Real-Time Multi-Class Brain Tumor Classification through Deep CNN Models of MRI Scans

Abstract

In this work, we propose the real-time system for multi-class brain tumor classificationbydeep CNNs on MR images. Methodology: The methodology starts with capturingimagedata from three public data sets (SARTAJ, Figshare, Br35H), which are integratedtoforma consistent set of 5712 images of MRI images grouped into 4 classes: glioma, meningioma, pituitary tumor, and no tumor. The data is split into training (4,570), validation(571), andtesting (571) sets. Data preprocessing methods such as contrast adjustment, parametrictransformation, and augmentation (e.g., rotation, flipping, and scaling) enhance imagequality and improve model generalization. Four deep learning models (ResNet50, InceptionV3, EfficientNetB2, and a custom CNN) are trained and tested based onaccuracy, precision, recall, and loss. ResNet50 had the highest accuracy (98.80%), followedbyEfficientNetB2 (92.18%), custom CNN (91.40%) , and InceptionV3 (83.53%). The resultsshow that the model architecture is important to the classification performance, andResNet50 performs the best because of its residual learning. To ensure practical utility, thetop-performing ResNet50 model was implemented using the Streamlit frameworkandhosted on HuggingFace Spaces, allowing users to predict MRI images in real-time througha user-friendly web interface. The proposed method provides valuable clinical decisionsupport and helps enhance diagnostic confidence and early treatment of neuro-oncology.