ITK Pre-Processing Algorithm Wrappers in Python/CurvatureFlowFilter

Curvature flow filter

This filter is implemented using the itkCurvatureFlowImageFilter class in ITK. The curvature flow filter is among the diffusion-like filters present in ITK. The other prominent filters are Gradient Anisotropic Diffusion, Curvature Anisotropic Diffusion, and Min-Max Curvature Flow Filter. The diffusion speed at a particular pixel (or voxel) is defined to be proportional to the curvature and the magnitude of the gradient at that pixel (or voxel). The regions of high curvature diffuse very quickly - this removes noise artifacts. The regions of less curvature (such as image edges) diffuse much slower, thereby preserving the features. On application of this filter, however, there would be a small amount of smoothing near the edges in the image.

Parameters:

• 1. TimeStep - this parameter refers to the time step involved in solving the partial differential equation in the algorithm. For 2D, the typical value to use is 0.125. In general, for n-dimensional images, the timestep should be restricted to be less than ^[1].

• 2. NumberOfIterations - this specifies the number of iterations that the solver must perform before returning a solution image. More the number of iterations, more smooth the image would be. A typical value for the number of iterations is 10^[1]. The effect of running the curvature flow algorithm for 5, 10, and 25 iterations on seg_orig can be seen in Figures 1, 2, and 3, respectively. The edges are reasonable preserved in Fig. 1, while smoothing can be observed in Fig. 2. The image given in Fig. 3 is over-smoothed, and has lost many details from the original image. Curvature flow image filter is probably not the best filter to apply for this image.

Back to main page.

Fig. 1: Curvature flow image filter applied on seg_orig, with timestep = 0.125, and number of iterations = 5.

Fig. 2: Curvature flow image filter applied on seg_orig, with timestep = 0.125, and number of iterations = 10.

Fig. 3: Curvature flow image filter applied on seg_orig, with timestep = 0.125, and number of iterations = 25.