Deep learning based 3D brain tumor segmentation

Abstract

Brain tumor segmentation plays a crucial role in medical image analysis and assists in the diagnosis, treatment planning, and monitoring of brain tumor patients. However, accurately segmenting brain tumors from multi-modal medical images remains a challenging task due to the complex and heterogeneous nature of tumors, tissues. In thisthesis,weproposea3DU-netdeeplearningmodelforthepurpose of having accurate 3D brain tumor segmentation from multi-modality data to take advantage of different levels of information that exist in theMRIsequences,inadditiontoreducingthediagnosistimeandhaving an automated process to address the problem from like-real data and scale down the human bias when dealing with such a sensitive task. The 3D U-net model is trained on the BRATS2020 data-set and evaluated with segmentation volumetric metrics. The model showed promising results on the majority of the test data 63 % that reached high IOU score for the whole tumor region and showed a tolerance to variationafterapplyingthetestonnoisydata. However,itmissclassifiedthe37%ofthetestdatabecauseoftheimbalancebetweenclasses. as a recommendation to tackle this problem; a customized approach to deal with imbalance conditions can be proposed.