JPH0792090A - Fluorescence antibody judging means - Google Patents

Abstract

PURPOSE:To automatically perform the precise kind judgement of an antibody in an image, and quickly judge the kind without causing an individual difference in the judgement result. CONSTITUTION:The fluorescence antibody judging device consists of an image input part 10 for inputting an image data including a subject to be judged an image memory part 12 for storing the image data inputted by the image input means 10, a subject extracting means 14 for extracting the subject to be judged from the image data stored in the image memory part 12, an input data forming means 16 for forming an input data by regulating the size of the subject extracted by the subject extracting means 14 to the size consisting of a prescribed number of picture elements, a neural network processing means 18 for conducting a neural network processing to the input data formed by the input data forming means 16, and a kind output part 20 for outputting the kind of the subject on the basis of the processing result of the neural network processing means 18.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a fluorescent antibody determination apparatus, and more particularly to a fluorescent antibody determination apparatus for determining the type of antinuclear antibody by the fluorescent antibody method.

[0002]

2. Description of the Related Art Antinuclear antibody is a normal immune system (function)
Is an antibody that reacts with the nuclear component of cells, which is present in the sera of patients with autoimmune diseases in which disorders occur in so-called allergies. When the normal immune system is impaired, an antibody that reacts with its own body components, that is, an autoantibody, is produced, which causes the disease. Such antinuclear antibody is detected by a fluorescent antibody method or the like.

That is, the patient's serum is reacted with the cell nucleus, a fluorescent label is added, and the antinuclear antibody is observed using a fluorescence microscope. At this time, if the antinuclear antibody emits fluorescence, the patient's serum is positive. Otherwise, it is determined to be negative.

In the conventional inspection by the fluorescent antibody method, the inspector makes a type judgment by directly visually observing the fluorescent microscope, or the image captured by the fluorescent microscope is displayed on a CRT through a CCD camera or the like and the inspector is displayed. The type was determined by visual inspection.

[0005]

However, in the above-mentioned conventional example, the type determination by visual observation is normally performed in a dark room because the fluorescence emitted by the antinuclear antibody is weak. Therefore, there is an inconvenience that the working environment is bad.

Further, even when a CRT or the like is used, the image obtained from the fluorescence microscope is dark, so that it is difficult for the inspector to perform visual processing. Therefore, there is a problem in that the determination result depends on the individual, and skill is required to accurately determine the type. In particular, for ambiguous images that are similar to a plurality of types, there is a significant individual difference in the determination result.

[0007]

An object of the present invention is to improve the inconvenience of the conventional example, and to automatically perform accurate type determination of an antibody in an image without visual inspection by an inspector. It is an object of the present invention to provide a fluorescent antibody determination device capable of quickly performing type determination without causing individual differences.

[0008]

Therefore, in the present invention, an image input section for inputting image data including an object to be judged, an image storage section for storing the image data input by this image input section, and this image storage section. Determination target object extracting means for extracting the determination target object from the image data stored in the unit, and the size of the determination target object extracted by the determination target object extracting means is normalized to a size consisting of a predetermined number of pixels. Input data creating means for creating input data by converting the input data, a neural network processing means for performing a neural network processing on the input data created by the input data creating means, and a processing result of the neural network processing means And a type output unit that outputs the type of the determination target object. This aims to achieve the above-mentioned object.

[0009]

The image input section inputs image data including the object to be judged through, for example, a CCD camera equipped in the fluorescence microscope. This image data is stored in the image storage unit. The determination target extraction means extracts one determination target from the stored image data. The extracted determination target has variations in size. Therefore, in order to make the sizes of the determination objects uniform, the input data creating means normalizes the size to, for example, the number of pixels of i × j. In this way, the input data is created. On the other hand, the neural network processing means has a neural network that has been preliminarily learned so as to output the type of the determination target object corresponding to the input data. Therefore, when input data is input to the neural network processing means, a signal corresponding to the type of the determination target object is output to the type output unit. The type output unit outputs the type of the determination target object based on the output from the neural network processing means.

[0010]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to FIGS.

In the embodiment shown in FIG. 1, an image input section 10 for inputting image data including an object to be judged, an image storage section 12 for storing the image data input by the image input section 10, and an image storage section 12 are stored. The determination target object extracting means 14 for extracting the determination target object from the image data thus obtained, and the size of the determination target object extracted by the determination target object extracting means 14 are normalized to a size consisting of a predetermined number of pixels. Input data creating means 16 for creating input data by this, and neural network processing means 18 for performing neural network processing on the input data created by the input data creating means 16.
And a type output unit 20 that outputs the type of the determination target based on the processing result of the neural network processing unit 18.

The image input unit 10 includes a fluorescence microscope 30 for observing fluorescence from an object to be judged, a CCD camera 32 for photoelectrically converting an image captured by the fluorescence microscope 30 and outputting it as analog image data, and a CCD camera 32. The A / D converter 34 converts the input analog image data into digital image data and outputs the digital image data to the image storage unit 12. An example of this digital image data is shown in FIG.
In FIG. 2, the digital image data is composed of I × J pixels, and S in the figure is one of the determination objects.

The image storage unit 12 is composed of an external storage device such as a floppy disk or a hard disk, or a RAM for video signals, and stores image data for each pixel as a density of, for example, 256 gradations.

The type output unit 20 is composed of a CRT or a printer which displays the type of the judgment object output from the neural network processing means 18.

The judgment object extracting means 14, the input data creating means 16 and the neural network processing means 18 are
For example, it can be realized by a computer and a program that operates the computer. The neural network processing means 18 can also be realized by a neurochip.

The determination object extracting means 14 sets the image data to "0" below a certain threshold value and to "1" above the threshold value according to the density for each pixel.
Then, from the binary image data thus obtained, only one determination target object is extracted and cut out by so-called labeling processing. The labeling process is a process in which a portion in which pixels of "1" are connected in a binary image is one component.

As shown in FIG. 3, the input data creating means 16 normalizes the cut-out decision object to a size composed of, for example, the number of pixels of i × j. In this way, the sizes of the objects to be judged are made uniform and become the input data of the neural network processing means 18. That is,
The input data is i × j data consisting of “0” or “1” as shown in FIG.

The neural network processing means 18 is
It has a neural network unit 36 as shown in FIG.

The neural network section 36 is composed of an input layer 36a to which the input data created by the input data creating means 16 is input, an intermediate layer 36b, and an output layer 36c. Each layer is made up of constituent elements called units, and a neural network is formed by connecting the units.

Each unit of the input layer 36a is composed of an intermediate layer 36a.
It is all associated with each unit of b. The number of units of the input layer 36a is the number corresponding to the number of pixels of input data, and can be set arbitrarily. In this embodiment, as shown in FIG. 5, i × j units are provided. In addition, each unit of the intermediate layer 36b corresponds to the output layer 3
All of them are connected to the respective units of 6c.

The number of units of the output layer 36c is set to the number of categories to be identified. In this embodiment, since the type of the determination target is determined, the output layer 36c is composed of five units as shown in FIG.

However, the coupling method between the input layer 36a and the intermediate layer 36b, the intermediate layer 36b and the output layer 36c, and the number of layers and the number of units of the intermediate layer 36b can be arbitrarily set according to the design.

Here, the number of units of the input layer 36a is N
i, the number of units of the intermediate layer 36b is Nh, and the number of units of the output layer 36c is No. The magnitude of the coupling from the unit n of the input layer 36a to the unit m of the intermediate layer 36b, that is, the coupling load is Winm, and the coupling load from the unit n of the intermediate layer 36b to the unit m of the output layer 36c is Whnm.

The input layer 36a performs a process of simply outputting the input data as it is to the intermediate layer 36b.
Further, the intermediate layer 36b performs a calculation using the following equation (1) based on the data from the input layer 36a. Here, Oin is the output from the unit n of the input layer 36a, and Ohm is the intermediate layer 36a.
The output from unit m of b, Thm, is the threshold in unit m of intermediate layer 36b.

[0025]

[Equation 1]

Here, f is a conversion function, and a sigmoid function is generally used as shown in FIG.

[0027]

[Equation 2]

Similarly, in the case of the intermediate layer 36b to the output layer 36c, the conversion represented by the following equation (3) is performed. Here, Ohn is the output from the unit n of the intermediate layer 36b, Oh
om is an output from the unit m of the output layer 36c, and Tom is a threshold value in the unit m of the output layer 36c.

[0029]

[Equation 3]

In such a configuration, in order to obtain the relationship between the input data of the image input to the input layer 36a and the output from the output layer 36c, that is, the result of the image type determination,
The bond strength between each unit is obtained by learning in advance.

In the image shown in FIG. 3, there is a difference in shape as shown in FIGS. 7 (A) and 7 (B). here,
When input data of a certain image is input to the input layer 36a, if the input data of this image is of the type shown in FIG.
7 is output from the first unit of the output layer 36c, and in the case of the type shown in FIG. 7B, the second unit of the output layer 36c is trained to output 1.

If such a learned network is used, when the input data of a certain unknown image is given, it is possible to determine the type by determining from which unit the maximum value is output. .

Next, the operation of this embodiment will be described with reference to the flowchart of FIG.

.. The patient's serum and cell nuclei are reacted with each other, and the determination target to which a fluorescent label is added is set on the fluorescence microscope 30.

.. The image obtained through the fluorescence microscope 30 is picked up by the CCD camera 32, photoelectrically converted, and output to the A / D conversion unit 34 as an analog signal.

.. The A / D conversion unit 34 is the CCD camera 3
The image data from 2 is input (step S10).

.. The A / D conversion unit 34 is the CCD camera 3
The analog image data from 2 is converted into digital image data, and the digital image data for I × J pixels is transferred to the image storage unit 12 (step S11).

Here, the image input from the CCD camera 32 is a color image, but since the fluorescence reaction of the antinuclear antibody can be represented only by the density value of green color, A /
The D conversion unit 34 targets only the green color signal.

.. In the image storage unit 12, the A / D conversion unit 3
The image data from 4 is stored as a combination of I × J pixels. Further, any one pixel has a gray scale of 2 ^P and has a density value in the range of 0 to 2 ^(P-1) . Therefore,
In this embodiment, since the density value is represented by 8 bits, 2 ⁸ = 2
56 gradations, that is, the density value is 0 to 255.

.. The determination target extraction means 14 sets the image data to be "0" below a certain threshold value and "1" above the threshold value according to the density for each pixel. Then, from the binary image data thus obtained, only one determination target object is extracted and cut out by so-called labeling processing (step S12).

.. As shown in FIG. 3, the input data creating unit 16 normalizes the cut-out determination object to a size including the number of pixels of i × j (step S13). Then, the i × j pixels are input as input data to the neural network unit 36 (step S1).
4).

.. The neural network unit 36 performs a predetermined process on the basis of the data of i × j pixels input to each unit of the input layer 36a to perform a predetermined process.
The processing result is output from c to the type output unit 20.

.. The type output unit 20 determines the type based on the output from the neural network unit 36. In this embodiment, the number of units of the output layer 36c is five. The output from each unit is 0,
If it is 0, 1, 0, 0, it is determined that the input data is (C) shown in FIG. 8 (step S15).

Further, the neural network unit 36 can also make a judgment for ambiguous types of input data. For example, 0.0, 0.65, 0.45, 0.0, 0. for input data for which it is difficult to determine (B) or (C) shown in FIG.
The output will be 0. That is, the input data is (B)
Alternatively, it can be determined to be a type close to (C), and a type close to (B) among them.

Next, the operation of the neural network unit 36 will be described with reference to the flowchart of FIG.

.. In the input layer 36a of the neural network unit 36, the i × j pieces of data shown in FIG. 4 are input from the unit 0 to the unit i × j shown in FIG. 5 (step S21).

Then, each unit of the input layer 36a outputs the input data to the intermediate layer 36b (step S22).

.. Each unit of the intermediate layer 36b is the input layer 3
Based on the data from 6a, the sum of inputs is calculated (step S24).

For example, in the case of the unit m, the sum Ym 'of the inputs is calculated using the following equation (4).

[0050]

[Equation 4]

.. Each unit of the intermediate layer 36b obtains the difference between the sum of inputs and the threshold value (step S25).

For example, in the case of the unit m, the difference Ym from the threshold value Thm is obtained using the following equation (5).

Ym = Ym'-Thm (5)

.. Each unit of the intermediate layer 36b obtains an output value by performing conversion by the function f of the equation (2) based on the difference between the sum of inputs and the threshold value (step S26).

For example, in the case of the unit m, the output value Ohm is obtained using the following equation (6).

Ohm = f (Ym) (6)

.. When the calculation of the output value is completed in all the units of the intermediate layer 36b (step S23), each unit of the output layer 36c obtains the sum of the inputs (step S28).

For example, in the case of the unit m, the sum Xm 'of the inputs is calculated using the following equation (7).

[0059]

[Equation 5]

.. Each unit of the output layer 36c obtains the difference between the sum of the inputs and the threshold value (step S29).

For example, in the case of the unit m, the difference Xm from the threshold value Tom is obtained using the following equation (8).

Xm = Xm'-Tom (8)

.. Each unit of the output layer 36c obtains an output value by performing conversion by the function f of the equation (2) based on the difference between the sum of inputs and the threshold value (step S30).

For example, in the case of the unit m, the output value Oom is obtained using the following equation (9).

Oom = f (Xm) (9)

.. The above process is repeated until the calculation of the output value is completed in all the units of the output layer 36c (step S27).

As described above, according to this embodiment,
Since the neural network processing means 18 makes a quantitative type determination, there is no risk of individual differences in the determination result due to different determiners, and it is possible to automate the type determination, and even a beginner can accurately and accurately It is possible to quickly determine the type of antinuclear antibody.

[0068]

According to the present invention, the type of antinuclear antibody can be judged by the neural network processing means without the visual inspection by the examiner. Therefore, even a beginner can perform accurate type determination automatically and quickly without causing individual differences in determination results.

Further, even for antinuclear antibodies having ambiguous shapes similar to a plurality of types, the ambiguity can be quantified, and accurate and reproducible judgment results can be obtained.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing image data obtained by the CCD camera of FIG.

FIG. 3 is an explanatory diagram showing a state where the image data of FIG. 2 is normalized.

FIG. 4 is an explanatory diagram showing a state in which input data of a neural network unit is created from the normalized image data of FIG.

FIG. 5 is an explanatory diagram showing a neural network unit according to an embodiment of the present invention.

FIG. 6 is an explanatory diagram showing a sigmoid function.

FIG. 7 is an explanatory diagram showing types of antinuclear antibodies.

FIG. 8 is a flowchart showing the operation of the embodiment of the present invention.

FIG. 9 is a flowchart showing the operation of the neural network unit according to the embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 10 image input unit 12 image storage unit 14 determination target object extraction unit 16 input data creation unit 18 neural network processing unit 20 type output unit

Claims (3)

[Claims] 1. An image input unit for inputting image data including an object to be judged, an image storage unit for storing the image data input by the image input unit, and an image data stored in the image storage unit. Determination target object extracting means for extracting the determination target object; and input data creating means for creating input data by normalizing the size of the determination target object extracted by the determination target object extracting means to a predetermined number of pixels. A neural network processing means for performing a neural network processing on the input data created by the input data creating means, and a kind output section for outputting the kind of the judgment object based on the processing result of the neural network processing means. An apparatus for determining a fluorescent antibody, which comprises: 2. The neural network processing means determines an input layer having a unit number equal to the number of pixels normalized by the input data creating means, and a number of types of determination objects output by the type output section. The fluorescent antibody determination apparatus according to claim 1, further comprising an output layer having an equal number of units. 3. The image input unit, a fluorescence microscope for observing fluorescence from an object to be judged, and a CC for photoelectrically converting an image captured by the fluorescence microscope and outputting it as analog image data.
3. The fluorescence according to claim 2, comprising a D camera and an A / D converter which converts analog image data input from the CCD camera into digital image data and outputs the digital image data to the image storage section. Antibody determination device.