3D U-NET FOR BRAIN TUMOUR SEGMENTATION

The brain is the most complex part of the human body that controls memory, emotions, touch, motor, thought, skills, vision, breathing, temperature and every process related to regulating the human body. Usually, the brain tumour can be classified as malignant or benign, and it can spread to other regions as sometimes not. (MRI) Magnetic Resonance Imaging is the most common exam to identify the tumours, and later the resection surgery as a decision that has to be made from neurosurgeon. The specialized doctor has to mark the tumour region precisely and a manual method can be high time-consuming work for a doctor. Deep Learning has improved and demonstrated efficiency in many tasks, in special, the (CNN)Convolutional Neural Networks has achieved the state-of-the-art in performing medical image segmentation tasks. Nowadays it is possible to segment tumours in any kind of shape, size and contrast through machine learning models, using MRI images. 3 D U-Net architecture was selected because it is specialized for brain tumours segmentations and has been successfully utilized for this type of tasks in the research field. Generally, U-Net architectures utilize fully convolutional networks with the encoder-decoder strategy, it has variants which are proposed to apply to the brain segmentation task. The traditional U-Net, the Attention U-Net a variant of the traditional and it differentiates by adding gates on the top of its architecture and the Residual U-Net which utilizes a shortcut connection as variation of the architecture. The three architectures are successful in medical segmentation tasks, the proposal is to identify the best performance and accuracy overall, specially when related to healthcare application it is critical to highly achieve accuracy in the training. The academic reference of this thesis is based on other researchers work which applied different models into the same task and solved the same problem in different way, some using different datasets and training strategy, from there it emerges the interest to understand how this U-Net variants works when is trained with same dataset and strategy. The dataset to be utilized is a public, open source available on Kaggle, which is very known for competitions for learning data science and AI activities, this dataset is called Brats 2021 and it contains MRI images scans divided and annotated by specialist’s neuro-radiologists, the images are in nii.gz format file and classified accordingly to the specific diagnosis