Fig. 1
Two CT images belonging to the same patient dataset. The intensity values of the cortical zone are different in the diaphysis (left) and the epiphysis (right). The cortical area is much thinner in the epiphysis (right)
2.2 Fractured Bone
Fractured bone tissue is more difficult to identify because it has some additional features to be considered. Due to the fact that bone fragments may have arbitrary shape and can belong to any bone in a nearby area, it is necessary to label all the fragments during the segmentation process. In some cases, this labelling requires expert knowledge. In addition, a priori knowledge can not be easily used because it is uncommon to find two identical fractures and therefore it is difficult to predict the shape of the bone fragments, specially in comminuted fractures. On the other hand, bone fragments are not completely surrounded by cortical tissue, since they have areas on the edges without cortical tissue due to the fracture. Finally, proximity between fragments and the resolution of the CT image may cause that different fragments appear together as one in the image. For this reason, smoothing filters should be used with caution. This type of filters can deform the shape of bone fragments and fracture zones or even remove small bone fragments. In some cases, it is necessary to detect the fracture zone of each fragment after its segmentation. The fracture zone is the area of the bone where the fracture occurs and is composed of trabecular tissue (Fig. 2). In situations in which bone fragments appear connected, it is difficult to accurately identify the fractured zone of each fragment. Therefore, post-processing can be necessary to delimit fracture zones in these situations.
Fig. 2
CT slices that represent some different simple bone fractures. Fracture lines are marked in red
The method applied in fractured bone identification depends on the fracture type. Based on the fracture line, a fracture can be classified as (Fig. 3): greenstick, transverse, oblique, spiral, avulsed, segmental and comminuted [7]. In a greenstick fracture (Fig. 4a) there are no fragments because the bone is not completely broken. Thus, labelling is not necessary. Since the fracture barely changes the shape of the bone, segmentation methods that are based on previous knowledge are available. Nevertheless, the edges of the fracture zone, composed of trabecular tissue, may require special processing. The detection of the fracture zone is specially complicated since the bone is not completely broken and trabecular tissue is very heterogeneous. Therefore, the fracture zone can be fuzzy in the CT image.
Fig. 3
Fractured bones classified by their fracture lines
Fig. 4
CT images that represent different simple fractures. (a) contains, among others, a greenstick fracture, since the bone is not completely broken. The remaining images contain simple fractures with (b) and without (c, d, e, f) bone displacement
Transverse, oblique and spiral fractures (Fig. 4b, c,d,e, and f) can be similarly treated during the segmentation. Despite of having different fracture lines, these types of fracture generate two fragments with similar shape. Labelling is necessary, but expert knowledge is not required. Segmentation methods that can be applied depend on whether or not there is displacement. If there is no displacement (Fig. 4c, d, e, and f), they can be processed as a greenstick fracture but considering that there are two fragments. These two fragments can be completely joined, hence an additional processing to separate them may be required. In order to detect fracture zones, the same issues applicable to greenstick fractures should be considered. In the case that there is displacement (Fig. 4b), the probability that both fragments are jointly segmented decreases and methods based on prior knowledge are almost discarded. In return, the fracture zone is easier to be identified. Avulsed fractures normally occur near a join thus the fracture zone is composed almost exclusively by trabecular tissue and the boundaries of the fragments are weak. This complicates the identification of the fracture zone because practically the entire fragment is surrounded by trabecular tissue. Segmental fractures are simple fractures that generate three bone fragments. Therefore, they can be treated as transverse or oblique fractures but considering that there are two distinct fracture regions. Comminuted fractures (Fig. 5) add some additional constraints, hence this is the type of fracture that is more complicated to be segmented. Comminuted fractures usually generate small fragments and bone may be deformed due to the fracture. This is because comminuted fractures are usually associated with crush injuries. In most cases, some fragments overlap in the CT image and require additional processing to be separated. Labelling is necessary and expert knowledge is strongly required to identify fragments. The detection of fracture zones is complicated in this case. Due to the complexity of the fracture, several fracture zones are generated. Since the relationship between fragments in this type of fractures is many-to-many, it can be necessary not only to identify fracture zones, but also to delimit which part of the fracture zone corresponds to each fragment. As mentioned before, some fragments can overlap due to the fracture and therefore post-processing and expert knowledge can be required to accurately identify fracture zones.
Fig. 5
CT images representing highly comminuted bone fractures
3 Currently Proposed Approaches
3.1 Healthy Bone
In recent years, many approaches have been proposed in order to segment bone tissue from CT images. Most of these methods are focused on the segmentation of a specific area. In [25] authors combine region growing, active contours and region competition to segment carpal bones. An expectation maximization algorithm has been utilized to segment phalanx bones [23]. The method requires a previously generated CT atlas. In [18], 3D region growing is used to segment the inferior maxillary bone from CT images. In order to fill holes in the segmented surface, a morphological operation of closing is used. Then, 3D ray casting is applied to segment the internal region of the bone by determining which points are inside of the outer shell. The segmented voxels are classified as cortical or trabecular bone using a fuzzy c-means algorithm. To improve the result, an adapted median filter allows to remove outliers. A 3D region growing method has also been used to segment bone tissue in [32]. Both the seeds and the threshold are calculated automatically. Since they use an unique threshold, some areas of bone are not segmented and they propose a method to fill them. This segmentation approach has been tested to segment skull and spine bones. A novel active contour model is utilized to segment bone tissue in [28]. The statistical texture method has also been proposed to segment mandible bones from CT images [19]. In [17] authors use a 3D deformable balloon model to segment the vertebral bodies semi-automatically. Graph cuts have also been used to segment vertebrae [2]. Previously, seeds are automatically placed using the matched filter and vertebrae are identified with a statistical method based on an adaptive threshold. Cortical and trabecular bone are then separated by using a local adaptive region growing method. In [15], Willmore flow is integrated into the level set method to segment the spinal vertebrae. Graph cuts have also been employed to segment the hip bone [16]. Most of these approaches can not be applied to the segmentation of fractured bone tissue because they take advantage of the prior knowledge of the shape of the bones.
Statistical methods are frequently used to segment bone tissue [3]. In this case, they use a generative model to classify pixels into cortical bone or another tissue. A learned model is constructed by modeling probability functions using Gaussian mixture models. Then, the learned model allows to assign a probability to each pixel and a maximum a-posteriori probability rule enables a crisp classification. In [12], a genetic algorithm is used to search the better procedure to segment bone tissue and to separate cortical and trabecular tissue. For that, the genetic algorithm requires previous expert information. Despite the results obtained, learning based methods can not be easily used to segment fractured bones because previous learning is not available in most cases.
Several methods are based on the fact that the shape and the anatomy of the bone are known [31]. In this work, an adaptive threshold method is utilized to segment bone tissue. However, the method can not be applied to segment bone fractures because it is based on the supposition that bone fragments are completely surrounded by cortical tissue, and this is not always true in the case of a fracture. All the revised works for segmenting healthy bone from CT images are summarized in Table 1.
Table 1
Summary of the works for identifying healthy bone which are described in this paper
Authors | Requirements | Interaction | Methods | Evaluation set | Achievements |
|---|---|---|---|---|---|
Sebastian et al. (2003) | – | Specify parameters | Region growing, active contours and region competition | Carpus | Combine the advantages of all the methods used |
Mastmeyer et al. (2006) | – | Set seeds and markers | 3D deformable balloon model | Vertebrae | Vertebra separation |
Battiato et al. (2007) | A learned model | Set the threshold | Gaussian mixture models | Knee | Cortical tissue pixels classification |
Ramme et al. (2009) | CT atlas | Place landmarks | Expectation maximization | Phalanxes | Semi-automatic segmentation |
Moreno et al. (2010) | – | Set the seed point | 3D region growing | Inferior maxilar | Bone tissue classification |
Zhao et al. (2010) | – | – | 3D region growing | Skull | Threshold and seeds automatically selected |
Aslan et al. (2010) | – | – | Graph cuts and region growing | Vertebrae | Automatic cortical and trabecular tissue classification |
Zhang et al. (2010) | – | – | Adaptive thresholding | Calcaneus and vertebra | Automatic segmentation |
Truc et al. (2011) | – | – | Active contours | Knee and heart | Bone contours extraction from CT and MRI images |
Nassef et al. (2011) | – | – | Statistical texture | Mandible | Identification of different bone tissues |
Janc et al. (2011) | Expert bone identification | – | Genetic algorithm | Mandible, skull and knee | Cortical and trabecular tissues separation |
Lim et al. (2013) | – | Set initial contours | Level set | Vertebrae | Deal with missing information |
Malan et al. (2013) | Previous manual segmentation | – | Graph cuts | Hip | Detailed tissue classification |
3.2 Fractured Bone
The methods applied to the segmentation of healthy bone could not be suitable for segmenting fractured bone. This is because, as seen in the previous section, fractured bone has different features. Moreover, the identification of fracture bone requires to carry out additional steps, such as labelling the fragments or splitting wrongly joined fragments. Currently proposed methods to perform these steps are described below.
3.2.1 Fragment Segmentation and Labelling
There are several papers that are focused on the identification of fractured bone. With this aim, threshold-based methods are used in most cases. The most basic threshold-based method consist in defining an intensity interval that corresponds to bone tissue and calculating the pixels in the image that belong to this interval [24]. The intensity interval can be defined manually or can be calculated from the information provided by the image. On the other hand, the interval can be used in the hole stack or can be defined for each slice. The second option is usually the most successful because, as seen in Sect. 2, intensity values differ between slices. Several works propose to use thresholding to segment fractured bone. In [20], ulna, radius and carpus are segmented to simulate a virtual corrective osteotomy. Therefore, the segmentation is performed on non-fractured bones and then the segmented bones are virtually cut. In order to separate bone from other tissues, an user-defined threshold is used. In [27], the area where the bones are located is detected using a threshold-based method. Then, they present manual and semi-automatic tools for interactively segmenting bone fragments. This toolkit includes separation, merge and hole filling tools to generate individually segmented fragments from the result of the threshold-based segmentation. Thus, the method achieves accuracy at the expense of requiring a lot of user intervention. A global fixed threshold method has been utilized in [26] to detect the trabecular bone fracture zone. Due to the difference of intensity values between slices, it is difficult to set a threshold that fits all the slices.
Region growing is a threshold-based method that allows to limit the segmentation to a specific area [8]. To that end, the algorithm requires to place seeds before starting the segmentation. The selection of the seed points can be performed manually or automatically. The manual placement of the seeds enables the labelling of the different bone fragments. Moreover, the algorithm also needs to define an intensity interval. As in the previous case, the interval can be defined globally or for each slice. Once the seeds have been placed and the interval has been defined, the algorithm check all their neighbouring pixels. If the intensity of a neighbouring pixel is outside of the defined interval, it is discarded. Otherwise, the pixel is included in the segmented area and its adjacent pixels are studied. The algorithm stops when there are no pixels to study. The result of the algorithm can differ depending on the criteria used to accept or discard pixels. The basic algorithm accepts a pixel if its intensity is inside the interval. This approach allows to detect small bone features but image noise can also be segmented. However, noise can be mostly reduced using smoothing filters. Therefore, this approach can be suitable for segmenting fractured bone. Other approaches decide to accept or discard a pixel based on the intensity value of its neighbours. The simplest option is to accept a pixel if all its neighbours have intensity values inside the interval. Another option is to use a criteria based on statistical values calculated from the neighbouring pixels. In this case, small features could be discarded. Thus, this variation could not be suitable for segmenting fractured bone.
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
