Neuroimage-based clinical prediction using machine learning tools

摘要

Classification of structural brain magnetic resonance (MR) images is a crucial task for many neurological phenotypes that machine learning tools are increasingly developed and applied to solve this problem in recent years. In this study binary classification of T1-weighted structural brain MR images are performed using state-of-the-art machine learning algorithms when there is no information about the clinical context or specifics of neuroimaging. Image derived features and clinical labels that are provided by the International Conference on Medical Image Computing and Computer-Assisted Intervention 2014 machine learning challenge are used. These morphological summary features are obtained from four different datasets (each N > 70) with clinically relevant phenotypes and automatically extracted from the MR imaging scans using FreeSurfer, a freely distributed brain MR image processing software package. Widely used machine learning tools, namely; back-propagation neural network, self-organizing maps, support vector machines and k-nearest neighbors are used as classifiers. Clinical prediction accuracy is obtained via cross-validation on the training data (N = 150) and predictions are made on the test data (N = 100). Classification accuracy, the fraction of cases where prediction is accurate and area under the ROC curve are used as the performance metrics. Accuracy and area under curve metrics are used for tuning the training hyperparameters and the evaluation of the performance of the classifiers. Performed experiments revealed that support vector machines show a better success compared to the other methods on clinical predictions using summary morphological features in the absence of any information about the phenotype. Prediction accuracy would increase greatly if contextual information is integrated into the system. (C) 2017 Wiley Periodicals, Inc.