Markov Clustering Based Feature Selection for Brain Tumor Segmentation from Multi-Spectral MRI Records

The automatic segmentation of medical images is very important in clinical practice, because it can reduce the costs of diagnosis and therapy planning processes. In order to have a really useful solution, we need to provide automatic algorithms that are both highly accurate and highly efficient. If we need to choose one of them, accuracy usually has the priority over efficiency, but the best way of treating the problem is providing a good trade-off between these two requirements. This paper proposes a feature selection algorithm that attempts to reduce the feature set used by a decision making algorithm such a way that the accuracy does not suffer any significant damage. The feature selection is performed by a Markov clustering algorithm, which receives as input information on all feature pairs regarding their frequency of use and the correctness of the decisions made while using them together in decision making. The proposed method is validated in a brain tumor segmentation problem based on multi-spectral MRI data, in a framework that uses ensemble learning built upon binary decision trees. The proposed method reduces the initial full set of 104 features to 34 without losing anything in terms of segmentation accuracy, thus contributing to the efficiency of the segmentation algorithm.