Automatic Segmentation of Extraocular Muscles Using Superpixel and Normalized Cuts

This paper proposes a novel automatic method to segment extraocular muscles and orbital structures. Instead of conventional segmentation at the pixel level, superpixels at the structure level were used as the basic image processing unit. A region adjacency graph was built based on the neighborhood relationship among superpixels. Using Normalized Cuts on the region adjacency graph, we refined the segmentation by using a variety of features derived from the classical shape cues, including contours and continuity. To demonstrate the efficiency of the method, segmentation of Magnetic Resonance images of five healthy subjects was performed and analyzed. Three region-based image segmentation evaluation metrics were applied to quantify the automatic segmentation accuracy against manual segmentation. Our novel method could produce accurate and reproducible eye muscle segmentation.