U-Net and Active Contour Methods for Brain Tumour Segmentation and Visualization Estera Kot Faculty of Electrical Engineering Warsaw University of Technology Warsaw, Poland kote@ee.pw.edu.pl Zuzanna Krawczyk Faculty of Electrical Engineering Warsaw University of Technology Warsaw, Poland krawczyz@ee.pw.edu.pl Krzysztof Siwek Faculty of Electrical Engineering Warsaw University of Technology Warsaw, Poland Krzysztof.Siwek@ee.pw.edu.pl Piotr S. Czwarnowski Nuclear Medicine Department Medical University Of Warsaw Warsaw, Poland pczwarnowski@wum.edu.pl Abstract—Gliomas are a type of brain and spinal tumour. They originate from glial cells that form the stroma of nerve tissue. Gliomas constitute about 70% of all intracranial tumours and are perceived as the leading cause of death due to brain tumours. This paper presents a comparison of three approaches for glioma detection, volume computation and 3D visualization. The classic approaches based on thresholding and active contour methods were compared with a deep learning implementation. The state-of-the-art model, named U-Net, was used to segment biomedical images which effectively removed bone images from computed tomography (CT) scans. To portion the tumour area, the Morphological Geodesic Active Contour method was applied. Model training was enriched by implementing a data augmentation strategy. Numerical results of tumour volume in cm3 were presented as well as a 3D visualization example. Keywords—Convolutional Neural Networks, U-Net, Brain Tumour, Glioblastoma, Active Contour, CT, PET I. INTRODUCTION Glioblastoma (GBM) are the most malicious primary brain tumours [1], with a 100% mortality rate and the fourth (the highest) histologic classification of the World Health Organization. The median lifetime for patients diagnosed with GBM is 14.6 months [2]. The objective of an established treatment plan is extending life expectancy and retaining quality of life and human functionality. In the Nuclear Medicine Department of the Medical University of Warsaw, at the Central Clinical Hospital, clinical trials are conducted using modern treatment plans for patients diagnosed with a glioblastoma. Following an examination, the standard medical procedure is tumour extraction and introduction of a post-extraction radio and chemistry therapy plan. L– radiation emitter is injected into the area of the postoperative lodge, e.g. 213Bi. Based on the CT and positron emission tomography (PET) scans, radiologists determine the amount of radiological substance, which simultaneously damages two strands of DNA. Radiologists and oncologists determine the volume of radiopharmaceutical inflicted, as well as the value of the probability of local cure TCP (Tumor Control Probability) based on microdosimetry methods, i.e. dosimetric measurements for radiation exposure. Excess volume has the potential to damage healthy brain cells, and insufficient volume will fail to eradicate cancer cells. In order to increase the effectiveness of the introduced method and launch a replicable and error- proof approach with a precisely calculated tumour volume, its localization and shape are required. The objective of this paper is to compare different methods to detect GBM, compute its volume and make a 3D visualization. The classical approach results of brain tumour segmentation were contrasted with a Convolutional Neural Networks (CNN) approach. II. LITERATURE REVIEW Automated segmentation of biological objects such as bone tissues or tumours is a challenging task due to high variability in the shape of such structures and unclear, blurred borders of the objects visible in medical data. Hence, the successful and accurate segmentation is still the subject of ongoing research. Approaches to segmentation of bone structures out of medical data, presented in the literature, include inter alia, variety of active-contour based methods [3]-[5], thresholding operations and region growing methods [6]-[7], morphological methods as well as more elaborate atlas-based methods [8]. Analogously, the state-of-the-art algorithms of tumour segmentation cover region-based methods, i.e. the semi- automatic region growing algorithm described in [9] or watershed algorithm [10]. Another frequently used class of methods is the deformable contour approach [11]-[12]. Because of weak contrast between the tumour and healthy tissue, classical approaches often result in over-segmentation of the detected object. The accuracy of deformable contour or region growing method is also usually dependent on its correct initialisation by the user. In recent years, the approach based on deep convolutional neural networks is becoming popular. The CNN application to medical images allow to obtain promising results both in the area of segmentation of bone structures [13]-[14] as well as in tumour recognition task [15]-[18], and it gives hopes for better automation of the problem. However, what must be emphasised, to obtain satisfactory results, an extensive training set of expert-labelled data is needed. 978-1-7281-6926-2/20/$31.00 ©2020 IEEE