Novel Robust Automatic Brain-Tumor Detection and Segmentation Using Magnetic Resonance Imaging

Computer-aided automatic brain tumor detection has been a very important biomedical engineering research problem for years. As the sizes and shapes of tumors vary dramatically from a subject to another, how to build a robust automatic brain-tumor detector is still quite challenging nowadays. There exist many signal processing approaches for automatic brain-tumor detection. In this work, we focus on the approach based on active contour models for magnetic resonance imaging (MRI) as it not only can detect the existence of one or more tumors but also can measure the size(s), delineate the shape(s), and find the location(s) at the same time. On the other hand, the aforementioned purpose cannot be achieved by a single deep-learning model easily. It would be quite difficult for the existing deep-learning-based methods to select the appropriate region(s) of interest so as to extract reliable features for brain-tumor detection since tumors of various sizes and shapes could be present in an MRI image. Here we propose a novel brain-tumor detection and segmentation scheme using active contouring in tandem with texture-based decision metric for MRI images. In our proposed new scheme, the active contouring mechanism is first employed to automatically find one or more suspected regions (image segments). Then, these suspected image segments are further evaluated using our proposed new texture-based decision-making criterion to determine whether any of them actually involves a true tumor. We also test the effectiveness of our proposed new scheme using the well-known Brain Tumor MRI Dataset and the area-under-curve (AUC) and the average correct segmentation area-ratio resulting from our new method are 0.9244 and 0.8019 respectively, which is much better than the existing deep-learning and active-contour based brain-tumor detection methods.