Gender Classifications Based On Structural MRI

The neuroimaging technique provides an effective tool for studying the mechanism and variation of the structure and function of the human brain.Machine learning techniques provide a novel way for effective analysis of neuroimaging data.A number of MRI-based studies have been performed to evaluate gender-related differences in brain structure.To note,all these studies were based on traditional statistical analyses,while statistical results reflect only "possibilities"(e.g.,the cortical thickness in a certain region of the males are thicker than that of females at a possibility of 99%),rather than the "truth".As a result,contradictory conclusions may sometimes be drawn based on statistical analyses,and it is impossible to judge which of the conclusions are correct based on statistical analyses.By introducing machine learning techniques,it is possible to evaluate which of these statistical results reflect the "truth":the gender differences that contribute to the effective identification of individuals’ gender may be more likely to be "true".In addition to evaluating the statistical results,gender classifications may also be helpful for investigating the neural basis and objective diagnosis of some psychiatric diseases that are prevalent in a certain gender(e.g.,ADHD,which is prevalent in the males,and depression,which is prevalent in the females).In this study,we performed gender classifications based on MRI data of large sample size.Specifically,we performed the study on MRI-derived parameters of 526 healthy adults(215 males,aged 22-35 years),including cortical surface area,cortical thickness,gray matter volume,folding index,mean curvature,Gaussian curvature and subcortical volume.Elastic net,stacked auto-encoder and random forests were used to decode individuals’ gender.The main results are as follows:Gender classifications based on elastic net:we use the elastic net to construct the classification models based on different brain structural parameters,and 10-fold cross validation was used to estimate the performance of the classifiers.The results showed that the accuracy based on the volume of subcortical segmentations was the highest(85.38%);with all 7 types of brain structural parameters included in the model,the classification accuracy was enhanced to 88.61%;and an accuracy of 90.50%when the asymmetry indices of the parameters of the cortical parameters were further included in the model.Gender classifications based on stacked auto-encoder:we use the stacked auto-encoder to construct the classification model based on different brain structure parameters,and 10-fold cross validation was used to estimate the performance of the classifiers.The results indicated that the classification based on the volume of subcortical segmentations was the highest(86.50%);with all 7 types of brain structural parameters included in the model,the classification accuracy was enhanced to 89.92%;classification accuracy reduced when the asymmetry indices of the parameters of the cortical parameters were further included in the model.We suggest that an effective fusion of the brain structural features may have already achieved with the use of stacked auto-encoder,and the inclusion of the asymmetry indices did not improve the classifications.Gender classifications based on random forests:we use the random forests to construct the classification model based on different brain structure parameters,and 5-fold cross validation was used to estimate the performance of the classifiers.The results showed that the accuracy based on subcortical volume was the highest(83.48%);with all 7 types of brain structural parameters included in the model,the classification accuracy was enhanced to 88.61%;classification accuracy did not change much when the asymmetry indices of the parameters of the cortical parameters were further included in the model.Overall,we performed three studies on gender classifications based on human brain structure MRI.The results show that such parameters as subcortical volume,surface area,mean curvature and gray matter volume are more specific for gender decoding as compared to folding index and Gaussian curvature.The parameters are complimentary to each other,as the combinations of several parameters can effectively improve the classification accuracies.In addition,the current results indicate that hemispheric asymmetries of human cortex are important aspect of gender-related differences in brain structure.The current results may be helpful for choosing proper brain structural parameters in studies based on MRI data,and may be helpful for the classification of psychiatric disease that are prevalent in a certain gender based on MRI-derived brain structural parameters.