A suspected pulmonary nodule detection method was proposed based on dot-filter and extracting centerline algorithm. In this paper, we focus on the distinguishing adhesion pulmonary nodules attached to vessels in two-dimensional (2D) lung computed tomography (CT) images. Firstly, the dot-filter based on Hessian matrix was constructed to enhance the circular area of the pulmonary CT images, which enhanced the circular suspected pulmonary nodule and suppresses the line-like areas. Secondly, to detect the nondistinguishable attached pulmonary nodules by the dot-filter, an algorithm based on extracting centerline was developed to enhance the circle area formed by the end or head of the vessels including the intersection of the lines. 20 sets of CT images were used in the experiments. In addition, 20 true/false nodules extracted were used to test the function of classifier. The experimental results show that the method based on dot-filter and extracting centerline algorithm can detect the attached pulmonary nodules accurately, which is a basis for further studies on the pulmonary nodule detection and diagnose.
Pulmonary nodules are small masses of tissue in the lung, are prevalent findings on chest and abdominal CT scans, and can be cancerous, though most of them are benign [
To date, many researchers all over the world are devoted to the study of the detection of attached pulmonary nodules, for example, nodule attached to vessels and the pulmonary wall. However, limitations occur in lung cancer imaging of distinguishing nodules attached to vessels from the normal blood vessels, which infiltrate the vessels surreptitiously. Using the corrosion morphology and expansion to segment the pulmonary nodules from the vessels resulted in the corrosion of the nodule thorn, which is another important index for malignant nodule valuation [
Pulmonary nodules are similar to spherical objects, and the lung CT images are 2D. In order to enhance the dot-like regions and depress the line-like regions quickly and effectively, an algorithm named dot-filter was proposed by Li et al. [
The process of the algorithms used in this paper was shown in Figure
The process of the algorithms used to recognize adhesion pulmonary nodules.
To a medical CT image, the enhancement filter of local structure was used extensively which is based on the shape of organization. On a 2D image, we used the dot model conforming to Gauss distribution to represent a nodule [
Here,
(a) simulate the model of Gaussian dot; (b) enhanced with one scale; (c) enhanced with multitude scales.
Li et al. [
The enhanced dot-filter is expressed by the following expression [
In CT images, if the semidiameter of one pulmonary nodule is
The steps of extracting dot with numbers of scales of dot-filter are as follows: According to the range of scale of the nodules we compute the value of For every Using Gaussian function convolve with 2D For every pixel, repeat Compute Compute Stop computing. Select the maximum of
In order to prove better the effect of using dot-filter with variety value of
Figure
As depicted above, we know that dot-filter can enhance the dot-like areas effectively. However, in the lung CT images, the ends and cross sections of vessels are also of dot-like shapes, which will be enhanced by using dot-filter, leading to many more false positives appearance. In order to prove that, we construct three types’ vessel models, such as single line model,
Models of vessels constructed. The regions marked by red in (a), (d), and (g) are dot-like regions, which will be enhanced by dot-filter. (a) the vessel of single line type; (b) smoothed by Gauss function; (c) enhanced by dot-filter; (d) the vessel of
Now we will use the Dot-Filter constructed above based on Hessian matrix to lung CT images, and the result is shown in Figure
Detection of the solitary nodules in a 2D CT image enhanced by dot-Filter. (a) original lung image; (b) lung segment extraction; (c) the right lung segment; (d) lung CT image smoothed by Gauss function; (e) detection of the solitary nodules; (f) lung CT image without nodules; (g) pulmonary segment extracted; (h) smoothed by Gauss function; (i) enhanced by dot-filter.
Figure
There are many algorithms used to extract the central line [
Different from the traditional algorithms of extracting central line, area skeleton can be defined by mean axle transforming (MAT). Describe an area whose profile is
We here give the mean of a refinement of two-value algorithm region: we suppose the value of the pixel in the region is 1, and the values of the pixels on the background are 0. The value of the pixels in the edge of the region is 1 and at least there is one pixel of which the value is 0. As 8 neighborhoods shown in Figure
(a) The arrangement of the neighborhood pixels used in the thinning algorithm; (b) the explanation of expression (
In Step 2, (a) and (b) remain unchanged, and (c) and (d) become
We apply step 1 to every pixel in the edge of the two-value region. If we violate (a) or (b), the value of the pixel we talk about is unchanged. Otherwise, we take it as the pixel that will be removed after we handle all the pixels of the edge. Then, we use step 2 the same way as step 1 till there is no pixel needed to be removed any more and stop the algorithm.
Take Figure
(a) chromosome image after segment; (b) image after Gaussian-filter; (c) the skeleton; (d) eight times for extinguishing the burr of the skeleton; (e) seven more times for extinguishing the burr.
Figure
If a line is expressed by
The difference between four algorithms for extracting center line in deviation and time consuming.
Algorithm for center-line extracted |
|
|
|
---|---|---|---|
Margin of linear least square fitting legitimate | 0.142 | 45.437 | 0.042 |
Symmetric moment fitting center method | 1.50 | 550.832 | 0.031 |
Block cancroids least squares fitting | 0.671 | 332.117 | 0.033 |
Algorithm used in this paper | 0.157 | 23.858 | 0.059 |
According to Table
In this paper, to accomplish the experiment combining dot-filter with the method proposed above we use six steps shown as follows. We first selected three lung CT images after extracting the lung segment in Figures As the value of vessels, tracheas, and nodules are bigger than the value of lung parenchyma, in order to decrease the computation, we extracted the soft tissue of the lung based on gray threshold which we defined as 130, which is obtained after many times of drawing histogram, shown in Figures To the tissues extracted in step We eliminated the tissues obtained in step Firstly we worked out the center of every suspected nodule (false positive) and then computed the value of The diameter of the nodules is 3 mm~30 mm, so we compared the diameter with
(a)~(c) original image; (d)~(f) soft tissue extraction of lung; (g)~(i) images after dot-filter; (j)~(l) images after seven times for extracting the skeleton; (m)~(o) attached nodule extraction; (p)~(r) is the three-dimension display of the objects.
Figures
Table
Databases of testing the method used in this paper.
Name | LIDC database | Supported by Jida Hospital |
---|---|---|
Number of CT images | 10 | 10 |
Pixel unit (volume)/mm3 | 0.6 × 0.6 × 0.6 | 0.6 × 0.6 × 0.6 |
Average number | 36 | 70 |
Number of adhesion nodules | 3 | 13 |
Image size/pixels | 512 × 512 | 512 × 512 |
Layer thickness | 1 mm | 1 mm |
For the CT images supported by the hospital it missed 3 adhesion nodules and missed none for the LIDC database. Table
The method used in this paper compared with current two methods.
Method | Number of false positives per set | Missing rate/(%) | Runtime/(min) |
---|---|---|---|
Literature [ |
8.6 | 33.3 | 2.8 |
Literature [ |
11.2 | 27.5 | 4.2 |
Method with only dot-filter | 34.8 | 18.7 | 1.2 |
Method with dot-filter and centerline extraction | 5.3 | 18.7 | 1.7 |
In this paper we first use 2D Hessian matrix to construct dot-filter constructed to extract dot-like region. In order to solve the problem that the dot-filter cannot detect attached pulmonary nodules, an algorithm based on extracting centerline was used. Results of experiment indicated that the method is easy and effective while extracting attached pulmonary nodules well. In the future, we will be devoted to extracting the lung nodules contacting pulmonary wall and ground glass opacity pulmonary nodules.
The authors declared that they have no conflict of interests regarding this work.