Detection of Brain Tumors from MRI Images using Convolutional Neural Networks

Abstract

There are more than 150 different brain tumors but can be grouped into two main types: primary and metastatic. Presently, magnetic resonance imaging (MRI) is the imaging of choice for the assessment of brain tumors. The objective of the study is to assess the diagnostic performance of convolutional neural networks in the evaluation of brain tumors from cranial MRI images. Different CNN models were applied to an anonymized publicly available MRI brain tumor dataset to assess the presence of brain tumors. Several pre-processing steps (image normalization, shuffling of images, image cropping, and geometric augmentation techniques) were applied to the MRI images. After a series of preliminary verification of various configurations, a base CNN model was developed with succeeding experiments being conducted to search for the optimum composition of neural network parameters. The best base CNN model configurations had very good performance results while the pre-trained architecture (VGG16 and ResN et50) generated excellent performance metrics. For complicated medical images such as cranial MRI, a deeper architecture is preferred over a shallower base model as the pre-trained models obtained much higher performance metrics. Nonetheless, the base model performed well particularly its sensitivity despite having a simpler though shallower architecture which indicates lower false negative results leading to fewer missed cases of patients with brain tumors. This indicates the base model's capability as a valid and reliable decision support tool. Hence, any of these CNN models can be incorporated routinely by physicians in their clinical practice to further augment their decision-making capability.