CONSISTENT AND ROBUST 4D WHOLE-BRAIN SEGMENTATION: APPLICATION TO TRAUMATIC BRAIN INJURY Christian Ledig ⋆ Wenzhe Shi ⋆ Antonious Makropoulos ⋆ Juha Koikkalainen † Rolf A. Heckemann ‡,⋄,∗ Alexander Hammers ∗,⋄ Jyrki L¨ otj¨ onen † Olli Tenovuo ◦ Daniel Rueckert ⋆ ⋆ Department of Computing, Imperial College London, UK † VTT Technical Research Centre of Finland, Tampere, Finland ‡ MedTech West, Institute of Neuroscience and Physiology, University of Gothenburg, Sweden ⋄ Division of Brain Sciences, Faculty of Medicine, Imperial College London, UK ∗ The Neurodis Foundation, CERMEP, Lyon, France ◦ Turku University Hospital, Turku, Finland ABSTRACT We propose a consistent approach to automatically segmenting lon- gitudinal magnetic resonance scans of pathological brains. Using symmetric intra-subject registration, we align corresponding scans. In an expectation-maximization framework we exploit the availabil- ity of probabilistic segmentation estimates to perform a symmet- ric intensity normalisation. We introduce a novel technique to per- form symmetric differential bias correction for images in presence of pathologies. To achieve a consistent multi-time-point segmenta- tion, we propose a patch-based coupling term using a spatially and temporally varying Markov random field. We demonstrate the su- perior consistency of our method by segmenting repeat scans into 134 regions. Furthermore, the approach has been applied to segment baseline and six month follow-up scans from 56 patients who have sustained traumatic brain injury (TBI). We find significant correla- tions between regional atrophy rates and clinical outcome: Patients with poor outcome showed a much higher thalamic atrophy rate (4.9 ± 3.4%) than patients with favourable outcome (0.6 ± 1.9%). Index Terms— brain MRI, longitudinal segmentation, EM op- timisation, temporal consistency, traumatic brain injury 1. INTRODUCTION The worldwide incidence of traumatic brain injury (TBI) cases is estimated at 6.8 million annually, representing a substantial public health burden [1]. Although the need for reliable assessment tools was already expressed 30 years ago [2], prognostic assessment re- mains a challenge, and standardised models to predict outcome of patients with moderate TBI remain unavailable [1]. To assist the de- velopment of such models, an accurate assessment and understand- ing of the structural changes happening during and after TBI is cru- cial. In [3] indications for brain volume loss following a TBI have been identified using tissue segmentation techniques on structural magnetic resonance (MR) scans and diffusion tensor imaging. Several methods have been published that address the problem of quantifying brain changes over time based on MR scans. One popular representative is “CLASSIC” [4], a method that uses adap- tive clustering for tissue segmentation of MR scans taken at multiple time-points. Another longitudinal segmentation method, presented This work is partially funded under the 7th Framework Programme by the European Commision (http://cordis.europa.eu/ist/). by Lorenzo et al. [5], is based on expectation-maximization (EM) optimisation and applied to 4D cardiac sequences. This asymmetric approach uses stationary temporal Markov random fields (MRF) and affine alignment with a probabilistic spatiotemporal atlas. Several other methods have been proposed that address primar- ily brain tissue segmentation such as [6] which uses level sets. Wolz et al. [7] simultaneously segment the hippocampus of longitudinal scans using graph cuts. Recent research suggests that a fully sym- metric process can significantly reduce bias, e.g. in atrophy measure- ment [8]. In addition the advantages of a longitudinal segmentation of several time points over multiple, single time point segmentation are well established [4, 7]. In this work we propose a novel method based on [9] for con- sistent segmentation of serial images into many anatomical regions, rather than tissue classes. Different from previous methods, we per- form image alignment, intensity normalisation and differential bias field correction in a symmetric fashion. To achieve consistency, we present an approach to perform differential bias field correction in the presence of significant pathology in the images. We further in- troduce a novel way to determine a spatially and temporally varying, fully data driven temporal coupling of the longitudinal segmentation based on MRF. To our knowledge, our method is the first consis- tent segmentation approach that segments longitudinal whole-brain scans into a large number of anatomical structures while being robust to pathology. Our experiments show quantitative evidence for im- proved segmentation consistency while maintaining high accuracy. We also demonstrate on a cohort of TBI patients that the proposed method is robust and has the potential to quantify imaging biomark- ers, specifically atrophy, that correlate well with clinical outcome. 2. METHOD 2.1. Spatial priors and symmetric longitudinal image alignment Assuming preprocessed, brain extracted and bias corrected images we derive subject specific probabilistic labels from M available at- lases using a multi-atlas label propagation approach. For the un- segmented images at n time points, I t at t =0 ...n - 1, we cal- culate individual transformations φ t m ,m =1 ...M by registering M manually labelled atlases to I t . For the image alignment, we employ MAPER [10], an approach that incorporates tissue proba- bility maps into the registration and relies on a non-rigid registration