Shape guided 3D active contour model for automatic and accurate MRI prostate segmentation

We present a new 3D region-based active contour model that is driven by an automatic and accurate object initialization scheme guided by the prior shape information. Because Laplace-Beltrami spectrum is isometric invariant, it is one of the best ways to represent the clinical shape of prostate. The general geometrical information (i.e. the elliptical or walnut shape for the prostate) of volumetric prostate datasets is discovered and extracted by taking the eigenvalues of its Laplace-Beltrami operator. Specifically, we use the heat kernel to link the spectrum with the fingerprint of the volumetric data. Through the selection of eigenvector, the entire volumetric dataset is automatically divided into two classes: ROI and non-ROI. The boundary of these two classes can be used as the initial contour of our interest objects. The 3D region-based contour is further refined by minimizing the mean separation energy in a local region. By the comparison to edge-based methods, region-based contour models have better robustness against the initial curve placement and less insensitivity to image noise. Our automatically shape guided 3D active contour model results in more prominent boundaries compared to the classical gradient function based model (`snake´). We also evaluate segmentation results of our new contour method against the manually segmentation images to calculate the values of average overlap, sensitivity and specificity. These numbers demonstrate that our automatic object contour initialization scheme guided by the shape information is an ideal segmentation tool for the large number of prostate MRI datasets without the clear boundaries.