Tissue response curve‐shape analysis of dynamic glucose‐enhanced and dynamic contrast‐enhanced magnetic resonance imaging in patients with brain tumor

Dynamic glucose‐enhanced (DGE) MRI is used to study the signal intensity time course (tissue response curve) after D‐glucose injection. D‐glucose has potential as a biodegradable alternative or complement to gadolinium‐based contrast agents, with DGE being comparable with dynamic contrast‐enhanced (DCE) MRI. However, the tissue uptake kinetics as well as the detection methods of DGE differ from DCE MRI, and it is relevant to compare these techniques in terms of spatiotemporal enhancement patterns. This study aims to develop a DGE analysis method based on tissue response curve shapes, and to investigate whether DGE MRI provides similar or complementary information to DCE MRI. Eleven patients with suspected gliomas were studied. Tissue response curves were measured for DGE and DCE MRI at 7 T and the area under the curve (AUC) was assessed. Seven types of response curve shapes were postulated and subsequently identified by deep learning to create color‐coded “curve maps” showing the spatial distribution of different curve types. DGE AUC values were significantly higher in lesions than in normal tissue (p < 0.007). Furthermore, the distribution of curve types differed between lesions and normal tissue for both DGE and DCE. The DGE and DCE response curves in a 6‐min postinjection time interval were classified as the same curve type in 20% of the lesion voxels, which increased to 29% when a 12‐min DGE time interval was considered. While both DGE and DCE tissue response curve‐shape analysis enabled differentiation of lesions from normal brain tissue in humans, their enhancements were neither temporally identical nor confined entirely to the same regions. Curve maps can provide accessible and intuitive information about the shape of DGE response curves, which is expected to be useful in the continued work towards the interpretation of DGE uptake curves in terms of D‐glucose delivery, transport, and metabolism.

interval were classified as the same curve type in 20% of the lesion voxels, which increased to 29% when a 12-min DGE time interval was considered. While both DGE and DCE tissue response curve-shape analysis enabled differentiation of lesions from normal brain tissue in humans, their enhancements were neither temporally identical nor confined entirely to the same regions. Curve maps can provide accessible and intuitive information about the shape of DGE response curves, which is expected to be useful in the continued work towards the interpretation of DGE uptake curves in terms of D-glucose delivery, transport, and metabolism.
K E Y W O R D S brain tumor, curve-shape analysis, D-glucose, dynamic contrast-enhanced MRI, dynamic glucose-enhanced MRI, glucoCEST

| INTRODUCTION
Chemical exchange-based detection relates to an emerging group of MRI approaches that exploit naturally occurring chemical exchange of protons between solute molecules and water. The technique enables the use of biodegradable compounds as MRI contrast agents. Dynamic glucose-enhanced (DGE) 1-6 MRI targets the exchangeable hydroxyl (OH) protons in glucose and employs dynamic image acquisition before, during, and after intravenous administration of D-glucose. DGE MRI can be based on chemical exchange saturation transfer (glucoCEST), 7-9 chemical exchange sensitive spin-lock (glucoCESL), 10,11 or T 2 relaxivity of glucose. 12,13 In addition to its biodegradability, the advantages of D-glucose include low cost and high accessibility, as well as the opportunity to obtain unique tissue information that is unavailable with other quantitative MRI techniques, for example, D-glucose delivery, uptake, and even metabolism. In resemblance to dynamic contrast-enhanced (DCE) MRI, DGE MRI can yield response curves in a tissue of interest, 1,14-16 but without the need for gadolinium injection. DGE and DCE MRI have previously been compared in mice and rats, 2,5 highlighting the potential to use DGE MRI for blood volume and/or vascular permeability imaging and glucose uptake in tumors.
Because of the growing interest in DGE imaging and its potential for perfusion-related or metabolic assessments, it is of interest to compare the time courses (tissue response curves) of DGE and DCE, and to investigate whether DGE MRI can provide information that is similar or complementary to DCE MRI. It should be emphasized that the tissue uptake kinetics of D-glucose differ from those of gadolinium. 1,[5][6][7] The challenge in kinetic modeling of DGE MRI is to include the intracellular uptake and metabolism of D-glucose to lactate or other metabolites. This leads to faster disappearance of the CEST signal over time [5][6][7] compared with, for instance, fluorodeoxyglucose in positron emission tomography (PET), where the phosphorylated product remains. D-glucose is transported over the blood-brain barrier (BBB) and the blood-cerebrospinal fluid (CSF) barrier and thus can cause signal enhancement in normal tissue and CSF. 2,6,17 Another challenge is the conversion from DGE signal change to concentration, which is complicated by different influencing factors such as pH, magnetization transfer effects other than CEST, as well as contributions from changes in the transverse relaxivity of D-glucose caused by OH proton exchange. 12,13 The other differences to DCE MRI are the nonzero baseline concentration of D-glucose, and the fact that the injected D-glucose amount is higher than for a true tracer and may affect the physiology. Furthermore, in current DGE protocols, a relatively large amount of D-glucose is injected slowly (over 1-4 min), making it difficult to estimate rate parameters for D-glucose uptake in tissue without local arterial input functions (AIFs) for correction. 2 In addition, the magnitude of the AIF is likely to show a different scaling factor to tracer concentration than the tissue response magnitude, because of different background signals and pH in blood versus tissue. The physiological response to glucose varies between individuals because of different insulin response and glucose metabolism, so the AIFs acquired in DGE experiments do not have a general shape and individual AIF measurements will probably be necessary. 14 Because of the low signal-to-noise ratio (SNR) of DGE, it is also difficult to distinguish the tissue signal peak from the baseline. 14 It is therefore not straightforward to apply existing DCE tracer kinetic models to DGE data. However, in a recent preclinical study performed by Dickie et al., two kinetic models for glucoCESL DGE MRI were introduced, 18 one based on an existing 13 C D-glucose uptake model 19 and another on free diffusion. Such models could in principle also be applied to glucoCEST DGE data. However, robust fitting of kinetic models requires adequate SNR, which may hamper this approach at lower field strengths.
Based on the above, a practical solution for quick and informative assessment of DGE data would be important. In previous DGE studies, [1][2][3]15,16,20 image evaluation has mainly been performed by calculating the area under the curve (AUC) of the DGE signal response, and/or by studying the response curves in regions of interest (ROIs). As a step towards an analysis method for DGE that may be clinically practical, here we introduce an approach based on response curve patterns. In DCE analysis, a concept referred to as time intensity curve (TIC) shape analysis [21][22][23] or "curve-ology" 24 has previously been introduced, based on the idea that the curve properties differ between tumor and normal tissue. DCE TIC shape analysis has mainly been used in breast imaging, 25,26 for example, to differentiate between benign and malignant breast tumors, 25 but also in prostate, 27 brain, 22 and the musculoskeletal system. 21 Nonmodel-based analysis in DCE MRI is less complex than tracer kinetic model-based analysis because there is no need for conversion from signal to concentration, choice of AIF, or assumptions regarding physiology. Instead, the tissue response curve is analyzed and described by summary parameters. The extracted parameters are not directly or entirely related to physiological properties (perfusion, leakage), but correlations between curve features and the underlying physiology do exist. 28,29 In the present study we propose a similar approach for DGE, using the type of shape as a representation of tissue physiology. In addition to conventional AUC analysis, we created color-coded curve maps showing the spatial distribution of different curve-shape types. Previous studies have successfully applied DGE and DGEρ MRI in brain tumor patients, 1,3,4,16,20,30 but the signal change is small, and it is therefore of interest to identify new postprocessing methods that can facilitate differentiation between tumor and normal tissue.
This study aimed to develop an analysis method for DGE based on tissue response curve patterns, and to apply this method to gain a deeper understanding of the DGE contrast properties, and the similarities and discrepancies between glucose and gadolinium response curves. This was accomplished by comparison of DGE and DCE curve-shape analysis of data from brain tumor patients scanned at 7 T.

| Patients and MR imaging
At Lund University Hospital, eight presurgical patients with suspected gliomas were enrolled in the study (Table 1). MRI was conducted using an actively shielded 7T scanner (Achieva, Philips, Best, The Netherlands) using a dual transmit head coil with a 32-channel phased-array receive coil (Nova Medical, Wilmington, MA, USA). The project was approved by the local ethics committee (The Regional Ethical Review Board in Lund), and written informed consent was obtained from all patients.
Three additional patients, scanned at the Johns Hopkins University (JHU), were included to increase the number of samples to 11. These patients were scanned on a passively shielded 7T Philips MRI scanner (Philips, Best, The Netherlands) using a head coil of the same model, and the same DGE and DCE protocols as for the Lund patients. DGE data from patients 9 and 10 have been published previously. 1 Exclusion criteria for DCE and DGE were sensitivity to gadolinium and D-glucose (diabetes), respectively. All patients fasted for 4-6 h prior to the examination. In Lund, the baseline venous blood glucose level was measured upon arrival using a blood gas analyzer (i-Stat, Abbot Scandinavia AB, Sweden), and a value between 3.9 and 7.5 mM (70 to 135 mg/dl) was considered normal. At JHU, the normal fasting glucose value was limited to less than 7 mM (< 126 mg/dl). Anatomical images (MPRAGE; 1 mm 3 isotropic resolution) were acquired at the start of the examination as well as after the DCE acquisition. The final diagnoses of all patients were confirmed by histopathology after surgical resection of the tumor.

| DGE imaging
To enable a reasonable temporal resolution (5.3 s), dynamic glucoCEST data were acquired at a single saturation offset of 1.2 ppm. Saturation was achieved using an equidistant train of 32 sinc-gauss pulses with peak B 1 = 1.96 μT, duration of 50 ms, and separation of 25 ms, followed by gradient-echo imaging with TR/TE/FA = 5 ms/1.48 ms/30 . A single transaxial slice with a thickness of 6 mm, field of view of 224 Â 224 mm 2 , T A B L E 1 List of included patients. Patients 1-8 were scanned in Lund and patients 9-11 were scanned in Baltimore and in-plane resolution of 2 Â 2 mm 2 was acquired dynamically. The total glucoCEST scan time was 15 min 54 s, giving 180 dynamics. Manual injection of 50 ml of D-glucose (50% dextrose, APL, clinical grade) was given intravenously 3 min into the glucoCEST scan, and the injection time was 1 min ± 12 s. A saline flush of 20 ml was given manually after completion of the D-glucose injection.

| DCE imaging
DCE imaging was performed using a 3D spoiled gradient-echo sequence with TR/TE/FA = 2.7 ms/1.3 ms/7 . The field of view was 224 Â 224 mm 2 , the in-plane spatial resolution was 2 Â 2 mm 2 , and 28 slices of thickness 3 mm were acquired. For two of the JHU patients (patients 9 and 10), this was 4 mm thickness and 23 slices. The total DCE scan time was 6 min 40 s, giving 100 dynamics. Gadoterate meglumine contrast agent (0.1 mmol/kg body weight, Dotarem; Guerbet, Paris, France) was used in Lund and gadoteridol contrast agent (20 ml, ProHance; Bracco Diagnostics, Monroe Twp, NJ, USA) was used at JHU. Using a power injector, the gadolinium contrast agent was injected at an injection rate of 5 ml/s, after collection of baseline images for 40 s. A saline flush of 20 ml was given at a rate of 5 ml/s after the gadolinium administration. The gadolinium injection was given approximately 45 min after the D-glucose injection.

| Postprocessing
All postprocessing and analysis were performed using custom-written Matlab (R2020a; MathWorks, Natick, MA, USA) scripts. Realignment of glu-coCEST images and motion correction of DCE images were performed using Elastix 31 rigid 2D and 3D transformation, respectively. DCE images and T 1 -weighted postgadolinium images were registered (resliced) to the glucoCEST images, using SPM12, 32 to match the orientation and spatial resolution of the glucoCEST images.
DGE and DCE images were calculated for each time point t using Equation (1).
where "+" refers to DGE and "À" to DCE, and S base is the average signal of the preinjection (baseline) images. The first image in each series was discarded. The obtained DGE and DCE response curves were smoothed, first by removing spikes (Matlab function "medfilt1") followed by a fivepoint moving average. AUC was calculated by using the Matlab trapezoid function to integrate all smoothed DGE or DCE curves in each voxel over a time interval, t = 0 to t = 6 min (or 12 min for DGE), normalized to the number of integrated images. The start of the injection (D-glucose or gadolinium) is referred to as t = 0 for both methods.
ROIs were drawn in resliced T 1w postgadolinium images in the lesion (gadolinium-enhancing region), contralateral tissue, and in contralateral normal-appearing white matter (WM). The ROIs in contralateral tissue were of similar size, placement, and shape to the lesion ROIs, and the WM ROIs were small to avoid including non-WM.

| Curve maps
Curve maps were calculated by classifying the response curve in each voxel as one of seven postulated curve types, creating a map that represents the spatial distribution of curve shapes (temporal signal responses). These different curve types were defined based on DCE literature, 23 and according to the following criteria: type 1, strictly decreasing (negative) signal; type 2, no signal change (i.e., curves that are close to zero [within the noise limits] or which include spikes that do not have any particular enhancement pattern); type 3, initial signal increase (uptake) followed by a decrease (washout); type 4 curves are shaped like type 3, but with a steeper initial slope. For DGE, a steeper positive initial slope represents a breakdown of the BBB, which allows one to separate normal brain and tumor tissue. In normal brain tissue, facilitated D-glucose transport causes a concentration reduction by a factor of 4-5 relative to the blood. Importantly, the glucose transport in normal tissue is slow, with a half-life of about 2 min. 19 However, during BBB breakdown in tumors, more D-glucose will leak in (concentration becoming similar to the vessel and signal going up) and D-glucose will enter the extravascular extracellular space (EES) faster, resulting in an increased slope and signal maximum. Similarly, types 5 and 6 curves represent a slower (type 5) or quicker (type 6) exponential-like (1-e -t ) increase, with a signal intensity that approaches or assumes a constant value during the latter part of the experiment. Finally, type 7 curves are characterized by a strictly increasing signal over the course of the experiment. The basic response curves are shown in Figure 1, together with an overview of the curve map calculation process.
The categorization of response curves into curve maps was accomplished by deep learning using Matlab on a Quad-Core i7 CPU. Basic response curves of 150 time points were created based on mathematical functions, to resemble typical DCE response curves used in DCE TIC shape analysis. Paired training and test datasets of 960,400 response curves each were created, with corresponding labels (a number from 1 to 7).
A paired validation dataset of 240,100 curves was also created. Except for type 2 curves, all curves in all three datasets were multiplied by an arbitrary scaling factor between 1% and 15%. In the type 2 curves, spikes (amplitude between 0% and ±20%) were added at up to 10 random time points to account for response curves without a particular pattern. Gaussian noise (Matlab function "awgn", SNR = 20 dB, signalpower = "measured") was then added to all curve types to create different versions of the synthetic response curves within each curve type. The modified synthetic response curves were placed in an image and the corresponding curve-type labels were placed in another image of the same size, representing a curve map. A bidirectional long short-term memory 33 (biLSTM) network was trained to classify the synthetic response curves as curve types (supervised training). The biLSTM network consisted of one biLSTM layer of 25 hidden units, one dropout layer with dropoutconstant 0.1, one fully connected layer of size 7, and one softmax layer for classification using cross-entropy as loss function. The learning rate was 0.001 and the network was trained for two epochs, upon which convergence was observed, using a minibatch size of 128. Training the network took 30 min, and a test accuracy of 99.9% was obtained, as shown in Figure 1.
The DGE and DCE patient response curves (i.e., the postinjection time interval without the baseline) were presented to the trained network voxel by voxel, and the output was the labels i.e., a curve map. Creating one curve map took about 2 s. The intensities of the DCE curves were first divided by 10 to enable using the same network for DGE and DCE. A custom-made color scale was created for images representing response curves on a voxel-by-voxel basis ( Figure 1).

| Analysis and statistics
WM, contralateral, and lesion ROIs were chosen to analyze AUC and curve maps for both DGE and DCE MRI. The average AUC value was calculated within each of these ROIs. The relative frequency of curve types, expressed as the fraction of the total number of voxels in each ROI, was calculated from the curve maps. Scatterplots of DCE AUC versus DGE AUC were created by averaging within the lesion for each patient after sorting the voxels according to curve type.
F I G U R E 1 Overview of the process for calculating curve maps. Seven different basic curve shapes were defined and subsequently modified (scaled and noise added) to create one training and one test dataset consisting of almost 1 million unique synthetic response curves each. A bidirectional LSTM network was trained on synthetic response curves and corresponding labels (curve types 1-7). Real patient response curves were then presented to the trained network, and each curve was classified, creating a curve map. Acc, accuracy; LSTM, long short-term memory AUC average values and curve-type distributions in lesion were compared with contralateral tissue and WM using two-sided Wilcoxon signed-rank tests. Differences with p less than 0.05 were considered statistically significant.

| Additional analyses
Both the injection and the total scan duration were longer for DGE than for DCE. In the analyses described above, the same time interval (6 min from the start of injection) was used for both methods. However, it is of interest to also study the DGE signal evolution over a longer time interval, so DGE curve maps and AUC maps were also calculated over the postinjection time interval of t = 0 to t = 12 min. Two patients (8 and 10) were excluded from this analysis because of severe motion in the latter part of the DGE scan.
The lesion ROI analysis described above considered only the gadolinium-enhancing part of the lesion. However, glucose might not enhance the same regions as gadolinium. This was evaluated in patients with lesions that showed heterogenous enhancement (e.g., because of necrosis) in and curve maps were analyzed as described previously, for both the 6-and the 12-min postinjection intervals. However, statistical analysis was not performed because of the small sample size in this group. Patients 4, 5, 6, and 11 were excluded from this analysis because of homogenous postgadolinium enhancement. Patients 8 and 10 were excluded because of a lack of 12-min maps, as described earlier.

| RESULTS
Curve maps and AUC maps of six patients, calculated over the first 6 min after the start of the injection, are shown in Figure 2. In the DGE curve maps of seven of the patients, the lesion included curve types (types 5, 6, and/or 7) that contrasted with surrounding tissue, as seen in patients 1, 2, 3, and 6. The lesion was not possible or was difficult to identify in the DGE AUC or curve maps of the other four patients, as seen in patients 4 and 5. The DCE curve maps showed a clear difference between lesion (curve type 5) and normal tissue (mainly curve type 2, with some contribution from curve type 4, where the latter is seen in blood vessels.). The averaged frequency of curve types in WM, contralateral tissue, and lesion over all patients is shown in Figure 4. For DGE 6-min curve maps, the most common curve types in all three tissue compartments were types 1 and 2. The fraction of curve types 4-6 was higher in the lesion than in the other two compartments, and the fraction of type 5 curves (16%) was significantly higher than in WM (p = 0.01) and contralateral tissue (p = 0.02). For DGE 12-min curve maps, the fraction of type 5 curves in lesion increased to 25%. For DCE, WM and contralateral tissue included mainly type 2 curves (93% and 78%, respectively), while the lesion had mainly type 5 curves (77%), followed by type 6 curves (12%).
DGE and DCE curves were classified as the same curve type in 26% and 20% of the voxels in the whole slice (background and skull excluded) and in the lesion, respectively, for the 6-min curve maps of the nine patients that were also included in the 12-min calculation. For the 12-min DGE curve maps, the DGE and DCE curves were assigned to the same curve type in 19% of the voxels for the whole slice, and 29% of the voxels for lesion. Figure 5 shows the curve distributions for two patients that had good spatial correspondence within the lesion between DGE and DCE curve maps. The resemblance between DGE and DCE in terms of curve type increased for the longer DGE time interval. Examples of response curves and their classifications are shown in Figure 6. The response curves were retrieved from patient 1 at voxels in the curve map within or close to the lesion, chosen to visualize one realistic shape for each curve type. The scatterplots in Figure 7 show averaged DGE and DCE AUC values within each curve type in the lesion ROI for both DGE postinjection time intervals of all patients. Figure 8 shows the curve-type distribution and averaged AUC in WM and in the gadolinium-enhancing and nonenhancing parts of the lesion of patients 1, 2, 3, 7, and 9. In DGE curve maps (6 and 12 min), the relative distribution of curve types was similar in enhancing and nonenhancing regions, but different from WM. The nonenhancing regions in DCE were dominated by type 2 curves, followed by type 5 and 6 curves. The averaged DGE AUC was similar in both tumor regions, whereas the DCE AUC in the nonenhancing region was more similar to WM than to the enhancing tumor region.

| DISCUSSION
To investigate the correspondence between tissue response curves obtained by DGE and DCE MRI, we calculated AUC maps as well as curve maps that show the spatial distribution of different response curve types. DGE enhancement (positive AUC that was higher in lesions than in contralateral tissue and WM; Figure 3) was seen in seven out of 11 patients. DGE failed to enhance the lesion (negative AUC) in two patients (patients 4 and 11), and the enhancement was ambiguous in two other patients (5 and 7), meaning that the AUC in the lesion was close to zero but higher than in WM or in contralateral tissue. The averaged DGE AUC was significantly higher in lesion than in contralateral tissue and WM ( Figure 3). In line with previous research, 1,3,16 our conclusion is that DGE can differentiate lesions enhanced in DCE from normal tissue in some of the cases, but not in others, most likely reflective of the lower SNR.  15 An important difference in uptake kinetics between D-glucose and gadolinium is the facilitated transport of D-glucose into cells, where it is metabolized. As a consequence of the latter, the curve representing D-glucose signal will reach its maximum later than the gadolinium curve. 6 In addition, the later part of the DGE curve may also include contributions from metabolic products of D-glucose. While extract studies have shown the contributions of phosphorylated intermediates to be negligible both in normal tissue 19 and tumors, 34 the merged broad OH signal of lactate in tumors may in principle contribute. 6 Overall, recent studies have shown that the signal in DGE is mainly caused by vascular and EES components, the latter enhanced by the low pH in the EES in tumors. 6,35 In addition, for the current protocol with 6-12 min of postinjection imaging, the blood glucose levels are not In the DCE curve maps, the lesion can be separated unambiguously from normal tissue. Type 5 was the most common curve shape in lesions, and was rarely found outside of the lesions. When comparing DGE and DCE curve maps visually, the two main differences were the presence of type 1 curves (negative signal) in DGE, and the more consistent occurrence of type 5 curves in the lesion in DCE maps. The fraction of type 1 curves increased in all tissue compartments when comparing the longer with the shorter DGE time interval. Negative DGE signal in WM has been observed in other DGE studies, 1,15,16,20 and possible explanations are field drift, a susceptibility-based frequency shift or a glucose-induced change in osmolarity, or in some cases even a combination of one of these effects with a motion-induced B 0 shift. Another obvious difference was the noisier appearance of all DGE maps. For DGE, the ROI curve-type distribution was more heterogenous in lesions than in normal tissue, with types 1 and 2 being the most common, although the presence of type 4, 5, and 6 curves was increased in lesions compared with WM and contralateral tissue. Type 1 and 2 curves may be present in voxels that also contain other curve types. At low SNR, the presence of types 1 and 2 can dominate the contribution to the average shape representing the voxel. An increase in SNR, e.g., by increasing the amount of D-glucose infused or using other pulse sequences, could improve the detectability of type 3-7 curves.
Different enhancement patterns were seen when comparing enhancing and nonenhancing parts of the lesion, where DGE sometimes gives a signal increase when DCE does not. This finding is consistent with previous studies. 1,3,4 An important difference between DGE and DCE is that glucoCEST is sensitive to pH. The lower pH in tumor EES, because of the presence of lactate, slows down the OH exchange rate, thus resulting in a higher glucoCEST effect in tumors compared with normal brain. 6 Furthermore, D-glucose can enter the brain and tumor cells via glucose transporters, while gadolinium does not cross an intact BBB. This facilitated D-glucose transport causes a concentration reduction by a factor of 4-5 in the normal brain tissue relative to the blood. However, during BBB breakdown in tumors, both D-glucose and gadolinium will freely enter the EES, resulting in an increased DGE and DCE effect. Another difference is the uptake of glucose in CSF, giving a relatively high glucoCEST effect in CSF, as shown by Huang et al. 17 This can explain some of the hyperintensities seen in CSF-rich regions in the DGE AUC maps in Figure 2.
DGE and DCE curves in the lesion were classified as the same curve type in 20% of the cases for the 6-min DGE curve maps, and this number increased to 29% for the 12-min ones. These results were in contrast to the results when studying the whole slice, where the same curve type was found in 26% of the voxels for 6 min and in 19% of the voxels for 12 min. Plausible reasons for this discrepancy are: (i) the different uptake kinetics of D-glucose and gadolinium, leading to different enhancement patterns; (ii) the negative signal (type 1) in DGE, especially in WM, which is rarely seen in DCE, but is consistent with previous results 15,16,20 ; (iii) the slower injection and thus delayed response in DGE; and (iv) the longer scan time and higher sensitivity to motion in DGE, 36 resulting in more motion-related artefacts (pseudoCEST effects). Typical pseudoCEST effects at CSF-tissue boundaries can be observed in the DGE AUC maps of patients 2 and 3 ( Figure 2). These effects can be partially reduced by motion correction if a data volume has been acquired, but it is important to note that pseudoCEST effects may be misinterpreted as true DGE enhancement in or around lesions situated close to CSF. 36 The scatterplots (Figure 7) revealed that curve types 5 and 6 were associated with the largest DCE signal change, while the typical magnitude of the DGE signal change was more similar for types 4, 6, and 7. This finding suggests that curve maps can be particularly valuable for DGE, where the enhancements seem to be of similar magnitude, potentially leading to less differentiation between regions with different curve shapes in the AUC maps, while for DCE, the AUC maps can indirectly give information about curve shape and thus tumor response. It should be noted that all F I G U R E 8 Relative frequency of curve types in ROIs for normal-appearing WM, nonenhancing and enhancing parts of the lesion, as identified in postgadolinium T 1w images, for patients 1, 2, 3, 7, and 9. Bottom right: averaged AUC in the same patients and ROIs. AUC, area under the curve; ROI, region of interest; WM, white matter synthetic curves except for type 2 were arbitrarily scaled before training the network. In DCE, no transport of gadolinium into the EES should occur in normal tissue. However, water exchange can introduce small signal changes in the curves. Normalization of the real curves would therefore equalize such unwanted curves to the much larger amplitude curves seen in lesions. However, in pure DGE studies, normalizing both synthetic and real curves might be beneficial, because also very small signal changes would be classified.
In this first attempt using typical DCE curve shapes, the results imply that DGE curve maps have the potential to facilitate differentiation of tumor from normal tissue based on differences in temporal enhancement patterns. They also suggest that curve maps have potential in the assessment of heterogenous lesions. However, more work is needed, and we anticipate that the curve map method can be improved as knowledge of DGE curves increases.

| Limitations and future aspects
To the best of our knowledge, this is the first application of voxel-wise curve-shape analysis to DGE MRI. The curve maps reflect differences in temporal patterns, in contrast to the commonly used AUC maps, which show the accumulated signal intensity in a chosen time interval. The curve maps can help to distinguish lesions from normal tissue, because different temporal enhancement patterns can be expected. While not achievable in the present study because of the limited number of patients examined, the method could also have the potential to differentiate between different lesions, tumor types, or grades. Since 2016, the World Health Organization has included molecular markers in the grading of tumors, and, recently, Gates et al. used machine learning with four inputs (K trans , cerebral blood flow, fractional anisotropy, and T 2 ) to predict Ki-67 maps. 37 One might speculate that a similar approach should be possible in glucoCEST by using the curve types together with other imaging parameters to identify surrogates for molecular markers. An advantage with curve-shape analysis is that it is less sensitive to variations in MRI scanning protocols. 23 The curve map approach may also give valuable input for the continued development of kinetic models for DGE MRI, and can be applied to further investigate the compartmental origin of the DGE signal by studying DGE response curves of glucose and glucose analogs from animal data.
The curve-shape classification method can be developed further. In this first demonstration, the curve types were generalized to be valid for both DCE and DGE MRI, and more refined versions could be used for either of the methods, for example, by including different color intensities (hues) for different signal intensities within a curve type. A future possible method to create curve maps could be to use a data-driven approach to identify different curve types in both DCE and DGE data, to compare the enhancement patterns between the two methods. Another approach could be to use the recent models of glucose uptake, 18 or similar models, to create a curve library for DGE. The curve classification algorithm can also be fine-tuned for finding AIFs, which are needed for model-based analysis. The deep learning-based curve map method is relatively quick and can easily be adapted for different purposes by retraining the network with different sets of basic curve types, depending on the intended use.
This can be particularly useful when comparing methods using different contrast agents, for example, sugars with different tumor uptake or metabolic properties, 6 as suggested above.
One obvious limitation of the current study is the small cohort, and more patients are needed to draw conclusions regarding the capability of DGE MRI to distinguish between, for example, different categories of lesions or tumor types. Because of the small CEST signal, motion artefacts and concomitant motion-induced B 0 artefacts are well-known problems in DGE imaging. 15,16,20,36 In this study, data were collected in an early stage of DGE MRI (2015-2017) using a single saturation offset acquisition, which hampers dynamic B 0 correction. Single-slice and single saturation offset acquisition allows a higher temporal resolution but increases the risk for motion artefacts and obstructs retrospective motion correction. Hence, we cannot rule out that parts of the DGE signal changes studied are affected, or in some cases even caused, by motion. Movement patterns or B 0 changes that result in a diverging DGE signal change over time could be erroneously interpreted as, for example, type 7 curves. In the future, this risk can be minimized by proper motion prevention and correction, for example, by avoiding single-slice acquisition. Multislice acquisition is, obviously, also required for a more reasonable volume coverage. The curve-shape classification can also be affected by partial volume effects, and these can be reduced by employing higher spatial resolution. Additionally, in dynamic approaches such as DCE and DGE MRI, the AIF will affect the tissue curve shape. Because the AIF may vary among individuals, this can be a confounding factor when comparing response curves and curve maps between individuals and between DGE and DCE. However, retrieving an accurate AIF can be a challenging process. Although a long infusion time relative to the time resolution might result in measuring a more accurate input function, 38,39 the AIF may still be affected by other confounding factors, such as partial volume effects.
In this study, the DCE imaging was performed $ 45 min after D-glucose injection, and the induced hyperglycemia may affect the vascular perfusion in lesions or normal tissue. However, Walker-Samuel et al. found no significant change in vascular perfusion following D-glucose administration in mice. 8 Furthermore, Hasselbalch et al., using the Kety-Schmidt technique, found no change in cerebral blood flow during acute hyperglycemia (15 mmol/l, clamp technique) in healthy humans. 40 It is therefore likely that the injected D-glucose does not significantly affect the DCE curve shapes.
This study was carried out at 7 T, where the DGE effect size is larger than at lower field strengths. It was reported by Xu et al. that the DGE effect at 7 T in a gadolinium-enhancing region of a glioma rim was approximately four times higher than at 3 T. 16 This conclusion was also reached by Bender et al. when using DGEρ at 3 T. 30 Therefore, one of the challenges with DGE MRI is the translation to clinical field strength (3 T), but the use of curve-shape analysis with deep learning may help in better analyzing these noisier curves. The analysis approach proposed in this study can be adapted and applied to any DGE data, and the results are expected to become more accurate and revealing as DGE data acquisition and image processing methods improve. One such improvement could be deep learning-based methods that increase signal-and contrast-to-noise ratios. The ideal subsequent step would be to apply the curve map method to multislice DGE datasets with sufficient temporal resolution that are both motion-and B 0 -corrected.

| CONCLUSIONS
DGE MRI enabled differentiation of lesions from normal brain tissue in humans based on tissue response curve-shape analysis and AUC analysis, and has the potential to facilitate radiological assessment of DGE and DCE images via visualization of characteristic temporal enhancement patterns. DGE and DCE enhancements were similar but not temporally identical and were not entirely confined to the same regions.