A novel multi-class brain tumor classification method based on unsupervised PCANet features

Brain tumors are one of the most severe tumors in the human body because of their nonlinear morphological and textural characteristics. The appropriate clinical practices selected by the surgeons for the patients can be improved by using automated brain tumor diagnosis systems based on magnetic resonance images (MRIs). Therefore, improving the efficiency of these systems plays an essential role in saving the lives of patients. In this paper, a novel multi-class brain tumor classification method is proposed. The proposed method comprises two modules: a simple unsupervised convolutional PCANet module is utilized for feature extraction and a supervised CNN module for feature classification. We modified the PCANet model by applying a nonlinear activation function on the convolutional feature maps. In addition, the learned convolutional PCA filters are computed across the 3D feature maps contrary to the traditional PCANet model, which apply PCA filters on the 2D maps. These modifications enhance the features' expressive power compared with the traditional PCANet and the deep CNN models. The unsupervised features extracted from the modified PCANet module are followed by a simple CNN classification module. The unsupervised PCANet does not require a large number of training data compared with its competitive supervised CNN. We apply grid-search optimization to obtain the optimal hyper-parameters for the classification module. Four public benchmark datasets containing 9581 MRI brain images are utilized to evaluate the performance of the proposed method. The datasets contain different classification tasks, number of samples, image sizes, contrast, and planes. Our proposed method achieved the highest performance compared with other state-of-the-art methods. The proposed method encourages medical imaging diagnosis and can solve the size limitation problem in medical datasets.