Automated Brain Tumor Classification in Tomographic Imaging Using a Hybrid CBAM-DenseNet121 Architecture

The exponential growth of medical imaging data, particularly in tomographic scans, necessitates robust artificial intelligence solutions to reduce the manual radiological workload, mitigate human fatigue, and minimize inter-observer variability. Aligning with recent advancements in deep learning and radiomics, this study proposes an advanced automated framework integrating a Convolutional Block Attention Module (CBAM) with a DenseNet121 backbone to precisely classify critical brain anomalies—specifically Gliomas, Meningiomas, and general brain tumors—from cross-sectional scans. The DenseNet121 base network was specifically selected and utilized for its optimal feature reuse capabilities, alleviation of the vanishing-gradient problem, and overall computational efficiency during training. To explicitly isolate critical pathological regions and effectively suppress irrelevant background noise, a dual-attention CBAM, operating sequentially across both spatial and channel dimensions, was dynamically embedded into the architecture. The proposed methodology was rigorously trained and validated on a comprehensive multi-class dataset of brain scans using standardized preprocessing techniques. The hybrid model demonstrated exceptional diagnostic performance, achieving a high overall validation accuracy of 96.67%. Detailed confusion matrix analysis further revealed superior class-specific precision, notably achieving an outstanding 99.51% accuracy in detecting Gliomas, followed closely by 95.34% for Meningiomas and 95.08% for other unclassified tumors. Exhibiting stable convergence curves with minimal validation loss, this AI-driven architecture provides a highly accurate, automated decision-support tool for medical practitioners. By significantly improving diagnostic confidence, acting as a reliable second diagnostic reader, and streamlining complex radiological workflows, the proposed CBAM-DenseNet121 model offers a highly viable and scalable framework for real-time clinical integration in tomographic analysis.