Guotai Wang's homepage

The aortic valve is significant in image-guided cardiac intervention and therapy. The GVF-snake is used for the segmentation of aortic valve in US image for the following reasons: 1) the cardiac valve has several boundary concavities and GVF-snake can converge to these concavities. 2) GVF-snake allows a simple initialization of the snake since it has a large capture range, which can reduce the user interaction. In this project, user need only draw a small circle in the center of the image as the initial counter and it will converge after less than 20 iterations. The initial circle can also be automatically generated. As a result, accurate segmentation can be obtained without user interactions. （C++ implementation）

Graphcut has been widely used in medical image segmentation. However, since adjacent organs have similar grayscales to those of liver, many segmentation methods can easily lead to over-segmentation. In this project, I tried to incorporate a shape prior into the graph cut method. The shape prior is employed to construct the weight of each edge, so that a shape energy is defined and incorporated in the energy function. Compared with the initial graphcut without shape prior, this method can obtain a smoother counter and exclude adjacent organs. Max-flow is used to solve the optimization problem.（C++ implementation）

Point correspondence between shapes play a critical role in ensuring the performance of statistical shape model. I tried different methods, including ICP, TPS-RPM, SPHARM, and AFDM. In my application of modeling liver shapes, which have a complex variation, I found AFDM has a better performance.(MATLAB implementation)

I developed a toolkit to facilitate the evaluation of medical image segmentation. Evaluation based on metrics including Dice coefficient, average symmetric surface distance (ASD), and Hausdorff distance were implemented. All the quantitative evaluation could be achieved by this toolkit. In addition, distance errors were visualized to display the mis-segmentation. For example, in the left figure, the shortest Euclidean surface distance was measured in millimeters and visualized. Blue color shows over-segmentation and red color shows under-segmentation.

Medical images exhibit transform sparsity, and they have a sparse representation in terms of spatial finite differences, their wavelet coefficients, or other transforms. In this project, a non-uniform random sampling method was used to measuring only a subset of image’s k-space. The wavelet coefficients, which could be treated as a sparse signal, were reconstructed from the unsampled image’s wavelet transform using L1 penalty, solved by iterative soft-thresholding. Then the initial image can be reconstructed by the wavelet coefficients. (MATLAB implementation)

I write a function to generate the Shepp-Logan phantom, the user can adjust different parameters such as Xoff, Yoff, Dx, Dy, Nx and Ny, so that images in different sizes and resolutions can be obtained. Then I write a function to realize Radom transformation. In addition, the projection data can also be acquired by using Joseph algorithm and blob approach, which can be compared with the result of Radom transformation. To backproject the data, a ramp filter is implemented on the projections and then the results are back projected. (MATLAB implementation)

In this project, I simulated the 3D cone-beam CT reconstruction. The whole progress consists of four parts:(MATLAB implementation)

• generate 3D phantoms with arbitrary ellipsoids
• generate cone-beam data
• filter the data with H_ramp(t)
• back project the filtered projection data

Monte Carlo simulations offer a flexible approach to photon transport in turbid tissues. It is based on macroscopic optical properties of small units of tissue volume. The most famous related work is that of Lihong Wang, who modeled the photon transport in multi-layered tissues. In this project, the Monte Carlo simulation deals with the transport of an infinitely narrow photon beam, perpendicularly incident on a tissue under which a vessel is parallel to the surface of the tissue. The tissue, which is infinitely wide and deep, and the vessel, which is infinitely long are described by the refractive index, the absorption coefficient, the scattering coefficient, and the anisotropy factor. This simulation can help a better understanding of the optical imaging of tissue with vessels. This work was mentored by Shanbao Tong.

In this project, I implemented an image processing software which can handle a lot of tasks: (C++ implementation)

• histogram analysis and segmentation based on threshold value (Otsu, Entropy)
• morphological operation such as opening and closing
• filtering and noise deducting
• edge detection (use different kernels such as sobel, pewitte, Roberts and user defined kernel)
• distance transform and skeletonization, followed by reconstruction

This project is the dissertation for my second major (Intelligence Science and Technology) when I was an undergraduate student. This project studied the application of texture features in face recognition and texture features based on gray level co-occurrence matrix (GLCM) from different directions were used. The correlation extracted from GLCM based on different offset in the same direction was fitted in to cubic polynomial curve and the factors of the curve were used to construct feature vectors. In the last step, classifiers based on support vector machine (SVM) were employed for identification. This work was mentored by Jingqi Yan.