Brain Tumor Detection Using Convolutional Neural Networks: A Comparative Study

Abstract

Using Magnetic Resonance Imaging (MRI) images to detect brain tumors by medical practitioners is mundane and prone to errors. Misdiagnosis of brain tumors can be life-threatening, so to lessen misdiagnosis, computational techniques can be used in concert with medical professionals. Deep learning approaches have been gaining popularity in modeling and developing systems for medical image processing that can detect abnormalities quickly. The methods proposed herein are based on Convolutional Neural Networks (CNN) trained on the 'BR35H::Brain Tumor Detection 2020' dataset. A custom CNN architecture was designed, followed by the utilization of transfer learning with four pre-trained models: InceptionV3, ResNet101, VGG19, and DenseNet169 and a comparative analysis of these architectures has been presented in this paper. The experimental results show that the DenseNet169 model outperformed other models with a training accuracy of 99.83 %, test accuracy of 99.66%, precision of 99.67%, and recall of 99.67%. Additionally, ResNet101 has a 95.92% test accuracy, VGG19 has a test accuracy of 97.83%, the custom architecture has a test accuracy of 98.16%, and InceptionV3 has the lowest test accuracy of 91.66%. It has been concluded that DenseNet169 provides better results for the classification of brain tumors than other models.