Unsupervised Grow-Cut: Cellular Automata-Based Medical Image Segmentation

This paper presents a new cellular automata-based unsupervised image segmentation technique that is motivated by the interactive grow-cut algorithm. In contrast to the traditional method which requires user-interaction to identify classes, the unsupervised grow-cut algorithm (UGC) starts with a random number of seed points and automatically converges to a natural segmentation. This is useful when deriving classes from large image datasets for applications such as region-based image retrieval. The algorithm has been tested on a subset of thirty medical images derived from the Image CLEF med database and 300 natural images from the Berkeley dataset. The unsupervised grow-cut algorithm has been compared against the Mean Shift method and Normalized Cut method. The segmentation outcome of the UGC algorithm is comparable with the other two methods. The number of classes derived by the UGC is independent of the number of initial seed points. Incorporating cellular automata makes the computational complexity of the algorithm independent of the dimension of the image and feature space.