Superpixel Segmentation of Brain MR Images Based on Probabilistic Weighted Geodesic Distance

While superpixel segmentation is a significant over-segmentation technique, it is extremely difficult to perform it on medical image for blurred boundaries and overlapping grayscales of different tissues. This paper bases on the characteristic of brain MR images, aims to deal with the problem of brain MR images superpixels generation. We make full use of the prior of brain MR images, combine the structure of the brain MR images and define a probability every pixel belongs to each class. An efficient method to compute boundary gradient based on classification probability is proposed. Based on the above, we propose a probability density weighted geodesic distance for superpixel segmentation. The fuzzy C-means method is applied as subsequent processing to obtain the classification of brain image. By comparing segmentation performance with the state-of-the-art methods in quantitative analysis, our algorithm generates more accurate segmentation boundaries.