Добавил:
Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
[2.1] 3D Imaging, Analysis and Applications-Springer-Verlag London (2012).pdf
Скачиваний:
12
Добавлен:
11.12.2021
Размер:
12.61 Mб
Скачать

468

P.G. Batchelor et al.

tive measurements of blood flow or electrical activation [72]; in image-guided interventions, segmentations are often required for visualization in surgical planning or guidance [63]. Segmentation techniques have been applied to delineate a wide variety of organs from medical imaging data acquired using a wide range of modalities.

Approaches to segmentation vary a lot in terms of their sophistication and the amount of user input required. Purely manual techniques allow users to outline structures using software such as the ANALYZE1 package. Manual segmentation can be very accurate, but time-consuming, and is subject to inter-observer variation or bias. Semi-automatic techniques allow the user to have some control or input into the segmentation process, combined with automatic processing using computer algorithms. Finally, fully automatic techniques require no user input and often make use of some prior knowledge to produce the segmentation. The following sections review a number of semi-automatic and fully automatic approaches to medical image segmentation. Our coverage of this huge research area is not exhaustive, but we provide a few examples that give an introduction to the field.

11.5.1 Semi-automatic Methods

In this section, we consider three semi-automatic approaches to segmentation based on thresholding, region growing and deformable models.

11.5.1.1 Thresholding

Perhaps the simplest example of user interaction in segmentation is the specification of a threshold. Typically, a thresholding operation involves comparing every intensity in the image to the threshold value and setting it to 1 if it is greater than or equal to the threshold, or 0 otherwise, thereby creating a binary image. Figure 11.13 illustrates this operation on cardiac MRI data. Software such as ITKSNAP can be used to interactively adjust the threshold to produce the desired result. Thresholding is commonly applied as one step in a segmentation pipeline, but the presence of noise, image artifacts and other structures in the image mean that it is rarely enough on its own to produce an accurate and reliable segmentation. This fact can be observed in Fig. 11.13b, in which several pixels outside of the left and right ventricles are set to 1, despite not being the target structures of the segmentation.

1Biomedical Imaging Resource, Mayo Foundation, Rochester, MN, USA or ITK-SNAP [84], http://www.itksnap.org.

11 3D Medical Imaging

469

Fig. 11.13 Thresholding an axial cardiac MRI slice.

(a) Original slice. (b) Result of thresholding operation

Fig. 11.14 Segmenting the left ventricle using region growing on the thresholded axial cardiac MR slice.

(a) Thresholded image with seed point in the left ventricle indicated by the red cross. (b) Result of region growing, overlaid onto original cardiac MRI slice

11.5.1.2 Region Growing

One technique that can be used to refine thresholded images, such as that shown in Fig. 11.13b, is known as region growing. Region growing involves user interaction in the form of specifying a seed point. This is a pixel in the image that is known to lie inside the structure being segmented. The region is then iteratively grown by adding neighboring pixels that have similar appearance. The concept of similarity needs to be defined: for example, by specifying a range of intensity values around that of the seed pixel. If region growing is applied to a thresholded image then, to be considered similar, pixels should have the same binary value as the seed pixels. In addition, the concept of neighborhood needs to be defined, with 4-neighborhoods (i.e. only directly adjacent pixels) and 8-neighborhoods (i.e. including diagonally adjacent) commonly used in 2D images. Figure 11.14 illustrates the operation of the region growing algorithm for segmenting the left ventricle from the cardiac MRI data introduced in Fig. 11.13. A seed point has been placed in the left ventricle (indicated by the red cross) and the region growing algorithm has included all pixels that were connected to this seed, excluding all others. The resulting segmentation is of the left ventricle only.

470

P.G. Batchelor et al.

Algorithm 11.1 REGION GROW on seed pixel Require: Initialize all pixels to unknown region

Set current pixel to be inside region for all neighboring pixels do

if it is inside image bounds and currently has an unknown region then Compute similarity to current pixel

if similar then

REGION GROW on neighboring pixel else

Set neighboring pixel to background end if

end if end for

Fig. 11.15 The problem of leaks in region growing. (a) An axial cardiac MRI slice. (b) Result of thresholding operation, seed point for region growing in the left ventricle indicated by the red cross. (c) Result of region growing. The segmentation has ‘leaked’ into the right ventricle

The region growing algorithm can be implemented in a number of ways. The simplest, although not the most computationally efficient, is a recursive implementation. There are many implementations and this is the same problem as flood-fill or area-fill in graphics or vision. Pseudocode for recursive region growing is given in Algorithm 11.1.

One problem that the region growing algorithm can encounter is that of leaks. This problem is illustrated in Fig. 11.15. This shows a different axial cardiac MRI slice from that used in Fig. 11.13 and Fig. 11.14. This time the region growing has ‘leaked’ the segmentation into the right ventricle. Therefore, in region growing, care must be taken when selecting threshold values and specifying the similarity term to avoid such cases. The region growing algorithm can be extended to 3D, in which case the neighborhood in Algorithm 11.1 will extend to 3D accordingly. However, in 3D the problem of leaks can be exacerbated.