Synopsis

Keywords: Data Analysis, CEST & MT, pH imaging

AcidoCEST MRI can measure the extracellular pH of the tumor microenvironment. We have further refined our “Bloch fitting” method, and shown that this analysis method can accurately and precisely measure pH without additional MRI information, with an accuracy of 0.03 pH units. In addition, we have developed a machine learning method that can classify pH as > 7.0 or < 6.5 pH units (PPV=0.94, NPV=096), and a machine learning regression method that can estimate pH with a mean absolute error of 0.031 pH units.

INTRODUCTION

AcidoCEST MRI is used to measure extracellular pH in the tumor microenvironment by evaluating CEST spectra of an exogenous agent that produces two CEST signals, with at least one pH dependent signal.1 The Bloch-McConnell equations modified to include pH as a fitting parameter can be used to analyze CEST MRI, known as “Bloch fitting”.2 However, these equations also fit for T1, T2, B1, and B0, raising questions about whether this multiparametric fitting is the best approach. These equations also fit for concentration, and temperature can also affect chemical exchange rates, raising concerns that concentration and temperature can affect the pH measurement.
More recently, machine learning methods have been used to evaluate CEST MRI.3 AcidoCEST MRI provides an intriguing opportunity for machine learning methods, because CEST spectra of an exogenous agent can be difficult to visualize by humans (so that human learning is often reduced to focusing on one saturation frequency) and yet the spectra are rich in information for machine learning. Both classification methods and regression methods can be employed to evaluate acidoCEST spectra.

METHODS and MATERIALS

To refine the analysis of acidoCEST MRI, we acquired 36,000 CEST spectra of iopamidol (Isovue™, Bracco Diagnostics) at 5 concentrations, 5 T1 relaxation times, 5 temperatures, 6 saturation powers, 6 saturation times, and 8 pH values (Figure 1). We also measured T1, T2, B1 and B0 of each sample, and carefully controlled and validated the sample temperature. We used an advanced QUESP method to measure chemical exchange rates of iopamidol over a range of pH, and the Bloch fitting method was recalibrated with these results. Bloch fitting was used to estimate pH using only CEST spectra or by including T1, T2, B1, and/or B0 measurements (16 combinations). We analyzed how the pH estimates from Bloch fitting were potentially dependent on concentration and temperature. Computation time was measured for the analyses with the 16 combinations.
The same experimental dataset was split into sets of training data (70%) and testing data (30%) to assess machine learning methods (Figure 2). We trained and validated the Linear Regression Classification (LFC) model and the Random Forest Classification (RFC) model to classify pH as > 7.0 or pH < 6.5. We evaluated the sensitivity, specificity, PPV and NPV for each classification. We also trained and validated a regression approach using a Random Forest Regression (RFR) model and a LASSO model to estimate pH between 6.2 – 7.4. We evaluated the mean absolute error of the pH measurements.

RESULTS

Our Bloch fitting method could accurately and precisely measure pH throughout the range of pH 6.2-7.3, using CEST results without the need for experimental T1, T2, B1, or B0 measurements, resulting in an accuracy of 0.03 pH units relative to a benchtop pH meter(Figure 3). As expected, computation time was longest for analyses that only included CEST results. The most accurate and precise pH estimates were obtained with saturation power of 3 mT and saturation time of 3 sec. The pH estimates were dependent on temperature, and independent of concentration and T1 time. The precision of the measurement was dependent on concentration, as evaluated with Lin’s Concordance Correlation Coefficient (LCCC).
These results show that the Bloch fitting method is robust for estimating pH with acidoCEST MRI. Obviating the need for experimental T1, T2, B1, or B0 measurements reduces total acquisition time, although lengthy analysis time is then required for the multiparametric Bloch fitting process. A concern remains that a biological milieu may interact with iopamidol, altering the pH measurement. However, the independence of the pH estimate from the Bloch fitting method on concentration and T1 mitigates this concern.
For the evaluation of machine learning methods, the RFC produced the best classifications at both pH thresholds of > 7.0 and < 6.5 pH unites, with a positive predictive value of 0.96 and negative predictive value of 0.94 for both thresholds (Table 1). The RFR measured pH with a mean absolute error of 0.031 pH units that was consistent throughout the pH range, which performed better than LASSO that showed a mean absolute error of 0.24 pH units that was biased against low and high pH (Figure 3). The most accurate and precise pH estimates were obtained with saturation power of 3 mT and saturation time of 3 sec.
The outstanding pH classifications with RFC machine learning can be used for clinical diagnoses. A pH > 7.0 is an outstanding threshold for differentiating cancer vs. non-cancer lesions. A pH < 6.5 is an excellent threshold for differentiating highly aggressive and metastatic tumors from more benign tumors. The outstanding accuracy and precision of pH estimates from RFR can be used to provide a more quantitative diagnosis of tumor status and treatment effect. Importantly, these machine learning methods are very fast, which addresses a pitfall of the Bloch fitting method.

CONCLUSIONS

This study concludes our development of the Bloch fitting method for acidoCEST MRI. This study demonstrates opportunities to apply machine learning to the clinical translation of acidoCEST MRI.

Acknowledgements

Our research is supported by the NIH/NCI through grants R01 CA231513 and P30 CA016672.