Purpose: To test and evaluate the performance of a machine learning (ML) algorithm in predicting pH from CEST spectra of iopamidol.
Methods: Two-hundred iopamidol samples were prepared with 5 concentrations (5-50 mM), 5 T₁ values (0.4-3.9 sec) and 8 pH levels (6.25-7.30, equally spaced). Thirty-six CEST scans were performed on each sample using a FISP acquisition sequence and a cross-combination of six saturation powers (0.5-6 μT) and six saturation times (0.5-6 sec) with a Bruker 7T preclinical scanner. The set of 36 CEST scans was repeated at five temperatures. For each CEST scan, 81 images were collected at different saturation frequencies. In addition, T₁, T₂, B₁ and B₀ maps were also acquired. All 36,000 CEST spectra from the 200 samples and the T₁, T₂, B₁ and B₀ data for each phantom sample were used to the train and validate the ML algorithm.
Results: ML is able to produce reliable pH prediction from iopamidol CEST spectra. As expected, the accuracy of pH prediction increases as the increase of saturation powers and times. However, to our surprise, the ML algorithm can achieve a 95% accuracy even for spectra acquired with 0.5 μT power and 0.5 sec time, which is not easily achievable by other analysis methods. This advantage of ML likely arises from that no assumptions need to be made about what parts of the spectra are useful. We also demonstrated that the pH prediction accuracy can be retained when the number of saturation frequencies is decreased from 81 to 27, therefore eliminating the need for data acquisition for redundant frequencies.
Conclusion: Employing a ML algorithm is an effective method in predicting pH based on iopamidol spectra across a wide range of pH and saturation parameters. The accuracy of pH prediction can be retained when the saturation frequency list is truncated.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Institutes of Health grant no. 1R01CA169774.