Segmentation of Brain Tissues from Infant MRI Records Using Machine Learning Techniques

The automatic segmentation of medical images is an intensely investigated problem, due to the quick rise of medical image data amount created day by day, which cannot be followed by the number of human experts. This paper searches for the most suitable classical machine learning method to be employed in the segmentation of various tissue types from volumetric multi-spectral MRI records of 6-month infant patients. Model training and model based prediction is performed using the 10 records of the train data set available at the iSeg-2017 challenge. All MRI records are treated with histogram normalization and feature generation, and then fed to six machine learning methods, which use them as train and test data according to the leave-one-out technique. The output of the classification algorithms is evaluated with statistical methods. The best segmentation accuracy is achieved by the random forest based approach, with a correct decision rate of 83.4%.