An intelligent brain tumor detection model using lightweight hybrid twin attentive pyramid convolutional network

Brain tumors (BTs) pose a serious threat to human health, and the optimized treatment and results depend on early and accurate detection. Although MRIs and other medical imaging technologies provide insightful information, it is still difficult to detect brain cancers in these images. In this research, hybrid lightweight deep learning framework (HybLwDL) for advanced BT diagnosis utilizing MRI images is proposed. To improve image quality, the HybLwDL framework pre-processes the MRI images using a Gaussian Bilateral Network Filter (GANF). Then, effective feature extraction and classification is done using lightweight hybrid twin-attentive pyramid convolutional network (LHTA-PCNet) model. In particular, LHTA-PCNet modified the twin-level attention (TwinL-A) module and the hybrid pyramid convolution (HPC) block, and introduced ResNet as the backbone network. The TwinL-A module is employed to successively extract various local and global representations from the channel and spatial domains, while the HPC block is intended to improve the acquisition of information at various scales. Stellar Oscillation Optimizer (SOO) is used to expand the classification accuracy by tuning the hyper-parameters of LHTA-PCNet. Furthermore, Grad-CAM is incorporated into the HybLwDL method to visualize and highlight the significant regions in MRI images that contribute to BT detection. The HybLwDL model is implemented using python platform. Deep learning-based approaches have drawn a greater interest because of their excellent computational efficiency and precision in tumor classification. With a high accuracy of 99.5% for the BT Detection 2020 dataset, the HybLwDL model can reliably classify BTs.