Pixel-Level Feature Space Modeling and Brain Tumor Detection Using Machine Learning

We present a feature learning technique that enhances the performance of a machine learning technique to detect brain tumor regions at pixel-level in a magnetic resonance imaging (MRI) brain scan. This technique utilizes the image filtering based feature extraction techniques (e.g., Laplacian, Gradient, and Sobel filters) to construct a feature space. It then maps the feature space to a response set that has predefined labels (manual or automated). Feature space and the response set are constructed using a reference frame of a volumetric MRI brain scan, and then used for developing a ML model. We applied the learned models, as the automated techniques to other frames of the MRI scan for detecting tumor and non-tumor regions. We adapted the Brain Tumor Segmentation (BraTS 2015) datasets to develop and validate the proposed computational framework. We also used the ground truth labels (or response sets) delivered by the BraTS 2015 datasets. We evaluated the support vector machine (SVM), random forest (RF), and artificial neural network (ANN) models using various quantitative and qualitative measures. We determined, based on precision-recall curve, that the RF model acquired 92% of the tumor detection skills of a perfect model, while ANN and SVM acquired 90% and 88% tumor detection skills of a perfect model.