Adaptação dos algoritmos ROCKET e MiniRocket para classificação de tumores cerebrais

Abstract

One of the main methods of brain tumor diagnosis is the magnetic resonance imaging of a patient’s head. Identifying a tumor’s type using those images is important for its adequate treatment, and one of the most relevant means of automating this process is the use of Machine Learning, in particular deep learning and convolutional neural networks. However, those models may be expensive in terms of training time and data. Therefore, this work proposed the adaptation of two time series classification algorithms to the image classification task: ROCKET and MiniRocket, which do not employ deep learning. We showed that, in spite of simplistic experimental procedures presented throughout most of the literature, both proposed algorithms exhibit state of the art performance and, furthermore, MiniRocket requires a fraction of convolutional networks’ training time to achieve great results.