Two-step majority voting of convolutional neural networks for brain tumor classification

Abstract

Brain tumor type classification is essential for determining further examinations. Convolutional neural network (CNN) model with magnetic resonance imaging (MRI) image input can improve brain tumor classification performance. However, due to the highly variable shape, size, and location of brain tumors, increasing the performance of tumor classification requires consideration of the results of several different CNN models. Therefore, we proposed a two-step majority voting (MV) on the results of several CNN models for tumor classification. The CNN models included InceptionV3, Xception, DensNet201, EfficientNetB3, and ResNet50; each was customized at the classification layer. The initial step of the method is transfer-learning for each CNN model. The next step is to carry out two steps of MV, namely MV on the three CNN model classification results at different training epochs and MV on the results of the first step. The performance evaluation of the proposed method used the Nickparvar dataset, which included MRI images of glioma, pituitary, no tumor, and meningioma. The test results showed that the proposed method obtained an accuracy of 99.69% with a precision and sensitivity average of 99.67% and a specificity of 99.90%. With these results, the proposed method is better than several other methods.