An Update on MR Spectroscopy in Cancer Management: Advances in Instrumentation, Acquisition, and Analysis

Abstract

MR spectroscopy (MRS) is a noninvasive imaging method enabling chemical and molecular profiling of tissues in a localized, multiplexed, and nonionizing manner. As metabolic reprogramming is a hallmark of cancer, MRS provides valuable metabolic and molecular information for cancer diagnosis, prognosis, treatment monitoring, and patient management. This review provides an update on the use of MRS for clinical cancer management. The first section includes an overview of the principles of MRS, current methods, and conventional metabolites of interest. The remainder of the review is focused on three key areas: advances in instrumentation, specifically ultrahigh-field-strength MRI scanners and hybrid systems; emerging methods for acquisition, including deuterium imaging, hyperpolarized carbon 13 MRI and MRS, chemical exchange saturation transfer, diffusion-weighted MRS, MR fingerprinting, and fast acquisition; and analysis aided by artificial intelligence. The review concludes with future recommendations to facilitate routine use of MRS in cancer management.

Keywords: MR Spectroscopy, Spectroscopic Imaging, Molecular Imaging in Oncology, Metabolic Reprogramming, Clinical Cancer Management

© RSNA, 2024

Summary

This review provides an update on MR spectroscopy (MRS) for applications in cancer, with a focus on new instrumentation, acquisition, and analysis methods, as well as recommendations to facilitate routine clinical use of MRS in cancer management.

Essentials

• ■ MR spectroscopy (MRS) is a noninvasive technique enabling in vivo quantification of tissue metabolites and is particularly well-suited for the assessment of metabolic abnormalities in cancer.

• ■ Recent advances in instrumentation and hardware, acquisition, and analysis using artificial intelligence pave the way for integration of MRS in routine cancer management.

• ■ To achieve full translation to the clinic, standardized pipelines for data acquisition and analysis, clear quantitative metrics, and routine reimbursement by insurance providers are necessary.

Introduction

MRI emerged as a powerful tool for medical imaging in the mid-20th century, and much of the early application and development focused on MR spectroscopy (MRS) (1). While contemporary clinical MRI, particularly in cancer, is largely focused on structural and morphologic imaging, the capability of MRS to facilitate chemical and molecular profiling of tissues without the need for exogenous contrast material, and in a localized, multiplexed, and nonionizing manner, lends itself particularly well to applications in cancer imaging (1). Metabolic reprogramming is a hallmark of cancer, and MRS is invaluable for evaluating potential biomarkers for cancer detection, prognostication, and treatment monitoring. Higher magnetic field strengths (2), multichannel phased-array coils (3,4), and improved pulse sequences (5) have all supported the integration of MRS as a viable clinical tool, setting the stage for a new era of MRS in routine cancer imaging.

In this review, we provide an update on the use of MRS for clinical cancer management. First, we provide an overview on conventional MRS and current metabolites of interest in cancer imaging. We then review recent advances in instrumentation, data acquisition, and analysis using artificial intelligence (AI) and how these may contribute to more widespread adoption of MRS in clinical practice. Finally, we conclude with an outlook for the future, including current challenges and recommendations for the clinical integration of MRS for cancer-associated applications.

Overview of Conventional MRS in Cancer

MRS is a noninvasive method used to quantify metabolic and biochemical processes. Relying on the same physical principles as analytical nuclear MR spectroscopy, MRS is implemented in vivo using MRI scanners (1,6). A localized MR spectrum is generated, with peaks corresponding to nuclei or functional groups within individual metabolites. The area under the spectral peaks in the frequency domain is directly proportional to the concentration of each metabolite (6). Each spectrum contains multiple peaks, and, as a result, MRS is inherently multiplexed. The most common nucleus in biologic and clinical MRS is ¹H (ie, proton), but spectra can be obtained from other MR-active nuclei, such as phosphorus 31 (³¹P), sodium 23 (²³Na), and carbon 13 (¹³C) (1). For ¹H MRS, water suppression is required to eliminate or reduce the large water peak (5). MR spectra can be acquired either from a single region, known as single-voxel spectroscopy (SVS), or from multiple voxels or regions, known as MR spectroscopic imaging (MRSI) (5).

Clinically available pulse sequences for in vivo MRS include stimulated-echo acquisition mode (ie, STEAM), point-resolved spectroscopy (ie, PRESS), localization by adiabatic selective refocusing (ie, LASER), and semi-LASER (ie, sLASER) sequences (5). sLASER is the recommended sequence for SVS from recent MRS consensus efforts (5). The use of spectral editing (eg, using the Mescher-Garwood point-resolved spectroscopy, or MEGA-PRESS, sequence) enables differentiation of overlapping peaks to selectively detect metabolites of interest, such as 2-hydroxyglutarate (2HG) or gamma-aminobutyric acid (6,7). Most SVS sequences can be adapted for MRSI by incorporating additional spatial localization steps (8). For a more comprehensive description of MRS, we refer the reader to books by de Graaf (6) and Barker et al (9), as well as recent MRS consensus recommendations for SVS and MRSI acquisition (5,8).

In the realm of cancer, quantification of metabolic changes has focused largely on a subset of ¹H MRS–detectable metabolites (1). A hallmark of most cancers is metabolic reprogramming, which facilitates the necessary precursors for proliferation as well as the high energy demand of cancer cells (10). The most common form of reprogramming is the Warburg effect, or the shift to aerobic glycolysis from oxidative phosphorylation even in the presence of oxygen (11). Consequently, cancer cells exhibit a unique phenotype characterized by increased glycolytic activity and aberrant phospholipid metabolism. MRS is uniquely suited to the assessment of these alterations, particularly given the high sensitivity of the local tumor microenvironment to both metabolic and inflammatory changes (12). Elevated total choline, primarily in the form of phosphocholines, is a well-established feature of many cancers and frequently observed under inflammatory conditions as well (1,13). N-acetyl-aspartate (NAA) is considered a surrogate for neuronal health and is typically decreased in brain tumors due to the loss or displacement of normal neuronal tissue. Typically, the ratio of total choline to total NAA is used in brain tumors, as decreased total NAA (encompassing both NAA and N-acetyl-aspartyl-glutamate) is generally coupled to an increase in total choline (14). Abnormally elevated lactate due to enhanced glycolysis has been quantified in multiple cancers, including high-grade gliomas (14) and prostate cancer (15).

While the Warburg effect is the most well-documented form of metabolic reprogramming in cancer, glucose alone cannot supply the nitrogen needed for protein and amino acid synthesis. As a result, many cancer cells develop an “addiction” to glutamine as a nitrogenous energy source (16), and prior work has demonstrated upregulation of glutamine synthesis in tumor cells (17). Furthermore, the glutamine pathway has gained traction as a target for cancer therapies, largely focused on glutaminase, for a wide range of cancers, including glioma, leukemia, melanoma, colorectal cancer, and triple-negative breast cancer (18). Importantly, glutamine and metabolites associated with glutaminolysis (eg, glutamate and glutathione) can be quantified with MRS. Ekici et al (17) reported an increase in both glutamate and glutathione with increasing glioma grade. In addition to brain tumors, elevated glutamine uptake and metabolism has been observed in nuclear MRI studies of multiple malignancies, including thyroid (19) and prostate (15). The overlapping peak resonances of glutamine and glutamate in MRS spectra is a key limitation when quantifying glutamine in vivo, but this can be ameliorated with ultrahigh-field-strength (UHF) MRS (>3 T) (see UHF MRI Systems). Additional MRS-detectable metabolites of interest include lipids, which provide insights into, for example, activation of lipogenic enzymes (1,20), and citrate, related primarily to prostate tumor aggressiveness (21). Decreased levels of citrate coupled to elevated levels of choline, spermine, and creatine is a consistent observation in ¹H MRS in prostate cancer and commonly reported as the (choline + spermine + creatine)-to-citrate ratio (21).

While many of these MRS-detectable metabolites are commonly up- or downregulated in cancers, most are not true molecular biomarkers, except for 2HG. 2HG is an oncometabolite produced as a result of mutations in the genes encoding isocitrate dehydrogenase enzymes (7). First identified 15 years ago in gliomas and acute myeloid leukemia, 2HG has more recently been detected in other cancers, including breast, colorectal, prostate, and lung (22). Detection of 2HG at 3 T is possible (7); however, the overlap of 2HG peaks and relatively low in vivo concentrations continue to be key challenges.

MR spectra acquired in benign and tumor tissue in the brain, prostate, breast, and liver are shown in Figure 1. While conventional MRS approaches have effectively showcased the applicability to oncology, the evolving landscape of medical imaging has introduced new capabilities and expanded applications of clinical MRS through advances in instrumentation, acquisition, and analysis facilitated by AI. Each of these emerging areas is expanded below, ending with recommendations for future clinical integration.

Figure 1:

Proton MR spectra acquired in vivo at clinical field strengths in benign and tumor tissue in brain, prostate, breast, and liver. (A) Brain spectra acquired with single-voxel MR spectroscopy (MRS) at 3 T in an individual diagnosed with isocitrate dehydrogenase–mutant astrocytoma. MRS voxel positions (white squares) in normal-appearing (left) and tumor (right) tissue are overlaid on an axial MR image. The raw (black), fitted (red), residual (orange), and 2-hydroxyglutarate (2HG)–fitted (pink) spectra are shown for both voxels. The tumor spectrum shows elevated choline (Cho) and decreased N-acetyl-aspartate (NAA) (red arrows) relative to normal-appearing tissue. The 2HG peak is characteristic of isocitrate dehydrogenase–mutant gliomas. Creatine (Cr) is visible in both spectra, and lactate (Lac) is distinguishable only in the tumor spectrum. Adapted and reprinted, under a CC-BY-4.0 license, from reference 83. (B) Multivoxel MR spectroscopic imaging (MRSI) scan acquired at 3 T in tumor (red) and benign (blue) prostate tissue in an individual with histologically confirmed prostate cancer. The MRSI grid is overlaid on a coronal MR image of the prostate. Raw (black) and fitted (red) spectra are shown. The tumor spectrum illustrates decreased citrate (Cit) and spermine (Spm) and increased Cho (red arrows) relative to benign tissue. Myo-inositol (mI) and taurine (Tau) are also visible in the spectra. Adapted and reprinted, under a CC-BY-4.0 license, from reference 21. (C) Sagittal postcontrast T1-weighted MR images of the breast showing single-voxel positions (white rectangles) and corresponding MR spectra containing Cho and overlapping Lac/lipid (Lip) peaks in an individual with biopsy-confirmed invasive ductal carcinoma (top) and another with benign papilloma (bottom). The spectrum acquired from carcinoma tissue shows an elevated Cho peak (red arrow) absent in the spectrum acquired from benign tissue, illustrating the ability for MRS to differentiate tumor and benign tissue. Adapted and reprinted, with permission, from reference 20. (D) Liver spectra acquired with single-voxel MRS at 3 T from an individual diagnosed with hepatocellular carcinoma (top) and a healthy volunteer (bottom). MRS voxel positions (white squares) are overlaid on axial MR images. The spectrum acquired from the malignant lesion shows the presence of Cho (red arrow) absent in healthy liver tissue. Adapted and reprinted, under a CC-BY-4.0 license, from reference 84.

Emerging Instrumentation and Hardware

Many limitations in spatial and spectral resolution of MRS as well as signal-to-noise ratio (SNR) are due to limitations in hardware. Herein, we describe the benefits of both UHF and hybrid PET/MRI systems for MRS, as well as current roadblocks for clinical cancer imaging.

UHF MRI Systems

UHF MRI systems provide a promising avenue for noninvasive imaging in oncology. UHFs above 3 T address limitations of lower field strengths, including overlapping peaks and low SNR. In contrast to 3 T, MRS at 7 T can resolve the spectral overlap of glutamine and glutamate peaks (classified together as Glx at lower field strengths) (2), resolve individual fatty acids (eg, the 2.1-to-1.3 ppm fatty acid ratio relevant to cervical cancer) (23), and improve specificity and sensitivity of 2HG detection (24).

Metabolite quantification using UHFs has been demonstrated in multiple in vivo studies, supporting a strong outlook for UHF MRI as a potential addition to oncologic imaging (8,23). At this time, 7-T systems are the most common UHF imaging systems, but human scanners have been built with field strengths of 9.4, 10.5, and 11.7 T (2). Clinical MRI and MRS at UHFs have primarily focused on brain and musculoskeletal applications (2). Cost is currently one of the largest barriers to the adoption of UHF. Additional challenges of MRS, including procedural consistency, chemical shift displacement errors, field inhomogeneity, susceptibility to motion, and safety (eg, high specific absorption rates), are exacerbated at higher field strengths. High-order shimming (greater than second degree) at UHF is required to overcome patient-specific field inhomogeneities. Technologies to address these challenges include resistive shim coils, each producing a spatial field distribution approximating a spherical harmonic, and resistive multicoil shim arrays, consisting of a multicoil array placed in proximity to the patient and facilitating local B₀ shimming (25). Parallel transmission using multichannel phased arrays has been proposed as a promising method to overcome radiofrequency field inhomogeneity by correcting the radiofrequency field (static parallel transmission) or flip angle (dynamic parallel transmission) to achieve uniform excitation (2). To address challenges of patient motion, prospective motion correction can be achieved using optical tracking cameras inside or outside the MRI scanner bore. Additionally, new coil designs with built-in cameras have been demonstrated to improve the line of sight and facilitate real-time motion correction (26). In the future, a major goal is the design of strategies and MRS acquisition protocols specific to UHF MRI systems to fully maximize their potential (5).

Hybrid PET/MRI Systems

PET/MRI, the hybrid modality integrating both PET and MRI scanners in a single imaging system, presents a valuable addition to medical imaging by providing complementary metabolic, structural, and morphologic information in a single scan (27). Metabolic imaging using both MRS and PET in the same scanning session can quantify glucose-related metabolites as markers for tumor progression, including glutamine, glutamate, 2HG, and lactate, allowing for a more comprehensive assessment of tumor progression and prognosis. For example, in clinically high-risk prostate cancer, the combined use of both MRS and fluorine 18 (¹⁸F) fluciclovine PET enabled improved discrimination of voxels depicting malignant neoplasms compared with either modality alone (28) (Fig 2). Although PET/MRI systems have yet to be widely implemented, the unique advantages of hybrid systems may lead to more widespread use. While hybrid systems combining PET with CT are more common, the use of an MRI system rather than CT avoids additional radiation exposure (27). This is particularly beneficial for patient populations who require repeated imaging or circumstances in which the concern for radiation burden is especially high (eg, pediatric patients).

Figure 2:

Simultaneous PET and MR spectroscopy (MRS) acquisition in high-risk prostate cancer. Hybrid PET/MRI allows simultaneous and complementary acquisition of metabolic, structural, and morphologic data. The combination of features extracted from PET and MRS enables better diagnostic performance of clinically high-risk prostate tumor voxels than either technique alone, highlighting the added value of combined PET and MRS. (A) Fluorine 18 fluciclovine PET image acquired in an individual with high-risk prostate cancer shows enhanced radiotracer uptake in the right lobe of the peripheral zone of the prostate, suggesting increased protein metabolism. Color bar represents mean standardized uptake value (SUV_mean). (B) Multivoxel MRS grid overlaid on a T2-weighted MR image. The inset in the top right corner shows the corresponding histologic slice with outlined lesion area (black outline). (C) MR spectra from the red voxels and in the same order as shown in the grid in B shows citrate (Cit) and choline + spermine + creatine (CSC) peaks in tumor regions. a.u. = arbitrary units. Reprinted, under a CC-BY-4.0 license, from reference 28.

Emerging Methods for Acquisition

While most clinical MRS currently relies on ¹H nuclei, emerging acquisition strategies enabling the study of new metabolites or metabolic pathways have expanded the MRS arsenal. Although not an exhaustive list, recent or reemerging methods most relevant to cancer imaging are surveyed below.

Deuterium MRS

A recently revisited MRI approach with broad applicability to cancer imaging is deuterium (²H) MRS, commonly referred to as deuterium metabolic imaging (DMI) (29). In contrast to ¹H MRS, DMI directly monitors metabolic flux, including the uptake of labeled substrates and their downstream metabolites (29). DMI does not require water suppression, and ²H metabolites have shorter T1 values and larger magnetic moments, facilitating shorter scanning time and higher SNR. In addition, the smaller gyromagnetic ratio of ²H compared with ¹H results in reduced sensitivity to magnetic field inhomogeneity (29), which is particularly beneficial at higher field strengths. Due to the low natural abundance of ²H (approximately 0.01%), DMI requires oral intake or intravenous infusion of deuterated substrates. Furthermore, most clinical and preclinical DMI studies have been performed at UHF, but DMI at 3 T has been demonstrated in healthy volunteers (30). Robust and optimized protocols for DMI in patients are needed at 3 T to accelerate clinical translation.

The most common ²H substrate is [6,6´-²H₂]glucose. Clinically, glucose dysregulation is typically quantified with PET using the glucose analog ¹⁸F-fluorodeoxyglucose; however, PET requires radioactive and ionizing substrates, only quantifies cellular glucose uptake, and does not provide downstream information on key metabolic byproducts (29). In comparison, in vivo quantification of Glx and lactate in both patients with glioblastoma and animal models after oral intake of [6,6´-²H₂]glucose has been demonstrated with ²H MRSI (29). De Feyter et al (29) demonstrated the use of the ratio of lactate over Glx as a measure of the Warburg effect. Increased ²H lactate-to-Glx signal was observed in tumor regions, indicating increased glycolysis (Fig 3). Additional glycolytic substrates, including ²H acetate, pyruvate, and 3-O-methylglucose, have also been explored in applications ranging from glioblastoma to liver cancer (29,31,32) (Table).

Figure 3:

Deuterium metabolic imaging (DMI) in an individual with glioblastoma after oral [6,6’-²H₂]glucose intake. (A) Standard-of-care structural MR images acquired in an individual diagnosed with glioblastoma in the right frontal lobe. Subtotal resection of the lesion was performed 9 months before the DMI examination. Images from left to right (displayed in radiologic orientation) include axial T2-weighted (T₂W) fluid-attenuated inversion-recovery (FLAIR) image, T1-weighted contrast-enhanced (T₁W CE) image, susceptibility-weighted imaging (SWI) scan, and diffusion-weighted imaging (DWI) scan. (B) MR spectroscopy voxels, acquired in tumor (1, black square) and healthy tissue (2, white square), are displayed on an axial T2-weighted MR image. Deuterium (²H)–labeled glucose DMI maps from the same section illustrate a relatively homogeneous distribution across the brain but show lower levels of ²H-labeled glutamine + glutamate (Glx), accompanied by a higher concentration of ²H-labeled lactate (Lac) in tumor regions. (C) ²H MR spectra from regions indicated in B show glucose (Glc) and Lac peaks in the tumor region and Glc and Glx peaks in healthy tissue. (D) The Warburg effect is represented by a three-dimensional rendering of combined structural MR images and Lac-to-Glx ratios obtained with DMI. The tumor region shows elevated rates of aerobic glycolysis, demonstrating the added value of DMI to standard-of-care imaging for glioblastomas. Adapted and reprinted, under a CC-BY-NC-4.0 license, from reference 29.

Emerging Methods for MR Spectroscopy Acquisition and the Associated Metabolic Probes Under Development for Cancer Applications

Fumarate, an intermediate in the tricarboxylic acid cycle and precursor of malate, can be imaged with DMI to assess tissue viability. In healthy cells, fumarate is metabolized relatively slowly; however, in necrotic tissue, fumarate rapidly gains access to intracellular fumarase via a leaky plasma membrane, resulting in a substantial increase in malate production. Consequently, the [2,3-²H₂]malate-to-[2,3-²H₂]fumarate ratio has been exploited as a proxy for cell death. Initial experiments administering deuterated fumarate were used to measure treatment-induced cell death in human breast and colorectal xenograft models and in a murine lymphoma model, resulting in high tissue contrast (33). DMI of the production of [2,3-²H₂]malate from [2,3-²H₂]fumarate was a more sensitive marker for early treatment response compared with both diffusion-weighted imaging and contrast-enhanced MRI when evaluated in patient-derived murine models of glioblastoma (34).

Increased choline uptake in tumors has been investigated with ²H-labeled choline substrates in human breast cancer and renal carcinoma xenografts, indicated by increased choline signal after infusion of [²H₉]choline that stabilizes over time (32). The ability to image metabolic flux with DMI is a key advantage over static measurements with ¹H MRS. Despite evidence of the viability and potential of DMI as a method for molecular and metabolic imaging in cancer (29), challenges include the need for specialized receiver coils operating at ²H frequency and, often, vendor-specific hardware to detect deuterium signals, along with the limited number of deuterated tracers currently available (35). For further information about DMI and ²H labeled substrates, we refer the reader to the review by Chen Ming Low et al (32).

Hyperpolarized ¹³C MRI and MRS

Hyperpolarized (HP) ¹³C MRI and MRS (36) enable the study of dynamic carbon metabolism by administering HP ¹³C–labeled probes in vivo, which facilitates assessment of metabolic flux due to the low natural abundance of ¹³C (1.1%), and overcome challenges of low polarization and smaller gyromagnetic ratio compared with other MR-active nuclei. Induced hyperpolarization is achieved by mixing free radicals with the ¹³C probe at a low temperature and high magnetic field strengths. This results in electron transfer to the probe, increasing polarization by a factor of approximately 10⁵ with a subsequent increase in MR signal (36,37). The contrast agent is injected intravenously, and spectroscopic or imaging sequences are applied to generate metabolic maps or MR images, respectively, of the HP metabolite and its metabolic byproducts. Dynamic ¹³C acquisition is of particular interest, enabling quantification of metabolic conversion rates (37). HP ¹³C is well-suited for cancer applications, particularly due to its ability to image the consequences of aberrant glycolysis and other metabolic reprogramming. The most widely used HP probe is [1-¹³C]pyruvate, typically used to track the ¹³C-pyruvate to ¹³C-lactate conversion related to the Warburg effect (37) (Fig 4). HP ¹³C-pyruvate is available as an investigational new drug for clinical studies, with most applications in oncology. Clinical studies have focused primarily on prostate, breast, pancreatic, and brain cancers (36–38) (Table).

Figure 4:

Hyperpolarized carbon 13 (¹³C) MR spectroscopy of pancreatic cancer using the [1-¹³C]pyruvate probe. Top row: From left to right, images correspond to an axial anatomic MR image acquired in an individual with a pancreatic tumor, the apparent diffusion coefficient (ADC) map, and a representative ¹³C MR spectrum. The exocrine pancreas (purple arrows) and the tumor (white arrows) are outlined in green in the anatomic image. Decreased apparent diffusion coefficient is present within the tumor tissue, suggestive of hypercellularity. The ¹³C MR spectrum acquired from the tumor region shows clear peaks for lactate, alanine, and pyruvate. Bottom row: ¹³C maps acquired using multivoxel MR spectroscopic imaging measured more than 35 seconds after injection of [1‐¹³C]pyruvate and overlaid onto anatomic images. From left to right, [1‐¹³C]lactate, [1‐¹³C]alanine, and [1‐¹³C]pyruvate maps are displayed, showing increased ¹³C signal in the tumor region for all three metabolites. Reprinted, with permission, from reference 38.

In addition to pyruvate, the use of HP ¹³C bicarbonate (H¹³CO₃^–) has been successfully used as a pH mapping tool in tumors. Recently, Mu et al (39) developed a clinically translatable method to produce an HP H¹³CO₃^– solution for imaging extracellular pH, currently under evaluation as an investigational new drug in a clinical trial for prostate cancer (ClinicalTrials.gov identifier NCT05851365). Additional probes with translational potential include HP ¹³C urea colabeled with nitrogen 15 (¹⁵N) to increase ¹³C relaxation times (ie, HP [¹³C,¹⁵N₂]urea), which is used in combination with HP [1-¹³C]pyruvate for simultaneous metabolic and perfusion imaging (40). This approach provides insight into both tumor progression and therapeutic response due to the often opposing changes in metabolism and perfusion in tumors (ie, decreased metabolism and increased perfusion in response to treatment). The dual-probe HP imaging technique has been validated in preclinical studies and demonstrated in a patient with prostate cancer (40). HP [1,4-¹³C₂]fumarate has been evaluated after anticancer therapy in animal models of lymphoma, breast, liver, and colon cancer. Increased fumarate to malate conversion, indicative of tissue necrosis, was observed, followed by decreased tumor volumes in all cancer models (37).

Similar to the dual-probe HP ¹³C and ¹⁵N approach, combined HP ¹³C and DMI is currently under investigation, as deuteration can extend ¹³C polarization lifetimes (ie, relaxation times). HP [5-¹³C,4,4-²H₂,5-¹⁵N]-l-glutamine was used to monitor glutaminase flux in models of pancreatic ductal adenocarcinoma (41). A similar approach was used in a murine lymphoma model, where deuterated ¹³C-labeled glucose ([U-²H,U-¹³C]glucose) was used to image flux across 11 steps of glycolysis in models of lymphoma and lung carcinoma (42), and in oligodendrogliomas to assess telomerase transcriptase activity, demonstrating the ability to noninvasively address telomere maintenance (43). The major limitation of HP ¹³C is the lack of hyperpolarizers at most clinical sites and the need for hardware operating at the ¹³C frequency. While other nuclei are not covered in this review, the need for specialized hardware, including coils tuned to each resonance frequency, is a requirement and often a barrier for all non-¹H nuclei (eg, ³¹P). Nevertheless, clinical studies confirm preclinical findings and highlight the potential of HP ¹³C MRI and MRS in cancer as a valuable tool for early response assessment to neoadjuvant therapy and subsequent clinical decision-making (36).

Chemical Exchange Saturation Transfer

Chemical exchange saturation transfer (CEST) is a molecular imaging contrast technique capable of detecting compounds with exchangeable hydrogen atoms between water and nonwater compounds (44). CEST involves the application of a radiofrequency pulse at the resonance frequency of the target moiety, resulting in a reduction of the water signal proportional to the concentration of the target compound, typically presented as a z-spectrum (45) (Fig 5).

Figure 5:

Principles of chemical exchange saturation transfer (CEST) and amide proton transfer (APT) images acquired in human brain and breast. (A) In CEST, solute protons (¹H) (blue) are saturated (ie, MR signal is nulled) after a radiofrequency (RF) pulse is applied at their resonance frequency (8.25 ppm for amide protons). Due to the exchangeable nature of these protons, the saturation is transferred to water (4.75 ppm) by exchanging solute protons for unsaturated protons (black) (k_sw represents the exchange rate of saturated solute protons to water; k_ws represents the return rate of unsaturated protons). (B) After a time delay (t_sat), the saturated water signal (S_sat) is attenuated relative to the unsaturated water signal (S₀). (C) The resulting z-spectrum (or CEST spectrum) is produced by repeated and selective saturation over a range of frequencies. The z-spectrum represents the ratio of saturated water signal normalized by the unsaturated water signal as a function of irradiation frequency. The reduction in water signal is an indirect measure of the concentration of solute, typically assigned as 0 ppm in z‐spectra. The new frequency of the exchangeable solute protons is defined with respect to water (ie, Δω). A–C reprinted, with permission, from reference 45. (D) APT is the most widely used form of CEST imaging. Axial structural MR image and APT image were acquired in an individual with a brain tumor, where increased APT signal in the tumor region compared with nontumorous tissue was observed. Adapted and reprinted, under a CC-BY-4.0 license, from reference 46. (E) APT images overlaid on a dynamic contrast-enhanced image acquired in an individual with breast cancer before the start of neoadjuvant chemotherapy (NAC) (top) and after the first cycle (bottom). The tumor region is outlined in green, and the APT maps show a signal decrease after one cycle of neoadjuvant chemotherapy, illustrating the potential of APT imaging for treatment monitoring in patients with breast cancer. Adapted and reprinted, under a CC-BY-4.0 license, from reference 47.

The most common in vivo application of CEST is amide proton transfer (APT), currently the only form of CEST imaging approved by the U.S. Food and Drug Administration (44). APT leverages the CEST effect of amide protons, primarily in proteins, with a saturation frequency of approximately 3.5 ppm (45). APT has primarily been applied to brain tumors (46) (Fig 5D). The potential added value of APT has also been demonstrated in breast cancer to improve tumor monitoring after neoadjuvant chemotherapy (47) (Fig 5E). In addition to widely used APT imaging, new CEST agents have emerged that may provide expanded insight for improved cancer management.

Glucose as a CEST agent, or glucoCEST, using both glucose and glucose analogs (eg, 2-deoxy-d-glucose, glucosamine, 3-O-methylglucose, 6-deoxy-d-glucose), has been investigated as a potential alternative to ¹⁸F-fluorodeoxyglucose. GlucoCEST does not require ionizing radiation and comes at a substantially lower cost. The GlucoCEST Imaging of Neoplastic Tumors consortium was created in 2020 to coordinate efforts to overcome the boundaries for clinical translation (48). The primary objectives include efforts to make glucoCEST technology available at clinical field strengths, establish detection thresholds in animal models, and expand on the toxicology, safety, and efficacy profiles of glucose and 3-O-methylglucose.

In addition to glucose, the selective imaging of glutamate with CEST, or gluCEST, is performed by applying selective saturation of the glutamate amine group at 3.0 ppm, which is well-separated from the MR frequency of water. GluCEST is proposed as the only noninvasive technique with the potential to assess changes in tumor glutamate levels, specifically outside the brain, and has been demonstrated in a mouse model of breast cancer to detect early response to treatment with glutaminase inhibition (49). Additional CEST agents that provide insights into tumor metabolism include acidoCEST to assess extracellular tumor acidosis (50,51) and lactate CEST, or LATEST, to evaluate lactate concentration (52) (Table).

CEST imaging, while a valuable technique, typically enables the quantification of one or two predefined compounds, and the exchange rate of the exchangeable protons must be smaller than their frequency offset from water to generate sufficient contrast. CEST quantification can be improved with UHF due to increased Larmor frequencies at higher field strengths. Additionally, due to the lack of labeling, CEST imaging is unable to quantify downstream metabolic byproducts as with other techniques. While APT imaging can be performed at 3 T, the applicability of other CEST modalities at clinical field strengths remains challenging, and much of the development work has been performed on UHF animal scanners. There is a need for more sensitive CEST agents that can be widely implemented on clinical scanners. For an in-depth review on CEST imaging, we refer the reader to the article by Goldenberg and Pagel (53).

Diffusion-weighted MRS

Diffusion-weighted MRS (DW-MRS) enables quantification of diffusion-related properties of metabolites, including those within the intracellular space. DW-MRS data are acquired by integrating diffusion-weighting gradients into the spectroscopic sequence of choice. Similar to diffusion-weighted imaging of water molecules, DW-MRS examines the same properties (eg, apparent diffusion coefficient) but for nonwater metabolites (54). Compared with MRS, DW-MRS has the potential to differentiate between intra- and extracellular signals at high diffusion weightings (55). Additionally, unique information can be derived from diffusion properties of metabolites, such as cell size and morphology, and properties of the intracellular environment (eg, viscosity, molecular crowding, tortuosity) (56). This may provide valuable insights about the heterogeneous tumor microenvironment with multiple types of tissue (eg, viable, necrotic, malignant, edema), each with unique properties.

To date, DW-MRS has been used to monitor treatment response of human breast cancer cell lines and murine melanoma cell lines when treated with the anticancer drug lonidamine (55) and to assess diffusion abnormalities in brain tumors (57). Despite current efforts to adapt DW-MRS sequences and postprocessing pipelines to clinical 3-T scanners (56), further research is needed to facilitate the widespread use of DW-MRS in cancer, particularly to overcome the challenges posed by the addition of diffusion-weighting gradients to MRS sequences (ie, higher motion sensitivity and corruption of MRS signal). The ability to directly probe the heterogeneous tumor microenvironment is a key motivator for continued exploration of cancer applications of DW-MRS.

MR Fingerprinting

MR fingerprinting is a promising quantitative technique facilitating simultaneous acquisition of multiple tissue property maps. Each “fingerprint” refers to the unique signal from each voxel acquired by modifying acquisition parameters over time. After acquisition, the resulting signal is mapped to a simulated dictionary containing all possible signals from different combinations of the tuned parameters, which enables signal decoding and generation of multiple property maps from a single acquisition (58). Initial studies have demonstrated the feasibility of MR fingerprinting for applications in cancer, including prostate, liver, and brain tumors (58); however, much of the work combining MRS with MR fingerprinting has focused on demonstration in healthy volunteers (59). Several technical difficulties, such as undesirable processing time or lack of robust motion correction, remain outstanding. In light of the valuable information provided by MR fingerprinting in a short acquisition time, successful integration of MRS with other quantitative metrics provided by MR fingerprinting is an important goal.

Fast Acquisition

New techniques in fast MRSI, largely focused on whole-brain imaging, have been developed to overcome obstacles, such as long acquisition times and the need for manual single-voxel placement. We describe two methods, both of which rely on echo-planar imaging for three-dimensional acquisition: echo-planar spectroscopic imaging, or EPSI, and spectroscopic imaging by exploiting spatiospectral correlation, or SPICE. EPSI was one of the earliest high-speed MRSI approaches (60), combining the principles of fast echo-planar imaging with MRS to acquire spatially resolved spectra in approximately 15 minutes. EPSI has been successfully used in neuro-oncology to discriminate between tumor true progression and pseudoprogression (61) and detect microscopic tumor infiltration undetectable with conventional MRI, which may aid in more precise radiation therapy (62).

SPICE is a model-based reconstruction method first proposed by Liang and Lauterbur in 1994 (63). Following acquisition of raw MRSI data (using multiecho or echo-planar imaging) and preprocessing, SPICE uses mathematical algorithms to exploit spatial and temporal spectral correlations and derive low-dimensional representations or subspaces from high-dimensional signals (8). Each subspace is predetermined from training data and enables extraction of spectral information from multiple voxels to produce a high-spatial-resolution image within 6 minutes, avoiding the need for water and lipid suppression (64). SPICE has been mainly applied to the brain (65) (Fig 6) but has also been demonstrated for prostate cancer imaging (66). While neither of these fast MRS methods are currently product sequences provided by any MRI vendor, whole-organ metabolic imaging in a clinically feasible time window has high potential value, as long scanning times are a current barrier to routine clinical MRS.

Figure 6:

Fast whole-brain acquisition of MR spectra using spectroscopic imaging by exploiting spatiospectral correlation, or SPICE. Whole-brain and high-resolution spectra were acquired in an individual with a low-grade glioma, with a total scanning time of 9 minutes. Left: The first and second rows of images correspond to axial T1-weighted magnetization-prepared rapid gradient echo contrast-enhanced (MPRAGE+C) and fluid-attenuated inversion-recovery (FLAIR) images, respectively. The third and fourth rows are metabolic maps of N-acetyl-aspartate (NAA) and choline (Cho)–to–creatine (Cr) ratios, respectively. Arrows point to the tumor region. Right: Representative spectra are displayed from voxel locations indicated in the MPRAGE+C image (green = normal, blue = edema, and red = glioma). Peaks corresponding to NAA, Cho, and Cr metabolites are identified. The three regions exhibit distinct spectral features. Compared with healthy tissue, the edema region displays a decrease in NAA, while the tumor region shows an increase in Cho-to-Cr ratio as well as a decrease in NAA. Reprinted, with permission, from reference 65.

Emerging Analysis Using AI

Over the past 3 decades, technologic advances in medical imaging and widespread access to computational power have paved the way for AI integration in both radiology and oncology, and efforts are underway to integrate them into routine cancer imaging (67). Of AI tools currently approved by the Food and Drug Administration, image-based algorithms account for 70%, with most applications related to radiology (67). We highlight the most viable applications of AI and machine learning in MRS for cancer applications, including pre- and postprocessing of MRS data, quality control, and multiparametric approaches for stratification.

AI-driven Improvements in MRS Pre- and Postprocessing

AI has high potential to streamline pre- and postprocessing pipelines for MRS data, including denoising, quality control, and feature selection. AI-driven models have demonstrated the ability to improve both SNR and quality control pipelines. For example, extracted features from MR spectra derived from a continuous wavelet transform were used in a support vector machine to discriminate signal from noise (68). Significant SNR increases were reported after denoising spectra acquired with a low number of repetitions. In addition to improvements in SNR, prior reports have also demonstrated the use of AI for spectral quality control. Kyathanahally et al (69) compared several combinations of methods for feature selection, feature extraction, and classification to detect poor-quality MR spectra acquired from brain tumor regions and emphasized the need for balanced data sets and multiclass approaches for spectral classification. Gurbani et al (70) trained a convolutional neural network that avoids the use of engineered features using EPSI data from patients with glioblastoma to determine spectral quality. They developed and implemented a whole-brain pipeline that can be used with existing operating systems on clinical MRI scanners to automatically determine spectral quality with processing times of approximately 2 minutes, allowing filtering of whole-brain EPSI data in real time. Notably, the importance of robust spectral quality control and the feasibility of using AI models for automatic whole-brain spectral filtering was demonstrated (Fig 7).

Figure 7:

Convolutional neural network (CNN) used to filter artifacts of MR spectroscopic imaging (MRSI) data acquired in individuals with glioblastoma. The use of artificial intelligence (AI) can enhance MR spectroscopy processing pipelines and data interpretability by enabling automated spectral quality control. A whole-brain pipeline was developed using a convolutional neural network to filter poor-quality MRSI spectra acquired in individuals with glioblastoma. The convolutional neural network outputs the probability of each spectrum to be of good quality and filters those that fall below a set quality threshold, removing them from the metabolic map. Left: Structural MR images with superimposed choline-to–N-acetyl aspartate (NAA) ratio color maps acquired in an individual with newly diagnosed left frontal glioblastoma following surgical resection of the tumor and before starting radiation therapy. Images and metabolic maps are shown before (no filter) and after convolutional neural network filtering (CNN filtered). The choline to-NAA ratio maps suggest residual tumor (high choline-to-NAA ratios indicated in red in the images) posteriorly and medially to the surgical cavity. Elevated choline-to-NAA ratios are also observed in the unfiltered maps in the anterior tip of the left frontal lobe (voxel A), in the necrotic tumor core (voxel B), and in the occipital lobe (voxel C). Right: Spectra from voxels A, B, and C in the left images are displayed. Voxel A was poor quality, attributed to its position near the periphery. Voxel B is in a necrotic region, and low signal-to-noise ratio suggests reduced metabolic activity. Voxel C displays an upfield artifact resulting in incorrect fitting of the NAA peak and subsequently erroneously high choline-to-NAA ratio. The convolutional neural network was able to detect and remove these three voxels that would otherwise be included in the analysis as areas with high choline-to-NAA ratio, highlighting the importance of spectral quality control and potential of AI for reducing false-positive findings in clinical decision-making. Reprinted, with permission, from reference 70.

AI-assisted Applications of MRS in Diagnosis, Prognosis, and Treatment Stratification

AI may be particularly well-suited to exploit the multiplexed nature of MRS, complementing information gleaned from MRI for tumor grading and treatment stratification. In gliomas, Ranjith et al (71) used MRS features to classify tumors as low- or high-grade. Five metabolite ratios (including NAA, choline, and creatine peaks) were used as features for four different machine learning algorithms, and their performance was compared with results from histopathology. The random forest approach resulted in the highest area under the receiver operating characteristic curve (0.91), suggesting that automated methods for glioma classification using MRS may complement conventional invasive methods.

Tiwari et al (72) combined features from both MRS and T2-weighted MRI to determine the appropriate treatment for patients with prostate cancer. They presented a computerized decision support system called Semi Supervised Multi Kernel Graph Embedding, or SeSMiK-GE, combining structural imaging and MRS data channels (ie, three-dimensional imaging features from T2-weighted images along with one-dimensional spectral features from MRS) for distinguishing benign versus cancerous and high– versus low–Gleason grade prostate cancer regions. Combined features from MRI and MRS acquired using SeSMiK-GE outperformed individual protocols (ie, T2-weighted MRI or MRS alone) or other integration strategies (ie, combining decisions from individual classifiers), supporting the value of incorporating AI-derived MRS features for patient stratification in prostate cancer. These studies reinforce the benefit of AI-supported methods used in conjunction with MRI data to automate analysis and exploit the full potential of MRS in both radiology and oncology.

Recommendations and Outlook

MRS has high potential for added value in the clinical setting, complementing existing medical imaging modalities. The ability of MRS to quantify metabolic processes is appropriately suited for applications in cancer given the well-characterized metabolic reprogramming in tumors; however, MRS is largely still considered an investigational technique in most U.S. health care settings. Outstanding challenges that must be addressed before routine clinical integration are discussed below.

Standardized Acquisition, Analysis, and Reporting Protocols

A major challenge in MRS is the lack of streamlined protocols on clinical MRI scanners. Measurement variability depends on many factors, including vendor, hardware capabilities, acquisition parameters, and operator, and harmonization is an important and immediate goal (73). To address these concerns, recommendations from recent MRS consensus efforts have been published (5), and AI harmonization approaches are being developed (73). As routine clinical protocols for MRS do not exist for all organs and cancer types, the widespread use of standardized MRS protocols and recommendations is essential for facilitating large-scale evaluation of the utility and applicability of MRS for specific use cases. For example, the current recommended sequence for single-voxel MRS is the sLASER sequence (5), with further consensus recommendations regarding analysis and quantification (74). Guidelines for B₀ shimming and quality control are detailed for both SVS and MRSI (75). Importantly, the use of minimum reporting standards for MRS studies (ie, MRSinMRS) is crucial for reproducibility and reliability of MRS results across the community, particularly between sites (76). Finally, consensus recommendations for clinical MRS of the brain (14,77,78) and increased use of MRS for both prostate (21) and breast (20) cancer represent important progress toward routine clinical MRS in cancer.

Quantitative Metrics

MRS is an intrinsically quantitative technique; however, MRS analysis is not easily streamlined, particularly between scanner vendors. In addition, specialized expertise is required for quality control and interpretation. Quantitative baseline metrics for healthy metabolite ranges are not readily available to the radiologist, despite extensive studies in healthy individuals. For example, while a cutoff such as 30% cerebral blood flow for infarcted tissue exists for perfusion imaging, the same metrics are not available for MRS, making interpretation of resulting spectra and metabolite concentrations challenging. Recent efforts such as the Quantitative Imaging Network, established by the National Cancer Institute to validate quantitative imaging tools in clinical oncology, represent key progress in this area. Several studies led by Quantitative Imaging Network investigators are focused specifically on the use of MRS in treatment planning for both brain and prostate cancers (79). Furthermore, toward routine and simplified clinical MRS analysis, a new web-based software package, BrainSpec (https://brainspecmed.com), was developed to postprocess, quantify, and interpret MRS data. This software, which achieved Food and Drug Administration breakthrough device status in 2022 for the quantification of 2HG, enhances the use of MRS in conjunction with other MRI modalities to facilitate a virtual biopsy (ie, noninvasive method[s] to characterize tissue composition for clinical diagnosis). Last, emerging methods such as radiomics, or metrics combining radiologic imaging with genomics and metabolomics, may further support the development of quantitative MRS-based metrics.

Reimbursement

A key barrier to the use of MRS in the diagnosis and treatment of cancer, among other diseases, is the lack of widespread insurance reimbursement. In the United States, the Centers for Medicare and Medicaid Services have determined as of 2021, no national coverage determination is appropriate for MRS. This presents a substantial financial burden for patients, as out-of-pocket costs for advanced imaging can be very high (80); thus, providers may be less likely to integrate MRS into standard of care. While some providers currently reimburse for clinical MRS scans, this is not routine. Insurance coverage for MRS is currently in the process of review, which may enable expanded clinical integration.

Horizon of Routine Clinical MRS

Due to the lack of routine clinical use, a coordinated effort between scientists, clinicians, and MRI vendors is necessary to overcome the current challenges hindering the widespread use of MRS and design protocols and methodologies for routine clinical use. With few exceptions, specific recommendations for MRS in most organs and cancer types are outstanding. Multisite clinical trials evaluating MRS for specific applications in oncology are an important aspect to progress in this area. For example, several ongoing clinical trials are evaluating the suitability of MRSI-guided radiation therapy for newly diagnosed glioblastomas (ClinicalTrials.gov identifier NCT03137888) and pediatric high-grade gliomas (NCT04908709), providing an avenue for both integration of MRSI into routine MRI protocols and evaluation across multiple sites. The eventual incorporation of these protocols into, for example, reporting and data systems such as those endorsed by the American College of Radiology, will be necessary. The current state of existing and emerging MRS technologies are represented in Figure 8. Importantly, all emerging acquisition methods outlined in this review have been, at minimum, demonstrated in healthy volunteers. Several (eg, HP ¹³C MRI) are being evaluated in clinical trials for applications in radiation oncology, and some are currently implemented more routinely at select clinical sites (eg, ¹H MRS).

Figure 8:

Diagram shows conventional and emerging MR spectroscopy (MRS) technologies and their current stage in clinical translation. ¹H SVS = proton single-voxel spectroscopy, ¹H MRSI = proton MR spectroscopic imaging, ³¹P MRS = phosphorus 31 MRS, HP ¹³C MRI/MRS = hyperpolarized carbon 13 MRI/MRS, CEST = chemical exchange saturation transfer, DMI = deuterium metabolic imaging, DW-MRS = diffusion-weighted MRS.

Conclusion

In summary, the recent efforts toward consensus in MRS implementation, improved hardware and MRI systems, advanced methods for expanding the metabolic imaging toolbox, and the integration of AI-based methods for analysis and automation support a positive outlook for expanded clinical integration of MRS. Streamlined and automated analysis (from pre- to postprocessing and stratification), standardized clinical workflows, and quantitative metrics for interpretation are immediate and necessary steps. As MRS is primed to provide crucial metabolic data, the outlook is promising for routine use of MRS across all stages of clinical cancer management.