enhanced residual attention cnn with squeeze-and-excitation blocks for brain tumor mri classification

Abstract— precise categorization of brain tumor varieties using magnetic resonance imaging (mri) scans is crucial for prompt diagnosis. it facilitates informed treatment decisions and enhances patient outcomes in neuro-oncology. the challenges include diverse tumor shapes, overlapping intensity patterns, and subtle distinctions among classes (e.g., meningioma, glioma, pituitary tumors). these factors complicate automated assessments. this study presents an enhanced residual attention convolutional neural network (ra-cnn) featuring squeeze-and-excitation (se) blocks. the se blocks dynamically modify channel-wise characteristics. this boosts the detection of spatial and global dependencies for intricate tumor attributes. residual links mitigate the issue of vanishing gradients, allowing for deeper and more computationally efficient architectures. tested on the figshare brain tumor mri dataset (3,064 images), the ra-cnn achieves a remarkable 94% classification accuracy. metrics for precision, recall, and f1-scores surpass 93% across all categories. it exceeds the performance of conventional cnns, nnu-net, and transformer-based frameworks. ablation studies confirm the effectiveness of se blocks and residual links. this methodology presents significant promise for clinical applications, assisting radiologists in accurate tumor classification and enhancing diagnostic processes.