Segmentation and classification of brain tumor images using statistical feature extraction and deep neural networks

Cancer happens to be the second leading cause of death globally. One of the deadliest types of cancers with a very high mortality rate is brain cancer, which is characterized by malignant tumors or neoplasm. The general approach to diagnosis is the analysis of brain image scans. However, mere manual inspection of the images can be non-conclusive, and the diagnosis may need to be preceded by a biopsy, after a surgery. The entire process of prognosis and a biopsy may result in the loss of valuable time and advancement of cancer. Hence, automated tools which can analyze brain scan images and classify them may aid the prognosis and speed up the actual course of action. This paper presents an approach based on segmentation, statistical feature extraction and classification based on deep neural networks. The images used in this work is brain Magnetic Resonance Imaging (MRI) scans. The two-dimensional discrete wavelet transform has been employed for denoising the raw data, followed by statistical feature extraction. Two deep neural network architecture has been proposed which are the Deep Bayes Network (DBN) and the Residual Network (ResNet). The performance of the deep neural network has been evaluated in terms of classification accuracy. The classification of brain MRI images has been done in three categories of images which are meningioma, glioma and pituitary tumors. It has been shown that the proposed method attains an average classification accuracy of 98.13% and 96.89 using DBN and the ResNet respectively, for the used dataset. A comparative analysis with existing work shows improvement in classification accuracy with respect to existing techniques.