Empirical Analysis of the pre-trained models for Brain Tumor Diagnosis

Brain tumors are a significant health problem worldwide requiring accurate and timely diagnosis for efficient treatment planning. Medical imaging, especially magnetic resonance imaging (MRI) and computed tomography (CT), is important towards recognizing and classifying brain tumors. Traditional approaches to classifying brain tumors typically require radiologists to manually review images, a process that is protracted, laborious and prone to mistakes. Continuing advances in medical imaging technology are optimizing the early detection and diagnosis of various diseases, including brain tumors. Convolutional neural networks (CNNs) are gaining importance in this tool set for automatic image analysis and classification. This research paper delves into the application of CNNs for brain tumor categorization, exploring their effectiveness in improving accuracy and efficiency in comparison to traditional methods. The models undergo training with a dataset consisting of 7023 MRI images of the human brain, categorized into four classes: meningioma, pituitary, glioma and no tumor. The networks implemented in our research are ResNet-50, EfficientNetB0, VGG16, AlexNet, and DenseNet. In particular, the pre-trained ResNet-50 model from CNN exhibited excellent results, with an accuracy during training reaching while the validation accuracy stood at $\mathbf{9 8 . 4 4 \%}$.