A rapid and accurate guanidine CEST imaging in ischemic stroke using a machine learning approach

<i>Objective.</i>Rapid and accurate mapping of brain tissue pH is crucial for early diagnosis and management of ischemic stroke. Amide proton transfer (APT) imaging has been used for this purpose but suffers from hypointense contrast and low signal intensity in lesions. Guanidine chemical exchange saturation transfer (CEST) imaging provides hyperintense contrast and higher signal intensity in lesions at appropriate saturation power, making it a promising complementary approach. However, quantifying the guanidine CEST effect remains challenging due to its proximity to water resonance and the influence of multiple confounding effects. This study presents a machine learning (ML) framework to improve the accuracy and robustness of guanidine CEST quantification with reduced scan time.<i>Approach.</i>The model was trained on<i>partially synthetic data</i>, where measured line-shape information from experiments were incorporated into a simulation framework along with other CEST pools whose solute fraction (<i>f</i>_s), exchange rate (<i>k</i>_sw), and relaxation parameters were systematically varied. Gradient-based feature selection was used to identify the most informative frequency offsets to reduce the number of acquisition points.<i>Main results.</i>The proposed model achieved significantly higher accuracy than polynomial fitting, multi-pool Lorentzian fitting, and ML models trained solely on synthetic or<i>in vivo</i>data. Gradient-based feature selection identified the most informative frequency offsets, reducing acquisition points from 69 to 19, a 72% reduction in CEST scan time without loss of accuracy.<i>In vivo</i>, conventional fitting methods produced unclear lesion contrast, whereas our model predicted clear hyperintense lesion maps. The strong negative correlation between guanidine and APT effects supports its physiological relevance to tissue acidosis.<i>Significance.</i>The use of partially synthetic training data combines realistic spectral features with known ground-truth values, overcoming limitations of purely synthetic or limited<i>in vivo</i>datasets. Leveraging this data with ML, enables robust quantification of guanidine CEST effects, showing potential for rapid pH-sensitive imaging.