A boundary optimisation algorithm for delineating brain objects from CT-scans

The authors present an image processing algorithm for contour refinement of anatomical objects from brain CT scans. The algorithm uses an optimisation function to detect the contour points and to determine the belief that a point is lying on the object boundary. This optimisation function takes into account the following image features: the amplitude of the gradient, the orientation change of the gradient along the contour and a curvature measure. The proposed technique consists of using the three image features as variables in a Fisher linear classifier. That means a linear transformation vector W^T =(α₁, α₂, α₃) must be determined to achieve a maximal separation of two sets of points-namely boundary points and non-boundary points. The training sets for the Fisher linear classifier are acquired by a manual selection process, yielding typical boundary points and non-boundary points for different brain objects. The authors also discuss experiments in which the technique is applied to CT-scans of normal human brains