Multilevel Segmentation and Classication of Ms Lesions

This paper focuses on the segmentation and classification of multiple sclerosis lesions in magnetic resonance images. As MRI is the primary tool used in the diagnosis of multiple sclerosis, there is substantial interest in developing an algorithm that will detect lesions from such images. We present a flexible framework in which segmentation and classification are integrated. We assume that we are given a training set of MRI images which contains manually labeled regions of MS lesions. The algorithm we used combines two effective techniques from the computer vision literature: graph-based bottom-up methods and top-down generative models. Results on synthetic data obtained from the McGill Brainweb simulator and real data acquired at 3T are presented. Introduction Multiple sclerosis (MS) is a demyelinating disease of the central nervous system (CNS). With MS, the body's immune system attacks myelin, which protects the nerve fibers in the CNS. The damaged myelin leads to scar tissue (sclerosis). Even though the causes of MS remain unknown and there is no cure as of today, much is known about the mechanisms involved in the disease process and several therapies have proven helpful. In order to better understand MS, magnetic resonance imaging (MRI) is often used during the diagnostic process. MRI is the best imaging technology for detecting the presence of MS lesions in different parts of the CNS. Presently, to fully diagnosis and quantify MS lesions from MRI images, it is common to have human experts manually delineate the lesions from the data. However, it is an extremely time consuming task and very often have significant intra and inter-observer variability. Therefore, there is a need to have a fully automatic detection of MS lesion. However, it is well known that lesions are relatively small, exhibit a wide variety of shapes, may have significant differences in location across patients, and have varying intensities and textures. Methods We present a framework in which segmentation and classification are integrated in a flexible manner. In particular, we will adopt the approach of [1]. We assume that we are given a training set of MRI images which contains manually labeled regions of MS lesions. Our algorithm combines two effective computer vision algorithms, namely a graph-based bottom-up method and a top-down generative model. Graph-based bottom-up methods are based on the concept of spectral clustering. Such algorithms are able to give quick answers but are very memory intensive and do not always give meaningful results. Top-down generative models are guaranteed to get an answer based on the posterior probability but are very computationally expensive and hard to design and train. The framework allows for different users to specify different types of features, e.g. intensity, spatial location, shape, and input channels to be used for classification. The algorithm is summarized as follows: 1) Based on the training set where each voxel corresponds to each node, generate the probabilities ) , ( j i m m P , prior probability that nodes i v and j v have model (class) labels i m and j m , and ) | ( i i m s P , the conditional probability of an observed feature i s at the node i v for the model i m . 2) At level 0 = t , initialize the graph 0 G with each voxel being a node 0 i v with 6-neighbor connectivity, the feature being the intensity = 0 i s Intensity at voxel 0 i v , and the weight |) | ( exp = 0 0 0 j i ij s s W − −θ , where θ is a user-defined parameter. 3) Choose representative set t R such that t i V v ∈ ∀ , with t V being the entire set of nodes and 1 < < 0 β , ∑ ∑ ∈ ∈ ≥ t j t j V v t ij R v t ij W W β . 4) Construct the graph at the next level 1 + t , ) , ( = 1 1 1 + + + t t t W V G by the following: a) Set t t R V = 1 + . b) Compute interpolation matrix ∑ + + ∈ = 1 1 / t t k V v t ik t ij t ij W W P where t t i V v ∈ and 1 1 + + ∈ t j V v . c) Compute the affinity at coarser level, . = 1 t t t t P W P W Τ + d) Compute the feature of node 1 1 + + ∈ t j V v , ∑ ∑ ∈ ∈ + = t t i t t i V v t ij V v t i t ij t i P s P s / 1 . e) Use the model-based probabilities to modify the affinity. For 1 1 1, + + + ∈ t j t i V v v , ) , | ( = ~ 1 1 1 1 + + + + t j t i ij t ij t ij s s X P W W and set 1 1 ~ = + + t t W W . 5) Set 1 + ← t t and repeat steps 2)-4) until 1 |= | t V or 0 = t W or T t = , where T is a user defined value. 6) Classify each voxel 0 0 i v by } , ,...., , { mode = 0 1