Adaptive kernel interpolation on 3D image reconstruction of breast ultrasound images

Three-dimensional ultrasound imaging is getting popular due to the ability to visualize volumetric representation of tissues and organs while having non-ionizing radiation properties. This project aims to reconstruct a 3D ultrasound image from series of BScan images. 3D ultrasound image allows better overview of whole organ. Holes will appear during the reconstruction process due to the reason that pixels will always be small when comparing to real world spatial measurement. The process of finding potential holes pixel by pixel takes huge computation cost. Besides, conventional interpolation technique uses fixed kernel size which does not tackle holes of different size. There will be some holes that are too far for the kernel to reach out and there will be holes that are just one pixel away from filled pixel. Thus, using a fixed sized kernel waste computation resources. A technique of reconstructing the image volume is introduced. Ultrasound probe will be rotated 360° around the breast to capture each image individually. Fan-like 3D image is formed with this technique of image acquisition. A new technique is introduced to find interested holes in a faster approach by convoluting the image with a smaller size kernel. This technique tackles the problem of finding holes pixel by pixel. This approach will reveal interested holes by tagging each holes with value to represent how far away the filled pixels are. Nonetheless, the existing technique that deals with hole-filling uses nearest neighbour interpolation with fixed kernel is slow and inefficient. The goal is to develop a holefilling technique that uses variable sized kernel on nearest neighbour and Gaussian interpolations. The improvement of the Gaussian interpolation technique is done with an addition of sigma filter. Gaussian interpolation uses values from nearest neighbour as local mean and computes the local variance accordingly. Sigma filter helps remove noise by eliminating values that difference the mean by more than two sigma. Results of proposed technique are interpreted quantitatively. Homogenous and nonhomogenous region are extracted and compared. Gaussian interpolation with sigma filter gives the best result in both homogenous and non-homogenous area. The quantitative study of each technique are compared in terms of standard deviation, average absolute difference, iteration, time, kernel distribution. The proposed methods produced 80% of the kernel that distributed in the kernel size of 1 and 2 in nearest neighbour interpolation. As for Gaussian interpolation, 50% of the kernels used are kernels of size 2 and 3, and 30% left in kernel size of 1 and 4. The iteration and time required for the proposed holes-finding and holes-filling technique has improved up to 4 times faster compared to conventional methods. The addition of sigma filter manages to suppress noise while keeping edge details. To conclude, a more efficient system is shown to reconstruct raw B-Scan image into 3D image that could be visualized using 3D visualization tools easily while improving the conventional way to interpolate the holes.

Preview Text FK 2018 158 - IR.pdf Download (1MB) | Preview

Actions (login required)

View Item