A Statistical Shape-Based Patient-Specific Anatomical Structure Model

A patient-specific anatomical structure model has been widely used in many medical applications. However, in practical applications, to effectively construct a patient-specific anatomical structure model is a challenge, the reasons are: (1) the manual marking process for landmark points is time-consuming and is prone to have false points; (2) the correspondence establishment is difficult; (3) the performance of the model is limited. Therefore, the purpose of this study is to automatically construct a patient-specific anatomical structure model to solve these difficulties. Firstly, the input data are preprocessed to enhance the region of interest in CT scan images. Then, the region of interest is regarded as a training specimen, and the triangle is used to mesh the training specimen. Meanwhile, vertices contraction strategy is introduced to iteratively contract triangle meshes, and the correspondences are established through improved B-spline free-form deformation. Finally, principal component analysis is used to generate the final patient-specific anatomical structure model. Experimental results on 30 pelvic CT scan images verify that the proposed method outperforms the compared methods.

However, the traditional methods for the anatomical structure model have several problems: (1) they require manual selection of landmark points and thus have lower accuracy. (2) the correspondence between the template specimen and the target specimens is difficult.
(3) the accuracy of the constructed model is limited.

B. PREVIOUS STUDIES
In order to solve these problems, some previous studies have proposed many methods to construct the patient-specific anatomical structure model. Kelemen Consequently, Dalal et al [14] developed the Landmark Sliding Method (SLIDE) to automatically mark landmark points. In SLIDE, target shapes were aligned to template shape, and the initial correspondence of the landmark points was evaluated based on Euclidean distance. Then, the landmark points were iteratively slid along the tangent planes of the landmark points to minimize the shape deformation and shape representation error. SLIDE could construct an anatomical structure model. However, it was necessary to continuously calculate the tangent planes of each landmark point during the sliding process. The process was complicated and only applicable to an image with a little number of landmark points. Based on the previous methods, Barratt     We model the slices of these images to construct the patient's anatomical structure. The "patient-specific anatomical structural model" means that we reconstruct the 3D geometric model of a patient with the statistical shape model based on the training dataset. The schematic diagram of our method is shown in Figure 1.
The image enhancement step is to remove the noise in the image, so that the region of interest in the target image is more prominent and the segmentation results can be more accurate.

A. DATA PROCESSING
Pre-processing methods are applied to CT images to enhance the region of interest. First, the region of interest is extracted from the CT images by using the region growing method [20]. Then, the Marching Cubes method [21] is employed to generate a surface mesh while the inter-slicer connectivity is maintained.
Finally, a mean filter is applied to each region of interest to reduce noise. The pre-processed regions of interest are used as training data for subsequent modeling.

B. VERTICES CONTRACTION
First, a template specimen is selected from training , , , The vertices contraction process is depicted in Figure 2. (2)   , ab VV is an edge; The vertices contraction is applied to all images in training data  .  S . We extend equation (1) in the form of coordinates, as shown in equation (2):

C. CORRESPONDENCE ESTABLISHMENT
where M is the number of main deformation modes.
G is selected by equation (17): where P is the percentage of the whole deformed models that we want the first G principal components to represent.

III. EXPERIMENTAL VERIFICATION
The performance of the proposed method is compared with the other methods on 30 pelvic CT scan images.

A. USABILITY EVALUATION
In order to verify the usability of the proposed method, the 30 pelvic 3D CT scan images are used. The CT resolution is 0.9mm on the plane and 1.5mm between slices. The left femur region of the pelvis is the region of interest. By processing the pelvic CT scan images, we obtain 30 left femur images and regard them as training data. To observe the training data more intuitively, we place 30 left femur images in a plane as shown in Figure 3.   Template S1 Meshed S1 Simplified S1 Target S2 Meshed S2 Simplified S2

Registration Result
Vertices contraction

Meshing
Vertices contraction FIGURE 7. Visualization Registration process between template specimen 1 S and target specimen 2 S . deformation models as shown in Figure 9.

PROPOSED METHOD
To evaluate the performance of our method, we compare our method with Eck, Abler and Plessers. We regard compactness, specificity, generality and representation error as evaluation indicators.
The comparison results are shown in Figure 11. From   Figure 14, and the results are shown in Table 2. Therefore, we can conclude that vertices contraction has batter mesh simplification effect than QECD and MCEC.

IV. DISCUSSIONS
Different from the compared methods, our method regards the statistical shape model as the basis. By [9] Liu, T., Qin, S., Zou, D., Song, W., and Teng, J.,