MobDenseNet: A hybrid deep learning model for brain tumor classification using MRI

Abstract

This paper presents MobDenseNet, an improved deep learning model that assists medical practitioners in diagnosing brain tumors accurately. The proposed MobDenseNet is developed using the concepts of existing deep learning models: MobileNetV1 and DenseNet; the model incorporates hyperparameter fine-tuning and feature fusion ensemble during the feature extraction phase, consolidating layers like batch normalization, dense layers in the classification step to classify brain tumors. The classification is done into multiple classes, including, gliomas, meningiomas, pituitary, and healthy brain. The performance of the proposed model is assessed on two benchmark datasets. The experiments consider 2757 training and 307 testing images for the first dataset of 3064 MRI images, available on Figshare, having classes of glioma, meningioma, and pituitary. The experiment for the second dataset, has 2937 training and 327 testing images with glioma, meningioma, pituitary, and no tumor classes. The model achieves 98.4 % accuracy, 99.9 % AUC, 98.6 % precision, 98.40 % recall, 98.5 % F1-score for the Figshare dataset, and 96.02 % accuracy, 99.4 % AUC, 96.3 % precision, 95.7 % recall and 95.9 % F1-score for the Sartaj Bhuvaji dataset, respectively. The proposed MobDenseNet shows better accuracy than the existing models considered in the research. To demonstrate the effectiveness of the proposed model on diverse and unseen data, cross-dataset evaluations are conducted, where the model is trained using the Figshare dataset and tested using the Sartaj Bhuvaji and two additional datasets. Results indicate that even for the cross-dataset scenario, the proposed model achieves acceptable classification accuracy and outperforms existing models of MobileNetV1 and DenseNet.