Fat fraction mapping using magnetic resonance imaging: insight into pathophysiology

Abstract

Adipose cells have traditionally been viewed as a simple, passive energy storage depot for triglycerides. However, in recent years it has become clear that adipose cells are highly physiologically active and have a multitude of endocrine, metabolic, haematological and immune functions. Changes in the number or size of adipose cells may be directly implicated in disease (e.g. in the metabolic syndrome), but may also be linked to other pathological processes such as inflammation, malignant infiltration or infarction. MRI is ideally suited to the quantification of fat, since most of the acquired signal comes from water and fat protons. Fat fraction (FF, the proportion of the acquired signal derived from fat protons) has, therefore, emerged as an objective, image-based biomarker of disease. Methods for FF quantification are becoming increasingly available in both research and clinical settings, but these methods vary depending on the scanner, manufacturer, imaging sequence and reconstruction software being used. Careful selection of the imaging method—and correct interpretation—can improve the accuracy of FF measurements, minimize potential confounding factors and maximize clinical utility. Here, we review methods for fat quantification and their strengths and weaknesses, before considering how they can be tailored to specific applications, particularly in the gastrointestinal and musculoskeletal systems. FF quantification is becoming established as a clinical and research tool, and understanding the underlying principles will be helpful to both imaging scientists and clinicians.

Introduction

Adipose cells have traditionally been viewed as simple, passive energy storage depots for triglycerides, which release energy in the form of fatty acids through lipolysis during times of metabolic stress. However, in recent years it has become clear that adipose cells are highly dynamic, and have important endocrine, metabolic, haematological, immune and structural functions. For example, adipose cells secrete a wide variety of hormones, growth factors, cytokines, matrix proteins and enzymes which interact with diverse organ systems including the hypothalamo-pituitary axis, pancreas, liver, kidneys, endothelium, skeletal muscle and immune system.¹ Leptin is the most well-known of these factors and has an important role in the regulation of satiety,² but a number of other “adipocytokines” also contribute to the regulation of food intake, metabolism, immunity and blood pressure homeostasis.³ These adipocytokines include adiponectin, which is thought to increase insulin sensitivity,⁴ and tumour necrosis factor-α, which is an inflammatory cytokine and may contribute to insulin resistance (IR)and the pro-inflammatory state which is observed in obese patients.⁵ Adipose tissue expresses all the components of the renin-angiotensin-aldosterone system, which contributes to the regulation of blood pressure and platelet function,¹ and secretes plasminogen activator inhibitor-1, which is involved in fibrinolysis and is altered in obesity.¹

Importantly, adipose cells in different locations may have different precursors, cytokine profiles and functions.^1,6 The primary site for adipose cells is in adipose tissue, which consists of subcutaneous adipose tissue, found immediately under the skin, and visceral adipose tissue which surrounds organs such as the liver, bowel, pancreas, kidneys and heart. Additionally, fat can accumulate within organs, including the liver, pancreas, heart, muscle and bone marrow; these fat deposits are described as “ectopic fat depots”.⁷ There is a growing appreciation of the importance of ectopic fat in the pathogenesis of disease;^8–10 e.g. hepatic fat is implicated in the pathogenesis of metabolic syndrome, and predicts IR independent of visceral adipose tissue volume.¹⁰ Similarly, bone marrow adipose tissue has a unique role in the regulation of haematopoiesis, bone turnover and systemic metabolism,¹¹ and derangements in marrow fat are seen in osteoporosis, obesity, diabetes, metastatic malignancy and haematological disorders. Increased awareness of the importance of these fat depots has led to a demand for noninvasive techniques for measuring organ fat content, in both clinical and research settings.

Proton MRI is ideally suited to fat quantification since the majority of the acquired signal originates from protons in water and fat molecules. Fat and water protons produce signals with slightly different frequencies because their chemical environments give rise to subtle differences in the local magnetic fields they experience; the MR signal, therefore, consists of multiple spectral components which have frequencies separated by a few parts per million (ppm). In a chemical shift-encoded MRI (CSE-MRI) experiment, the interference of these components leads to oscillation in the signal over time, and repeated sampling of the signal at varying echo times (TEs) enables quantification of the fat and water signals.¹² The fat fraction(FF) is defined as the signal arising from fat protons divided by the sum of the signals from fat and water protons. The FF can be used as a quantitative image-based indicator of biological and pathological processes—an imaging biomarker.¹³

Recent years have seen a rapid expansion in the number MR methods available for measuring the FF, each with its own strengths and weaknesses. These methods are being used in clinical trials and in clinical practice,^14–16 but approaches to fat quantification are inconsistent. In clinical practice, simple “in phase and out of phase” imaging is commonly used, but can be misinterpreted and may be inaccurate in the presence of iron or bone.^9,17 In this work, we aimed to highlight the strengths and weaknesses of different MR techniques for fat quantification, and to explain potential confounding factors which may introduce bias into FF measurements. Furthermore, we describe a wide variety of existing and potential applications of FF mapping, including established applications in the liver and more recently introduced applications in bone, muscle and heart. We will argue that, quite apart from its role in passive energy storage, the presence and quantity of fat in a tissue can directly inform us about underlying pathophysiological processes in a wide variety of diseases.

MR Methods for Fat Quantification

Dixon’s method

The original imaging method for fat quantification (“simple proton spectroscopic imaging”) was described by WT Dixon.¹² In Dixon’s method, magnitude images were acquired at both “in phase (IP)” and “opposed phase (OP)” TEs, at which times water and fat would constructively and destructively interfere (Figure 1). Water-only and fat-only images (Figure 2) could then be obtained at each voxel by adding and subtracting the two images, respectively.

Figure 1. .

Principles of Dixon imaging. If the image is acquired when the water and fat have the same phase (a), the signals from water (W) and fat (F) constructively interfere, and the total signal S = W + F. If the image is acquired when water and fat are in opposed phase (b), W and F destructively interfere and S = W− F. (c) Shows data from a single voxel in normal bone marrow (which contains both water and fat), using a gradient echo-based Dixon acquisition with very closely spaced echo times. This data explicitly shows the signal oscillation over time as the fat and water signals dephase, come back into phase,and then dephase again. There is also a progressive reduction in the height of the IP peaks with increasing  echo timeowing to signal decay (in this case with the time constant T₂*). 

Figure 2. .

Examples of CSE-MR images. In a two-point CSE-MRI (Dixon) experiment, images are acquired at IP and OP TEs. Addition and subtraction of these images produces water only (WO) and fat only (FO) images, respectively. Fat fraction (FF) maps can then be generated from the WO and FO images as described above.

However, Dixon’s original method suffered from a major drawback. To overcome problems with background magnetic field (B₀) inhomogeneity, Dixon used only magnitude images for the reconstruction (signal phase data were discarded). Any tissue consisting of either pure fat or pure water would, therefore, have approximately equal signal on the IP and OP images. As a result, fatty tissues (such as subcutaneous fat) have high signal on the water images but low signal on the FO images, meaning that the FF measured would be (incorrectly) close to 0%. This phenomenon can potentially cause misinterpretation in clinical practice when in-phase and opposed-phase images are visually compared, since lesions consisting of pure fat exhibit little signal drop-out on opposed images. Furthermore, this is a severe limitation if CSE-MRI is used quantitatively, since measured FF values will be inaccurate or heavily biased by fat-water ambiguity.

Extended two-point methods

To overcome the fat-water ambiguity problem, several authors extended Dixon’s method to enable resolution of the ambiguity between water-dominant and fat-dominant tissues.^18,19 These extended methods involve the acquisition of complex-valued images (i.e. images with both magnitude and phase data) at in phase and out of phase echo times; the phase data is used to distinguish between fat-dominant and water-dominant tissues.

Although these extended two-point methods are a substantial advance on Dixon’s original method, they are somewhat susceptible to the effects of B₀ inhomogeneity, since it can be difficult to differentiate between phase shifts owing to B₀ effects and phase shifts due to chemical shift. One approach to reducing the effect of B₀ inhomogeneity is to combine these two-point methods with phase-unwrapping or region-growing algorithms (fat-water separation is usually straightforward if the phase can be successfully unwrapped).^18,19 Unfortunately, these algorithms are somewhat complex and tend to fail in areas of signal cancellation, particularly where the fat and water signals have similar magnitude.

Three-point methods

To eliminate the need for complex region-growing methods and to improve estimation of B₀ inhomogeneity, Glover and Schneider introduced a three-point CSE-MRI method enabling direct estimation of field inhomogeneity to remove its effects from the signal.²⁰ In this method, one IP echo and two OP echoes are used; any phase shift between the two opposed phase echoes can be attributed to B₀ inhomogeneity, which can be used to correct the acquired signals. Glover and Schneider’s method was subsequently extended to allow greater flexibility with regard to echo times by Xiang and An²¹, and a three-point method using a similar approach has recently been introduced for whole body imaging.²² These three-point methods form the basis of a number of manufacturers' methods for rapid fat-suppressed imaging, primarily due to their speed and relative robustness to artifacts.^23,24 However, where FFs need to be calculatedly accurately, a different approach is necessary.

Maximum -likelihood estimation

The extended two-point and three-point methods described above were only intended to separate two species (water and fat). In practice, the two-species assumption leads to inaccuracies in fat quantification, since human fat actually consists of a number of spectral fat components (the “main” fat peak comes from hydrogen atoms in methylene residues, but methyl, allylic and olefinic residues also contribute substantially to the measured signal).²⁵ Furthermore, these methods do not account for T₂* decay, which can introduce a further source of bias.^25,26

To account for the effects of spectral fat complexity and T₂* decay and thereby reduce bias, a maximum-likelihood method can be used to find the parameters in a pre-specified signal model which will most closely fit the acquired data.²⁷ This approach was first proposed by Xiang and An, who used a non-linear least squares (NLLS) method with a signal model assuming multiple spectral fat components. Subsequently, a more general maximum-likelihood method—echo asymmetry and least squares estimation [IDEAL (iterative decomposition of water and fat with echo asymmetry and least squares estimation)]—was proposed and developed by Reeder et al^28,29 The IDEAL method enables an arbitrary choice of echo times (and, importantly, can be easily applied to acquisitions with more than three echoes) and arbitrary numbers of spectral components. IDEAL can also be adapted to account for T₂* decay, which substantially improves the accuracy of fat quantification in the presence of liver, and has been shown to be an accurate measure of fat content both in phantoms and in patients with hepatic steatosis.^9,25 IDEAL has formed the basis of a number of commercially available algorithms for fat quantification designed for liver imaging, and has been widely used in medical imaging research.^30–33

Nonetheless, IDEAL also has limitations. The critical step in the IDEAL algorithm is the estimation of the B₀ field map; however, this always involves a degree of ambiguity, and an incorrect B₀ estimation may occur in areas of large inhomogeneity. Modifications to the IDEAL method using region-growing or graph-cut methods can be used to improve performance in the presence of B₀ inhomogeneity,^34,35 but fat-water swaps remain difficult to prevent entirely, and algorithm development remains an area of active research.

MR spectroscopy

The “gold standard” modality for fat quantification is magnetic resonance spectroscopy (MRS), which typically uses point-resolved spectroscopy or stimulated echo acquisition mode single-voxel sequences.^9,33,36 Spectra can be used to determine both the fat fraction and the specific composition of the fat, including fatty acid composition parameters. However, MRS cannot be used to generate FF maps, and CSE-MRI methods, which allow for high-resolution imaging, are therefore more attractive for medical applications. This is particularly important in the measurement of organ fat where distribution may be spatially heterogenous.

The fat fraction as a biomarker: properties and confounders

Two characteristics of primary importance when measuring FF are accuracy and precision.³⁷ Accuracy can be viewed as the correctness of a biomarker measurement in comparison to a reference standard (an accurate biomarker demonstrates low bias and high linearity), whilst precision is a measure of measurement consistency.^13,37 The accuracy of FF measurements can be determined using fat-water phantoms, typically consisting of emulsions of fat and water which are solidified using agar or thickened using carrageenan.^25,38 Measured FF values can then be directly compared with known fat fraction values in the phantom. Precision (repeatability and reproducibility) of measurements is typically determined using repeat scans, often across multiple scanners and sites.³⁹

Depending on the acquisition, FF measurements may be biased by T₁ and T₂/T₂* relaxation, the presence of multiple fat peaks noise, phase errors, B₁-inhomogeneity and J-coupling.¹³ The bias may be minimized by optimizing acquisition parameters, and/or by using post-processing correction strategies.

T₁-bias can be minimized by using a low flip angle (typically 5° or less) and long repetition times. For gradient echo acquisitions, an important consideration is the number of echoes acquired, since later echoes are subject to a longer period of T₂* decay.⁴⁰ Using maximum-likelihood estimation, T₂* decay can be incorporated into signal models and, therefore, corrected for—this is particularly important in organs containing significant quantities of iron or mineralized bone where T₂* will be short.^26,41 To minimize the effect of the spectral complexity of fat, a pre-defined fat spectrum can be assumed. All CSE-MRI algorithms are also dependent on the choice of echo time, echo spacing and the presence of noise.⁴²

If all potential sources of bias are eliminated, the FF can then be described as a proton density fat fraction (PDFF)—this is the ratio of unconfounded fat signal to the sum of the unconfounded fat and water signals.⁴³FF measurements which are still biased by one or more of these factors may be referred to as signal FFs (sFF). Accurate measurement of PDFF potentially enables comparison of values between scanners, and across multiple sites. Data from multisite, multivendor studies suggest that the reproducibility of “confounder-corrected” PDFF measurements is excellent in the liver,^39,44 although further work is required to examine reproducibility for other applications. Reproducibility is a major strength of CSE-MRI and makes this technique attractive for use in research and clinical practice.

Practical implementation: choosing a technique

Deciding on an imaging acquisition and post-processing method for FF mapping can be a somewhat bewildering process, owing to the large variety of acquisition and post-processing methods available. The most important choice is whether to use a manufacturer-implemented solution, with inline processing on the scanner, or to acquire raw data for offline processing. On most modern scanners, users should have access to simple two- or three-point methods. The major manufacturers also offer dedicated packages for PDFF measurement, which measure and correct for T₂* decay and other confounds, and produce PDFF maps inline. The major advantage of this approach is ease of use; the disadvantage is that the details of the processing methodology are generally not available to the user. Conversely, acquiring raw magnitude and phase images enables offline processing using a chosen reconstruction method, but is dependent on the availability of local expertize and may not be feasible for some centres. However, a number of commonly used methods for post-processing, including analytical three-point methods, versions of the IDEAL algorithm and graph cut-based methods, are available through the International Society of Magnetic Resonance in Medicine fat-water toolbox.¹³ This toolbox enables the user to try different algorithms on their own datasets, enabling them to find a method which works well with their data and suits their specific requirements. At our institution UCLH/UCL), we typically use a 3D multiecho gradient echo acquisition from which we are able to reconstruct both PDFF and R₂* maps using the manufacturer PDFF product, and also extract raw magnitude and phase images for offline-reconstruction. Typical acquisition parameters for manufacturer-implemented sequences for PDFF measurement are given in Table 1; similar parameters can also be used if raw complex data is acquired. For abdominal applications, these parameters allow for good anatomical coverage with a scan time short enough for acquisition within a breath-hold.

Table 1.

Typical imaging parameters for liver PDFF imaging using manufacturer-implemented PDFF products

Parameter	Philips	Siemens	GE
PDFF manufacturer-supplied package	mDixon Quant	DIXON FQ in Liver Lab package	IDEAL IQ
Sequence variant	3D spoiled gradient echo	3D spoiled gradient echo	3D spoiled gradient echo
Imaging time	Breath-hold (<20 s)	Breath-hold (< 20 s)	Breath-hold (< 20 s)
TR	Shortest (5–10ms)	Shortest (5–10ms)	Shortest (5–10ms)
Number of echoes	6	6	6
TE of first echo (TE1)	Shortest (~0.8–1.5ms)	Shortest (~ 0.8–1.5ms)	Shortest (~ 0.8–1.5ms)
Echo spacing (ΔTE)	Shortest (~0.8–1.5ms)	Shortest (~ 0.8–1.5ms)	Shortest (~ 0.8–1.5ms)
Flip angle	3°	3°	3°
Parallel imaging factor	2	2	2
Number of averages	1	1	0.5
Number of shots	–	–	2
Reconstructed images	Fat-only image Water-only image PDFF map T2* map	Fat-only image Water-only image PDFF map T2* map	Fat-only image Water-only image PDFF map T2* map

IDEAL, iterative decomposition of water and fat with echo asymmetry and least squares estimation; PDFF, proton density fat fraction; TE, echo time; TR, repetition time.

How many points?

The ideal situation is to use an acquisition which allows for true PDFF measurement, which generally involves acquiring more than three echoes (at our institution we use six). However, depending on the clinical problem, meaningful FF measurements may be obtained using simple two- or three-point methods. For example, patients with haematological malignancies undergo relatively large changes in FF compared to normal marrow (and with treatment) and a simple sFF measurement is sufficient to demonstrate a clinically meaningful effect.⁴⁵ Conversely, in patients with hepatic steatosis, relatively small variations in FF can be clinically relevant, whilst T₂* effects can cause a significant bias—a true PDFF measurement is, therefore, required.⁹ The choice of acquisition, therefore, depends on the range of values expected in normal tissue, the size of the change expected in disease and the post-processing software which is available to the user.

Echo times

When setting up a CSE-MRI acquisition, the first echo time (TE₁) and echo spacing (ΔTE) need to be optimized. In general, using a shorter ΔTE reduces the sensitivity of the technique to B₀ inhomogeneity (effectively increasing the spectral resolution), thereby improving the quality of fat-water decomposition.¹³ However, a balance needs to be struck between image resolution (higher resolution necessitates longer readouts, and therefore larger ΔTE), and fat-water decomposition quality. At our institution, we typically use a multiecho acquisition with TE₁ of 0.5–1.5 ms and ΔTE 0.8–1.5 ms.

Applications

Hepatic steatosis and the metabolic syndrome

Metabolic syndrome (MetS) is a group of risk factors for cardiovascular disease, diabetes and stroke, which includes obesity, dyslipidaemia, hypertension and raised fasting blood glucose.⁴⁶ Although current definitions of MetS do not include hepatic steatosis, the association between liver fat and MetS is becoming increasingly clear.^47–49 Furthermore, non-alcoholic fatty liver disease (NAFLD) is the most common cause of chronic liver disease in the Western world.⁵⁰ As such, there is a clinical need for simple, non-invasive tools for measuring liver fat. Hepatic steatosis (increased liver fat content) was one of the earliest applications of the Dixon method, and remains one of the most important.

Hepatic steatosis may be either primary (in the case of NAFLD), or secondary, e.g. owing to alcoholic liver disease, drug-toxicity, hepatitis, pregnancy and parenteral nutrition. NAFLD is defined by the presence of hepatic steatosis in the absence of any secondary cause, and lies on the benign end of a spectrum with non-alcoholic steatohepatitis (NASH), which is a more aggressive disease entity associated with cirrhosis, hepatocellular carcinoma and end-stage liver disease.

Underpinning MetS is the phenomenon of insulin resistance (IR), whereby increasing insulin levels are required to achieve a normal metabolic response, or normal insulin levels fail to provide a normal metabolic response.⁵¹ Accumulation of fat within hepatocytes is driven by IR through increases in peripheral lipolysis and de novo lipogenesis, and through reductions in beta-oxidation of fatty acids within hepatocytes.^52,53 Accumulation of intracellular lipid within hepatocytes is the first part of the “two-hit” hypothesis,⁵⁴ after which increased vulnerability to oxidative stress, lipid peroxidation, inflammation and mitochondrial dysfunction in combination cause chronic hepatocyte injury, fibrosis and NASH.⁵⁵ The widespread prevalence of NAFLD has thus become a major healthcare focus, not only to try to avoid the much more aggressive NASH phenotype which may evolve if untreated, but also as a potential therapeutic target to address cardiovascular disease and diabetes mellitus.

The diagnosis of steatosis can be made histologically, but both the European Association for the Study of the Liver⁵⁶ and the American Association for the Study of Liver Disease⁵⁷ use MRI-based criteria to define the condition. For example, the EASL guidelines use a threshold of > 5% as measured using PDFF, or > 5.6% by MRS. MRS has been seen as the non-invasive gold-standard for quantitative assessment of liver fat,⁸ but the technical complexity associated with MRS acquisitions and the spatial heterogeneity of hepatic steatosis have led to the widespread use of imaging-based methods for assessment of steatosis.⁵⁸

Numerous studies have now demonstrated the accuracy of PDFF measurements using CSE-MRI as compared with MRS,^8,59 histology^60–62 or both,^{24,60,63–65} and the reproducibility of PDFF measurements is excellent—Hernando et al³⁹ found an overall intraclass correlation coefficient of 0.999 on repeat scans across sites, field strengths and vendors (example of PDFF of normal liver and NAFLD shown in Figure 3). Similarly, a large meta-analysis including data from > 1000 patients has demonstrated 95% Bland-Altman limits-of-agreement of ± 4% for liver PDFF vs MRS, with liver PDFF reproducibility coefficients of 4.3% across different scanner vendors.⁶⁶ PDFF measurements are strongly correlated with quantitative histopathological assessment of liver fat.^61,63,67,68 Furthermore, longitudinal studies have suggested that PDFF may be more sensitive than histological quantification for early steatosis.^69,70

Figure 3. .

PDFF maps of a comparable axial slice through the abdomen in two subjects. Despite having comparable subcutaneous and intraperitoneal fat volumes, there is differing intravisceral fat with (a) low hepatic and pancreatic fat and (b) NAFLD, with heterogeneous left and right steatosis and pancreatic interlobular fatty infiltration. The pancreas is arrowed on both subjects, and regions of interest have been placed on the liver to demonstrate differences in hepatic PDFF.

The data from these validation studies have supported the use of liver PDFF measurements in clinical trials. PDFF measurements have recently been used as endpoints in double-blind randomized, placebo-controlled trials of colesevelam and ezetimibe for NASH,^14,69 and of insulin glargine and liraglutide for poorly controlled Type2 diabetes.⁷¹ These studies highlight the value of PDFF measurements as primary outcome measures for high quality randomized controlled trials evaluating the next generation of treatments for NASH, NAFLD and diabetes.

With a progressively established role for liver PDFF measurements in the research setting, there are several challenges facing routine clinical implementation. Hepatic steatosis is a spatially heterogeneous process,⁷² and there is some debate about how to condense the information stored in imaging maps into simple indices that can be used by clinicians.^71,73,74 Quantitative liver imaging remains in its infancy, but PDFF has a justified role alongside other quantitative methods (such as T₁ mapping) in multiparametric MRI.⁷⁵ One of the major advantages of PDFF acquisitions is that T₂* maps, which have a role in the quantification of liver iron, can be generated simultaneously. This means that two quantitative parametric maps reflecting different aspects of liver physiology can be generated quickly and simply.

Pancreatic fat – the missing link?

Qualitative assessment of pancreatic fat accumulation is readily undertaken using anatomical MRI, but it can be difficult to separate “fatty infiltration” from small volume “fatty replacement” as these can often occur simultaneously.⁷⁶ Structural imaging changes in pancreatic steatosis include accumulation of extra-, inter- and intralobular fat, in addition to intracellular accumulation of fat, yet the clinical significance of each of these processes remains unknown. As a result, some groups have recently begun to use quantitative MRI to investigate the role ofpancreatic fat.

Pancreatic steatosis research has been hampered by confusing nomenclature (e.g. pancreatic lipomatosis, fatty pancreas, fatty infiltration), but the more recent label of non-alcoholic fatty pancreas disease is gradually gaining acceptance.⁷⁷ Non-alcoholic fatty pancreas disease can occur secondary to several causes (age, viral infections including HBV and HIV, drug-related causes, haemochromatosis and congenital causes), but it is the association with metabolic causes (obesity and diabetes mellitus) that has driven interest in this area of research.⁷⁸

Histologically, pancreatic steatosis is a heterogeneous process characterized by increased intracellular lipid accumulation followed by rising pancreatic adipocytes within pancreatic tissue.⁷⁹ A recent meta-analysis quantified normal mean pancreatic fat at 4.5 ± 0.9% (based on pooled data across multiple MR methods and imaging modalities),⁸⁰ but precise consensus thresholds have yet to be defined. It is possible that triglyceride overload may contribute to beta-cell dysfunction in Type 2 diabetes, although a causal relationship between pancreatic steatosis and diabetes remains unproven.⁸¹ Current evidence regarding the relationship between pancreatic fat, body mass index, diabetes, MetS and IR is somewhat conflicting and the nature of these links remains unclear.^82,83 Some large-scale studies have demonstrated relationships with IR,^84,85 but others have failed to find a consistent link between age, body mass index, incidence of Type 2 diabetes.^84,86,87

Histological validation studies suggest reasonable agreement between histological measures and PDFF, although the repeatability of the measurement is arguably suboptimal (Yoon et al⁸⁸ found Bland-Altman 95% limits of agreement of ± 17% for pancreatic PDFF measurement). There are several specific methodological issues which may contribute to problems with repeatability in the pancreas, including the presence of motion artefact, differences in the composition of fat compared to liver and the fact that pancreatic fat is separated into extra-, inter- and intralobular compartments. Just as in the liver, T₂* decay needs to be accounted for (and may also be a useful biomarker in its own right), particularly in iron-deposition disorders such as haemochromatosis.^89,90

Osteoporosis, obesity and the fat-bone connection

Osteoporosis and obesity are increasingly important public health problems, and are associated with a large burden of morbidity and mortality. Although these diseases are seemingly disparate, both can be traced to the dysfunction of a common progenitor cell—the bone marrow mesenchymal stem cell (MSC). MSCs can differentiate into either adipocytes or osteoblasts depending on their cytokine environment and stromal interactions.⁹¹ Furthermore, adipocytes themselves (which express a variety of paracrine and endocrine factors) may actively influence MSC differentiation and, therefore, affect bone remodeling.⁹¹

Multiple studies have demonstrated an inverse relationship between marrow fat content and bone mineral density or bone formation rate,^92–95 such that osteoblast activity and bone formation rates are both decreased at sites of high marrow fat.^96,97 This has led to a the suggestion that osteoporosis is the “obesity of bone”.⁹¹ Several animal studies have indicated that adipose cells may directly reduce bone mass through the action of leptin.^98,99 However, the precise role of marrow fat in the regulation of bone formation in humans remains unclear, and changes in marrow fat content may partly reflect a passive response to bone loss.^{11,91,100,101} It may be that changes in marrow fat contribute to the pathogenesis of certain subtypes of osteoporosis, but not others.¹¹

In patients with obesity per se, both visceral and bone marrow fat are again adversely associated with bone microarchitecture.¹⁰² Visceral fat may negatively impact on bone health via modifying insulin-like growth factor-1 (IGF-1) levels; vertebral bone marrow fat in postmenopausal females is negatively associated with BMD and IGF-1, but positively associated with visceral fat.¹⁰³ Since bone and fat cells have the same mesenchymal precursor, IGF-1 may influence BMD through its effect on MSC differentiation (into fat or bone lineages).¹⁰³ Accordingly, mice with IGF-1 suppression have fatty infiltration of the marrow and liver and markedly reduced BMD, although they are not obese.¹⁰⁴

Impairments in bone microarchitecture are also observed in diabetic patients, who have an increased fracture risk.¹⁰⁵ Again, this may occur because obesity and IR cause osteoblast and osteoclast dysfunction and, therefore, reduced bone turnover.^105,106 Studies of patients with Type2 diabetes have not revealed significant differences in marrow FF compared to controls, although multiple studies have found that diabetic patients have a greater proportion of unsaturated fat.^107,108 Conversely, the proportion of unsaturated fat in the marrow is thought to decrease in osteoporosis, although the significance of this finding is unclear.¹⁰⁹

Paradoxically, bone marrow fat is increased in anorexia nervosa despite losses in overall body fat.^110,111 A possible explanation is that marrow adipose tissue increases represent a homeostatic response designed to increase appetite and promote insulin sensitivity;¹¹ alternatively, marrow adipose tissue may be maintained in spite of overall adipose tissue loss owing to its role in haematopoiesis and bone turnover.¹¹

FF as a marker of malignant infiltration in bone

In multiple myeloma (MM), the rapid development of novel therapeutic strategies requires improved disease staging techniques. Whole-body MRI (WB-MRI) is now widely available, relatively cheap and establishing itself as a first-line imaging modality in MM.^112,113 CSE-MRI may be used as a fast“anatomical” imaging modality and enables the simultaneous generation of up to four image types using a single sequence (in phase, out of phase, WO and FO), which are used for qualitative interpretation. Additionally, quantitative FF measurements may be used as a marker of marrow composition, and are thought to reflect the proportion of haematopoietic (red) and fatty (yellow) marrow, which may be altered by malignant infiltration.¹¹⁴ Patients with symptomatic MM have significantly lower PDFF measurements than those with asymptomatic disease.¹¹⁴ Furthermore, recent work analysing sFF measurements in focal MM lesions suggests that sFF measurements may stratify patients according to their depth of therapeutic response.⁴⁵ sFF measurements appear to be better discriminator of response than alternative tumour volume or apparent diffusion coefficient (ADC) values.⁴⁵

CSE-MRI may also be used to detect skeletal metastases in patients with solid tumours, and is reasonably sensitive and specific (70.8 and 89.1%, respectively).¹¹⁵ FF measurements are also reduced in patients with a malignancy but without skeletal involvement, possibly reflecting red marrow proliferation FF owing to systemic effects of malignancy or drugs.¹¹⁶

FF as a measure of skeletal inflammation

An emerging application for PDFF measurements is in patients with spondyloarthritis (SpA), in whom quantification of inflammation is important for guiding disease-modifiying and biological therapy.

MRI is commonly used to diagnose and monitor axial inflammation in SpA,¹¹⁷ but current clinical imaging protocols consist mostly of conventional T₁ weighted, T₂ weighted and short tau inversion recovery spin echo sequences, which can be subjectively interpreted by a radiologist to give a semi-quantitative inflammatory “score”.¹¹⁸ However, visual scoring of inflammation suffers from relatively poor reproducibility - Arnak et al¹¹⁹ reported kappa coefficients of 0.85 for the presence of bone marrow oedema, and 0.61 for global diagnosis of SpA using MRI- and is, therefore, not widely used in clinical practice. PDFF measurements have, therefore, been investigated as an objective tool for quantifying active inflammation (i.e. bone marrow oedema) and chronic inflammation (which causes an increase in marrow fat content known as fat metaplasia) (Figure 4).¹²⁰ PDFF measurements are reduced in areas of active inflammation (likely owing to an increase in water content caused by the inflammatory exudate), but increased in areas of chronic inflammation, possibly owing to the effect of inflammatory cytokines on MSC differentiation.¹²⁰FF measurements in bone marrow are accurate in fat-water phantoms and have been found to offer excellent reproducibility (reported intraclass correlation coefficient values for PDFF in bone vary from 0.87 to 0.98).^{120, 121} They may, therefore, provide a more reliable measurement of inflammatory severity than visual interpretation. A further advantage is that simultaneously generated R₂* (i.e. 1/T₂*) maps are thought to report on BMD, and may provide a measure of new bone formation and bone destruction (both of which are key processes in SpA).¹²⁰ 

Figure 4.

 Emerging application of PDFF and R₂* measurement in spondyloarthritis. Areas of active inflammation (a, c) demonstrate a reduction in fat content, but no change in R₂*. Areas of chronic inflammation (fat metaplasia) (b, d) demonstrate an increase in fat content, and a reduction in R₂* which may indicate local osteoporosis. 

FF in neuromuscular diseases

Neuromuscular diseases involve a range of pathological processes including inflammation, muscle atrophy and fat infiltration, all of which may contribute to functional disability. The lack of sensitive, responsive outcome measures has been a barrier to the development of novel therapies for neuromuscular diseases, and there has, therefore been a drive to develop imaging markers of disease activity and muscle damage.

Morrow et al¹²² showed that whole muscle FF measurements were altered in patients with both Charcot-Marie-Tooth disease and inclusion body myositis, highly responsive to change and correlated with conventional functional measures.¹²² Different imaging biomarkers appear to reflect different disease processes—Yao et al¹²³ showed that muscle T₂ and fat-corrected T₂ measurements were correlated with visual inflammation scores, whilst FF measurements correlated with damage scores, in patients with idiopathic inflammatory myopathy. Technical developments in CSE-MRI methods may enable simultaneous quantitation of inflammation and damage. Recently, Janiczek et al³¹ described a technique for simultaneous quantitation of T₂ of water, T₂ of fat and fat fraction in a patient with inclusion body myositis, which revealed disease-specific patterns of fat infiltration and oedema. Similarly, in patients with myotonic dystrophy, Hiba et al demonstrated a clear relationship between disease severity and FF measurements in the tibialis anterior.¹²⁴ There is preliminary evidence that subclinical progression may be reflected by changes in FF.¹²² FF measurements also agree closely with histological measures of fat infiltration.¹²⁵ Given the increases in fat content which invariably accompany muscle damage, FF measurements are likely to have an increasingly important role in therapeutic trials and, ultimately, in the clinic.

Cardiac and vascular applications of CSE-MRI

Intramyocardial fat may arise in both ischaemic and nonischaemic cardiomyopathies, and is associated with a poor prognosis.¹²⁶ So-called “lipomatous metaplasia” is common in ventricular scar tissue following myocardial infarction,^127–129 and the transformation of compact scar into compressible “sliding” adipose tissue may worsen ventricular function, frequently resulting in severe heart failure.¹²⁹ In patients with dilated cardiomyopathy, fat deposition is associated with increased end-diastolic and end-systolic volumes, decreased left ventricular ejection fraction and increased late gadolinium enhancement (a marker of fibrosis).¹³⁰ A causal link between myocardial fat and functional impairment is currently unproven, although accumulated fat (either within adipocytes or the myocytes themselves) may be directly toxic to the myocardium. It has been suggested that lipid accumulation may sometimes be a precursor to myocardial infarction, potentially due to metabolic changes resulting from low-grade ischaemia, although this suggestion remains unproven.¹³¹

Cardiac fat quantification studies have relied on multiecho Dixon-like methods, which are sensitive to small concentrations of myocardial fat.^28–35 These methods are typically ECG-gated, which enables acquisition of data during diastole and a subsequent improvement in resolution compared to non-gated methods.¹³² Although “conventional” fat suppression using chemical shift selective pulses can also be used in the heart, Dixon-like methods offer superior fat-water decomposition in the presence of B₀ inhomogeneity.¹³² Similar to imaging other organs, T₁ and T₂* decay and noise can potentially confound FF measurements, but can be corrected for.¹³³ Once this correction has been performed, PDFF measurements using Dixon MRI agree closely with MRS measurements.¹³³

Finally, FF and R₂* measurements may be valuable as markers of plaque composition in patients with carotid atherosclerosis – Koppal et al⁴¹ showed that FF measurements reflect lipid composition, whilst R₂* measurements reflect intraplaque haemorrhage.

Perspectives and Conclusion

Far from being an inert energy store, adipose cells are highly physiologically active and have an extensive role in both normal physiology and disease. CSE-MRI is rapidly emerging as the primary clinical and research tool for fat quantification, and has a multitude of existing and emerging applications. Although abdominal fat quantification in obesity and metabolic syndromes are among the most obvious, there are a multitude of applications for CSE-MRI in the muscle, bone, heart and vessel wall.

Nonetheless, there are several barriers to the widespread use of CSE-MRI in clinical practice. Although most manufacturers now offer CSE-MRI packages, the bewildering array of options, confusing nomenclature and overall cost can discourage radiology departments from using these methods. Technical problems with transferring FF maps on to picture archiving and communication system software may also reduce the practicality of using CSE-MRI. However, as awareness and availability of CSE-MRI increase, these issues are likely to resolve. Novel post-processing technologies such as organ segmentation and machine learning may also help the radiologist to maximize the information which can be extracted from FF maps. Looking forward, we anticipate that FF mapping will become a mainstay of clinical radiology, and will be used to objectively characterize disease in a wide variety of organ systems. The FF is here to stay.

FUNDING

This work was undertaken at UCLH/UCL, which receives funding from the Department of Health's the National Institute for Health Research (NIHR) Biomedical Research Centre funding scheme. The views expressed in this publication are those of the authors and not necessarily those of the UK Department of Health. MHC and SP are supported by the NIHR University College London Hospitals BRC. TJPB is supported by Arthritis Research UK Grant 2,1369. MDC is an NIHR Clinical Lecturer.