Segmentation of Brain Tumour Using UNET Architecture

Abstract

Gliomas is the most widely recognized essential brain cancer with particular degrees of aggressive, variable prediction and different heterogeneous histological sub-regions including: edema, necrotic center, enhancing and non-enhancing tumour core. Exact tumour division of sub-areas plays a crucial role for diagnosis, treatment planning and risk factor identification. This study was conducted by using the MultiResUnet architecture for computerized segmentation of gliomas in multimodal MRI scans (FLAIR, T1CE, T2). Previously, the images were enhanced through various preprocessing stages, then they were fed by the MultiResUnet network. From the 3D database gave by BraTS 2019, we extracted 2176 images to train and 1056 images to assess our network. The segmentation model performance was qualitatively evaluated with the accuracy and under the Dice score metrics. Our study suggested a model which Dice score accomplished 0.78, 0.82 and 0.88 on the training set and 0.63, 0.73 and 0.86 on the validation set for the enhancing tumour, the tumour core, and the entire tumour, separately. In general, the system sectioned the entire tumour more precisely than it accomplishes for the core of tumour or enhancing tumour. They are significantly harder to segment because of the closeness between all areas. We plan to further our network to improve the testing phase results and use sub-regions tumour on Radiomic Features extraction for tumour investigation and survival expectation.