JP2021020015A - Pulmonary nodule clarification method by two-dimensional cellular automaton - Google Patents

Abstract

To extract a faint pulmonary nodule shadow by using a two-dimensional cellular automaton from a binary error diffusion image for a pulmonary nodule clarification image, and clarify it with an adaptive rank filter.SOLUTION: A pulmonary nodule clarification method includes: a step for creating a pulmonary nodule clarification image from a chest X-ray image; a step for creating a background noise suppression image; a step for creating a binary error diffusion image by applying an error diffusion method to the pulmonary nodule clarification image; a step for creating an image in an initial state on the basis of ranking of a pixel value in a predetermined area acquired by using a rank filter for the binary error diffusion image; a step for extracting a faint pulmonary nodule shadow as a set of binary points by executing state transition by a two-dimensional cellular automaton, and creating a binary point image; a step for creating an adaptive rank pulmonary nodule clarification image by compositing the binary point image, the pulmonary nodule clarification image, and the background noise suppression image; and a step for outputting the adaptive rank pulmonary nodule clarification image.SELECTED DRAWING: Figure 2

Description

The present invention relates to a method for clarifying lung nodules using a two-dimensional cellular automaton.

The mortality rate from lung cancer is still on the rise, and early detection and treatment are important issues based on the results of stage and 5-year survival rate (Non-Patent Document 1). However, lung cancer is not a little overlooked in chest X-ray interpretation in daily clinical practice and medical examination (Non-Patent Documents 2 and 3), and lung nodules are extracted from chest plain X-ray photographs (hereinafter referred to as "chest X-ray images"). Various computer-aided detection (or Diagnosis) (hereinafter, also simply referred to as “CAD”) methods have been proposed for detection (Non-Patent Documents 4 to 11), and expectations are rising.

However, in the detection of pulmonary nodules, pulmonary vascular shadows and their tangent images pose a problem as false positive candidates, but those that suppress pulmonary vascular shadows from a single chest X-ray image have not been well known so far. Therefore, the applicants used a two-dimensional histogram from a single chest X-ray image to obtain false positive shadows such as hilar pulmonary vascular shadows and their tangent images, and concentrations of clavicle, ribs, and peripheral pulmonary vascular shadows. We propose a pulmonary nodule clarification method that relatively clarifies the true pulmonary nodule by suppressing changes, and a pulmonary nodule clarification method that suppresses background noise (Patent Documents 1 and 2 and Non-Patent Document 12). And 13).

The background noise suppression pulmonary nodule clarification method contrasts the true pulmonary nodules in the clarification image, making it easier to point out. However, some of the pale lung nodule shadows with low contrast may not be conspicuous due to weak extraction as points and may not be pointed out (Non-Patent Document 13).

Japanese Unexamined Patent Publication No. 2017-018339 JP-A-2018-000312

Ministry of Health, Labor and Welfare: 2017 (2017) Overview of vital statistics (fixed number), statistical table Table 7 Gender mortality and mortality rate by simple classification of causes of death (per 100,000 population) http://www.mhlw.go .jp / toukei / saikin / hw / jinkou / kakutei17 / index.html September 7, 2018 Soda H, Tomita H, Kohno S, et al: Limitation of annual screening chest radiography for the diagnosis of lung cancer. A retrospective study, Cancer 72: 2341-2346, 1993 Shah PK, Austin JHM, White CS, et al: Missed non-small cell lung cancer: radiographic findings of potentially resectable lesions visible only in retrospect. Radiology 226: 235-241, 2003 Nakagawa K, Oosawa A, Tanaka H, et al: Clinical effectiveness of improved temporal subtraction for digital chest radiographys. Proc SPIE 4686: 319-330, 2002 Tsuyoshi Kawaguchi, Yoshitomi Harada, Ryoichi Nagata, et al .: Alignment method for contralateral difference of chest X-ray image. Med Imag Tech 28 (5): 351-361, 2010 Akira Sawada, Yoshinobu Sato, Naoji Kido, et al .: Detection of lung mass shadows on chest X-ray images-Multi-resolution filters, use and performance analysis of energy difference images-Med Imag Tech 17 (1): 81-91, 1999 Tadashi Oda, Nao Osamu Kido, Itsu Shono, et al .: Development of automatic detection system for nodular shadows using time-dependent difference images in chest radiographs D-11 J87-D-ll (Journal of the Institute of Electronics, Information and Communication Engineers) 1): 208-218, 2004 Suzuki K, Abe H, MacMahon H, et al: Image-processing technique for suppressing ribs in chest radiographs by means of massive training artificial neural network (MTANN). IEEE Trans Medical Imaging 25 (4): 406-416, 2006 Takeshi Hara, Hiroshi Fujita, Jin Yoshimura, et al .: Automatic detection of nodular shadows on chest radiographs-application of genetic algorithms-, Med Image Tech 15 (1): 73-81, 1997 Junji Morishita, Shigehiko Katsurakawa, Kunio Doi: Removal of false positive shadows by shape feature analysis of lung nodular shadows on chest radiographs, Journal of Japanese Society of Radiological Technology 57 (7): 829-836, 2001 Mikako Hiura, Naoji Kido, Itsu Shono: Development of a nodular shadow extraction method for plain chest X-ray images. Med Imag Tech 23 (4): 250-258, 2005 Yoshitomi Harada, Tatsuhachi Nomura, Hidetoshi Miyake: Pulmonary nodule clarification method of chest plain X-ray using a two-dimensional histogram, Med Imag Tech 35 (4): 239-249, 2017 Yoshitomi Harada, Hidetoshi Miyake: Pulmonary Nodule Clarification Method with Suppressed Background Noise, Med Imag Tech 36 (5): 221-230, 2018

The lung nodule clarification method by the two-dimensional cell automaton according to the embodiment of the present disclosure uses a two-dimensional cell automaton from a binary error diffusion image with respect to the pulmonary nodule clarification image obtained by the conventional pulmonary nodule clarification method. The purpose of this study is to extract pale lung nodule shadows existing in the hilar region and lung field as a set of black spots and clarify them with an adaptive rank filter.

The lung nodule clarification method according to the embodiment of the present disclosure is a first method of creating a pulmonary nodule clarification image by adjusting the brightness value of a false positive shadow extracted from a chest X-ray image using a two-dimensional histogram. The second step of creating a background noise suppression image in which the density change in the normal shadow of the lung nodule clarification image is suppressed as the background noise using the step and multiple resolution analysis, and the error diffusion method for the lung nodule clarification image. Initial state based on the third step of applying and creating a binary error diffusion image and the rank of pixel values within a predetermined area obtained by using a rank filter on the binary error diffusion image. By performing the state transition by the two-dimensional cell automaton for the fourth stage of creating the image of the above and the image of the initial state, the pale lung nodule shadow is extracted as a set of binary points and the binary point image is obtained. By combining a binary point image, a lung nodule clarification image, and a background noise suppression image to clarify the faint lung nodule shadow in the fifth step of creating the lung nodule clarification image. It is characterized by having a sixth stage of preparation and a seventh stage of outputting the prepared adaptation rank lung nodule clarification image.

In the first step, it is preferable to reduce the false positive shadow by adjusting the brightness value of the false positive shadow.

The second stage is the stage of extracting shadows using multi-resolution analysis by wavelet transform, and the stage of performing multi-resolution analysis to the depth of the first level of the lung nodule clarification image, excluding the low frequency side of the first level. It is preferable to include a step of compressing the components contained in the two levels to a predetermined size in a signed state, and a step of reconstructing the image by wavelet inverse transform so that the luminance value falls within a predetermined range.

The third stage is a stage in which the lung nodule clarification image has a predetermined filter size and the opening process is performed with a filter having a uniform pixel weight, and a stage in which the error diffusion method is applied to the image after the opening process. , Are preferably included.

In the fourth step, a rank filter of a predetermined size is used from the error diffusion image, and the black pixel is the one whose rank is equal to or higher than the first value, and the white pixel is the one whose rank is higher than the first value. It is preferable to include a step of creating an image in an initial state.

In the first state where the number of black pixels in the vicinity of the pixel of interest and Moore, which is in the vicinity of 8 of the pixel of interest, is 1 to 3, the fifth stage is a stage in which the pixel of interest is a black pixel and the vicinity of Moore. In the second state where the number of black pixels is four, the value of the pixel of interest is left as it is, and in the third state other than the first and second states, all the pixels of interest are white pixels. And the first generation value λ in the equation (λ (X) = A / B) representing the function λ (X) of the survival rate in the Xth generation, where the first to third states are one generation. It is preferable to include a step of extracting a pale shadow as a set of black pixels by repeating the generation change until the absolute value of the difference between (1) and the value λ (X) of the Xth generation becomes the threshold Th. .. However, A is the number when 1 to 3 pixels are black pixels in the Moore neighborhood + the number when 4 pixels are black pixels in the Moore neighborhood and the attention pixel is also a black pixel, and B is all. The number of transitions (total number of pixels excluding 1 pixel on the outer circumference).

The sixth step is to replace the black pixels of the binary point image with the pixels of the lung nodule clarification image at the same coordinates and the white pixels of the binary point image with the pixels of the background noise suppression image at the same position. It is preferable to include a step of synthesizing the two and a step of smoothing the combined image by using an average value filter.

According to the lung nodule clarification method by the two-dimensional cell automaton according to the embodiment of the present disclosure, a two-dimensional cell is obtained from a binary error diffusion image with respect to the lung nodule clarification image obtained by the conventional pulmonary nodule clarification method. Pale lung nodule shadows present in the hilar and lung fields can be extracted as a set of black spots using an automaton and clarified by an adaptive rank filter.

It is a figure which shows the pulmonary nodule clarification image and the coordinate system. It is a flowchart for demonstrating the flow of the lung nodule clarification method which concerns on embodiment of this disclosure. It is a figure which shows the background noise suppression image. It is a figure which shows the Moore neighborhood. It is a figure which shows the point image. It is a figure which shows the adaptation rank pulmonary nodule clarification image. (A) is a diagram showing a conventional background noise suppression pulmonary nodule clarification image, and (b) is an example of an adaptive rank pulmonary nodule clarification image by the lung nodule clarification method according to the embodiment of the present disclosure (background noise: It is a figure which shows no change / nodule (an example which improved). (A) is a diagram showing a conventional background noise suppression pulmonary nodule clarification image, and (b) is an example of an adaptive rank pulmonary nodule clarification image by the lung nodule clarification method according to the embodiment of the present disclosure (background noise: It is a figure which shows no change / nodule (an example which improved). (A) is a diagram showing a conventional background noise-suppressed pulmonary nodule clarification image, and (b) is an example of an adaptive rank pulmonary nodule clarification image by the lung nodule clarification method according to the embodiment of the present disclosure (background noise: It is a figure which shows the minor change / nodule (example of no change). It is a figure which shows the state transition simulation of a black dot by a generation change, (a) is a figure which shows 256 gradation gray scale, (b) is an error diffusion image, (c) is a figure which shows the initial state. It is a figure which shows the state transition simulation of a black dot by a generation change, (a)-(e) is a figure which shows the 1st to 5th generation, (f) is a figure which shows the 10th generation. It is a figure which shows the adaptation rank pulmonary nodule clarification image of a different generation, (a) is an original image, and (b) is a figure which shows the 1st generation. It is a figure which shows the adaptation rank pulmonary nodule clarification image of a different generation, (a) is a figure which shows the 2nd generation, (b) is a figure which shows the 3rd generation. It is a figure which shows the adaptation rank pulmonary nodule clarification image of a different generation, (a) is a figure which shows the 4th generation, (b) is a figure which shows the 10th generation. It is a figure which shows the state transition up to the 8th generation of the isolated black dot (1 pixel) in the initial state. It is a figure which shows the change of the λ value from the 1st to 200th generation.

Hereinafter, a method for clarifying lung nodules by a two-dimensional cellular automaton according to the present invention will be described with reference to the drawings. However, it should be noted that the technical scope of the present invention is not limited to those embodiments and extends to the inventions described in the claims and their equivalents.

[Target image]
Image processing of 154 chest mass images, matrix dimensions 2048 x 2048 (pixel dimensions 0.175 mm), and gradation number 4096 (12 bits) in the standard digital image database (Reference 1) created by the Japanese Society of Radiological Technology (JSRT). A lung created based on Non-Patent Document 12 for an image converted so that the matrix size is 512 × 512 (pixel size 0.7 mm) and the gradation is 8 bits using the software ImageJ 1.46r (Reference 2). A nodule clarification image (hereinafter, also simply referred to as “clarification image”) was used. Specifically, a lung nodule clarification image was created by adjusting the brightness value of the false positive shadow extracted from the chest X-ray image using a two-dimensional histogram. By adjusting the brightness value of the false positive shadow, the false positive shadow could be reduced.

FIG. 1 shows a pulmonary nodule clarification image and a coordinate system. As shown in FIG. 1, using a coordinate system in which the upper left corner is the origin and the horizontal and vertical axes are the x-axis and the y-axis, respectively, the luminance value 0 is "black", 255 is "white", and the luminance value 255 is the maximum. Let it be the brightness value. In addition, the cells in the two-dimensional cellular automaton correspond to each pixel, and black and white are used as the "live" and "dead" states, respectively. Furthermore, here, the faint pulmonary nodule shadows existing in the hilar region and the lung field are extracted.

[Proposed method]
FIG. 2 shows a flowchart for explaining the flow of the lung nodule clarification method according to the embodiment of the present disclosure. In the lung nodule clarification method according to the embodiment of the present disclosure, first, in step S101, a pulmonary nodule clarification image is created from a chest X-ray image using a two-dimensional histogram. Next, in step S102, a background noise suppression image in which the density change in the normal shadow of the clarification image is suppressed as the background noise is created by using the multi-resolution analysis (Non-Patent Document 13). Next, in step S103, the error diffusion method is applied to the clarification image to create a binary error diffusion image. Next, in step S104, an image in the initial state is created based on the rank of the pixel values in the predetermined region obtained by using the rank filter for the binary error diffusion image. Next, in step S105, a two-dimensional cellular automaton is used to perform a state transition on the image in the initial state to extract a pale lung nodule shadow as a set of binary points to create a binary point image. Next, in step S106, an adaptive rank lung nodule clarification image is created by synthesizing a binary point image, a pulmonary nodule clarification image, and a background noise suppression image in order to clarify a faint lung nodule shadow. .. Next, in step S107, an adaptive rank lung nodule clarification image is output.

[Creating a background noise suppression image]
Based on Non-Patent Document 13, the density change observed in normal shadows such as clavicle, ribs, and peripheral pulmonary blood vessels is defined as background noise, and multi-resolution analysis by wavelet transform (for example, Daubechies wavelet (N = 5)) is used. , Extract those shades. In the lung nodule clarification method according to the embodiment of the present disclosure, multi-resolution analysis is performed up to the first level depth (for example, depth level -5) of the clarification image, and the low frequency range of the first level (-5) is performed. The components contained in the second level (for example, level-3) excluding the side are compressed to a predetermined size (for example, 1/4) in a signed state.

In the restoration, the image is reconstructed by the wavelet inverse transform so that the luminance value falls within a predetermined range (for example, 0 to 255). The reconstructed image is shown in FIG. 3 as a background noise suppression image.

[Creation of error diffusion image]
On the other hand, in order to make the lung nodule shadow stand out in isolation from the surroundings, the clarified image has a predetermined filter size (for example, 21 × 21 pixels) based on Non-Patent Document 13 and has a uniform weight (square structure). Perform the opening process with a filter. Further, the error diffusion method is applied to the image after the opening process, and a binary error diffusion image is created from the clarification image.

[Create initial state]
Next, from the error diffusion image, using a rank filter of a predetermined size (for example, 3 × 3 pixels), a black pixel is defined as a value having a rank equal to or higher than the first value (for example, 5), and other values. An image in a binary initial state is created in which the one indicating is a white pixel.

[Search for faint shadows by state transition]
Since the difference between the light shadow and the brightness value around it is small, the set of points is vaguely expressed as sparse in the binarization by the pseudo gradation expression of the error diffusion method. Therefore, a pale shadow is extracted using the following three local conditions. Specifically, all the pixels in the initial state are processed according to the following three states.
State 1: When the number of black pixels in the attention pixel and its 8 neighborhoods (Moore neighborhood: FIG. 4) is 1 to 3, the attention pixel is defined as a black pixel (hereinafter, also referred to as “black dot”).
State 2: When the number of black pixels in the Moore neighborhood is 4, the value of the pixel of interest remains unchanged.
State 3: In cases other than states 1 and 2, all the pixels of interest are white pixels (hereinafter, also referred to as “white spots”).

Let states 1 to 3 be one generation (birth, survival, and death, respectively), and the difference between the first generation value λ (1) and the Xth generation value λ (X) in the following equation (1). The above operation is repeated until the absolute value becomes equal to or higher than the threshold value Th (| λ (X) −λ (1) | ≧ Th). By repeating the alternation of generations, pale shadows are extracted as a set of sunspots.
λ (X) = A / B (1)
However,
A: The sum of the number when 1-3 pixels are black dots near Moore and the number when 4 pixels are black dots and the pixel of interest is also a black dot near Moore B: Total number of transitions (1 pixel on the outer circumference) Total number of pixels excluding)
X: Generation λ was used as a function of survival rate in each generation (hereinafter, also referred to as “λ value”).

[Create point image]
Next, from the set of black spots and white spots extracted by the search, black spots of rank 4 or higher, white spots of rank 5 or higher, black spots (luminance value: 0), and other points are selected from the error diffusion image. A binary point image (hereinafter, also simply referred to as “point image”) with white points (luminance value: 255) is created (FIG. 5).

Finally, the black dots in the point image are used as the luminance values of the lung nodule clarification image, and the white dots are used as the luminance values of the background noise suppression image. That is, the black dots of the point image are replaced with the pixels of the lung nodule clarification image having the same coordinates, and the white dots of the point image are replaced with the pixels of the background noise suppression image having the same coordinates. In addition, an average value filter (3 x 3 pixels) is used to smooth the composited image to create an adaptive rank lung nodule clarification image.

FIG. 6 shows an adaptive rank lung nodule clarification image by the pulmonary nodule clarification method according to the embodiment of the present disclosure, and FIGS. 7, 8, and 9 show a conventional background noise suppression pulmonary nodule clarification image and the present disclosure. An example of an indication rank pulmonary nodule clarification image by the pulmonary nodule clarification method according to the embodiment is shown. The circles (◯) in FIGS. 7 (a), 8 (a), and 9 (a) indicate the positions where the true nodules are included.

[Experiment]
For the experiment, a lung nodule clarification image to which the pulmonary nodule clarification method (Non-Patent Document 12) was applied to the original image of 154 chest tumor images in the standard digital image database of the Japanese Society of Radiological Technology (JSRT) was used. There was. In the 154 examples of the original image, the position of the nodule is shown in advance, and depending on the difficulty of detecting the nodule, 1 (extremely difficult), 2 (very difficult), 3 (difficult), 4 (relatively easy), 5 ( It is classified into 5 levels (easy). In addition, there are 25, 29, 50, 38, and 12 cases of images of each level, respectively, but in the lung nodule clarification method according to the embodiment of the present disclosure, level 5 (easy) and level 1 (extremely difficult). The following experiments were performed on 117 cases of levels 2 to 4, which are likely to cause clinical problems, excluding cases.

[Experiment 1: Background noise suppression method for clarifying lung nodules]
Each target area (hereinafter, also referred to as “proposal image”) spread within both the image according to Non-Patent Document 13 and the adaptation rank lung nodule clarification image (hereinafter, also referred to as “proposal image”) according to the lung nodule clarification method according to the embodiment of the present disclosure. The standard deviation σ of the average brightness distribution in ROI: Region of Interest) was calculated and compared. The size of the ROI used here was 16 × 16 pixels, which is large enough to fit between the ribs (Table 1).

[Experiment 2]
Two diagnostic imaging specialists compared and evaluated the proposed image and the background noise-suppressed pulmonary nodule clarification image for the difference in the suppression of background noise around the nodule (Table 2a). When the background noise suppression is evaluated as no change, it is evaluated as "no change", when it is evaluated as slightly changed, it is evaluated as "minor change", and when it is evaluated as significantly changed, it is evaluated as "major change". Table 2a shows the respective ratios in 117 cases.

In addition, regarding the appearance of pale lung nodule shadows, the proposed image was compared with the background noise-suppressed lung nodule clarification image, and those evaluated as "improved" by both of them were evaluated as "background noise-suppressed lung nodule clarification image < "Proposed image", and those who evaluated that only one of them improved was "No change" ("Background noise suppression lung nodule clarification image = Proposed image"), and none of them evaluated that it improved. Was "worse" ("background noise suppression lung nodule clarification image> proposed image"), and the ratio of the cases evaluated at each stage in 117 cases was summarized (Table 2b). ). The execution time of this method was 1803 [ms] on average for a PC having an operating frequency of 2.2 GHz per image in 154 cases. The proposed method averaged about 4881 [ms] even when combined with the pulmonary nodule clarification method (Non-Patent Document 12).

[Discussion]
[Sunspot state transition simulation]
FIG. 10A (a) shows a range (a range indicated by an arrow (⇔) in FIG. 10A (a)) of a gray scale of 256 gradations and a brightness value (about 128 to 160) of a light shadow to be extracted. FIG. 10A (b) shows an error diffusion image of FIG. 10A (a), and FIG. 10A (c) shows the initial state. The black spots existing in the extraction range in the initial state grow even larger due to the black spots created in the first generation by the two-dimensional cellular automaton (FIG. 10B (a)). In the second generation, the inside of the large-grown sunspot dies and the outer circumference remains, or it is connected to the nearby quadrangle that has grown (Fig. 10B (b)). In the third generation, black spots are created in the luminance range to be extracted, and the density of black spots is increasing (Fig. 10B (c)). Even in the 4th generation, the density of black spots in the brightness range to be extracted is still high (Fig. 10B (d)), but the black spots show a geometric pattern every time the generation is repeated (Fig. 10B (e)). , A new pattern is created by connecting with a nearby pattern (Fig. 10B (f)).

If the brightness range of the faint lung nodule shadow is linearly extracted from the histogram of the clarification image, a lot of noise is included and it becomes conspicuous. Therefore, it was necessary to have the property of repeating birth, survival, and death, such as a simulation that non-linearly searches for pale lung nodules based on local conditions by a two-dimensional cellular automaton.

[Initial state and adaptive rank filter]
In the proposed method, a binary initial state was created by using a rank filter having a rank of 5 or more as a black point and other values as white points. If the shadow has a stronger contrast than the surroundings, the lung nodules can be sufficiently extracted as a set of black spots even if rank 5 or higher is used. However, it was necessary to include rank 4 in the case of light shadow extraction. Here, if a rank 4 or higher is adopted as the initial state, many shadows such as ribs remain, and the nodular shadow cannot be separated from the surroundings. Further, the brightness values of the original images indicating rank 4 and rank 5 are similar, and it is considered that the shadow indicating rank 4 exists near rank 5. Therefore, by examining the vicinity of rank 5, it is considered that false positives can be suppressed as much as possible and pale shadows can be extracted.

Therefore, in the proposed method, a rank 5 or higher is adopted as the initial state, and a faint shadow indicating rank 4 is searched in the vicinity using a two-dimensional cellular automaton. Among the sets extracted as black points by the two-dimensional cellular automaton, the points showing a rank of 4 or more in the error diffusion image are regarded as pale lung nodule candidates, and the rank filter is applied so that the rank of 5 or more is adopted in other points. It was used in a variable manner.

[About state transition]
Figures 11A to 11C show proposed images for different generations. Although the true lung nodule can be emphasized in the first generation of FIG. 11A (b), the false positive shadow (indicated by the arrow (→) in FIG. 11A (b)) is more than the true lung nodule shadow. Was conspicuous. In the 2nd and 4th generations, true nodules and false positive shadows are suppressed and inconspicuous (FIGS. 11B (a) and 11C (a)). As the generations change further (Fig. 11C (b)), false positive shadows become more prominent (indicated by the arrow (→) in Fig. 11C (b)), and black spots increase in the lung nodules and the shadows around them. , There is no difference in brightness between the two, and pale lung nodules become less noticeable. However, in the third generation shown in FIG. 11B (b), true nodules are emphasized and false positive shadows are inconspicuously suppressed. This is because the pale shadows to be extracted exist near the initial state, and the points showing true lung nodules rather than the false positive shadows are densely emphasized in the state transitions caused by several generational changes.

That is, as shown in FIG. 12, focusing on the black spots (1 pixel) isolated from the surroundings in the initial state, there are no black spots in the initial state in the third generation. However, in the proposed method, nodule shadows can be extracted as a set of sunspots in the state transition of the third generation. If the sunspots exist in isolation from the surroundings (shadows with a higher brightness value than the extraction target), it is weak to extract the shadows as a set of points in the alternation of generations of the third generation. On the other hand, if another sunspot exists in the vicinity (shadow of the extraction target), the sunspot is generated even in the alternation of generations of three generations due to the interaction, and a set is formed. It is considered that such a phenomenon due to the alternation of generations was influenced by the state of pixels in the vicinity of the local area. As described above, it is considered that the two-dimensional cellular automaton strengthens or weakens the shadow by the transition of the set state of the sunspots by the interaction with the neighboring pixels.

[About λ value]
Langton's λ parameter (Reference 3) used in artificial life is used as a measure for evaluating the behavior and complexity of cellular automata (the value in this method is λ = 0.361). However, in the lung nodule clarification method according to the embodiment of the present disclosure, the value of λ was used as the survival rate function of each generation in order to observe the behavior of the two-dimensional cellular automaton.

FIG. 13 shows the change in the λ value from the first generation to the 200th generation. As the alternation of generations progresses, the λ value takes a form close to a saturated curve. A and b in FIG. 13 show a transient state in which the λ value fluctuates sharply and a steady state in which the fluctuation is small, respectively. The increase in the λ value in the transient state depends on the contrast of the original image, but in each image, the stationary portion converges to about 0.4. Even in the vicinity of the 200th generation, the depiction of true lung nodules remained, but many false nodules occurred.

Also, from the state transition simulation results of the sunspots in Section 2, when the generations are changed for about 10 generations, the λ value (survival rate) rises sharply, but the expanded sunspots occupy the search area, resulting in noise. There is a risk of increasing. Therefore, as a result of investigating the life-and-death state of points in the alternation of generations, false positives were suppressed to some extent in the generation in which the λ value was close to the initial state and the survival rate (number of sunspots) increased significantly, and among the remaining sunspots. It was likely that pale lung nodules would be extracted. In the experiment, pale lung nodule shadows were extracted as a set of sunspots even in a few generational changes from the initial state to the third generation. This is thought to be because the dominant information was inherited by the cells that survived from birth even with a small number of generational changes.

[Termination conditions]
When the pale lung nodular shadows are simply extracted using a low rank (ranks 1 to 4) filter, the pale shadows can be extracted but may contain many false positive shadows. On the other hand, if only the true lung nodule shadow can be correctly extracted as a set of points by alternation of generations, there is no big difference compared to the total number of extraction points born and survived in the initial state or the first generation, and false nodules are also present. It is considered that there are few.

Therefore, in the proposed method, the generation change is repeated within a range (threshold value Th) in which the λ value does not increase too much, based on the λ value of the first generation. The λ value, which ends with a generational change of about 3 generations, was sufficient in consideration of the calculation time as long as the number of surviving cells of the 1st generation increased or decreased by about 3%. Therefore, in the lung nodule clarification method according to the embodiment of the present disclosure, the threshold value of the termination condition is set to Th = 0.03. However, as the boundary condition, all the outer peripheral cells are set to black (raw) and initialized every time.

As a result, the alternation of generations was about 3.41 generations on average, and it was possible to suppress false extraction to some extent and extract pale true lung nodules existing in the hilar and lung fields as a set of points.

[Experimental results]
[Result of Experiment 1]
In the proposed image, the compression rate of the background noise suppression image to be combined is set to be 1.0 higher than that of the background noise suppression lung nodule clarification method. This is because, in the proposed method, the number of points showing rank 4 (used for the synthesis process) increases due to the search. However, when the average luminance distribution was compared with the background noise suppression pulmonary nodule clarification method, the difference was less than 0.15 (Table 1). Further, even when the background noise suppression pulmonary nodule clarification method and the compression ratio were compared, the standard deviation was σ = 42.36, which was almost the same value. As a result, it was possible to clarify the faint pulmonary nodule shadow while suppressing the background noise to the same extent as the background noise suppression pulmonary nodule clarification method.

[Result of Experiment 2]
In the proposed method, there was no change in background noise due to emphasis on pale lung nodules in 76.1% of the images compared to the background noise suppression pulmonary nodule clarification method. On the other hand, 23.9% of the lung nodules on the outside of the thorax and the lung nodules overlapping the rib edges had white spots and black spots appearing as noise.

However, no significant increase in noise was observed that affected nodule detection (Table 2a). In addition, the appearance of nodule shadows was equal to or better than that of the background noise suppression pulmonary nodule clarification method (Table 2b). It is considered that this is because the background noise was suppressed to some extent and the pale shadow was clarified because the neighborhood was searched based on the background noise suppression pulmonary nodule clarification method in order to extract the pale lung nodule shadow.

For the detection of lung nodules, many methods such as CAD have been proposed for classifying lung nodules by a hyperplane from many features of the lung nodules. On the other hand, in the proposed method, it was possible to extract faint shadows even by using a mechanism called emergence in the field of artificial life (local interaction between pixels affects the entire image and determines its state). .. This suggests that the lung nodule shadow may have emergent properties that cannot be separated simply by the threshold value.

By clarifying the faint pulmonary nodule shadows present in the hilar and lung fields by the proposed method, the results are improved compared to the background noise suppression pulmonary nodule clarification method, and it is considered that it is fully applicable clinically.

[Summary]
According to the lung nodule clarification method using a two-dimensional cellular automaton according to the embodiment of the present disclosure, nodule candidates are extracted from a lung nodule clarification image using a two-dimensional cellular automaton and an adaptive rank filter is used to adaptively use a rank filter. The pale lung nodules present in the area and lung field could also be relatively clarified.

(Reference 1) Shiraishi J, Katsuragawa S, Ikezoe J, et al: Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. AJR 174: 71- 74, 2000
(Reference 2) Schneider CA, Rasband WS, Eliceiri KW: NIH Image to ImageJ: 25 years of image analysis. Nature Methods 9: 671-675, 2012
(Reference 3) Langton CG: Computation at the Edge of Chaos: Phase Transitions and Emergent Computation: Physica D, 42, pp.12-37, 1990

Claims (7)

The first step of creating a lung nodule clarification image by adjusting the brightness value of the false positive shadow extracted from the chest X-ray image using a two-dimensional histogram, and
The second step of creating a background noise suppression image in which the density change in the normal shadow of the lung nodule clarification image is suppressed as the background noise by using the multi-resolution analysis, and
The third step of applying the error diffusion method to the lung nodule clarification image to create a binary error diffusion image, and
The fourth step of creating an image in the initial state based on the rank of the pixel values in the predetermined region obtained by using the rank filter for the binary error diffusion image, and
The fifth step of creating a binary point image by extracting a pale lung nodule shadow as a set of binary points by performing a state transition with a two-dimensional cellular automaton on the image in the initial state.
A sixth to create an adaptive rank lung nodule clarification image by synthesizing the binary point image, the pulmonary nodule clarification image, and the background noise suppression image in order to clarify the pale lung nodule shadow. Stages and
The seventh step of outputting the created adaptive rank lung nodule clarification image, and
A method for clarifying lung nodules, which is characterized by having. The lung nodule clarification method according to claim 1, wherein the first step reduces false positive shadows by adjusting the brightness value of the false positive shadows. The second step is
The stage of extracting shadows using multi-resolution analysis by wavelet transform,
A step of performing multi-resolution analysis to the depth of the first level of the lung nodule clarification image and compressing the components contained in the second level excluding the low frequency side of the first level to a predetermined size in a signed state. ,
The stage of image reconstruction by wavelet inverse transform so that the brightness value falls within a predetermined range, and
The pulmonary nodule clarification method according to claim 1 or 2, which comprises. The third step is
The stage of performing the opening process with a filter having a predetermined filter size and a uniform weight on the lung nodule clarification image, and
At the stage of applying the error diffusion method to the image after the opening process,
The pulmonary nodule clarification method according to any one of claims 1 to 3, which comprises. In the fourth step, using a rank filter of a predetermined size from the error diffusion image, black pixels are defined as those having a rank equal to or higher than the first value, and white pixels are defined as other values. The lung nodule clarification method according to any one of claims 1 to 4, which comprises a step of creating an image of an initial state of values. The fifth step is
In the first state where the number of black pixels in the vicinity of the pixel of interest and Moore, which is in the vicinity of 8 of the pixel of interest, is 1 to 3, the step of setting the pixel of interest as a black pixel and
In the second state where the number of black pixels is 4 in the Moore neighborhood, the step of leaving the value of the pixel of interest as it is and
In the case of the third state other than the first and second states, the stage in which all the pixels of interest are white pixels and
Letting the first to third states be one generation, the first generation value λ (1) and the Xth generation value λ (X) in the following equation representing the function λ (X) of the survival rate in the Xth generation. ) And the stage of extracting a faint shadow as a set of black pixels by repeating generational change until the absolute value of the difference becomes equal to or higher than the threshold Th.
The pulmonary nodule clarification method according to any one of claims 1 to 5, which comprises.
λ (X) = A / B
However,
A: Number when 1 to 3 pixels are black pixels in the Moore neighborhood + Number when 4 pixels are black pixels in the Moore neighborhood and the attention pixel is also a black pixel B: Total number of transitions (outer circumference 1) Total number of pixels excluding pixels) The sixth step is
By replacing the black pixel of the binary point image with the pixel of the lung nodule clarification image having the same coordinates and replacing the white pixel of the binary point image with the pixel of the background noise suppression image at the same position. The stage of synthesizing both,
The step of smoothing the combined image using an average value filter and
The pulmonary nodule clarification method according to claim 6, which comprises.