Classifying malignant brain tumours from 1H-MRS data using Breadth Ensemble Learning

In neuro oncology, the accurate diagnostic identification and characterization of tumours is paramount for determining their prognosis and the adequate course of treatment. This is usually a difficult problem per se, due to the localization of the tumour in an extremely sensitive and difficult to reach organ such as the brain. The clinical analysis of brain tumours often requires the use of non-invasive measurement methods, the most common of which resort to imaging techniques. The discrimination between high-grade malignant tumours of different origin but similar characteristics, such as glioblastomas and metastases, is a particularly difficult problem in this context. This is because imaging techniques are often not sensitive enough and their spectroscopic signal is overall too similar. In spite of this, machine learning techniques, coupled with robust feature selection procedures, have recently made substantial inroads into the problem. In this study, magnetic resonance spectroscopy data from an international, multi-centre database were used to discriminate between these two types of malignant brain tumours using ensemble learning techniques, with a focus on the definition of a feature selection method specifically designed for ensembles. This method, Breadth Ensemble Learning, takes advantage of the fact that many of the frequencies of the available spectra convey no relevant information for the discrimination of the tumours. The potential of the proposed method is supported by some of the best results reported to date for this problem.