Fuzzy Multi-channel Clustering with Individualized Spatial Priors for Segmenting Brain Lesions and Infarcts

Abstract

Abstract. Quantitative analysis of brain lesions and ischemic infarcts is becoming very important due to their association with cardiovascular disease and normal aging. In this paper, we present a semi-supervised segmentation methodology that detects and classifies cerebrovascular disease in multi-channel magnetic resonance (MR) images. The method combines intensity based fuzzy c-means (FCM) segmentation with spatial probability maps calculated from a normative set of images from healthy individuals. Unlike common FCM-based methods which segment only healthy tissue, we have extended the fuzzy segmentation to include patient-specific spatial priors for both pathological conditions (lesions and infarcts). These priors are calculated by estimating the statistical voxel-wise variation of the healthy anatomy, and identifying abnormalities as deviations from normality. False detection is reduced by knowledgebased rules. Assessment on a population of 47 patients from different imaging sites illustrates the potential of the proposed method in segmenting both hyperintense lesions and necrotic infarcts.

Automated identification of brain tumors from single MR images based on segmentation with refined patient-specific priors

“…Our point here is that the segmentation procedure used in ALI can be optimized for a given type of brain damage by explicitly incorporating any prior knowledge that can be provided by theoretical or other empirical data. For example, in the context of brain tumors, it is perfectly possible to define subject-specific priors that could be derived from other contrast images that are likely to be available in the clinical setting (e.g., contrast-enhanced images, FLAIR, T2-weighted or DWI images, see Prastawa et al, 2003 ), or even include more than one extra class to model the different types of the abnormal tissue (e.g., see Zacharaki et al, 2012 ). Indeed, the accuracy of ALI might be boosted by incorporating other contrast images because any additional information in those images may help the segmentation procedure to optimally differentiate between abnormal and normal brain tissue.…”