Scopus Indexed Publications

Paper Details


Title
Advancing Brain Tumor MRI Classification with Interpretable Attention-Enhanced Deep Learning

Author
, Amit Kirti Mondal,

Email

Abstract

Neuro-oncology relies on magnetic resonance imaging (MRI) brain tumor classification since automated methods can greatly improve diagnosis. Despite advances in deep learning, long-range spatial interactions, model interpretability, and computational efficiency for realtime deployment remain obstacles. This paper introduces an attentionenhanced DenseNet201 network with Multi-Head Self-Attention (MHSA) to solve these constraints. DenseNet201 emphasizes dense feature reuse, and MHSA focuses on concentrated areas surrounding the image's spatial frame and improves discrimination. MRI images were prepared by rescaling, intensity normalization, and contrast-limited adaptive histogram equalization. Data augmentation via rotations, translations, horizontal and vertical flips, and zoom operations strengthened model generalization. Gradient-Weighted Class Activation Mapping (Grad-CAM), an explainable artificial intelligence method, produced class-specific saliency maps that visualized tumoral locations and gave clinical interpretability. Test accuracy of 0.9966, macro-average F1-score of 0.9966, Cohen κ value of 1.0000, and consistent five-fold cross-validation accuracy of 0.9734 demonstrate superior performance. Individual MRI scan inference takes less than 24 milliseconds per image, making it suitable for real-time use. Comparative assessments of contemporary convolutional neural networks and transformer-based architectures indicate that the DenseNet201 + MHSA framework effectively classifies brain tumors with commendable accuracy and interpretability, establishing it as a viable therapeutic tool.


Keywords

Journal or Conference Name
IEEE International WIE Conference on Electrical and Computer Engineering/WIECON-ECE

Publication Year
2025

Indexing
scopus