[Literature Reading] Segmentation and Image analysis of abnormal lungs at ct:current approaches, challenges, and future Trends

Summary:

Segmentation of the lungs, which is the process of determining the boundary between the lungs and the surrounding chest, is an important step in image analysis. However, almost all of the image segmentation methods are only available for very little or no pathological conditions. When a degree or a large number of diseases or abnormalities of the challenging shape or appearance exist in the lungs, due to inaccurate segmentation, the computer-aided detection system may well be unable to depict these abnormal areas. In particular, such as pleural effusion, real change, and a large number of inaccurate segmentation have greatly limited the use of image processing methods in clinical research. In this paper, a critical summary of the method of lung Parenchyma segmentation based on CT image is presented, especially the accuracy and performance of the anomaly and typical pathological examination results are emphasized.

The existing segmentation methods are mainly divided into 5 categories: ① based on the ② region-based ③ shape-based ④ adjacent anatomy guidance (neighboring anatomy–guided) ⑤ based on machine learning methods. The feasibility and shortcomings of each class are illustrated and exemplified by the most common lung abnormalities observed on CT images. In the overview, the practical application and developing techniques of the methods introduced in combination with the practice of radiologists are introduced in detail.

Introduction:

The importance of image segmentation in image reconstruction and image analysis: (1) The most basic step of lung image analysis is the segmentation of the organ of Interest (lung); In this step, the organ is detected, and the anatomical boundaries are depicted automatically or manually, and (2) errors in organ segmentation can result in subsequent identification of diseased areas and a variety of other clinically quantitative false information, so accurate segmentation is required.

What is Object segmentation?

The purpose of medical image segmentation is to extract quantitative information about organs or lesions of interest in tissues (such as volumetric data, morphological data, structural patterns, etc.). In general, a segmentation problem can be seen as consisting of two key tasks: target recognition and target contour.

Target recognition is the determination of the position of the target in the image, and the target contour is the spatial range and composition of the target in the image.

There is a huge challenge in the division of the lungs. Voids and real changes, for example, result in imprecise positioning of the boundary. Similarly, the presence of pneumothorax or pleural effusion on CT images may distort the results of automatic segmentation, resulting in incorrect segmentation.

Image Segmentation Methods for abnormal lungs  

Before describing the lung segmentation method and evaluating its performance under different pathological conditions, first introduce the common imaging patterns encountered in lung CT images to fully understand the performance of the segmentation method and the difficulty difference between the specific type and the location of abnormal imaging patterns.

thresholding-based Methods

The threshold-based segmentation method is the simplest and easiest to understand segmentation method, and the threshold segmentation method usually creates a binary segmentation based on the image attenuation value (atenuation values). The threshold program attempts to determine the attenuation value, called the Threshold (S), to create attenuation values for all image elements of the partition, satisfying the threshold interval. The process of threshold segmentation:

The threshold-based segmentation method is simple and effective for comparing the differences between specific areas of the image. Moreover, these methods can produce better results for the segmentation of CT images (attenuation value, in Hounsfeld unit measurement, there is a good definition range between different organizations of CT images. However, the threshold-based segmentation technique usually does not take into account the spatial characteristics (lungs) of the target object. In addition, this technique is more sensitive to noise and imaging artifacts than other methods of lung segmentation.

An overview of the thresholding-based segmentation method:

For small lesions or normal lung images, the image can be segmented by selecting the appropriate threshold value. However, for some lesion areas, although the lesion area and the lungs have the same threshold, the threshold interval usually removes the lesion from the lung image. Figure6 Show examples of pleural effusion and consolidation, which may require a variety of morphological operations or artificial false positive removal processes to correct the resulting segmentation

region-based Methods

The basic condition of region-based segmentation method is that neighboring pixels and an area have similar gray values. The most typical area-based segmentation method is the regional growth method, where a pixel is compared to its neighboring pixels, and if predefined criteria are met, the pixels belong to the same classification. Compared with the threshold segmentation method, the region growth method can produce more accurate and effective segmentation because of the inclusion of regional criteria (spatial information). In the application of CT image segmentation, based on image segmentation method area (especially region growth), the spatial neighborhood information and region terms are strengthened, and the processing efficiency and robustness attenuation change are useful (and under the mild pathological condition caused by imaging artifact). Region-based segmentation is shown in general Method 7.

In addition to regional growth methods, other region-based segmentation techniques such as watershed transformations, graph cut,random walks and fuzzy connectedness (not known how to translate) are also presented in a large number of literatures.

Watershed segmentation is a mathematical morphology segmentation method based on topological theory, whose basic idea is to think of images as topological landforms on geodetic topography, the gray value of each pixel in the image indicates the altitude of the point, each local minimum and its affected area is called the catchment basin, and the boundary of the basin is formed watershed. The concept and formation of a watershed can be explained by simulating the immersion process. At each local minimum surface, pierce a small hole, then immerse the whole model slowly into the water, with the deepening of immersion, each local minimum of the affected domain slowly outward expansion, at the confluence of two basins to build a dam, that is, the formation of a watershed. But watershed segmentation has a fatal weakness, that is, prone to segmentation, noise and fine texture is very sensitive, so that it often produces serious over-segmentation results. Therefore, the watershed segmentation method is seldom chosen to deal with lung segmentation.

Compared with watershed method, graph cut and random walk method are based on graph segmentation method. Because of its high segmentation precision, it is considered to be the global optimal segmentation. In the graph cut algorithm, boundary and texture information is used to create a minimum energy function to achieve segmentation purposes, and the probability of each pixel is computed by the random walk concept. Although graph cut and random walk method can produce effective and accurate segmentation for CT image segmentation of lung, but the degree or a large number of pathological regions exist, the resulting segmentation effect is not good.

Five well-established region-based segmentation methods is Brie?y summarized in Table 2, along with their most commonly used Criteria for lung segmentation

Region-based lung segmentation is often applied to seed point scheme (Figure 8). In this case, the seed point of a target area that is considered to be the most representative is first identified. Seed points can be selected automatically or manually. Once the seed point is selected, the predefined neighborhood criteria are applied to obtain the desired region. Different methods have different criteria for the determination of lung boundaries. One possible criterion is to grow the area until the lung edge is detected. Another example, area uniformity, can be used for convergent segmentation.

An area-based approach used to describe the contours of the pathological conditions of air and homogeneous content (such as voids).

Region-based lung segmentation often has false negatives and therefore requires further post-treatment. Such as. A failure case that results from applying region-based segmentation.

Some post-processing is summarized as follows: control the parameters based on region segmentation, and remove the inner noise, and smooth the image between image segmentation. In addition, artifacts are manually eliminated before the contour algorithm is performed. Another option is to remove artifacts and lesions from the adjacent areas of the lungs, and it seems more feasible to have a contour algorithm that cuts the area of the lungs from the CT image and defines a new area of interest on the area without artifacts.

the fuzzy connectedness method is compared with Graph-cut, random walk and the Region-growing segmentation methods has better robustness.

shape-based Methods

The prior shape information has been widely used in image segmentation. Shape-based technology requires the use of Atlas-based Methods (Atlas The Atlas Guidance Method (atlas-guided) is used to segment medical images using existing standard templates. ) and model-based methods to determine the lung boundary.

atlas-based Methods

An atlas includes a CT image template and a label for the corresponding thoracic region.

model-based Methods

The model-based method is also used to make use of the prior shape information, in order to better adapt to the changeable shape, based on the model method, the optimization process is used to match the appearance model of statistic shape and lung image. Basically, the expected shape and local gray structure of the target in the image are used to deduce the segmentation process. When the best model match is found, the split ends and the contour of the edges is drawn. The model-based approach is top-down, taking into account both global and local variables of shape and texture, so this approach is effective in dealing with abnormal lung segmentation problems. (The statistics of the model affect the result of the segmentation)

Snakes, Active contours, and level sets

neighboring anatomy–guided Methods

The image is segmented by using the spatial context information of the neighborhood target. The basic idea of segmentation of lungs by neighborhood organization limits the search space of optimal boundary search and automatically removes some false positives.

Machine learning–based Methods

The machine learning approach is to build a system and to learn from the data.

The machine learning method is used to predict the abnormalities of the lungs by extracting feature extracts from the data.

[Literature Reading] Segmentation and Image analysis of abnormal lungs at ct:current approaches, challenges, and future Trends