Using a Saliency-Driven Convolutional Neural Network Framework for Brain Tumor Detection

The process of diagnosing brain tumors from magnetic resonance imaging (MRI) is often time-consuming. Thus, a rapid analyses through an automated system could help improve the treatment possibilities and optimize hospital resources. This paper proposes a method for the classification of brain tumors by means of pre-selecting the tumor region. We estimate the region of interest using attention algorithms and input it to a neural network that classifies the regions as tumoral and non-tumoral. Pre-selecting the region of interest, instead of using the entire image, produced a final classification accuracy of 91.68%, 92.58%, 92.69%, and 93.4% with the models Resnet18, Resnet34, VGG16, and Alexnet, respectively. Once the dimensional space of the input image is reduced, the neural networks is able to capture additional details of the tumor regions during the training stage. This study demonstrates the importance of saliency maps for identifying tumor regions in magnetic resonance images.