Explainable Deep Ensemble of Brain Tumor Classification based on MRI

Abstract

Heterogeneity of cerebral neoplasms, loss of information labeling, and inter-class similarity make it difficult not only to differentiate the tumor clinically but also to establish separation of subtypes, taking Magnetic Resonance Imaging (MRI) into consideration. It presents a neurotumor multiclass classifier, which has a probabilistic soft voting ensemble as the classification method, based on the transfer learning model using pre-trained Convolutional Neural Networks (CNNs), using VGG16, ResNet50 and fine-tuned EfficientNetB0 as the basic classifiers. The experimental data is divided into four different categories: glioma, meningioma, pituitary and no visible tumor. The evaluation is carried out with the use of conventional performance indicators, which include recall, f1-score, confusion matrices, and a micro-averaged ROC analysis. Grad-CAM predicts clinical confidence and interpretability. The fine-tuned EfficientNetB0 model was found to have the highest test accuracy of 73.35%, compared to the equal-weight soft-voting ensemble with test accuracy 70.81%. The results indicate that the ensemble outperformed the ResNet50 and was marginally better than the VGG16, but still not better than the best EfficientNetB0 model. More frequent misclassification patterns, especially class confusion, are compared, and suggestions are made to enhance the sensitivity of the tumor.