Best Set of Relative Volume Features
Relative volume features of CSF, GM, WM were fitted to all four classes of classifiers and were evaluated by f1-score and auc, precision and recall. When comparing them with each other, the prior two of f1-score and auc were taken into account. The aucs of GM for all models were around 0.50, and the micro-averages of f1-score for all models were 0.51. The aucs of WM for all models were around 0.52, and the micro-averages of f1-score for all models were 0.51. The aucs of CSF for all models were around 0.55, and the micro-averages of f1-score for all models were 0.51. Thus, the best set of relative volume feature should be CSF relative volume. Details could be checked by source code file attached to the report.
Model Selection for Relative Volume Features
When comparing the models for relative volume features, best set of relative volume feature-CSF was considered. As the best set parameter for SVC was {'C':1, 'gamma': 0.001, 'kernel': 'rbf'} , parameters of SVC(kernel=RBF) were set as that for evaluation for models trained by relative volume features. For Linear SVC, the auc were around 0.55 but with relative large variance, and the f1-scores were around 0.51. For SGD model, the aucs were around 0.55 but with relative large variance when performing repeats by cross-validation method, and the f1-scores were around 0.51. For DecisionTree mdoel, the aucs were around 0.50, and the f1-scores were around 0.50. For SVC(kernel=Polynomial), the aucs were around 0.55 but also with relative small variance, and the f1-scores were around 0.50. For SVC(kernel=RBF), the aucs were around 0.55 but with relative small variance, and the f1-scores were around 0.50. Therefore, DecisionTree model was suggested to be the most inappropriate one when taking relative volume features as training data.
Model Selection for Principal Greymatter Map Features
As the best set parameter for SVC was {'C':10, 'gamma': 0.001, 'kernel': 'rbf'} , parameters of SVC(kernel=RBF) were set as that for evaluation of models trained by principal greymatter map features. For Linear SVC, the auc were around 0.98, and the f1-scores were around 0.92. For SGD model, the aucs were around 0.97, and the f1-scores were around 0.91. For DecisionTree mdoel, the aucs were around 0.69, and the f1-scores were around 0.69. For SVC(kernel=Polynomial), the aucs were around 0.94, and the f1-scores were around 0.49. For SVC(kernel=RBF), the aucs were around 0.98, and the f1-scores were around 0.64. Therefore, DecisionTree model should be the most inappropriate one when taking relative volume features as training data. Therefore, the relative good models for principal greymatter map features was suggested to be Linear SVC and SGD model, which are all linear separation models. Decision Tree model was suggested to be relativly poorly performed.
Volume-based Classification Approach V.S. Image-based Classification Approach
When comparing these two approaches, best set of features as well as models of each one was selected. For volume-based classification approach, relative CSF volumes were chosen to be features to fit SVC or Linear SVC or SGD model. The aucs were around 0.55 and the f1-scores were around 0.51. For image-based classification approach, Linear SVC model was chosen to be trained. The aucs were around 0.98 and the f1-scores were around 0.92. Thus compared to volume-based classification approach image-based classification approach was suggested to be a relative good approach.
Conclusions
In this report, two approaches for relative volume feature and principal feature extracted from greymatter maps on the basis of MRI brain scans were conducted to train classification models which can predict genders with obtained new images.
Different classifiers were tested and verified by cross-validation and their average accuracy were principally evaluated by f1 score and aucs. Grid searches were performed to find optimized parameters for SVC. Relative volume features were compared to each other and relative CSF volume was found to be the best feature set in the first approach. Using CSF relative volume feature to explore classifers for the volume-based approach, it was suggested that Decision Tree model was relative not appropriate. Principal greymatter map features were used to explore classifiers for the image-based approach, it was suggested that Linear SVC and SGD model were relative good models while Decision Tree classifer was as well relative inappropriate model. Overall, the comparison between two approaches found that Image-based classification method performed better than volume-based classification method.