International Journal of Scientific Engineering and Science Volume 5, Issue 9, pp. 45-50, 2021. ISSN (Online): 2456-7361 45 http://ijses.com/ All rights reserved Brain Tumour Segmentation Using JeisloNet - a Unet Architecture Oluwole Abiodun Adegbola 1 , Peter Olalekan Idowu 2 , Tolu Lydia Adebisi, Joshua Adeleke 4 , Demilade Oludide Babajide 5 , John Adedapo Ojo 6 1 Department of Electronic and Electrical Engineering (EEE), Ladoke Akintola University of Technology (LAUTECH), Ogbomoso, Nigeria 2, 3, 4, 5, 6 Department of EEE, LAUTECH, Ogbomoso, Oyo State, Nigeria Abstract— Manual medical images segmentation is very tedious task. Accurate segmentation of brain magnetic resonance images (MRI) is a critical phase in measuring the irregularities in brain structure. In recent years, deep learning has gained popularity for its efficiency in brain image segmentation. In particular, Unet architecture is being deployed in several biomedical fields for segmentation. It has contributed immensely to solving clinical problems, which includes accurate segmentation of desired feature, efficient processing, and analysis of biomedical images, therefore enhancing an improved accuracy in biomedical images diagnosis and prognosis. In this research, JesloNet, a Unet architecture was developed for the automatic segmentation of abnormal tissues, brain tumours in MR scan images. MR scanned brain images were obtained from The Cancer Image Archives, the images were preprocessed and split in train and test set in the ratio 80:20. The experiment results showed that the Unet model, JeisloNet achieved good performance with the following results in Dice Coefficient Index (DSC) 0.9931, Mean IOU 0.9321, Global Accuracy 0.9928, and Error rate 0.0072. The result was also compared with other methods in the literature an JeisloNet performed better, hence, it can be adopted for other medical image segmentation. Keywords— Brain tumour, unet, segmentation, magnetic resonance image, biomedical image. I. INTRODUCTION Brain and nervous system cancer happen to be prominent cause and reason of death today, having a world ranking of number 10 [1] [2]. Over the years, there has been a rapid growth in brain cancer cases [2], this therefore calls for the need of an automated diagnosis systems which help to detect early-stage tumour and reduce the intervention of clinicians. In the past, accurate diagnosis of brain cancer was a laborious and tedious task for neurologists leading to the invention of modalities such as X-ray, Magnetic Resonance Imaging (MRI) and the computed tomography (CT) which provides visualization of the brain structure, thereby simplifying the task of brain cancer diagnosis. However, MR images are more preferable as they provide detailed information about the tumour position, type, size, and better contrast images with higher spatial resolution, thus making it an essential tool in treatment, diagnosis, and monitoring of disease [3]. MR imaging is widely used for medical images segmentation such as liver segmentation, brain tumour segmentation, and breast tumour segmentation among others, an important procedure for automated diagnosis system. Automated segmentation method is therefore necessary since manual segmentation which involves labelling of pixel/voxel can be difficult a task as well as large time consumption. The conventional brain tumour MR images protocol includes T1-weighted imaging, T2-weighted imaging (including Fluid Attenuated Inversion Recovery i.e., FLAIR) and gadolinium-enhanced T1-weighted imaging sequences. These structural MR images provide a quality prognosis and diagnosis in most cases [4]. Initially, traditional method such as boundary extraction, regionbased segmentation, threshold-based segmentation and clustering- based segmentation was used for brain tumour segmentation, but over time these traditional methods became difficult to learn manual features when large data are used, therefore being unable to assist neurologist in accurate diagnosis or analysis and treatment of disease [5]. In recent times, Deep Learning (DL) methods are becoming popular in the area of medical images segmentation due to the aptitude to efficiently process enormous quantity of data and extraction of useful image features, it allows direct learning of complex features from the original data making it a requisite for medical image analysis [5]. DL technology is a subclass of Machine Learning (ML) technology that acquires data representations by improving abstraction levels [6]. Among different DL models that exist, Convolutional Neural Networks (CNNs) has shown outstanding success in almost all sophisticated task of computer vision such as image segmentation, object detections and image classification, through the advent and success of AlexNet developed by Krizhevsky et al [7] which revolved the field of computer vision from customary ML algorithms towards CNNs. According to researchers, the perception of CNNs was an innovation. It began from the discovery of Hubel and Wiesel elucidating that in the primary visual cortex there are simple and complex neurons, these simple structures such as oriented edges are the basic step of visual processing. Hubel and Wiesel work greatly motivated Kunihiko Fukushima and he developed a multi-layered neural network named Neocognitron [8] with the use of simple and complex neurons which could identify outlines in images and spatial invariance [9]. Using the Neocognitron idea the LeNet [10] was developed by Yann LeCun which was used for identifying handwritten digits, LeNet however became the first CNN built [11]. The breakthrough by Krizhevsky et al. [7] was a result of a supervised and overseen training exercise of a huge network having eight layers and parameters in millions on the