BrainResMambaVision-18 An Transformative approach for Enhanced Multi-class Classification

Abstract

The classification of brain tumor classes is a crucial task, as handcrafted feature extraction yields limited accuracy. The proposed BrainResMambaVision-18 classifies the tumor into multi-classes of Meningioma, Glioma, and Pituitary tumors to overcome the problem of handcrafted features. BrainResMambaVision-18 leverages long-sequence global dependencies similar to Vision Transformer. The MambaVision uses the weights and biases of finetuned ResNet-18 to capture local features efficiently. The proposed framework employs the features by providing a multi-stage algorithm that uses a mixer block that includes State Space Models (SSM) to improve global context modeling. The classification accuracy is enhanced by 3% from the existing state-of-the-art models, and the accuracy obtained is 93.00%, 96.42%, and 98.81% for binary-class classification, 3-class classification, and 4-class classification, respectively.