Abstract
Tumor acidosis is an important biomarker for aggressive tumors, and extracellular pH (pHe) of the tumor microenvironment can be used to predict and evaluate tumor responses to chemotherapy and immunotherapy. AcidoCEST MRI measures tumor pHe by exploiting the pH-dependent chemical exchange saturation transfer (CEST) effect of iopamidol, an exogenous CT agent repurposed for CEST MRI. However, all pH fitting methodologies for acidoCEST MRI data have limitations. Herein we present results of the application of machine learning for extracting pH values from CEST Z-spectra of iopamidol. We acquired 36,000 experimental CEST spectra from 200 phantoms of iopamidol prepared at five concentrations, five T1 values, and eight pH values at five temperatures, acquired at six saturation powers and six saturation times. We also acquired T1 , T2 , B1 RF power, and B0 magnetic field strength supplementary MR information. These MR images were used to train and validate machine learning models for the tasks of pH classification and pH regression. Specifically, we tested the L1-penalized logistic regression classification (LRC) model and the random forest classification (RFC) model for classifying the CEST Z-spectra for thresholds at pH 6.5 and 7.0. Our results showed that both RFC and LRC were effective for pH classification, although the RFC model achieved higher predictive value, and improved the accuracy of classification accuracy with CEST Z-spectra with a more limited set of saturation frequencies. Furthermore, we used LASSO and random forest regression (RFR) models to explore pH regression, which showed that the RFR model achieved higher accuracy and precision for estimating pH across the entire pH range of 6.2-7.3, especially when using a more limited set of features. Based on these results, machine learning for analysis of acidoCEST MRI is promising for eventual in vivo determination of tumor pHe.