Brain tumor

In recent years, brain tumors become the leading cause of death in the world. Detection and rapid classification of this tumor are very important and may indicate the likely diagnosis and treatment strategy. In this paper, we propose deep learning techniques based on the combinations of pre-trained VGG-16 CNNs to classify three types of brain tumors (i.e., meningioma, glioma, and pituitary tumor). The scope of this research is the use of gray level of co-occurrence matrix (GLCM) features images and the original images as inputs to CNNs. Two GLCM features images are used (contrast and energy image). Our experiments show that the original image with energy image as input has better distinguishing features than other input combinations; accuracy can achieve average of 96.5% which is higher than accuracy in state-of-the-art classifiers.

The accuracy of brain tumor detection and segmentation are greatly affected by tumors’ location, shape, and image properties. In some situations, brain tumor detection and segmentation processes are greatly complicated and far from being completely resolved. The accuracy of the segmentation process significantly influences the diagnosis process, such as abnormal tissue detection, disease classification, and assessment. However, medical images, in particular, the Magnetic Resonance Imaging (MRI), often include undesirable artefacts such as noise, density inhomogeneity, and partial volume effects. Although many segmentation methods have been proposed, the accuracy of the segmentation results can be further improved. Subsequently, this study attempts to provide very important properties about the size, initial location and shape of tumors known as Region of Interest (RoI) to kick-start the segmentation process. The MRI consists of a sequence of images (MRI slices) of a particular person and not one image. Our method chooses the best image among them based on the tumor size, initial location and shape to avoid the partial volume effects. The selected algorithms to test our method are Active Contour and Otsu Thresholding algorithms. Several experiments are conducted in this research using the BRATS standard dataset that consist of 100 samples. These experiments comprised of MRI slices of 65 patients. The proposed method is evaluated by the similarity coefficient as a standard measure using Dice, Jaccard, and BF scores. The results revealed that the Active Contour algorithm has higher segmentation accuracy when tested across the three different similarity coefficients. Moreover, the achieved results of the two algorithms verify the ability of the proposed method to choose the best RoIs of the MRI samples.