03-05-2012, 10:50 AM
Hemorrhage Slices Detection in Brain CT Images
Hemorrhage Slices Detection in Brain CT Images.doc (Size: 38.5 KB / Downloads: 35)
Abstract
Multi-slice Computer Tomography (CT) scans are widely used in today’s diagnosis of head traumas. It is effective to disclose the bleeding and fractures. In this paper, we present an automated detection of CT scan slices which contain hemorrhages. Our method is robust towards various rotation, displacement and motion blur. Detection of these pathological slices will be useful for further diagnosis and retrieval.
1. Introduction
Multi-slice Computer Tomography (CT) scans are widely used in today’s diagnosis of head traumas. It is effective to disclose the bleeding and fractures. Also, it is cheap and thus affordable to public. A huge amount of CT images are produced in modern hospitals. For example, in Singapore, the Neuroradiology Service at the National Neuroscience Institute (NNI), Tan Tock Seng Hospital, performed over a thousand CT scans, with each scan consisting of 20 slices, in the two-year period of 2003 to 2005 as a result of hospital admission for mild head injured patients [1]. Detection of abnormal slices is useful for further processing such as diagnosis and retrieval. The usefulness can be viewed as the following scenarios: (a) Doctors may confirm their diagnosis with the help of computer-aided segmentation and effective retrieval of other similar cases. (b) Medical students may learn the diagnosis by examples retrieved and enhancements of the hemorrhage segmented [2].
In this paper, we present an automated detection of CT scan slices which contain hemorrhages. We will first discuss some related work. Then we will present the detection method for abnormal brain slices. Next we will show and analyze the result of our detection method. Lastly, we will give our conclusion and discuss the future work.
2. Related work
A content-based 3D neuroradiologic image retrieval system is developed at the Robotics Institute of CMU. In the preliminary results of this system, Y.Liu et al. proposed a symmetric detection approach to detect brain lesions [2][3]. The methods will check the symmetry for each slice. If the mid line of the brain shifts, a hemorrhage is considered to be presented. They also assume the pathological brain absent the symmetry property. Thus if some abnormal region appears only at one side, it is considered as a hemorrhage. A similar method was also adapted by T.Hara et al. [4] In their approach, they detect mid line according to the skull contour. Mid line detection was also adapted by T.Chan [5]. However, all of these studies only consider the slices of encephalic region. In real situation, the slices will also include the nasal cavity and sometimes do not follow the sequential order. For example, some slices rescanned will be appended after all slices and some slices even interleave each other. In addition, the mid line detection approach will not work if the trauma is too severe. For example, some severe patients have such a fatal injury that the mid line does not present clearly as shown in figure 3 in the next section. Moreover, if the patient moves his/her head or does not place his/her head in a correct angle, the scanning angle will not be perfect. Thus the image result will not be symmetric even his/her brain is healthy. In the following sections, we will present a method that detects the encephalic slices and hemorrhages without using mid line estimation.
978-1-4244-2175-6/08/$25.00 ©2008 IEEE
3. The detection method
The detection method consists of two parts. The first part splits the scan slices into encephalic region and nasal cavity region (Figure 1). The second part focuses on the encephalic region and detects the abnormal slices. In both parts, we apply the same method by using wavelet and Haralick texture model.
Figure 1. Left: Nasal Cavity; Right: Encephalic region
3.1. Split slices into encephalic region and nasal cavity region.
From Figure 1 we see that the nasal cavity region is much more complex than the brain region. Thus the effective methods used in the brain region may not work dealing with the nasal cavity, for example, some segmentation methods such as simple background removal. Therefore, we will first separate them.
The CT scans normally are done sequentially, from bottom upwards. However, there are cases of scanning from the top downwards. Moreover, some rescanning of certain parts is required when the previous attempt failed. Thus, the image sequence may be out of order. Hence, we cannot simply assume that the first slice of encephalic region is next to last nasal cavity slice.
By visual inspection of the two regions, we can see that texture is a good criterion to distinguish them.
Wavelet transform has been widely applied in image processing, especially, texture analysis. In this paper, we use Haar wavelet family in favor of its simplicity and efficiency. The transform can be viewed as transforming a signal from the original space to a new space which is spanned by orthogonal wavelets, which are described by the basis function. Once transformed, we will split the signal into different waves with different frequencies. If we treat the image as a 2D signal, then by wavelet transform, we shall get high frequency components, which corresponds the textures and low frequency components, which corresponds the homogenous intensity field. The former is represented by detail coefficients and the latter is stated in the approximate coefficients. Thus, we are able to operate the texture or the homogeneous field, for example, the hemorrhage in the brain CT scan, by focusing on detail or the approximate coefficients.
In this part, we focus on texture analysis. The first step is to decompose the image four times using Haar wavelet transforms and then reconstruct the signal by only using the detail coefficients and obtain a texture map (Figure 2). More details of the wavelet transform may be found in Refs [6] and [7].
Figure 2. Left: Image; Right: Texture map
In the Second step, we calculate the histogram of the intensity distribution and record the arithmetic mean, the standard deviation, the skewness and the kurtosis of the distribution, totally 4 measurements.
Thirdly, we employ a variation of Haralick model texture descriptors [8] and add the 8 values into the feature vector such as entropy, energy, contrast, homogeneity, sum mean, variance, maximum probability and the correlation. The detail calculation may refer to Ref [8].
Thus, the feature vector contains 12 feature values.
To select good features, we employ entropy calculation from information theory. We discard the feature “sum mean” whose entropy is greater than 6, leaving the remaining 11 features for training.
Lastly, we use SVM with a linear kernel to learn the model. The model is also tested and the results will be shown in later section.
3.2. Detect abnormal encephalic slices.
The encephalic region is where the brain locates. The head traumas normally occur in this region. Thus, we will only focus on encephalic slices.
First of all, we define the abnormal brain versus normal brain. As shown in figure 3, the normal brain has a homogenous distribution of gray matter, clear texture on the brain such as fissures, and regular skull-dural outlines. The abnormal brain will have hemorrhage which appears brighter than the gray matter in the image. The hemorrhage occurs in different location will be considered as different types of head traumas. Table 1 and Figure 3 will concisely describe 5 types of hemorrhages we will deal with in this paper such as EDH, SDH, SAH, ICH, and IVH. We define the CT slice which contains hemorrhages as abnormal and CT slices with no hemorrhages as normal. Next we shall present the steps of the method