Multiclass Classification of Brain Tumor for MR Images Using Shallow Autoencoder Based Neural Network

Abstract

Brain tumor is an abnormal growth of cells, that may be cancerous or non-cancerous. The earlier prediction, identification, and classification of tumor is essential for rapid diagnosis. In brain MRI, the size and location of tumors can be diverse for different patients. Because of the increased flow of patients in scan centers, patients must now wait for a long time to collect their reports from the radiologists, as it ends up taking the radiologists a long time to classify the images. The proposed methodology in this work can classify tumors from MR brain images into three categories. At first, a shallow autoencoder network is designed for image reconstruction. The encoder segment is made up of three convolutional layers, and in decoder segment, four layers are used for reconstruction. Autoencoder offer excellent noise robustness and feature reduction thereby reducing the possibility of over-fitting. Secondly, to perform classification, an additional convolutional layer is added to the encoder part of neural network along with 2$\times$2 filter. The features extracted from the encoder part were given to a single layer dense neural network and finally testing is performed on SoftMax layer for the classification. The developed algorithm was trained and evaluated on the Cheng dataset, and achieved an accuracy of 95.26%. The developed methods’ outcomes outperform well than the conventional techniques.