BrainNeuroNet: advancing brain tumor detection with hierarchical transformers and multiscale attention

Brain tumor (BT) refers to abnormal proliferation of brain tissues and accurate detection is complex for radiologists. The accurate and early detection impact is based on the shapes, intensities, and location of the tumors. Therefore, a novel BrainNeuroNet approach is proposed for identifying BT through the collaborative teacher-student network. This proposed collaborative model is designed according to the combination of Hierarchical DConv (HD) transformer and MultiScale Attention (MSA) Network. The brain images are collected from Brain Tumor MRI and Br35H datasets and are preprocessed by using different processes such as image resizing, normalization, and enhancement for image quality enhancement. The HD transformer model is taken as a teacher network that extracts global features in the preprocessed brain images. In this HD transformer model, the complex patterns are captured by asymmetric convolution, and the computational complexity is decreased by shifting window technique. The MSA network is the student network that extracts local features in the preprocessed images and this model has a dilated convolutional block for decreasing parameters and retaining nodal information. The collaborative learning model is applied to concatenate both local and global extracted from the preprocessed images, and loss function trained the concatenated features for the prediction of BTs. The process of testing the model’s effectiveness is employed to demonstrate the efficiency of the BrainNeuroNet model by using significant tumor detection-related performance measures. The comparative study revealed that the BrainNeuroNet model attained a better detection accuracy of 98.63% rather to existing approaches.