Hsin-Yu Chen^{1}, Chia-Min Chen^{1}, Teng-Yi Huang^{1}, and Tzu-Chao Chuang^{2}

In this study, we
present a systematic approach to derive effective MR biomarkers of cerebral
cortical thickness using machine learning methods and a large-scale database.
Three neuroanatomical parcellation schemes for assessing region cortical
thickness were compared. The results supported using the Desikan–Killiany atlas^{1}
of FreeSurfer produced robust results of age and gender predictions in normal
subjects.

**Purpose**

**Methods and materials**

We used the publicly available IXI data set (566 T1-weighted volumes,
254 males and 312 females, average age is 48.40 ± 16.48, age range is 19.98～86.32 yr). The high resolution 3D T1 volumes were processed using
FreeSurfer with three parcellation schemes: Desikan–Killiany (Aparc) protocol^{1} for 68 labels, Destrieux (A2009s) protocol^{2,3}
for 148 labels and the Desikan–Killiany–Tourville (DKT) protocol^{4} for
62 labels. The average CT values of the cortical labels were stored in a
spreadsheet for machine-learning analysis in the Python environment.

We assessed the accuracies of two predictive models, age versus CT and
gender versus CT^{5}, to identify the optimal parcellation schemes.
Figure 1 displays the block diagrams of the data analysis. For age versus CT, we
obtained results using linear
regression and random forest regression (200 trees) with the linear model: $$AGE \sim VOLUME + GENDER + \sum_iT(labels_{i})$$

where T(labels_{i}) is the average thickness in each labels, and VOLUME is the intracranial volume. We used two-fold cross-validation with 1000 times randomly
shuffles in training and testing data sets and measured the root mean square error
of the predicted ages to evaluate the performance of predictions.

$$RMSE = \sqrt{\frac{\sum (AGE_{true}-AGE_{predicted} )^{2}}{N}}$$

For gender versus CT, we classified the genders of the subjects using the binomial generalized linear model (bGLM) with the model formula: $$GENDER \sim VOLUME + AGE + \sum_i T(labels_{i})$$ We used two-fold cross-validation with 1000 times of randomly shuffles in training and testing data sets. We compared the predicted genders obtained from the testing data sets with the ground truth and varied the threshold of gender classification to calculate the average area under curve (AUC) of the receiver operation curve (ROC) of the 1000 cross-validations.

**Discussion and Conclusions**

[1] Desikan, R. S., Segonne, F., Fischl, B., Quinn, B. T., Dickerson, B. C., Blacker, D., et al. An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage. 2006; 3, 968–980.

[2] Destrieux, C., Fischl, B., Dale, A., and Halgren, E. Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature. Neuroimage. 2010;53, 1–15.

[3] Destrieux, C., Halgren, E.,Fischl, B., M.I. Sereno Variability of the human brain studied on the flattened cortical surface. Soc. Neurosci. 1998;p. 1164 Los Angeles, CA.

[4] Klein, A., Tourville, J. 101 labeled brain images and a consistent human cortical la-beling protocol. Front. Neurosci. 6, 171.

[5] Tustison, NJ, Cook, PA, Klein, A,Song, G, Das, SR, Duda, JT, Kandel, BM, van Strien, N, Stone, JR, Gee, JC, Avants, BB Large-scale evaluation of ANTs and FreeSurfer cortical thickness measurements. NeuroImage99. 2014;166–179.

[6] Breiman, L. Random forests. Machine Learning, 2001; pp. 5–32.

Figure 1. Block diagrams of age prediction
(top) and gender prediction (lower).

Figure 2. The RMSE distributions of age prediction using linear
regression (left) and random forest (right).

Figure 3. Average
ROC curve for gender prediction using binomial generalized linear model (left)
and random forest (right) methods. The values were averaged from 1000
permutations using bGLM and random forest.

Table 1. The mean and standard deviation of
RMSE values of age prediction.

Table 2. The mean AUC values for gender prediction