• No results found

Medical Image Segmentation

cians that need specific analysis of the airways to engineers that can use the software as a starting point for other applications.

• Although not yet capable of segmenting all the airways on a CT image due to limits on CT resolution, the results obtained using the algorithm show that the method is comparable with other results.

• Unlike other algorithms present in the literature, the airway segmentation method presented here proved reliable across different types of CT scanners and parameters. Moreover, the algorithm can be easily modified, adding new options as new datasets become available.

4.2

Medical Image Segmentation

In computer vision, image segmentation is the process of partitioning an image into salient and non-overlapping regions to help identify, classify, analyse or sim- ply recognize objects or other relevant information. The idea underlying this process is to assign a specific label to pixels that share a certain characteris- tic [157]. In this way, all the pixels with similar characteristics, such as colour, texture or intensity can be gathered into different regions in order to analyse, in an easier manner, an object of interest or a part of the image. In fact, adjacent re- gions can be significantly different with respect to some characteristics [158, 159]. This technique is widely used in the field of medical imaging, as the resulting labels can be combined to create 3D reconstructions of the regions of interest, using algorithms that reconstruct new data points starting from a discrete set of known points, called interpolation algorithms. An example of an interpolation algorithm is the marching cubes algorithm [160], which forms an imaginary cube taking eight neighbour locations at a time. The algorithm then determines what polygon is necessary to represent the part of the isosurface passing through this cube. The individual polygons are then fused into the desired surface.

In the last three decades, several automated image segmentation techniques have been developed. An important aspect to bare in mind is that an optimal segmentation method does not exist, and each situation may require a different

technique. In particular, the technique to be used for segmentation is specific to application, imaging modality and type of body part to be studied. As an example, brain segmentation requires methods that are different from those re- quired in the thorax. Also, another important aspect to consider when selecting a segmentation algorithm is the possible presence of noise on the image, which my be due to partial volume effects, motion or ring artefacts, and noise related to sensors and electronic systems. Here, some of the most common techniques for segmentation used in medical imaging are briefly presented [161].

4.2.1

Thresholding

Thresholding represents the simplest method of image segmentation. It consists in applying an intensity threshold value and adding all the pixels with an intensity above or below this threshold. This way, the intensities of the image are binarized into two classes, with all the pixels above the threshold grouped in one class and the remaining pixels in another class [162]. This threshold can be chosen by analysing the histogram of the image, a graphical representation of the image intensities distribution. This method is suitable to segment regions or objects with intensities that stand out from the rest of the image. The main issue with this technique is that usually it not simple to choose a single threshold to label a region and, therefore, it often requires user interaction (i.e., the threshold value has to be chosen and evaluated by the user). Therefore, thresholding is often used as a first step in more complicated segmentation techniques, in order to have a first approximate separation. A typical example of usage of this technique is as the first step in lung segmentation from CT images. This way, the soft tissue of the lung can be roughly separated from the surrounding chest structures with higher HU values (see Table 2.1) [163]. However, simple thresholding can not be used in operation such as airway segmentation, as image noise and the high number of voxels with intensities similar to air in the lung tissue (as in patients with emphysema) would not allow a proper distinction of airways from the surrounding tissues.

4.2 Medical Image Segmentation

4.2.2

Edge Based Segmentation

Edge based segmentation is a method based on the identification of boundaries which separate the different regions. To this end, discontinuities in grey level close to uniform grey level regions are considered edges of the region itself. Edge based algorithms use edge detecting operators based on gradient (derivative) functions, such as Prewitt, Sobel, Roberts (1st derivative), and Laplacian (2nd derivative). Problems with this technique are the possible presence of noise or weak edges that can greatly affect the segmentation results [161]. For this reason, edge-based algorithms, like thresholding techniques are often used in conjunction with other techniques for complete segmentation. Examples of edge based segmentation algorithms are edge relaxation [164] and Hough transform based [165].

4.2.3

Region Growing Segmentation

Region growing algorithms involve extracting pixels within an image region that are connected based on some predefined criteria [161]. The segmentation starts from an initial “seed region” (one or more pixels) and check a neighbourhood re- gion of pixels to assess whether these neighbours satisfy the predefined condition. For 3D images, two approaches can be used to check neighbours; face connectivity and full connectivity. With a face connectivity approach, 6 neighbours connected to the seed pixels are considered (four pixels surrounding the seed on the same slice and two pixels corresponding to the seed on the previous and next slices), while with full connectivity, the 26 pixels surrounding the seed are checked. Pix- els that satisfy the condition are added to the initial region and their neighbours are checked in turn. The process continues as long as new pixels are added to the region. A representation of region growing segmentation is included in Figure 4.1. There exists several region growing segmentation techniques depending on (i) the selected criteria to add a pixel to the growing region, (ii) the connectivity indicating the neighbourhood size, and (iii) the strategy used to add neighbouring pixels. The main disadvantage of this method is that, oftentimes, the seed region and the inclusion criteria have to be set manually. Moreover, if several regions need to be extracted, different region seeds must be selected. Examples of region growing segmentation methods are the “Threshold Connected”, which evaluates

(a) (b)

Figure 4.1: Region growing segmentation scheme. (a) A seed pixel (red) is se- lected and the algorithm starts checking neighbouring pixels (blue arrows). (b) The pixels around the grown region (red) are assessed.

if the pixels’ intensity value is inside a specific interval, and “Otsu segmentation”, which tries to minimize the error of misclassification of the pixels by finding a threshold that classifies the image into two clusters. Otsu segmentation seeks to minimize the area under the histogram due to one cluster which lies on the other cluster’s side of the threshold [61]. For the development of the described airway segmentation method, a region growing technique was chosen for the simple ini- tialization process, the little user interaction required, as well as the good results showed in the literature.

4.2.4

Level Set Segmentation

The level set segmentation approach was proposed for the first time in 1988 [166] and refers to a numerical technique that tracks the evolution of contours and surfaces in an image. It uses and solves a partial differential equation (PDE) to facilitate the segmentation. In particular, a contour is embedded as the zero level set of a higher dimensional function called the level-set function, which is then evolved under the control of the PDE. At any time, the evolving contour can be obtained by extracting the zero level-set function from the output. This technique is widely used in medical image segmentation because of its capacity for modelling and handling complex shapes and topological changes, as well as its computational efficiency in 3D images. In a typical approach, a contour is

4.2 Medical Image Segmentation initialized by the user and is then evolved until it fits the form of an anatomical structure in the image. The main problem of level set methods is that choos- ing a proper initialization contour (speed function) may be complicated in some applications. As well as in region growing segmentation, many different imple- mentations and variants of the level set segmentation have been proposed [167]. A typical example of a level set segmentation method is the fast marching algo- rithm, that starts from an initial position on the front, and systematically moves the front forward one grid point at a time, successively solving the Eikonal equa- tion [168]. In the case of airway segmentation, a fast marching approach requires the definition of a very complicated speed function, as shown in Schlath¨olter et al [169], that can highly increase the computational complexity of the system.

4.2.5

Atlas Based Segmentation

A fourth approach to medical image segmentation is to exploit the anatomical features of the structures or region of interest with a repetitive form or geometry, using atlas based segmentation methods. If an atlas or template is available, these techniques can be very useful. The atlas is usually generated by comparing images of several subjects in order to extract a model explaining the variation in the shape of the structure itself. This generated model is then used as a reference frame for the segmentation of new images [162]. However, the creation of the atlas needs registration of all the images from several patients as well as a probabilistic representation of the registered data. Moreover, due to anatomical variability, an accurate construction of the atlas can be quite complex and expert knowledge is often required in building the database to be used for the atlas construction. Therefore, atlas based segmentation methods are generally used for the segmentation of structures that tend to remain stable over the population of study. An example of this technique lies in magnetic resonance (MR) brain imaging for different types of segmentation [170].