Enhanced Brain Tumor Classification using Squeeze and Excitation Attention-Guided Modified EfficientNetB2 Architecture

Brain tumors, spanning from benign to malignant forms, present a substantial medical challenge due to their capacity to disrupt crucial brain functions. Conventional diagnostic techniques, reliant on the manual interpretation of MRI images, often exhibit limited accuracy and efficiency. Recent research endeavors show promise but encounter difficulties in extracting essential features and providing model interpretability. In this study, an innovative approach to classify brain tumors is introduced, leveraging a modified EfficientNetB2 architecture guided by a Squeeze and Excitation Attention (SEA) block. The original EfficientNetB2 architecture undergoes fine-tuning, with its top layers being tailored for this specific investigation. Subsequently, after feature extraction from the modified EfficientNetB2, the SEA block is integrated to enhance the emphasis on relevant features while suppressing irrelevant ones. An empirical assessment is conducted on a merged dataset, resulting in an impressive accuracy rate of 97.98%, surpassing other state-of-the-art methods. Furthermore, Grad-CAM visualization is employed to reveal the specific image regions exerting influence on the model’s predictions.