Automatic computer aided analysis algorithms and system for adrenal tumors on CT images

Abstract

BACKGROUND:

The adrenal tumor will disturb the secreting function of adrenocortical cells, leading to many diseases. Different kinds of adrenal tumors require different therapeutic schedules.

OBJECTIVE:

In the practical diagnosis, it highly relies on the doctor’s experience to judge the tumor type by reading the hundreds of CT images.

METHODS:

This paper proposed an automatic computer aided analysis method for adrenal tumors detection and classification. It consisted of the automatic segmentation algorithms, the feature extraction and the classification algorithms. These algorithms were then integrated into a system and conducted on the graphic interface by using MATLAB Graphic user interface (GUI).

RESULTS:

The accuracy of the automatic computer aided segmentation and classification reached 90% on 436 CT images.

CONCLUSION:

The experiments proved the stability and reliability of this automatic computer aided analytic system.

Keywords

Adrenal tumor sparse representations automatic segmentation tumor classification

1. Introduction

The adrenal glands are important endocrine organs. It consists of two parts: cortex and medulla. The adrenal tumor will lead to serious cardiovascular complications. According to the location of the tumor, the adrenal tumor can be divided into cortex tumor and medulla tumor. Aldosterone-producing adenomas and Cushing adenomas are the most common cortex tumor and pheochromocytomas is the most common medulla tumor. Since different kinds of adrenal tumors require various therapeutic schedules, an accurate diagnosis for the adrenal tumor is of great importance.

In the clinical practice, the diagnosis is based on the hormone test and the imaging detection [1, 2, 3, 4]. In the imaging detection, for its outstanding resolution, fast speed and high detection rate, the multiphase spiral Computed Tomography (CT) imaging is most widely used [5, 6, 7]. There are two phases in the multiphase spiral CT imaging: scanning phase and enhanced phase. Scanning phase is normal CT imaging while enhanced phase is the imaging after the injection of iodinated contrast medium. With the help of contrast medium, CT images in enhanced phase are able to improve the image contrast and show the rich vessels in the tumors.

Figure 1 shows the typical multiphase spiral CT images of aldosterone-producing adenomas, Cushing adenomas and pheochromocytomas. The tumors are in the red dashed boxes. Normally, the aldosterone-producing adenomas have small diameter, smooth boundary and homogeneous internal structure [8]. The density of aldosterone-producing adenomas is low because it contains much lipid. The pheochromocytomas usually have large diameter and high density. Hemorrhage, calcification, cystic degeneration or necrosis often occurs in most pheochromocytomas. These interior degeneration parts contain few vessels, so contrast medium has no influence on them. The diameter and density of Cushing adenomas are between these of aldosterone-producing adenomas and pheochromocytomas. Thus the intensity of Cushing adenomas in multiphase CT images is also between the other two kinds of tumors.

Figure 1.

Multiphase CT images of adrenal tumors. (a)–(c) are the scanning phase of an aldosterone-producing adenoma, a Cushing adenoma and a pheochromocytoma. (d)–(f) are their enhanced phase.

The diagnosis of adrenal tumors is based on these characteristics. Usually, multiphase spiral CT images contain hundreds of slices and the abnormalities occupy only several slices. In order to discriminate the adrenal tumor in one patient, clinicians have to review all slices. It is time-consuming and the result highly relies on the doctor’s experience. Less experienced doctors may not find out atypical tumors or delineate precise contours. Such kind of the diagnosis is often inefficient and has poor repeatability. Thus proposing an automatic computer aided analysis system for adrenal tumors can provide the accurate and reliable diagnostic basis for cure and operation scheme, greatly improving the efficiency and accuracy.

An automatic computer aided analysis system is composed of three parts: tumor segmentation, feature selection and classification, in which the tumor segmentation is the first and key step. Segmentation of the adrenal tumor must overcome the following two difficulties. The first is the inhomogeneous intensity in the tumor. The second is the indistinct contour of the tumor.

The segmentation algorithms for CT images can be roughly divided into four groups: region-based, edge-based, statistics-based and deformable models. The region-based method is sensitive to noise and the tumor is easy to be over-segmented [9, 10]. The edge-based method, generally utilizing gradient information, usually needs image filter and cannot handle weak edge [11]. Statistics-based model, which makes use of Markov random field, only takes the pixel information of neighborhoods into consideration. Although it is not sensitive to noise, but it is time consuming and sometimes arrives at the local minimum point [12]. Among all segmentation methods based on deformable model, the level set method has proven effectiveness in CT images segmentation. By adjusting or modifying the energy function, the final segmentation results are usually acceptable [13, 14, 15]. As to the inhomogeneous intensity and indistinct contour of the adrenal tumors, the localized region-based level set method (LRLSM) can handle these two difficulties very well [16, 17, 18, 19], but it highly depends on the manually delineated initial contour, which is still a heavy burden for doctors and sometimes leads to not very reliable segmentation results.

In recent years, lots of research focused on the sparse representation. Compared with traditional signal representation, the sparse representation is more flexible and concise. It has proven to be a powerful tool in image processing, including super-resolution construction, inverse problem and compressive sensing [20, 21, 22, 23]. In these issues, images are often divided into patches. For a specific overcomplete dictionary, the differences of sparse coefficients between each patch indicate different intrinsic structures. This can be utilized in the initial contour extraction of level set method.

For the adrenal tumor classification, normally the features of the tumor are extracted and selected first. Then these features are imported into classifiers. Shape features, gray features and textural features are frequently used in the tumor classification. For the classifiers, Support Vector Machine (SVM) and Radial Basis Function (RBF) network are widely used for their outstanding differentiation and flexibility [24, 25].

In this paper, a novel automatic computer aided analysis method of adrenal tumor on CT images is proposed, including the automatic segmentation based on the multiscale sparse representations and the classification of three kinds of adrenal tumor. In the rest of this paper, Section 2 describes the automatic segmentation and classification algorithms. Then the experimental results are presented in Section 3. Discussions and Conclusions are drawn in Section 4.

2. Materials and methods

2.1 Automatic segmentation based on the multiscale sparse representations

2.1.1 Training dictionary sensitive to patches with edges

Figure 2.

Sparse representation.

Sparse representation problem can be represented as follow according to Fig. 2.

$\displaystyle{\min\limits_{\bm{D},\bm{X}}}\left\|{\bm{Y}-\bm{DX}}\right\|_{2}^% {2}\quad\quad s.t.\quad\left\|{{\bm{x}}_{i}}\right\|_{0}\leqslant T_{0}\quad\forall i$ (1)

where Y is the set of training patches, D is the overcomplete dictionary, X is the set of sparse representation vectors x ${}_{i}$ corresponding to Y, and $T_{0}$ is the sparsity. $\parallel\cdot\parallel_{2}$ and $\parallel\cdot\parallel_{0}$ are the $l^{2}$ -norms and $l^{0}$ -norms respectively. In the constraint condition $\left\|{\bm{x}_{i}}\right\|_{0}\leqslant T_{0}$ , an appropriate solution to D and X need to be found to minimize $\left\|{\bm{Y}-\bm{DX}}\right\|_{2}^{2}$ .

There are generally two kinds of dictionaries: the fixed dictionary and the learned dictionary. The learned dictionary has better flexibility and adaptability if the training method and training set is appropriate. In this paper, the K-means Singular Value Decomposition (KSVD) [26] algorithm is employed. The KSVD algorithm is an iterative algorithm. Each iteration is divided into two stages: the sparse coding stage and the codebook updating stage. In the sparse coding stage, Orthogonal Matching Pursuit (OMP) algorithm [27] is used to compute the representation set X corresponding to the training set Y. In the codebook updating stage, the singular value decomposition (SVD) is used to update the atoms in the dictionary D. The iterations end until Eq. (1) reaches the condition of convergence.

As mentioned in Section 1, sparse representation can help obtain the initial contour of the tumor by recognizing whether image patches contain tumor contour. For a specific dictionary, the sparse coefficients of two kinds of patches are quite different and can be easily classified. By training the dictionary with image patches containing tumor contour like patches 1–3 in Fig. 3a, a dictionary sensitive to patches with edges is obtained.

Figure 3.

Sparse representation coefficients of six example patches. (a) Example patches 1–6 in green boxes. The contour of tumor is in red. Patches 1–3 contain the tumor contour. Patches 4–6 do not contain the tumor contour. (b) The red, blue and green lines are the sparse coefficients of patches 1–3 respectively. (c) The red, blue and green lines are the sparse coefficients of patches 4–6 respectively.

Figure 4.

Tumor’s ROI and its binary map. (a) The ROI where the tumor contour is in red. (b) The binary map of the ROI.

According to Fig. 3, the sparse coefficients of two patches are quite different. The first sparse coeffi- cient’s value of the former one is much larger than that of the latter.

Moreover, by using different scales of dictionary, more image information can be obtained. Small scale patches have high resolution and lager scale patches have more texture information [28, 29]. Thus 3 scales of dictionary are used to extract the initial contour automatically in this paper.

2.1.2 Initial contour

After the region of interest (ROI) is chosen, it is divided into overlapped patches. By using three scales of dictionary which are trained in Section 2.1.1, the sparse coefficients of each patch are calculated. With a simple threshold for the first sparse coefficient, patches can be sorted into two parts. By replacing the center pixel value with 1 for patches containing the tumor contour and 0 for patches containing no tumor contour, the ROI is mapped into a binary image. Figure 4 is the example of a binary map. Then, after some simple morphological operations and region grow method, a precise initial contour of the adrenal tumor is acquired, which is benefit for further level set based segmentation.

2.1.3 Localized region-based level set method

The initial contour of the adrenal tumor is applied to the LRLSM, which aims to minimize the local energy of the contour. The local energy function is described in Eq. (2):

$\displaystyle F[u_{x},v_{x},\Phi(y)]=[I(y)-u_{x}]^{2}H[\Phi(y)]+[I(y)-v_{x}]^{% 2}\{1-H[\Phi(y)]\}$ (2)

Where $I(y)$ is the intensity of the pixel $y$ , $u_{x}$ and $v_{x}$ are mean intensities of local pixels inside and outside the contour, respectively.

Figure 5.

Minimum of the energy function. (a) The adrenal tumor in red. (b) Hypothetic right segmentation result in red. (c) Hypothetic wrong segmentation result in red.

The local region is defined as follows:

$\displaystyle\Gamma(x,y)=\left\{{\begin{array}[]{ll}1,&\left\|{x-y}\right\|<r% \\ 0,&\textit{otherwise.}\\ \end{array}}\right.$ (3)

where $x$ is the pixel on the contour and $y$ is the other pixel, $\left\|{x-y}\right\|$ is the distance between the pixel $x$ and $y$ , $r$ is the radius of the local region.

$\Phi(y)$ is a signed distance function according to the distance of the pixel $y$ in the ROI to the initial contour. The pixels inside the contour can be specified by the following approximation of the smoothed Heaviside function:

$\displaystyle H[\Phi(y)]=\left\{{\begin{array}[]{ll}1,&\Phi(y)<-\varepsilon\\ 0,&\Phi(y)>\varepsilon\\ \frac{1}{2}\left[{1+\frac{\Phi(y)}{\varepsilon}+\frac{1}{\pi}\sin\left({\frac{% \pi\Phi(y)}{\varepsilon}}\right)}\right],&\textit{otherwise.}\\ \end{array}}\right.$ (4)

where the positive constant $\varepsilon$ is set to 1.2 according to the clinical experience.

Obviously, pixels outside the contour can be defined as $1-H[\Phi(y)]$ .

As shown in Fig. 5a, some black regions are between the tumor and other tissues. The red line is the tumor contour, and the yellow dashed boxes are the example patches where there is a black region between the tumor and other tissues. Figure 5b and c is the simplified module of such patches, and the red lines in b and c are the tumor contour which are level 0 during the evolution of the LRLSM.

Assuming the gray part in Fig. 5b and c is the tumor with the intensity as 1, the black part is the background with the intensity as 0, the white part is neighbor tissue and its intensity is 4. The red line is level 0. When the contour is evolved to Fig. 5b, the energy function’ calculation is 4, and when the contour is evolved to Fig. 5c, the energy function’ calculation is 0.5. Thus the right initial contour will evolve to the wrong result.

To solve this problem, the energy function is refined as follows:

$\displaystyle F_{\textit{new}}[u_{x},v_{x},\Phi(y)]=tk[I(y)-u_{x}]^{2}H[\Phi(y% )]+t(1-k)[I(y)-v_{x}]^{2}\{1-H[\Phi(y)]\}$ (5) $\displaystyle\quad+\,(1-t)[I(y)-u_{x}]^{2}H[\Phi(y)]+(1-t)[I(y)-v_{x}]^{2}\{1-% H[\Phi(y)]\}$

where $k$ is a penalty coefficient to ensure the internal homogeneity of the segmentation result and $t$ is the recognition index of CT images’ the background.

$\displaystyle t=\left\{{\begin{array}[]{ll}1,&v_{x}<105\\ 0,&\textit{otherwise.}\\ \end{array}}\right.$ (6)

Thus the global energy of the local region is:

$\displaystyle E_{L}[\Phi(x)]=\int\limits_{\Omega_{x}}\delta[\Phi(x)]\int% \limits_{\Omega_{y}}\Gamma(x,y)F_{\textit{new}}[u_{x},v_{x},\Phi(y)]\textit{dydx}$ (7)

where $\delta[\Phi(x)]$ is the derivative of $H[\Phi(x)]$ :

$\displaystyle\delta[\Phi(x)]=\left\{{\begin{array}[]{ll}1,&\Phi(x)=0\\ 0,&\left|{\Phi(x)}\right|<\varepsilon\\ \frac{1}{2\varepsilon}\left[{1+\cos\left({\frac{\pi\Phi(x)}{\varepsilon}}% \right)}\right],&\textit{otherwise.}\\ \end{array}}\right.$ (8)

Here $\delta[\Phi(x)]$ ensures the selection of every point $x$ near the initial contour.

By taking the length energy term with a parameter $\lambda$ into consideration, the final formulation of the LRLSM is solved as follows:

$\displaystyle\frac{\partial[\Phi(x)]}{\partial t}=t\delta[\Phi(x)]\int\limits_% {\Omega_{y}}\Gamma(x,y)\delta[\Phi(x)]\{k[I(y)-u_{x}]^{2}-(1-k)[I(y)-v_{x}]^{2% }\}dy$ $\displaystyle\quad(1-t)\delta[\Phi(x)]\int\limits_{\Omega_{y}}{\Gamma(x,y)% \delta[\Phi(x)]\{[I(y)-u_{x}]^{2}-[I(y)-v_{x}]^{2}\}dy}$ (9) $\displaystyle\quad+\,\lambda\delta[\Phi(x)]\textit{div}\left({\frac{\nabla\Phi% (x)}{\left|{\nabla\Phi(x)}\right|}}\right)$

where $\nabla$ is the gradient operator and div( $\cdot$ ) is divergence.

Through the initial contour in Section 2.1.2 and Eq. (2.1.3), the final segmentation result is obtained.

2.2 Feature extraction and classification

Based on the characteristics between the three kinds of adrenal tumor introduced in Section 1, three shape features, two gray features and three textural features of the adrenal tumors are extracted.

2.2.1 Shape features

Shape features describe the size and shape of the tumor which are the main diagnostic basics. The most aldosterone-producing adenomas are elliptical and most pheochromocytomas and Cushing adenomas are more close to circle. As for the tumor size, normally, the aldosterone-producing adenomas have small diameter. The pheochromocytomas are the largest. The sizes of Cushing adenomas are in median. Thus shape features are significant for classification. Three shape features are chosen in this paper.

1) The largest tumor diameter $F_{1}$ [30]

$\displaystyle F_{1}={\max\limits_{s,g\in C}}\left\|{s-g}\right\|$ (10)

where $C$ is the pixel set of the segmentation result; $s$ and $g$ are the pixel coordinates; $\left\|{s-g}\right\|$ is the distance between $s$ and $g$ .

2) The tumor area $F_{2}$ [30]

$\displaystyle F_{2}=\left|{SE}\right|$ (11)

where $|SE|$ is the number of pixels in the segmentation result.

Figure 6.

Mean radius of the tumor.

3) The variance of the mean radiuses of the tumor $F_{3}$

As shown in Fig. 6, the intersection of the yellow lines is the barycenter of the tumor. The tumor is divided into 16 parts equally. The mean radius of each part is calculated:

$\displaystyle r_{i}=\frac{1}{\left|{D_{i}}\right|}\sum\limits_{s\in D_{i}}{% \left\|{s-m}\right\|}$ (12)

where $D_{i}$ is the divided blocks; $m$ is the coordinates of the barycenter; $s$ is the coordinates of the pixel on the tumor contour.

The variance of the mean radius is:

$\displaystyle F_{3}=\frac{1}{16}\sum\limits_{i=1}^{16}{r_{i}^{2}}-\left({\frac% {1}{16}\sum\limits_{i=1}^{16}{r_{i}}}\right)^{2}.$ (13)

All shape features were extracted from CT images with the same resolution. If different images are obtained with different resolutions, shape features can be scaled to the same criterion.

2.2.2 Gray features

Gray features describe the intensity of the tumor in CT images. The three kinds of tumor have similar intensity in scanning phase while their intensity varies in enhanced phase. Pheochromocytomas are brightened observably and inhomogeneously. Cushing adenomas are often brightened homogeneously. Aldosterone-producing adenomas are only brightened a little. Two gray features are chosen in this paper.

1) Mean gray level $F_{4}$

$\displaystyle F_{4}=\frac{1}{\left|{SE}\right|}\sum\limits_{g_{i}\in SE}{g_{i}}$ (14)

where $g_{i}$ is the gray level of the pixels in the tumor.

2) Gray level variance $F_{5}$

$\displaystyle F_{5}=\frac{1}{\left|{SE}\right|}\sum\limits_{g_{i}\in SE}{g_{i}% ^{2}}-\left({\frac{1}{\left|{SE}\right|}\sum\limits_{g_{i}\in SE}{g_{i}}}% \right)^{2}.$ (15)

The smaller $F_{5}$ is, the smoother the tumor is.

2.2.3 Texture features

Texture features describe the internal homogeneity, flatness and the characteristics between neighbor pixels of the image. Gray Level Concurrence Matrix (GLCM) is widely used in the describing the texture features of CT images. Defining GLCM as $P(i,j,d,\theta)$ , where $i$ , $j$ is gray level, $d$ is the distance, $\theta$ is the direction. Usually $d$ is 1and $\theta$ is 0.

1) Energy of GLCM $F_{6}$

$\displaystyle F_{6}=\sum\limits_{i=1}^{N_{g}}{\sum\limits_{j=1}^{N_{g}}{[P(i,j% ,d,\theta)}}]^{2}.$ (16)

The energy of GLCM describes the homogeneity of the internal intensity and the thickness of the texture.

2) Contrast of GLCM $F_{7}$

$\displaystyle F_{7}=\sum\limits_{i=1}^{N_{g}}{\sum\limits_{j=1}^{N_{g}}{\left|% {i-j}\right|^{2}P(i,j,d,\theta)}}.$ (17)

The contrast of GLCM describes the definition and the shade of the texture.

3) Autocorrelation of GLCM $F_{8}$

$\displaystyle F_{8}=\sum\limits_{i=1}^{N_{g}}{\sum\limits_{j=1}^{N_{g}}{ijP(i,% j,d,\theta)}}.$ (18)

The autocorrelation of GLCM describes the similarity of the texture directions.

2.2.4 Tumor classification

Here, two classical machine learning methods SVM and RBF are employed for tumor classification. The eight extracted features are input to the SVM and RBF network for classification.

The SVM is based on the structural risk minimization principle. Utilizing a nonlinear kernel and the optimal splitting plane, it finds an optimal decision hyperplane by supervised training to classify linear and nonlinear features [31]. The SVM aims at finding an optimal splitting plane ${\bm{w}}^{\text{T}}{\bm{x}}+b=0$ , where the distances of the samples at both side of this plane $\left|{{\bm{w}}^{\text{T}}{\bm{x}}^{p}+b}\right|\geqslant 1$ . Here, the samples which meet the equal condition are named as support vectors.

The RBF network consists of the input layer, the hidden layer and the output layer. The input and output layer are linear transformation layers, and the hidden one is a nonlinear transformation layer [32]. The Gaussian kernel $\varphi=\exp\left({-\frac{\left\|{\bm{x}-\bm{t}}\right\|^{2}}{\delta^{2}}}\right)$ is frequently chosen as the nonlinear transformation kernel.

The advantage of SVM is that it can handle the local optimal problem and has excellent performance in solving small samples and high dimension modeling problems. The speed of SVM is usually faster than ANN classifier. However, SVM cannot deal with outliers and noises, i.e., its performance will decrease sharply when the data set either contains outliers or was contaminated by noises [33]. RBF is capable of solving nonlinear problems. It is less sensitive to noises and has better robustness than SVM. However, the selection of the classification centers is complex.

3. Experiments and results

3.1 Experimental materials

We conduct the experiments on a clinical adrenal tumor dataset, including 86 cases of Aldosterone-producing adenomas, 121 cases of pheochromocytomas and 11 cases of Cushing adenomas. Each case contains a scanning phase image and an enhanced phase image. CT images are transformed into 8-bit grayscale images where the Window width is 200 Hounsfield Unit (HU), and the window level is 40 HU. The segmentation golden standard is the experienced doctor’s manually delineated results. Each case is diagnosed in histopathology, regarding as the classification golden standard.

3.2 Automatic segmentation

In the dictionary training stage, the training set consisted of 900 patches containing the tumor contours from CT images. The patch size was 3 $\times$ 3, 5 $\times$ 5 and 7 $\times$ 7. The dictionary size was 9 $\times$ 36, 25 $\times$ 100 and 49 $\times$ 196. The threshold of the sparse representation coefficients to each dictionary is 0.12, 0.30 and 0.38 respectively.

The segmentation quality metrics is the accuracy ( $A C$ ) and the mean absolute distance (MAD):

$\displaystyle AC=\frac{TP+(1-FP)}{2}\times 100\text{\%}$ (19)

where $T P$ is the true positive and $F P$ is the false positive:

$\displaystyle TP=\frac{{\left|{SE\cap GS}\right|}}{{\left|{GS}\right|}}\times 1% 00\text{\%}$ (20) $\displaystyle FP=\frac{{\left|{SE-SE\cap GS}\right|}}{{\left|{GS}\right|}}% \times 100\text{\%}$ (21)

where $G S$ and $S E$ are the pixel sets of the golden standard and the segmentation result, respectively. $|\cdot|$ is the number of pixels in the set. $A C$ is a symmetric similarity index. It is 1 when the tumor is perfectly segmented and 0 for no overlap.

Let $A$ and $M$ be the contour of the segmentation result and the golden standard, respectively:

$\displaystyle A=\{a_{i}|i=1\cdots K\}$ (22) $\displaystyle M=\{m_{n}|n=1\cdots N\}$ (23) $\displaystyle\textit{MAD}=\frac{1}{K}\sum\limits_{i=1}^{K}{d(a_{i},M)}$ (24)

Where $d(a_{i},M)$ is the distance between the pixel $a_{i}$ and the contour $M$ :

$\displaystyle d(a_{i},M)={\min\limits_{n}}\left|{a_{i}-m_{n}}\right|$ (25)

MAD is the mean distance between two contours. It is 0 when the tumor is perfectly segmented.

As shown in Fig. 7, the modified new energy function restrains the level set method from evolving to neighbor tissue, thus the segmentation result is closer to the golden standard. Table 1 is the statistical results of $A C$ and MAD of 436 CT images. Compared with the original energy function, the modified one brings more than 1% improvement of the $A C$ , and the MAD decreases to 0.89, demonstrating the new energy function is more suitable for the segmentation of adrenal tumor.

Table 1

Average evaluation indexes of both methods in the segmentation of 436 CT images

	$A C$ $\pm$ std (%)	MAD $\pm$ std
Original energy function	91.33 $\pm$ 2.91	0.94 $\pm$ 0.28
New energy function	92.45 $\pm$ 2.55	0.89 $\pm$ 0.21

Figure 7.

(a) The golden standard. (b) The automatically extracted initial contour. (c) The segmentation result with original energy function of LRLSM. (d) The segmentation result with modified energy function of LRLSM.

3.3 Feature extraction

Table 2 is the mean and variance of eight features in three kinds of tumor at scanning phase. From the table we can see, the pheochromocytomas’ size is much bigger than aldosterone-producing adenomas and slightly bigger than Cushing adenomas. The pheochromocytomas and Cushing adenomas are very close to circle in shape. So in the shape features, the pheochromocytomas and Cushing adenomas are similar. They are quite different from aldosterone-producing adenomas. As to the gray features, all tumors’ intensity at scanning phase is low. Pheochromocytomas’ intensity and the intensity variance are slightly higher than the other two. It is hard to distinguish three kinds of tumor by the intensity of images. As to the texture features, Cushing adenomas and aldosterone-producing adenomas are very similar, the differences of three feature are all less than 0.01. From the above, aldosterone-producing adenomas and pheochromocytomas are easy to differentiate. Cushing adenomas are close to aldosterone-producing adenomas in intensity and texture and close to pheochromocytomas in size. The results are in accord with the clinical observation.

Table 2
The mean and variance of eight features in three kinds of tumor at scanning phase

	Tumor	Tumor	Radius	Mean	Intensity	GLCM	GLCM	GLCM
	diameter	area	variance	intensity	variance	energy	contrast	autocorrelation
Aldosterone-producing	10.12 $\pm$	83.54 $\pm$	10.08 $\pm$	85.11 $\pm$	6.44 $\pm$	0.07 $\pm$	2.04 $\pm$	0.24 $\pm$
adenomas	1.23	3.21	1.21	2.32	0.88	0.01	0.11	0.01
Pheochromocytomas	25.26 $\pm$	302.39 $\pm$	7.45 $\pm$	89.37 $\pm$	7.50 $\pm$	0.11 $\pm$	1.15 $\pm$	0.51 $\pm$
	3.24	27.83	0.81	2.11	1.51	0.01	0.04	0.02
Cushing adenomas	20.24 $\pm$	270.04 $\pm$	7.14 $\pm$	86.48 $\pm$	6.79 $\pm$	0.08 $\pm$	2.03 $\pm$	0.23 $\pm$
	1.76	3.53	0.77	1.92	0.78	0.02	0.08	0.01

Table 3

The mean and variance of eight features in three kinds of tumor at enhanced phase

	Tumor	Tumor	Radius	Mean	Intensity	GLCM	GLCM	GLCM
	diameter	area	variance	intensity	variance	energy	contrast	autocorrelation
Aldosterone-producing	11.81 $\pm$	90.08 $\pm$	10.52 $\pm$	95.47 $\pm$	7.89 $\pm$	0.08 $\pm$	1.71 $\pm$	0.31 $\pm$
adenomas	1.14	3.46	1.29	3.18	0.75	0.01	0.10	0.02
Pheochromocytomas	27.18 $\pm$	288.11 $\pm$	7.91 $\pm$	209.51 $\pm$	17.22 $\pm$	0.12 $\pm$	0.82 $\pm$	0.52 $\pm$
	2.98	30.67	0.77	11.32	5.31	0.01	0.02	0.01
Cushing adenomas	19.34 $\pm$	251.24 $\pm$	6.03 $\pm$	164.15 $\pm$	9.77 $\pm$	0.07 $\pm$	1.87 $\pm$	0.23 $\pm$
	1.56	3.32	0.67	7.74	2.44	0.01	0.07	0.02

Table 3 is the mean and variance of eight features in three kinds of tumor at enhanced phase. Compared to Table 2, tumors don’t have much change in shape features. For the gray features, pheochromocytomas increase more than 100. Cushing adenomas also increase a lot, and aldosterone-producing adenomas have a little enhancement. For the texture features, it is obvious that the contrast and autocorrelation of the GLCM become easier to distinguish than that in scanning phase, especially between the aldosterone-producing adenomas and Cushing adenomas.

Comparing Tables 2 and 3, texture feature have high discrimination at enhanced phase. Also the differences of the gray feature at scanning phase and enhanced phase are more discriminative than these in each phase. Thus, reserving the shape and texture features at enhanced phase, replacing the gray features by the differences at two phases. The new feature sets are shown in Table 4.

Table 4

The mean and variance of new feature sets

	Tumor	Tumor	Radius	Mean	Intensity	GLCM	GLCM	GLCM
	diameter	area	variance	intensity	variance	energy	contrast	autocorrelation
Aldosterone-producing	11.81 $\pm$	90.08 $\pm$	10.52 $\pm$	10.36 $\pm$	1.45 $\pm$	0.08 $\pm$	1.71 $\pm$	0.31 $\pm$
adenomas	1.14	3.46	1.29	3.22	0.77	0.01	0.10	0.02
Pheochromocytomas	27.18 $\pm$	288.11 $\pm$	7.91 $\pm$	120.14 $\pm$	9.72 $\pm$	0.12 $\pm$	0.82 $\pm$	0.52 $\pm$
	2.98	30.67	0.77	6.24	2.07	0.01	0.02	0.01
Cushing adenomas	19.34 $\pm$	251.24 $\pm$	6.03 $\pm$	77.67 $\pm$	2.98 $\pm$	0.07 $\pm$	1.87 $\pm$	0.23 $\pm$
	1.56	3.32	0.67	4.20	1.42	0.01	0.07	0.02

From Table 4, the most discriminative features are tumor diameter, tumor area and mean intensity, which is anastomotic with the clinical diagnosis.

3.4 Classification

The new feature set of three kinds of tumor are imported into SVM and RBF network for classification. Here, SVM is used as a two class classifier and RBF network are used as a three class classifier. The training and testing use leave-one-out method. Tables 5 and 6 are the classification results of the SVM and RBF network respectively.

According to Table 5, the classification accuracies of SVM are almost higher than 90%, indicating the distinction of extracted features. By Table 6, RBF network also presents good classification results, only the Cushing adenomas is a little lower, which is 81.82%. The reason is the Cushing adenomas are similar in size and intensity enhancement with the other two tumors. Also the training samples of Cushing adenomas are much smaller, affecting the training stage of the RBF network. Taking one with another, RBF network can handle the classification of the three kinds of tumor and present valuable suggestion for clinic diagnosis.

Table 5
SVM classification results

Accuracy	Aldosterone-producing	Pheochromocytomas (%)	Cushing adenomas (%)
	adenomas (%)
Aldosterone-producing adenomas and	95.35	95.04	$\backslash$
pheochromocytomas
Aldosterone-producing adenomas and	93.02	$\backslash$	90.91
Cushing adenomas
Pheochromocytomas and Cushing	$\backslash$	92.56	81.82
adenomas

Table 6

RBF network classification results

	Aldosterone-producing adenomas (%)	Pheochromocytomas (%)	Cushing adenomas (%)
Accuracy	89.53	90.08	81.82

3.5 MATLAB GUI interface

All the above methods are incorporated into a computer aided system. By using MATLAB GUI, the tumors segmentation and classification algorithms are integrated on the graphic interface to improve the efficiency of the clinical diagnosis.

Figure 8 is the interface of the automatic computer system. The segmentation result is shown on the CT image. The feature extraction and classification results are shown in the upper left corner.

Figure 8.

The interface of the automatic computer system. (a) The initial interface. (b) The result interface.

4. Conclusions

This paper proposed a novel automatic computer aided analysis algorithms on CT images to detect and classify the adrenal tumors. To overcome the problem that the adrenal tumor are often have adhesion with the adjacent tissue, weak edges and inhomogeneous interior, a multi-scale sparse representation and patch processing method was deployed to automatically obtain the initial contour. Then, the initial contour was applied in the modified LRLSM to finish the automatic segmentation. The experiments had proved that this automatic segmentation algorithm was efficient and accurate and appropriate for the adrenal tumor’s segmentation. Three shape features, two gray-scale features and three textural features (totally eight features) were extracted according to the differences of pheochromocytomas, aldosterone-producing adenomas and Cushing adenomas in CT images. The features at scanning phase and enhanced phase were combined to obtain more discriminative new feature sets. The new feature sets were input to SVM and RBF Network for classification. The experiments have proved that these two classifiers can identify different tumors by the input features and the accuracy is about 90%, indicating the stability and reliability. Finally, all algorithms were integrated into a system and conducted a graphic interface by using MATLAB GUI to improve the efficiency of the clinical diagnosis. Our system is promising in the practical diagnosis of adrenal tumor.

In the future, we will focus on the recognition of the benign and malignant adrenal tumor. We shall also integrate the recognition algorithms into the existing analysis system.

Conflict of interest

None to report.

Footnotes

Acknowledgments

This work is supported by the National Natural Science Foundation of China (61401102), and the Science and Technology Commission of Shanghai Municipality (Grant 14YF1400300).

References

Young

, Jr. The incidentally discovered adrenal mass. New England Journal of Medicine, 2007, 356(6): 601-610.

Szolar

Korobkin

Reittner

, et al. Adrenocortical carcinomas and adrenal pheochromocytomas: mass and enhancement loss evaluation at delayed contrast-enhanced CT. Radiology, 2005, 234(2): 479-485.

Grumbach

Biller

Braunstein

, et al. Management of the clinically inapparent adrenal mass (“incidentaloma”). Annals of Internal Medicine, 2003, 138(5): 424-429.

Sun

Zhou

Liu

, et al. The evalution of imaging diagnosis of adrenal tumors (report of 385 cases). Chinese Journal of Urology, 2004, 25(4): 224-226.

Johnson

Horton

Fishman

. Adrenal mass imaging with multidetector CT: pathologic conditions, pearls, and pitfalls. Radiographics, 2009, 29(5): 1333-1351.

Zhang

Wang

. Clinical study of MSCT and MPR on diagnosis of hyperfunctional adrenal gland neoplasm. Journal of China Medical University, 2008, 37(3): 411-413.

Sun

Liao

. Study of 64-slice dynamic contrast-enhanced CT in common adrenal tumors. Journal of China Clinic Medical Imaging, 2007, 18(10): 693-695.

Tsirlin

Shama

, et al. Pheochromocytoma: a review. Maturitas, 2014, 77(3): 229-238.

Liu

. Liver CT image segmentation by local entropy method. Computer Application and System Modeling (ICCASM), 2010 International Conference on. IEEE, 2010: pp. V11-591-V11-594.

10.

Liu

Wang

Zhang

. Liver cancer CT image segmentation methods based on watershed algorithm. Computational Intelligence and Software Engineering, 2009. CiSE 2009. International Conference on. IEEE, 2010: pp. 1-4.

11.

Casti

Mencattini

Salmeri

, et al. Estimation of the breast skin-line in mammograms using multidirectional Gabor filters – Computers in Biology and Medicine. Computers in Biology & Medicine, 2013, 43(11): 1870-1881.

12.

Hanzouli

Lapuyade-Lahorgue

Monfrini

, et al. PET/CT image denoising and segmentation based on a multi observation and a multi scale Markov tree model. Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC), 2013 IEEE, 2013: pp. 1-4.

13.

Farag

Munim

HEAE

Graham

, et al. A novel approach for lung nodules segmentation in chest CT using level sets. IEEE Transactions on Image Processing, 2013, 22(12): 5202-5213.

14.

Zidan

Ghali

Hassamen

, et al. Level set-based CT liver image segmentation with watershed and artificial neural networks. Hybrid Intelligent Systems (HIS), 2012 12th International Conference on. IEEE, 2012: pp. 96-102.

15.

Huang

Dione

Compas

, et al. Contour tracking in echocardiographic sequences via sparse representation and dictionary learning. Medical Image Analysis, 2014, 18(2): 253-71.

16.

Kao

Gore

, et al. Implicit active contours driven by local binary fitting energy. Proceeding of IEEE Conference on Computer Vision and Pattern Recognition, 2007: pp. 1-7.

17.

Lankton

Tannenbaum

. Localizing region-based active contours. IEEE Transactions on Image Processing, 2008, 17(11): 2029-2039.

18.

Zhang

Song

Zhang

. Active contours driven by local image fitting energy. Pattern Recognition, 2010, 43(4): 1199-1206.

19.

Tang

Guo

Wang

, et al. Adaptive Co-segmentation of pheochromocytomas in CECT images using localized level set models. IEEE Journal of Biomedical and Health Informatics, 2015, 20(2): 549-562.

20.

Yang

Wright

Huang

, et al. Image super-resolution via sparse representation. IEEE Transactions on Image Processing, 2010, 19(11): 2861-2873.

21.

Sapiro

Mallat

. Solving inverse problems with piecewise linear estimators from Gaussian mixture models to structured sparsity. IEEE Transactions on Image Processing, 2012, 21(5): 2481-2499.

22.

Wang

Bresler

Ntziachristos

. Compressive sensing for biomedical imaging. IEEE Transactions on Medical Imaging, 2011, 30(5): 1013-1016.

23.

Zhang

Wang

, et al. Exponential wavelet iterative shrinkage thresholding algorithm with random shift for compressed sensing magnetic resonance imaging. IEEJ Transactions on Electrical & Electronic Engineering, 2015, 10(1): 116-117.

24.

Furey

Cristianini

Duffy

, et al. Support vector machine classification and validation of cancer tissue samples using microarray expression data. Bioinformatics, 2000, 16(10): 906-914.

25.

Foody

Mathur

. A relative evaluation of multiclass image classification by support vector machines. IEEE Transactions on Geoscience and Remote Sensing, 2004, 42(6): 1335-1343.

26.

Aharon

Elad

Bruckstein

. K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation. IEEE Transactions on Signal Processing, 2006, 54(11): 4311-4322.

27.

Pati

Rezaiifar

Krishnaprasad

. Orthogonal matching pursuit: recursive function approximation with applications to wavelet decomposition. 27th Asilomar Conference on Signals, System and Computers, 1995: pp. 40-44.

28.

Liao

Gao

Lian

, et al. Sparse patch-based label propagation for accurate prostate localization in CT images. IEEE Transactions on Medical Imaging, 2013, 32(2): 419-434.

29.

Wang

Chen

Ran

. Image super-resolution reconstruction with content based dual-dictionary learning and sparse representation. Chinese Journal of Scientific Instrument, 2013, 34(8): 1690-1695.

30.

Zhang

Wang

, et al. Fruit classification using computer vision and feedforward neural network. Journal of Food Engineering, 2014, 143(6): 167-177.

31.

Chih

. A comparison of methods for multiclass support vector machines. IEEE Transactions on Neural Networks, 2002, 13(2): 415-425.

32.

Meng

. Face recognition with radial basis function (RBF) neural networks. IEEE Transactions on Neural Networks, 2002, 13(3): 697-710.

33.

Wang

Yang

Zhang

, et al. Identification of green, oolong and black teas in China via wavelet packet entropy and fuzzy support vector machine. Entropy, 2015, 17(10): 6663-6682.

Automatic computer aided analysis algorithms and system for adrenal tumors on CT images

Abstract

BACKGROUND:

OBJECTIVE:

METHODS:

RESULTS:

CONCLUSION:

Keywords

1. Introduction

2.1 Automatic segmentation based on the multiscale sparse representations

2.1.1 Training dictionary sensitive to patches with edges

2.1.3 Localized region-based level set method

2.2.1 Shape features

3. Experiments and results

3.1 Experimental materials

3.2 Automatic segmentation

Table 2 The mean and variance of eight features in three kinds of tumor at scanning phase

Table 5 SVM classification results

Conflict of interest

Footnotes

Acknowledgments

References

Table 2
The mean and variance of eight features in three kinds of tumor at scanning phase

Table 5
SVM classification results