On the accuracy of grey-level gradients for display and classification in medical imaging

Gradients in a 3D image indicate boundaries of regions with different grey values. They are used for computer assisted detection of the boundaries as well as for local shape description for display. Gradients can be approximated in different ways from the 3D image. In our work we used a discrete approximation of the partial derivatives of a 3D gaussian convoluted with the image scene. In this paper we present results of investigations on the behaviour of this approximation in the presence of noise.