Classification Forests and Markov Random Field to Segment Chronic Ischemic Infarcts from Multimodal MRI

Abstract

Abstract. Accurate identification of ischemic lesions and brain atrophy is critical in the management of stroke patients and may serve as an important biomarker in studying post-stroke depression. In this paper we present an automated method to identify chronic ischemic infarcts in gray matter and gray/white matter partial volume areas that may be used to measure the amount of tissue loss due to atrophy in the area. The measure of tissue loss may then be used as a potential biomarker in analyzing the relation between stroke and post-stroke depression. The automated segmentation method relies on Markov random field (MRF) and random forest based classifications. The MRF classification identifies the possible lesion areas from the fluid attenuated inversion recovery (FLAIR) magnetic resonance (MR) images. Thereafter, the multimodal (T1-, T2-weighted, FLAIR, and apparent diffusion coefficient (ADC)) MR images of the possible lesion areas are fed into the classification forests along with other context-aware features and probability maps of the gray and white matter regions. The results of classification from the MRF and the classification forests are finally combined using connected component analysis to identify the final lesion area. The accuracy of the method in identifying infarcted regions from multimodal MR images has been validated on 17 patient datasets with a mean accuracy of 99%, a mean positive predictive value (PPV) of 75% and a mean negative predictive value (NPV) of 99% and a volume correlation of r = 0.98.