A Machine Learning Approach That Measures Extracellular pH Within an In Vivo Tumor Model Using acidoCEST MRI

Quantitative mapping of extracellular tumor pH (pHe) using acidoCEST MRI offers a method to characterize the tumor microenvironment. Conventional analysis of acidoCEST MRI relies on fitting the Bloch-McConnell equations to experimental CEST spectra. However, this "Bloch fitting" method is computationally intensive. Recent work has demonstrated that machine learning can accurately predict pH from CEST spectra of phantoms containing iopamidol, providing an alternative to Bloch fitting for acidoCEST MRI analysis. In this study, we evaluated the ability of machine learning models to estimate extracellular tumor pHe using acidoCEST MRI in a 4T1 murine breast cancer model. Thirty-eight tumor-bearing mice were imaged using a CEST-FISP acquisition, and CEST spectra were extracted along with T₁, T₂, B₁, and B₀ maps. Pixel-wise pH values derived from Bloch fitting served as reference values. We trained and tested three machine learning models: (1) random forest regression (RFR) using CEST spectra and T₁, T₂, B₁, and B₀ maps; (2) RFR using only CEST spectra; and (3) a convolutional neural network (CNN) using only CEST spectra. The RFR model using only CEST spectra achieved the best performance, with mean absolute percentage errors of 0.29% (training) and 0.79% (testing), corresponding to 0.020 and 0.054 pH units, respectively. Spatial pHe maps generated by the RFR methods closely matched those from Bloch-McConnell fitting, whereas CNN-derived maps showed compressed pH ranges and reduced correlation with reference pH values. These findings demonstrate that RFR provides a computationally efficient alternative to Bloch fitting for in vivo acidoCEST MRI.