Accelerating GluCEST Imaging Using Deep Learning For B₀ Correction

Abstract

Purpose.

Glutamate Chemical Exchange Saturation Transfer (GluCEST) MRI is a noninvasive technique for mapping parenchymal glutamate in the brain. Because of the sensitivity to field (B₀) inhomogeneity, the total acquisition time is prolonged due to the repeated image acquisitions at several saturation offset frequencies, which can cause practical issues such as increased sensitive to patient motions. Since GluCEST signal is derived from the small z-spectrum difference, it often has a low signal-to-noise-ratio (SNR). We proposed a novel deep learning (DL)-based algorithm armed with wide activation neural network blocks to address both issues.

Methods.

B₀ correction based on reduced saturation offset acquisitions was performed for the positive and negative sides of the z-spectrum separately. For each side, a separate deep residual network was trained to learn the nonlinear mapping from few CEST-weighted images acquired at different ppm values to the one at 3 ppm (where GluCEST peaks) in the same side of the z-spectrum.

Results.

All DL-based methods outperformed the “traditional” method visually and quantitatively. The wide activation blocks-based one showed the highest performance in terms of Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR), which were 0.84 and 25dB respectively. SNR increase in regions of interest were over 8dB.

Conclusion.

We demonstrated that the new DL-based method can reduce the entire GluCEST imaging time by ~50% and yield higher SNR than current state-of-the-art.

1. Introduction

Chemical Exchange Saturation Transfer (CEST) is an MR based imaging method that can image endogenous[1–8] or exogenous compounds [9] containing protons exhibiting a suitable exchange rate with bulk water. CEST techniques offer excellent in-plane spatial resolution, as well as higher sensitivity for their target molecules than 1H MR spectroscopy [10]. The CEST effect has been used to measure relative concentration of glutamate [4], the major excitatory central nervous system neurotransmitter, which plays a central role in brain function in health and disease [11–18]. It is possible to image glutamate only at “ultra-high field”, since its constituent amine group has the exchange rate (5500±500 s⁻¹) comparable to chemical shift difference (Δω) at 7T magnetic field strength. Glutamate weighted CEST (GluCEST) has been demonstrated in vivo in various murine disease models [14–18], as well as in human clinical applications in epilepsy, schizophrenia and brain glioma [11–13]. Recently, GluCEST was documented to have a high degree of within-day and long term (0–6 months) reproducibility in healthy volunteers [18, 19]. However, CEST imaging is sensitive to both B₀ and B₁ inhomogeneities, requiring post-processing to correct for the errors introduced by both types of inhomogeneities. A widely used approach to correct for B₀ inhomogeneity in CEST images involves acquiring CEST images at a series of frequency offsets covering the expected B₀ inhomogeneity range and then synthesizing the B₀-corrected image in a pixelwise manner, informed by a separately acquired B₀ field map [20]. This need to acquire CEST images at several offset frequencies makes CEST acquisition very time consuming, potentially leading to patient discomfort and imparting high sensitivity to motion artifacts. A significant acceleration in data acquisition and robustness could be achieved by reducing the number of CEST acquisitions required to correct for B₀ inhomogeneity.

One possible solution is to use a practical six-offset dataset to calibrate the whole spectrum. To accelerate the B₀-inhomogeneity correction process, one can build a model of the z-spectrum which to be used for future CEST imaging with fewer offset acquisitions. Since each subject would need a model, a full range of different offset acquisitions will be required for each subject. An alternative approach is to learn the z-spectrum patterns from a representative subset of fully acquired CEST data and apply the learned model to correct B₀ inhomogeneity for new CEST imaging data with only a small number of offset acquisitions. When appropriately trained, the model is generalizable to different subjects, so there is no need to build a model for each individual subject. This process can be done using deep machine learning algorithms [21], which have shown astonishing success in many complex real-world applications, including autonomous driving [22], natural language processing [23], and other areas of medical imaging [24–27]. In CEST imaging, DL has been used to synthesize high field CEST signal from data acquired at low field [28]. In this study, we proposed a DL-based method to correct B₀ inhomogeneity for glutamate-weighted (GluCEST) CEST imaging with an aim to significantly reduce the number of different frequency offset acquisitions but without sacrificing B₀ inhomogeneity correction accuracy. While accelerating the imaging process often comes with reduced SNR, we expect to see increased SNR due to the CEST signal manifold learning and the constrained location convolution embedded in DL.

Our contributions in this paper include: 1) We proposed a first-of-its-kind DL-based CEST imaging B₀-inhomogeneity correction method with a demonstrated application to GluCEST imaging to accelerate the total scan time; 2) We developed an initial-interpolation-based map-wise deep neural network; 3) We optimized an existing DL neural network by incorporating the wide activation residual learning network blocks from the Wide-activation Deep Super-Resolution network (WDSR) [29].

2. Method

2.1. Glutamate Chemical Exchange Saturation Transfer (GluCEST)

Human subject studies were approved by the University of Pennsylvania Institutional Review Board. Signed informed written consent forms were obtained from all volunteers before the experiment. Seven healthy volunteers (6 males, 1 female) aged 28 to 66 years old (45 ± 14.54 years) participated in the study and were imaged using a 7T Siemens scanner (Erlangen, Germany) with a Siemens volume coil transmit/32-channel phased-array receive coil. 29 scans were collected from seven subjects, with twenty scans being used as training data sets and nine scans being used as the test data sets [19].

2D GluCEST images of an axial slice were acquired using the pulse sequence published in [4]. The slice was positioned 3 to 4mm above the corpus callosum. After the initial scan on a subject, a coregistration program, ImScribe (https://www.med.upenn.edu/cmroi/imscribe.html) [19], was used to identify the slice location for any subsequent scans. GluCEST imaging parameters were as follows: slice-thickness=5 mm, in-plane resolution =1×1mm2, matrix size=256×256, gradient-echo readout with TR=7.4 ms, TE=3.5 ms, read-out flip angle=10°, averages=2, shot TR=8000 ms, shots per slice=1. Saturation was performed with a train of 99.8-ms Hanning-windowed saturation pulses with an inter-pulse gap of 0.2 ms and B1rms=3.06μT and total duration of 800 ms. The saturation flip angle was 3772°. CEST images were acquired at varying saturation offset frequencies from ±1.8 to ±4.2 ppm (relative to the water resonance) with a step size of 0.2 ppm, and offset frequencies of ±20 ppm and ±100 ppm were also collected to generate the magnetization transfer ratio (MTR) map. The B₀ map was acquired at the same imaging slice position using the water saturation shift reference (WASSR) sequence [30]. WASSR images were collected from ±0 to ±1 ppm (step size=0.1 ppm) with a saturation pulse of B_1rms=0.29μT, duration=200 ms and the same imaging parameters as mentioned above. Gray and white matter segmentation was performed based on the magnetization transfer ratio map, using a K-means cluster algorithm with the number of segments set to 3 (gray, white, and CSF) [31]. The traditional GluCEST contrast was calculated as the ratio of the difference between an image obtained with saturation pulses on-resonant with the frequency of the exchangeable amine protons (+3 ppm downfield), and an image acquired with saturation pulses on-resonant to the frequency of −3 ppm (upfield from water), according to the following equation:

$$G{luCEST}_{asym(\mathrm{\Delta }\mathrm{\omega }=3ppm)}=\frac{M_{sat(-3ppm)}-M_{sat\left(3ppm\right)}}{M_{sat(-3ppm)}}\mathrm{*}100$$

Where $M_{sat(\pm 3ppm)}$ are the magnetizations obtained with saturation at$\mathrm{\Delta }\mathrm{\omega }=\pm 3ppm$ offset from the water resonance which need to be interpolated from the acquired data at different off-resonance positions due to the inhomogeneity of B₀.

2.2. DL-based B₀ correction and GluCEST quantification

Figure 1 illustrates the entire DL-based B0 calibrated GluCEST quantification (DL-B0GluCEST) process. Instead of using all 26 CEST-weighted images (from ±1.8 to ±4.2 ppm with a step size of 0.2 ppm), DL-B0GluCEST was set up to generate the final GluCEST ratio with data from only 1, 3, 5 or 7 pairs of saturation frequencies acquired at the downfield and upfield positions. While it is possible to build a deep network to interpolate the CEST weighted signal at +3ppm from those at arbitrarily located ppm locations, the model would need an additional channel to accept the offset locations as input, which will require additional network parameters. To simplify the model, we used an initial-interpolation module (shown as 2-A, 2-B in Figure 1) to shift all voxels of a CEST image acquired at a particular saturation offset (and subject to the spatially varying B0 field) to the same nominal offset location. Thus, the subsequent DL module can be still based on a standard convolutional neural network (CNN) which is known to be able to improve SNR because of the convolutional process involved at each layer. Once the 3ppm and −3ppm CEST data were produced by the two DL networks, the final GluCEST contrast ratio was calculated according to the equation above.

Figure 1.

The schematic illustration of the proposed DL-based B₀ calibrated GluCEST quantification neural network. Multiple pairs of images at different frequencies and B₀ offset map are used to interpolate images at same nominal offset locations as input of DL based module. From this module, the estimated images at ±3 ppm are predicted respectively. The combination of positive and negative estimations calculates the final B0 corrected GluCEST contrast ratio.

2.3. Network structure used in DL-B₀GluCEST

Network Architecture

As shown in Figure 2, the DL based module in DL-B₀GluCEST is a deep residual network with wide activation inside. The backbone of the network is the vanilla residual network while all the residual blocks are replaced by WDSR blocks. The optimal number of WDSR blocks (8), the number of filters (32) and expansion ratio (4×) for WDSR filters are determined from the original WDSR paper [29]. For CEST-weighted data at different positive and negative saturation offsets (the Z-spectrum values), patches of the initial-interpolated images were used as the input. The training reference was the GluCEST weighted data at 3 ppm estimated by the current method [19]. Each input channel was associated with a separate convolutional layer. The output of all input layers was concatenated into one channel as the input to the successive layer. Following each convolutional layer were eight consecutive WDSR blocks. In each block, wide activation was used to retain the high-frequency tissue boundary information. Following the eight consecutive blocks, another convolutional layer without any activation function was attached to the end to get the B₀ corrected 3 ppm image patches with additional input from the second layer.

Figure 2.

The architecture of DL based module used in DL-B₀GluCEST is an enhanced deep residual network. The first layer consists of 32×3×3 convolutional filters for each input image. After concatenating them as one channel and going through another layer with 32×3×3 convolutional filters, the subsequent layers include 8 consecutive WDSR blocks, which each contain two convolutional layers (128×3×3 and 32×3×3) and one activation function layer. Another 1×3×3 convolutional layer was attached to the end to get the B₀ corrected 3 ppm image with additional input from the concatenated layer with 1×3×3 convolution. (a×b×c indicates the property of convolution. a is the number of the filters and b×c is the kernel size of one filter).

Wide Activation.

Wide activation (more filters) was used to preserve more information from the input before the rectified linear unit (ReLU) activation layer. Adding more layers might impart a similar improvement but will increase the computational burden. It has a slim identity mapping pathway with wider (4×32) channels before activation in each residual block.

Training.

Training was implemented by using patches. Among 29 scans, first 20 of them were selected for training and the rest for testing. With original image size of 256*256 for each scan, patch with a size of 128*128 and stride of 8 was cropped and totally 8381 patches were generated. 300 epochs and a batch size of 4 were used. The ADAM optimizer with an initial learning rate of 0.001 was used to train the network. After 120 and 180 epochs, the learning rate was reduced by 9 and 19 times, respectively. Another training with data augmentation was implemented as a side experiment. Data augmentation was from noise corrupted data. Two Gaussian noise (mean=0, std=10 and mean=0, std=20) was applied to the input GluCEST while leaving corresponding reference unchanged, besides the original 20 scans. The number of training samples increased by three times while the network and hyperparameter kept the same.

Experimental Setup and Evaluation.

DL-B₀GluCEST was compared to a popular DL model, Unet [32] without using wide activation blocks. DL-B₀GluCEST was trained and validated for 3, 5 and 7 input image pairs separately to show the stability and consistency of the algorithm. For DL-B0GluCEST with 1/3 input pairs, the CEST images acquired at ±3/±2, 3, 4 ppm were used respectively. For DL-B0GluCEST with 5 input pairs, those at ±2.4, 2.8, 3, 3.4, 3.8 ppm or at ± 2.2, 2.6, 3, 3.2, 3.6 ppm were used as the input pairs. For the 7-input-channel run, the images acquired at position from ±1.8 to 4.2 ppm with a step of 0.4 ppm were used. Structural similarity index (SSIM), peak signal-to-noise ratio (PSNR) and contrast-to-noise ratio (CNR) were calculated as the performance indices. CNR was measured by the ratio of the subtraction between the mean value of a gray matter (GM) region-of-interest (ROI) and a white matter (WM) ROI and standard deviation of a WM ROI. Both ROI were extracted from segmentation mentioned in section 2.1. All DL experiments were performed using framework of Keras and Tensorflow running on a Ubuntu16.04 system with NVIDIA Tesla P100 and GTX 1080.

3. Results

Figure 3 shows the GluCEST maps of one representative subject. The results produced by traditional method [19] are used as reference. As shown in the first row, simply reducing acquisition time to 7/5/3/1 pairs CEST-weighted images with conventional method dramatically lowers SNR of the results. DL is a proper approach to overcome this issue. As compared to the traditional method in the second row, all DL methods produced high quality results in terms of tissue structure and contrast with same or less acquisition time except WSDR-1-pair. We did the experiment with single image input at 3 ppm, which should be the closest to the reference. As shown in Figure 3, this model preserves the structure roughly, however, quantitatively it is worse than any other DL methods based on multiple inputs even worse than Unet. As a result of this model, SSIM is 0.82, PSNR is 23.74, and CNR is 0.45. Therefore, the WSDR-1-pair cannot be considered as an eligible solution. The number of multiple inputs to DL-B₀GluCEST with WDSR showed minor effects on the output. All DL-B₀GluCEST showed few voxels with negative GluCEST ratio. Therefore, reducing acquisition time from 13 pairs to 3/5/7 pairs are allowed to accelerate B₀ correction duration. Considering the visual and quantitative performance, instead of using all 13 pairs CEST-weighted images (from ±1.8 to ±4.2 ppm with a step size of 0.2 ppm), DL-B₀GluCEST used 5 pairs of CEST-weighted images to get the final GluCEST contrast ratio, which saved 61.5% acquisition time.

Figure 3.

GluCEST ratio maps of a subject were calculated by different methods. Reference refers to the one produced by traditional method [19]. Reference-13/7/5/3/1-pair refers to the results by the same method with different pairs of input images. WDSR shows the DL methods and 1/3/5/7-pair shows the number of input image channels. WDSR-5-pair-02 indicates an alternate selection of 5 saturation offsets as input. Details like which saturation offsets were selected can be found in Experimental Setup and Evaluation in section 2.3.

Figure 4 shows the validation indices of all assessed methods. Except WDSR-1-pair model, WDSR outperformed Unet dramatically in terms of SSIM (p=0.04 with conducting a one-way ANOVA test shown in Supporting Information Table S1 and post hoc shown in Supporting Information Table S2), which are all greater than 0.84. WDSR-5-pair and WDSR-7-pair had slightly better SSIM than WDSR-3-pair. All DL methods yielded nearly the same PSNR (p=0.928 with conducting a one-way ANOVA test). The results suggested that WDSR and Unet both have good estimation in terms of image quality while WDSR surpassed Unet with perceived change in structural information. DL produced higher CNR than the traditional method at the region of interests that compared contrast between some subcortical gray matter and white matter.

Figure 4.

All DL based methods were compared with the reference estimated by current method [19] quantitatively using SSIM (A), PSNR (B), and CNR (C). The results of SSIM shows DL methods have statistically significant difference whereas PSNR and CNR (without reference) do not.

In order to validate the effectiveness of the model, a 3-fold cross validation was created for DL-B₀GluCEST with WDSR-5input and the results are illustrated in Figure 5. The 29 scans were randomly slipped into 3 groups and each of them was used for testing. As shown in Figure 5, there is no significant difference between groups in terms of SSIM, PSNR and CNR via one-way ANOVA test. The results are slightly better than the ones in Figure 5 since the scans were shuffled this time instead of selecting first 20 scans as training data. It shows consistent performance for DL-B₀GluCEST.

Figure 5.

Cross validation of WDSR-5input were evaluated by SSIM (A), PSNR (B) and CNR (C). All the results do not show statistically significant difference by any metric. It validates the consistency of DL-B₀GluCEST.

Figure 6 quantitatively demonstrates the efficacy of DL based methods for predicting the quantitative CEST signal by plotting the linear regression lines and scatter diagrams between the reference values and the predicted ones. Each red dot denotes a voxel of the GluCEST map. The x-axis represents the reference and the y-axis is the predicted value. The green line was the fitted relationship between the reference values (the pseudo ground truth) and the predicted ones, while the blue dashed line has a perfect slop of 1, indicating a perfect prediction with respect to the reference. The smaller the angle between the green and blue lines is, the better performance of the prediction model. In consistent with the performance indices measured by SSIM, PSNR, and CNR, WDSR based methods still outperformed Unet in Fig. 4. The number of input channels is insensitive to the results for WDSR models. In addition, the reference vs predicted value coupling (R2≈0.9) for the training samples was better than that for the testing samples (R2≈0.65) since the training data were seen by the model during training. Nevertheless, the high reference vs predicted value coincidence for the testing data suggested that the models were able to predict GluCEST properly. In Supporting Information Table S3 and S4, we listed the voxelwise R square value of both training and testing subjects by all possible methods. Two subjects (sub6 and sub9) showed relatively low R square values, though their DL-generated images still displayed reasonable image contrast similar to what shown in Figure 7. The major image discrepancy across the reference and the DL methods came from the area marked by the red rectangles. The majority of the voxels within these brain regions had negative values in the reference map, which might be caused by low B1 but were certainly inaccurate. Intriguingly, the same voxels showed positive GluCEST signal in the DL output. While this result demonstrates the capability of DL for suppressing the inaccurate GluCEST quantifications, future experiments with B1 shift correction will be necessary to confirm that.

Figure 6.

Scatter regression plots for a sample of training data (the first row) and testing data (the second row), for all samples of training data (the third row) and testing data (the fourth row) of all DL-B0GluCEST methods. The red points were scatter points of GluCEST prediction versus ground truth voxelwise. The green line was the linear regression fitting based on those points and blue dashed line was the golden standard that prediction equals ground truth, which means the closer distance of the two line, the better performance of the models.

Figure 7.

The prediction results of two testing subjects with low R square value were evaluated. The major image discrepancy across the reference and the DL methods came from the area marked by the red rectangles.

Figure 8 illustrates the capability of the DL-based model for input perturbation with Gaussian noise visually and quantitatively. Different random noise with the same mean and standard deviation (mean=0, std=10) was added to different input CEST images which were subsequently sent to the DL model (without retraining) to get the predicted CEST signal. The same experiment was repeated with a higher level of noise (mean=0, std=20). More pepper noise can be seen in the network output when stronger noise was added to the input. Interestingly, adding the noise contaminated input to the training samples significantly improved the CEST signal prediction fidelity, suggesting an efficient way to augment the training samples and to improve the model performance by artificially creating more training data via adding noise to the input.

Figure 8.

Effect of input perturbation on predicted results compared with the reference and clean prediction for one testing sample depends on different training data. The first row shows the results based on training without data augmentation. The second row shows the results from the integration training sets of clean data and noised data (mean=0, std=10 and mean=0, std=20).

4. Discussion

In the paper we implemented DL as a novel approach to B₀ inhomogeneity correction of CEST data. The proposed method enabled a total acquisition time reduction to 46%, 61%, and 77% using 3, 5 and 7 pairs of input images respectively. To get the reasonable results visually and quantitatively, 5 pairs of input images is the preferable option to balance acquisition time and reconstruction accuracy. There is no significant difference between the performance of 5-pair and 7-pair models while 3-pair model is not visually consistent through different subjects.

Correcting B₀ inhomogeneity of GluCEST data using the conventional approach is an image synthesis task, which relies on pixelwise interpolation. Initially we did the primary experiments based on a pixelwise DL model. The model was able to get the structural information of GluCEST contrast but not accurate for the magnitude. In contrast, DL-based approach employs nonlinear image prior modeling. The realistic image priors can be learned from a large number of training images and predict the B₀ corrected GluCEST contrast for test data (or a newly acquired data set). The receptive fields in CNN enable the model to extract features to capture the full characteristics of image structures. This makes a DL based approach more efficient than a traditional approach.

Also, the dataset only includes 29 scans from 7 subjects, all with a similar slice profile. It is more reliable to contain more input channels to overcome the potential overfit when the dataset expands. The network parameters, such as number of filters, layers, and the expansion ratio were all set up following the relevant literature [29] instead of using separate optimization for the current application. More exploration of network parameters may improve the performance of DL based B0 inhomogeneity correction. In the future, the acquisition strategy can be adapted based on the minimal need of the DL based approach. This may include collecting fewer but asymmetric frequency offset. Additionally, inspired by the results with Gaussian noise, training the model with augmented data (with some noise or manual B0 shift) is another option to expand the training set with better resilience to the noise. In terms of potential T2 or glutamate change, the altered glutamate content alone would not require retraining our model as that would just scale the part of asymmetry plot without affecting underlying nonlinear functional form between Bo corrected CEST and B0 offset values. An implicit assumption in our model is that the T2 dependence of asymmetry curve for GluCEST is not significant as the water peak is very far from nominal glutamate peak at 3 ppm. However, this assumption may not hold true if similar AI framework is attempted for CEST effect from other metabolites, such as lactate, creatine, etc resonating closer to water peak (lactate at ~0.5, creatine at 1.8 ppm).

Furthermore, the proposed method does not take into account the B₁-variabity across voxels and therefore, the current network requires B₁-profile of test data to be similar to B₁-profile of training data set. Future work will extend our DL based framework to account and correct for B₁-inhomogeneity as well.

5. Conclusion

We proposed a DL based framework for correcting B₀ inhomogeneity for GluCEST imaging using fewer acquisitions, which has the potential to reduce CEST acquisition time by >60%. The assessed DL-B₀GluCEST networks were largely insensitive to the number of input frequency offset images and yielded higher SNR than traditional method. We envision that a similar framework can be extended to include correction for B₁ inhomogeneity in future work.

Supplementary Material

supplement tables

Supporting Information Table S1 The quantitative results of mean SSIM, PSNR, and CNR for different DL-based methods (A) and cross validation of three groups in terms of same performace indices.

Supporting Information Table S2 The post hoc tests of ANOVA test for SSIM (A), PSNR (B), and CNR (C) were calculated. Unet-5-pair has significant difference with WDSR-5/7-pair model in terms of SSIM.

Supporting Information Table S3 Voxelwise R^2 value of each training subject

Supporting Information Table S4 Voxelwise R^2 value of each testing subject