Brain Tumor Segmentation Using Deep Neural Networks and Survival Prediction

Abstract

Magnetic resonance image (MRI) is widely applied to the brain tumor diagnosis and treatment. Approximately 35 million MRIs are performed annually worldwide in recent years. Manual segmentation and extraction of the tumor area from MRIs are time-consuming. The emergence of deep-learning algorithms offers new opportunities to automate the medical images processing and analysis with high accuracy. In the study, we built deep-learning models for brain tumor segmentation. The MRI files were first preprocessed through the procedures including reorientation, de-noising, bias-correcting, skull stripping, and co-registration. Then, the two deep-learning algorithms, DeepMedic and 3D U-Net, were used for tumor segmentation model construction. Different from the sequential DeepMedic model, 3D U-net has an encoding and decoding patch that allows shortcut connections from layers with equal resolution in the encoding path to the layers in the decoding path and can provide the high-resolution features. The dice coefficient (DICE), a most commonly used metric for validating medical volume segmentations, was adopted for performance evaluation. Our DeepMedic model achieved DICE of 0.802, whereas 3D U-Net achieved 0.876 for overall segmentation. Moreover, we built a linear regression model using shape features including size and surface area of different segmented tumor tissue sections from the results of 3D U-Net model along with clinical data of age and gross total resection status for patient survival prediction. Compared to using the clinical data alone, we found that combining shape features improved the prediction of overall survival by 7%. We further increased the overall survival prediction accuracy by an additional 9% by replacing the shape features that were not significantly correlated with survival with some selected texture. Our work provided models for automated brain tumor segmentation and patient survival prediction.