4881

Whole-Brain R1 mapping predicts occupational Mn air exposure: a support vector machine approach
David A Edmondson1,2, Sébastien Hélie3, and Ulrike Dydak1,2

1School of Health Sciences, Purdue University, West Lafayette, IN, United States, 2Department of Radiology and Imaging Sciences, Indiana University School of Medicine, Indianapolis, IN, United States, 3Department of Psychological Sciences, Purdue University, West Lafayette, IN, United States

Synopsis

Manganese (Mn) is a neurotoxin that can lead to symptoms similar to Parkinson’s disease. Welders exposed to welding fume can accumulate quantities of Mn in their brain eliciting T1 contrast effects. Mn exposure estimates are useful for determining a welder’s risk for symptoms, but quantifying Mn in the brain would be more beneficial. While R1 (1/T1) is proportional to local Mn accumulation, the relationship is likely non-linear, complicating interpretation of R1. Therefore, we propose a support vector machine model using whole-brain R1 maps to predict classes as determined by group, Mn air exposure, and excess brain Mn accumulation.

Introduction

Mn, a T1 contrast agent, accumulates in the brains of welders when being exposed to occupational levels of Mn contained in welding fume. Consequently, T1-W images taken during magnetic resonance imaging (MRI) scans (Figure 1) reveal hyperintensities primarily in the basal ganglia, but also in cortical grey and white matter. Although it is widely assumed that R1 (1/T1) is proportional to Mn accumulation, previous studies1,2 suggest that the relationships between Mn exposure, Mn accumulation, and changes in R1 are likely non-linear. To better understand these relationships, we used R1 maps to predict Mn exposure and Mn deposition in the brain using a series of support vector machine models.

Methods

Data was used from a prior study involving 89 subjects (57 welders, 32 controls). Scans were performed on a 3T GE Signa MRI scanner (GE Healthcare) using an 8-channel head coil. Whole brain 3D R1 relaxation maps were calculated from two spoiled gradient echo images resulting in a quantitative measure of Mn deposition in the brain3. R1 maps were then segmented into 192 regions using Freesurfer. The distribution of R1 in each region (Figure 2) was summarized by five statistics: mean, variance, skew, minimum, and maximum. Three targets were predicted using R1: study group (welder or control), 8-hr time weighted average of Mn exposure (MnTWA), and Excess Mn (MnEX) in the brain. MnTWA was calculated using acquired air samples from the workplace and accounting for factors that reduce the total amount of Mn inhaled into the lungs4. MnEX was calculated from a biokinetic model5 that predicted Mn accumulation in the brain given MnTWA and a detailed work history for each subject. Finally, data analysis was performed using Python’s Scikit-Learn machine learning module6. A support vector machine was used for classification (SVC) at increasing thresholds in MnTWA and MnEX for each summary statistic using a penalty term to account for unbalanced classes. Leave one out cross-validation was used to measure accuracy of each SVC. Finally, for each threshold, SVCs for each summary statistic were aggregated and the final prediction was determined by a majority rule. Accuracy was used as a final performance measure.

Results

Individual and aggregated prediction accuracies are found in Table 1. We found that while individual statistics performed slightly better than chance, the aggregated SVC model for predicting welder or control (study group) was about 74% accurate. Overall, we had better prediction accuracy for SVC models targeting MnTWA and MnEX (Figure 3) in differing threshold levels. For MnTWA, accuracy greater than 80% was found for air exposure levels of 0.13, 0.14, and 0.15 mg/m3. MnEx performed a little poorer overall but had accuracies greater than 80% for a threshold greater than or equal to 10 mg.

Discussion

Results indicate that we have the ability to use R1 maps to predict Mn exposure in subjects with reasonably high accuracy. While a SVC model was able to discriminate between welders and controls fairly well three-quarters of the time, better accuracies were found when we separated the groups based on exposure or excess Mn in the brain. In the MnTWA model, subjects exposed to air Mn concentrations greater than 0.11 mg/m3 had larger signatures that allowed the SVC to classify with greater accuracy, suggesting that Mn accumulation in the brain can be readily seen in the brain at these levels. When looking at separate statistics, minimum is a better predictor in lower thresholds while skew and maximum are better predictors in higher thresholds. The slightly lower level of performance of the MnEX model may reflect a difference in the storage location of Mn in the brain. Mn that is bound to macromolecules or stored in cells may not readily affect R1 as well as free Mn7,8. Therefore, while our model may predict the amount of excess Mn in the brain, it does not predict the amount of free Mn and thus, a more complex biokinetic model may be necessary.

Conclusion

This study is a representation of how MRI and machine learning can be used to synergistically analyze a set of data with low signal. We found that R1 can be used to differentiate between subjects at different exposure levels. Because our SVC’s prediction accuracy increased in the higher MnTWA threshold values, the results suggest that R1 as a marker for Mn accumulation in the brain may not have a linear relationship with Mn accumulation in the brain. To further elucidate this relationship, future studies should take into account whether Mn in the brain is in a free or bound state.

Acknowledgements

This study was supported by NIEHS F31 ES028081 and NIEHS R01 ES020529.

References

  1. Ma RE, Ward EJ, Yeh C-LL, et al. Thalamic GABA levels and occupational manganese neurotoxicity: Association with exposure levels and brain MRI. Neurotoxicology. 2017;64:30-42. doi:10.1016/j.neuro.2017.08.013.
  2. Lee EY, Flynn MR, Du G, et al. T1 relaxation rate (R1) indicates nonlinear mn accumulation in brain tissue of welders with low-level exposure. Toxicol Sci. 2015;146(2):281-289. doi:10.1093/toxsci/kfv088.
  3. Deoni SCL. High-resolution T1 mapping of the brain at 3T with driven equilibrium single pulse observation of T1 with high-speed incorporation of RF field inhomogeneities (DESPOT1-HIFI). J Magn Reson Imaging. 2007;26(4):1106-1111. doi:10.1002/jmri.21130.
  4. Ward EJ, Edmondson DA, Nour MM, Snyder S, Rosenthal FS, Dydak U. Toenail Manganese: A Sensitive and Specific Biomarker of Exposure to Manganese in Career Welders. Ann Work Expo Heal. 2017;62(1):101-111. doi:10.1093/annweh/wxx091.
  5. Leggett RW. A biokinetic model for manganese. Sci Total Environ. 2011;409(20):4179-4186. doi:10.1016/j.scitotenv.2011.07.003.
  6. Pedregosa F, Varoquaux G, Gramfort A, et al. Scikit-learn: Machine Learning in Python. J Mach Learn Res. 2012;12:2825-2830. doi:10.1007/s13398-014-0173-7.2.
  7. Troughton JS, Greenfield MT, Greenwood JM, et al. Synthesis and evaluation of a high relaxivity manganese(II)-based MRI contrast agent. Inorg Chem. 2004;43(20):6313-6323. doi:10.1021/ic049559g.
  8. Tromsdorf UI, Bigall NC, Kaul MG, et al. Size and surface effects on the MRI relaxivity of manganese ferrite nanoparticle contrast agents. Nano Lett. 2007;7(8):2422-2427. doi:10.1021/nl071099b.

Figures

Table 1: Prediction Accuracy Results for Support Vector Machine Classification of R1. A) Prediction accuracy for individual SVC models for each of the five statistical parameters of 192 region of interest R1 distributions. Aggregated refers to using all five SVCs and determining the class based on a majority rules. B) Prediction accuracy for aggregated SVC models for every threshold chosen for excess brain Mn (MnEX) and 8-hr time weighted average Mn air exposure (MnTWA).

Figure 1: T1-W images for Welder and Control. The welder (left) has vivid T1-hyperintensity in the basal ganglia, specifically the globus pallidus, as well as in white matter tracts. The control (right) is shown on the same contrast scale for reference.

Figure 2: Example Distributions of R1 in two regions of interest. Above are two regions of interest as segmented using Freesurfer with examples of a welder’s and a control’s R1 distribution in the regions. Five statistical parameters were taken from each distribution for use for separate SVC models: mean, variance, skew, minimum, and maximum. These are chosen classic examples of our hypothesis where the welder, with excess Mn accumulation in the brain, will have higher R1s, resulting in a right-ward shift of their distribution.

Figure 3: Prediction Accuracy Results from SVC Analysis A) For predicting classes based on air Mn concentration (MnTWA) thresholds, the combined SVC model shows relatively stable predictability across all thresholds, but begins to peak after 0.11 mg/m^3 . B) The combined SVC model has more variability in its predictions of classes based on Excess Brain Mn (MnEX) compared to MnTWA. One possible cause is poor performance of the mean SVC while the minimum SVC performs well in the lower range of thresholds. Skew performs well in the upper ranges before the combined SVC model peaks at around 80% accuracy.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
4881