Deep Dive into MRI: Exploring Deep Learning Applications in 0.55T and 7T MRI

MRI operates by using an intense magnetic field that aligns the protons in the tissues of a person. The spins of protons align parallel to it when it is present. The protons are displaced momentarily from their normal position when they receive an RF pulse. After the completion of the RF pulse, the protons again realign with the magnetic field and emit energy [3]. This emitted energy is captured by MRI sensors and used to create images which represent body structures as shown in Figure 1.

The magnetic field intensity of MRI systems influences their performance by changing the signal-to-noise ratio (SNR), which in turn affects the quality of a resultant image. According to their field strength, MRI systems are rated as low-field, high-field or ultra-high-field systems.

Low-Field MRI - 0.55T

Low-field MRI systems, like the 0.55T MRI, function at a decreased magnetic field strength, leading to reduced SNR and CNR. This often leads to a lowering of spatial resolution and a longer duration for scanning. However, advancements in technology have improved the capabilities of low-field MRI. Advanced hardware and sophisticated image reconstruction techniques greatly improve the diagnostic capabilities of modern 0.55T MRI systems [5], [4]. Low-field MRI’s open design provides a major benefit by enhancing patient accessibility and comfort, particularly for individuals with claustrophobia. Moreover, low-field MRI systems are usually more cost-effective, needing minimal shielding and consuming less power, ultimately cutting down operational expenses [6].

High-Field and Ultra-High-Field MRI - 7T and above

To achieve enhanced image resolution and a more rapid scan time, physicians use MRI systems with high magnetic fields, such as 1.5T and 3T [7]. At the same time, ultra-high field MRI systems that focus mainly on 7T and above are leading the way in terms of MRI technology. Consequently, the first 7T clinical MRI scanner was launched in 2017, marking a significant advance in this technology that is now applicable at both clinical and research levels [8], [9]. This results in an increase in spatial resolution and SNR, allowing the use of smaller voxels and the presentation of more detailed anatomical structures [7]. This proves advantageous, especially when dealing with neurological diseases, along with musculoskeletal and cardiovascular images. However, UHF MRI has its limitations. The higher intensity magnet can cause dizziness or nausea, as well as artefacts in the resulting images. In addition, processing and storing the large amounts of data produced requires considerable computing power [7], [8].

Integration of Deep Learning with MRI

Incorporating DL methods into MRI technology is a breakthrough, providing new methods to enhance image quality, speed up imaging processes, and enhance diagnostic precision. DL algorithms are being used more in MRI to address disabilities in imaging speed and resolution, leveraging their abilities in pattern recognition and data analysis [2]. DL techniques can enhance image quality for 0.55T MRI systems, boosting SNR and CNR to compete with higher field systems in image output. In higher field strengths DL can assist in handling extensive data, minimising artefacts, and enhancing image reconstruction processes for UHF MRI systems [2], [6], [9]. Therefore, these advancements ensure better diagnostic accuracy, easier patient access, and increased efficiency in imaging, ultimately influencing the direction of MRI in clinical settings.

This subsection presents studies on the application of DL with MRI scans for body composition and abdominal studies. The studies explored here are shown in Table LABEL:tab1. Rusche et al. [14] performed a comparative study in 2022 on subject comfort during MRI examination using the latest low-field MRI system with a wider bore aperture (0.55T) compared to the 1.5T MRI system. The research involved a group of 50 patients who had undergone an MRI scan on both machines involved in the survey. Those patients were required to assess their sense of space, noise level, and comfort as well as coil comfort and overall impression of the exam. The findings indicated that the majority of patients regarded, the 0.55T MRI scanner as better relating to sense of space, noise levels and general comfort. These noise levels were also measured and showed that 0.55T had markedly reduced noise for all body regions compared to 1.5T:

• •

  Brain MRI: 83.8 ± 3.6 dB (0.55T) vs. 89.3 ± 5.4 dB (1.5T), p = 0.04

• •

  Spine MRI: 83.7 ± 3.7 dB (0.55T) vs. 89.4 ± 2.6 dB (1.5T), p = 0.004

• •

  Hip MRI: 86.3 ± 5.0 dB (0.55T) vs. 89.1 ± 1.4 dB (1.5T), p = 0.04

The study found that the larger aperture and lower noise levels of 0.55T MRI scans contribute to patient’s comfort during the procedure, reducing anxiety levels.

In 2023, Nayak et al. [15] investigated whether it would be possible to profile body composition using a whole-body MRI system operating at low-field strengths. Through scans on healthy adults, they refined parameters such as echo time, resolution, and inversion angle. A test-retest study with 10 volunteers confirmed the viability of the optimised protocol. Despite the differences in optimal parameters compared to higher field strengths, image quality remained comparable, without significant issues. Consequently, the study concludes that body composition MRI at 0.55T, with optimised parameters, is feasible for assessing conditions like metabolic diseases and obesity.

In 2024, Ramachandran et al. [16] made a comparative analysis of abdominal MRI images taken from healthy subjects and patients using a 0.55T MRI. The study involved 15 healthy subjects and 52 patients who were scanned by abdominal MRI scans at 0.55T. Additionally, images were taken of healthy individuals at 1.5T and the images were available for comparison on 28 of the 52 patients. Two radiologists independently assessed and scored the quality of the images using a four-level rating scale. The images acquired at 0.55T were considered satisfactory, although some sequences may require further optimisation. The radiologists were able to respond to clinical queries for the majority of the patient cases. While the diagnostic quality was satisfactory, certain sequences at 0.55T received lower scores compared to those at higher field strengths. Moreover, the average scan time at 0.55T was found to be longer than that at higher field strengths, which may present a challenge in clinical settings.

Tian et al. [17] introduced a real-time MRI technique at 0.55T using spiral out-in-out-in sampling combined with balanced steady-state free precession (bSSFP) to separate water and fat effectively. This method, tested on healthy volunteers, successfully visualised wrist ligament gaps during motion, enhanced the clarity of epicardial fat and coronary arteries in the heart, and improved the water-fat contrast for observing small bowel motility in the abdomen. The flexible sequence can be customised for various temporal and spatial resolutions, thereby demonstrating its adaptability for a multitude of clinical and research applications.

Dietzel et al. [18] investigated the potential of using a low-field MRI system operating at 0.55T for breast imaging. The study involved a multiparametric breast MRI protocol applied to 12 participants, using a next-generation scanner (MAGNETOM Free.Max). The protocol included dynamic T1-weighted, T2-weighted, and diffusion-weighted sequences optimised for the low-field setting. Two experienced radiologists independently evaluated the images and confirmed that all examinations yielded diagnostic-quality images without significant artefacts. The typical imaging phenotypes were effectively visualised, and the total scan time ranged from 10 minutes and 30 seconds to 18 minutes and 40 seconds. The results of the study indicated that low-field breast MRI at 0.55T is a viable option for clinical use, delivering diagnostic-quality images within a reasonable examination time. This approach is expected to result in reduced procedural costs and facilitate the integration of MRI systems directly within breast clinics, potentially enhancing accessibility and efficiency in breast cancer diagnostics.

Kim et al. [19] proposed a novel approach to accelerate MR cholangiopancreatography (MRCP) scans involving DL-based reconstruction techniques at both 3T and 0.55T field strengths. The research team utilised retrospectively undersampled data acquired at 3T to train a variational neural network. They then evaluated the performance of their proposed method against conventional techniques like parallel imaging and compressed sensing. Furthermore, they developed a self-supervised DL reconstruction method and tested their approach in both retrospective and prospective accelerated scenarios. The results demonstrated a reduction in the average acquisition time with the variational network while maintaining an image quality comparable to that of conventional acquisition methods. The variational network exhibited superior performance in preserving the sharpness and visibility of hepatobiliary ducts, with promising adaptability to MRCP at 0.55T. This study showed the potential of DL-based reconstruction to enhance the efficiency and quality of MRCP imaging, paving the way for accelerated clinical applications.

Low-field MRI has the capability to diagnose various musculoskeletal (MSK) conditions. Researchers have investigated the potential benefits and limitations of using low-field MRI systems for MSK imaging applications. The studies explored here are presented in Table LABEL:tab2. In 2023, Breit et al. [20] assessed the potential of a 0.55T MRI system for lumbar spine imaging, comparing it to a standard 1.5T system. The study included 14 volunteers who underwent imaging on both systems, with additional sequences acquired at 0.55T using advanced post-processing techniques. Although the 0.55T images were rated slightly lower in quality compared to the 1.5T images for signal/contrast, resolution, and visualisation, the overall quality of the examination remains good, with a high level of agreement between radiologists. The application of advanced post-processing techniques at 0.55T data showed potential for enhancing image quality while also reducing the necessary acquisition time.

In 2024, Lavrova et al. [21] conducted a study that compared the image quality, variability between reader’s interpretations, and diagnostic effectiveness of standard lumbar spine MRI sequences acquired at 0.55T versus those acquired at the more conventional field strengths of 1.5T and 3T. The study involved 665 image series from 70 studies, which were evaluated by two neuroradiologists for overall imaging quality, the presence of artefacts, and the ability to visualise anatomical features using a four-point Likert scale. Despite some sequences at 0.55T receiving lower scores, all were considered acceptable for diagnostic use. There was a low level of agreement between readers on individual scores, but a high level of agreement on diagnoses, indicating the potential of 0.55T MRI for detecting routine spine pathologies.

The study by Schlicht et al. [22] aimed to evaluate the performance of a DL-based post-processing technique for enhancing lumbar spine MRI scans acquired at 0.55T. The primary goals were to assess the resulting image quality and the potential reduction in acquisition time offered by this advanced method. A comparison was made between standard DL-based reconstruction and advanced reconstruction (ADLR) in the context of lumbar spine imaging of 18 patients. A visual evaluation by radiologists using a Likert scale demonstrated superior image quality, resolution, and assessability of anatomical structures with the ADLR technique. A quantitative analysis revealed that ADLR resulted in higher contrast ratios in fluid-sensitive sequences. Furthermore, ADLR resulted in a significant reduction in image acquisition time, with a decrease of 44.4% from 14:22 minutes to 07:59 minutes. These results show that advanced DL-based reconstruction algorithms can improve image quality and reduce the time needed to take a lumbar spine MRI at 0.55T.

Seifert et al. [23] evaluated the image quality and metal artefact intensity between 0.55T and 1.5T MRI in patients who had undergone posterior fusion surgery with spinal implants. The study involved 50 patients and employed metal artefact reduction protocols used in clinical practice for the evaluation of images. The findings revealed that while contrast and resolution were rated lower at 0.55T compared to 1.5T MRI, the ability to evaluate the spinal canal was equally effective at both field strengths. However, mild susceptibility artefacts were observed at 0.55T. Furthermore, the study indicated that while the visibility of anatomical structures near metal implants reduced as the implant length increased in 1.5T MRI scans, the visibility stayed consistent and unaffected by implant length in the 0.55T MRI scans. This suggests that the latter may offer a superior assessment of near-metal anatomy, particularly in cases involving multi-level spinal implants.

In 2023, Chaudhari et al. [24] conducted a study investigating the feasibility of real-time MRI of the moving wrist using a high-performance 0.55T system. Conventional MRI systems are constrained by limitations in real-time imaging due to the generation of artefacts and power deposition. The study participants underwent wrist motion manoeuvres during MRI scans with temporal resolutions below 100 milliseconds. The objective of the study was to visualise tissues to evaluate dynamic wrist instability. The results demonstrated minimal distortion and image artefacts in real-time MRI scans, with static MRI providing a complement to the assessment. In conclusion, the study demonstrated that high-temporal-resolution real-time MRI in conjunction with static MRI is a viable approach with a 0.55T system, offering enhanced evaluation of wrist dysfunction and instability.

Lim et al. [25] investigated the potential of real-time MRI to observe the dynamics of speech production using a 0.55T MRI system. The study involved healthy adult volunteers and employed a range of pulse sequences, including bSSFP and gradient-recalled echo. The results demonstrated that both imaging sequences were capable of effectively tracking and visualising the intricate tongue motion patterns during the rapid articulation of consonant-vowel syllables. Nevertheless, bSSFP demonstrated superior SNR in the vocal tract articulators and offered enhanced image quality and SNR efficiency, particularly when utilising the shortest readout duration. This result indicates that bSSFP outperforms gradient-recalled echo in terms of image quality, even when compared to higher field strengths. The study highlights the potential of using high-performance 0.55T MRI systems for speech-production real-time MRI, with spiral bSSFP demonstrating the capacity to provide enhanced SNR efficiency and image quality, while avoiding the common artefacts.

Schmidt et al. [26] conducted a study to assess the diagnostic image quality of knee MRI performed at a low magnetic field strength of 0.55T in comparison with the 1.5T MRI. They conducted MRI scans on 20 volunteers using both scanners. Two radiologists assessed the MRI sequences, scoring the overall image quality, level of image noise, and diagnostic utility using a five-point rating scale. Furthermore, the study involved assessing the potential presence of pathologies affecting the menisci, ligaments, and cartilage within the scanned regions. The study demonstrated that while image quality for specific sequences at 0.55T was comparable to that of 1.5T MRI, it exhibited a slight decline for other sequences. Nevertheless, the diagnostic capability for meniscal and cartilage pathologies remained comparable between the two field strengths. Moreover, the tissue contrast ratios were comparable across both. The inter-observer agreement on image quality was satisfactory, particularly about the identification of pathologies. Therefore, the implementation of DL-reconstructed images at 0.55T produced knee MRI scans yielded image quality comparable to that obtained from the higher 1.5T field strength MRI scanners.

A complementary study by Donners et al. [27] was conducted to assess the efficacy of the new-generation 0.55T knee MRI in comparison to 3T knee MRI for patients experiencing acute trauma and knee pain. The research involved 25 patients presenting with symptoms to obtain knee MRIs on both 0.55T and 3T systems. DL image reconstruction algorithm was employed only on the 0.55T system, resulting in enhanced image quality. Two radiologists independently reviewed all images, assessing image quality parameters, detecting MRI findings, and evaluating confidence levels in reporting these findings. The study demonstrated that while image quality was rated higher for 3T MRI, there was agreement between 0.55T and 3T MRI in detecting various knee lesions. Nevertheless, there was a lower degree of agreement and reader confidence in the detection of low-grade cartilage and meniscal lesions in the 0.55T MRI compared to the 3T MRI. This suggests that while the 0.55T knee MRI system, combined with DL image reconstruction, can reliably detect and grade joint lesions in symptomatic patients, it may face challenges in accurately evaluating low-grade cartilage and meniscal lesions when compared to the higher 3T MRI field strength.

Bhattacharjee et al. [28] addressed the feasibility of applying DL segmentation models, originally trained on 3T MRI data, directly to low-field 0.55T knee MRI scans of healthy controls. The objective of this study is to evaluate the performance of DL algorithms for quantifying knee biomarkers at 0.55T compared to 3T. Seven healthy subjects underwent bilateral knee MRI acquisitions on both 3T and 0.55T scanners. DL segmentation models were applied to segment knee bones and cartilage, and the segmentations were qualitatively and quantitatively evaluated. The results showed that the DL models exhibited a satisfactory capability to accurately segment and delineate bones and cartilaginous structures at 0.55T, although segmentation quality varied between different knee regions. While bone segmentation performance was generally better at 3T, cartilage segmentation was comparable between 0.55T and 3T, particularly for the femoral and tibial regions. However, the patellar region posed challenges, likely due to lower SNR at 0.55T. Despite these challenges, the study demonstrated the technical feasibility of using DL segmentation models trained on higher-field MRI data for knee imaging at 0.55T. This approach could facilitate knee biomarker quantification in low-field MRI settings.

In 2024, Plesniar et al. [29] assessed the diagnostic accuracy of 0.55T MRI in identifying periprosthetic joint infection (PJI) among patients with symptomatic total hip arthroplasty (THA). The analysis included MRI scans from three groups of patients: Group A consisted of individuals with confirmed PJI, Group B included patients with non-infected THA, and Group C was a subgroup of Group B with aseptic loosening of the implant. The radiologists evaluated the MRI scans independently, identifying various markers of infection. Capsule oedema, intramuscular oedema, and joint effusion were identified as the most effective discriminators for the diagnosis of PJI, exhibiting high sensitivity and specificity when tested individually and in parallel. Moreover, these markers demonstrated significance in differentiating between PJI and aseptic loosening. The results indicate that 0.55T MRI, particularly with its capacity to identify oedema in the joint capsule and adjacent muscles, has the potential to to aid in detecting PJI in symptomatic patients who have undergone THA.

Wiesmueller et al. [30] aimed to evaluate the usefulness of 0.55T MRI for diagnosing vestibular schwannomas (VS) in the internal auditory canal (IAC). The IAC of 56 patients with unilateral VS was imaged using both 1.5T and 0.55T MRIs. Several MRI sequences were independently assessed by two radiologists for image quality, visibility of VS, diagnostic confidence and artefacts that may be present while a second independent reading involved a direct comparison of images at both field strengths. The findings revealed that there was no substantial difference in image quality between the scans when considering factors such as visibility, diagnostic confidence, or the presence of artifacts. Additionally, direct comparison of the images did not indicate any significant difference in the ability to discern lesions or the level of confidence in making diagnoses. Consequently, low-field MRI demonstrated satisfactory diagnostic image quality and proved to be feasible in evaluating VS within IAC.

According to Kopp et al. [31], the 0.55T MRI for temporomandibular disorders (TMDs) was evaluated in comparison with the 1.5T standard MRI system. Temporomandibular Joint Disorder is a condition that is common causing functional restrictions and pains, therefore it often needs an MRI examination to assess the temporomandibular joint (TMJ) in detail. Low-field MRI has not yet been tested as a means of evaluating TMJ despite being a commonly used method in TMJ assessment using MRI imaging at 1.5T and 3T. On the same day, 17 patients with suspected intra-articular TMDs underwent both 0.55T and 1.5T MRI scans. Two senior readers independently assessed the image quality of the MRI scans while focusing on disc morphology and position as well as osseous joint morphology for each TMJ: by way of a four-point Likert scale, image quality and the degree of artefacts were evaluated for both field strengths. The results indicate that compared to those using 1.5T MRI, image quality at 0.55T was poor for disc morphology, and osseous joint morphology but almost similar for disc position between two field strengths. However, most cases were diagnosed correctly at both 0.55T and 1.5T since minor image artifacts were more common at 0.55T in comparison with 1.5T. This suggests that 0.55T MRI may help in diagnosing TMDs.

This subsection presents studies on the application of DL with brain MRI scans for neuroimaging purposes. The studies explored here are presented in Table LABEL:tab3. In 2021, Campbell-Washburn et al. [32] investigated the feasibility of using magnetic resonance fingerprinting (MRF) on a high performance 0.55T MRI system for rapid multi-parametric quantification. MRF is known for its versatility and ability to provide quantitative imaging at multiple field strengths. The study aims to demonstrate the potential of MRF for whole brain mapping on a modern 0.55T MRI system, which could improve examination efficiency and enable multi-contrast imaging on lower field strength systems. A whole-brain 3D stack-of-spirals fast imaging with steady-state precession (FISP) MRF sequence was implemented and quantification was validated using a standardised MRI phantom. The results showed strong correlations between MRF values and standard phantom measurements, with high precision despite the lower SNR at 0.55T. The study concludes that the proposed method is feasible at 0.55T, suggesting the potential for high-quality imaging at lower cost, making it attractive for various imaging scenarios.

In 2024, Kung et al. [33] conducted a study to assess the feasibility of performing diffusion tensor brain imaging at a lower magnetic field strength of 0.55T compared to the more commonly used 3T. Data were collected from a small group of five healthy subjects using both 0.55T and 3T scanners, with a focus on achieving high-resolution imaging with adequate SNR. To overcome the challenge of reduced SNR at 0.55T, a signal enhancement technique called SNR-enhancing joint reconstruction was applied. The study found that after applying this reconstruction method, the diffusion tensor parameters obtained at 0.55T correlated strongly with those obtained at 3T, demonstrating the feasibility of performing high-resolution diffusion MRI at lower field strengths. In addition, test-retest analysis indicated improved repeatability of diffusion tensor parameters with the SNR-enhancing reconstruction at 0.55T. These results suggest that, with appropriate noise reduction strategies, high-quality diffusion MRI of the human brain is achievable at 0.55T.

This subsection presents studies on the application of DL with MRI scans of the heart and blood vessels. The studies explored here are shown in Table LABEL:tab4. In 2020, Restivo et al. [34] investigated the benefits of using a 0.55T MRI for cardiac cine imaging. The researchers aimed to improve the SNR efficiency of long-read acquisitions, focusing on spiral and echoplanar imaging (EPI) techniques. They developed two high-efficiency bSSFP acquisition methods, a spiral in-out acquisition and an EPI acquisition incorporating cancellation techniques. The study included simulations, phantom images and images of 12 healthy volunteers to evaluate SNR and image quality. The results demonstrated that the spiral in-out bSSFP method performed well and produced high-quality images, while EPI bSSFP was more sensitive to movement and flow artefacts. The proposed method at 0.55T achieved a significant increase in myocardial SNR compared to the reference acquisition. The study demonstrated that the efficient spiral in-out bSSFP technique provides high-quality cardiac cine images and significant SNR recovery using a high-performance 0.55T MRI system.

Varghese et al. [35] developed a comprehensive cardiovascular magnetic resonance (CMR) exam protocol on a 0.55T MRI system, to overcome the challenges presented by lower field strength and slower gradient performance. The researchers constructed a prototype CMR exam on an ultra-wide bore 0.55T MRI system by leveraging compressed sensing-based techniques to compensate for the inherently lower SNR and slower imaging times. The study demonstrated that compressed sensing-based techniques enabled the acquisition of images of sufficient quality for the assessment of cardiovascular structure, function, and flow. These techniques also enabled spatial and temporal resolution suggestions comparable to those achievable with higher field strength systems.

In 2024, McGrath et al. [36] investigated the use of bSSFP in a 0.55T low-field MRI system for cardiac assessment. The researchers employed a free-running radial acquisition technique in conjunction with cardiac and respiratory auto-regulation methods to obtain reliable blood flow measurements and generate high-quality images of the heart. By implementing a phase-contrast bSSFP sequence in 0.55T and 1.5T MRI systems, the researchers were able to exploit the flow sensitivity of bSSFP and the phase cancellation properties. This approach enabled tests to be conducted without the need for ECG, thereby reducing examination times while maintaining accuracy. The in vivo data collected from healthy subjects demonstrated that the bSSFP method with phase contrast is comparable to measurements obtained with gradient-echo techniques, thereby confirming its viability in low-field MRI systems.

The ability of 12 lead Electrocardiogram (ECG) monitoring in a 0.55T MRI system for improved monitoring of the heart during an MRI scan was evaluated by Kolandaivelu et al. [37]. The 12-lead ECG is a basic instrument for observing cardiac ischemia and arrhythmia in many hospital settings. However, the strong magnetic field generated by the MRI equipment can distort the ECG signal and make it less understandable due to its non-interpretability. The objective of this study was to assess whether a 0.55T MRI scanner could provide superior ECG interpretation compared to higher field strengths. Experiments were conducted on pigs and healthy human volunteers, with ECGs being recorded inside different MRI scanners at various positions relative to the scanner. The study was designed to assess the occurrence and extent of ECG error and variation, with a particular focus on clinically relevant features. The results demonstrated that ECG distortion was significantly lower at 0.55T compared to higher field scanners, with distortion approaching acceptable levels even at the scanner table’s home position. Consequently, the 0.55 T MRI scanner can facilitate ECG interpretation and reduce the need for uncomfortable patient repositioning, thus increasing the feasibility of ECG monitoring during interventional MRI procedures.

Subsequent studies have concentrated on interventional cardiac MRI procedures. Basar et al. [38] conducted a study to analyse the susceptibility artefacts created by metallic markers and cardiac catheterisation devices in a 0.55T MRI system. These artefacts play a pivotal role in the visualisation of passive devices during MR-guided catheterisation. The researchers examined the field strength and orientation dependence of these artefacts to gain insight into their variations with different MRI field strengths. Gradient echo images were employed at 0.55T, 1.5T, and 3T to quantify the volume of artefacts produced by various types of passive susceptibility markers made from nitinol and stainless steel. Furthermore, the visibility of catheterisation devices currently available for clinical use at 0.55T and 1.5T was compared in phantoms and in vivo using real-time bSSFP imaging. The results demonstrated that the size of susceptibility artefacts is dependent on field strength and orientation. It is noteworthy that stainless steel markers with high susceptibility produced field-independent artefacts at various field strengths, suggesting magnetic saturation below 0.55T. Consequently, the size of the artefacts can be adjusted to ensure consistent imaging signatures of interventional devices at different MRI field strengths.

Rusche et al. [39] conducted a comparative study to assess the effectiveness of 0.55T MRI in imaging stroke. The study involved 27 participants, including stroke patients and control subjects, who underwent brain scans with both types of scanners. The data was analysed in three stages: qualitative assessment, lesion detection, and segmentation of lesion volume. The results demonstrated that the detection of lesions in diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) sequences was equivalent between the two scanners. The detection of lesions was also comparable. Noise levels were lower in the 0.55T sequences for DWI/ADC, but the 1.5T sequences were superior in other aspects. There was no difference in infarct volume between the two scanners. Thereby, low-field MRI at 0.55T could be a viable alternative for stroke imaging.

In a recent publication, Yildirim et al. [40] presented a novel interventional MRI guidewire designed for cardiovascular catheterisation at 0.55T. The guidewire is constructed from a metallic material and exhibits a continuous image profile from its shaft to its tip. The researchers conducted electromagnetic simulations and real-time MRI visibility tests to ensure that the guidewire’s shaft and tip exhibited continuous visibility. Subsequent tests of the guidewire’s RF-induced heating and mechanical performance demonstrated its safety and effectiveness. The detachable connector facilitates the replacement of the catheter while maintaining the position of the guidewire, thereby enhancing its utility in clinical settings. This innovative design allows for the safe navigation of luminal structures and cardiac chambers, offering similar performance to X-ray guidewires during MRI catheterisation.

In 2024, Uzun et al. [41] introduced a new method for tracking MRI intervention devices using markers printed with conductive ink, which are controlled by alternating current. The efficacy of the markers was evaluated on MRI-compatible specimens that are usually difficult to visualise during MRI scans. The visibility of the markers was evaluated using gradient echo, bSSFP and turbo spin-echo sequences with variable current parameters (amplitude, frequency) on 0.55T and 1.5T MRI scanners. A current supply circuit compatible with MRI technology was developed to regulate the current applied to the markers. The results demonstrated that adjusting the current parameters enhanced the visibility of the markers in all three sequences, thereby increasing their visibility. Furthermore, the study demonstrated that this method effectively increased the visibility of a customised nitinol needle in both in vivo and post-mortem animal experiments. In conclusion, current-controlled markers printed in conductive ink represent a simple and effective solution for tracking MRI intervention devices, adaptable to various parameters of the pulse sequence by adjusting the applied current.

This subsection explores studies on the application of DL in MRI examinations of the lungs and airways. The studies explored here are shown in Table LABEL:tab5. In 2020, Campbell-Washburn et al. [42] examined the potential of a 0.55T MRI for imaging lung diseases and compared its image quality with that of clinical CT scans. Despite the limited use of MRI in lung imaging compared to CT, recent advancements in ultrashort echo time (UTE) gradient-echo imaging show promise for generating high-quality T1-weighted lung images. The study describes a 0.55T low-field MRI system with extended upper-field homogeneity, which has the potential to reduce susceptibility artefacts in lung imaging. The study involved 24 participants with common lung anomalies. The researchers obtained high-quality structural MRI images of the lung in an average time of 11 minutes. The results demonstrated the ability to effectively detect a range of lung abnormalities in a low-field MRI. However, it proved more challenging to distinguish between diffuse diseases. Nevertheless, the study demonstrates that high-performance 0.55T MRI is promising for assessing common lung diseases, with strong correlations observed between nodule sizes measured on CT and MRI.

Bhattacharya et al. [43] proposed an ML method for segmenting cystic structures in the lung using MRI, with a focus on patients with lymphangioleiomyomatosis (LAM). Taking advantage of advances in low-field MRI technology, the study offers a radiation-free alternative to conventional CT imaging. A DL pipeline incorporating generative adversarial networks (GANs) and modified UNet architectures is used to accurately identify lung cysts and tissues in MRI scans. By using a CycleGAN for MRI to CT synthesis and subsequent segmentation with a modified residual UNet, the approach achieved good results, validated by high average Dice scores. The study included data from 65 LAM patients, with MRI and CT scans used for training and validation, demonstrating the impact of MRI-based cyst quantification methods. These findings have the potential to transform diagnostic practice by providing safer and more efficient tools for cyst analysis in clinical settings.

In 2022, a study was conducted by Azour et al. [44] to assess the efficacy of non-contrast 0.55T MRI in detecting and characterising ground-glass opacities and fibrosis in post-COVID patients. The MRI images were compared to clinically acquired chest CT images. The study involved 64 participants with a history of COVID-19 pneumonia, with 20 of them also undergoing chest CT scans. Two thoracic radiologists reviewed and scored the images for various parameters, including the presence, extent, distribution, and presence of fibrosis-like features. The inter-reader agreement for the CT images was found to be high, while that for the MRI images was found to be reasonable to moderate. The degree of agreement between CT and MRI scans was found to range from moderate to substantial. Despite the challenges associated with lung MRI, such as signal characteristics and respiratory motion, low-field 0.55T MRI has demonstrated potential for the detection and characterisation of lung abnormalities.

Javed et al. [45] developed a high-resolution imaging technique for pulmonary imaging at 0.55T. Their approach involved using a stack-of-spirals UTE sequence for continuous data acquisition. The acquired data was grouped into a stable respiratory phase using self-navigator signals. The corrections for trajectory errors and concomitant field artefacts were applied during image reconstruction. The study evaluated image quality using SNR, image sharpness, and qualitative reader scores in healthy volunteers, patients with COVID-19 infection, and patients with lung nodules. The technique yielded diagnostic-quality images with good SNR. The technique incorporates respiratory binning, which enhances image sharpness, and field corrections, which facilitate the observation of anatomical details in areas that are not situated in the centre of the image. The technique maintained image quality despite a slight reduction in SNR for shorter scan times. Inline image reconstruction and artefact correction can be achieved in less than five minutes. Consequently, the technique developed integrates several advanced methods to enable the acquisition of high-quality pulmonary images with a 1.75 mm isotropic resolution on a 0.55T MRI system within a scan time of 8.5 minutes. It employs an efficient stack-of-spirals imaging approach, coupled with respiratory binning to account for breathing motion, field correction to mitigate artifacts arising from field inhomogeneities, and trajectory correction to compensate for errors in the data acquisition trajectory.

In 2023, Seemann et al. [46] introduced a time-resolved 3D MRI technique capable of measuring the transient dynamics of water in the lungs during both resting and physically performing conditions. The objective was to detect early signs of heart failure manifested through exercise-induced dyspnea due to lung water accumulation. The study was conducted at 0.55T and involved 15 healthy subjects and two patients with heart failure. The results demonstrated that lung water density (LWD) exhibited an increase during exercise, particularly in healthy individuals, while patients exhibited slower accumulation rates. Furthermore, regional differences in the distribution of LWD between the anterior and posterior lungs were observed. These findings provide insights into the variations in lung water dynamics under different physiological conditions.

In their study, Li et al. [47] developed an improved low-field lung MRI technique using 0.55T MRI that utilises the transverse relaxation rates (R2 and R’2) of the lung parenchyma. The study was conducted on healthy adult volunteers and employed a bespoke single-slice spin-echo multi-echo (ES-MCSE) pulse sequence triggered by ECG on a prototype 0.55T whole-body MRI scanner. The results demonstrated that while the average R2 was comparable to that of 1.5T MRI, the R’2 was notably lower, indicating the potential for accurate measurements in 0.55T MRI. This research highlights the importance of the design and optimisation of new imaging methodologies for low-field lung MRI.

This subsection explores studies on the application of DL in fetal and pregnancy MRI scans. The studies explored here are presented in Table LABEL:tab6. In 2023, Ponrartana et al. [48] investigated fetal assessment with the 0.55T MRI system. The researchers presented an approach for achieving diagnostic-quality fetal MRI. In the past, fetal MRI was conducted on higher-field systems, such as 1.5T or 3T. Nevertheless, the advent of low-field strength 0.55T MRI systems presents several advantages, including enhanced field homogeneity, and a decrease in acoustic noise. These advantages enhance patient comfort and safety. The study outlines the experimental methods employed, including data collection on a 0.55T MRI system using fast sequence acquisition to mitigate the impact of fetal movement. The fetal MRI protocol incorporates a range of sequences, in addition to supplementary techniques such as DWI, hydrography, and cine imaging. Collectively, these techniques enable a comprehensive assessment of fetal health. The research demonstrates the potential of low-field MRI systems for conducting high-quality fetal MRI, providing a promising avenue for fetal imaging with an improved patient experience.

Verdera et al. [49] assessed the reliability and feasibility of low-field fetal MRI to evaluate anatomical and functional measurements in pregnant women. The study employed a 0.55T MRI scanner and a 20-minute protocol. The study included healthy pregnant women and women with pregnancy-related abnormalities, with a total of 79 fetal MRI scans performed. The 20-minute protocol, which included fast-spin-echo anatomical sequences, quantitative T2* and diffusion, produced key measurements such as biparietal diameter, transcerebellar diameter, lung volume and cervical length. These measurements were assessed by two radiologists and an obstetrician experienced in MRI. The results demonstrated concordance with large cross-sectional control studies for the anatomical measurements, and high inter-observer correlations were observed for the biparietal diameter. The functional characteristics, including the placental and brain T2* values and the placental ADC values, demonstrated a robust correlation with gestational age. The study demonstrated that the 20-minute low-field fetal MRI protocol reliably produces structural and functional measurements of the fetus and placenta during pregnancy.

In addition, Slator et al. [50] investigated the utilisation of diffusion MRI at 0.55T to generate comprehensive maps of placental structure and function. The study involved 57 placental MRI scans acquired using a combined T2* diffusion technique capable of simultaneously capturing multiple diffusion preparations and echo times. The results demonstrated that the quantitative parameter maps at 0.55T exhibited a high degree of similarity to those obtained at higher field strengths. This suggests that placental MRI at lower-field strengths is a viable and reliable option. The researchers identified several advantages of utilising lower-field strength MRI for placental imaging, including cost-effectiveness, ease of deployment, and greater patient comfort due to the wider bore.

Furthermore, Payette et al. [51] investigated the potential of fetal MRI at 0.55T in conjunction with advanced image analysis techniques for clinical examination. The study introduced a pipeline for automated segmentation and quantitative analysis, addressing some of the limitations of 0.55T fetal MRI applications. The researchers used this imaging pipeline to acquire multi-echo dynamic sequences of the fetal body. These dynamic sequences were subsequently reconstructed into high-resolution 3D volumes, providing both structural and quantitative information. Researchers developed a neural network model using a semi-supervised training approach to automatically segment fetal body volumes into ten different organ regions. The model demonstrated high accuracy in identifying and separating the various fetal organs from the volumetric data. The results demonstrated a robust correlation between T2* values and gestational age in key organs, including the lungs, liver, and kidney parenchyma. It is noteworthy that the pipeline demonstrated efficacy across a broad range of gestational ages (17-40 weeks) and exhibited resilience to motion artefacts. The study showed that low-field fetal MRI, when integrated with advanced image analysis methods, can offer comprehensive quantitative assessments and represents a viable option for clinical scanning applications. This approach holds promise for enhancing perinatal healthcare by providing valuable insights into fetal development and monitoring.

In a study published by Silva et al. [52], a real-time motion correction method was developed to improve functional fetal MRI scans. A 3D UNet was employed to track the position of the fetal brain in 125 datasets, thereby enabling real-time correction of the acquisition geometry. The method demonstrated high localisation accuracy, even for datasets of 0.55T, and was effective for fetuses under 18 weeks gestational age. This technique enhanced the reliability of fetal MRI and facilitated the characterisation of fetal brain development, with the possibility of clinical applications in conditions such as pre-eclampsia and congenital heart disease.

In 2024, the same author [53] developed a fully automated real-time planning system for fetal brain MRI on a 0.55T scanner to broaden the accessibility of fetal MRI. The method involved the use of DL to automatically identify crucial brain landmarks on whole-uterus echo planar imaging scans. The successful detection of these anatomical landmarks facilitated the subsequent automatic planning and guidance of radiological image acquisitions. The pipeline was trained on over 120 datasets and subsequently evaluated for both accuracy and quality. The prospective testing of the system in nine fetal subjects demonstrated that real-time planning could be performed without any latency. The accuracy of landmark detection ranged from 4.2 ± 2.6 mm for the fetal eyes to 6.5 ± 3.2 mm for the cerebellum. The quality of the planned images and the diagnostic images were found to be comparable to those produced by manual planning. The successful implementation of real-time automatic planning has the potential to simplify fetal MRI acquisition and increase its availability in non-specialist centres.

In a recent publication, Uus et al. [54] presented a novel approach to fetal MRI imaging that involves real-time 3D reconstruction of the fetal brain and body using a 0.55T MRI scanner. The researchers sought to address the fetal motion artefacts, which often compromise image quality and diagnostic accuracy. The method involved integrating advanced slice-to-volume reconstruction techniques into the Gadgetron framework, a platform designed for MRI data processing. By combining automated deformable and rigid reconstruction methods with DL algorithms, the research team has developed a workflow capable of processing motion-corrupted MRI data in real-time. This innovation not only allows for the generation of high-resolution 3D images during the scan but also enables motion correction, thereby enhancing the diagnostic utility of fetal MRI.

This subsection analyses studies on the application of DL in brain MRI scans. The studies explored here are shown in Table LABEL:tab7. In 2019, Jung et al. [55] developed a DL method to enhance the visualisation of perivascular spaces (PVS) in human brain MRI. PVS is associated with a number of brain disorders, but their quantification is challenging due to their thin and indistinct appearance. The proposed method utilises a complex 3D CNN with dense, hopping connections, which enables accurate prediction of the enhanced image. The networks effectively exploit contextual information from low to high-level features, resolving the issue of gradient disappearance in deep layers. The research team evaluated the method using 17 7T MRI images through double cross-validation. The results demonstrated that the proposed network significantly outperforms previous approaches in increasing the visibility of the PVS.

In 2020, Zong et al. [56] investigated the morphology of PVS and enclosed blood vessels in healthy adults aged 21-55 years using ultra-high-resolution MRI at 7T. The researchers observed age-dependent changes in PVS characteristics, including an increased count and volume fraction in the basal ganglia, as well as spatial heterogeneity. The use of carbogen breathing induced dynamic alterations in PVS morphology, indicating that this system is sensitive to physiological changes. These results provide valuable insights into the glymphatic system and its potential implications for brain health.

Meliadò et al. [57] introduced a DL approach for image-based subject-specific assessment of local SAR, which is crucial for ensuring patient safety during MRI procedures, especially at 7T. In the past, SAR assessment has been conducted using offline numerical simulations with generic body models, which has led to uncertainties due to inter-subject anatomical variability. The proposed method employs a CNN trained on subject-specific maps and corresponding local SAR distributions, thereby enabling more accurate and efficient SAR estimation. By incorporating additional penalties for SAR underestimation errors in the loss function, their model achieves remarkable results in both in silico and in vivo validations, demonstrating excellent qualitative and quantitative agreement with ground truth data. This approach not only enhances the accuracy of SAR assessment but also reduces examination protocol time by nearly 25%, thus improving patient safety and MRI efficiency in clinical practice.

Yang et al. [58] introduced the Automated Hippocampal Subfield Segmentation Toolbox (CAST), a deep CNN designed for the segmentation of hippocampal subfields in 7T and 3T MRI datasets. The CAST model is adaptable, enabling researchers to train it with their data and accommodating various modalities. The results demonstrated that CAST exhibited comparable segmentation accuracy to a standard method on the 7T and 3T datasets. The CAST method reliably segments subfields, prioritising efficiency and accuracy, particularly in 7T MRI applications. The toolbox employs a multi-scale deep CNN to capture global structural information while efficiently using computational resources. Therefore, CAST markedly enhances the reliability of subfield segmentation and ensures consistent segmentation quality across diverse datasets.

Bazin et al. [59] developed a method for measuring image sharpness in 7T MRI images. The aim of the study was to investigate the association between subject movement and image sharpness, with a particular focus on studies of brain structure. The method was assessed in 24 healthy volunteers, including individuals of varying ages. The results demonstrated a correlation between motion and sharpness enhancement, with motion correction exhibiting a pronounced impact on sharpness. Following motion correction, the full width at half maximum was decreased from 0.88 mm to 0.70 mm, resulting in a 2.0 times smaller voxel volume. This method provides valuable insights into the impact of motion on image quality in 7T MRI studies.

Kim et al. [60] presented DCN-Net, a DL framework designed for the accurate and rapid segmentation of cerebellar dentate nuclei in 7T diffusion MRI. The DCN-Net employs a hybrid loss function to address the issue of highly imbalanced data, and it learns the probability of each label independently during training, including those for dentate and intermediate nuclei. To address the limited availability of labelled data, auto-training strategies are employed. These strategies generate auxiliary labels for unlabelled data using DCN-Net, which has been trained with manual labels. The experimental results, based on 60 individuals, demonstrated that DCN-Net outperforms other segmentation tools and deep neural networks in terms of accuracy and consistency.

In 2021, Demirel et al. [61] presented a new method for enhancing 7T MRI technology, achieving a 20-fold acceleration in imaging while maintaining high image quality. This development is important for enhancing spatial and temporal resolution in brain imaging, and for accurately mapping neural activity. The researchers addressed the limitations of traditional techniques, which suffer from aliasing and noise at higher rates. The study demonstrates significant improvements over existing methods by employing a self-supervised DL reconstruction method that uses only undersampled datasets. This offers comparable functional precision and temporal effects to standard 10-fold accelerations.

Trutti et al. [62] presented a probabilistic atlas of the human ventral tegmental area (VTA) using 7T MRI data, addressing the need for anatomically atlases of subcortical structures. While traditional neuroanatomical atlases are invaluable for localising brain regions using functional MRI, they often lack coverage of subcortical nuclei, limiting their utility for deep brain imaging. The researchers developed a multimodal 7T MRI protocol for enhanced imaging and delineation of the VTA, leveraging ultra-high field resolution. This was followed by the creation of a comprehensive probabilistic VTA atlas.

In a subsequent study, Morris et al. [63] utilised high-resolution 7T MRI to investigate the integrity of the VTA in individuals with and without depression and anxiety disorders. By employing quantitative MRI and dopamine-related signal mapping, they were able to delineate the VTA and correlate its structural integrity with measures of intrinsic and extrinsic motivation assessed through cognitive tasks. The results demonstrated that subjects with mood and anxiety disorders exhibited larger VTA volumes but lower signal intensity within the VTA compared to healthy controls. This indicates reduced structural integrity of the dopaminergic VTA in these individuals. It should be noted that VTA integrity was not directly associated with self-reported symptoms of depression or anxiety. However, it did show a correlation with objective measures of extrinsic motivation, suggesting a link between neural integrity and behaviour in clinical and non-clinical groups.

In 2022, Miletić et al. [64] conducted a study using 7T functional MRI to investigate the tripartite model of the subthalamic nucleus during a speeded decision-making task. The tripartite model, which suggests cognitive, limbic, and motor subregions within the subthalamic nucleus, is typically supported by anatomical studies and research on clinical patients. In this study, 34 healthy participants were asked to complete a random-dot motion task with varying levels of difficulty and choice payoffs. The objective was to differentially engage the cognitive and limbic networks while recording motor responses using the left and right index fingers. The analysis of the signals in manually delineated subthalamic nucleus subregions revealed that all segments responded similarly to the experimental manipulations. The study found no evidence to support the existence of distinct functional subregions, challenging the tripartite model by demonstrating uniform activity across the subthalamic nucleus subregions in response to the task.

Rodrigues et al. [65] proposed a benchmark for hypothalamus segmentation on 7T MRI images, addressing the challenges of high variability and subjectivity in manual segmentation due to unclear morphological landmarks. The study presents a diverse dataset of 1381 individuals from various sources. The researchers proposed a teacher-student model with segmentation and correction blocks, which they demonstrated to have improved generalisation ability. The method was trained on the datasets, achieving Dice coefficients of 0.83 on familiar datasets and 0.74 on the unseen dataset. This demonstrates the effectiveness and reliability of the approach, offering a valuable tool for research in automated hypothalamus segmentation on ultra-high-resolution 7T MRI images.

Wei et al. [66] developed CaNes-Net, a cascaded nested network designed for segmenting 3T MRI images. The authors addressed the issue of low tissue contrast and anatomical variability in 3T images by utilising nested networks that iteratively refine the segmentation. The network is trained using the higher contrast of 7T MRI images. To address the issue of misalignment between 3T and 7T images, the researchers have incorporated a correlation coefficient map. This map enables the segmentation accuracy to be improved, with results that are superior to those achieved by popular tools. The CaNes-Net is constituted of two nested networks: The Nes-Net handles the initial segmentation, while another Nes-Net refines the segmentation using tissue-specific distance maps. This iterative process allows for the gradual improvement of the segmentation results, ensuring optimal results are achieved. The CaNes-Net was tested on 18 adults and the ADNI dataset, demonstrating reduced segmentation errors and superior performance compared to widely used methods. This highlights the effectiveness of incorporating 7T images in the segmentation process.

In 2023, Doss et al. [67] conducted a study to develop a DL network for accurately segmenting the basal nucleus of Meynert in 3T MRI scans. This structure is important for cognitive function and arousal. The small size, patient variability, and low contrast on 3T MRI make accurately segmenting this structure a challenging task. The researchers employed a dataset including 21 healthy individuals, each with paired 3T and 7T MRI scans. The 7T scans were annotated, resulting in accurate labels for the 3T images. A 3D-Unet CNN was constructed and validated using a 5-fold cross-validation approach. The model demonstrated superior performance compared to the standard probabilistic atlas, with higher Dice coefficients and reduced centroid distances observed in both healthy subjects and patients with temporal lobe epilepsy.

Zhang et al. [68] developed a model, the quantitative Recurrent Inference Machine (qRIM), which addresses the challenges of long scan times in 7T MRI. This model employs DL techniques to reconstruct images and map values from sparse data. Integrated into a Recurrent Inference Machine, qRIM solves the inverse problem iteratively, making use of data redundancy and prior information from the relaxometry model. Experimental evaluations using ultra-high-resolution multi-echo gradient echo data have demonstrated that qRIM markedly reduces reconstruction error in comparison to sequential reconstruction and parameter fitting methods. Moreover, it outperforms other architectures, such as U-Net and Compressed Sensing, and can be used to investigate human subcortex activities across all age groups. Although the quantitative end-to-end variational network exhibits slightly inferior reconstruction quality in comparison to qRIM, it nevertheless outperforms alternative methods. In general, qRIM demonstrates superior performance with increasing acceleration factors and benefits the reconstruction quality of sparser data.

Quantitative susceptibility mapping (QSM) is a non-invasive MRI technique that makes it possible to measure the spatial distribution of magnetic susceptibility in an object. Chen et al. [69] evaluated various QSM methods to assess radiation-induced cerebral microbleeds and basal ganglia at both 3T and 7T MRI field strengths. The study tested four background phase removal methods and five dipole field inversion algorithms, finding that 7T MRI provided a lower noise level compared to 3T. The QSIP (Quantitative Susceptibility mapping by Inversion of a Perturbation) and VSHARP (Sophisticated Harmonic Artifact Reduction for Phase) + iLSQR (iterative Least Squares with QR factorisation) methods yielded the best results for white matter homogeneity and vein contrast, with QSIP also giving the highest cerebral microbleeds contrast. The results demonstrated that VSHARP + iLSQR produced the most accurate susceptibility values across all regions, closely matching the ground truth. This research provides guidance on selecting the most appropriate QSM processing pipeline for specific applications and MRI field strengths.

In 2023, Ramadass et al. [70] investigated the challenges and solutions to achieving accurate segmentation of the whole brain using 7T structural MRI. The study showed the potential of 7T MRI to enhance the comprehension of brain function due to its superior resolution and contrast. However, existing segmentation tools encounter difficulties when confronted with inhomogeneous image contrast on 7T MRI. To address this issue, the researchers proposed a solution: registering the segmentation of the whole brain in 3T to 7T MRI and adjusting the model accordingly. This approach led to a significant reduction in the average relative difference in label volumes and an improvement in the Dice similarity coefficient, indicating a more accurate segmentation. The results demonstrate that fine-tuning the model is an effective method for improving whole-brain segmentation in 7T MRI, thereby enabling more detailed and accurate neuroimaging analyses.

Arledge et al. [71] developed a DL approach to automatically generate vascular pharmacokinetic parameter maps from dynamic contrast-enhanced MRI images in mouse brain tumour models. The approach was validated using 7T MRI scans performed on glioma in nude mice. The researchers trained a CNN on dynamic contrast-enhanced images to capture multiscale features. The CNN demonstrated promising results, accurately estimating vascular pharmacokinetic parameters and matching conventional pharmacokinetic models. Furthermore, the DL approach was successfully applied to a mouse model of breast cancer brain metastases, indicating its potential for assessing vascular permeability in different types of brain tumours.

Brink et al. [72] developed a method for personalising local SAR predictions for parallel transmit (PTx) neuroimaging at 7T, utilising a T1-weighted dataset. The researchers addressed the challenge of regulatory compliance and RF performance limitations presented by intersubject variability in SAR. By employing DL, they generated subject-specific body models, thereby overcoming the limitations of the conventional "one-size-fits-all" approach. The CNN developed by the researchers trained on multi-contrast 7T MRI data, achieved accurate segmentations, addressing the intensity bias and contrast variations that are typical at 7T. The efficacy of personalised local SAR predictions was demonstrated through evaluation in PTx configurations.

In a subsequent study, Gokyar et al. [73] presented a DL-based method for predicting SAR distributions in the human head for 7T MRI PTx systems. At 7T, the high intensity of the magnetic field results in a non-uniform deposition of electromagnetic energy, which can lead to the formation of local high SAR hot spots. To address this issue, the researchers developed a multichannel 3D CNN designed to predict local SAR maps and peak spatial SAR levels. The CNN architecture demonstrated superior prediction accuracy, with an average reduction in overestimation errors of 20% compared to 152% with traditional methods. The DL model demonstrated a reduction in prediction uncertainties and was able to perform predictions in a relatively short time.

In their study, Liu et al. [74] proposed a new unsupervised learning method, self-labelled encoder-decoder (SLED), to enhance the reconstruction of myelin water image (MWI) maps from multi-echo gradient echo sequences. The objective of this study was to address the challenges associated with stability, accuracy, and SNR in adapting MWI data. The imaging data was obtained utilising a 7T MRI system designed for preclinical research applications. The implementation of the SLED network enabled the researchers to demonstrate an improvement in the estimation of myelin water fraction (MWF) compared to the traditional approach. Experiments conducted on ultra-high resolution MWI data from rat brains and a simulated MWI phantom demonstrated that SLED exhibited superior stability and accuracy in MWF estimation. The generated maps exhibited less noise and demonstrated greater tolerance to low SNR data. It is of particular significance that SLED’s unsupervised, self-labelling approach represents a distinctive alternative for the analysis of MWI.

In 2023, Tsuji et al. [75] investigated the use of DL to enhance image quality and reduce acquisition time in preclinical 7T MRI, focusing on brain images of rats. The researchers employed a residual dense network (RDN) to reconstruct images with a low number of excitations (NEXs), to match the quality of high NEX images while reducing scan time. By training the RDN with low NEX images as input and high NEX images as the reference, they achieved significant improvements in SNR, peak SNR, and SSIM. The results demonstrated that the enhanced images produced by the RDN exhibited higher quality, with the SNR for NEX = 2 increasing from 10.4 to 32.1, surpassing even the high NEX (NEX = 12) SNR of 19.6. Furthermore, the method reduced the scanning time by 83%, demonstrating its potential for efficient and high-quality MRI imaging.

In a more recent study, Jha et al. [76] examined the diffusion MRI and its application in the assessment of white matter regions of the brain. The researchers proposed a novel approach to transform diffusion MRI data acquired at 3T to achieve a quality comparable to that obtained at 7T. This is achieved by combining the trapezoidal rule with graph-based attention modules. This transformation is of significant importance, as higher magnetic field strengths, such as 7T, afford greater tissue contrast and spatial resolution, thereby enabling more accurate estimation of structural details within the brain. However, due to the cost of 7T scanners, there is a pressing need for algorithms to enhance the quality of 3T diffusion MRI data. The proposed CNN architecture was designed to resolve this limitation and the results confirm that it is a viable alternative.

An et al. [77] developed a DL-based method for fully automated segmentation of ischemic stroke lesions in 7T MRI data acquired from mouse models. This approach addresses the labour-intensive and variable nature of manual segmentation by trained experts. Utilising a dataset of 382 MRI scans, the researchers split the data into 293 scans for training and 89 for evaluation. The evaluation of the automated segmentation technique involved comparing its output with manual segmentation by calculating Dice coefficients and assessing the accuracy of lesion volume measurements. Additionally, the automated method was further tested on an independent dataset with different imaging characteristics to assess its robustness. The automated segmentation technique produced segmentation masks that appeared smooth, compact, and realistic, achieving a high level of agreement with the manual segmentation performed by human experts. Remarkably, the Dice scores surpassed those between human experts in previous studies, indicating the method’s potential to eliminate human bias and standardise lesion segmentation across various research studies and centres.

This subsection examines studies on the application of DL in muscle and skeletal MRI scans. The studies presented here are shown in Table LABEL:tab9. In 2021, Belyk et al. [85] conducted a study investigating the coordination of vocal and respiratory control in the human brain. The researchers focused on the larynx motor cortices involved in vocal-motor control and employed a 7T functional MRI to localise and understand the function of these cortical regions. The study involved 13 participants who were required to complete a series of tasks involving singing and whistling. In contrast to previous assumptions, the study found that both the dorsal and ventral larynx motor regions are activated during whistling, despite the decreased engagement of the larynx. This indicates that both areas exhibit a shared property of laryngeal-respiratory integration.

Tanaka et al. [86] sought to discern whether the initial reduce in the blood oxygenation level-dependent signal contains information that can be decoded to identify facial expressions. The researchers employed multivoxel pattern analysis and hyperalignment techniques on 7T functional MRI data to analyse brain activity during a facial expression discrimination task. The results demonstrate that the decoding accuracy in the bilateral amygdala exceeds chance levels 2 seconds after face onset. This indicates that the initial reduction in blood oxygenation level-dependent signals contains more precise information about task events and cognitive processes.

In 2022, Terraciano et al. [87] developed a 7T MRI protocol using delayed gadolinium-enhanced MRI of cartilage to quantitatively assess post-traumatic osteoarthritis (PTOA) in a rabbit model. Following unilateral anterior cruciate ligament (ACL) transactions in rabbits, MRI examinations were performed at various intervals using a modified delayed gadolinium-enhanced MRI of cartilage protocol to study PTOA development. Voxel-based maps were generated to evaluate the changes in cartilage. The study demonstrated a decline in the delayed gadolinium-enhanced index in the medial compartment of the knee following ACLT, indicative of cartilage degeneration. Histological analysis confirmed the presence of PTOA in ACLT knees, with osteophytes observed exclusively in these knees. This approach demonstrates the feasibility of detecting and monitoring PTOA using quantitative 7T MRI, which has the potential to facilitate therapeutic interventions and advance our understanding of PTOA progression.

This subsection analyses studies that aim to solve the problem of the inhomogeneity of the MRI magnetic field. The studies presented here are shown in Table LABEL:tab10. In a study published in 2021, Haast et al. [88] examined the impact of MP2RAGE B1+ sensitivity on T1 reproducibility and hippocampal morphometry in 7T MRI. MP2RAGE stands for Magnetization-Prepared 2 RApid Gradient Echo. The sequence is designed to provide high-resolution anatomical images with improved contrast and reduced bias from factors such as magnetic field inhomogeneities and transmit field (B1+) variations. The study, conducted across multiple sites, aimed to validate the effects of specific MP2RAGE parameter choices on image quality and cortical T1 estimates, while also evaluating their consequences on hippocampal segmentation results. The results of the study underline the importance of optimising MP2RAGE sequences for neuroimaging studies, particularly in the context of 7T MRI, highlighting the need for standardisation and harmonisation of MRI protocols at different imaging sites.

In 2022, Ma et al. [89] introduced a DL approach aimed at mitigating transmit-B1 (B1+) artefacts in MRI without relying on PTx hardware. The method involved training a deep neural network to predict PTx-style images, and it demonstrated promising results in improving diffusion tensor imaging analysis. This was achieved by effectively restoring signal dropout and improving image quality. Furthermore, the trained DL model was able to successfully predict PTx images for new subjects, thereby demonstrating its potential to reduce B1+ artefacts even in the absence of PTx resources.

Hunger et al. [90] developed a DL approach, named deepCEST 7T, to enhance the speed and precision of chemical exchange saturation transfer MRI parameter mapping at 7T. This method employs a neural network to address the pronounced B1 inhomogeneity challenges that are characteristic of 7T. The network is trained to predict parameters from a single scan using 7T in vivo and B1 maps as input. This process results in a reduction in scan time to 6 minutes and 42 seconds while maintaining the ability to correct for both B0 and B1 inhomogeneities. Xu et al. [91] presented the autoHOS, an automated high-order shimming method tailored for addressing B0 inhomogeneity in MRI and MR spectroscopy at 7T. This approach combines DL based brain extraction with image-based high-order shimming, eliminating manual ROI selection. In vivo studies at 7T have demonstrated that the autoHOS method enhances image and spectral quality in comparison with the traditional linear and manual shimming methods.

Zhao et al. [92] developed a distortion-free whole-cerebrum 3D pseudo-continuous arterial spin labelling (pCASL) sequence at 7T to overcome B1/B0 inhomogeneities. By optimising pCASL labelling parameters, background suppression pulses, and an accelerated Turbo-FLASH readout, they achieved high-resolution pCASL imaging with whole-brain coverage and detailed perfusion information. The proposed technique demonstrated excellent test-retest repeatability and improved SNR compared to similar sequences at lower field strengths, offering promising prospects for cerebral perfusion imaging at ultrahigh field strengths.

In 2024, Kilic et al. [93] proposed an approach to mitigate B+1 inhomogeneity in multi-channel transmit arrays at 7T using unsupervised DL with CNNs. This approach differs from previous methodologies, which relied on supervised training, by employing unsupervised training and CNNs for multi-channel B+1 maps. By concatenating B+1 maps along the spatial dimension, the method enables shift-equivariant processing, which is suitable for CNNs. The training process employs a physics-driven loss function, which is designed to minimise the discrepancy between the Bloch simulation and the target magnetisation. The proposed method has been shown to offer superior performance compared to traditional approaches in the domain of static PTx design. The method offers enhanced speed and resilience while preventing the formation of local phase singularities.

In 2022, Plumley et al. [94] presented a new approach to address the sensitivity of head movement in adapted PTx pulses used for uniform excitation profiles at 7T. The researchers proposed a solution that involves reshaping pulses in real time using a DL framework. To train their model, simulated data is used, and conditional GANs are employed to predict B+1 distributions when there is head displacement. The accuracy of the predicted maps is evaluated against true B1 maps, considering both magnitude and phase. The study found that the predicted B+1 maps aligned well with the true maps and exhibited lowererror than motion-related errors in the majority of evaluations. Furthermore, the use of the predicted maps to redraw the pulses resulted in a significant reduction in the worst-case error of the inversion angle due to movement. Consequently, the proposed framework offers a solution for the real-time prediction of B+1 maps, which can be employed to redesign personalised pulses and mitigate errors related to head movement in PTx applications.

In their latest publication, Motyka et al. [95] have developed a DL approach to anticipate changes in the brain’s B0 field caused by subject movement during 7T MRI scans. The method utilises a 3D U-net trained on 7T MRI data to accurately predict B0 field alterations based on variations in head position. By fine-tuning the network weights to individual subjects, they achieved superior performance compared to traditional methods, enhancing spatial resolution. This study presents a promising solution for mitigating B0 inhomogeneities induced by subject motion, offering potential applications in retrospective or real-time correction during 7T MRI examinations.

This subsection explores studies on the application of DL in cardiac MRI examinations. The studies presented here are shown in Table LABEL:tab11. In 2021, Ankenbrand et al. [96] explored the transformative potential of DL for the automatic segmentation of cardiac cine MRI images. Employing transfer learning techniques, the researchers adapted existing models to the realm of 7T MRI. By leveraging a publicly available segmentation model to annotate a dataset and subsequently training a neural network to segment the left ventricle and myocardium, the study revealed that transfer learning significantly boosted model performance, even when applied to a reduced dataset. Furthermore, this research provided practical guidelines and resources to facilitate similar endeavours in the future.

Van der Brick et al. [97] presented the principles and design of the ZOOM@SVDs research, which aims to determine measures of cerebral small vessel dysfunction using 7T MRI as a new method for small vessel disease (SVD). The study includes an observational cohort of patients with cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy, sporadic SVDs, and healthy controls. The study uses a 7T brain MRI to evaluate features of small vessel function such as its reactivity, cerebral punching artery flow and pulsatility. The evaluation included clinical and neuropsychological assessments, as well as conventional 3T MRI to evaluate conventional imaging markers of SVD. In conclusion, by comparing measures of small vessel dysfunction between patients and controls and correlating them with clinical and MRI manifestations of SVD, ZOOM@SVDs identified novel markers of cerebral small vessel function.

In a complementary study, Garcia-garcia et al. [98] introduced Vessel Distance Mapping (VDM) to assess blood supply and its relationship to cognition. The researchers aimed to overcome previous controversies regarding the connection between cerebral blood flow and cognitive function. VDM provides a quantitative approach to evaluating vascular patterns about surrounding structures, extending the previous binary classification into a continuous spectrum. Using 7T MR angiographic imaging, the researchers manually segmented hippocampal vessels in older adults with and without cerebral SVD. VDM metrics, which indicate vessel distances, were found to correlate with cognitive results in individuals with vascular pathology but not in healthy controls. The study highlights that both vessel pattern and vessel density contribute to cognitive resilience. VDM provides a reliable method for vascular mapping, enabling the investigation of various clinical research questions related to cerebral blood flow and cognition.

In 2023, Chowdhury et al. [99] conducted an investigation into the detection of R peaks in ECG waveforms that had been distorted by the magnetohydrodynamic effect during MRI scans, with a particular focus on 3T and 7T magnetic fields. The researchers proposed the Self-Attention MHDNet model, which was designed to accurately identify R-peaks from distorted ECG signals. The model demonstrated high recall and precision rates, with 99.83% and 99.68% in 3T settings, and 99.87% and 99.78% in 7T settings, respectively. This improved detection capability can enhance the precision of functional cardiovascular MRI by ensuring the reliable triggering of the pulse. The study demonstrates the potential of integrating machine learning techniques to mitigate ECG signal distortions in the MRI environment, thereby facilitating enhanced patient monitoring and diagnosis.

In 2024, Kolmann et al. [100] sought to improve the reproducibility of image segmentation for myocardial tissue analysis in a large animal model of myocardial infarction using CMR imaging at 7T. The researchers discovered that traditional automated segmentation approaches, designed for human imaging at clinical field strengths, encountered difficulties when applied to preclinical data and ultra-high field strengths. To address this issue, the researchers re-trained a DL model using their preclinical 7T data and tested it on a separate dataset. The DL segmentation showed excellent agreement with manual segmentation, as evidenced by a high Pearson correlation coefficient, intraclass correlation coefficient and low coefficient of variability for ejection fraction. The Dice scores for left ventricular and myocardial segmentation were also high. This approach shows great promise for improving the analysis of cardiac function, particularly in large animal models of myocardial infarction.

This subsection explores studies on the application of general DL, to generate and reconstruct reliable MRI images. The studies presented here are shown in Table LABEL:tab12. In 2018, Nie et al. [101] proposed a technique to generate medical images using an FCN from a source image. The researchers used adversarial learning strategies to improve the realism of the generated images. In addition, a loss function based on the image gradient difference was implemented to prevent the generation of blurred target images. The method proposed in this study demonstrated both accuracy and robustness in experiments performed on a variety of datasets. Consequently, these tasks, such as the synthesis of CT images from MRI scans and the generation of 7T MRI images from 3T MRI scans, demonstrated superior performance compared to existing approaches.

In a study addressing the issue of accessibility to expensive 7T MRI scanners, Qu et al. [102] conducted a study. A DL network was developed to synthesise high-quality 7T MRI images from their 3T counterparts. The network incorporates information from both the spatial and wavelet domains, including a novel wavelet-based affine transformation layer. The network’s use of the wavelet transform for effective multi-scale reconstruction improves tissue contrast and anatomical detail in synthesised 7T images. The results show the effectiveness of the proposed method, which outperforms existing techniques and offers a promising solution to the limited accessibility of 7T MRI scanners.

In 2024, Cabini et al. [103] presented a DL model to enhance the reconstruction of T1 and T2 maps obtained through MRF on a 7T MRI machine. The implementation of a hyperparameter optimisation strategy enabled the model to achieve enhanced reconstruction accuracy and speed for preclinical investigations. In comparison to traditional methodologies, the proposed DL approach demonstrated a reduction in reconstruction errors. This was achieved through the optimisation of neural network architectures, model structures, and supervised learning algorithms. Furthermore, it enhanced computational efficiency, rendering it suitable for real-time applications. The results demonstrated that the DL method outperformed dictionary-based methods, resulting in a reduction of the mean percentage relative error. These findings indicate the potential of the DL model to significantly advance MRI reconstruction techniques.

Eidex et al. [104] proposed a hybrid CNN-transformer model to generate high-resolution 7T ADC maps from multimodal 3T MRI data. This addresses the limitations of 7T MRI, such as limited availability and higher costs. Their Vision CNN-Transformer model, which combines Vision Transformer blocks for global context and convolutional layers for fine details, was validated using the Human Connectome Project Young Adult dataset. The model was trained on 130 cases, validated on 20, and tested on 21. The Vision CNN-Transformer model achieved a PSNR of 27.0 ± 0.9 dB, SSIM of 0.945 ± 0.010, and MSE of 2.0E-3 ± 0.4E-3, outperforming other methods and reducing training time by 50%. The findings demonstrated that the Vision CNN-Transformer model was capable of effectively synthesising 7T quality ADC maps from 3T data, which could potentially enhance disease diagnosis and intervention.

Although the integration of DL into MRI is very promising, it is not without its challenges. The demand for substantial computational resources, the complexity of training algorithms, and the potential for algorithmic bias represent significant considerations that must be addressed [105]. Additionally, the high field strengths inherent to 7T MRI can aggravate side effects and image artefacts, requiring the implementation of advanced management and correction techniques. For 0.55T MRI, the primary challenge lies in achieving image quality comparable to higher field strengths. Although DL techniques have shown the potential to fill this gap, further validation and optimisation are needed to guarantee clinical reliability [106]. Moreover, the complexity of DL models and the need for extensively annotated datasets present obstacles for both 0.55T and 7T MRI.

Looking ahead, future research efforts should focus on optimising DL algorithms adapted to specific MRI modalities, addressing computational challenges and improving the robustness and reliability of DL-based imaging techniques. Furthermore, the clinical validation of these integrated technologies must be extended to enable widespread adoption and implementation in routine clinical practice [106], [107]. To advance the integration of DL into MRI, future research should prioritise the following areas:

• •

  Data Sharing: Establishing large-scale, multi-centre collaborations to create comprehensive, annotated datasets will be crucial for training robust DL models that can be generalised effectively.

• •

  Model generalisation: Developing and validating DL models that can be generalised across diverse populations and clinical contexts, ensuring widespread applicability and equitable access to these cutting-edge imaging solutions.

• •

  Cost-Benefit Analysis: Conducting detailed cost-benefit analyses to evaluate the economic impact of implementing advanced DL techniques in resource-constrained settings, particularly for 0.55T MRI, which holds potential for expanding access to high-quality imaging in poorly served regions [107].