Hybrid Pooling Based Convolutional Neural Network for Multi-class Classification of MR Brain Tumor Images

Abstract

AbstractThis paper aims to improve the conventional convolutional neural network (CNN) for classification problem of the MRI brain tumor images by generalizing the pooling operations that play a central role in spatial dimensionality reduction. A hybrid pooling operation based on max and average pooling is developed to learn and adopt the complex and variable patterns of the tumor regions without discarding or distorting the information. The proposed hybrid pooling operation boost up the invariance properties when used in place of average or max pooling. In addition, the performance of the proposed hybrid pooling based CNN is evaluated by performing a ten-fold cross-validation method. Results indicated that the proposed network structure achieves a mean accuracy, precision, recall, and F1-score of 99.48%, 99.49%, 99.68%, and 99.42% respectively. The computation time during the training stage is increased slightly by 4.5% and 5.2% compared to the conventional max and average pooling based CNN model. However, the classification accuracy is improved by approximately 0.56% and 0.49% from the single max and average pooling based model respectively. Thereby, this incremental classification accuracy could add an effective decision-support to the radiologists or physicians for brain cancer treatment.KeywordsBrain tumorConvolutional neural networkImage classificationMax-poolingAverage-poolingHybrid pooling