Brain Tumor Segmentation using 3D U-Net with Hyperparameter Optimization

For the sake of proper diagnosis and treatment, accurate brain tumour segmentation is required. Because manual brain tumour segmentation is a time-consuming, costly, and subjective task, effective automated approaches for this purpose are generally desired. However, because brain tumours vary greatly in terms of location, shape, and size, establishing automatic segmentation algorithms has remained challenging throughout the years. Automatic segmentation of brain tumour is the process of separating abnormal tissues from normal tissues, such as white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF). Brian segmentation needs typically to be carried out for different image modalities in order to reveal important metabolic and physiological information. These modalities include positron emission tomography (PET), computer tomography (CT) image and magnetic resonance image (MRI). Multimodal imaging techniques (such as PET/CT and PET/MRI) that combine the information from multiple imaging modalities contribute more for accurate brain tumour segmentation. In this work, we introduce a deep learning framework for automated segmentation of 3D brain tumors that can save physicians time and provide an accurate reproducible solution for further tumor analysis and monitoring. In particular, a 3D U-Net was trained on brain MRI data obtained from the 2018 Brain tumor Image Segmentation (BraTS) challenge. Three optimizers (RMSProp, Adam and Nadam) and three loss functions (Dice loss, focal Tversky loss, Log-Cosh loss functions) were used. We demonstrated that some loss functions and optimizers combinations perform better than other ones. For example, using the Log-Cosh loss function along with RMSProp optimizer resulted in the highest Dice coefficient, 0.75. Indeed, we also optimized the network hyperparameters in order to enhance the segmentation outcomes. These results demonstrate the feasibility and effectiveness of the proposed deep learning scheme with optimized hyperparemeters and appropriate selection of the optimizer and loss function.