Abdominal aortic aneurysm (AAA) is an enlargement of the abdominal aorta due to a weakened aortic wall. If left untreated, AAA will expand over time, increasing the risk of aortic wall rupture. Accurate measurements of AAA are required for appropriate planning of repair strategies. The size and shape of AAA is important for the diagnosis, endovascular procedure planning and post-operative evaluation. High resolution, 3D black blood MRI using fast-spin-echo (FSE) sequences with variable flip angle train (SPACE) is advantageous for AAA imaging given its high scan efficiency and intrinsic black blood effect¹. Automatic segmentation methods allow quick and reproducible morphology quantification, and it is especially favorable for 3D imaging, which has large amount of thin slices. This study aims to propose a novel registration based geometric active contour model to segment vessel wall of AAA from 3D black blood MRI.

(a) MR Imaging: Six patients with AAA disease underwent 3D T1 weighted black blood MRI (DANTE-SPACE)¹ in a Siemens 3T Skyra scanner. The acquisition parameters were: TR/TE = 800~1040ms/20ms; echo train length (ETL) 60; 44~60 coronal slices with 1.3mm slice thickness; 32cm*32cm filed of view and 256*256 matrix; band width: 781 Hz/pixel; echo spacing: 3.58 ms; duration of the echo train: 215ms. (b) Image segmentation and analysis: Lumen and outer wall were segmented using an in-house software developed in MATLAB. The proposed algorithm used the theory of image registration, which outputs the relative movement of the pixels between two images after registration. The first image was segmented in advance, then the second image can be segmented according to the relative movement to the first one. Similarly, the third image can also be segmented after registering the second and the third image. Therefore, all images in a sequence can be processed one by one. In our study, we proposed a new registration based geometric active contour model that can refine and estimate the segmentation. The algorithm is focus on the minimization of the following energy function:

$$E=rE_{registration}+aE_{gradient}+sE_{similarity}$$

where $$$E_{registration}$$$ controls the the registration of two neighboring images² and $$$E_{gradient}$$$ controls the contour of the segmentation moving to the local minimization of the image. $$$E_{similarity}$$$ represents that the segmentation result should be similar to the reference image. Besides, the weight $$$r$$$, $$$a$$$ and $$$s$$$ correspond to positive parameter that balances the influence of this three terms.

$$$E_{registration}$$$ can be minimized by solving the following gradient flow^3,4:

$$\frac{\partial\phi }{\partial t}=\mu(\triangle\phi-div(\frac{\nabla\phi}{\left|\nabla \phi \right|})+\lambda\delta(\phi)div(g\frac{\nabla\phi}{\left|\nabla \phi \right|}))+\upsilon{g}\delta(\phi)$$

where $$$g$$$ is the edge indicator function. The first term is a penalty term used to penalize the deviation of $$$\phi$$$ from a signed distance function during its evolution, and the other two terms represent the gradient flow of the energy function.

The segmentation procedure contains two parts: lumen and outer wall segmentation. Manually segmented lumen and outer wall contours at the middle level of the AAA were used as initial contours in the first equation, then the images were segmented slice by slice. During the whole segmentation process, human interactions were minimized.

The results obtained on six patients showed that the proposed segmentation was comparable with manual segmentation. And quantitatively, the average Dice value⁵ of inner wall and outer wall are 91.75% and 91.22%, respectively, demonstrating good segmentation accuracy. Sample segmentation result is shown in Figure 1, where the lines in blue and in red represent the results obtained from our method and the radiologist, respectively. Good agreement between automatic and manual segmentation is noted. Reconstructed volumes of two AAAs (in light gray) are shown in Figure 2. The comparison of AAA segmentation obtained from our method (in blue) and the manual segmentation (in red) is shown in Figure 3. The good accuracy and fast speed of our proposed method make it a potential tool for clinical routine use. Our study indicates that the proposed segmentation method can quantify vessel wall dimensions reliably with good agreement with manual segmentation. The reduced segmentation time makes it potential for clinical applications.