Typical and minimum voxel sizes for various field strength MRI machines?

Typical and minimum voxel sizes for various field strength MRI machines?

We are searching data for your request:

Forums and discussions:
Manuals and reference books:
Data from registers:
Wait the end of the search in all databases.
Upon completion, a link will appear to access the found materials.

Hoping somebody can give me some real world numbers on the typical capabilities of 3T and 7T MRI machines in terms of there minimum voxel size (specifically in relation to there use in fMRI studies if that makes a difference)? The Wikipedia article on MRI gives a rather non committal range of 1mm to 5mm on a side, but I have been reading the term High-Resolution fMRI in a few places without any qualification being placed on it.

This is a slightly older post, but the Wikipedia article is basically correct. It is a non-committal range of 1mm - 5mm. Typically high resolution fMRI will probably be <= 1mm on a side, but is typically going to be higher field scanners (>>= 3T) or using slightly different sequences than the standard fMRI sequence.

Best thing to do would be to do a quick survey of the literature (e.g., and get an intuition from there. I think what you will find is it will depend a lot of sequence, sequence parameters and main magnet field strength.


The middle cerebral artery occlusion (MCAO) rat model is the most widely used animal model for investigating the pathophysiological mechanisms that underlie human brain injury after ischemic stroke and for developing effective therapeutic approaches to the disease. 1,2 Diffusion-weighted imaging (DWI) of magnetic resonance imaging (MRI) can identify ischemic lesions with cytotoxic edema from very early on after MCAO, especially within the first 5� min after arterial occlusion. 3 In addition, DWI can be used to visualize the location and volume of ischemic lesions and provide a temporal evolution of focal cerebral ischemia in the rat brain. 2 𠄴

Several quantitative DWI studies have used dedicated, small-bore, high Tesla (T) MRI scanners with strong gradient coils for evaluating small animals with experimental stroke however, 3.0-T clinical MRI scanners can provide comparable spatial resolution with the use of small animal coils. 4 𠄶 However, the use of 3.0-T MRI scanners in quantitative DWI studies has some disadvantages, such as a prolonged scan time and scanner-dependent variation. 7,8 The apparent diffusion coefficient (ADC) values of normal brain tissue and ischemic lesions in various animals have wide ranges, overlap, vary greatly over time and show inter-subject variation. 7,8 The situation becomes complicated because ADC values in hyperacute (earlier than 2𠄳 h) focal cerebral ischemia have very wide ranges and overlap those of the normal brain. Additionally, time-consuming spatial resolution is needed to assess rat brain infarction with 3 T magnet, resulting in the decrease of SNR and the exacerbation of the situation. Therefore, to assess hyperacute focal cerebral ischemia in rats, multiple ADC maps should be normalized to overcome the substantial variations, but signal intensity normalization has rarely been attempted using T1- or T2-weighted imaging maps. 9,10 This may be because ipsilateral focal lesions make normalization difficult, and complicated techniques should be required such as adjusting or matching histograms. 11

To evaluate focal cerebral ischemic lesions on MRI, complex automatic or semi-automatic segmentation techniques are available, such as a multi-parametrical regression model, three-dimensional (3D) histograms of multi-modality imaging, and techniques based on tissue classifications 12 � however, these techniques are not commonly used because of their limitations in acquiring perfusion maps or other time-consuming structural images in hyperacute studies. Threshold-derived pixel or voxel-wise segmentation based on ADC intensity is most commonly used. 12,16,17 This technique uses a cut-off value of the ADC threshold (also known as the viable threshold, 530 (휐 𢄦  mm 2 /s)), 12,18 � which has been verified by pathologic correlation studies using tetrazolium chloride (TTC) staining. The cut-off value of the ADC threshold of 530 reflects post-ischemic alterations after 12 to 24 h. 12,21,22 Other pathologic staining can show hyperacute ischemic damage in rat brains at 2𠄴 h after MCAO. 22,23 Especially for the hyperacute focal ischemic lesion on DWI (ADC map) very early after MCAO (within 1 h), a novel segmentation technique may be needed that has been validated by pathologic evidence. The present study observed that ischemic MCA territories in rats at 1 h after MCAO showed pallor areas on hematoxylin and eosin (Hɮ)-stained pathology, and we identified a valid, viable threshold that was correlated with Hɮ-stained pathology. Such a threshold has rarely been reported in previous work at this earlier hyperacute stage.

Here, we present a novel voxel-wise lesion segmentation technique for the assessment of hyperacute focal cerebral ischemia in rats at 1 h after MCAO on 3.0-T DWI, which was verified by gross and high-magnification light-microscopic examination of Hɮ pathology.

Perfusion MR Imaging


Perfusion MR imaging methods exploit signal intensity changes that occur with the passage of a tracer such as gadopentetate dimeglumine (28). As the gadolinium-based contrast agent passes in high concentration through the microvasculature, susceptibility-induced T2 and T2* relaxation occurs in surrounding tissues, which is seen as a decrease in signal intensity. The decrease in signal intensity is assumed to be linearly related to the concentration of gadolinium-based contrast agent in the microvasculature (29). The change in the signal intensity of the tissues is quantitatively related to the local concentration of gadolinium-based contrast agent and is converted into a curve of concentration versus time. By applying tracer kinetics to the concentration-time curve of the first passage of the bolus of gadolinium-based contrast agent, relative cerebral blood volume and cerebral blood flow maps are calculated (28).


Rapid imaging during the first pass of gadopentetate dimeglumine is performed by using a T2*-weighted echo-planar imaging sequence. This sequence can acquire 8–10 (or 16–20) sections covering the entire brain in 1 or 2 seconds. A series of 60 such multisection acquisitions is performed before, during, and after injection of gadopentetate dimeglumine. The injection of gadopentetate dimeglumine is started after the 10th run, followed by a 20-mL flush of normal saline solution at the same rate (28). Software is available that allows data analysis and calculation of various maps, such as relative cerebral blood volume, cerebral blood flow, time to peak, mean transit time, and permeability maps (Fig 11 ). The resolution of perfusion images is compromised to get the temporal resolution for making perfusion measurements.

Clinical Applications

Tumor growth is dependent on angiogenesis. Susceptibility-weighted MR perfusion imaging depicts microscopic capillary-level blood flow and hence serves as an indicator of angiogenesis in tumors (30,31). MR perfusion is commonly used in tumor grading (hypervascular tumors are often high grade), stereotactic biopsy guidance, monitoring response to therapy, and differentiation of radiation necrosis from recurrence (30,32–34). MR perfusion imaging is used in hyperacute stroke to determine potentially salvageable tissue in the penumbra so that appropriate action, such as intravenous or intraarterial thrombolysis, can be taken in a timely manner. Perfusion and diffusion together are the best predictors of salvageable tissue. A perfusion abnormality without a diffusion abnormality around the ischemic core indicates the penumbra of tissue that is salvageable (35). Generally, mean transit time is prolonged and relative cerebral blood flow is decreased in stroke. Relative cerebral blood volume is decreased in irreversibly injured areas and increased in spontaneous reperfusion. Other clinical applications of MR perfusion imaging include determinations of brain perfusion in moyamoya disease, vasculitis, and Alzheimer disease (36).


Low field MRI? The rationale behind it has been around for quite some time, traditionally perceived as a mean to reduce cost or to provide open access to patients suffering from claustrophobia. Over the past 30 years, scientists have supported low-field MRI on multiple occasions and brought facts that corroborate clinical relevance [1𠄴]. Yet, low field MRI has not spread. Reasons that have been invoked are diverse and have led to numerous debates. From the manufacturer point-of-view, the current business model in MRI results in higher margins allowing to increase profit [5]. From a clinical and academic point-of-view, the quest for higher and higher spatial resolution has led radiologists and scientists worldwide to always push toward high- and ultra-high field MRI research, eventually dominating over all others in peer reviewed journals [6]. One sure thing is that the statistics of MRI sales over the last two decades have certainly helped closing that debate. Nowadays high-field MRI sales (B0≥ 1.5 T) represent about 85% of the market size in Europe and North America [7]. One of the main misconceptions is that low-field MRI translates into poor image resolution, often associated with poor image quality. It is important, as scientists, to state that this concept is purely and simply wrong. Magnetic field strength has by no means ever been a limit to an achievable image resolution. A brief jump into the early days of MRI is enough to appreciate the tremendous leap in image quality that was made for a given field strength (Figure 1). More recent work from Choquet et al. [9] has reported MRI of different mouse body parts in vivo at 0.1 T (ߤ.3 MHz) with down to 100 × 100 × 750 μm 3 voxel size, more than 10 years ago (Figure 2). Sensitivity though, and how far the signal lies above the detection chain's noise floor will tell about one's capability to achieve a given resolution in the minimum amount of time. Hence time really is the argument at stake when considering lower field options. Indeed, lower field strengths result in lower bulk magnetization of nuclear spins leading in turn to a reduced sensitivity. Assuming a fixed noise floor in the detection chain, the decreased total magnetic moment brings the maximum signal detectable closer to the latter and the overall signal-to-noise ratio (SNR) drops. One main alternative to compensate for this loss is signal averaging. It is generally accepted that n averages will produce an SNR gain of n . Hence time is currently the true limit to a wide spread of low-field MRI, due to lower overall NMR sensitivity. However, this is also a matter of perspective. Why time considerations have become key in clinical diagnosis has to be contextualized in the current landscape of MRI. Most hospitals currently host one or two MRI scanners, which cost roughly scales with magnetic field (~$1M/T). As such MRI units are expensive and used for the imaging of all body parts, they are likely to represent a bottleneck in clinical workflows. Hence, the time needed for one scan has to be short in order to scan as many patients as possible within a day. Nowadays, no one can afford a machine that would perform slower than the state-of-the art because there is such a high demand for non-invasive radiation-free diagnosis. Yet, other than applications where speed is truly paramount such as for cardiovascular applications, or patients with a life-threatening risk, fast imaging is only required due to scanners being a low-volume/high price equipment. One could argue that this quest for speed is not as relevant if numerous low-field, low-cost devices were to be used (high-volume, low-price). After all, if the price of a scanner is divided by two and the acquisition time multiplied by two, then the cost per unit time stays the same, and the same number of patients can be scanned within the same time slot. Only cost for personnel would increase. The situation in China is a good illustration of this point: the high population density requires a higher density of MR units, and mid-field MR units represent about 50% of the market size [10], vs. 6% in Europe and North America [7]. As a consequence, depending on the market ability to embrace such a change in paradigm, that approach would naturally reduce pressure on acquisition times. Most importantly, rather than acquisition time or image resolution, one key aspect in the democratization of low-field MRI resides in its value. Most likely, if manufacturers and end users would foresee higher value in low-field MRI solutions, there would be a straightforward path toward mass adoption. Value though is a complex concept that finds different resonances across populations and cultures. One should agree on a simple description of value as being the ratio of a benefit over a cost [11]. For low-field technology to be truly visible and adopted, its value in MRI diagnosis would hence need to be increased. Two approaches then allow to increase value 1-reduce cost, 2-increase benefits, or both at the same time. Considering cost, the last two decades have already shown that relevant diagnosis can be achieved in lower-cost lower-field devices [12�]. Interest though, is hard to trigger if value increases only slightly and yet no broad adoption has followed since. For example, commercially available low-field scanners rely on permanent magnet technology that can weigh up to 13 tons. Thus, their individual cost (that includes siting) has never reached a point where value goes through the roof and triggers such a cultural change. Maybe the economic pressure on health expenses will change the current landscape with populations worldwide aging and growing, but this has been a long-heard argument never followed by action. Eventually, it appears challenging to spark interest in both radiologists and academics only by lowering cost as this is often perceived as leading to less potent technology, except maybe when the research is directed toward developing countries. The latter field of research is considered a niche though, and if cost is one key to selling in these countries, “low-cost” alone will never replace tailor-made solutions to country-specific needs. The alternative to increased value is then to increase benefits. Nowadays, MRI units require specific siting from their heavy weight and intense magnetic field strength, and shielding from magnetic and electromagnetic disturbances. MRI systems are known to be incompatible with most devices unless they are specifically made MRI-compatible. Increasing accessibility by means of a (much) lower footprint, little siting requirement, or enhanced compatibility certainly is a path toward increased benefits, and hence increased value. Now, what key element is driving such heavy weights, compatibility aspects, and ultimately the cost of MRI machines nowadays? Magnetic field is. As a result, low magnetic field MRI could very likely bring high value from both decreased costs and enhanced benefits. But how low can we go? In this manuscript, we aim to provide a fresh view on this old debate in MRI.

Figure 1. T2-weighted MR brain images acquired at 1.5 T in (A) 1986. Image reused, with permission, from Zimmerman et al. [8] and (B) 2009 (authors' database).

Figure 2. MRI images of the mouse acquired at 0.1 T using a FISP sequence and dedicated coils for the whole-body and tail. (Top) Whole body: field-of-view (FoV) of 110 mm and in plane resolution of 430 × 430 μm 2 . The acquisition time was 30 min for 30 slices of ߡ mm thickness. (Bottom) Tail: field-of-view (FoV) of 6.4 mm and in plane resolution of 100 × 100 μm 2 . The acquisition time was 1 h 30 min for 26 slices of ~ 750 μm thickness. Images modified, with permission, from Choquet et al. [9].

Comparison of electric field strength and spatial distribution of electroconvulsive therapy and magnetic seizure therapy in a realistic human head model

This study examines the strength and spatial distribution of the electric field induced in the brain by electroconvulsive therapy (ECT) and magnetic seizure therapy (MST).

The electric field induced by standard (bilateral, right unilateral, and bifrontal) and experimental (focal electrically administered seizure therapy and frontomedial) ECT electrode configurations as well as a circular MST coil configuration was simulated in an anatomically realistic finite element model of the human head. Maps of the electric field strength relative to an estimated neural activation threshold were used to evaluate the stimulation strength and focality in specific brain regions of interest for these ECT and MST paradigms and various stimulus current amplitudes.

The standard ECT configurations and current amplitude of 800–900 mA produced the strongest overall stimulation with median of 1.8–2.9 times neural activation threshold and more than 94% of the brain volume stimulated at suprathreshold level. All standard ECT electrode placements exposed the hippocampi to suprathreshold electric field, although there were differences across modalities with bilateral and right unilateral producing respectively the strongest and weakest hippocampal stimulation. MST stimulation is up to 9 times weaker compared to conventional ECT, resulting in direct activation of only 21% of the brain. Reducing the stimulus current amplitude can make ECT as focal as MST.

The relative differences in electric field strength may be a contributing factor for the cognitive sparing observed with right unilateral compared to bilateral ECT, and MST compared to right unilateral ECT. These simulations could help understand the mechanisms of seizure therapies and develop interventions with superior risk/benefit ratio.


The present study compared SVR, RVR and GPR with different morphometric input to perform brain age prediction. A total of eight models was assessed. The wide range of methods used in previous studies makes it challenging to disentangle the direct effect of model choice and other factors, such as the characteristics of the data set. In our study, we showed that the type of data input is generally more important than the choice of model, but various other aspects like data set size and processing time available should be considered when choosing a model. In Figure 2, we provide a decision tree that may help inform the model choice. This decision tree is based on the sequence of steps a researcher would typically take when designing a brain age study and is informed by the results of the present investigation. It is important to note that these results, and therefore our recommendations, are based on the UK Biobank. For example, our recommendations regarding the sample size and computational resources may be dependent on the characteristics of this specific data set. However, we believe that the general idea that some models require considerably more training data and computational resources than others can be generalised to other data sets.

Based on the literature, our first hypothesis was that all models would perform with MAE values below 5 years. With scores ranging from 3.7 to 4.7 years in the CV as well as the independent test set, this hypothesis was confirmed. These findings are generally in line with existing studies using a comparable setup, where MAE values in CV and independent data sets tend to fall between 3.9 and 6.2 years and 4.8 and 7.1 years, respectively (see Table S1 for an overview of related studies).

Our models showed moderate-high positive associations between age and predicted age (r ≈ .7 for all models) and they accounted for 40–50% of variation in new data (prediction R 2 ≈ .4–.5). Whilst these values are relatively high, the associations were lower than previous brain age studies that reported r values above .9 (Cole, Leech, & Sharp, 2015 Franke et al., 2010 Gutierrez Becker, Klein, & Wachinger, 2018 Kondo et al., 2015 ). The latter studies have in common that they covered a wider age range, including young people. In these age groups, the ongoing brain maturational changes make the task of brain age prediction easier. It therefore is possible that the limited and older age range in our sample along with the greater heterogeneity because of our unique data set size contributed to the lower—though still relatively high—r values of our models.

Our second hypothesis was that region-based models would outperform voxel-level ones due to the “curse of dimensionality” and high level of redundancy in the latter data, for example, high spatial correlations between voxels. This hypothesis was not confirmed, as there was no significant difference between the region- and voxel-based models (without PCA) in CV. Nonetheless, it appeared that dimensionality reduction through PCA could successfully remove redundancy to the extent that voxel-based models with PCA performed significantly better than the region-based models. This finding suggests that some of the age-related heterogeneity might be lost if the MRI data are summarised as regional volumes using FreeSurfer software. One previous study compared region- and voxel-level data input for GPR, but there was no clear difference in performance based on data type only (Gutierrez Becker et al., 2018 ). Comparing previous studies using either region- or voxel-level data as input also does not point at either type of data preprocessing being more suited for brain age prediction using SVR, RVR or GPR (Table S1).

Our third hypothesis was that RVR would perform best regardless of data input type, because it is seen as the “most popular” algorithm for brain age prediction (Cole et al., 2019 ). This hypothesis was not confirmed. Although voxel-based RVR without PCA showed the lowest MAE overall with

3.7 years, the difference to the other models was not statistically significant due to its high variance. The analysis of training set size also suggested that many iterations underfitted to the training set, which likely caused this variance. Therefore, we cannot conclude that RVR is the best model choice for brain age prediction regardless of data input. Previous studies on RVR or SVR that did not show a clear superior model (Table S1). Only two studies seem to have directly compared these two methods. For example, in Kondo et al. ( 2015 ), RVR performed slightly better than SVR in terms of MAE (4.50 and 4.73 years after dimensionality reduction, respectively). In Franke et al. ( 2010 ), RVR also performed slightly better than SVR after dimensionality reduction (4.98 vs. 5.10 after dimensionality reduction) but not without dimensionality reduction (5.23 vs. 5.14 without dimensionality reduction). This coincided with our findings, where PCA improved SVR but not RVR performance. However, neither of the previous studies assessed the significance of the difference, and we did not find a statistically significant difference between SVR and RVR if trained on the same data.

In terms of the GPR model, performance did not differ to SVR and RVR if trained on the same data. This confirms findings from a previous study where RVR and GPR were compared (Aycheh et al., 2018 ). However, there is little data available on this comparison and especially GPR on region-based data seems to be rare in the brain age literature. Our region-based GPR model had a smaller MAE than Gutierrez Becker et al. ( 2018 ), but higher than Aycheh et al. ( 2018 ). The MAE of the voxel-based GPR model with PCA is lower than previous comparable models by >1 year (Cole et al., 2015 , 2018 Monté-Rubio, Falcón, Pomarol-Clotet, & Ashburner, 2018 Table S1).

While MAE values of our models were generally low, their weighted MAE scores of 0.14 and above were notably higher than in other studies on SVR, RVR and GPR, where the scores tend to fall between 0.07 and 0.09 (Table S1). This is likely due to the smaller age range used here, as detailed in the limitations below. Although weighted MAE has not been formally validated as a measure of model performance, taking into account the age range of the training and test set is a useful exercise. A potential reason for the relatively high weighted MAE scores in our study might be greater heterogeneity in our sample due to the very large data set of >10,000 subjects, while the largest comparable study had around 3,000 subjects (Valizadeh et al., 2017 ). The acquisition of 10,000 subjects in one scanner will likely take place over a much longer time period than smaller data sets, so the acquired images will also be affected by changes in the scanner environment. These scanner effects might further contribute to the heterogeneity of our sample. In short, while the large data set is a clear strength of our study, it might compromise the comparability of our results to other studies in terms of weighted MAE.

Our models showed relatively high negative correlations between chronological age and BrainAGE in the CV iterations as well as the independent test set (approx. −0.7 for all with the exception of voxel-based SVR without PCA). This finding suggests that the models were equally and highly affected by regression to the mean (Le et al., 2018 ), although it is unclear why voxel-based SVR may be less affected by this. Whilst the high confounding effect of chronological may be seen as a limitation of our study, we believe it does not affect the direct comparison of models, which was our primary objective. Nonetheless, future studies should revisit these models and include the correction for age in the training. Various types of correction have been proposed in recent years (Beheshti, Nugent, Potvin, & Duchesne, 2019 Cole et al., 2020 de Lange & Cole, 2020 Le et al., 2018 ).

In a clinical context, it is crucial for a model to generalise to data from different scanners, because the parameters and environment of a scanner can introduce considerable bias. It is important to note that the independent data set in the present study was acquired on a different scanner with the same acquisition parameters, so future studies should address how our models would perform in other independent data sets acquired using different scanners and acquisition parameters. Our models generalised well to the independent test set. Indeed, the region-based models or the voxel-level models with PCA performed slightly better in the independent data set than in the CV set by approx. 0.3 and 0.1 years, respectively. Statistical significance between CV and generalisation performance was not assessed. These findings suggest the promise of these models for real-world application. In previous brain age prediction studies that compared model performance in an independent data set against the CV test, the models usually performed worse in the former (Cole et al., 2018 Franke et al., 2010 Lancaster, Lorenz, Leech, & Cole, 2018 Liem et al., 2017 see Table S1). However, similar to our findings, two studies also showed comparable performance in both (Cole et al., 2015 Le et al., 2018 ). Performance differences in an independent data set can likely be explained by sample characteristics, such as the similarity of this sample and the training data. In our case, the independent test set was acquired using the same acquisition protocol on a different scanner and the subjects came from the same population as the CV set. Noise and homogeneity should thus be similar between the samples. However, the independent test set appeared to be significantly older and it contained a slightly higher proportion of women (57%, see Table 1). So far, it is unclear whether sex has a considerable effect on brain age prediction, but this factor may have contributed to the performance differences between the sites in our study.

As expected, the analysis of training set size showed that larger sample sizes generally led to better prediction performances, though MAE scores did plateau with increasing training set size. For the region-based models, RVR required only half the training sample size than the other two to make predictions better than chance level, suggesting its suitability for studies where the sample size is limited. The analysis of training set size for GPR showed a sharp decrease in performance (i.e., higher MAE) at the smaller training set sizes, which might indicate overfitting to the training sample in the smaller samples. To our knowledge, no other studies have systematically evaluated the impact of training set size analysis to brain age prediction. In one case, Franke et al. ( 2010 ) assessed the effect of training set size by running separate RVR models on the full training data set (N = 410), half the data set (N = 205) and a quarter of the data set (N = 103). The MAE decreased from the smallest training set (5.6 years) to the largest set (4.9 years), which coincided with our results.

While our investigation was based on healthy brain ageing, it is important to ponder the potential implications of our findings for studies in clinical populations. One of the most promising uses of brain age prediction is its relevance and use as a biomarker. It could, for example, be implemented as an individualised marker of brain health in diagnostic tools. The main idea is to quantify the deviation between predicted and chronological age. Brains that are predicted to be older than their true age might suggest aberrant age-related changes and be associated with disease (Cole & Franke, 2017 ). Previous studies have assessed BrainAGE in various neurological and psychiatric disorders and they demonstrated that different stages of Alzheimer's disease as well as schizophrenia can present as accelerated ageing in the brain (Franke et al., 2010 Franke & Gaser, 2012 Gaser et al., 2013 Kaufmann et al., 2019 Koutsouleris et al., 2014 , 2015 Nenadić et al., 2017 Schnack et al., 2016 ). One of the necessary characteristics of a biomarker is its reliability. Therefore, future studies could adopt a longitudinal design to (a) further examine the reliability of the brain age prediction methods through test–retest setups in single or multi-scanner experiments, (b) learn more about the brain changes in health and disease, and (c) explore if brain age is a useful marker of treatment success in clinical trials.

The present study had three main limitations. First, whilst our data set size was quite large, the age range of 47–73 was smaller than most studies in the literature (e.g., Ashburner, 2007 Cole et al., 2018 , Cole et al., 2015 Franke et al., 2010 Gutierrez Becker et al., 2018 Le et al., 2018 Madan & Kensinger, 2018 Wang et al., 2014 Table S1). Furthermore, we excluded non-white ethnicities from the analysis because of data availability. These two factors imply that our models cannot be applied to data sets with ages or ethnicities that were not included in the training sample. Second, whilst the present study explored a wide range of methodological choices in terms of machine learning models and data input, there are several other methods that could be assessed in the future. For example, we did not explore nonlinear regression models, because we were interested in the interpretability of the models. Nevertheless, Ashburner ( 2007 ) directly compared the performance of RVR using either a linear and radial-basis kernel and found performance improvements with some configurations of the nonlinear one, so this seems to be an interesting area for future research. In addition, deep convolutional neural networks have shown to have a high accuracy when predicting brain age (Cole et al., 2017 Ito et al., 2018 Peng, Gong, Beckmann, Vedaldi, & Smith, 2021 ). Third, the present study was based on the use of a single neuroimaging modality. Our models could likely be improved by using multimodal input data. Previous research has shown that even combining different morphometric features, such as cortical thickness, surface area and/or curvature information, can improve model performance (Valizadeh et al., 2017 Wang et al., 2014 Zhao et al., 2018 ), because they may carry potentially complementary information about brain age. Similarly, Gutierrez Becker et al. ( 2018 ) achieved better performance of their GPR model when combining voxel-level and region-level features than looking at them separately, and Liem et al. ( 2017 ) were the first to combine structural and functional MRI in brain age prediction to achieve better performance. Multimodal data sets could also integrate conventional health assessments of ageing, which might improve the performance and generalisation of the models, making them a promising avenue for future brain age research (Cole et al., 2018 ).

MRI physics

The physics of MRI are complicated and much harder to understand than those underpinning image generation in plain radiography, CT or ultrasound.

What follows is a very abbreviated, 'broad strokes' description of the process. Essentially, the process can be broken down into four parts:

For a more detailed description of each part of the process, please refer to the links scattered throughout this introduction and at the bottom of the page.


The patient is placed in a static magnetic field produced by the magnet of the MRI scanner. In living tissues there are a lot of hydrogen atoms included in water molecules or in many different other molecules. The proton, the nucleus of hydrogen, possesses an intrinsic magnetization called spin. The spin magnetization vector precesses (rotates) around the magnetic field at a frequency called the Larmor frequency, which is proportional to the magnetic field intensity. The resulting magnetization of all protons inside the tissues aligns parallel to the magnetic field. The parallel magnetization scales with the magnetic field intensity, basically at 3 T it will be twice the value obtained at 1.5 T. Additional preparation sequences can also be performed to manipulate the magnetization and so the image contrast, e.g. inversion preparation.


During the image acquisition process, a radiofrequency (RF) pulse is emitted from the scanner. When tuned to the Larmor frequency, the RF pulse is at resonance: it creates a phase coherence in the precession of all the proton spins. The duration of the RF pulse is chosen such that it tilts the spin magnetization perpendicularly to the magnetic field. When a receiving coil (an electrical conductor) is put in the vicinity of the tissue, the transverse magnetization, that still rotates as the Larmor precession, will generate an electric current in the coil by Faraday induction: this is the nuclear magnetic resonance (NMR) signal.

The NMR signal is attenuated due to two simultaneous relaxation processes. The loss of coherence of the spin system attenuates the NMR signal with a time constant called the transverse relaxation time (T2). Concurrently, the magnetization vector slowly relaxes towards its equilibrium orientation that is parallel to the magnetic field: this occurs with a time constant called the spin-lattice relaxation time (T1). The contrast in MR images originates from the fact that different tissues have, in general, different T1 and T2 relaxation times as this is especially true for soft tissues, it explains the excellent soft tissue contrast of MRI.

Spatial encoding

Spatial encoding of the MRI signal is accomplished through the use of magnetic field gradients (smaller additional magnetic fields with an intensity that linearly depends on their spatial location): spins from protons in different locations precess at slightly different rates. The portion of the gradient coils and the associated current that is perpendicular to the main magnetic field cause a force (Lorentz force) on the coils. The gradients are turned on and off very quickly in this process causing them to vibrate and producing the majority of the acoustic noise during MR image acquisition.

Signal acquisition

When using magnetic field gradients, the obtained NMR signal contains different frequencies corresponding to the different tissue spin positions and is called the MRI signal. After sampling, the analog MRI signal is digitized and stored for processing, which consists of a separation of the signal contributions from different spatial locations represented by pixels in the final image. This is achieved by a mathematical operation called a Fourier transform.

Standard exam

Multiple image sets are obtained in a standard examination protocol (which varies from facility to facility). Exam times vary according to the part of the anatomy being studied, pathology expected, radiologist preferences, and the scanner hardware and software used. Occasionally, a contrast medium may be used to enhance images, this will also usually prolong the scan time. Typically, exams are ordered without and with contrast for comparison purposes. Very rarely, and only in certain circumstances are exams ordered with contrast only. After the MRI exam the patient is removed from the scanner and given post-procedure instructions (information about contrast medium and/or sedation if used).

Typical and minimum voxel sizes for various field strength MRI machines? - Psychology

OBJECTIVE. The objectives of our study were to assess the evidence for the diagnostic efficacy of 3-T MRI for meniscal and anterior cruciate ligament (ACL) injuries in the knee using arthroscopy as the reference standard and to compare these results with the results of a previous meta-analysis assessing 1.5-T MRI.

MATERIALS AND METHODS. The online Cochrane Library, MEDLINE, and PubMed databases were searched using the following terms: MRI AND ((3 OR three) AND (Tesla OR T)) AND knee AND arthroscopy AND (menisc* OR ligament). Patient demographics, patient characteristics, MRI scanning details, and diagnostic results were investigated. The methodologic quality of the included studies was assessed using the revised Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tool. A meta-analysis of studies using 3-T MRI was performed, and the results were compared with a previous meta-analysis of studies using 1.5-T MRI.

RESULTS. One hundred one studies were identified by the search strategy, and 13 studies were included in our review. Twelve studies were considered to have level 1b evidence, and one study was considered to have level 2b evidence. All 13 studies had high methodologic integrity and low risk of bias using the QUADAS-2 tool. The studies included 1197 patients with a mean age of 41.9 years. Ten of the 13 studies were eligible for meta-analysis. The mean sensitivity and mean specificity of 3-T MRI for knee injuries by location were as follows: medial meniscus, 0.94 (95% CI, 0.91–0.96) and 0.79 (95% CI, 0.75–0.83), respectively lateral meniscus, 0.81 (95% CI, 0.75–0.85) and 0.87 (95% CI, 0.84–0.89) and ACL, 0.92 (95% CI, 0.83–0.96) and 0.99 (95% CI, 0.96–1.00). The specificity of 3-T MRI for injuries of the lateral meniscus was significantly lower than that of 1.5-T MRI (p = 0.0013).

CONCLUSION. This study does not provide evidence that 3-T scanners have superior diagnostic efficacy for meniscal damage and ACL integrity when compared with previous studies of 1.5-T machines.

MRI has high sensitivity and high specificity for diagnosing ligamentous, meniscal, and articular injuries and degeneration of the knee [1–7]. MRI scanners with greater field strengths have been developed [8–10] and may have potential for improving the diagnostic efficacy of this imaging modality.

The Tesla is a quantitative measurement of the field strength of a magnet. In 2002, 3-T MRI scanners became commercially available in the United States. More powerful scanners have been developed, albeit not for human clinical imaging, with field strengths of up to 11 T [11, 12].

The more powerful 3-T scanners are replacing the older 1.5-T scanners, which have been the clinical standard in hospitals for many years [13]. The 3-T scanners generate twice the magnetic field strength of the 1.5-T scanners, thus increasing the signal intensity detected from any given tissue, with only marginal increases in background noise. This increased field strength provides a greater signal-to-noise ratio and produces clearer images with enhanced resolution or allows faster scanning at a conventional resolution [14, 15].

There are numerous technical considerations regarding high-field-strength MRI scanners such as relaxation kinetics, magnetic susceptibility effects, and safety considerations. New acquisition protocols are being developed to examine the effect of variations in field strength, slice orientation and thickness, voxel size, pixel bandwidth, and scanning algorithms. There is no consensus about the optimum protocol for musculoskeletal imaging with 3-T scanners [16], whereas there is some agreement about the optimum protocol for musculoskeletal imaging with 1.5-T scanners [17–19].

Individual clinical tests for meniscal and anterior cruciate ligament (ACL) injuries have wide and varied reported sensitivities, ranging from 38% to 85% [20–22]. The sensitivity of a thorough clinical assessment for medial meniscal and ACL injuries is approximately 85–95% [22–26]. The role of MRI in patients with these injuries is to confirm the clinical diagnosis and exclude occult mechanical, degenerative, or neoplastic pathologic entities [27]. The reference standard for diagnosing intraarticular soft-tissue pathologic entities of the knee is direct visualization at arthroscopy [7, 28, 29], which carries a risk of morbidity and mortality. Noninvasive investigations minimize the need for invasive procedures and allow planning of therapeutic interventions, the benefits of which justify the exposure to the potential risks of surgery.

The aim of this study was to assess the evidence for the diagnostic efficacy of 3-T MRI for meniscal and ACL injuries in the knee and examine the evidence for its superiority over 1.5-T MRI.

A search of the online Cochrane Library, MEDLINE, and PubMed databases was conducted using the following terms: MRI AND ((3 OR three) AND (Tesla OR T)) AND knee AND arthroscopy AND (menisc* OR ligament). No limitations were placed on patient sex, patient age, publication date of the article, or language of the article. The bibliographies of all articles were screened to identify any other published studies that the primary search may have failed to identify.

All articles were assessed to identify studies for inclusion. Studies were included if 3-T MRI had been used to diagnose medial meniscal, lateral meniscal, or ACL injuries and if the MRI findings were correlated with arthroscopic findings both prospective and retrospective studies were eligible for inclusion.

Studies were excluded if MRI field strengths other than 3 T were used, if a new scanning protocol for 3-T MRI was used without inclusion of the results of a previously established control protocol, if nonhuman subjects were used, and if the full text of the article or a translation of the full text was not available in the English language. Case reports, review articles, and comments about existing studies were excluded.

All articles that were included underwent a detailed review and assessment for the following data: number of subjects studied, male-to-female ratio, number of knees imaged (i.e., if patients with bilateral injuries were included), laterality, average age of sample group, time between MRI and arthroscopy, type of MRI scanner, performance values for 3-T MRI (i.e., accuracy, sensitivity, specificity, positive predictive value, and negative predictive value), interobserver correlation, and reliability. If the diagnostic applicability of a new 3-T protocol was being evaluated, the reference protocol data were used and any index test data were excluded that is, the data from the current standardized protocol were used, and the data from the experimental protocol data were excluded.

The methodologic quality of the included studies was assessed using the revised Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tool [30]. Eleven criteria in four separate domains (i.e., patient selection, index test, reference standard, and flow and timing) are used to critically appraise the methodology of a study. Each domain is assessed for risk of bias, and the first three domains (i.e., patient selection, index test, and reference standard) are also assessed in terms of the applicability of the study for inclusion in our review [30]. The assessment was performed independently by two authors, and disagreements were resolved by the senior author. The level of evidence of the included studies was also determined.

Studies that reported true-positive, false-positive, true-negative, and false-negative values were included. Studies in which these values could be derived by working backward from the sensitivity and specificity values were also included in the meta-analysis. Because of the lack of randomization in the studies yielded by our search, nonrandomized trials were included. This methodologic technique has previously been described by Shrier et al. [31] they stated that meta-analyses including observational studies can yield informative and meaningful results [31].

Statistical analysis was performed using a statistical package (SPSS, version 5.2.7, IBM), and the level of significance was set at p < 0.05. The meta-analysis was conducted using Review Manager software (version 5, The Cochrane Collaboration). This software package summarizes data and creates appropriate forest plots for graphical presentation. A random-effects model was used to reduce bias from the potential systemic errors of the included studies, and an inverse variance method was used for the weighting of each study. Homogeneity across the studies was assessed and represented by χ 2 and I 2 statistical significance for χ 2 was set at p < 0.10. The independent t test was used to compare our findings with the findings of a previous meta-analysis of 1.5-T MRI studies.

Because of the nonclinical nature of this study, approval by the institutional review board was not deemed necessary.

Figure 1 shows the results of the search strategy and the application of the preferred reporting items for systematic reviews and meta-analyses (PRISMA) flow diagram. The database search identified 101 studies. Reviewing bibliographies of those identified by the search yielded an additional three relevant articles. After the inclusion criteria were applied, 21 articles were eligible for inclusion. Eight studies were removed due to the exclusion criteria (no full text or translation in the English language, n = 3 articles use of non–3-T MRI, n = 3 inability to use results because of inappropriate methods, n = 1 and review article of different 3-T MRI protocols, n = 1), resulting in 13 studies included for the final review [32–44].

Fig. 1 —Preferred reporting items for systematic reviews and meta-analyses (PRISMA) flow diagram show search strategy used to obtain studies for our review and meta-analysis. The following terms were used to search online Cochrane Library, MEDLINE, and PubMed databases: MRI AND ((3 OR three) AND (Tesla OR T)) AND knee AND arthroscopy AND (menisc* OR ligament).

Table 1 shows the results of the QUADAS-2 tool for the 13 studies: Seven studies scored 11 points or 100% [33, 36–40, 42], four studies scored 10 points or 91% [32, 35, 43, 44], and two studies scored 9 points or 82% [34, 41].

Twelve of the 13 studies (92%) had a low risk of bias from all areas of their methodology ( Fig. 2 ). The Schub et al. [44] study was the only exception. In that study, bias with regard to the index test was unclear ( Fig. 3 ) because patients who had previously undergone arthroscopic knee surgery were not excluded from their study, which may have affected the specificity of the test.

Fig. 2 —Graphic shows overall risk of bias among all 13 studies according to results of revised Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) [30].

Fig. 3 —Graphic shows individual assessment of bias of all 13 studies using revised Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) [30] methodologic assessment tool.

We classified 12 of the 13 studies (92%) [32–43] as having level 1b evidence because they are validating cohort studies using a reference standard as a reference test. One study [44] was deemed as having level 2b evidence because it was more of an exploratory study. A summary of the studies, including the publication details, levels of evidence, and QUADAS-2 scores, is shown in Table 1.

The 13 studies included a total of 1197 patients. The largest sample group was composed of 250 patients [37] and the smallest, 24 patients [38]. Seven hundred twenty-five patients were male, and 472 were female. In 13 patients, bilateral investigations were performed, so 1210 knees were reviewed. The mean age of the patients was 41.9 years. The MRI examination in 11 of the 13 studies were reviewed by a minimum of two experienced musculoskeletal radiologists, with a mean number of radiologist reviewers among the 13 studies of 2.6. The mean time between MRI and arthroscopic surgery was 47.4 days.

All 13 studies focused on the diagnosis of medial meniscal injuries, 12 of the 13 studies assessed lateral meniscus injuries, and five of the 13 studies assessed ACL injuries.

Ten studies were suitable to allow meta-analysis of their data [32–34, 37–43]. The three excluded studies did not report their data in suitable detail that would permit pooling, as stated in the Materials and Methods section [35, 36, 44]. Heterogeneity was calculated using χ 2 and I 2 statistics to permit pooling of study results for meta-analysis. All 10 studies provided stratified data for medial and lateral meniscal injuries. Only three studies provided data relating to ACL injuries that were suitable for meta-analysis [32, 40, 42]. The results of these studies and the extracted data are summarized in Table 2.

Figure 4 shows the forest plots for the studies included in the meta-analysis the calculated sensitivity and specificity for each study, including 95% CIs, are provided. For medial meniscal injuries, the mean sensitivity of 3-T MRI was 0.94 (95% CI, 0.91–0.96) and the mean specificity was 0.79 (95% CI, 0.75–0.83). For the detection of lateral meniscal injuries, 3-T MRI had a mean sensitivity of 0.81 (95% CI, 0.75–0.85) and a mean specificity of 0.87 (95% CI, 0.84–0.89). Finally, for ACL injuries, 3-T MRI had a mean sensitivity of 0.92 (95% CI, 0.83–0.96) and a mean specificity of 0.99 (95% CI, 0.96–1.00). The meta-analysis results for the pooled data of the included studies are summarized in Table 3.

Fig. 4 —Forest plots show sensitivity and specificity values for detection of medial meniscal, lateral meniscal, and anterior cruciate ligament (ACL) injuries using 3-T MRI. TP = true-positive, FP = false-positive, FN = false-negative, TN = true-negative.

The summary ROC plots show that 3-T MRI had the greatest diagnostic accuracy for ACL injuries, followed by medial meniscal injuries, and finally lateral meniscal injuries ( Fig. 5 ). Diagnostic odds ratio (OR) values were greatest for ACL injuries, then medial meniscal injuries, and finally lateral meniscal injuries. These results show that 3-T MRI has better discriminatory test performance for ACL injuries than for meniscal injuries. Likewise, 3-T MRI had better discriminatory test performance for medial meniscal injuries than for lateral meniscal injuries.

Fig. 5 —Summary ROC plots for diagnostic accuracy of 3-T MRI for medial meniscal, lateral meniscal, and anterior cruciate ligament (ACL) injuries. Dashes are line of no demarcation.

Two previous meta-analyses of the diagnostic accuracy of 1.5-T MRI used arthroscopy as a reference standard. These meta-analyses were performed by Crawford et al. [7] and Oei et al. [45] and examined 43 and 29 studies, respectively (Table 4).

Crawford et al. [7] did not provide enough data to allow meaningful statistical comparison with our data therefore, the results of our study are compared with only the results of Oei et al. [45]. A weighted independent t test was used to examine whether there was a significant difference between the mean sensitivity and mean specificity of 1.5-T MRI and 3-T MRI for medial and lateral meniscal injuries. There was no significant difference between 1.5-T studies and 3-T studies for the detection of medial meniscal injuries (sensitivity, p = 0.444 specificity, p = 0.460). In addition, there was no significant difference between the sensitivity of 1.5-T MRI and that of 3-T MRI for lateral meniscal injuries (p = 0.527). The specificity of 3-T MRI for lateral meniscal injuries was significantly different from that of 1.5-T MRI (p = 0.0013) however, this difference was in favor of 1.5-T MRI, which had a specificity of 0.957 (95% CI, 0.946–0.968) [45] compared with 0.87 for 3 T (95% CI, 0.84–0.89).

The purpose of this study was to perform a systematic review of the published peer-reviewed literature assessing the diagnostic efficacy of 3-T MRI in detecting intraarticular knee injuries compared with arthroscopic findings as the reference standard.

All of the studies included in our review were of high quality. Twelve of the 13 studies are diagnostic studies with level 1b evidence because they are studies validating a procedure against the current reference standard (i.e., arthroscopic findings) and had good methodologic quality. One study had level 2b evidence because it failed to have the same methodologic principles as the others, but it used the same reference test. For the studies that were investigating the effects of changes or attempted optimization of 3-T MRI protocols, we used only the data of the reference protocol (i.e., the current standard against which the altered protocols were compared) and ignored the results from the index protocol. We used this strategy to ensure a higher degree of standardization of the included studies. All studies also scored highly on the QUADAS-2 methodologic assessment tool, with only one study exhibiting possible bias of the index test [44].

The results of our meta-analysis show the excellent diagnostic ability of 3-T MRI for detecting medial meniscal, lateral meniscal, and ACL injuries. The sensitivities for diagnosing medial meniscal, lateral meniscal, and ACL injuries were 0.94, 0.81, and 0.92, respectively the specificities were 0.79, 0.87, and 0.99, respectively.

When comparing our results with those of a previous meta-analysis of 1.5-T scanners [45], we omitted ACL injuries because there were too few results to achieve a reliable statistical comparison. We found that the only significant difference was in favor of the 1.5-T machines. As stated in the Results section, the specificity of 3-T MRI and the specificity of 1.5-T MRI were significantly different for lateral meniscal injuries (p = 0.0013). One suggested hypothesis is that lateral meniscal tears are overdiagnosed on 3-T MRI because of the heterogeneous appearance of the anterior horn toward the anterior root insertion [46]. Overdiagnosis would obviously result in an increased number of false-positive results.

Likelihood ratios (LRs) were calculated for the 1.5-T scanners using the data of Oei et al. [45]. The negative LR was comparable for the 1.5-T data and the 3-T data: 0.08 and 0.09 for the medial meniscus, respectively, and 0.22 and 0.26 for the lateral meniscus. The positive LR values were 8.04 and 4.37 for the medial meniscus and 18.44 and 5.82 for the lateral meniscus for 1.5-T MRI and 3-T MRI, respectively. The 1.5-T machines had a higher positive LR for both the medial and lateral menisci compared with the 3-T scanners. Therefore, although the probability of both 1.5-T and 3-T scanners correctly excluding meniscal tears is equivocal, there is a higher probability of a tear being found at arthroscopy if it was diagnosed on a 1.5-T scanner than if it was diagnosed on a 3-T scanner.

Van Dyck et al. [42] and Grossman et al. [33] compared routine knee MRI at 1.5 T and 3 T and used arthroscopy as the reference standard. Grossman and colleagues looked at two separate cohorts: one undergoing MRI on a 1.5-T scanner and one undergoing MRI on a 3-T scanner. Van Dyck et al. [42] used a single cohort, scanning each participant twice, once at 1.5 T and then at 3 T. Both groups of investigators found that the diagnostic efficacy of 3 T was not significantly different compared with 1.5 T. Magee and Williams [47] reported overall sensitivity of 0.96 and specificity of 0.97 for 3-T MRI in diagnosing meniscal tears. They concluded that “3 T compares favorably in sensitivity and specificity with studies performed at 1.5-T field strength or lower” [47]. They did not perform a direct comparison of the diagnostic efficacy of 1.5-T MRI and 3-T MRI instead, they compared their results at 3 T with the results of previous studies performed at 1.5 T. They stated that the results of the comparative studies performed at 1.5 T had sensitivities of 80–100%, which contradicts their conclusions. They also combined the results of the medial and lateral meniscus together when calculating sensitivity and specificity values, making it difficult to comment on their findings.

The first commercial 1.5-T MRI machines became available in 1983 [9]. The 1.5-T software protocols for image acquisition and processing have undergone extensive refinement, compared with the more recent introduction of the 3-T scanners. Four of the studies included in our study group examined alternative protocols and coil arrangements for 3-T MRI and the effect on diagnostic accuracy [35–38]. Further refinement of this technology may enhance the diagnostic efficacy of 3-T MRI. With the improvement of software packages allowing more complex algorithms to process and reformat data, it is likely that the diagnostic capacity of 3-T MRI will improve. Whether this improvement will translate into a reciprocal improvement in 1.5-T scanners is uncertain.

Comparing our results with those of previous studies does not provide any evidence that the diagnostic efficacy of 3-T MRI is superior to that of 1.5-T MRI for diagnosing meniscal or ACL injuries. The 1.5-T scanners cost $1–1.5 million, whereas the 3-T scanners cost $2–2.5 million (exclusive of service, maintenance, warranty extension, and so on) [48]. The 1.5-T scanners already have excellent diagnostic ability at detecting meniscal and ACL injuries [7, 45], so the advancements in scanner technology may provide disproportionately marginal diagnostic gains in comparison with the higher costs. There is evidence for improved diagnosis of chondral abnormalities of the knee [49] and of abnormalities in other body systems, but these topics are beyond the scope of this study.

The improved field strengths and consequent increase in signal intensity may be used to produce higher-resolution images. The 3-T MRI machines can also produce images of the same quality as the 1.5-T machines but in markedly reduced times. The 3-T machines are able to perform examinations faster than their 1.5-T counterparts, allowing a greater number of examinations to be performed in a set time period. These factors combined with the value of health care reimbursement, running costs, maintenance, higher throughput of patients, and the possibility of improved diagnostic efficacy in imaging other body parts, such as neuroimaging, must be considered in any cost analysis.

The six steps of a radiologic diagnostic sequence are data acquisition using a diagnostic system processing of the data by software to create an image display of the image on a high-resolution monitor light from the monitor striking the radiologist's retina, triggering impulses along the optic nerves Processing of these nerve impulses in the radiologist's visual cortex and the ability of the radiologist to form a clinical opinion from the stimulated visual cortex. There is the question of whether the chase for stronger field strengths will result in improved diagnostic accuracy or whether the optimal performance has already been reached because of limitations in the last three components—that is, the radiologist's retina, visual cortex, and clinical opinion.

The limitations of this study are dependent on the limitations of the studies included. Although 10 studies were able to have their data pooled for medial and lateral meniscal injuries, only three were suitable for ACL injuries. The results for ACL injuries are therefore more open to bias than the results for the meniscal injuries.

All studies except one had level 1b evidence. That study had level 2b evidence [44] the evidence was deemed to be of slightly lower quality because these investigators recruited pediatric patients and did not exclude patients who had previous knee surgery. Children have increased meniscal vascularity that can mimic the MRI signal intensity of intrasubstance degeneration and tears [50]. Both of these factors can increase the bias of the results because there may be more false-positive findings. We still included this study to create a comprehensive review however, the data were not suitable for qualitative analysis.

A meta-analysis looks at data over a certain period of time for, in this case, an area of medical technology that is continuously evolving. The most recent article included in our meta-analysis of 3-T machines was published in 2013 as a result, this context must be taken into account when considering the conclusions drawn from the results. Improvements in technology and their benefits will often occur long before the results are independently investigated and published. It would be interesting to see how the results of this study would compare with a meta-analysis of peer-reviewed studies published from 2013 through 2018.

Most of studies included in our meta-analysis adopted a consensus approach to the musculoskeletal radiologists' interpretation of a study. This approach represents a clinically abnormal approach to analyzing MRI studies, but it helps remove any bias of interpretive abilities for the purpose of the included studies. Intraobserver agreement was calculated by only Jung et al. [35] and Kijowski et al. [37] they found that there was good-to-excellent agreement for all pathologic findings. There was no statistical difference in agreement reported by either study.

The results of this study show that 3-T MRI scanners have excellent diagnostic efficacy for ACL and meniscal injuries. However, the diagnostic studies published through 2013 do not provide any evidence that 3-T scanners are superior when compared with a previous meta-analysis of studies performed using 1.5-T machines [45]. In fact, our analysis shows that the specificity of 3-T MRI is lower than that of 1.5-T MRI with regard to the diagnosis of lateral meniscal tears. Advances in technology and software developments may improve the diagnostic efficacy of 3-T MRI scanners in the future to a point at which it is greater than that of 1.5-T scanners.

Access options

Get full journal access for 1 year

All prices are NET prices.
VAT will be added later in the checkout.
Tax calculation will be finalised during checkout.

Get time limited or full article access on ReadCube.

All prices are NET prices.

Low-Cost High-Performance MRI

Magnetic Resonance Imaging (MRI) is unparalleled in its ability to visualize anatomical structure and function non-invasively with high spatial and temporal resolution. Yet to overcome the low sensitivity inherent in inductive detection of weakly polarized nuclear spins, the vast majority of clinical MRI scanners employ superconducting magnets producing very high magnetic fields. Commonly found at 1.5–3 tesla (T), these powerful magnets are massive and have very strict infrastructure demands that preclude operation in many environments. MRI scanners are costly to purchase, site and maintain, with the purchase price approaching $1 M per tesla (T) of magnetic field. We present here a remarkably simple, non-cryogenic approach to high-performance human MRI at ultra-low magnetic field, whereby modern under-sampling strategies are combined with fully-refocused dynamic spin control using steady-state free precession techniques. At 6.5 mT (more than 450 times lower than clinical MRI scanners) we demonstrate (2.5 × 3.5 × 8.5) mm 3 imaging resolution in the living human brain using a simple, open-geometry electromagnet, with 3D image acquisition over the entire brain in 6 minutes. We contend that these practical ultra-low magnetic field implementations of MRI (<10 mT) will complement traditional MRI, providing clinically relevant images and setting new standards for affordable (<$50,000) and robust portable devices.


  1. Tramaine

    And is not infinitely distant :)

  2. Ptah

    Message deleted

  3. Kelly

    It seems to me it is very good idea. Completely with you I will agree.

  4. Renneil

    I apologize, but in my opinion you are wrong. Enter we'll discuss it. Write to me in PM, we'll talk.

  5. Nosh

    This is doubtful.

  6. Carrick

    I accept it with pleasure. In my opinion, this is relevant, I will take part in the discussion.

Write a message