Phosphocreatine (PCr), a high-energy phosphate compound prevalent in muscle and brain tissue, is vital for cellular energy buffering and transport, especially in tissues with fluctuating energy demands like skeletal muscle, cardiac muscle, and brain. PCr measurement offers unique insights into cellular energetics and shows promise in evaluating mitochondrial function, identifying peripheral arterial disease and heart failure, and understanding neurodegenerative and muscle diseases. Phosphorus-31 magnetic resonance spectroscopy (³¹P MRS) is the established method for noninvasive PCr detection and quantification, providing information about pH, inorganic phosphate, and adenosine phosphates. However, its low sensitivity limits spatial resolution and extends acquisition times, hindering widespread use. Furthermore, ³¹P MRS isn't readily available on most clinical MRI scanners due to additional hardware costs. This necessitates the development of a routine diagnostic test for noninvasive PCr quantification and mapping with clinically relevant resolution and scan time. Chemical exchange saturation transfer (CEST) MRI, a sensitivity-enhancing technique, leverages interactions between exchangeable protons in low-concentration molecules and water protons. It's compatible with standard clinical MRI scanners, but its clinical translation has been hampered by lower field strengths (1.5T and 3T), reducing frequency shifts and contrast-to-noise ratio. Artificial neural networks (ANNs) are increasingly employed to extract features from large datasets and create predictive tools. This study explores the use of ANNs for CEST quantification (ANNCEST) to accurately predict metabolite concentration, exchange rate, and B₁/B₀ homogeneity from Z-spectra. The study validates ANNCEST using simulations and phantom data at 3T, optimizes PCr CEST acquisition for human skeletal muscle, and applies ANNCEST to quantify PCr concentration, exchange rate, and B₀ and B₁ maps on a clinical 3T MRI scanner. Finally, it validates ANNCEST by measuring PCr depletion and recovery in exercised skeletal muscle, comparing results with ³¹P 2D MRS. The study aims to establish ANNCEST as a sensitive and efficient method for PCr detection and quantification.

The literature extensively covers the importance of phosphocreatine (PCr) in cellular energy metabolism and its role in various physiological processes and diseases. ³¹P MRS has been the gold standard for non-invasive PCr quantification, but its limitations in spatial resolution and accessibility on clinical MRI scanners have motivated research into alternative methods. Chemical exchange saturation transfer (CEST) MRI, utilizing the interaction between exchangeable protons and water protons, has emerged as a promising technique. Studies have explored CEST MRI for detecting various metabolites, but its application to PCr mapping at clinical field strengths (1.5T and 3T) has been limited by lower sensitivity and challenges in robust quantification. The use of artificial neural networks (ANNs) in medical imaging is rapidly expanding, with applications in image reconstruction and analysis. This research builds on the success of ANNs in other imaging modalities, applying them to the complex problem of CEST quantification, aiming to improve accuracy and efficiency.

The study employed a feed-forward artificial neural network with one input layer (Z-spectral intensity at different saturation offsets), one output layer (quantification results), and multiple hidden layers (optimized to seven for adequate capacity and to avoid overfitting). A sigmoid transfer function was used in the hidden layer, and input/output scales were normalized. The network was trained using a scaled conjugate gradient backpropagation algorithm, with early stopping to prevent overfitting. The training data was divided into training (80%), validation (15%), and test (5%) sets. A modified performance function combined mean squared normalized error (mse) and the mean of the sum of squares of network weights and biases (msw) for regularization, preventing overfitting and promoting smoother network responses. Training stopped when the gradient fell below 10⁻⁷, the validation error fell below 10⁻⁴, or the maximum epoch number (10⁶) was reached. Training data for the neural network were generated using Bloch-McConnell equations, incorporating B₁/B₀ inhomogeneity and noise. For numerical simulations and phantom experiments, Z-spectra were simulated over a frequency offset range of 0.5–4 ppm with 50 offsets, with PCr CEST peaks at 1.95 ppm and 2.5 ppm. The exchange rate ratio between these peaks was set to 1:2.19 based on phantom measurements. For in vivo human skeletal muscle mapping, the frequency offset range was 1.3–3.5 ppm (50 offsets), with a CEST peak at 2.5 ppm. T₁ and T₂ values for water and PCr protons were determined experimentally. Gaussian white noise was added to mimic real-world conditions. The Bloch equations were used for numerical simulations, generating ground-truth maps of concentration, exchange rate, and B₀. Bloch fitting was performed to compare with ANNCEST. Phantom experiments were conducted on a 3T Bruker Biospec system, using PCr phantoms of various concentrations. A continuous wave saturation module and turbo spin echo (TSE) sequence were used. B₀ maps were obtained via the water saturation shift referencing (WASSR) method. T₁ and T₂ maps were acquired using inversion recovery and CPMG sequences, respectively. Human skeletal muscle imaging was performed on a 3T Philips MRI system using a 16-channel knee coil. High-resolution T₂-weighted images were collected for anatomical referencing. A continuous wave saturation module (0.6 µT, 800 ms) and a single-shot TSE sequence were used for PCr mapping. B₀ and B₁ maps were obtained using dual-echo and DREAM techniques, respectively. In vivo data was used to train a separate neural network for ANNCEST application. The Polynomial and Lorentzian line-shape fitting (PLOF) method served as a comparison method. The robustness of ANNCEST to T1 and T2 variations was also assessed using simulated data. Finally, in-magnet plantar flexion exercise was performed on subjects, with PCr mapping using both ANNCEST and ³¹P 2D MRS. Correlation analysis compared PCr maps from both methods. Image registration and downsampling were employed to address motion artifacts and resolution differences.

ANNCEST demonstrated high accuracy in predicting PCr concentration, exchange rate, and B₀/B₁ inhomogeneities from simulated and phantom data, outperforming Bloch equation fitting, particularly in low-concentration regions. In phantom experiments, ANNCEST showed excellent correlation (R = 0.9989) between predicted and ground truth PCr concentrations. Optimization of CEST parameters (0.6 µT saturation power, 800 ms saturation length) yielded maximal PCr contrast in human skeletal muscle. ANNCEST provided high-quality PCr maps of human skeletal muscle, showing robustness against B₀ and B₁ inhomogeneities. PCr concentrations obtained by ANNCEST (31.9 ± 2.0 mM in gastrocnemius medial, 31.7 ± 3.3 mM in soleus, 30.8 ± 4.1 mM in tibialis anterior, and 30.9 ± 3.9 mM in peroneus) were consistent with previous reports. The quantified exchange rate (164 ± 36.8 Hz) was also consistent with literature. ANNCEST-derived B₀ and B₁ maps showed high similarity to those from established methods. ANNCEST demonstrated robustness against T₁ and T₂ variations in both water and PCr protons. In the exercise study, ANNCEST accurately detected PCr depletion and recovery in the gastrocnemius muscle, strongly correlating with ³¹P 2D MRS data (R = 0.813, p < 0.001). The PCr recovery time constant obtained using ANNCEST was consistent with previously reported values.

ANNCEST offers a significant advance in noninvasive PCr mapping, addressing the limitations of ³¹P MRS. The use of ANNs effectively handles the complexities of CEST quantification, mitigating the effects of B₀ and B₁ inhomogeneities. The ability to rapidly acquire and process data within 1.5 minutes enables real-time analysis, crucial for dynamic studies of PCr metabolism during exercise and recovery. ANNCEST leverages the discernible PCr CEST peak to accurately quantify PCr concentration, exchange rate, and field inhomogeneities simultaneously. The robustness to T₁ and T₂ variations further enhances the reliability of the technique. The strong correlation between ANNCEST and ³¹P 2D MRS results validates its accuracy and clinical potential. This method's compatibility with widely available clinical MRI scanners makes it a cost-effective and accessible tool for diagnosing muscle-related and neurological diseases.

ANNCEST, a novel technique combining CEST MRI with artificial neural networks, provides a rapid, robust, and widely accessible method for high-spatial-resolution PCr mapping in human skeletal muscle. Its strong correlation with ³¹P MRS, robustness to field inhomogeneities and T₁/T₂ variations, and short scan time make it a promising tool for clinical applications. Future research should focus on reducing scan time further (e.g., by employing fewer saturation offsets or faster CEST sequences) and on separating the effects of pH and temperature on the exchange rate.

The accuracy of ANNCEST relies heavily on the quality and representativeness of the training data. While the study used extensive simulated data incorporating field inhomogeneities and noise, limitations might arise from not completely capturing all possible in vivo variations. The current temporal resolution may be insufficient for capturing very rapid dynamic changes in PCr metabolism. Although the study addressed T₁ and T₂ variations, the influence of other physiological factors (e.g., pH) on the exchange rate requires further investigation.