Brain tumor segmentation based on 2D U-Net using MRI multi-modalities brain images

A brain tumor is a mass of abnormal cells that grow uncontrollably in the brain or central spine. Glioblastoma is the most common malignant primary brain tumor, accounting for over 48.6 \% of all malignant primary brain tumors, and 14.5 \% of all primary brain tumors. Regardless of the diagnosis, the morbidity and mortality represented by a tumor located in the brain remain significant. They have a 5-year relative survival rate of about 7.2 \%, which means that only 7.2 \% of people diagnosed with glioblastoma will live 5 years after diagnosis; the median survival time is about 8 months. Detecting the presence of brain tumors using magnetic resonance images in a rapid, precise, and reproducible manner is a difficult but still necessary task. Brain tumor segmentation is a key factor for the analysis of gliomas, it involves the delineation of the different tumor regions including peritumoral edema, enhancing tumor, and tumor core. Currently, different methodologies in the literature implement the tumor segmentation technique to improve the diagnosis and treatment plan. The proposed methodology consists of an automated segmentation of brain tumors with the use of neural networks in particular U-Net and Attention U-Net, which are mainly used for fast and accurate segmentation of biomedical images. The BraTS 2020 dataset is used in this study to test the egmentation performance of the suggested technique. An accuracy of 0.9950 was obtained for both models, and a sensitivity of 0.9931 and 0.9891 for the U-Net and Attention U-Net models. A peritumoral edema, enhancing tumor, and tumor core dice similarity coefficient of 0.8453, 0.6950, and 0.7429 respectively has been achieved, for the U-Net model. For the Attention U-Net model, a dice score of 0.8829 0.7233, and 0.8090 were obtained. The study presents and discusses further quantitative and qualitative assessments. Results show that both approaches have considerable potential and can be employed in clinical practice in the segmentation of various sub-regions of brain tumors.

• Autor principal: Tene Hurtado, Daniela Verónica (author)
• Formato: bachelorThesis
• Idioma: eng
• Publicado em: 2021
• Assuntos: Glioma; Segmentación de tumores cerebrales; U-Net; Redes neuronales; Brain tumor segmentation; Multimodal MRI; BraTS dataset; Neural networks
• Acesso em linha: http://repositorio.yachaytech.edu.ec/handle/123456789/460