Integrating model based and data driven approaches for the automatic segmentation of cardiac short-axis cine MRI recordings

A robust model-based segmentation method for automatically detecting left ventricular endocardial borders in cardiac MRI short-axis slices is proposed. The method is evaluated with expert drawn contours on an extended dataset (83 cases). The contour detection involves the four steps: (a) A region of interest enclosing the left ventricle is computed using a size invariant circular Hough transform. (b) Segmentation is performed by an adaptive threshold detection method taking model-based information into account. (c) Papillary muscles are included into the endocardial border by merging scattered regions and interpolating the resulting contour with approximating periodic splines. (d) Outliers and partially wrong contours are finally removed by applying a three dimensional median filter.