Deep Learning-Based Approaches for Brain Tumour Segmentation and Classification

Abstract

Brain tumours are caused by the abnormal growth of cells in the brain. This occurs mainly due to genetic changes or exposure to X-ray radiation. When the tumours are detected early, they can be removed via surgery. The tumour can be removed through radiotherapy and chemotherapy if the removal of the tumour through surgery affects the survival rate. There are two main classifications of tumours: malignant or cancerous and benign or non-cancerous. Deep learning techniques are considered as they require more minimal human intervention than machine learning; they are built to accommodate huge amounts of unstructured data, while machine learning uses traditional algorithms. Though deep learning takes time to set up, the results are generated instantaneously. In this review, the authors focus on the various deep learning techniques and approaches that could detect brain tumours that were analysed and compared. The different types of deep learning approaches investigated are convolutional neural network (CNN), cascaded CNN (C-CNN), fully CNN and dual multiscale dilated fusion network, fully CNN and conditional random field, U-net convolutional network, fully automatic heterogeneous segmentation using support vector machine, residual neural network, and stacked denoising autoencoder for brain tumour segmentation and classification. After reviewing the algorithms, the authors have listed them based on their best accuracy (U-net convolutional network), dice score (residual neural network), and sensitivity score (cascaded CNN).