Automated Localization of Brain Tumors in MRI Using Potential-K-Means Clustering Algorithm

The manual localization and precise segmentation of brain tumours from magnetic resonance images (MRI) is time-consuming and error-prone. In T2 and FLAIR MRI, tumours appear as bright areas of higher signal intensity than their surroundings. In this paper we view the intensity of a pixel as equal to its "workload" and employ an unsupervised learning algorithm called potential-K-means that generates a balanced distribution of the pixels into clusters of approximately equal total intensity. The algorithm is based on an analogy with the gravitational force exerted by masses. This balancing requirement introduces a search bias that tends to generate either small clusters of higher intensity pixels, which overlap with the tumour area, or large clusters of lower intensity pixels. We evaluate the proposed algorithm on the publicly available brain tumour image segmentation (BRATS) MRI benchmark by comparing the centre of the cluster that overlaps with the tumour, with the centre of the tumour in the corresponding ground truth segmentation. We compare the proposed algorithm with the well-known K-means and with the Force clustering algorithm by Calamari et al. (2009), which follows a different Physics analogy, but it is also based on a balancing criteria. Experimental results show that K-means is not suitable for tumour localization and that potential-K-means and Kalantari's approach are comparable. However, the performance of Kalantari's approach is highly dependent on a parameter whose value needs to be set a priori, but without an informed way of doing so, which makes the present proposed method more practical.