Enhanced MobileNetV2 with an Attention Mechanism for Real-Time MRI-Based Brain Tumor Classification: A Deep Learning Approach

Abstract

This work presents the main steps toward develop- ing a real-time MRI-based classification system, building upon our previous work in deep learning-based classification of brain tumors. We enhance the model scalability and achieve higher computation speed by applying an efficient and lightweight architecture - MobileNetV2. Importantly, it features the inclusion of an attention mechanism, explicitly sharpening the model’s at- tention to relevant image elements, and mixed precision training that maximizes processing speed along with memory use. This approach boosts the robustness and accuracy of tumor detection while concurrently reducing training times. Data augmentation strategies have been refined to make models more generalizable at lower computational cost. Moreover, the learning rate dynamic adjustments are carefully tuned to make the convergence stable and improve the effectiveness of model training more effectively. Our results show significant improvements compared to earlier versions by achieving better recall rates and precision, which remain important metrics in clinical applications where prompt- ness and accuracy of diagnosis is tantamount. We were able to achieve higher stability while maintaining an accuracy of 98% and similar high baselines. The work performed here can be used as a strong aid to improve diagnosis in healthcare and is a giant leap in the use of deep learning technologies for medical imaging. Index Terms—Brain Tumor, CNN Architecture, Deep Learn-
ing, Classification, ResNet50, VGG16, Transfer learning.