Deep CNN-Based Multi-Grade Brain Tumor Classification with Enhanced Data Augmentation

Abstract

Innovations in the field of medical imaging techniques, especially in magnetic resonance imaging (MRI), have significantly enhanced diagnostic capabilities. However, the accurate classification of brain tumors from MRI scans remains a difficult task due to the subtle variations between different tumor types and the presence of non-tumorous regions. The primary challenges in automated MRI classification include the high variability in tumor appearance, similarities between benign and malignant tumor features, and the inherent imbalance in medical datasets. The proposed model architecture includes multiple convolutional layers with normalizing batches and removing outliers to enhance generalization and control for overfitting. The dataset was artificially expanded using data augmentation techniques like flipping, zooming, and rotating from 5,712 original images to 142,800 images, allowing the model to learn from a more diverse set of examples. The model demonstrated promising results, obtaining a training precision of 99% and a validation accuracy of 91.5% after 50 epochs, suggesting effective learning and generalization capabilities.