Enhancing Brain Tumor Classification Using AdaDensenet

Abstract

Early identification is critical since brain tumors harm over 300,000 people annually and cause over 250,000 deaths. Although models such as VGG16, ResNet50, InceptionNetV3, and DenseNet201 exhibit potential, they struggle with dataset bias and flexibility. With global average pooling, transfer learning, sophisticated data augmentation, and attention techniques to improve flexibility for practical application, our suggested model, AdaDensenet, expands upon DenseNet201. AdaDensenet achieved 98.86% accuracy in a complete examination of 7,023 MRI scans from Figshare, SARTAJ, and Br35H (identified as pituitary, glioma, meningioma, or tumor). In addition to accuracy, the model achieves classwise AUC-ROC values of 1.00 and macro and weighted precision, recall, and F1-score of 0.99, demonstrating outstanding discrimination and balanced performance across tumor types. These results position AdaDensenet as a strong candidate for real-time clinical decision support systems.