BrainNet: a fusion assisted novel optimal framework of residual blocks and stacked autoencoders for multimodal brain tumor classification

A significant issue in computer-aided diagnosis (CAD) for medical applications is brain tumor classification. Radiologists could reliably detect tumors using machine learning algorithms without extensive surgery. However, a few important challenges arise, such as (i) the selection of the most important deep learning architecture for classification (ii) an expert in the field who can assess the output of deep learning models. These difficulties motivate us to propose an efficient and accurate system based on deep learning and evolutionary optimization for the classification of four types of brain modalities (t1 tumor, t1ce tumor, t2 tumor, and flair tumor) on a large-scale MRI database. Thus, a CNN architecture is modified based on domain knowledge and connected with an evolutionary optimization algorithm to select hyperparameters. In parallel, a Stack Encoder–Decoder network is designed with ten convolutional layers. The features of both models are extracted and optimized using an improved version of Grey Wolf with updated criteria of the Jaya algorithm. The improved version speeds up the learning process and improves the accuracy. Finally, the selected features are fused using a novel parallel pooling approach that is classified using machine learning and neural networks. Two datasets, BraTS2020 and BraTS2021, have been employed for the experimental tasks and obtained an improved average accuracy of 98% and a maximum single-classifier accuracy of 99%. Comparison is also conducted with several classifiers, techniques, and neural nets; the proposed method achieved improved performance.