Radiomic Features to Predict Overall Survival Time for Patients with Glioblastoma Brain Tumors Based on Machine Learning and Deep Learning Methods

Machine Learning (ML) methods including Deep Learning (DL) Methods have been employed in the medical field to improve diagnosis process and patient’s prognosis outcomes. Glioblastoma multiforme is an extremely aggressive Glioma brain tumor that has a poor survival rate. Understanding the behavior of the Glioblastoma brain tumor is still uncertain and some factors are still unrecognized. In fact, the tumor behavior is important to decide a proper treatment plan and to improve a patient’s health. The aim of this dissertation is to develop a Computer-Aided-Diagnosis system (CADiag) based on ML/DL methods to automatically estimate the Overall Survival Time (OST) for patients with Glioblastoma brain tumors from medical imaging and non-imaging data. This system is developed to enhance and speed-up the diagnosis process, as well as to increase understanding of the behavior of Glioblastoma brain tumors. The proposed OST prediction system is developed based on a classification process to categorize a GBM patient into one of the following three survival time groups: short-term (months), mid-term (10-15 months), and long-term (\u3e15 months). The Brain Tumor Segmentation challenge (BraTS) dataset is used to develop the automatic OST prediction system. This dataset consists of multimodal preoperative Magnetic Resonance Imaging (mpMRI) data, and clinical data. The training data is relatively small in size to train an accurate OST prediction model based on DL method. Therefore, traditional ML methods such as Support Vector Machine (SVM), Neural Network, K-Nearest Neighbor (KNN), Decision Tree (DT) were used to develop the OST prediction model for GBM patients. The main contributions in the perspective of ML field include: developing and evaluating five novel radiomic feature extraction methods to produce an automatic and reliable OST prediction system based on classification task. These methods are volumetric, shape, location, texture, histogram-based, and DL features. Some of these radiomic features can be extracted directly from MRI images, such as statistical texture features and histogram-based features. However, preprocessing methods are required to extract automatically other radiomic features from MRI images such as the volume, shape, and location information of the GBM brain tumors. Therefore, a three-dimension (3D) segmentation DL model based on modified U-Net architecture is developed to identify and localize the three glioma brain tumor subregions, peritumoral edematous/invaded tissue (ED), GD-enhancing tumor (ET), and the necrotic tumor core (NCR), in multi MRI scans. The segmentation results are used to calculate the volume, location and shape information of a GBM tumor. Two novel approaches based on volumetric, shape, and location information, are proposed and evaluated in this dissertation. To improve the performance of the OST prediction system, information fusion strategies based on data-fusion, features-fusion and decision-fusion are involved. The best prediction model was developed based on feature fusions and ensemble models using NN classifiers. The proposed OST prediction system achieved competitive results in the BraTS 2020 with accuracy 55.2% and 55.1% on the BraTS 2020 validation and test datasets, respectively. In sum, developing automatic CADiag systems based on robust features and ML methods, such as our developed OST prediction system, enhances the diagnosis process in terms of cost, accuracy, and time. Our OST prediction system was evaluated from the perspective of the ML field. In addition, preprocessing steps are essential to improve not only the quality of the features but also boost the performance of the prediction system. To test the effectiveness of our developed OST system in medical decisions, we suggest more evaluations from the perspective of biology and medical decisions, to be then involved in the diagnosis process as a fast, inexpensive and automatic diagnosis method. To improve the performance of our developed OST prediction system, we believe it is required to increase the size of the training data, involve multi-modal data, and/or provide any uncertain or missing information to the data (such as patients\u27 resection statuses, gender, etc.). The DL structure is able to extract numerous meaningful low-level and high-level radiomic features during the training process without any feature type nominations by researchers. We thus believe that DL methods could achieve better predictions than ML methods if large size and proper data is available