Accurate Pathology Segmentation in FLAIR MRI for Robust Shape Characterization

Abstract

Shape analysis of pathology requires an accurate initial segmentation. However, in magnetic resonance images (MRI) of the brain, an artifact known as partial volume averaging (PVA) pathology severely impedes segmentation accuracy. Traditional MRI brain segmentation techniques rely on Gaussianmixture models to handle noise and PVA, or high-dimensional feature sets that exploit redundancy in multispectral datasets. Unfortunately, model-based techniques have limited performance on images with non-Gaussian noise distributions and pathology, and multispectral techniques do not make efficient use of imaging resources. For robust segmentation, a generalized PVA modeling approach is developed for FLAIR MRI with white matter lesion (WML) pathology that does not depend on predetermined intensity distribution models or multispectral scans. Instead, PVA is estimated directly from each image using an adaptively defined global edge map constructed by exploiting a mathematical relationship between edge content and PVA. The final PVA map is used to segment WML with sub-voxel accuracy. Using the highly accurately segmented WML, shape analysis experiments were conducted to characterize the types of lesions in the brain. Currently, WML are divided into periventricular white matter lesions (PVWML) and deep white matter lesions (DWML) and radiologists differentiate between them manually. It is important classify these two types of WML since the pathogenic mechanisms between them provide clues regarding the pathophysiology of many diseases (such as MS, stroke, etc.). In this work, we used boundary-based and Fourier descriptors to automatically classify the WML into PVWML and DWML classes. A supervised, linear discriminant classifier was used, where a leave-one-out training and testing strategy was employed. It was found that circularity features alone provided the highest classification rate (90%).