Volume rendering in medical ultrasound imaging based on nonlinear filtering

This paper explores the application of nonlinear filtering techniques in volume rendering of medical ultrasound data. Volume rendering methods based on an initial probabilistic classification have been used with success on Computed Tomography and Magnetic Resonance data, but have shown to fail in ultrasonic studies. Several considerations have to be made in ultrasound imaging, because of a substantial noise content and other limitations in the imaging system. A major concern is to render the boundaries between different tissues. Boundaries can be detected by computing local gradients in the volume, but these computations will generally be very noise sensitive. Noise can be reduced by applying linear filtering schemes such as the moving average filter, but this will blur the boundaries. We have investigated how nonlinear filtering schemes could increase the robustness of the gradient estimates. Recently, several researchers have proposed nonlinear filters, which, although they have different motivation, can be modeled as an anisotropic diffusion process. The filters are based on adaptive weighting of pixels in a minimal filter window, and they are shown to smooth within homogeneous regions, while enhancing boundaries. By iterating these filters, the signals will be segmented into piecewise constant regions. These filtering schemes tend to produce artificially sharp boundaries. In this paper we discuss how the edge enhancement property of the proposed nonlinear filters can be avoided. We propose a modified nonlinear filtering scheme which smoothes within homogeneous regions while leaving significant monotone transitions unaltered. The proposed filtering schemes are compared with alternative filtering models. Finally, we integrate the filtered data volumes in a volume rendering pipeline and illustrate examples from clinical studies with 3-dimensional ultrasound.