Skip to main content
  • Original research
  • Open access
  • Published:

Prediction of multiple pH compartments by deep learning in magnetic resonance spectroscopy with hyperpolarized 13C-labelled zymonic acid

Abstract

Background

Hyperpolarization enhances the sensitivity of nuclear magnetic resonance experiments by between four and five orders of magnitude. Several hyperpolarized sensor molecules have been introduced that enable high sensitivity detection of metabolism and physiological parameters. However, hyperpolarized magnetic resonance spectroscopy imaging (MRSI) often suffers from poor signal-to-noise ratio and spectral analysis is complicated by peak overlap. Here, we study measurements of extracellular pH (pHe) by hyperpolarized zymonic acid, where multiple pHe compartments, such as those observed in healthy kidney or other heterogeneous tissue, result in a cluster of spectrally overlapping peaks, which is hard to resolve with conventional spectroscopy analysis routines.

Methods

We investigate whether deep learning methods can yield improved pHe prediction in hyperpolarized zymonic acid spectra of multiple pHe compartments compared to conventional line fitting. As hyperpolarized 13C-MRSI data sets are often small, a convolutional neural network (CNN) and a multilayer perceptron (MLP) were trained with either a synthetic or a mixed (synthetic and augmented) data set of acquisitions from the kidneys of healthy mice.

Results

Comparing the networks’ performances compartment-wise on a synthetic test data set and eight real kidney data shows superior performance of CNN compared to MLP and equal or superior performance compared to conventional line fitting. For correct prediction of real kidney pHe values, training with a mixed data set containing only 0.5% real data shows a large improvement compared to training with synthetic data only. Using a manual segmentation approach, pH maps of kidney compartments can be improved by neural network predictions for voxels including three pH compartments.

Conclusion

The results of this study indicate that CNNs offer a reliable, accurate, fast and non-interactive method for analysis of hyperpolarized 13C MRS and MRSI data, where low amounts of acquired data can be complemented to achieve suitable network training.

Introduction

In living species, extracellular pH (pHe) is an important physiological parameter that is tightly regulated by intrinsic buffer systems. Locally, deviations from the systemic pH are often caused by pathologies, such as cancer, inflammation, infection, ischemia, renal failure or pulmonary disease [1,2,3]. Since pHe can play a critical role in disease progression [4] and can influence therapeutic success [5], many efforts have been undertaken to develop a quantitative non-invasive pH imaging technique [3, 4, 6]. However, there is no clinical routine method available for spatial quantification of pHe, rendering it still an important target in biomedical imaging.

Magnetic resonance-based pH imaging methods offer high spatial resolution without limitations on the penetration depth and without involving ionizing radiation. In addition, conventional 1H MRI offers high anatomical soft tissue contrast that can be overlaid on top of pH images. MRI-based pHe imaging techniques that have been applied in vivo require the use of exogenous molecules and rely either on their pH-dependent chemical exchange saturation transfer (CEST) or on their pH dependence of chemical shifts [1]. Utilizing endogenous molecules, the intracellular pH (pHi) can be measured by pHi-dependent proton exchange from amide groups of intracellular proteins [4].

Magnetic resonance-based detection of biochemical and physicochemical quantities by exogenous molecules was revolutionized by dissolution dynamic nuclear polarization (DNP) which lifts nuclear spin polarization to a so-called hyperpolarized state leading to a sensitivity gain of more than four orders of magnitude [7]. Hyperpolarized [1-13C]pyruvic acid is currently being used in clinical studies to examine its use for metabolic imaging of cancer, as well as in the brain and the heart [8,9,10]. Several pH-sensitive molecules have been hyperpolarized and been used for in vitro pH mapping including 13C, 15N, 31P, 89Y and 129Xe spin-1/2 nuclei [11]. Only two of those have so far been applied for pH imaging in vivo: hyperpolarized 13C-labelled bicarbonate [3, 12] and hyperpolarized [1,5-13C2]zymonic acid (ZA) [13] as well as its deuterated variant [1,5-13C2,3,6,6,6-D4]zymonic acid (ZAd) [14].

With hyperpolarized bicarbonate, pHe is being determined by the signal intensity ratio of the CO2 and HCO3 peaks, while the pHe determination with ZA works via spectral analysis of the peak position, i.e. the chemical shifts. Chemical shift-based pHe detection offers the unique advantage compared to intensity-based pH detection that multiple pHe compartments within one imaging voxel can be resolved if their spectral peaks are separable, e.g. for resolving different pHe compartments in the kidney [13]. For intensity-based pHe detection, on the other hand, multiple pH compartments within one imaging voxel result in one signal intensity ratio, allowing only the determination of an average voxel pH. The concept of chemical shift-based detection of quantitative physiological measures using hyperpolarized magnetic resonance sensors has, besides for detection of pHe, also been used to quantify zinc [15], calcium/magnesium/iron ions [16], temperature [17], or ligand-receptor interactions [18].

Quantification of these measurements with hyperpolarized NMR sensors is done via analysis of the peak positions of the respective molecular sensors. Typically, the NMR spectra and all respective peaks are fitted via an optimization procedure giving the peak positions and amplitudes. However, such line-fitting procedures are error-prone in cases of low signal-to-noise ratio (SNR) and peak overlap, e.g. for multiple pHe compartments within the kidneys [13]. In recent years, deep learning has shown its potential for magnetic resonance spectroscopy (MRS) and magnetic resonance spectroscopic imaging (MRSI) data in several applications to improve analysis of noisy data with interfering signals [19, 20]. Among these, artificial neural networks (ANN) demonstrated their value for spectroscopy analysis in medicine by classifying lung cancer tissue based on 1H MRS [21] or denoising of brain 1H MRS [22]. Furthermore, it was shown that convolutional neural networks (CNN) and multilayer perceptrons (MLP) can be trained to classify specific chemical compounds in various spectroscopy data sets [23, 24]. Nevertheless, we hypothesize that there is an advantage in applying a CNN for spectral analysis, as this class of network is invariant under frequency shifts of the entire spectrum which can be caused by B0 inhomogeneities.

We also hypothesize that transfer learning with real mice kidney data could improve the performance for our deep learning model. Transfer learning and domain adaptation have been used to adapt the model trained by one data distribution to the target data domain [25, 26], especially when the target domain data is limited [27]. Our target domain data, 13C-labelled zymonic acid kidney spectra, are by definition of the animal study and experimental efforts limited in size. Only one or two PRESS spectra or one CSI data set, still containing only a few single voxel spectra from kidneys, can be obtained from a single imaging experiment.

In this work, we investigate whether deep learning can improve the prediction of multiple pHe compartments from magnetic resonance data using hyperpolarized ZA. For this task, we evaluate the performance of a CNN compared to a MLP as well as to conventional line fitting on both a single type of data (synthetic) and real data adaptation (a mix of real and synthetic data). For deep learning evaluation, both real data using line fitting as a gold-standard for evaluation of pHe compartments as well as synthetic data with known pHe compartments are used.

Methods

Neural network architecture

We implemented a multi-output regression convolutional neural network (CNN) and a multilayer perceptron (MLP), as shown in Fig. 1. The neural networks learned to map nuclear magnetic resonance spectra to specific pHe values of a specific number of distinct pHe compartments, of which were three for our specific case of in vivo kidney data of healthy mice.

Fig. 1
figure 1

Schematic diagram of the neural network architecture for the a convolutional neural network (CNN) and the b multilayer perceptron (MLP). Both the proposed CNN and MLP consist of 4 feature extraction layers, for which they learnt a mapping between the input spectra and multiple pHe compartments. To compare the performance of CNN and MLP in correctly predicting pHe compartments, the architectures were set to a similar number of weights, ≈ 8000. The length of the spectrum or feature maps, which are used as the input to each next convolutional or dense layer, are shown in the square brackets. In CNN, the lengths are scaled logarithmically. CNN: The number of filters is 4, 4, 8, and 8. The sizes of the convolutional kernel are shown in the round brackets. MLP: The number of neurons is 16, 16, 32, and 32. Dense layers are represented with half (MLP) or quarter (CNN) the number of nodes, except for output layers

The proposed CNN (Fig. 1a) consists of 4 sequential hidden layers for feature extraction. Each layer consists of a 1D convolutional kernel, a rectified linear unit (ReLU) activation function, and max pooling. The input into each layer was first convolved with the sliding kernel with stride of 1 and with length of 300, 150, 50 and 20. The initial weight (kernel) was a random value drawn from a truncated normal distribution. The convolutional kernel length decreased along the layers for extracting sub-regional features [28] as the input was downsampled from 1024 to 512, 256, 128 and 64 due to max pooling. The number of filters for feature extraction increased from 4, 4, to 8 and 8. ReLU activation functions were used in each layer to provide sparsity and thus robustness to small changes in input such as noise [29], as noise is inevitably present within the acquired spectra. The pooling layer was used to reduce the tensor size which could potentially merge semantically similar features [28].

Dropout was applied in the last feature extraction layer as a regularization to prevent over-fitting [30], for which it randomly dropped out 10% of the weights during training. Before the output layer, these feature maps were then flattened and passed to a dense layer with hyperbolic tangent activation function [31], and the outputs were then mapped into the target pHe range 6.32 to 7.44.

The proposed MLP (Fig. 1b) consists of 4 fully connected dense layers, followed by max pooling. Each dense layer has a filter size 16, 16, 32 and 32, for which the output of each node is connected to all of the input nodes in the next layer. To compare their performance in mapping pHe compartments, the architecture of CNN and MLP were designed such that the number of weights in both neural networks were similar, at approximately 8000.

Modelling of NMR pHe spectra

Relative to the urea peak position, the chemical shifts ZA5 and ZA1 can be described as a function of pHe by the following scaled logistic function [13]:

$${\text{ZA}}_{i} \left( {{\text{pH}}_{{\text{e}}} } \right) = {\text{ZA}}_{i,\min } + \frac{{\delta_{i} }}{{1 + 10^{{\left( {{\text{pK}}_{{\text{a}}} - {\text{pH}}_{{\text{e}}} } \right)}} }}$$
(1)

where ZA5,min = 12.57 ppm, ZA1,min = 8.52 ppm, δ1 = 2.57 ppm, δ5 = 5.13 ppm and pKa = 6.90 [13]. Using a Lorentzian peak model, the spectrum can be described by the following equation:

$$x\left( f \right) = \mathop \sum \nolimits_{i}^{N} \frac{{a_{i} \cdot w_{i}^{2} }}{{w_{i}^{2} + \left( {x + {\text{ZA}}_{i} \left( {{\text{pH}}_{{\text{e}}} } \right)} \right)^{2} }}$$
(2)

where x(f) is the spectrum, N is the number of peaks (in our case n = 7; 6 zymonic acid peaks and one urea peak), \(a_{i}\) is the peak amplitude, \(w_{i}\) is the full width at half maximum, and ZAi(pHe) is the corresponding chemical shifts of ZA5 and ZA1 peaks found by Eq. 1. For the 13C-urea peak, ZAi(pHe) is set to zero.

Hyperpolarized 13C magnetic resonance spectroscopy

Hyperpolarization

27 mg [1,5-13C2,3,6,6,6-D4]zymonic acid [14], 1.7 mg Ox063 trityl radical (GE Healthcare, Chicago, IL, USA) and 24 µL DMSO were vortexed for 35 min. The mixture was added to a DNP sample cup and frozen in liquid nitrogen. 25 µL of a sample, containing 10 M 13C-urea in glycerol, 30 mM Ox063 and 1.5 mM DOTAREM (Guerbet, Villepinte, France) was subsequently added on top of the frozen layer and also frozen in liquid nitrogen. The sample cup was then polarized at 1.2 K for three hours by irradiation with a microwave frequency of 94.169 GHz and a power of 100 mW using a HyperSense® DNP Polarizer (Oxford Instruments, Abingdon, UK). Dissolution was performed with 2.99 ml D2O containing 80 mM TRIS, 0.3 mM EDTA and 50 mM NaOH, resulting in solutions containing 50 mM hyperpolarized zymonic acid and 100 mM urea with a pH 6.7 ± 0.4.

Hyperpolarized 13C-magnetic resonance spectroscopy

All experiments were performed on a horizontal bore small animal 7 T magnet MRI scanner (Agilent/GE) MR901 with Bruker AVANCE III HD electronics and a 31 mm 1H/13C volume resonator (RAPID Biomedical, Rimpar, Germany). Experiments were performed on seven healthy C57BL/6 mice (Charles River, Wilmington, MA, USA) in accordance with pertinent laws and regulations and approved by an ethical review board (Regierung von Oberbayern, Munich, Germany, Approval Number ROB-55.2-2532.Vet_02-17-177). Prior to imaging, animals were anaesthetized with 1.5–2.5% Isoflurane (v/v) in oxygen as a carrier gas, a tail vain catheter was placed and animals were positioned together with a 4.4 M [1-13C]lactate phantom for B1 calibration inside the magnet. Breathing rate (40 ± 7 min−1) and animal temperature (37.0 ± 0.6 °C) were monitored with an ECG trigger unit (RAPID Biomedical) and an MR-compatible temperature monitoring system Model 1030 (SA Instruments Inc, Stony Brook, New York, NY, USA) respectively. Kidneys were located using 1H RARE with FOV 32 × 32 mm2, slice thickness 1 mm, matrix size 128 × 128, repetition time 4000 ms, effective echo time 48 ms, RARE factor 12, 10 averages. Following manual B1 calibration using 13C-FID acquisitions with non-selective excitation by a 1 ms block pulse of varying RF power and fitting of the resulting signal vs. excitation power curves, the hyperpolarized solution was injected and 13C acquisitions started 5 s after end of injection. Single voxel 13C spectroscopy used PRESS on single kidneys with typical parameters: Total echo time 13.9 ms, total scan time 531 ms, excitation pulse flip angle 90°, refocusing pulse flip angle 180°, receive bandwidth 2000 Hz, 1024 points, voxel size 5 × 5 × 7 mm. Hyperpolarized 13C-MRSI was performed using FIDCSI with typical parameters: total scan time 14 s, FOV 28 × 24 mm2, slice thickness 5 mm, matrix size 14 × 12, repetition time 83.1 ms, flip angle 15°, receive bandwidth 3200 Hz, 256 spectral points. Overall, nine CSI data sets and six PRESS data sets were acquired and used for training and testing.

Data analysis and conventional line fitting

All data processing was performed in MatLab (The Mathworks Inc., Natick, MA, USA). For PRESS acquisitions, spectra were line-broadened by 5 Hz and phased manually. For CSI acquisitions, no line-broadening was applied, and magnitude spectra were averaged across both kidneys. C1- and C5-peaks of zymonic acid and of urea were identified by a standard automatic peak picking algorithm in MatLab for each pHe compartment and selection was inspected manually. Peaks were fitted according to the model described in Eq. 2 where peak height, position relative to urea, and a uniform peak width for all compartments were fitted as free parameters using a built-in non-linear least squares algorithm. Second, the corresponding pHe value was fitted from the relative zymonic acid peak distance to urea according to Eq. 1. Pairs of zymonic acid peaks were grouped into pH compartments and for each detected pHe compartment, a mean pHe was calculated which was weighted by the signal intensities from both the C1- and the C5-peaks. Common values for peak linewidths in Hz and ratios of signal amplitudes of the different kidney compartments were also extracted for generation of synthetic spectra for the training of the neural networks.

Evaluation and training of the neural networks

Data sets

Spectra synthetization

Due both to the fact that in vivo experiments are necessarily small in sample size for ethical reasons, and that in vivo hyperpolarized 13C experiments are labour intensive, synthetic data was generated for the purpose of training the neural networks. To avoid over-fitting, noise was included in the spectral synthetization and was performed based on the following model:

$$X\left( f \right) = b_{0} + \varepsilon + S \cdot x\left( f \right)$$
(3)

where X(f) is the synthetic spectrum, b0 is a constant baseline, \(\varepsilon\) is the additive noise, S is the SNR scale factor, x(f) is the set of Lorentzian peaks for the 3 metabolite peaks: urea, ZA5 and ZA1. Figure 2a shows the distribution of the 3 pHe compartment values of the synthetic spectra, which were initiated from a normal distribution respectively in the ranges: 7.33–7.44; 6.96–7.15; 6.32–6.78, which were found from the measurements carried out according to “Hyperpolarized 13C-magnetic resonance spectroscopy” section. The urea peak positions and widths were initiated from a normal distribution with a standard deviation of 0.580 ppm starting from 163 ppm, to represent potential B0 inhomogeneities and the peak widths were initiated between 30 and 70 Hz (0.397–0.927 ppm) to account for variations in shim quality. The corresponding chemical shifts ZA5 and ZA1 peaks were then found by Eq. 2 using the Lorentzian peak model, and a set of basis peaks for the 3 metabolites peaks urea, ZA5 and ZA1 was then generated (Fig. 2b). The ratios for the urea, ZA5 and ZA1 peak amplitudes were set to be 4:1:2, which represents the in vivo signal ratios of zymonic acid peaks detected in the different pHe compartments compared to urea. Gaussian noise and baseline were added to the set of combined basis peaks to take the noise from MRS acquisition into account. The range of SNR scale factor was 2 to 7 and baseline was set from − 0.2 to 0.2, both drawn from the normal distribution (Fig. 2c). 10,020 spectra were synthesized, 10,000 spectra were used for training and 20 spectra were used for testing.

Fig. 2
figure 2

Spectra synthetization was performed to generate three pHe compartment spectra for the CNN and MLP training data. a The pHe values for each of the three compartments (7.33–7.44; 6.96–7.15; 6.32–6.78) are normally distributed. b An example of the generated spectrum by 3 compartment pHe values. c Signal-to-noise (SNR) ratios are applied ranging from 2 to 7 for the synthetic spectra. For simplification, only SNR 2, 5, and 7 are shown. Kidney spectra were augmented by applying five-scale Gaussian denoising to increase the real training data size. d Example of the original and denoised spectra. For simplicity, only the first and the fifth scale-denoised spectra are shown. An enlarged version of the spectra (green box) is shown on the right

Kidney data and augmentation

To adapt the neural networks to our target domain of mice kidney spectra, we generated 40 augmented spectra as training data based on the eight acquired mice kidney spectra by applying a five-scale Gaussian denoising (scale factors 1.5, 1.2, 1, 0.8, 0.5) (Fig. 2d). The original eight spectra were used for testing.

Experiments

We set up four experiments to evaluate the performance of CNN and MLP on mixed training data: CNNmix, MLPmix, and single type of training data: CNNsyn, and MLPsyn. A total of 10,000 spectra were used (Training: 8500, Validation: 1500). CNNmix and MLPmix consisted of 9960 synthetic spectra, 40 augmented kidney data from PRESS and ROI-averaged CSI acquisitions, whereas CNNsyn and MLPsyn consisted of 10,000 synthetic datasets. All augmented spectra were set to size of 1024, the magnitude of all spectra was normalized between 0 and 1 for training and testing.

A total of 28 spectra were used for testing, which included 20 augmented spectra and eight real mice kidney spectra. Generation of augmented spectra and pre-processing of in vivo kidney spectra was implemented in MatLab.

Network training

Both the CNN and the MLP were trained with a batch size of 200 spectra and 400 epochs. Training progress was achieved by minimizing the sum of mean-square error loss of three pHe compartments: \(L = L_{{\text{cortex }}} + L_{{\text{medulla }}} + L_{{{\text{ureter}}}}\), where Li is the L2 norm of the difference between the predicted pHe compartment from the input spectra, and the ground truth pHe compartment. The loss was then back-propagated for updating the weight kernel for each layer using NADAM (Nesterov-accelerated Adaptive Moment Estimation) optimizer [32]. Both networks were implemented in Keras using TensorFlow as the backend [33]. The training time for both neural networks was approximately five minutes, both training and testing were performed on a NVIDIA Tesla P100 GPU.

Synthesis of line fitted pH values and neural network predicted pH values into a combined pH map

Based on the network performance results from tests on synthetic and real kidney data, the best performing network is chosen for neural-network-assisted improvement of pH mapping in healthy mice kidneys. For this purpose, supervised line fitting was performed voxel-wise on seven CSI data sets from four mice for which the correct number of fitted compartments and fit quality was assessed. Spectra which were fitted with three pH compartments were extracted. For each image, a segmentation mask was created, indicating voxels either containing three pH compartments (“1”) or less than three compartments (“0”; corresponding to 0, 1 or 2 pH compartments), and fed voxel-wise into the best-performing neural network. pH maps for each CSI data set were then generated where the segmentation mask-positive area pH values were replaced by the predicted pH values from the network for the respective compartment, resulting in hybrid pH maps which are composed of pH values either based on line fitting or neural network predictions. The corresponding mean pH maps were calculated by averaging all compartmental pH values.

Results

13C spectroscopic data

Spectra from 13C-acquisitions of healthy kidney (top “Input” in Fig. 1) show the urea peak (164 ppm), six zymonic acid peaks (173–178 ppm) and the C5-peak of parapyruvate-hydrate (179 ppm), a decay product of zymonic acid. For zymonic acid, three pairs of C1 and C5 peaks can be grouped unambiguously to a single pHe compartment, and each compartment corresponds respectively to the three anatomical regions of the kidney, namely the cortex (red stars), the medulla (green diamonds) and the ureter (blue triangles). Multiple measurements on three mice return consistent pHe values for the cortex (pHe = 7.38 ± 0.03, n = 13), the medulla (pHe = 7.06 ± 0.06, n = 11) and the ureter (pHe = 6.53 ± 0.16, n = 9).

Network training

The training losses of CNNmix, CNNsyn, MLPmix and MLPsyn over 400 epochs are shown in Fig. 3. While the CNNs rapidly converge to their respective limit, the MLPs’ minimal loss after 400 epochs remains higher compared to the CNNs, having not yet reached a converging limit. Interestingly, for both networks, this behaviour is independent of the training data set.

Fig. 3
figure 3

The training losses for four models: CNNmix, CNNsyn, MLPmix and MLPsyn, stably converge to a minimal loss. While the CNNs achieve absolute minimum loss after 400 epochs, the MLPs remain at a considerably higher loss having not yet reached their absolute minimum at this training stage

Comparison of performance of neural networks on augmented test data and conventional line fitting

To validate the network outputs and perform a comparison of the trained networks and conventional line fitting performed by an expert MRS scientist, 20 synthetically generated spectra with known ground truth pHe compartments were blind-fitted by the conventional line fitting and pH analysis routine. In cases of sufficient pH difference between compartments together with sufficient SNR, all peaks can be detected reliably with a high accuracy of the predicted pHe values (Fig. 4a).

Fig. 4
figure 4

a Conventional line fitting of a synthetic kidney spectrum with three fitted pHe compartments indicated by coloured markers (compartment 1: red stars, compartment 2: green diamonds, compartment 3: blue triangles) and comparison of calculated pHe values and ground truth data is shown as inset resulting in good agreement. b Conventional line fitting on a noisy spectrum and low compartment intensities of the medulla and the ureter. Three compartments can only be partially detected and agreement with ground truth is rather poor

In cases of low SNR and low intensity for one or multiple compartments relative to a third one, the conventional line fitting only partially allows for detection of pHe compartments with less accuracy when comparing to the ground truth pH values (Fig. 4b). In the next step, the same twenty spectra were analyzed with the two CNN models CNNmix and CNNsyn and the two MLP models MLPmix and MLPsyn which were all previously trained either with a mixed (augmented + synthetic spectra) or with synthetic spectra only. The performance of the conventional line fitting and the four network models on the synthetic spectra test set relative to each other was compared by linear regression of the predicted and the ground truth pH value for each compartment. The adjusted coefficient of determination R2 and slopes of the linear regression β to compare statistical and systematic uncertainties respectively are listed in Table 1. The dominant pHe compartment of the cortex is best predicted by the CNNmix (β = 1.01), however with greater uncertainty (R2 = 0.78) compared to conventional line fitting (R2 = 0.85) or CNNsyn (R2 = 0.90). The pHe compartment of the medulla is best and equally well predicted by CNNmix and CNNsyn compared to all other methods, however with greater systematic uncertainty (βCNNmix = 1.30, βCNNsyn = 1.26) and lower scattering (RCNNmix2 = 0.91, RCNNsyn2 = 0.92) compared to the cortex. Interestingly, for the pHe compartment of the ureter, line fitting achieves equivalent statistical uncertainty (R2 = 0.99) compared to the CNNs (RCNNmix2 = 0.98, RCNNsyn2 = 0.99) while outperforming them regarding systematic error (βFit = 1.02). The performance of the MLPs is very poor for the cortex and the medulla compartment and only modest for the ureter compartment. This observation of poor MLP performance is also visualized in the modified Bland–Altman plots (Fig. 5), indicating that the MLPs have deviations of more than 0.2 pH units for the ureter compartments for some spectra. Furthermore, the good performance of both trained CNNs for all compartments is verified while for the conventional line fitting some systematic underestimation of the cortex pHe can be observed.

Table 1 Top: Evaluation of the prediction accuracy by compartment-wise comparison of the adjusted coefficient of determination R2 derived from a linear regression of ground truth and pH values predicted by conventional line fitting (model “Modelling of NMR pHe spectra” and “Data analysis and conventional line fitting” sections) and the neural networks after application to 20 synthetic test spectra; Bottom: Linear slope coefficients β derived from linear regressions to evaluate prediction bias
Fig. 5
figure 5

Modified Bland–Altman plots showing the difference between predicted and ground truth pH values from synthetic kidney test spectra against the ground truth pH for each compartment. Black dashed lines indicate the mean difference and grey dotted lines indicate the 95% confidence interval for this deviation

Comparison of performance of neural networks on real kidney data

To evaluate the suitability for routine spectroscopic data analysis, all four networks were tested on eight real mice kidney spectra. A comparison to the values obtained by conventional line fitting as a pseudo ground truth is visualized with Bland–Altman plots (Fig. 6). Following the trend observed for the synthetic test data, the CNNmix network outperforms all other networks, showing the predicted pHe values to deviate less than 0.1 pH units from the fitted data for all compartments. In contrast to testing on synthetic spectra, a CNN network being trained only with synthetic data CNNsyn, shows only poor performance when tested on real data with predicted pHe values deviating up to 0.3 pH units from conventional fit values and decreasing performance from compartment 1 (cortex) to compartment 3 (ureter). For compartment 2 (medulla) and 3 (ureter), MLPmix also achieves better agreement with fitted pHe values. Analogous to testing on synthetic test spectra, MLPsyn shows the worst agreement with conventional line fitting, exceeding 0.1 pH units mean difference for compartment 3 (ureter).

Fig. 6
figure 6

Conventional Bland–Altman plots showing the difference between predicted and conventionally fitted pHe values from spectra of hyperpolarized 13C acquisitions on mice kidney plotted against the mean pHe from both analysis methods. Black dashed lines indicate the mean difference and grey dotted lines indicate the 95% confidence interval for this deviation

Hybrid pH mapping by voxel-wise combination of pH values reconstructed from CNNmix and line fitting

Based on the performance measurements for synthetic and real kidney spectra, CNNmix was chosen for application in pH mapping of healthy mice kidney (Fig. 7a). An exemplary segmentation mask for CSI data matching the anatomy in Fig. 7a based on supervised line fitting to distinguish voxels with three pH compartments and voxels with less than three pH compartments is shown in Fig. 7b (white: three pH compartments, black: less than three pH compartments).

Fig. 7
figure 7

a Anatomical axial T2-weighted image of mice kidney encircled with white ROIs. For shim improvements, the mouse, as well as a [1-13C]lactate-phantom for B1 calibration are covered with carbomer gel. b Segmentation mask for a 14 × 12 CSI data sets acquired on the anatomy in a. White areas indicate voxels corresponding to spectra with three pH compartments, black areas indicate voxels of which spectra contain two or less compartments. c Top row: Individual pH compartment maps for the cortex or a physiological pH compartment (compartment 1), a slightly acidified compartment, mainly the medulla (compartment 2) and a strongly acidified compartment of the ureter (compartment 3) as derived from supervised line fitting. The mean pH map represents the un-weighted average of all three pH compartment maps. Bottom row: Compartment maps derived from line fitting where white areas in the segmentation mask have been replaced by voxel-wise predicted pH values from CNNmix. The mean pH map displays the average pH value of the respective number of pH compartments

pH mapping based on supervised line fitting (Fig. 7c, top row) reveals a globally present physiological pH compartment (top left), a heterogeneous, slightly acidic second pH compartment which can be attributed to the medulla (top middle-left) and a third pH compartment corresponding to the ureter (top middle-right). Voxel-wise compartment-averaging generates kidney-specific pH contrast. Substitution of line fitted pH values by values predicted by CNNmix shows slightly more basic pH values predicted by the neural network compared to the line fitting. For the medulla compartment, network predictions appear to be more homogeneous compared to line fitted maps. For the ureter compartment, line fitted as well as neural network predicted maps agree well with each other. The mean pH map based on these hybrid compartment pH maps shows good inter- and intra-kidney homogeneity in mean pH values compared to line fitted maps.

For quantitative comparison, pH compartments derived from line fitting and neural networks were both averaged across individual kidneys for multiple acquisitions on different animals (Fig. 8). pH values derived from line fitting show lowest inter- and intra-subject variation for the cortex (pHcortex,fit = 7.41 ± 0.02, n = 14) while pH values for the medulla (pHmedulla,fit = 7.09 ± 0.10, n = 14) and the ureter (pHureter,fit = 6.70 ± 0.13, n = 14) are distributed across larger pH ranges while all compartments can be well separated from each other based on pH. pH compartments predicted by the CNNmix agree well with line-fitted compartments, despite the cortex (pHcortex,CNN = 7.43 ± 0.01, n = 14) and the medulla (pHmedulla,CNN = 7.13 ± 0.04, n = 14) exhibiting overall slightly more basic pH values compared to the line-fitted ones. For the ureter, no relevant difference can be observed (pHureter,CNN = 6.72 ± 0.04, n = 14). In agreement with lower intra-subject variations as seen in compartment maps in Fig. 7, the inter- and intra-subject variations of compartment pH values are lower for the neural-network-predicted pH values, while the values for each subject are in good agreement relative to the compartment-specific standard deviation (black crosses are corresponding to the same kidney in the same subject in Fig. 8).

Fig. 8
figure 8

pH compartments averaged over regions of interests in single kidneys of mice. pH values derived from line fitting show stronger variations between single kidneys and different subjects compared to predictions from CNNmix.pH values for each compartment show good agreement between line fitting and neural network predictions for individual kidneys (example indicated by black crosses)

Discussion

Modelling and fitting of hyperpolarized 13C spectroscopic data

Analogous to published data on spectra of zymonic acid in kidney of healthy rats [13], several pH compartments can be detected in hyperpolarized 13C acquisitions of zymonic acid in healthy kidney of mice. However, out of three detected pHe compartments which could be assigned to cortex, medulla and ureter, the latter two often suffer from lower compartment signal due to limited perfusion of the agent within the short acquisition time [13] or enhanced signal decay in case of injury or pathological alterations [34,35,36]. Consequently, line fitting with conventional methods becomes challenging when trying to resolve the pHe compartments of the kidney. For these reasons, manual spectral pre-processing together with fitting of a linear combination of individual Lorentzian peaks while minimizing the free parameters to amplitude, peak position, and peak width (FWHM) was chosen as an appropriate fitting routine despite inherent SNR limitations. For these reasons, manual spectral pre-processing together with fitting of a linear combination of individual Lorentzian peaks while minimizing the free parameters to amplitude, peak position, and peak width (FWHM) was chosen as an appropriate fitting routine despite inherent SNR limitations. Additionally, other standard fitting routines for MR spectra such as LCModel [37] or AMARES [38], were either unsuitable, unstable, or of no significant benefit in this work.

For instance, for AMARES it was already shown that this algorithm suffers from unstable fitting when the peak frequencies are allowed as free parameters [39]. In addition, zymonic acid spectra on kidney are less sparse than the previously fitted pyruvate- and lactate-containing spectra.

LCModel is a standard fitting routine for magnetic resonance spectra which allows excellent peak quantification for 1H metabolites. Nevertheless, this method also has several limitations regarding the application to the data of this work. LCModel predominantly aims to quantify spectral peaks which requires the input of a set of basis spectra of high spectral quality and good SNR. In such cases, the peak positions are fixed, and only minor peak shifts due to eddy currents and magnetic field inhomogeneities are tolerated. This contrasts with hyperpolarized 13C acquisitions using zymonic acid in which SNR is typically modest, spectral resolution limited and quantification not necessary. As zymonic acid peaks strongly shift with pHe, a suitable set of basis spectra would require multiple zymonic acid spectra at different pHe values which either requires a high amount of basis components or an inherent limitation in the measured pH accuracy limited by a small basis set. In addition, since C1- and C5-peak intensities vary relative to each other in different acquisitions, combined modelling as one basis spectrum for a fixed pH is difficult. Furthermore, LCModel requires well separated peaks for proper differentiation which is not the case for the densely packed pH compartment peaks as seen in Fig. 1.

CNN and MLP performance

In our study, both CNNmix and CNNsyn outperform MLPmix and MLPsyn in predicting all three pH compartments in synthetic test data. Here, the CNNs have better accuracy and less uncertainty as shown in the regression analysis. Interestingly, for the cortex and medulla compartments, CNNmix was giving a better prediction accuracy as compared to the conventional line fitting method. While the CNN and the MLP have a similar number of weights (≈ 8000), the CNN used kernels in convolutional layers to perform elementwise multiplications to inputs while the MLP used densely connected neurons. When applied systematically across the entire input spectra, these convolutional kernels could extract spectral features such as the metabolite peaks distances, as the kernels account for the values on neighboring pixels. Because of the weight sharing that occurs when the convolutional kernels slide across the spectrum [40, 41], the CNN becomes less susceptible than the MLP to spectral variance or drifts in spectral peak positions which can be caused by B0 inhomogeneities. However, the choice of specific neural network depends on the type of learning tasks and features to be extracted, as previous studies showed MLP performed well in classification [42, 43], while CNN also demonstrated good performance in image segmentation [44] or classifications [45]. Some even explore the synergies of MLP and CNN networks [46, 47].

Performance of neural networks trained with mixed data

We showed that it is possible to train models on a limited amount of real data by transfer learning, whilst most of the training data were synthetically generated based on a spectral model for [1,5-13C2]zymonic acid and 13C-urea. When tested with synthetic data, CNNsyn performed better than CNNmix in predicting medulla and ureter pH as shown in the linear regression analysis and modified Bland–Altman plots. Especially for the medulla compartment, both line fitting and MLPsyn and MLPmix show poor performance what might be due to higher sensitivity to low SNR. As there is no absolute ground truth for the kidney in healthy mice, we compared the neural network predictions with the results in line fitting as a pseudo ground truth. We found CNNmix had the most consistent and comparable results to line fitting, as it has the smallest difference compared to other models, and its 95% confidence level is also smaller than CNNsyn in cortex and ureter (see Fig. 5). Moreover, MLPmix are more comparable to line fitting than CNNsyn, and MLPmix has generally smaller mean difference.

Hybrid pH map generation by combination of line fitting and CNNmix predictions

Combination of line fitting and neural networks appears to improve pH mapping in kidneys of healthy mice. Based on compartmental pH maps, line fitting appears to be only quantitatively robust for the cortex while the medulla and the ureter show considerable inter- and intra-kidney pH variability, the latter being physiologically rather unreasonable. We assume that this high variability stems from the low SNR of zymonic acid peaks corresponding to these two compartments. Substitution of voxels in pH maps corresponding to spectra containing three pH compartments by neural network predictions results in more homogeneous compartment maps while quantitatively still agreeing with line-fitted compartments. This suggests a superior performance of the neural network compared to the line fitting approach for low SNR compartments. Furthermore, as the network is predicting pH compartments voxel-wise based on individual spectra, it has to be pointed out that the predictions of neighboring voxels are independent from each other and therefore the observed spatial homogeneity of compartments therefore indicates a good robustness of the pH predictions. In addition, high quantitative prediction accuracy is suggested by the observation that cortex compartments are systematically predicted with a higher pH value compared to the line fitting method, which agrees with the observation that line fitting systematically underestimates the cortex pH when evaluating the method performance for artificial spectra of known pH.

Data size

These observations suggest that the real and augmented data might consist of spatially independent features, such as the noise during the spectra acquisition, which is crucial to train a more accurate model. Neural networks usually require a large amount of training data, and the number of training data depends on the complexity of the tasks and features to extract. However, generating a larger data set is challenging for hyperpolarized 13C MRSI. In vivo spectra obtained by preclinical studies are limited in size for ethical reasons: the number of animals should be as low as possible. Additionally, the experimental efforts are rather large. Also, with regard to application for hyperpolarized 13C acquisitions in humans, data set size is critical as clinical trials currently performed with this imaging technique are typically limited to 5–100 patients [48]. Efforts to obtain larger amounts of data might involve the generation of databases but, especially for imaging using hyperpolarized 13C-labelled zymonic acid, this is at an early stage. Nevertheless, in our study, we showed improved network performance by including less than 0.5% real augmented data (40 augmented spectra out of 10,000 training spectra), an amount that can be realistically generated from single preclinical studies.

Future works

In this study, we found that the convolutional layers enable the network to better extract spectral features in the spectra. Future works could extend the application of convolutional layers to denoise the spectra or automating peak picking. Moreover, the neural networks here only predict the chemical shift of the spectra—they do not yet consider the signal intensity as in the conventional line fitting method. An extension the current approach might also predict signal intensity, which could then allow a more direct calculation of weighted-average pH maps. In addition, for imaging of cancer or unknown tissue, networks could be trained to predict the correct number of pH compartments and using this information to selectively pass the spectra through other networks which predict the correct pH values.

Conclusion

Two different types of neural networks trained once with a fully synthetic data set and once with a mixed data set, containing real and synthetic data, were each evaluated for prediction of pH compartments from hyperpolarized 13C acquisitions of zymonic acid on kidney in healthy mice. CNNs trained with a mixed set of augmented and synthetic spectra show the ability to accurately predict multiple pH compartments in hyperpolarized 13C spectra. This network achieves the best results out of all tested networks and its performance competes with or outperforms conventional line fitting being supervised by humans. The trained network can be used to improve pH mapping by segmentation-based substitution of line fitted pH values by neural network predictions. Therefore, small amounts of experimental data and appropriate neural network and training method choice allows fast, accurate, and reliable evaluation of hyperpolarized 13C magnetic resonance spectroscopic acquisitions for pH measurements in kidney. Using appropriate training data sets and slightly modified output layers of the networks to account for different amounts of detected pH compartments, the presented concept could potentially be applied to other organs or tumours.

Availability of data and materials

The data sets used and analyzed during the current study are available at https://github.com/ryanayf/KidNeYronal

References

  1. Gillies RJ, Raghunand N, Garcia-Martin ML, Gatenby RA. pH imaging. A review of pH measurement methods and applications in cancers. IEEE Eng Med Biol Mag. 2004;23:57–64. https://doi.org/10.1109/MEMB.2004.1360409.

    Article  PubMed  Google Scholar 

  2. Adrogue HJ, Madias NE. Management of life-threatening acid-base disorders. Second of two parts. N Engl J Med. 1998;338:107–11. https://doi.org/10.1056/NEJM199801083380207.

    Article  CAS  PubMed  Google Scholar 

  3. Gallagher FA, Kettunen MI, Day SE, Hu DE, Ardenkjaer-Larsen JH, Zandt R, Jensen PR, Karlsson M, Golman K, Lerche MH, et al. Magnetic resonance imaging of pH in vivo using hyperpolarized 13C-labelled bicarbonate. Nature. 2008;453:940–3. https://doi.org/10.1038/nature07017.

    Article  CAS  PubMed  Google Scholar 

  4. Hashim AI, Zhang X, Wojtkowiak JW, Martinez GV, Gillies RJ. Imaging pH and metastasis. NMR Biomed. 2011;24:582–91. https://doi.org/10.1002/nbm.1644.

    Article  PubMed  PubMed Central  Google Scholar 

  5. Parks SK, Chiche J, Pouyssegur J. Disrupting proton dynamics and energy metabolism for cancer therapy. Nat Rev Cancer. 2013;13:611–23. https://doi.org/10.1038/nrc3579.

    Article  CAS  PubMed  Google Scholar 

  6. Aime S, Castelli DD, Terreno E. Novel pH-reporter MRI contrast agents. Angew Chem. 2002;41:4334–6. https://doi.org/10.1002/1521-3773(20021115)41:22%3c4334::AID-ANIE4334%3e3.0.CO;2-1.

    Article  CAS  Google Scholar 

  7. Ardenkjaer-Larsen JH, Fridlund B, Gram A, Hansson G, Hansson L, Lerche MH, Servin R, Thaning M, Golman K. Increase in signal-to-noise ratio of > 10,000 times in liquid-state NMR. Proc Natl Acad Sci USA. 2003;100:10158–63. https://doi.org/10.1073/pnas.1733835100.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  8. Kurhanewicz J, Vigneron DB, Ardenkjaer-Larsen JH, Bankson JA, Brindle K, Cunningham CH, Gallagher FA, Keshari KR, Kjaer A, Laustsen C, et al. Hyperpolarized 13C MRI: path to clinical translation in oncology. Neoplasia. 2019;21:1–16. https://doi.org/10.1016/j.neo.2018.09.006.

    Article  PubMed  Google Scholar 

  9. Cunningham CH, Lau JY, Chen AP, Geraghty BJ, Perks WJ, Roifman I, Wright GA, Connelly KA. Hyperpolarized 13C metabolic MRI of the human heart: initial experience. Circ Res. 2016;119:1177–82. https://doi.org/10.1161/CIRCRESAHA.116.309769.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  10. Grist JT, Miller JJ, Zaccagna F, McLean MA, Riemer F, Matys T, Tyler DJ, Laustsen C, Coles AJ, Gallagher FA. Hyperpolarized 13C MRI: a novel approach for probing cerebral metabolism in health and neurological disease. J Cereb Blood Flow Metab. 2020;40:1137–47. https://doi.org/10.1177/0271678X20909045.

    Article  PubMed  PubMed Central  Google Scholar 

  11. Hundshammer C, Düwel S, Schilling F. Imaging of extracellular pH using hyperpolarized molecules. Isr J Chem. 2017;57:788–99. https://doi.org/10.1002/ijch.201700017.

    Article  CAS  Google Scholar 

  12. Korenchan DE, Flavell RR, Baligand C, Sriram R, Neumann K, Sukumar S, VanBrocklin H, Vigneron DB, Wilson DM, Kurhanewicz J. Dynamic nuclear polarization of biocompatible 13C-enriched carbonates for in vivo pH imaging. Chem Commun (Camb). 2016;52:3030–3. https://doi.org/10.1039/c5cc09724j.

    Article  CAS  Google Scholar 

  13. Düwel S, Hundshammer C, Gersch M, Feuerecker B, Steiger K, Buck A, Walch A, Haase A, Glaser SJ, Schwaiger M, et al. Imaging of pH in vivo using hyperpolarized 13C-labelled zymonic acid. Nat Commun. 2017;8:15126. https://doi.org/10.1038/ncomms15126.

    Article  PubMed  PubMed Central  Google Scholar 

  14. Hundshammer C, Duwel S, Kocher SS, Gersch M, Feuerecker B, Scheurer C, Haase A, Glaser SJ, Schwaiger M, Schilling F. Deuteration of hyperpolarized 13C-labeled zymonic acid enables sensitivity-enhanced dynamic MRI of pH. ChemPhysChem. 2017;18:2422–5. https://doi.org/10.1002/cphc.201700779.

    Article  CAS  PubMed  Google Scholar 

  15. Suh EH, Park JM, Lumata L, Sherry AD, Kovacs Z. Hyperpolarized 15N-labeled, deuterated tris(2-pyridylmethyl)amine as an MRI sensor of freely available Zn2+. Commun Chem. 2020. https://doi.org/10.1038/s42004-020-00426-6.

    Article  PubMed  PubMed Central  Google Scholar 

  16. Mishra A, Pariani G, Oerther T, Schwaiger M, Westmeyer GG. Near-infrared photoacoustic imaging probe responsive to calcium. Anal Chem. 2016;88:10790–4. https://doi.org/10.1021/acs.analchem.6b03546.

    Article  CAS  PubMed  Google Scholar 

  17. Schilling F, Schroder L, Palaniappan KK, Zapf S, Wemmer DE, Pines A. MRI thermometry based on encapsulated hyperpolarized Xenon. ChemPhysChem. 2010;11:3529–33. https://doi.org/10.1002/cphc.201000507.

    Article  CAS  PubMed  Google Scholar 

  18. Keshari KR, Kurhanewicz J, Macdonald JM, Wilson DM. Generating contrast in hyperpolarized 13C MRI using ligand-receptor interactions. Analyst. 2012;137:3427–9. https://doi.org/10.1039/c2an35406c.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  19. Chen D, Wang Z, Guo D, Orekhov V, Qu X. Review and prospect: deep learning in nuclear magnetic resonance spectroscopy. Chemistry. 2020;26:10391–401. https://doi.org/10.1002/chem.202000246.

    Article  CAS  PubMed  Google Scholar 

  20. Das D, Coello E, Schulte RF, Menze BH. Quantification of Metabolites in Magnetic Resonance Spectroscopic Imaging Using Machine Learning. In: Medical image computing and computer assisted intervention—MICCAI 2017. Springer Singapore, 2017; p 462–70.

  21. Brougham DF, Ivanova G, Gottschalk M, Collins DM, Eustace AJ, O’Connor R, Havel J. Artificial neural networks for classification in metabolomic studies of whole cells using 1H nuclear magnetic resonance. J Biomed Biotechnol. 2011;2011: 158094. https://doi.org/10.1155/2011/158094.

    Article  CAS  PubMed  Google Scholar 

  22. Lee HH, Kim H. Intact metabolite spectrum mining by deep learning in proton magnetic resonance spectroscopy of the brain. Magn Reson Med. 2019;82:33–48. https://doi.org/10.1002/mrm.27727.

    Article  CAS  PubMed  Google Scholar 

  23. Chatzidakis M, Botton GA. Towards calibration-invariant spectroscopy using deep learning. Sci Rep. 2019;9:2126. https://doi.org/10.1038/s41598-019-38482-1.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  24. Ghosh K, Stuke A, Todorovic M, Jorgensen PB, Schmidt MN, Vehtari A, Rinke P. Deep learning spectroscopy: neural networks for molecular excitation spectra. Adv Sci (Weinh). 2019;6:1801367. https://doi.org/10.1002/advs.201801367.

    Article  CAS  Google Scholar 

  25. Pan SJ, Yang Q. A survey on transfer learning. IEEE Trans Knowl Data Eng. 2010;22:1345–59. https://doi.org/10.1109/tkde.2009.191.

    Article  Google Scholar 

  26. Guan H, Mingxia L. Domain adaptation for medical image analysis. a survey. 2021; arXiv:2102.09508.

  27. Inoue T, Chaudhury S, De Magistris G, Dasgupta S. Transfer Learning from Synthetic to real images using variational Autoencoders for precise position detection. In: Proceedings of 25th IEEE international conference on image processing (ICIP), pp 2725–29.

  28. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521:436–44. https://doi.org/10.1038/nature14539.

    Article  CAS  PubMed  Google Scholar 

  29. Glorot X, Bordes A, Bengio Y. Deep sparse rectifier neural networks. In: Proceedings of fourteenth international conference on artificial intelligence and statistics, p 315–23

  30. Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting. J Mach Learn Res. 2014;15:1929–58.

    Google Scholar 

  31. LeCun Y, Bottou L, Orr GB, Müller KR. Efficient BackProp. In: Neural networks: tricks of the trade. Berlin: Springer; 1998. p. 9–51.

    Chapter  Google Scholar 

  32. Dozat T. Incorporating Nesterov momentum into Adam. In: Proceedings of international conference on learning representations 2016—workshop track.

  33. Chollet F, et al. e. Keras. GitHub. https://github.com/fchollet/keras. Accessed 11 June 2021.

  34. Reed GD, von Morze C, Verkman AS, Koelsch BL, Chaumeil MM, Lustig M, Ronen SM, Bok RA, Sands JM, Larson PE, et al. Imaging renal urea handling in rats at millimeter resolution using hyperpolarized magnetic resonance relaxometry. Tomography. 2016;2:125–35. https://doi.org/10.18383/j.tom.2016.00127.

    Article  PubMed  PubMed Central  Google Scholar 

  35. Grist JT, Mariager CO, Qi H, Nielsen PM, Laustsen C. Detection of acute kidney injury with hyperpolarized [13C, 15N]Urea and multiexponential relaxation modeling. Magn Reson Med. 2020;84:943–9. https://doi.org/10.1002/mrm.28134.

    Article  CAS  PubMed  Google Scholar 

  36. Laustsen C, Stokholm Norlinger T, Christoffer Hansen D, Qi H, Mose Nielsen P, Bonde Bertelsen L, Henrik Ardenkjaer-Larsen J, Stodkilde Jorgensen H. Hyperpolarized 13C urea relaxation mechanism reveals renal changes in diabetic nephropathy. Magn Reson Med. 2016;75:515–8. https://doi.org/10.1002/mrm.26036.

    Article  CAS  PubMed  Google Scholar 

  37. Provencher SW. Estimation of metabolite concentrations from localized in vivo proton NMR spectra. Magn Reson Med. 1993;30:672–9. https://doi.org/10.1002/mrm.1910300604.

    Article  CAS  PubMed  Google Scholar 

  38. Vanhamme L, van den Boogaart A, Van Huffel S. Improved method for accurate and efficient quantification of MRS data with use of prior knowledge. J Magn Reson. 1997;129:35–43. https://doi.org/10.1006/jmre.1997.1244.

    Article  CAS  PubMed  Google Scholar 

  39. Topping GJ, Heid I, Trajkovic-Arsic M, Kritzner L, Grashei M, Hundshammer C, Aigner M, Skinner JG, Braren R, Schilling F. Hyperpolarized 13C spectroscopy with simple slice-and-frequency-selective excitation. Biomedicines. 2021. https://doi.org/10.3390/biomedicines9020121.

    Article  PubMed  PubMed Central  Google Scholar 

  40. LeCun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to document recognition. Proc IEEE. 1998;86:2278–324. https://doi.org/10.1109/5.726791.

    Article  Google Scholar 

  41. Yamashita R, Nishio M, Do RKG, Togashi K. Convolutional neural networks: an overview and application in radiology. Insights Imaging. 2018;9:611–29. https://doi.org/10.1007/s13244-018-0639-9.

    Article  PubMed  PubMed Central  Google Scholar 

  42. Yun J, Park JE, Lee H, Ham S, Kim N, Kim HS. Radiomic features and multilayer perceptron network classifier: a robust MRI classification strategy for distinguishing glioblastoma from primary central nervous system lymphoma. Sci Rep. 2019;9:5746. https://doi.org/10.1038/s41598-019-42276-w.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  43. Naraei P, Abhari A, Sadeghian A. Application of multilayer perceptron neural networks and support vector machines in classification of healthcare data. In: Proceedings of 2016 future technologies conference (FTC).

  44. Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. In: Medical image computing and computer-assisted intervention—MICCAI 2015. vol 9351. Cham: Springer; 2015. p 234–41.

  45. Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Commun ACM. 2017;60:84–90. https://doi.org/10.1145/3065386.

    Article  Google Scholar 

  46. Lai Z, Deng H. Medical image classification based on deep features extracted by deep model and statistic feature fusion with multilayer perceptron(). Comput Intell Neurosci. 2018;2018:2061516. https://doi.org/10.1155/2018/2061516.

    Article  PubMed  PubMed Central  Google Scholar 

  47. Ahsan MM, Alam TE, Trafalis T, Huebner P. Deep MLP-CNN model using mixed-data to distinguish between COVID-19 and non-COVID-19 patients. Symmetry. 2020. https://doi.org/10.3390/sym12091526.

    Article  Google Scholar 

  48. National Library of Medicine. ClinicalTrials.gov. 2021.

Download references

Acknowledgements

We acknowledge support from Dr. Geoffrey J. Topping for help with setting up hyperpolarized 13C acquisition protocols. Further, we acknowledge support from Dr. Christian Hundshammer for help with zymonic acid polarization and synthesis and help from Sandra Sühnel with animal experiments.

Funding

Open Access funding enabled and organized by Projekt DEAL. This research was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation, Sonderforschungsbereich (SFB) 824, subprojects A7 and Z3, Grant Number 391523415), the Young Academy of the Bavarian Academy of Sciences and Humanities and the European Union’s Horizon 2020 research and innovation program under Grant Agreement No 820374.

Author information

Authors and Affiliations

Authors

Contributions

WYRF and JGS designed and implemented the software code of the neural networks. MG acquired MRSI data. WYRF and MG interpreted and compared data from MRSI and neural network data. MG performed supervised line fitting and generated segmentation masks for hybrid pH maps. All authors analyzed the data in this study, BHM and FS designed the study, WYRF, MG, JGS and FS drafted the manuscript. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Franz Schilling.

Ethics declarations

Ethics approval and consent to participate

Experiments were performed in accordance with pertinent laws and regulations and approved by an ethical review board (Regierung von Oberbayern, Munich, Germany, Approval Number ROB-55.2-2532.Vet_02-17-177).

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Fok, WY.R., Grashei, M., Skinner, J.G. et al. Prediction of multiple pH compartments by deep learning in magnetic resonance spectroscopy with hyperpolarized 13C-labelled zymonic acid. EJNMMI Res 12, 24 (2022). https://doi.org/10.1186/s13550-022-00894-y

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s13550-022-00894-y

Keywords