Enhanced Cerebral Vascular Segmentation with Harmonic Constraints Jorge Miguel Silva 1 jorge.miguel.ferreira.silva@ua.pt Augusto Silva 1 augusto.silva@ua.pt Pedro Vilela 2 ferrovilela@hospitaldaluz.pt 1 IEETA - Institute of Electronics and Informatics Engineering of Aveiro Aveiro University, Campus do Santiago 2 Hospital da Luz Abstract Stroke is the main cause of complex disability and the second leading cause of death worldwide. A fast diagnosis is crucial for avoiding patient death or after effects. In the emergency service, the cerebral Computed Tomographic angiography is one of the first diagnostic lines often accom- panied by perfusion studies. As such, in this paper we propose a novel cerebral vascular segmentation methodology based on the perfusion stud- ies with voxel based harmonic constraints. 1 Introduction Stroke is the largest cause of complex disability in adults and the second leading cause of death worldwide [1]. Every year, there are 15 million stroke incidents of which only two thirds manage to survive [2, 3]. It is described by brain cell death result- ing from blood flow blockage or rupture of a vessel in the brain. A rapid and correct diagnosis is critical to avoid patient death or after effects. One of the first diagnostic procedures used to evaluate if a patient had a stroke is to perform a cerebral Computed Tomographic (CT) angiog- raphy often comprising perfusion studies. A CT perfusion study requires fast imaging technologies and provides a series of contrast enhanced brain volumes along a series of predetermined time points. After iodinated con- trast medium has been administered intravenously to the patient [4], it is possible to assess visually and computationally the various flows that travel through the cerebral vascular structure. With this in mind, in this paper we decided to investigate the possibil- ity of computationally performing a segmentation of the vascular struc- ture considering the intrinsic dynamics of the perfusion studies. Each perfusion study is composed by 24 volumetric datasets (512x512x81x24 voxels), distributed typically along 46 seconds. 2 Methods The segmentation pipeline can be divided into 3 major steps: • 3D Registration of the volumes to compensate for the patient’s in- voluntary head movements along the time course of the perfusion studies. • The raw Segmentation of the study to obtain the mask of cerebral vascular structure and apply it to obtain the flow of contrast along the blood vessels. • Finally, perform a harmonic decomposition on the masked CT for each voxel of the cerebral vascular structure to refine the segmen- tation process based on the relative harmonic weighting of the Fourier components. In order to implement this pipeline, the Mevislab framework 1 was firstly tested. However, due to difficulties in automating the solution, we chose to use Python 3 and some of its libraries to implement the registration and segmentation steps leaving the visualization burden to Mevislab. 2.1 3D Registration Co-registration is the process of trying to spatially align one dataset with another. In order to align a given volume V (x) to a fixed volume V F (x), a spatial transformation T (x) is required. This transformation can be: rigid, hybrid or completely non-rigid. As shown in Equation 1, the goal 1 http://www.mevislab.de/ of the transformation T (x) is to minimize a Cost function C(x), which is determined by a similarity metric between a given volume V (x) and the fixed volume V F (x). The higher the similarity between volumes, the lower the cost function C(x) value assuming a monotonic behavior. ˆ T (x)= arg min T C( V F (x), V (T (x))) (1) Several methodologies were tested to perform co-registration, how- ever we chose an affine transformation to register the 3D volumes since it better minimized the cost function C(x). The similarity metric was the Mutual Information [5] and the optimization strategy implemented was similar to that used by ANTS [6]. To implement the solution we used the dipy 2 Python library. Firstly we created a similarity metric by specifying the number of bins to be used in the discretization of the joint and marginal probability distri- bution functions (PDF) - the selected value was 32 - and the percentage of voxels to be used for computing the PDFs. All voxels were used in order to achieve the best registration possible. Furthermore, in order to avoid getting stuck at local optima, and thus to accelerate convergence, we opted to use a multi-resolution strategy by building a Gaussian Pyramid with the default settings. These settings consist of 3 resolutions with 10000 iterations at the coarsest resolution, 1000 iterations at the medium reso- lution and 100 iterations at the finest. Additionally, in order to build the Gaussian Pyramid, the fixed volume was firstly smoothed at each level of the pyramid using a Gaussian kernel. Next, we defined the sub-sampling factors so that the coarsest image had a quarter of the original size, the middle resolution had half of the original size, and the image at the finest scale had the same size as the original image. Finally, we used these configurations to instantiate the registration class, and perform the Affine Transform (translation, rotation, scale and shear). 2.2 Segmentation After the 3D registration, segmentation of the blood vessels took place. For this process we selected the intra-cranial region and performed the segmentation of the vascular structure. Figure 1 shows the different stages of this segmentation [7]. In order to select the intra-cranial space, we looked at the Hounsfield Scale, the standard reference scale used for measuring radiodensity in CT scanning. As shown by Table 1, the range of bone tissue goes from 200 to 3000 HU. On the other hand, blood with contrast ranges from 100 HU to 500 HU. Tissue Hounsfield Units (HU) Bone 200 - 3000 Blood with Contrast 100 - 500 Blood 40 Table 1: Table of Hounsfield Units for different tissues Since our first step was to select the intra-cranial region, we started by applying a threshold to all the volumes of the exam. A threshold at 500 Hounsfield Units (HU) was applied in order to detect bone and ignore other soft tissues. To the resulting binary image, a 3D morphological opening operation with a spherical structural element was performed to eliminate small objects in the intra-cranial region of the volume. Next, the volume was labeled and a mask constructed by selecting the labeled intra-cranial region. This mask was applied to the original registered ex- amination, thus creating an exam containing only the interior of the skull. 2 http://nipy.org/dipy/