See also article by Yang et al in this issue.
Farid GharehMohammadi, PhD, is a postdoctoral fellow at Mayo Clinic, specializing in the exploration of deep learning applications within the realm of musculoskeletal imaging, focusing on both diagnosis and authentication. His PhD in computer science with a concentration in meta/multimodel classification is from the University of Georgia. He has published more than 30 journal and conference papers. He received the Outstanding Achievement Award for deep learning at CSCI 2019. He is the associate editor for CSCE 2022–2023.
Ronnie A. Sebro, MD, PhD, is an internationally renowned statistician, geneticist, and musculoskeletal radiologist. His research focuses on using artificial intelligence and machine learning to improve diagnosis of osteoporosis, sarcomas, and other musculoskeletal conditions. He has published more than 100 scientific journal papers and lectured nationally and internationally. He is deputy editor of Radiology: Artificial Intelligence, associate editor for BMC Musculoskeletal Disorders, and associate editor for the Journal of Imaging Informatics in Medicine.
The demand for MRI has continued to increase over the past 20 years (1). The typical MRI study may take 20 to 60 minutes, depending on the body parts imaged and sequence protocols used. The MRI scan time is defined as the sum of the preprocessing, acquisition, and postprocessing times (2,3) and can be calculated from the Digital Imaging and Communications in Medicine (DICOM) MRI scan files. Most MRI scanners can image only 20 to 40 patients daily. As a result, patients may have to wait for weeks or months. Decreasing the MRI scan time would allow imaging of more patients and potentially decrease MRI waiting times.
Artificial intelligence (AI)–based solutions to decrease the MRI scan time related to MRI scan reconstruction can be classified as decreasing preprocessing times or postprocessing times. In this issue of Radiology: Artificial Intelligence, Yang et al showed that postprocessing solutions were better than preprocessing solutions to decrease MRI scan times (3). Yang et al used deep learning applications to assess two important clinical metrics: (a) MRI scan time and (b) MRI room time. The latter was defined as the sum of the patient change time, interscan delay time, and MRI scan time, calculated using the institution’s radiology information system.
Undersampled or noisy MRI data takes less time to obtain, and deep learning reconstruction (DLR) algorithms are able to reduce MRI scan time by reconstructing high-quality images from this poorer quality MRI data. Yang et al compared DICOM-based and k-space–based DLR algorithms to the conventional reconstruction algorithm to evaluate how these three methods affected MRI scan time and MRI room time.
The DICOM-based reconstruction algorithm is a commercially available tool designed for denoising head, spine, neck, and knee MRI and for sharpness enhancement in noncontrast head MR images (3). The k-space–based reconstruction algorithm is a commercially available tool integrated into the scanner’s native reconstruction pipeline and requires access to the raw k-space data. The conventional reconstruction algorithm is the standard imaging technique and uses no AI-based algorithms to create MRI scans.
When considering MRI scan times, the DICOM-based reconstruction algorithm decreased mean MRI scan time between 6 minutes (brain MRI without contrast material) and 24 minutes (knee MRI without contrast material). The k-space–based reconstruction algorithm decreased mean MRI scan time between 3 minutes (cervical spine MRI without contrast material, lumbar spine MRI without contrast material) and 11 minutes (knee MRI without contrast material). Over the same period, the mean MRI scan time decreased using the conventional method between 1 minute (brain MRI without contrast material, cervical spine MRI without contrast material) and 8 minutes (knee MRI without contrast material).
When considering MRI room times, the DICOM-based reconstruction algorithm decreased the mean MRI room time between 5.5 minutes (cervical spine MRI without contrast material) and 23 minutes (knee MRI without contrast material). Several MRI room times were not statistically significantly changed, including those for brain MRI with contrast material, brain MRI without contrast material, thoracic MRI without contrast material, shoulder MRI without contrast material, and pelvic MRI with contrast material. The k-space–based reconstruction algorithm decreased mean MRI room times between 1 minute (lumbar MRI without contrast material) and 21.5 minutes (thoracic MRI without contrast material). No significant differences in MRI room times were noted for cervical spine MRI without contrast material or pelvic MRI with contrast material after incorporating the k-space reconstruction algorithm. MRI room times were largely unchanged using the conventional method; however, there was an 8-minute decrease in the average MRI room time for lumbar spine MRI without contrast material.
This study represents a real-life study of the impact of integration of AI-based MRI reconstruction algorithms into the clinical workflow and showed that there was some decrease in MRI scan times and MRI room times for some studies, assuming that there were no changes in the MRI protocols, technologists, or other workflow. This advancement shows that AI has the capacity to decrease MRI scan time and MRI room time, potentially to enable more patients to undergo MRI scans in the future. The novelty of Yang et al’s work lies in the utilization of two U.S. Food and Drug Administration–approved DLR algorithms (4,5) within clinical settings to expedite MRI scan and room times. This application of AI demonstrates significant promise in reducing MRI scan times, particularly beneficial for multicenter health care institutions seeking to reduce scan times and increase patient throughput.
Other DLR techniques for MRI have resulted in faster or enhanced image acquisition (6). Rudie et al (7) evaluated the effectiveness of a DLR algorithm on 32 consecutive patients (25 for tumor assessment, six for workup for headaches, and one for cognitive dysfunction) and found a 45% scan time reduction with no significant clinical image degradation. Rastogi et al (8) showed that DLR algorithms for undersampled MRI data could decrease MRI scan times by a factor of 10 while retaining image quality. Hyun et al (9) used a DLR algorithm based on U-Net architecture to reconstruct high-quality images from undersampled k-space data; they decreased MRI scan time using only 29% of the k-space data without compromising image quality.
There are a few limitations noted in this study. The first is that there is limited MRI vendor coverage. The study focused solely on two MRI vendors. Future investigations should include a broader range of vendors for a more comprehensive assessment. Not all MRI protocols were included in the study; therefore, it is difficult to assess whether there is any potential bias based on the MRI protocol. Future directions should encompass a broader range of MRI protocols to evaluate the robustness and generalizability of DLR analysis across different imaging settings. The study was evaluated based on a condition per MRI protocol, a minimum sample size of 200 cases, and excluded rare MRI cases. It would be interesting to see how these DLR methods affect scan times and room times for MRI studies of the fingers, toes, elbows, wrists, and temporomandibular joints. Future research should include a wider spectrum of anatomic regions.
Another factor that may contribute to MRI room time is patient habitus, patient age, and patient frailty. It was not immediately obvious whether the patients prior to DLR algorithm implementation were comparable to those patients after DLR algorithm implementation. The analysis lacked patient-level clinical data. Future investigations should incorporate patient-level information to facilitate a more comprehensive understanding of the clinical implications of DLR and should include larger and more diverse patient cohorts. Finally, the DLR methods utilized in the study are proprietary commercial algorithms, which limits the ability for an independent researcher to completely replicate the study.
References
- 1. Chang Y , Li Z , Saju G , Mao H , Liu T . Deep learning-based rigid motion correction for magnetic resonance imaging: a survey . Meta Radiol 2023. ; 1 ( 1 ): 100001 . [Google Scholar]
- 2. Sartoretti E , Sartoretti T , Binkert C , et al . Reduction of procedure times in routine clinical practice with Compressed SENSE magnetic resonance imaging technique . PLoS One 2019. ; 14 ( 4 ): e0214887 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Yang A , Finkelstein M , Koo C , Doshi AH . Impact of deep learning image reconstruction methods on MRI throughput . Radiol Artif Intell 2024. ; 6 ( 3 ): e230181 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. U.S. Food and Drug Administration . 510(k) Clearance No. K191688. https://www.accessdata.fda.gov/cdrh_docs/pdf19/K191688.pdf. Published September 2019. Accessed March 2024.
- 5. U.S. Food and Drug Administration . 510(k) Clearance No. K213717. https://www.accessdata.fda.gov/cdrh_docs/pdf21/K213717.pdf. Published June 2022. Accessed March 2024.
- 6. Kiryu S , Akai H , Yasaka K , et al . Clinical impact of deep learning reconstruction in MRI . RadioGraphics 2023. ; 43 ( 6 ): e220133 . [DOI] [PubMed] [Google Scholar]
- 7. Rudie JD , Gleason T , Barkovich MJ , et al . Clinical assessment of deep learning-based super-resolution for 3D volumetric brain MRI . Radiol Artif Intell 2022. ; 4 ( 2 ): e210059 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Rastogi A , Brugnara G , Foltyn-Dumitru M , et al . Deep-learning-based reconstruction of undersampled MRI to reduce scan times: a multicentre, retrospective, cohort study . Lancet Oncol 2024. ; 25 ( 3 ): 400 – 410 . [DOI] [PubMed] [Google Scholar]
- 9. Hyun CM , Kim HP , Lee SM , Lee S , Seo JK . Deep learning for undersampled MRI reconstruction . Phys Med Biol 2018. ; 63 ( 13 ): 135007 . [DOI] [PubMed] [Google Scholar]