A Practical Approach to Automated Segmen- tation of Brain Tumours in MRI Practicum report Aleksandar Miladinović (Student id: 1227841) aleksandar.miladinovic@univie.ac.at Introduction Expert brain tumor identification on multi-modal Magnetic Resonance (MR) images is a very time-consuming process for medical experts. Therefore, throughout the last decade, significant effort has been invested in the development of novel approaches applying com- puter-aided techniques for brain tumor segmentation. Automated brain tumor segmentation aims to separate and label different tumor tissues, including: (1) active tumor cells, (2) ne- crotic core, and (3) edema from normal brain tissues of Gray Matter (GM), White Matter (WM), and Cerebrospinal fluid (CSF). Even though experts from the brain tumor research area can accurately characterize and identify brain tissue abnormalities, the automated pro- cess of tumor segmentation is not straightforward. Many accurate approaches are only evaluated on a single data type coming from a particular brain tumor type, and thus, being far away from a practical clinical application [4] The aim of this work is to evaluate already existing frameworks and recent techniques of automated brain tumor segmentation in realistic clinical settings. We will use the algo- rithms on a coherent study data set and evaluate how they perform based on a small anno- tated data set. The algorithm choice will be based on quantitative evaluations from [5,8]. Additionally, internal requirements for the selection will be defined (i.e. degree of user supervision, ro- bustness on the various data, etc.). The dataset consists of multi-contrast MR clinical imag- es from the repository of Medical University of Vienna. The robustness of the selected approach will be evaluated using test data from [5,8]. The performance measure includes segmentation accuracy of each tumor sub-region defined in 1-3, brain tumor type and the severity of the pathological case. Requirements and Approaches selections The evaluation process has been based on the on the extensive literature research and re- sults from different competitions where different authors can demonstrate their approaches and techniques. The primary source for approach pre-selection is based on the performance evaluation in BRATS 2013 benchmark [5][8]. Due to time constraints, the secondary crite- ria for algorithm pre-selection is the availability of the technical implementation. The goal practicum is practically oriented. Thus, only currently available solutions have been taken into consideration. The third criterion for the evaluation is the compatibility of the ap- proach with existent and clinical most used MRI (T1,T2,T1c and FLAIR) sequences. The selection approaches were: (1) BraTumIA[5][1][2] 1 , (2) GLISTR[3] 2 , (3) EMS MENZE [6] 3 1 https://www.nitrc.org/projects/bratumia/ 2 https://www.cbica.upenn.edu/sbia/software/glistr/ 3 https://bitbucket.org/s0216660/brain_tumor_segmentation_em/overview