times in orbital imaging.

F I G U R E 2
A 46-year-old male patient underwent an orbital MRI scan on a 3.0-Tesla scanner because of nonspecific visual disturbances.Standard T2-weighted axial imaging of the orbit on the left with 2-mm slice thickness compared to accelerated images reconstructed with deep learning on the right side with above-average sharpness and image quality.DL, deep learning; TSE, turbo spin echo.F I G U R E 3 T2-weighted coronal MRI images of the orbit of a 51-year-old male patient.The accelerated images on the right side show significantly better delineation of the eye muscles as well as the optic nerve compared to the standard image on the left.Overall, the accelerated imaging shows fewer artifacts, better image quality, and sharper delineation of the anatomical structures.DL, deep learning; TSE, turbo spin echo.In axial contrast-enhanced fat-saturated T1-weighted imaging, the median was 4 (IQR 2-4) for T1 DL and 3 (IQR 3-4) for T1 S (p < .001); in coronal contrast-enhanced fat-saturated T1-weighted imaging, the median was 4 (IQR 3.5-4) for T1 DL and 3.5 (IQR 2-4) for T1 S (p < .05).The extent of artifacts was rated significantly less in accelerated imaging than in standard imaging in axial T2-weighted imaging (p < .001),coronal T2-weighted imaging (p < .001),and axial and coronal contrast-enhanced fat-saturated T1 imaging (p < .001for both).
5 (IQR 3-4) for T1 DL versus 4 (IQR F G U R E 4 Sixty-one-year-old woman with phthisis bulbi on the right side after neoadjuvant ruthenium brachytherapy.Accelerated T1-weighted turbo spin echo contrast-enhanced fat-saturated coronal images reconstructed with deep learning on the right side demonstrate improved sharpness and less image noise as well as sharper delineation of anatomic structures compared to the standard image on the left side.DL, deep learning; TSE, turbo spin echo.F I G U R E 5 Sphenoid wing meningioma with infiltration of the left lateral orbit.Note the sharper delineation of the orbital structures displaced by the meningioma in accelerated, contrast-enhanced imaging with fat saturation and a scan time reduction of 68% (1:38 minutes vs. 5:06 minutes) on the right side compared to standard imaging without deep learning on the left side.DL, deep learning; TSE, turbo spin echo.3-4) for T1 S (p < .05)was found in axial plane and a median of 3 (IQR 3-4) for T1 S versus 4 (IQR 3-4) for T1 DL (p < .001) was found in coronal plane.
radiologists.Although there was a strong interrater agreement, image quality assessments can be influenced by individual biases, potentially affecting the reliability of the results.Future studies could be advantageous by investigating additional sequences, especially those incorporating ultrathin-slice imaging.It is also crucial to take into account recent research indicating that DL might result in specific artifacts, such as "banding artifacts," characterized by streaking patterns aligned with the phase-encoding direction, as observed in DL-accelerated musculoskeletal MRI34 and spine MRI.13In addition, simultaneous scans of standard conventional and AI sequences could potentially increase the risk of artifacts due to eye movement caused by longer scanning times.It is therefore even more pleasing that no adverse effects of this kind occurred in the AI sequences and that no DL-specific artifacts were detected in the scans performed in the current study.Nevertheless, this study represents an initial clinical exploration of the DLbased acceleration technique in orbital MRI, and it presents promising findings.In conclusion, our study successfully showcased the clinical feasibility of utilizing DL for TSE image reconstruction in orbital MRI with standard T2-weighted and contrast-enhanced T1-weighted sequences.The DL TSE images displayed exceptional quality and enhanced diagnostic accuracy, while also achieving a remarkable 69% reduction in examination time.As a result, the DL technique holds great potential for enabling ultrafast orbital MRI.Furthermore, in the future, this technique can be extended to other sequences, paving the way for rapid and precise imaging in diverse clinical scenarios.