Diffuse Axonal Injury Lesion Segmentation Using Contouring Algorithm Olga V. Senyukova 1 , Valeriy E. Galanine 1 , Andrey S. Krylov 1 , Alexey V. Petraikin 2 , Tolibdjon A. Akhadov 2 , Sergey V.Sidorin 2 1 Faculty of Computational Mathematics and Cybernetics, Lomonosov Moscow State University, Moscow, Russia 2 Children's Clinical and Research Institute Emergency Surgery and Trauma, Moscow, Russia osenyukova@graphics.cs.msu.ru Abstract Diffuse axonal injury (DAI) is a common type of brain damage induced by trauma. Accurate detection and quantification of DAI lesions is essential for assessment of a patient’s state and making a prognosis of a traumatic disease. This study is devoted to semi- automated segmentation of DAI lesions on T2*-weighted brain magnetic resonance (MR) images. A radiologist outlines the regions containing lesions, and the lesions are further automatically delineated inside these regions. A proposed algorithm for automated segmentation of DAI lesions inside specified regions is based on contouring algorithm which treats the image as a 3D topological map where a pixel’s intensity corresponds to its height. It builds isolines of an intensity function. Lesion contours are among these isolines. In order to distinguish true lesion contours from other closed contours obtained by contouring algorithm machine learning approach is exploited. A labeled training base with positive (lesions) and negative (non-lesions or lesion parts) examples of closed contours is used to train the classifier. The algorithm was evaluated with real T2*-weighted brain MRI images. Keywords: Diffuse Axonal Injury, Brain MRI, Brain Lesion Segmentation, Contour Levels. 1. INTRODUCTION Recently automated and semi-automated algorithms for medical image processing are widely used by radiologists during evaluation of a patient’s state and diagnostics of various diseases. In this paper we introduce a semi-automated method for detection of diffuse axonal injury (DAI) lesions on T2*-weighted brain MRI images. Diffuse axonal injury occurs in 50% cases of severe traumatic brain injury. Occurrence of DAI lesions in midbrain and brainstem is the most common cause of coma and subsequent disability. An accurate and convenient method for detection of DAI lesions will help a radiologist to estimate current stage of the disease and make a prognosis of a further disease flow. T2*-weighted MRI was chosen among other MRI modalities because it allows to identify DAI lesions visually rather clearly. The T2*-weighted MRI sequence is the most sensitive to the magnetic susceptibility induced by static field inhomogeneities, arising from paramagnetic blood breakdown products in DAI lesions. They look like small dark spots, usually in brain tissue, sometimes connecting each other. So some automatization can be added to the process. Fully automated algorithms for brain lesion detection usually lack accuracy and stability. True lesions can be easily confused with other formations, for example, blood vessels. Therefore, we developed a semi-automated algorithm. A radiologist interactively specifies rectangular regions of interest (ROI) which are likely to contain DAI lesions. After this marking procedure each MRI section of the current stack may include none, one or several square regions (Figure 1). Then the lesions inside these specified regions are segmented fully automatically. So there is a trade-off between speed and accuracy. Figure 1: ROI selected on T2*-weighted MRI. A proposed algorithm for DAI lesion segmentation, i.e. delineation of lesion contours, is based on contouring algorithm which is aimed to obtain closed contours which may potentially correspond to true lesion contours. In order to distinguish lesion contours from other closed contours machine learning is applied. The rest of the paper is organized as follows. A brief review of previous work is given in section 2. Contouring algorithm purpose and its application to a current problem is described in section 3. Section 4 is devoted to selection of closed contours obtained by contouring algorithm, which correspond to DAI lesions contours, via support vector machine (SVM) classifier. Implementation details and comparison with other methods is given in Section 5. Section 6 includes conclusion and discussion. 2. PREVIOUS WORK There are several works devoted to processing of MR images of brain with DAI. For the most part they deal with medical and practical aspects of the experiment rather than with image processing algorithms. Paper [8] presents a method for quantification of severity of DAI by diffusion tensor images (DTI). Several specific characteristics