SEGMENTATION OF SUBCORTICAL STRUCTURES AND THE HIPPOCAMPUS IN BRAIN MRI USING GRAPH-CUTS AND SUBJECT-SPECIFIC A-PRIORI INFORMATION Robin Wolz, Paul Aljabar, Daniel Rueckert * Department of Computing Imperial College London London, UK Rolf A. Heckemann, Alexander Hammers Division of Neuroscience and Mental Health MRC Clinical Sciences Centre Imperial College London London, UK ABSTRACT We propose a general framework for segmentation of subcortical structures and the hippocampus in magnetic resonance brain images based on multi-atlas label propagation and graph cuts. The label maps obtained from multi-atlas segmentation are used to build a subject-specific probabilistic atlas of a structure of interest. From this atlas and an intensity model estimated from the unseen image, a Markov random field-based energy function is defined and opti- mized via graph cuts. Compared to a previously proposed approach, our method does not rely on manual training of the intensity model and is applied to five subcortical structures and the hippocampus. We used this approach to segment the hippocampus on 60 images from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and achieved an average overlap (Dice coefficient) of 0.86 with the man- ually delineated reference segmentations. Index Terms— structural MR images, atlas-based segmenta- tion, graph cuts, subcortical structures, MRF 1. INTRODUCTION The accurate and robust segmentation of subcortical brain structures and the hippocampus in magnetic resonance (MR) images is an in- creasingly important step in many clinical applications, such as as- sisting in the diagnosis of schizophrenia or Alzheimer’s disease. Al- though much research has been published in this area [1, 2, 3, 4, 5], no method has established itself in routine clinical use. One well- validated approach relies on combining the segmentations obtained from non-rigidly aligning multiple manually labeled atlases with the target image [3]. The final label at each voxel is determined by ap- plying vote-rule decision fusion. This method makes no use of the target intensity information. Considering such information, how- ever, potentially results in further improvements to the quality of multi-atlas segmentation. Combining prior knowledge of the intensity and spatial distribu- tion of an object of interest in the contextual framework of a Markov random field (MRF) is an established technique for brain segmenta- tion (e.g. [1, 4, 6]). In these approaches spatial information in the form of a probabilistic atlas and an estimation of the probability dis- tribution of the target structure’s intensities are used to formulate an energy function. Introduced by Greig et al. [7] and proposed as a generic method for finding the global optimum for labeling tasks in ∗ This project is partially funded under the 7th Framework Programme by the European Commission (http://cordis.europa.eu/ist/) computer vision by Boykov et al. [8], graph cuts have been widely used for optimization in this area. Recently, two brain segmentation methods based on MRFs and graph cuts have been introduced: Song et al. [9] proposed a method for tissue class segmentation of 2D MR images. Their spatial prior is defined as a probabilistic atlas that is affinely registered to the target image. The intensity distributions of white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF) are modeled using Gaussian distributions. Another promis- ing approach, proposed by van der Lijn et al. [10] for segmenting the hippocampus, can be considered an extension of the multi-atlas segmentation approach of [3] and tackles the previously described problem as follows: instead of directly fusing the individual seg- mentations obtained from registering multiple atlases to the target image, they are used to build a probabilistic atlas which is combined with statistical intensity models for foreground and background to formulate an energy function to be minimized. A limitation of this method is the reliance on a strictly controlled training of its statistical intensity model where a Gaussian distribution for the hippocampus and a Parzen estimate of the background distribution are defined on the manually labeled atlas images. This approach requires the use of identical MR sequences for the atlas (training images) and target (subject images). In this paper we propose a generalized framework for the seg- mentation of subcortical brain structures and the hippocampus in MR images which overcomes these problems: We directly estimate the Gaussian distribution for the foreground from the target image. Furthermore we use a spatially varying mixture of Gaussians (MOG) model for the background in order to better model the different back- ground parts surrounding a structure of interest. We have extended the method to five subcortical structures and the hippocampus and evaluated it on 60 images from the Alzheimer’s Disease Neuroimag- ing Initiative (ADNI) study [11]. 2. METHOD The task of segmenting an image I into structures of interest can be described as assigning a label fp ∈L to each voxel p ∈ I .A MRF-based energy function can be formulated as E(f )= λ X p∈I Dp(fp)+ X {p,q}∈N Vp,q (fp,fq ), (1) where N is a neighborhood of voxels and f is the labeling of I [8]. The data term Dp(fp) measures the disagreement between a prior probabilistic model and the observed data. Vp,q (fp,fq ) is