EfficientNet and mixed convolution network for three-class brain tumor magnetic resonance image classification

Abstract

The classification of brain tumor images is the prevalent task in computer-aided brain tumor diagnosis. Recently, three-class classification has become a superlative task in brain tumor type classification. The existing models are fine-tuned for a single dataset, and hence, they may exhibit displeasing results on other datasets. Thus, there is a need for a generalized model that can produce superior performance on multiple datasets. In this paper, we have presented a generalized model that produces similar results on two datasets. We have proposed an EfficientNet and Mixed Convolution Network model to perform a three-class brain tumor type classification. We have devised a mixed convolution network to enhance the feature vector extracted from pre-trained EfficientNet. The proposed network consists of two blocks, namely, separable convolution and residual convolution. We have utilized a Gaussian dropout layer before the softmax layer to avoid model overfitting. In our experiments, two publicly available datasets (BTDS and CPM) are considered for the evaluation of the proposed model. The BTDS dataset has been segregated into three tumor types: Meningioma, Glioma, and Pituitary. The CPM dataset has been divided into three glioma subtypes: Glioblastoma, Oligodendroglioma, and Astrocytoma. We have achieved an accuracy of 98.04% and 96.00% on BTDS and CPM datasets, respectively. The proposed model outperforms existing pre-trained models and state-of-the-art models in vital metrics.