Shape Models for Segmentation MIT Artificial Intelligence Lab and Brigham and Women's Hospital Overview Medical Image Segmentation is the process of labeling voxels of a scan according to the anatomical structure or tissue type. Using image intensity alone can cause segmentation errors (left). The addition of prior shape knowledge fills in the gaps of missing or noisy image information (right). Construction of Shape Models A statistical shape model is constructed from a set of rigidly aligned training examples. Seven vertebrae (T3-T9) are shown on the left, and a Gaussian model is fit to the shapes. The computed "mean" vertebra is shown on the right. Modes of Shape Variance The primary mode of shape variance is shown on the right, warping about the mean shape in green. Notice the primary mode captures changes in shape as the vertebrae go up the spine from T9 to T3. The body of the vertebra gets smaller as the process extends. Segmentation The segmentation process consists of a surface evolving towards the maximum likelihood shape given the prior model and the image information. Three orthogonal views are shown. The yellow highlight shows the estimated model pose. The red contours are slices through the evolving surface. The 3D model is also shown as the segmentation progresses.