Superpixel-based ROI segmentation Clause Samples

Superpixel-based ROI segmentation. To create anatomically meaningful structures within the tissue such as ducts or lobules, the superpixels classified as epithelium (generated in Task 5.2) have to be clustered. To achieve this we use an algorithm which has 3 major steps. (1) Using Relative Neighborhood Graph (RNG) to identify coarse clusters of neighboring superpixels. (2) Using Delaunay Triangulation (DT) to find spatially homogeneous sub-clusters within clusters in the first step. (3) Finding the outer contour of each sub- cluster representing the boundary of the identified ROI. STEP 1: Our algorithm initially builds an RNG on the entire set of superpixels labeled as epithelium. By imposing a threshold on the maximum edge connecting two neighboring superpixels we cluster the image into coarse clusters of neighboring superpixels. These coarse clusters can contain multiple ductal or lobular structures lying close to each other. STEP 2: We employ a graph-pruning technique to find spatially homogeneous sub-clusters within the coarse clusters identified in the previous step. For this reason we initially build a DT on the superpixels composing a coarse cluster (See Fig. 2b). Our clustering algorithm identifies and removes the inter-cluster simplexes within the DT. Inter-cluster simplexes are the ones which connect two or multiple sub-clusters (e.g. ducts) to each other. By pruning these simplexes we can identify separate clusters of points within the coarse cluster. To identify these, for each simplex in the DT, 3 measures are computed. (1) Perimeter of the simplex. (2) Elongation of the simplex. (3) The standard deviation of the elongation measures over the set of , where denotes the set of simplexes neighboring . Finally, the entire simplex analysis is captured in a criterion function , which is equal to the multiplication of the three measures. For each simplex s in DT, if is bigger than a predetermined threshold the simplex is removed from the graph. As a result of this step, only positive nodes and edges of the graph remain. By using depth-first search we then infer the number of isolated sub-clusters and correspondingly the list of superpixels in each of them. Fig. 2c shows the result of this clustering. STEP 3: In the last step we find the outer contour of each sub-cluster. This is done by computing the concave hall of the subgraph which is built on the set of superpixels inside each sub-cluster. An accurate delineation of the ROI is obtained by taking the union of the binary masks of th...