JP2616520B2 - Medical image display device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a medical image for displaying an image of a subject taken by using an image diagnostic apparatus such as an X-ray CT apparatus or a nuclear magnetic resonance tomography apparatus (hereinafter referred to as an MRI apparatus). The present invention relates to a display device, and more particularly to a technique for adjusting the contrast of a displayed image.

[0002]

2. Description of the Related Art Contrast adjustment of an image displayed on a monitor such as a CRT display is performed by setting a window level (hereinafter abbreviated as WL) and a window width (hereinafter abbreviated as WW) in an image density histogram. Have been done. Actually, a trackball for adjusting WL and WW is provided on an image display device on which a monitor is installed, and a doctor or a technician who interprets an image operates the trackball and displays the image on the monitor. The contrast of the displayed image is adjusted. In this case, the density histogram is not displayed, and the setting of WL and WW by the operation of the trackball is processed inside the apparatus and reflected on the display image.
Hereinafter, WW and WL will be described with reference to the drawings.

FIG. 10 exemplifies a density histogram of an image of a subject captured by an image diagnostic apparatus. As is well known, a density histogram is created by extracting the density value of each pixel of an image and counting the density value. A CRT display will be described as an example of a device that displays an image represented by such a density histogram. Since the display gradation value of a CRT display is generally 256 gradations, the width of the density value on the horizontal axis of the density histogram is divided into 256 from the minimum value to the maximum value. The above WW is a parameter for arbitrarily setting the range of the density value, and WL indicates a central value of the WW.

[0004] The contrast adjustment is performed in a direction to clarify the difference in brightness between a part to be interpreted (for example, a lesion part) and a surrounding part. Now, suppose that it is desired to clarify the difference in brightness between the area A and the density histogram shown in FIG. As described above, the density value of the density histogram is divided into 256 as a whole. However, if the density value of the area A is divided into 256, the density difference in the area A becomes remarkable and the light-dark difference becomes clear. That is, WW may be set to the area A, and WL may be set to the center value. Conventionally, the WW and WL are set while arbitrarily operating the above-described trackball and observing the change in the contrast of the displayed image.

In this method, it is necessary to adjust the contrast of each displayed image by operating a trackball. In particular, since an image captured using an MRI apparatus has a significantly different density value depending on an imaging method, an imaging parameter (such as a pulse sequence) and an imaging site, it is essential to adjust contrast of each displayed image. . However, there is an increasing demand for shortening the examination time in order to improve the examination throughput and reduce the burden on the patient. In many cases, it is sufficient if the contrast can be adjusted to some extent. Therefore, a technique for automatically setting the contrast has been developed.

In the automatic contrast setting of an image picked up by the MRI apparatus, first, it is determined that the image is appropriate on average according to the imaging method, the imaging parameters, and the imaging site.
A function expression (hereinafter referred to as an evaluation expression) for performing density conversion for obtaining values of W and WL is obtained. Then, a density histogram of the picked-up image is created, and given to the obtained evaluation expression, WW and WL are obtained. The doctor or the operator only needs to select the evaluation formula according to the imaging method, the imaging parameters, and the imaging region, and a certain degree of appropriate contrast adjustment is automatically performed. Of course, when fine contrast adjustment is desired, manual fine adjustment as described above is performed. However, otherwise, troublesome manual contrast adjustment can be omitted, and frequently used automatic contrast adjustment is performed. This is a setting method.

[0007]

SUMMARY OF THE INVENTION However, MRI
The imaging method and the imaging parameters in the apparatus are quite various, and deriving an evaluation formula for all of them is a serious task. In addition, new imaging methods and imaging parameters are being developed every day with the aim of shortening the imaging time, and it is possible to immediately respond to images captured with new imaging methods and imaging parameters. There is a drawback that you can not. Not only for MRI equipment, but also for X-ray CT equipment, it is a difficult task to derive each evaluation formula according to the type of X-ray intensity and scan position exposed to the subject. I have.

The present invention has been made in view of such circumstances, and has as its object to provide a medical image display device capable of automatically setting contrast while solving the above-mentioned disadvantages.

[0009]

The present invention has the following configuration to achieve the above object. That is, the present invention is a medical image display apparatus for setting a window width and a window level, which are parameters for contrast adjustment of an image captured by an image diagnostic apparatus, for a display image, Means for creating a density histogram of an image captured by a device, a neural network for calculating a weighted sum of the created density histogram data and outputting the window width and the window level, and a weight of the neural network. And learning means for learning and determining based on the density histograms of various images.

[0010]

According to the present invention, the learning means learns and determines the weight of the neural network based on the density histograms of various images. The neural network uses the weight determined as a result of the learning to calculate the weighted sum of the density histogram data of the image, and outputs the window width and the window level. Here, the output window widths and window levels are values calculated using the weights obtained as a result of learning the density histograms of various types of images. It will be a corresponding appropriate value.

[0011]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram illustrating a schematic configuration of the medical image display device according to the present embodiment. Histogram creation means 2 for obtaining a density histogram of an image of a subject imaged by an image diagnostic apparatus 1 such as an X-ray CT apparatus or an MRI apparatus, and a weighted sum calculation of the created density histogram data to obtain values of WW and WL (Hereinafter referred to as a neural network 3), a random number generator 4, learning means 5 for determining the weight of the neural network 3 by learning, and WW, WL calculated by the neural network 3. And a CRT display 7 for displaying an image.

As shown in the conceptual model of FIG. 2, the neural network 3 includes a plurality of neurons N (electronic components for performing the same load sum processing as the nerve cells) in a layered form, and It is configured by network connection with a bond called a bond C. In this example, a three-layer model of an input layer A, an intermediate layer B, and an output layer C is used. The input layer A includes 30 neurons NA1 to NAn, and the intermediate layer B includes 20 neurons NB1 to NBm. The output layer C is provided with two neurons NC1 and NC2. The number of layers and the number of neurons are not specified,
Any number may be set. The reason why the number of neurons in the output layer C is two is that these neurons NC1 and N
This is because the output value of C2 corresponds to WW and WL. Note that the neural network 3 may be configured by hardware or may be realized by software.

The operation of the medical image display device according to the present embodiment is roughly divided into two flows. First, (A) the weight of the synaptic connection C of the neural network 3 is determined by using the learning means 5, and (B) the weight of the learned synaptic connection C is set in the neural network 3 to set the actual weight. Adjust the contrast of the image. The neural network learning of (A) will be described below.

[A-1] An image used for learning is input to the histogram creating means 1. It is desirable to input as many learning images as possible as many as possible. For example, if an MRI apparatus is used as a medical diagnostic device, various images obtained by changing an imaging method and imaging parameters are used. The histogram creating means 1 extracts a predetermined number (for example, 4096 points) of pixel values from the input image and counts the number of pixels having the same value.
A density histogram as shown in FIG. 10 is obtained. Here, the reason why a predetermined number of pixel values are extracted is to make the number of samplings for creating the density histogram constant regardless of the size of the image.

Then, the range from the minimum value to the maximum value of the density value of the density histogram normally divided into 256 is divided into 30. This is to make the density value correspond to the number of neurons NAn in the input layer A of the neural network 3.
Divide into 30.

Next, the histogram creating means 1 normalizes the created density histogram so that it can be easily learned by the neural network 3. The purpose of learning is to uniformly optimize the W
It is to create a neural network that can set W and WL. Since the density histograms of all the patterns are learned, only the patterns are learned by defining the maximum and minimum values of the density values of the density histogram and the maximum and minimum values of the count number. The normalization of the density histogram is a process for that. For example,

Each count is converted such that the maximum count Y1 on the vertical axis of the density histogram becomes "1" as shown in FIG. 3 (a). Similarly, the maximum density value X2 on the horizontal axis divided into 30 is defined as “1”, and the minimum density value X1 is defined as “0”. Figure (b) is the result of standardization.
Is a density histogram as shown in FIG. When the above-described normalization is performed on the density histogram shown in FIG. 4A, the result becomes as shown in FIG.

[A-2] Learning process of the neural network 3 by the learning means 5 will be described below with reference to the flowcharts of FIGS. These flowcharts show the flow of one learning process.

First, in step S1 of FIG. 5, a random number is generated from the random number generator 4 shown in FIG. 1 to the neural network 3, and the random number value is weighted by the synapse connection C connecting each neuron N. And With the weight set here as an initial value, the following learning process is executed to determine the final weight. Note that, for simplicity of description, as shown in FIG. 7, each layer focuses on the i-th neuron.
Here, i of the neuron NAi in the A layer changes from 1 to n, i of the neuron NBi in the B layer changes from 1 to m,
The i of the neuron NCi in the C layer changes from 1 to 2.

The weight of the synapse connection that connects the neuron NAi of the input layer A and the neuron NBi of the intermediate layer B set by the random number generator 4 is represented by Wii, and the neuron NBi of the intermediate layer B and the neuron NCi of the output layer C are represented by Wii. Is represented by Wji.

The density histogram of a certain image standardized by the histogram generation means 2 of FIG. 1 is converted into each histogram data (density value × count) by the neuron N of the input layer A.
Input to Ai. XAi
This is set as the neuron value of the neuron NAi (step S2).

In step S3, a neuron value XBi of each neuron NBi of the intermediate layer B is calculated by the following equations (1) and (2). XBi ′ = Σ (XAi × Wii) (1) In the expression (1), 式 on the right-hand side finds the sum when i is changed from 1 to n. First, a weighted sum with the weight Wi of the synaptic connection is calculated. XBi = f (XBi ') = 1 / (1 + Exp (XBi')) (2) Next, using the equation (2), the load sum calculated by the equation (1) is calculated by the sigmoid function f. (X) to calculate the neuron value XBi of the neuron NBi. The sigmoid function is f
(X) = 1 / (1 + Exp (X)).

At step S4, similarly to step S3, the neuron value X of each neuron NCi of the output layer C is obtained.
Calculate Ci. That is, (3), which is similar to Eqs. (1) and (2),
(4) The neuron value NCi is calculated using the equation. XCi ′ = Σ (XBi × Wji) (3) XCi = f (XCi ′) = 1 / (1 + Exp (XCi ′)) (4)

Each of these calculated neuron values, XA
i, XBi, and XCi are stored in a memory (not shown) (step S5).

In the neural network 3, as described above, the neuron values NCi (i = 1, 2) propagated from the input layer A and appearing on the output layer C are set to WW and WL, respectively. Weight of synaptic connection W
Since ii and Wji are values set by random numbers, the neuron value NCi output here does not give optimum WW and WL. In the following processing, W
Learning based on W and WL is performed to determine the weight of the synaptic connection. This learning method is based on a generally used back vacation method (back propagation method).

In step S6, the image represented by the density histogram input in step S2 is contrast-adjusted by some specialists (doctors, surgeons, etc.), and the WW, Average value ww, wl of WL
Ask for. Then, this is output to the neuron NCi of the output layer C.
To enter.

In step S7, the neuron N of the output layer C
The difference between the Ci value XCi and ww, wl is calculated, and the error Eci of the neuron NCi is obtained to calculate the learning signal DCi. Eci = NCi-ww (wl). In the case of the neuron NCi of the output layer C, the error Eci calculated here becomes the learning signal DCi as it is (DCi = Eci).

In step S8, an error Ebi to be borne by each neuron NBi of the intermediate layer B is obtained to calculate a learning signal DBi. The error Ebi is calculated using the following equation (5). Ebi = Σ (Wji × DCi) (5) As shown in FIG. 8, the expression (5) represents the neuron NBi of the intermediate layer B and the neuron NCi of the output layer C (i = 1,
2) and weights Wj1 and Wj2 of the synaptic connection connecting
This is an equation for calculating the weighted sum with the learning signal DCi of the neuron NCi of the output layer C. This is the error Ebi to be borne by the neuron NBi.

Next, the neuron N is calculated using the following equation (6).
A learning signal DBi for Bi is calculated. DBi = Ebi × f ′ (XBi ′) (6) f ′ (XBi ′) in the expression (6) is obtained by adding the differential expression f ′ (X ′) of the sigmoid function in the step S3. The equation is obtained by substituting the obtained weight sum XBi '= Σ (XAi × Wii) of the neuron NBi and multiplying the obtained result by the error Ebi obtained in step S8. Thus, the learning signal DBi of each neuron NBi of the intermediate layer B is calculated.

In step S9, similarly, the learning signal DAi is calculated by obtaining the error Eai to be borne by each neuron NAi of the input layer A. That is, the following equations (7) and (8) similar to the equations (5) and (6) are used. Eai = Σ (Wji × DBi) (7) DAi = Eai × f ′ (XAi) (8)

Referring to FIG. Neuron N in input layer A
The error Eai to be borne by Ai was calculated in step S8, with the synaptic convenience weight Wii (i = 1 to m of the second character) connecting the neuron NAi and the neuron NBi (i = 1 to m) of the intermediate layer B. Learning signal DBi of neuron NBi
(I = 1 to m). The learning signal DAi is obtained by multiplying the obtained error Eai by the derivative f ′ (XAi) of the sigmoid function, where XAi is obtained by multiplying the value of the histogram data (density value × count number) input to the neuron XAi. Can be

In step S10, the weight Wi of the synaptic connection C is calculated using the calculated learning signal and the neuron value.
i, Wji are changed. The following equations (9) and (10) are used to change the weight. ΔWii (n + 1) = α (DBi × XAi) + β (ΔWii (n)) (9) In the expression (9), the symbol n indicates the number of changes, α is a learning constant set arbitrarily, β indicates an arbitrarily set stabilization constant. ΔWii (n + 1) indicates the value of the weight changed this time, and ΔWii (n) indicates the value of the weight changed last time. Now, since it is the first learning, there is no weight value changed last time, and ΔWii (n) is zero. That is, the neuron NAi of the input layer A and the neuron NBi of the intermediate layer B
Is changed to a value calculated by α (DBi × XAi).

.DELTA.Wji (n + 1) =. Alpha. (DCi.times.XBi) +. Beta. (. DELTA.Wji (n)) (10) This equation (10) is also the same, and the neuron NBi of the intermediate layer B and the neuron NCi of the output layer C Is changed to a value calculated by α (DCi × XBi).

As described above, by changing the weight of the synapse connection which connects the neurons of each layer once, 1
The learning of the time ends. In the second learning, the value of the weight obtained by the above-described processing is ΔWii (n), ΔWii (n) in Equations (9) and (10).
It is calculated by substituting into Wji (n). In the process of repeating this many times, the values of the errors Eai, Ebi, and Eci are monitored, and α and β of the equations (9) and (10) are set so that the values of the errors are reduced.

As described above, the purpose of this learning is to uniformly and optimally set WW and WL for density histograms of all patterns. Therefore, a plurality of images represented by different density histograms are prepared, and these images are trained many times to determine final synaptic connection weights Wii and Wji. At this time, it is desirable to perform learning by changing the input order of different images. For example, when ten images are input one after another and learning is completed once, the input of the ten images is changed in the second learning. This is because the neural network 3 learns the arrangement of the images and determines the weight of the synapse connection in accordance with the learning.

When the contrast of an image is actually adjusted by using the learned neural network 3,
It must be assumed that images having different density values are input at random. For this purpose, it is preferable to perform learning by changing the arrangement of images. For this learning, the present inventor prepared 206 MRI images imaged using different imaging methods and different imaging parameters, and performed 30,000 times learning while changing the arrangement of each image. As a result, by using the set synaptic connection weights Wii and Wji, even if images having different density values were randomly input, appropriate averages were obtained for each of them.

When the learning process is completed and the weights Wii and Wji of the synapse connection of the neural network 3 are determined, the contrast of the actual image is adjusted using the neural network 3 (see the above-mentioned code (B)). Described process). [B-1] First, the data of the display image is input to the histogram creating means 2 shown in FIG. 1 to create a standardized density histogram, which is input to the neural network 3.

[B-2] The neural network 3
The processing from step S2 to step S4 in the flowchart shown in FIG.
The values of 1, 2 and XC1 and XC2 are obtained. These neuron values XC1 and XC2 become the values of WW and WL. In step S6 of the flowchart shown in FIG.
The values of w and wl are input to the neurons NC1 and NC2 of the output layer C. At this time, if ww is input to the neuron NC1 for learning, the value XC1 of the neuron NC1 becomes the WW to be obtained, and the value XC2 of the neuron NC2 is obtained. Sought WL
Value.

However, the WW and WL obtained here are WW and WL for the standard density histogram.
Since the density histogram of the image actually displayed on the CRT display 6 has not been standardized,
The obtained WW and WL are converted so as to correspond to them. This is performed using the following equations (11) and (12). WW '= WW (X2-X1) (11) WL' = WL (X2-X1) + X1 (12)

Symbols X1 and X2 in the above equation indicate the minimum value and the maximum value of the density value shown in FIG. The histogram creating means 1 sets the width “X2−X1” of the minimum value and the maximum value of the density histogram to “1-0 = 1” for normalization.
, The obtained WW is multiplied by (X2−X1) in order to restore this. The value WW 'becomes the window width for the actual image. Similarly, WL is multiplied by (X2−X1), and X1 is further added. This defines X1 to be "0" as shown in FIG. 3B, and adds X1 to correct this. The obtained value WL 'is the window level for the actual image.

The inventor described that the present inventors prepared 206 MRI images and performed 30,000 learnings while changing the arrangement of each image. The processing time when WW ′ and WL ′ are obtained is described below for reference. Approximately 0.29 seconds, 256 × 256 for an image with 128 × 128 pixels
The size of the image is also 0.29 seconds, and the size of the image of 512 × 512 is about 0.43 seconds. In such a short time, appropriate WW 'and WL' were obtained on average.

The obtained WW 'and WL' are sent to the contrast setting section 6 shown in FIG. The contrast setting unit 6 sets WW ′ and WL ′ for the input image and
The output is displayed on the RT display 7.

When the doctor or the surgeon looks at the displayed image and determines that further contrast adjustment is necessary, manual fine adjustment is performed as described in the conventional example, but there is no particular need for fine adjustment. In this case, the contrast is automatically set without any manual operation. In addition, even if fine adjustment is made, since appropriate contrast adjustment is performed in advance on average,
Its operation becomes easy.

[0044]

As is apparent from the above description, the medical image display apparatus of the present invention determines the weight of the neural network by learning based on the density histogram of various kinds of images, and uses the determined weights to generate the image. The weighted sum calculation is performed on the density histogram data of (1) to obtain the window width and the window level, which are the parameters of the contrast setting. Therefore, appropriate window widgets and window levels can be obtained for images represented by different density histograms. Unlike conventional models, there is no need to derive an evaluation formula according to the type of image, and learning is performed based on the density histogram of various types of images. Can be set automatically.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a schematic configuration of a medical image display device according to one embodiment of the present invention.

FIG. 2 is a diagram conceptually modeling the configuration of a neural network.

FIG. 3 is a diagram illustrating a density histogram normalization process.

FIG. 4 is a diagram illustrating a normalization process of a density histogram.

FIG. 5 is a flowchart showing the flow of one learning process.

FIG. 6 is a flowchart following FIG. 5;

FIG. 7 is a diagram conceptually modeling a part of the configuration of a neural network.

FIG. 8 is a model diagram of a neural network showing a state of learning by the back vacation method.

FIG. 9 is also a model diagram of a neural network showing a state of learning by the back vacation method.

FIG. 10 is a density histogram for explaining a window width and a window level which are parameters for adjusting the contrast of an image in the related art.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Image diagnostic equipment 2 ... Histogram creation means 3 ... Neural network 5 ... Learning means

Claims (1)

(57) [Claims] 1. A medical image display device for setting a window width and a window level, which are parameters for contrast adjustment of an image captured by an image diagnostic device, for a display image, wherein the image diagnostic device Means for creating a density histogram of the image captured in the above, a weighted sum of the data of the generated density histogram, a neural network that outputs the window width and the window level, and a weight of the neural network. A medical image display device comprising: learning means for learning and determining based on density histograms of various images.