The chemical exchange saturation transfer (CEST) effect observed in brain tissue in vivo at the frequency offset 3.5 ppm downfield of water was assigned to amide protons of the protein backbone. Obeying a base-catalyzed exchange process such an amide-CEST effect would correlate with intracellular pH and protein concentration, correlations that are highly interesting for cancer diagnosis. However, recent experiments suggested that, besides the known aliphatic relayed-nuclear Overhauser effect (rNOE) upfield of water, an additional downfield rNOE is apparent in vivo resonating as well around +3.5 ppm. In this study, we present further evidence for the underlying downfield-rNOE signal, and we propose a first method that suppresses the downfield-rNOE contribution to the amide-CEST contrast. Thus, an isolated amide-CEST effect depending mainly on amide proton concentration and pH is generated.
The isolation of the exchange mediated amide proton effect was investigated in protein model-solutions and tissue lysates and successfully applied to in vivo CEST images of 11 glioblastoma patients.
Comparison with gadolinium contrast enhancing longitudinal relaxation time–weighted images revealed that the downfield-rNOE-suppressed amide-CEST contrast forms a unique contrast that delineates tumor regions and show remarkable overlap with the gadolinium contrast enhancement.
The power of chemical exchange saturation transfer (CEST) MR imaging as a marker for tumor diseases is the indirect detection of intracellular compounds like functional metabolites [1-4] or mobile proteins [5, 6] in vivo by use of the exchange processes with water pool protons . Two distinct saturation transfer (ST) effects apparent in vivo are attributed to protons of mobile proteins: at 3.5 ppm the backbone amide signals with their base catalyzed proton transfer (APT) and at -3.5 ppm the nuclear Overhauser enhancement (NOE) mediated aliphatic proton magnetization transfer (so called exchange-relayed NOE or relayed-NOE (rNOE) ST) [8, 9]. For these protein ST effects, several interesting correlations have been shown that might play a role in vivo and especially in pathologies: dependence on intracellular pH [5, 10-13], protein concentration [8, 13], or protein folding [14, 15] and aggregation states . Already, the use for brain tumor detection [6, 17-20], grading , and possible differentiation of tumor recurrence and radiation necrosis  has been shown to be feasible by protein-based saturation transfer MRI.
In a previous study, we showed that Lorentzian fit analysis optimized for low rf irradiation amplitude (low-B1) CEST spectra at 7 Tesla (T)  allowed for separation of the aliphatic rNOE-ST signal at -3.5 ppm and the amide-CEST signal at 3.5 ppm . Furthermore, we demonstrated that influences of changes in relaxation are strong, but can be corrected using the relaxation-compensated metric that yields the apparent exchange-dependent relaxation rate (AREX) [11, 24].
However, even after correction of all relaxation influences, the identification of the CEST effect at +3.5 ppm with amide protons is still not accurate: “under” the amide peak exists an additional rNOE contribution originating most likely from aromatic protons of covalent bound C-Hn groups which resonate around +3.5 ppm, but also from nonexchanging amide protons within hydrogen bounds [25, 26]. A first hint to these contribution is given in MRS . Furthermore, a study of Jin and Kim showed a clear downfield rNOE in the protein bovine serum albumin (BSA) by using deuterium to quench exchange in vitro, as well as the existence of a large residual CEST signal at +3.5 ppm even in stroke regions with low pH an thus low amide CEST contribution . This residual signal was also observed in ref. , using the relaxation compensated AREX approach. A third evidence for these downfield rNOE-ST effects of proteins is the behavior of the ST resonance upon unfolding: in a previous study of Goerke et al.  it was shown that rNOE signals quench under protein unfolding, whereas exchanging signals increase. Moreover, it was shown that the selective peak at +3.5 ppm actually increases, but the background signal decreases and thus shows a clear rNOE like behavior. Experiments presented in this article performed in protein solutions and tissue homogenates at different pH show the existence and relative size of this downfield NOE contribution to the CEST effect at +3.5 ppm.
Consequently, an ST effect at +3.5 ppm with suppressed downfield rNOEs should yield a more pure amide proton CEST effect with correlations to protein concentration and pH.
However, separation of two contributions at the same resonance frequency is not a simple task. Still, the similar behavior of downfield rNOE and upfield aliphatic rNOE in proteins upon protein unfolding and pH changes  encouraged us to a very first approach: to use the easier accessible aliphatic rNOE effect as an estimator for the downfield NOE contribution.
This implicitly assumes a ratio between these two rNOE effects that remains relatively constant within each tissue type and is relatively independent of pH. Namely, we assume a rNOE-ratio rrNOE such that rNOEdownfield = rrNOE ·rNOEupfield. Validity of this assumption was tested in tissue homogenate experiments and is presented herein.
With such a ratio, which was obtained by tissue homogenate experiments, the relaxation-compensated and downfield-NOE-suppressed CEST effect at +3.5 ppm, called dnsAREX(+3.5 ppm) here, can be defined as
In this study, we want to share the following discovery: If this calculus is applied to CEST data obtained in brain tumor patients, a unique contrast is generated that shows structures in the tumor area similar to the gadolinium contrast-enhanced T1-weighted image contrast. The downfield-NOE-suppressed AREX evaluation was prospectively applied on spectrally highly resolved 7T CEST data of one histologically proven glioblastoma patient before surgery. Retrospectively, we applied the method to the whole patient collective of 10 glioblastoma patients of our previous relaxation-compensated CEST study (24). A comparison of the dnsAREX contrast with contrast enhanced T1-weighted imaging and other conventional MR contrast is performed. Details of this calculus and the underlying relaxation-compensation are discussed.
BSA model solutions were prepared with a concentration of 2.5% (w/v) in phosphate buffered saline (PBS) (1/15 M) at different pH values at 25 °C.
The tissue homogenates were made from white matter of pigs brain tissue was chilled on ice till sample preparation; before homogenization it was diluted by 1:2 with distilled water. To homogenize the tissue, a 7 mL Wheaton™ Dounce tissue grinder was used first with a loose pestle (gap approximately 112 µm) followed by a tight pestle (gap approximately 50 µm). The homogenate was then diluted by PBS buffer to yield the final sample with 1/3 tissue fraction and 2/3 PBS fraction with different pH. The pH of the final solution was then measured again using a pH meter. To reduce T1 and thus measurement time, Magnevist™ (20 µM Gadopentetate Dimeglumine) was added to each sample to set up a Gadopentetate Dimeglumine concentration of 20 µM.
Eleven patients with newly diagnosed and histologically proven glioblastoma were enrolled in this study and were examined before surgery. Approval of the local ethics committee after written informed consent was obtained. Data of 10 patients were acquired in a previous study  and retrospectively evaluated with the new method. One additional female patient was examined prospectively. Additional patient data can be found in the Table S1.
CEST Spectroscopy at 14.1T
A Bruker Avance II spectrometer (AV-600 from 2001) manufactured by Bruker (Bruker BioSpin GmbH, Karlsruhe-Rheinstetten, Germany) was used for high-resolution CEST spectroscopy. The system operates at a field strength of B0 = 14.1T, corresponding to a resonance frequency of 600.13 MHz for 1H nuclei. A 5-mm and 8-mm probe were used for rf irradiation and signal acquisition. The temperature of the samples was set with an internal heating and cooling device to 25 °C or 37 °C.
Z-spectra were obtained after saturation defined by the mean saturation RF amplitude B1, and the saturation duration tsat = 12 s. In the case of cw CEST one rectangular pulse and in the case of pulsed CEST Gaussian-shaped pulses of duration tp = 15 ms and DC = 60% were used for presaturation. Integration of the water resonance in the range of ±0.7 ppm yielded the signal M(Δω). Before repetition of the pulse sequence a waiting interval of 1 s was implemented. For normalization, a fully relaxed signal (M0) without saturation was acquired. T1-weighted spectra were obtained by a saturation-recovery pulse sequence using a cw, rectangular, low-RF-power pulse of length 15 s to saturate the water signal. The integrated water resonances in the range of ±0.7 ppm were fitted for 15 different recovery times TI between 10 ms and 20 s to determine the R1W of water.
Water-exchange (WEX) spectroscopy enables the detection of magnetization transfer pathways between protons in solutes and water [29, 30]. The WEX signals built up during an intermediate time interval called the mixing time (TM). Fitting of the WEX signal dynamics as a function of TM (assuming ksw > kws, R1s, R1w) :
enables determination of ksw. The WEX II sequence was implemented on the 14.1T spectrometer as described by Goerke et al .
Conventional MRI at 3T
Conventional MR clinical exams were acquired at 3T with the following protocol parameters: T1-weighted gadolinium contrast-enhanced MRI (gdce-T1) [echo time (TE) = 4.04 ms; repetition time (TR) = 1710 ms; field of view (FOV) 256 × 256; resolution, 512 × 512; slice thickness, 1 mm], T2-fluid attenuated inversion recovery (FLAIR) (TE = 135 ms; TR = 8500 ms; FOV 230 × 172; resolution, 256 × 192; slice thickness, 5 mm), T2-weighted MRI [turbo spin echo (TSE), TE = 86 ms; TR = 5550 ms; FOV, 229 × 172; resolution, 384 × 230; slice thickness, 5], and diffusion imaging yielding the apparent diffusion coefficient (ADC) (echo planar readout, TE = 90 ms; TR = 5300 ms; b = 0 mm2/s, and b = 1200 mm2/s, FOV, 229 × 229; resolution, 130 × 130; slice thickness, 5 mm). All images were co-registered with the gdce-T1 using an automatic multi modal rigid registration algorithm in MITK .
Details of the correlation analysis of CEST with the 3T data are given in the Figures S2–S6.
MRI at 7T
Measurements were performed on a 7T (297.2 MHz) whole body MRI scanner (MAGNETOM 7T; Siemens Healthcare, Erlangen, Germany) using a custom-developed CEST sequence based on two-dimensional (2D) centric-reordered gradient echo (GRE) and a single-channel transmit (Tx)/24-channel receive (Rx) 1H head coil. In vivo CEST images were obtained after saturation by a train of 150 Gaussian-shaped radiofrequency (RF) pulses with (tp = 15 ms, td = 10 ms, Duty-Cycle (DC) =60 %, tsat = 3.75 s) for 2 B1 amplitudes, 0.6 µT and 0.9 µT. The B1 amplitude refers here to the average amplitude of the pulse or B1 = flip angle/(γ⋅tp). Thus, the average amplitude of the used pulse train is DC⋅B1, the average power  of the used pulse train is 1.2⋅DC⋅B1. Frequency offsets were distributed with higher sampling around the CEST pools: from ± 4 ppm to ± 3 ppm in steps of 0.1 ppm, from ± 2.75 ppm to ± 2 ppm in steps of 0.25 ppm, ± 1.8 ppm to ± 1.2 ppm in steps of 0.1 ppm, ± 0.5 ppm, ± 0.25 ppm, and 0 ppm. Semi-solid magnetization transfer (MT) was sampled using the following offsets in ppm: ± 300, ± 100, ± 50, ± 20, ± 10, ± 8, ± 5, and ± 4.25. An additional normalization image M0 was acquired after saturation at -300 ppm, after 12 s waiting time for complete relaxation. The imaging parameters of the single-slice readout were as follows: TE = 3.76 ms; TR = 7.6 ms; field of view, 200 × 175; matrix, 128 × 112; slice thickness 5 mm, in-plane resolution 1.56 × 1.56 mm2; flip angle, 10 °. The CEST images were manually co-registered with gdce-T1.
T1 mapping was achieved by fitting T1-weighted images of a saturation recovery GRE (single-shot, TE = 1.36 ms; TR = 7.6 ms; readout was followed by a relaxation time of 5500 ms; FOV, resolution, and slice thickness equal to CEST) with 10 equally spaced recovery times from 0.5 s to 5 s using MATLAB 2014b.
B0 and B1 Correction
For the B0 correction, water saturation shift reference [WASSR ] images were collected from -1.0 to 1.0 ppm with a step size of 0.1 ppm, a saturation pulse train of 1 Gaussian-shaped RF pulse with tp = 25 ms, DC = 50% and B1 = 0.2 µT using the same sequence as used for CEST imaging and identical readout parameters except for an additional waiting time of 3 s before each preparation pulse. To correct for B1 inhomogeneities, Z-spectra at different reference amplitudes were acquired. B1 correction was performed using the “contrast-correction” method as described by Windschuh et al  yielding corrected images with CEST contrasts corresponding to 0.6 µT saturation amplitude. The required B1 mapping was realized using a magnetization-prepared GRE sequence (TE = 2.42 ms; TR = 5 s; 5 shots; field of view, 180 mm2; matrix, 128 × 128; in-plane resolution, 1.4 × 1.4 × 5 mm3; flip angle, 8 °).
CEST scans at two B1 amplitudes required 4:07 min each, the WASSR scan 1:14 min, the B1 map 10 s and the T1 map 1:20 min. Total measurement time for B1-corrected and relaxation-compensated 2D-CEST is, therefore, 10:45 min.
CEST Effect Separation and Relaxation Compensation
Z-spectra were fitted pixel-wise by a multi-Lorentzian fit described in Windschuh et al , which yields a label (Zlab) and a reference Z-spectrum (Zref) for each CEST effect. So for amides the reference is given by
where Lamide is the Lorentzian dedicated to the amide peak at 3.5 ppm. Based on these reference values, the following CEST contrasts were generated for amides, amines and upfield rNOE:
The spillover and MT corrected inverse difference :
The spillover, MT and T1 corrected AREX contrast [7, 11, 37, 38], which is also referred to as relaxation-compensated CEST contrast in the document:
The conventional asymmetry analysis is also performed (based on nonfitted data):
Removal of Downfield rNOE Contributions to the +3.5 ppm Resonance Using the Aliphatic rNOE
We assume that downfield and upfield aliphatic rNOE have a specific ratio, which we call the rNOE-ratio here: rrNOE = AREXdownfield/AREXupfield. Consequently, we define an downfield-NOE-suppressed AREX signal (dnsAREX) at the 3.5 ppm resonance by
By measurement of the rNOE-ratio in brain tissue lysates at low pH we obtained rrNOE of approximately 0.2. This value was then used for the evaluation of the patient data. As the rNOE-CEST contribution in the Z-spectrum might be B1 dependent, also rrNOE might change with B1, thus we used the same saturation scheme in vitro as in vivo, and want to point out that our derived value is only valid under this saturation condition.
For calculation of region of interest (ROI) -averaged dnsAREX-spectra, first a reference Z-value was generated pixel-wise by using the Lorentzian fit of direct water saturation and the semi-solid MT yielding Zref(Δω). Then AREX was calculated for each offset using the measured Z-spectrum in each pixel Z(Δω): Subsequently, the dnsAREX-spectrum could be calculated using
Figure 1 shows the WEX data of BSA at different pH. Independent of pH there is a contribution of downfield protons that are associated with aromatic protons or hydrogen bound amide protons. The rates of these exchange processes can also be estimated by WEX, it turns out that the downfield rNOE transfer rate is 3.4 Hz, hence approximately 1.6 time larger than the aliphatic rNOE transfer rate of 2.1 Hz. Again these rates, and thus also their ratio, do not change significantly with changes in the pH. This means that the rNOE-ratio that we want to derive in CEST is not only a proton number ratio, but a combined effect of the relative proton fraction f and the effective exchange rate k.
Determination of the rNOE-ratio
To determine the factor rrNOE that is most similar to human brain in vivo, pig brain homogenates were used (Fig. 2). The homogenized state allows for changing pH (Fig. 2a) and thus to extrapolate for the downfield NOE contribution (Fig. 2b). Assuming the amide proton transfer to be small at pH = 5.85, this yields an upper estimation of the rNOE-ratio of AREX(+3.5 ppm; pH = 5.85)/ AREX(-3.5 ppm; pH = 5.85) ≈ 0.16. If we assume an exponential increase with pH (Fig. 2b) and fit the 3.6 ppm resonance by a⋅(eb⋅pH-1) + c, the constant c becomes 0.004 and a lower rrNOE estimation c/AREX(-3.6 ppm) ≈ 0.1. As rNOE has also a slight pH dependency, and also amines might affect the signal at lower pH, we cannot determine the rrNOE ratio very precisely, but we can estimate that it should be approximately between 0.1 and 0.2 for the given saturation parameters.
To verify the NOE-ratio determination at 14.1T using different pH is feasible and also translatable to 7T, three control experiments were performed to check (i) if CEST effects, especially the rNOE saturation transfer, change during the homogenization process, (ii) if the homogenate is stable during the measurement time, and (iii) if the rNOE saturation transfer effects change with the static B0 field. Figure 3 shows the results of these control experiments; Figure 3a shows that the lysis process has only very small effect on the CEST signals; loss of the structure especially seem to affect the aliphatic NOE only slightly, Figure 3b shows that the CEST effects in the homogenate are relatively stable over the measurement time, and rrNOE determined in the pH measurement (Fig. 2) might be underestimated by 10–20% due to the slight increase of aliphatic signals with time. Figure 3c shows the AREX spectrum of BSA (pH = 6.58) at different magnetic field strengths; in principle it shows that all CEST effect increase for the same bandwidth of saturation pulses, which is most probably due to reduced spectral resolution at lower field strengths and thus a labeling of more proton species at each frequency offset. Still, the NOE-ratio seems to be only slightly affected by B0: if we use AREX(4 pmm)/AREX(-4 ppm) as an estimation we get a slightly higher value of rrNOE at 7T (rrNOE = 0.55) compared with 9.4T (rrNOE = 0.54) and 14.1T (rrNOE = 0.48). Again this means that the NOE-ratio might be 10–20% higher than estimated at 14.1T. Thus the tissue homogenate yields at least a good guess of the rrNOE of white matter (WM) in vivo, however, the derived value has some uncertainty, approximately rrNOE = 0.2 ± 0.1 and for the in vivo approach the influence of these variations must be investigated (see Figure S7). We want to point out, that this value is just valid for the given saturation parameters, and might vary with B1 amplitude or saturation time.
Transfer to In Vivo Data
The downfield-NOE suppressed evaluation was performed in vivo with rrNOE = 0.2 in Eqs. (6) and (7) (other values were tested in the Supporting Informations). Figure 4 shows that this also removes most of the downfield NOE in this case: Figures 4ab show the AREX- and dnsAREX-spectra according to Eq. (7) of brain tissue of a gadolinium contrast-enhancing (gdce) ROI and a normal appearing WM (nWM) ROI. AREX signals at +3.5 ppm show no significant difference between gdce and nWM. However, after suppression of the downfield rNOE contribution estimated by the aliphatic rNOE using dnsAREX, the baseline seems to be partially removed and the peak at 3.5 ppm shows to be increased in the tumor tissue compared with nWM.
Figure 5 shows multi-parametric MR images of a glioblastoma patient. The tumor affected area shows a changed signal in all imaging contrasts. In contrast to MTRasym, the relaxation-compensated AREX(3.5 ppm) shows a hyperintense region within the tumor area similar to the gdce-T1 ring enhancement and areas of restricted diffusion on ADC. However, this hyper-intensity is not significant compared with contralateral tissue (Fig. 6). In contrast, the downfield-rNOE suppressed amide CEST contrast shows this feature clearly (Fig. 5h) and also significantly compared with nWM (Fig. 6a). Interestingly, the “ring enhancement” in dnsAREX displays more pronounced on the medial tumor border. In this region, gdce-T1 only shows a diffuse enhancement, while ADC maps display a pronounced ADC decrease. This finding might potentially be interpreted as tumor infiltration in that region. In contrast to MTRasym all relaxation-compensated contrasts show lower values in the necrotic area compared with the gdce region. In agreement with our previous study, the AREX(+3.5 ppm) is not significantly changed whereas the aliphatic rNOE-CEST effect AREX(-3.5 ppm) might be a helpful marker as it shows significantly lower signals in the whole tumor affected region compared with nWM and normal appearing gray brain matter (nGM).
Figure 7 shows the results of a retrospective application of dnsAREX on data of 5 patients of a previously published patient collective of histologically proven glioblastoma patients where the relaxation-compensated AREX was performed. (Find images of the residual five patients in the Figure S1). As illustrated dnsAREX show structures with a remarkable correlation with the gdce-T1 ring-enhancement in some of the patients. Of interest, some hyper-intensities in dnsAREX go in line with a diffuse gadolinium (Gd) enhancement. This is also the case in the patient shown in Figure 5. To check for stability of these features against changes in the used rNOE-ratio, the evaluation was repeated with different ratios (see Figure S7) with the result that changes from rrNOE = 0.1 to 0.3 do not affect the principal features of the dnsAREX contrast.
Comparison with ADC maps reveals that hyper-intensities in dnsAREX sometimes correlate with decreased ADC values (pink arrows in Figure 8). In addition, dnsAREX is more intense in regions were the gdce is diffuse. Both findings hint toward possible identification of beginning infiltration and malignization of the tumor. A pixel-wise quantitative correlation over the whole collective of the contrasts dnsAREX as well as AREX(+3.5 ppm) with ADC and gdce-T1 maps was performed within a larger ROI containing the tumor area (Figure 9, details can be found in the Figures S2–S6). It turns out that dnsAREX has only a weak negative correlation with ADC (Fig. 9c). Also AREX shows a negative correlation with ADC (Fig. 9a). However, while AREX shows only a very weak correlation with gdce-T1 (Fig. 9b), dnsAREX (Fig. 9d) shows a stronger positive correlation with gdce-T1 compared with AREX. The removal of the downfield-NOE by dnsAREX decreases the correlation with ADC and increases the correlation with gdce-T1.
In this study, we introduced a relaxation-compensated and downfield-rNOE-suppressed amide CEST evaluation method called dnsAREX(3.5 ppm) which was applied prospectively in one glioblastoma patient and retrospectively applied to relaxation-compensated CEST measurements of 10 glioblastoma patients reported in a previous study .
Our primary finding is that the downfield-rNOE-suppressed amide CEST evaluation increases the contrast between normal appearing white matter and tumor areas, and in some cases the observed structures show remarkable similarity with the gadolinium ring enhancement. While AREX(-3.5 ppm) drops in the whole tumor area and AREX(+3.5 ppm) is not significantly changed between normal and gdce regions, the downfield-rNOE-suppressed AREX(3.5 ppm) delineates the ring enhancement with significantly higher intensities than contralateral WM.
In a previous study it was stated : “…the peak at + 3.5 ppm, dubbed “amide” CEST here, has an amide contribution and a rNOE contribution. We think that the aromatic rNOE part is decreasing similar to the aliphatic rNOE, but the exchanging part of the +3.5 ppm peak is actually increasing in the tumor. This results in the almost unchanged AREXamide signal.” We want to note that at +3.5 ppm in addition to aromatic protons also amide protons in hydrogen bounds might contribute to a pH independent saturation transfer effect [25, 26], still, the current study gives more evidence for this mixed signal hypothesis, as the increasing amide part could be isolated by suppression of the downfield rNOE using the aliphatic rNOE (Eq. (6)). However, there are still some limitations of this interpretation:
From a physicists point of view, a limitation of our approach is the assumption of the constant rNOE-ratio. First of all, this ratio might be different in GM and WM which might explain the higher signals of dnsAREX in GM (Fig. 2); and thus it could also change in pathologies. Especially, if the two rNOE signals have different origins. Still we were able to show by our in vitro experiments that within one tissue type this ratio is relatively constant upon macroscopic changes of the tissue. In this context we also want to stress the nonspecificity of CEST: the observed rNOE ST effects origin from all types of mobile proteins, thus selective changes will have only a small effect on the ratio, whereas global changes in the proteome will have similar affect to the aromatic and aliphatic rNOE effects as they both appear in all types of proteins. We think it is relatively improbable that downfield and upfield rNOEs originate from complete disjunctive mobile protein types. Thus, the suppression by simple linear subtraction of a fraction of the aliphatic rNOE is not the end of the story of downfield rNOE suppression, but must be understood as a first order attempt that yields a remarkable contrast. In the end we could show that the principle features of the dnsAREX contrast do not change upon intermediate changes (in the range of ±0.1) in the rNOE-ratio (Figure S7).
Moreover, we want to reference to more sophisticated methods that can be used to filter the downfield rNOE without this assumption, e.g., CERT [39, 40] or VDPM-CEST , as well as transfer rate edited-CEST  that all have the power of exchange rate or T2 filtering. As rNOE-CEST effects have lower exchange rates and shorter T2 than amides, one could filter them without making use of the aliphatic rNOE, similar to previous efforts in WEX spectroscopy .
These approaches are also promising to solve the issue that the hyper-intensities of dnsAREX are significant compared with nWM, but not always significant compared with nGM signals (see Figure 7). A contrast displaying exclusively tumor areas hyper-intensely would be desirable.
The presented tissue homogenates form an interesting model solution, close to the in vivo situation, but with direct access to intracellular parameters such as pH or relaxation parameters and many more which were not yet investigated herein. On the other hand, BSA as used in Figure 1 might be a good exemplary protein showing CEST effects, however, it may not be a good reference sample for APT simulation as the amide contribution is rather small; a model closer to in vivo APT effects might be egg white as previously used [40, 44, 45]. In the model solutions we also added gadolinium based contrast agents and the question arose if these are changing the rNOE processes directly. A phantom experiment of BSA with different gadolinium concentrations (Figure S8) revealed that AREX compensates the T1 change for both, chemical exchange and rNOE effects and direct changes of rNOE are negligibly small.
dnsAREX and MTRasym
After all the efforts to separate the rNOE and ‘amide’ peaks from the background as described in our previous study, the mixing of signals of -3.5 ppm and +3.5 ppm introduced by Eq. (6) again reminds of amide proton transfer (APT) imaging using MTRasym. One might speculate that MTRasym/T1 should yield the same contrast: Left and right side are subtracted and the scaling by T1 can also be removed by the division. However, this is not the case as aliphatic rNOEs outweigh the amide contribution at this low B1 level, visible in Figures 4 and 5 and thus MTRasym/T1 (Fig. 10a) looks still most similar to MTRasym (Fig. 5e). Of interest, in this patient, the ring enhancement is also depicted in the T1 map (Fig. 10b) and in the linear difference metric MTRLD (Fig. 10c). However, in contrast to the outcome of dnsAREX, this was not the case for all patients . Even though MTRLD might be influenced by relaxation, it should stay on the list of evaluated contrasts as it sometimes shows these features.
The uncorrected MTRLD(-3.5 ppm) (Fig. 10d) shows a similar contrast as AREX(+3.5 ppm) (Fig. 5f). Concerning R1 correction Li et al  came up with a nice verification of AREX by injection of gadolinium contrast agents in a tumor with blood–brain barrier (BBB) breakdown and subsequently measuring CEST; it turned out that AREX remains stable upon induced R1 changes.
Altogether, our findings might be an explanation why MTRasym-based APT, which is normally acquired using higher saturation power, yields a good contrast in tumors: (i) As stated by Xu et al  upfield and downfield rNOEs might be different for low power, but melt together to one broad peak, due to more efficient labeling of short T2 components. (ii) At higher power, also the exchange-mediated contribution might still increase, as exchange rates of amides (average approximately 30 Hz as measured in brain tissue) are a factor 5–10 higher than rNOE rates (average approximately 2–7 Hz), but also the amine contribution rises. Thus it is plausible that, if well optimized, MTRasym at higher power intrinsically compensates for the downfield rNOE contribution.
This might be interesting especially for the application at clinical field strengths, as at 3T the isolation of different effects in the Z-spectrum is difficult. It should be further investigated, if using the above mentioned filtering techniques yields a robust and downfield-NOE-suppressed amide-CEST contrast also at 3T based on the simpler MTRasym approach.
Origin of Contrast
Even if the downfield rNOE contribution is suppressed, it is still not perfectly clear where the remaining contrast originates from. As for previous amide-CEST studies, it could be explained by an increased intracellular pH [5, 17] or increased protein content . On the one hand, Yan et al  showed the intracellular protein content can be increased in tumors, on the other hand, the shown correlations with ADC (Fig. 9) could hint to increased protein content due to increased cellularity . Still, as we found this strong correlation with the gdce, it could also be a direct influence of the BBB breakdown, like a CEST effect of extracellular proteins or of blood compartments  or increased temperature or changed pH due to this influence. However, the observed heterogeneity speaks more for an origin not directly linked to BBB breakdown.
There are more and more efforts to bring CEST into the clinical decision process for brain cancer diagnosis [17, 21, 50-52]. A CEST contrast with the features presented in this work, is of special interest as reliable identification of infiltrative tumor progression that is not necessarily accompanied by enhancement on T1-weighted images is still a challenge faced by neuro-oncologic imaging [53, 54]. In addition, a contrast agent free technique with access to the presented insight would simplify clinical routine not only because of avoidance of allergic reactions and general toxicity of Gd, but also by omitting accumulation of Gd in patients . Even if the contrast structures of gdce could be also identified by CEST, there might still be more information on a cellular level accessible by CEST. Especially, the reported differences of increased downfield-rNOE-suppressed amide CEST effects and the contrast-enhancing areas encourage further investigations to determine the potential of downfield-rNOE-suppressed and relaxation-compensated CEST to identify areas of tumor infiltration. Finally, a contrast that correlates with pH might provide valuable additional information on the tumor microenvironment and progression of disease [56-59].
In the presented work, it was shown that at the amide proton frequency 3.5 ppm as well rNOE-CEST effects are apparent when using low-B1 saturation where slow transfer rates dominate. Comprehensive in vitro experiments showed that there is a constant ratio between downfield and upfield rNOE that is in first order independent on pH. This allowed for establishing a method that suppresses the downfield rNOE contribution to the amide-CEST effect by using the upfield aliphatic rNOE-CEST signal. This downfield-rNOE-suppressed amide-CEST method was verified in tissue homogenate experiments and applied to spectrally highly resolved relaxation-compensated CEST data of 11 glioblastoma patients. Precise co-registration and image fusions with contrast-enhanced imaging revealed that the downfield-rNOE-suppressed amide-CEST contrast forms a unique contrast that delineates tumor regions and shows remarkable overlap with the gadolinium contrast enhancement. These additional features were not visible in uncorrected CEST images or other contrast-agent-free conventional MRI. Thus, the presented correction forms an important step toward the application of protein-CEST at 7 Tesla in vivo.
We want to thank Dr. Manfred Jugold and Dr. Karel Klika for their support with the protein solution and tissue homogenate NMR measurements.