Machine Learning Algorithms for Radiogenomics: Application to Prediction of the MGMT promoter methylation status in mpMRI scans

Glioblastomas are the most aggressive and destructive forms of solid brain tumors of the central nervous system. Despite aggressive multimodal treatment approaches, the overall survival period is reported to be less than 15 months after diagnosis. MGMT gene silencing is a potentially proper predictive element in determining the mortality rate for glioblastoma patients undergoing chemotherapy. Analyzing the correlation between different medical image characteristics and MGMT promoter methylation status through machine learning tools could play an essential role in the automatic aided diagnosis approach. By doing data preprocessing and transformation, this thesis aims to extract features from data provided by the Radiological Society of North America (RSNA) and investigate this relationship with various classification methods like Logistic Regression (LR), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP). Using Pyradiomics, an open-source python package, 1153 features have been extracted from the images, their segmented forms, and Laplacian of Gaussian (LoG) and Wavelet filters to get features with more details. The Extreme Gradient Boosting (XGBoost) classifier is used since it is desirable to reduce the number of features to improve performance and identify the optimal number for the most relevant results. The model used nested cross-validation to prevent information leakage and obtain better outcomes by finding the best set of hyperparameters for the chosen models. This approach is also exclusively applied to different MRI sequence types such as Fluid Attenuated Inversion Recovery (FLAIR), T1-weighted pre-contrast (T1w), T1-weighted post-contrast (T1Gd), and T2-weighted (T2) in order to point out the importance of each for final user support and future research purposes. Different statistical metrics such as accuracy, F1 score, and confusion matrix are considered for each classification algorithm. This study demonstrated acceptable performance by the proposed feature extraction, feature selection methods, and machine learning classification algorithms. Although the deep learning approach would result in better performance metrics, considering computational load and time-space trade-off, using radiomics features and performing classification lead to valuable results in quicker time for the final user. Undoubtedly, better results can be obtained by accessing more extensive data, which points out the importance of data quantity. Analyzing different optimal features also would be a good starting point for future research to focus on the most critical aspect of brain tumor MRI images.