Random forests based recognition of the clinical labels using brain MRI scans

There has been a great interest in the systems that predict clinical labels from the brain images automatically for the last decade since it is a very important task that helps clinicians for decision making. In this study, clinical labels of the structural brain magnetic resonance (MR) images are predicted automatically using the random forests ensemble method. Morphological measurements like volume and thickness that exhibit details of the brain structures are used as input. Structural T1-weighted MR images of the patients with a relevant clinical phenotype are used in this study. Training images are obtained from the 150 patients and the system is tested using the images of the 100 patients. 5-fold cross validation is used for the training, determining the hyperparameters of the random forests and performance evaluation. Accuracy, the area under receiver operator characteristic curves, specificity and sensitivity are used as the performance metrics of the proposed system. Results obtained from the experiments proved that random forests can be used successfully for the identification of the clinical labels using the structural brain MR images.