Impact of shared dense connectivity and channel width on convolutional block attention module for regional MRI-based brain tumor classification

MR imaging is a widely used imaging technique for diagnosing brain-related issues. Different tumor types in MR images often share similar visual characteristics, leading to misclassification. This research focused on the impact of multiple shared dense channel attention (MSDCAT) with varying dimensionality on the CBAM(convolutional block attention module) architecture, which is further used for the classification of MR Images for detecting brain tumors. The major objectives addressed in this study are to enhance feature extraction capabilities through multiple shared dense layers, the effect of channel reduction ratio, and to enable efficient information flow across multiple layers. The proposed model leverages the dense connectivity and feature reuse properties of Dense block to extract discriminative features from multi-modal MRI images. The model includes 4 shared dense layers on the channel attention module in conjunction with the spatial attention module in a sequential manner. The structured dense block with a transition layer is also included in the initial pathways. The model is evaluated on varying scenarios of shared dense layers and multiple channel reduction ratios on different standardized databases. Testing of the proposed system on the Figshare database, together with the Kaggle database, demonstrated promising outcomes and produced strong accuracy rates with specific sensitivity and specificity measurements. The model reached 99.70% accuracy in the Figshare database while achieving 99.90% accuracy in the Kaggle dataset. Our findings demonstrated the feature extraction capability of the proposed approach in accurately classifying brain tumors on both datasets and highlighting the potential of Multi-Layered shared dense layers in accurately extracting the channel attention features from the MRI(Magnetic Resonance Imaging) images.