A Comparative Study of DenseNet121, VGG16, and Custom CNNs for Brain Tumor Classification using MRI Images

The categorization of brain tumors from magnetic resonance imaging (MRI) scans is a condemnatory step in the identification and management of brain diseases. With develpment in machine learning and specially deep learning, convolutional neural networks (CNNs) have shown impactful execution in the automatic identification of medical images. This paper compares three CNN-based models for classifying brain tumors: DenseNet121, VGG16, and a custom designed CNN. Both DenseNet121 and VGG16 utilize the power of transfer learning, while custom CNN is built from scratch for this specific task. The dataset used comprises MRI images categorized into glioma, meningioma, pituitary tumors, and no tumor. The models are evaluated on the basis of accuracy, loss, and generalization. The results reveal that DenseNet121 achieved an outstanding accuracy of 99%, VGG16 reached 98%, while the custom CNN yielded 96%. However, the custom CNN showed significant signs of overfitting during training, which will be thoroughly examined in this study. The paper further explores the benefits of transfer learning and discusses the tradeoffs in using pretrained models versus custom-built networks. This comparative analysis provides insights to the best practices to utilize deep learning models in medical image classification.