SAlexNet: Superimposed AlexNet using Residual Attention Mechanism for Accurate and Efficient Automatic Primary Brain Tumor Detection and Classification

Accurate classification of brain tumors is crucial for informing clinical diagnoses and guiding patient treatment plans. It is one of the most aggressive tumors, leading to a short life expectancy. However, the classification of brain tumors is a challenging task due to the heterogeneity, complexity, and variability of brain tumors. In this work, we propose Superimposed AlexNet (SAlexNet-1 and its extension SAlexNet-2) to detect the malignancy of primary brain tumors (Glioma, Meningioma, and Pituitary) by incorporating three enhancements: (1) fusing Hybrid Attention Mechanism (HAM), (2) dense feature extraction by replacing initial convolution 11 × 11 layer with multiple convolution 3 × 3 layers for extra non-linearity alleviating parameter burden, and (3) pretraining the encoder path on a correlated dataset via semi-transfer learning (STL) enhancing model performance. HAM provides more comprehensive and accurate feature representations. In this study, we evaluated the performance of our proposed SAlexNet models on two publicly available extensive datasets for multi-class and binary classification tasks. Our results show that SAlexNet-1 achieved an accuracy of (98.78 ± 0.80 %) and (98.07± 0.02 %) on the multi-class and binary classification datasets, respectively. In comparison, SAlexNet-2 achieved outstanding accuracy of (99.69 ± 0.22 %) and (99.17 ± 0.00 %) on the multi-class and binary classification MRI datasets, respectively. The STL-based SAlexNet-2 surpassed previous literature with complex models and techniques, achieving an accuracy of (99.20 ± 0.01 %). Furthermore, we provided a comprehensive analysis of current state-of-the-art tumor classification methods on the same dataset, demonstrating the effectiveness of our approach. Enhanced tumor classification accuracy enables better diagnosis, treatment planning, and patient outcomes.