Explainable hybrid CNN model using DenseNet121, and EfficientNetB0 for brain tumor detection and classification

Abstract

• A new hybrid method is proposed for detecting and classifying brain tumors. • The LBP extracts local features, while EfficientNetB0 extracts global features. • The model provides interpretability using local interpretable model-agnostic explanations (LIME). • We are leveraging LIME to interpret the hybrid model results, which will help us better understand the decision-making process and enhance trust. • The proposed model is superior to existing models such as DenseNet121, ResNet101, InceptionV3, and MobileNet. Early detection and classification of brain tumors are vital, given their potential to be life-threatening. Nevertheless, the wide range of histological variations makes it challenging for artificial intelligence models to generalize across different tumor types. Recent breakthroughs in artificial intelligence have successfully utilized various deep learning approaches to accurately detect and categorize brain malignancies. Hence, these approaches must be precise and effectively interpret their findings, as misdiagnosis can result in inappropriate treatment and prolonged recovery. Therefore, we introduce an explainable hybrid convolutional neural network that utilizes transfer learning using DenseNet121 and EfficientNetB0. This model does not rely on features extracted from a standalone network, thereby enhancing its ability to classify various types of brain tumors. In addition, a new classifier was used, comprising fully connected layers and a SoftMax layer. The approach was evaluated using two publicly available datasets. The first dataset has three classes of brain tumors: meningioma, glioma, and pituitary. On the other hand, the second one has two classes: non-tumor and tumor scans. The model achieved 99.13% and 99.56% accuracy, 98.94% and 99.56% precision, 98.98% and 99.56% recall, and 98.96% and 99.56% F1-score on three-class and two-class datasets, respectively. The proposed approach generally outperformed prior research in classifying brain tumors on MRI scans, and the implemented explainable model increased confidence.