Abstract:
The brain tumors are the irregular growths of cells in the brain that may impair the brain neurological functions and cause extreme consequences when not recognized at early stages. Magnetic resonance imaging (MRI) is becoming a common procedure in the evaluation of brain tumors due to its ability to provide detailed anatomy images of the intracranial structures. Due to the blistering development of deep learning, automated MRI analysis has become one of the promising tools in clinical decision-making. We train and test two image-classification models, supervised to detect brain-tumors, in this work; a custom convolutional neural network (CNN) and a transfer-learning model which is built upon the existing architecture VGG16. Both the models are trained and tested on the very same MRI dataset with the same conditions of the experiment. In our experiments, the custom CNN obtained a classification of 99% and VGG16 model, 95%. Such a distinction implies that a brain MRI-specific network can be better than an off- the-shelf pre-trained model, such as VGG16. Practically, these small and well-designed CNNs can be used to assist radiologists and offer fast and consistent predictions of tumors based on MRI, which can contribute to the more accurate diagnosis and eventually positively influence patient treatment.
