Investigating the Influence of the Kernel Size on the Performance of Three-Dimensional Ultrasound Volume Reconstruction Methods

Ultrasound imaging is commonly used in a wide range of medical procedures. In fact, conventional two-dimensional (2D) ultrasound systems provide cross-sectional ultrasound images, called B-mode images, of the scanned three-dimensional (3D) anatomy. Freehand 3D ultrasound offers an attractive approach to extend the capabilities of conventional 2D ultrasound systems by mapping the acquired B-mode images into the 3D space and synthesizing a 3D ultrasound volume of the scanned anatomy. However, the synthesized ultrasound volume usually includes holes that are created due to the irregular spatial distribution of the B-mode images. Hence, 3D interpolation methods have been proposed to estimate the gray-level intensities of the empty voxels inside the holes. These methods often employ a kernel to estimate the gray level intensities of the empty voxels based on the neighboring voxels that have known gray-level intensities. In general, these methods assume that the size of the kernel is known a priori and it can cover all holes in the synthesized ultrasound volume. However, in real-life freehand 3D ultrasound imaging procedures, the sizes of the holes might vary drastically, which impose the need to adjust the sizes of the employed kernels. This paper presents an experimental study to investigate the effect of varying the kernel size on the interpolation accuracy and execution time of two well-studied 3D interpolation methods. The results indicate that the best possible performance of the 3D interpolation methods can be achieved when the kernel size is close to the size of the hole under consideration. Furthermore, the results indicate that setting the size of the kernel to values larger than the hole degrades the interpolation accuracy and increases the execution time of the 3D interpolation methods. The results reported in the current study can be employed to develop adaptive 3D interpolation methods that enable high interpolation accuracy and low execution time.