Attention transformer mechanism and fusion-based deep learning architecture for MRI brain tumor classification system

•A hybrid deep learning approach is proposed to classifying different tumors in MR images using the Transformer module (TM) and the self-attention unit (SAU).•Using cross-fusion mechanism, our novel model generates fine-grained features by fusing local and global features together.•The lightweight and optimized version of deep transfer learning (iResNet) helps to overcome the problem of overfitting.•Multiple tumors in MR images were classified with 99.30 % accuracy.