Arithmetic optimization-based K means algorithm for segmentation of ischemic stroke lesion

Stroke is a life-threatening medical condition which blocks arteries supplying blood to brain. Stroke lesion delineation from brain MRI by clinical experts consumes longer time and are subjected to variation among multiple clinical observers. There is a huge demand for developing automated system which extracts out stroke lesion from brain MRI slices. Performance lag in clustering-based segmentation algorithm adopted for ischemic stroke lesion segmentation is due to their inability in determining the number of clusters before segmentation and random cluster center initialization. These problems may result in clustering steps get trapped in local optimal solution and not progresses towards reaching the global optimal solution. The proposed AOK-Means addresses the above issues. The number of clusters for each MRI slice is found by determining the break in point of the aggregated arithmetic mean of sub-images of whole brain MRI slice, where number of sub-images is equal to the number of clusters. Among randomly generated population of initial cluster centroids based on number of gray levels and probability distribution of number pixels at each gray level, the best cluster centroid is reached by maximizing Tsallis and Otsu entropy functions through arithmetic optimization algorithm. These optimal cluster centroids are taken as initial cluster centers of K means clustering algorithm for segmentation of Diffusion Weighted MRI images. Segmentation results are most promising with average accuracy 0.989762 +/- 0.010668, sensitivity 0.695454 +/- 0.164761, Dice 0.764438 +/- 0.133147 and precision 0.919855 +/- 0.113664 in Otsu entropy-based segmentation and average accuracy 0.986679 +/- 0.013351, sensitivity 0.635716 +/- 0.158888, Dice 0.708926 +/- 0.144569 and precision 0.904033 +/- 0.126629 in Tsallis entropy-based segmentation.