Brain tumor categorization from imbalanced MRI dataset using weighted loss and deep feature fusion

Deep learning-based brain tumor classification from brain magnetic resonance imaging (MRI) is a significant research problem. The research problem encounters a major challenge. The training datasets used to develop deep learning algorithms could be imbalanced with significantly more samples for one type of tumor than others. This imbalance in the training dataset affects the performance of tumor classification using deep learning models as the classifier performance gets biased towards the majority class. The article addresses the challenge of training data imbalance by proposing a novel class-weighted focal loss and studies the effects of weighted loss functions on feature learning by convolutional neural networks (CNN). However, finding optimal class weights is a challenge and the predictions of CNN trained using weighted functions could be biased. The article presents two approaches to improve the performance of the expert system: deep feature fusion and majority voting on classifier predictions. In the first approach, the deep feature fusion concerns the fusion of deep features extracted from CNN models trained using separate loss functions. The fused deep features are classified using proven models, such as support vector machine (SVM) and k-nearest neighbours (KNN). In the other approach, a majority voting is performed on the predictions for three different feature sets extracted from CNN models trained using separate loss functions. The majority voting uses the same classifier upon three different feature sets. The proposed approaches show a significant improvement in brain tumor predictions over a state of the art method based on CNN trained using cross-entropy loss. The classification errors between the majority class and the minority class samples are reduced considerably in the proposed strategies. The experiments are evaluated using the Figshare dataset, and the performance improved for the metrics: accuracy, precision, recall, balanced accuracy and F-scores.