Squeeze and Excitation Attention Meets Modified EfficientNetB0 Architecture: Multi-Class Brain Tumor Classification Using Explainable Artifical Intelligence

Brain tumors, spanning a spectrum from benign to malignant, pose a formidable medical challenge due to their potential to disrupt critical brain functions. Traditional diagnostic methods, relying on human interpretation of MRI images, frequently suffer from accuracy and efficiency limitations. Despite promising research, it encounters challenges in feature extraction and model interpretability. In this study, an innovative approach to brain tumor classification was introduced. A modified variant of the EfficientNetB0 architecture was leveraged, and reinforced by a Squeeze and Excitation Attention (SEA) block. The upper layers of the original EfficientNetB0 architecture were fine-tuned and customized to suit the specific investigation. Subsequently, the tailored EfficientNetB0 was employed to extract features and seamlessly integrate the SEA block. This integration amplified the emphasis on pertinent features while dampening irrelevant ones. The approach underwent a comprehensive empirical assessment on a combined dataset, yielding a remarkable accuracy rate of 99.70%, surpassing other state-of-the-art methods. Moreover, Grad-CAM visualization was employed to precisely pinpoint areas within the images that exerted a substantial influence on the model's predictions.