Modular Neural Networks for detection and classification of brain cancer images

Among the various oncological pathologies, brain cancer continues to be one of the most wide-spread, as well as lethal, diseases. Within this paper we extended a previous work leveraging the early exits strategies to reduce the computa- tional time needed to detect and classify brain tumor MRI images namely say glioma, meningioma, and pituitary. The dataset we used consisted of 3264 2-D MRI images and 4 classes. Due to the small number of images, various data aug- mentation techniques were used to increase the size of the dataset. Our proposed methodology consists not only of data augmentation, but also of various techniques of image denoising, skull strip-ping, cropping and bias correction. In our working proposal, we implemented a custom CNN and we explored both a static and dynamic solution of Early Exit procedure to achieve the smallest computational effort. The purpose of this document is to distinguish between normal and abnormal pixels and classify them more rapidly but with a comparable level of accuracy.