JPH03102477A - Radial image processing device - Google Patents

Abstract

PURPOSE:To further improve accuracy while using a device by correcting a data processing by a neural network and inputting corrective information to perform re-learning to the neural network after packaging a learned neural network to the device. CONSTITUTION:The learned neural network is packaged in an arithmetic part, and the final image processed with the data of a radiation field, an image reading condition, and abnormal shading, etc., obtained by computation is outputted to an output part, and is displayed as a visible image. When a display image is an inappropriate one, the corrective information to correct the data processing by the neural network and to perform the re-learning are inputted to the neural network with a corrective input means to obtain an appropriate image. In such a way, when the data processing by the neural network is performed inappropriately after packaging the neural network, a user can correct the processing and perform the re-learning. Thereby, the accuracy can be improved further while using the device.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Application in Industry) The present invention relates to a radiation image processing device, and more specifically, to a radiation image processing device, and more particularly, to a radiation image processing device, which uses radiation image data to recognize an irradiation field, and to judge a book reading based on pre-read image data. The present invention relates to a radiation image processing apparatus that uses a neural network to perform data processing such as determining image reading conditions, determining image processing conditions, and detecting abnormal shadows.

(Prior Art) It is practiced in various fields to read a recorded radiation image to obtain an image signal, perform appropriate image processing on the image signal, and then reproduce and record the image. For example, an X-ray image is recorded using an X-ray film with a low gamma value designed to be compatible with later image processing, and the X-ray image is read from the film on which it is recorded and converted into an electrical signal. By performing image processing on this electrical signal (image signal) and then reproducing it as a visible image in a photocopy, etc., it is possible to obtain a reproduced image with good image quality performance such as contrast, sharpness, and graininess. 5 6 (see Japanese Patent Publication No. Sho BI.-5193).

In addition, the applicant has proposed radiation (X-rays, α-rays, β-rays, γ-rays, etc.)
When exposed to a ray, electron beam, ultraviolet ray, etc., a part of this radiation energy is accumulated, and then when exposed to excitation light such as visible light, a stimulable phosphor that exhibits stimulated luminescence according to the accumulated energy ( Radiation image information of a subject such as a human body is recorded on a sheet of stimulable phosphor, and this stimulable phosphor sheet is scanned with excitation light such as a laser beam to make it shine. The resulting stimulated luminescence is read electronically to obtain an image signal, and based on this image data, a radiation image of the subject is recorded as a visible image on a recording material such as a photographic light-sensitive material, a CRT, etc. A radiation image recording and reproducing system has already been proposed (Japanese Unexamined Patent Publication No. 55-1.
No. 2429, No. 5B-11395 No. 55-16347
No. 2, No. 58-1041345, No. 55-11H4
0 etc.).

This system has the practical advantage of being able to record images over a much wider range of radiation exposure than conventional silver halide radiography systems.

In other words, in stimulable phosphors, it is recognized that the amount of emitted light that is stimulated and emitted by excitation after storage is proportional to the amount of radiation exposure over a wide range. Even if the amount of radiation exposure varies considerably, the amount of stimulated luminescence light emitted from the stimulable phosphor sheet is read by a photoelectric conversion means by setting the reading gain to an appropriate value and converted into an electrical signal. By using this electrical signal to output the radiation image as a visible image to a recording material such as a photographic abrasive or a display device such as a CRT, a radial image that is not affected by fluctuations in the radiation exposure amount can be obtained. be able to.

In the above system, the stimulable phosphor sheet I is scanned in advance with a low-level light beam before reading it under optimal reading conditions depending on the dose of radiation applied to the stimulable phosphor sheet and obtaining an image signal. Then, pre-reading is performed to read the outline of the radial image recorded on this sheet, the pre-reading image signal obtained by this pre-reading is analyzed, and then the sheet is irradiated with a high 78 level light beam and scanned. However, there is also a system configured to perform actual reading in which an image signal is obtained by reading the radiation image under optimal reading conditions.

Here, reading conditions are a general term for various conditions that affect the relationship between the amount of stimulated luminescence light and the output of the reading device during reading, such as reading gain and scale factor that determine the relationship between human output. Alternatively, it refers to the power of excitation light during reading.

Also, the high level and low level of the light beam are, respectively.
The energy of the light beam irradiated per unit area of the sheet or the energy of stimulated luminescence light emitted from the sheet depends on the wavelength of the light beam (
(has a wavelength sensitivity distribution), the energy of the light beam irradiated per unit area of the sheet is weighted by the wavelength sensitivity, and then the weighted energy is large or small, and the level of the light beam is changed. Methods include methods of using light beams of different wavelengths, methods of changing the intensity of the light beam itself emitted from a laser light source, etc., and methods of changing the intensity of the light beam by inserting or removing an ND filter, etc. in the optical path of the light beam. Various known methods can be used, such as a method of changing the scanning density by changing the beam diameter of the light beam, and a method of changing the scanning speed.

In addition, regardless of whether the system performs this pre-reading or the system that does not, the obtained image signal (including the pre-read image signal) is analyzed and the optimal image processing conditions are determined when applying image processing to the image signal. Some systems let you decide. The term "image processing conditions" as used herein is a general term for various conditions when performing processing on an image signal that affects the gradation, sensitivity, etc. of a reproduced image based on the image signal. The method of determining the optimal image processing conditions based on this image signal is not limited to systems using stimulable phosphor sheets, and for example, image signals are obtained from radiation images recorded on recording sheets such as conventional X-ray films. It is also applied to the system.

The calculation for determining reading conditions and/or image processing conditions (hereinafter referred to as reading conditions, etc.) based on the above image signals (including pre-read image signals) is performed by statistically processing a large number of radiographic images in advance. The algorithm has been determined based on the results (see, for example, Japanese Patent Laid-Open No. 60-185944 and Japanese Patent Laid-Open No. 61-280163).

This conventionally used algorithm generally calculates a histogram of an image signal, and calculates various values such as the maximum value, minimum value, and value of the image signal at the point where the frequency of appearance of the image signal is maximum on the histogram. The feature points are determined, and the reading conditions, etc. are determined based on the feature points.

However, in recent years, a concept called neural networks, which is completely different from the above-mentioned algorithms, has emerged and is being applied to various fields.

This neural network can communicate between each unit within the neural network by inputting information (teacher signal) on whether the output signal output when a certain input signal is given is a correct signal or an incorrect signal. It is equipped with an error back propagation learning function that corrects connection weights (synaptic connection weights), and by repeatedly ``learning'',
This can increase the probability of outputting a correct answer when a new signal is input. (For example, rD.E.R.
umelhart. G. E. Hinton and R.
j. William Tls: Learning repr.
esentati. ons by back-prop
aga. tf. ngerrors. Nature, 82
3-9, 533-356.1986aJ, r Hideki Aso 2-back baggation Computrol N
0.24 53-60J, r Kazuyuki Aihara, Neural Computer Tokyo Denki University Press).

This neural network can also be applied to determining reading conditions, etc., and reading conditions, etc. can be output by inputting an image signal etc. to the neural network.

In addition, in recent years, in systems using the above-mentioned X-ray film, stimulable phosphor, etc., especially in systems configured for medical diagnosis of the human body, reproduced images with good image quality suitable for simple observation (diagnosis) have been developed. In addition to image recognition, automatic image recognition has been carried out (for example, see Japanese Patent Laid-Open No. 125481/1981).

Automatic image recognition here refers to the operation of extracting a desired pattern from a complex radiographic image by performing various processes on the image data. , an operation to extract, for example, a shadow corresponding to a tumor from a very complex image containing a mixture of circular patterns.

A target pattern (for example, a tumor shadow) in such a complex radiographic image (for example, a chest X-ray image of a human body)
By extracting the pattern and reproducing and displaying a visible image clearly showing the extracted pattern, it is possible to assist the observer in observation (for example, assist the doctor in diagnosis).

(Problem to be solved by the invention) When the above neural network is used to determine reading conditions, etc., it becomes possible to gradually determine correct reading conditions, etc. by repeating ゜learning'' in advance. thing,
For example, in a system that handles X-ray images of human shoulders, various variations such as right shoulder and left shoulder (image inversion), enlarged and reduced images, upright and sideways and inverted positions, and positional deviations are detected. In order to construct a neural network that can determine the correct reading conditions even if there is an image in which The storage capacity of the storage device for storing the weights of is also required to be enormous, and there is also the problem that learning must be repeated an extremely large number of times when performing ゜learning'.

Therefore, in view of the above-mentioned problems, the present applicant has developed a system for radiographic image reading conditions that can determine correct reading conditions using a neural network equipped with a relatively small number of units even when input images vary. /Or proposed an image processing condition determination device. (Patent Application No. 102015/1999) This is applied to a system that uses a stimulable phosphor sheet and performs so-called pre-reading13 14 . The excitation light is applied to the stimulable phosphor sheet again based on the first image signal representing the radiation image obtained by applying the excitation light and reading the stimulated luminescence light emitted from the stimulable phosphor sheet. The reading conditions for obtaining a second image signal representing the radiation image by irradiating the stimulable phosphor sheet and reading the stimulated luminescence light emitted from the stimulable phosphor sheet and/or the image A radiographic image reading condition and/or image processing condition determination device for determining image processing conditions for performing processing includes: a storage means for storing a standard pattern of a radiographic image; a signal conversion means for converting a radiographic image into the standard pattern and obtaining a converted image signal carrying the converted radiographic image; The present invention is characterized in that it includes a condition determining means consisting of a neural network as an output.

Another reason is that it is not limited to stimulable phosphor sheets and requires image processing conditions.
A radiation image processing condition determination device that determines image processing conditions for performing image processing on an image signal based on an image signal representing a radiation image, comprising: a storage means for storing a standard pattern of a radiation image; a signal converting means for converting a radiographic image carried by an image signal into the standard pattern and obtaining a converted image signal carrying the converted radiographic image; and a signal converting means that receives the converted image signal and outputs the image processing conditions. The invention is characterized by comprising a condition determining means comprising a neural network.

However, for example, radiological images of the human body are extremely complex;
In particular, in the case of diagnostic equipment, even if the reading conditions and image processing conditions are determined by the data processing capacity learned in advance by a neural network as described above, users such as specialized doctors may You may want to correct this. In such a case, even after the neural network has been installed in the device, it is desirable to have the neural network re-learn at a hospital or the like.

Regarding the abnormal shadow detection device, Japanese Patent Application Laid-open No. 62-
125481, for example, a chest X-ray image of a human body is scanned using a specific filter that does not change depending on the position on the image, a circular pattern and a linear pattern are extracted, and this circular pattern is used as a tumor candidate or a linear pattern. A device is described that displays blood vessels as shadows of blood vessels.

However, radiological images of the human body are extremely complex; for example, in a chest X-ray image, when a tumor shadow appears just beside the ribs, the pattern differs when it appears between the ribs. As mentioned above, devices with simple configurations often fail to recognize patterns that are not tumors, or recognize them as tumors.For example, when automatic image recognition is performed in a medical system, doctors When a doctor observes a reproduced image and makes a diagnosis, his or her attention tends to be drawn to the automatically recognized patterns, and if there is an omission in the automatic recognition, there is a high possibility that the doctor will miss it, which poses a serious problem.

In order to avoid this, it is conceivable to devise a filter that scans the radiation image to extract all patterns that are considered to be candidates for tumors or the like. However, if all patterns considered to be tumor candidates are extracted in this way, many patterns (noise) that are not tumors will also be extracted. Although this is better than missing something, too much noise reduces the reliability of the automatic recognition system and can even make the diagnosing doctor fatigued.

In addition, for example, doctors who observe (diagnose) chest X-ray images recorded on X-ray film, which have not been automatically recognized until now,
Based on previous knowledge and experience, it is a fact that no matter where the tumor is located on the image, the tumor can be extracted quite accurately even if the pattern 17 18 is slightly deformed.

On the other hand, even if the image forming observation systems are similar to each other,
The average properties of images may vary depending on the facility where the system is installed, and judgments may vary slightly depending on the observer observing the image. For example, chest
In X-ray imaging diagnostic systems, the energy and irradiation amount of X-rays at the time of imaging differs depending on the hospital where the system is installed, and the average chest X-ray image is slightly different. Depending on the doctor who observes the X-ray image, subtle discrepancies may occur, for example, in determining whether the X-ray image is normal or whether re-examination is necessary.

Therefore, in automatic image recognition, it is desirable to reduce noise as much as possible by performing advanced processing based on past knowledge that cannot be performed with simple filtering processing from candidates such as tumors extracted without omission. In addition, this automatic image recognition system accumulates further knowledge depending on the facility where it is installed and the observer, and is forced to re-learn to be more suitable for the facility, observer, and other users.
It is desirable to have a system that changes itself.

In view of these demands, the present invention provides neural network I.
Even after the system is implemented in a device, in hospitals, etc., users may inappropriately change the reading conditions, image processing conditions, or detected abnormal shadows determined by the processing capacity obtained through pre-learning of the neural network. If it is determined that this is the case, it can be corrected and re-trained, and as a result,
It is an object of the present invention to provide a radiation image processing device whose accuracy can be further improved during use.

(Means for Solving the Problems) A radiation image processing apparatus according to the present invention recognizes an irradiation field, determines image reading conditions for main reading based on pre-read image data, determines image processing conditions, and detects abnormalities based on radiation image data. In a radiation image processing device that performs data processing such as shadow detection using a neural network, 19 20 After implementing the trained neural network in the device,
An input means is provided so that the user can input corrective information to the neural network even after implementation so that the user can correct the data processing by the neural network and have it re-learn if the data processing by the neural network is inappropriate. It is characterized by having the following.

More specifically, a radiation image representing the radiation image obtained by shining excitation light onto a stimulable phosphor sheet on which a radiation image is recorded and reading stimulated luminescence light emitted from the stimulable phosphor sheet. Based on the first image signal, the stimulable phosphor sheet is irradiated with excitation light again and the stimulated luminescent light emitted from the stimulable phosphor sheet is read to obtain a second image signal representing the radiation image. In a radiation image reading condition and/or image processing condition determination device for determining reading conditions for the actual reading and/or image processing conditions for performing image processing on the obtained second image signal, a standard pattern of the radiation image is determined. a storage means for storing; a signal conversion means for converting the radiation image carried by the first image signal into the standard pattern to obtain a converted image signal carrying the converted radiation image; condition determining means comprising a neural network that receives the converted image signal as input and outputs the reading condition and/or the image processing condition; and based on the reading condition and/or the image processing condition output by the condition determining means. a display means for displaying a radiation image read and/or image-processed by the display means as a visible image; and a display means for displaying correction information for converting the visible image displayed by the display means into a more appropriate visible image to the condition determining means. and an input means for inputting, and the condition determining means includes:
Based on the correction information input from the input means,
Parallel image reading conditions and/or image processing conditions, characterized in that they have a learning function that changes the calculation of a neural network that uses the converted image signal as an input and outputs the reading conditions and/or the image processing conditions. A determining device 21 22 is provided.

Further, in a radiation image processing condition determination device for determining image processing conditions for performing image processing on an image signal based on an image signal representing a radiation image, a storage means for storing a standard pattern of a radiation image is provided. , a signal converting means for converting a radiographic image carried by the image signal into the standard pattern and obtaining a converted image signal carrying the converted radiographic image; and using the converted image signal as input and the image processing conditions. a condition determining means comprising a neural network whose output is the reading condition and/or the reading condition outputted by the condition determining means;
or display means for displaying a radiation image read and/or image-processed based on the image processing conditions as a visible image; and correction for converting the visible image displayed by the display means into a more appropriate visible image. an input means for inputting information into the condition determining means, the condition determining means comprising:
Based on the correction information input from the input means,
The present invention provides a radiation image processing apparatus characterized by having a learning function for changing the calculation of a neural network that receives the converted image signal as an input and outputs the reading conditions and/or the image processing conditions.

Furthermore, in the abnormal shadow detection device that detects an abnormal shadow on the radiation image based on image data representing a radiation image of a subject, a first method that extracts candidates for the abnormal shadow based on the image data is provided. candidate extracting means; feature amount calculating means for calculating a plurality of feature amounts for each candidate based on the image data in the vicinity of the candidate extracted by the first candidate extracting means; a second candidate extraction means for extracting a candidate with a high probability of being the abnormal shadow from among the candidates based on the plurality of obtained feature quantities; a display means for displaying a visible 23 24 image clearly showing the candidate based on the image; and an input means for inputting information as to whether the candidate clearly shown in the visible image is the abnormal shadow; The second candidate extracting means has a learning function for changing the calculation for extracting the highly probable candidate from the candidates extracted by the first candidate extracting means, based on the information input from the input means. An object of the present invention is to provide an abnormal shadow detection device characterized by the following features.

Here, the second candidate extracting means may include a neutral network having an error backpropagation learning function using the information as a teacher signal.

Regarding the above-mentioned "standard pattern of radiographic images," the standard pattern is not limited to a specific pattern in the present invention, but may be determined based on the design philosophy, etc. This is a vehicle clause that can be determined arbitrarily.

Furthermore, in the present invention, "converting to the standard pattern" is not limited to a specific operation, but includes, for example, inversion of the radiation image carried by the first image signal,
Refers to rotation, position adjustment, enlargement, reduction, etc.

(Operations and Effects) The present invention provides a radiographic image processing device using a neural network as described above, in which a user inputs correction information into the neural network after the learned neural network is installed in the device. Since the input means is provided so that the neural network is not properly processed, the user can correct it and re-learn it, thereby improving accuracy during use. It can be improved further.

(Example) Hereinafter, the structure of the present invention will be explained with reference to the drawings.

FIG. 1 is a block diagram for explaining the basic configuration of the present invention.

25 inputted into the radiation image processing device by the input unit 26 The radiation image data is sent to a calculation section within this device.

This calculation unit is equipped with a trained neural network, which recognizes the irradiation field based on the input image data, determines the image reading conditions for main reading based on the pre-read image data, and determines the image processing conditions. , data processing such as abnormal shadow detection is performed using the neural network. A final image processed using data such as the irradiation field, image reading conditions, image processing conditions, and abnormal shadows calculated here is output to the output unit and displayed as a visible image.

A user such as a doctor who uses this radiation image processing device looks at the image displayed on the display device and does nothing if the image is appropriate. If the displayed image is inappropriate, correction information for correcting and relearning the data processing by the neural network is input into the neural network I network by the correction input means in order to make it appropriate.

In this way, after implementing a neural network, the user can correct and retrain the data processing by said neural network if it is inappropriate, thereby further improving accuracy during use. Next, embodiments of the present invention will be described in detail with reference to the drawings. Here, an example will be described in which an X-ray image of the shoulder of a human body is handled using the above-mentioned stimulable phosphor sheet.

FIG. 5 is a schematic diagram of an example of an X-ray imaging apparatus.

X-rays 3 are irradiated from the X-ray source 2 of this X-ray imaging device 1 toward the shoulder 4a of the human body 4, and the X-rays 3a that have passed through the human body 4 are irradiated onto the stimulable phosphor sheet 11. As a result, a transmitted X-ray image of the shoulder 4a of the human body is accumulated and recorded on the stimulable phosphor sheet 11.

2A and 2B are diagrams schematically showing an example of a shoulder X-ray image stored and recorded on a stimulable phosphor sheet.

Figures 2A and 2B show X-ray images of the right shoulder and left shoulder, respectively, and in addition to the subject area 5 where the shadow of the human body is recorded, the X-rays 27 28 that are not illuminated by the subject 11 are directly exposed to the stimulable phosphor sheet A direct X-ray section 6 is formed that irradiates one area.

FIG. 6 is a perspective view showing an example of an X-ray image reading device and an example of a computer system incorporating an example of the present invention. This system uses the aforementioned stimulable phosphor sheet and performs pre-reading.

The stimulable phosphor sheet 1 on which the X-ray image has been recorded is first scanned with a weak light beam to emit only a portion of the radioactive ray energy accumulated in the stimulable phosphor sheet 1 and read it in advance. It is set at a predetermined position in the pre-reading means 100 that performs.

The stimulable phosphor sheet 11 set at a predetermined position is conveyed (sub-scanning) in the direction of arrow Y by a sheet conveying means 13 such as an endless belt driven by a motor l2.
be done. On the other hand, a weak light beam E5 emitted from a laser light source 14 is driven by a motor 23 and is reflected and deflected by a rotating polygon mirror 16 that rotates at high speed in the direction of the arrow. The light is incident on the stimulable phosphor sheet 1l and main-scanned in the direction of arrow X, which is substantially perpendicular to the direction of sub-scanning (direction of arrow Y). This light beam l5 of the stimulable phosphor sheet 11
From the irradiated area, stimulated luminescence light 19 is emitted with an amount of light according to the radiographic image information stored and recorded, and this stimulated luminescence light 19 is guided by a light guide 20 and directed to the photo 1
・Detected photoelectrically by multiplier (photomultiplier tube) 2. The light guide 20 is made by molding a light-guiding abrasive material such as an acrylic plate, and has a linear entrance end face 2Da that extends along the main scanning line on the stimulable phosphor sheet 11. The light-receiving surface of the photomultiplier 21 is coupled to the annularly formed exit end surface 20b. The light guide 2 from the end face 20a of the person I11
The photostimulated luminescent light source 9 that has entered the interior of the light guide 20 travels through the interior of the light guide 20 through repeated total reflection, exits from the exit end face 20b, and is received by the foI multiplier 21, where the photostimulated luminescent material 9 represents a radiation image. The amount of emitted light 19 is converted into an electrical signal by a photomultiplier 21.

The analog output signal S output from the photomultiplier 21 is logarithmically amplified by the logarithmic amplifier 2B, digitized by the 29 30 A/D converter 27, and converted into a pre-read image signal S.
p is obtained. The signal level of this pre-read image signal Sp is
It is proportional to the logarithm of the amount of stimulated luminescence light emitted from each pixel on sheet ↓1.

In the above-mentioned pre-reading, reading conditions such as the voltage value applied to the photomultiplier 2 and the amplification factor of the logarithmic amplifier 26 are set so that the radiation energy accumulated in the stimulable phosphor sheet 11 can be read over a wide range. etc. are stipulated.

The obtained pre-read image signal Sp is sent to the computer system 4
It is input to 0. This computer system 40 includes an example of the present invention, and includes a main body part 41 in which a CPU and an internal memory are built-in, and a drive part 4 into which a floppy disk as an auxiliary memory is inserted and driven.
2. It consists of a keyboard 43 for an operator to input necessary instructions to the computer system 40, and a CRT display 44 for displaying necessary information.

In this computer system 40, the reading conditions for main reading, that is, the sensitivity and contrast 1. for main reading, are determined as described later, and the voltage value to be applied to the photomultiplier 21', for example, is determined according to the sensitivity and latitude. The amplification factor etc. of the logarithmic amplifier 26' are controlled.

Contrast here corresponds to the light intensity ratio of the most intense stimulated luminescence light to the weakest stimulated luminescence light that is converted into an image signal during actual reading, and sensitivity is the ratio of the amount of stimulated luminescence light to the weakest stimulated luminescence light that is converted into an image signal during actual reading. This refers to the photoelectric conversion rate that determines the level of image signal generated from stimulated luminescence light.

The stimulable phosphor sheet hlL' for which pre-reading has been completed is set at a predetermined position in the main reading means 100', and the sheet 11' is scanned by a light beam 15' that is stronger than the light beam used for the above-mentioned pre-reading, and the pre-read image signal is read. An image signal is obtained under the reading conditions determined based on Sp, but the actual reading means 1
Since the configuration of 00' is substantially the same as the configuration of the pre-reading means 100, the components corresponding to the respective components of the pre-reading means 100 are indicated by adding a dash to the number used in the pre-reading means 100. , the explanation is omitted.

31 32 The image signal SQ obtained by being digitized by the A/D converter 27' is sent to the computer system 40 again.
is input. Appropriate image processing is performed on the image signal SQ within the computer system 40, and the image signal subjected to this image processing is sent to a reproduction device (not shown), and an X-ray image based on this image signal is reproduced and displayed in the reproduction device. Ru.

Next, in the computer system 40, the pre-read image signal Sp
A method for determining reading conditions for main reading based on the following will be explained.

Since we are dealing with shoulder X-rays here, Figure 2A,
As shown in FIG. 2B, an image that is approximately left and right reversed may be obtained. Therefore, the image of the right shoulder (2nd A) is
) or the left shoulder image (FIG. 2B). In this embodiment, the image of the right shoulder (FIG. 2A) is considered to be the standard pattern according to the present invention.

3A and 3B are diagrams representing a standard pattern and an inverted pattern, respectively, stored in advance in computer system 40. FIG.

These standard patterns and inverted patterns are divided into a first region 7 having an average value of the pre-read image signal Sp corresponding to the object part 5 of the X-ray image shown in FIGS. 2A and 2B, and a second region
and a second region 8 having an average value of the pre-read image signal Sp read from the direct X-ray unit 6 of the X-ray image shown in FIGS.

When the pre-read image signal Sp is input into the computer system 40, this pre-read image signal Sp and the third
Pattern matching is performed with each of the standard pattern shown in Figure A and the inverted pattern shown in Figure 3B, and whether the input pre-read image signal Sp is a signal representing the X-ray image of the right shoulder or the X-ray image of the left shoulder is determined. It is determined whether the signal represents.

Here, this pattern matching is the standard pattern,
The signals carrying the inverted patterns (FIGS. 3A and 3B) are respectively Ss. When SR, the square value of the difference with the pre-read image signal Sp (Ss Sp
) 2. (SR Sl)) z is calculated and added over the entire image to create the 33 34 quantity QS. QR is QS = Σ (Ss 3p)2
=-(1) QR 1Σ (SR Sll) 2
...(2) is obtained, and the X-ray image carried by the pre-read image signal sp is calculated using these ffiQs,Q
It is determined that the image corresponds to the smaller value of R.

If the image carried by the pre-read image signal Sp determined in this way is an X-ray image of the left shoulder (FIG. 2B), the operation of flipping (inverting) the X-ray image on the pre-read image signal Sp I do. As a result, the neural network shown below always receives the image of the right shoulder in FIG. 2A, or the inverted image of the left shoulder with a pattern almost similar to this image.

By adjusting the input image pattern to a predetermined standard pattern (in this case, the right shoulder pattern) and inputting it to the neural network, the number of units that make up the neural network can be reduced. , the storage capacity for weighting coefficients indicating the degree of connection between units can be reduced, and learning can be performed in a relatively short time.

FIG. 4 is a diagram showing an example of a neural network equipped with an error backpropagation learning (pack propagation) function. As mentioned above, error backpropagation learning (backpropagation) is, by comparing the output of the neural network with the correct answer (teacher signal), sequentially calculates connection weights (synaptic connection way 1) from the output side to the input side.
・) is a “learning” algorithm that corrects

As shown in the figure, the first part of this neural network 1 work
The layer (input layer), beat 2 layer (intermediate layer), and third layer (output layer) are each composed of n1 units, n2 units, and 2 units. Each signal Fl, F2, input to the first layer (input layer)
...+Fnl is the pre-read image signal Sp corresponding to each pixel of the X-ray image (in the case of the left shoulder, it is an inverted image), and the two outputs y? from the third layer (output layer). r
! /2 is a signal corresponding to each sensitivity and contrast 1 and contrast 1 during main reading. The unit 1 of the first (layer i number I) is 35 ui, each input to the unit ukl is Xkl, each output is 3'7, the weight of the connection from u: to 7''1 is Wkli+
1, and each unit 7 has the same characteristic function 1. At this time, the input x: and output y: of each unit u: are:
7 = f (X7) (5). However, each unit u1 (+=
1.2, -=, ns ) Each input F1 +
F2 + "'F., 1 is inputted as it is to each unit u: (]=1,2+...,n1> without weighting. The n1 input signals F1, F2,
-, Fn+ is the final output y? while being weighted by the weight w k k + l of each connection 36 . , y2, and thereby the reading conditions (sensitivity and contrast) for actual reading are determined.

Here, a method for determining the weight WkTI of each of the above-mentioned connections will be explained. First, the weight Wk k of each connection is determined by random numbers.
An initial value of +l is given. At this time, input F
1 ”'- F n Even if 1 rewards generously, output Y”
+rY! It is preferable to limit the range of the random number so that r is a value within a predetermined range or a value close to this.

A number of X-ray images of the right shoulder or X-ray images of the left shoulder for which the optimal reading conditions are known are read as described above, and a stimulable phosphor sheet containing the recorded stimulable phosphor sheets is obtained to obtain a pre-read image signal Sp. In the case of a line image, the obtained pre-read image signal Sp is inverted, so that the n1 inputs F1+FZ+
...+Fnl is calculated. These 01 inputs F1, .
F2,...+Fnl are input to the neural network shown in FIG. 437 38 ? , the output y〒 of each unit l■u〒 is monitored.

Once each output yi is determined, the final output yi'
, yz and the teacher signal (sensitivity y and contrast bar) based on the correct reading conditions for this image.
is a learning coefficient and]l;J: a coefficient to be learned.

Here, 2 is required. In order to minimize these square errors E1 and E2, the weight W' of each connection is determined as follows.
+'+'' is corrected.The output of y will be described below, and y: is omitted here because it is the same as yi.

To minimize the squared error E1, this E1 is and (
4.) From the formula, Xl'' - ΣW1,゛ ·y, so the formula (9) becomes as follows.

Here, from equation (6), it is a function of (4+' w:: +1, so there are three 40) If we transform equation (1i) using equation (5), we get this equation (16). According to W?? (i.
1-. 2. nx) are modified.

Here, from equation (3), f' (x) −f (x) (1−f (x)
), ...(13) Therefore, by substituting equations (4) and (5) into equation (7), f' (X?) = YI' · (1).

Putting k=2 in equation (10) and substituting it into equation (10), we get y+) (12) (Work 4〉 (]4) Now, from equation (13), f' (X:) 1 Y'+ ” (1 '!l'4
) ・-(19) Therefore, by substituting this equation (Eng. 9), equations (1.2), and (14) into equation (18), ...(15) This equation (Eng. 5) Substituting into equation (8), Wl l =W
＋? η・(Vll' d)・y? (1 y?) ・y′ ...(16) 41 42 (Place itel(-1 in the equation 10) and change the equation (2o) to (1
o) Substituting into the formula, 1(y1一冫)・yj・(1−ba)・y”・(1−y锫)・C? *y”...(2
1) Substituting this equation (2{) into equation (8), replacing k − 1, we get W}〒=W{〒−η・(yi' y?)・y
?・(1 y:)・Pa・(I Y?)・y}゜W''
1...〈
22〉, and w? was corrected by equation (16). ? (i
-1.2. ...n1) is substituted into this equation (22),
W}ding(i-1.2,..., nl;j-1.
2.・Conz) will be corrected.

Theoretically, by using equations (16) and (22) and setting the learning coefficient η sufficiently small and increasing the number of learnings sufficiently, the weight w k k + l of each connection can be set to a predetermined value. Although it is possible to focus, it is not realistic to make the learning coefficient η too small because it slows down the progress of learning. On the other hand, if the learning coefficient η is set to a large value, the learning may oscillate (the weights of the connections described above may not converge to a predetermined value). Therefore, in practice, an inertia term as shown in the following equation is added to the connection weight correction amount to suppress vibration, and the learning coefficient η is set to a somewhat large value.

(For example, D.E. Rumelhart, Gi.E.H.
tnton and R. J. Williams:Le
earning internal representative
errors by error prol)agati
on In Parallel Distribution
dProcessing, Volume 1. J. L
．． McClell and. D. E. Rumel
ha. rt and The PDP Research
hGroup. MITPress, 1986b J) ΔW (t+1) - α・Δw7';+1(B +where ΔWii" (1) is the connection weight Wii"1 after modification in the t-th learning to the connection weight before modification. Weight W
7 7 ” represents the amount of correction after subtracting 43 44 α is a coefficient called an inertia term.

For example, α-0, 9η-0 as the inertia term α and the learning coefficient η.
．． Modification (learning) of the weight W〒1゛1 of each connection using 25
For example, 200,000 times, and then the weight of each connection w
k k + l is fixed at the final value. At the end of this study you will have two outputs! /;'+3'2 is a signal that correctly represents the sensitivity and contrast during actual reading.

After the learning is completed, a pre-read image signal Sp representing an X-ray image is determined as the reading condition for main reading, and this is input to the neural network shown in FIG.
The resulting output)'? +! /: is the reading condition (sensitivity and contrast) for the actual reading of the X-ray image.
It becomes a signal representing. Since this signal is obtained after learning as described above, it accurately represents the reading conditions for actual reading.

It should be noted that the neural network described above is not limited to a three-layer structure, and it goes without saying that it may have even more layers. Furthermore, it goes without saying that each layer can be configured with an arbitrary number of units depending on the number of pixels of the input pre-read image signal Sp, the accuracy of the required reading conditions, etc.

The voltage applied to the photomultiplier 21' of the main reading means 100', the amplification factor of the amplifier 2B', etc. are controlled according to the reading conditions determined by the neural network as described above, and the main reading is performed according to the controlled conditions. It will be done.

The neural network that can determine the reading conditions in this way is installed in the X-ray image reading device and has the function of determining the reading conditions as described above.

This X-ray image reading device sends the image signal obtained by reading based on the reading conditions determined above to the playback device (not shown) as described above, and inputs it again to the computer system. , Here, the image obtained by the main reading is displayed on the CRT display 44. Therefore, the user can use the keyboard 43 to cause the bi=-ral network in this device to perform error backpropagation learning again by looking at the image obtained from this book reading.

As a result, even after implementation, the neural network can be further trained by doctors in the field to improve the accuracy of reading condition determination.

In the above embodiment, the X-ray image of the shoulder is processed as shown in FIG. 3A before being input to the neural network. Third
By performing pattern matching with the pattern shown in Figure B, it is determined whether the image is a standard image (image of the right shoulder) or an inverted image (image of the left shoulder), and in the case of an inverted image (image of the left shoulder) Although the image was reversed to match the standard image (image of the right shoulder), the present invention is not limited to only shoulder images even when dealing with horizontal reversal; It can also be applied to systems that handle images of limbs, heads facing right/left, images of abdomens, etc.

Furthermore, the present invention is not limited to left-right reversal;
For example, if the stimulable phosphor cartridge is set diagonally during photography and an image signal carrying a tilted image is obtained as a result, or if the stimulable phosphor cartridge is set diagonally during reading,
The direction of the image may be different, and the image signal may represent a sideways or inverted image, and if the image signal is obtained, it may be rotated to correct the image to its normal orientation. This may be used when correcting images with different enlargement ratios (reduction ratios) such as indirect X-ray photography, or when a desired subject is recorded at the edge of the image, it may be necessary to place the subject in the center of the image. The position may also be adjusted. Further, adjustments may be made for a plurality of combinations of these.

Further, in the above embodiment, the pre-reading means 100 and the main reading means 10
0' are configured separately, but as mentioned above, the configurations of the pre-reading means 1.00 and the main reading means 1.00' are substantially the same, so the pre-reading means 100 and the main reading means 100' They may also be used together. In this case, after pre-reading, the stimulable phosphor sheet 11 may be moved back once and scanned again to perform main reading.

47 48 When the pre-reading means and the main reading means are used, it is necessary to switch the intensity of the light beam between the pre-reading and the main reading. Any arbitrary method can be used, such as a method of switching the light intensity itself.

Further, in the above embodiment, a device for determining reading conditions for reading a book using the computer system 40 was explained.
During the actual reading, reading is performed under predetermined reading conditions regardless of the pre-read image signal Sp, and the computer system 40 sets the image processing conditions when performing image processing on the image signal SO based on the pre-read image signal Sp. Alternatively, the computer system 40 may obtain both the reading conditions and the image processing conditions.

Furthermore, although the above embodiment describes a radiation image reading device that performs pre-reading, the present invention does not perform pre-reading.
It can also be applied to a radiation image reading device that performs reading equivalent to the above-mentioned main reading from the beginning. In this case, when reading, an image signal is obtained by reading under predetermined reading conditions, and based on this image signal, the computer system 4
Image processing conditions are determined within 0, and image processing is performed on the image signal according to the determined image processing conditions.

Furthermore, similar ideas can be applied to field recognition and computer diagnosis (abnormal shadow detection).

In other words, in order to avoid irradiating radiation to parts that are not necessary for observation of the subject, the radiation range is limited so that the radiation is irradiated only to the necessary parts of the subject and a part of the recording sheet 1. Photography may be performed using an irradiation field aperture, but in such cases, before analyzing the image data to determine reading conditions and image processing conditions, the irradiation field must be recognized and the image data within the irradiation field must be adjusted. Based on this, it is necessary to determine reading conditions and image processing conditions. The method of the present invention can also be applied to a system that uses a neural network to recognize this irradiation field and detect abnormal shadows, which will be described in detail below.

Next, an embodiment in which the present invention is applied to an abnormal shadow detection device will be described in detail.

4 50 FIG. 7 is a block diagram clearly showing the structure of the shadow detection apparatus according to the first embodiment.

Image data S1 representing a radiation image of a subject is input to first candidate extraction means 101.

In the first candidate extraction means 101, the input image data S1- is subjected to relatively simple processing such as real space filtering processing, and a candidate C of an abnormal shadow in the radiographic image is extracted. .

Here, "abnormal shadows" are those that are not seen in standard shadows.
For example, this refers to shadows such as tumors, calcifications, pleural thickening, pneumothorax, etc. in chest X-ray images, and this first candidate extraction means 1
01 does not need to extract all of these abnormal shadows; for example, II! Only the tumor shadow may be extracted as an abnormal shadow.

If the first candidate extraction means 101 fails to extract candidates for cumulative shadows, they will remain until the end. Therefore, the extraction method used by the first candidate extraction means 101 is as follows. , some noise (extracting something that is not the target abnormal shadow (for example, a tumor shadow) as a common shadow candidate, or a common shadow candidate that is not actually an abnormal shadow extracted in this way) It is desirable to use a method that can extract all the actual chiasm shadows even if they are mixed in.

'A 7i' obtained by the first candidate extracting means 101;
The shadow candidate C is input to the feature calculation stage 102 together with the image data S1. Feature amount calculation means]02 calculates a plurality of feature amounts Fi t F2 r . . . , 4 for each candidate C based on the image data S1 in the vicinity of this candidate C.
n is found.

Here, "feature quantity" is a general term for various quantities that reflect the certainty of each candidate C as an abnormal shadow, such as the area of each candidate C, the irregularity of the shape, and the image data within each candidate C. the average value of the image data, the variance value of the image data, the ratio of the average value of the image data in each candidate C to the average value of the surrounding image data (
contrast) etc. There are no particular limitations on what kind of feature quantity to obtain with this feature quantity calculation means 102 or how many feature quantities to obtain.
1 52 It is determined by considering the type of shadow, the accuracy to be extracted, the calculation time, etc.

The plurality of feature quantities F1r F2 , . Second candidate extraction means 103
Now, the input features fiFx. F2.
...Based on Fn, first candidate extraction means 101
A candidate T with a high probability of being an abnormal shadow (hereinafter, this candidate T may be referred to as an abnormal shadow T) is extracted from the abnormal shadow candidate C extracted in step. Although the method for extracting this abnormal shadow T is not limited to a specific method in this embodiment, for example, the above-mentioned plurality of features JTtF
1. FZ, ..., Fn are input, and ffi represents the degree of probability that the above candidate C is an abnormal shadow.
A neural network whose output is E is provided to obtain a quantity E representing the degree of the above-mentioned probability of each candidate sleeve C, and based on this quantity E, it is determined whether or not candidate C is a different hall shadow T. .

In addition, the above multiple features ffiFx+Fz+ ・・・・・・
171 is not limited to being used as an input to a neural network, but can also be used as an NN method (near
Various methods can be used, such as a method of simply comparing a threshold value with each feature amount FIF2, . . . , Fn, and a combination of these methods.

The data representing the abnormal shadow T obtained as described above is input to the display means 104 together with the image data S1. The display means 104 displays a visible image in which the common shadow T is clearly shown. Note that the method of displaying the abnormal shadow T is not limited to a specific method; for example, a method of displaying the abnormal shadow T with an arrow on the visible image, a method of displaying the abnormal shadow T by changing the color or density, a method of displaying the abnormal shadow T at the edge of the visible image, etc. , a method of displaying the location of the abnormal shadow T using letters, symbols, etc. may be used.

Further, the abnormal shadow detection apparatus of this embodiment includes an input means 105.
is provided. This input means 105 includes display means 1
The observer inputs the result observed by the observer in the visible image displayed in step 04, and information (2) as to whether or not the abnormal shadow T clearly shown in the visible image was truly an abnormal shadow.

53 54 This information is input to the second candidate extraction means 103. This second candidate extracting means 103 extracts the input information I
It has a learning function that changes the calculation for extracting the abnormal shadow T from the candidate C based on the following.

This learning function is not limited to a specific one in this embodiment, but for example, the second candidate extraction means {0
3 is equipped with a neural network having an error backpropagation learning function using the above information T as a teacher signal, and by inputting the above information T, connection weights (synaptic connection weights) are calculated. "Learning" to correct is performed (for example, rD.E.Rum
Elhart, G. E. }1inton and R.
J. Willial1s: Learning repr
esenta-tfons by back-prop
agating errors, Nature. 323
-9,583-356.1986a J, r Hideki Aso Two Back Propagation Computrol N
0.24 53-60J, r Kazuyuki Aihara, Neural Computer Tokyo Denki University Press).

It should be noted that each function of the abnormal shadow detection apparatus of this embodiment does not need to be clearly distinguished for each function of each block 101 to 105 shown in FIG.
If it has the functions 1 to 105, it is included in the main embodiment, and for example, while performing height space filtering processing to extract the abnormal shadow candidate C (the function of the first candidate extraction means 101 ), and at the same time the area of abnormal shadow candidate C,
The abnormal shadow detection apparatus of this embodiment also includes the calculation of feature quantities Fl, F2, .

Further, as an aspect of the present embodiment, in addition to the functions of each block 101 to 105 shown in FIG. C
When calculating the abnormal shadow T from, for example, this anatomical information is used as the plurality of features EtF1+ F2+...
..., Fn is input to the neural network, and the output of the neural network (ffiE representing the probability that candidate C is an abnormal shadow) is subjected to threshold processing to determine that candidate C is "A 7i"; shadow T. The anatomical information is used to determine the threshold value for determining whether or not the above-described feature values F1 + F2, . . .
..., may be considered together with Fn.

Here, "anatomical information" refers to information that appears on radiographic images.
Refers to information regarding the behavior of the subject, specifically, for example, the lung field in a chest X-ray image.
Part 1 refers to location information such as ribs, heart, diaphragm, etc.

Since the abnormal shadow detection device of this embodiment has the above-mentioned configuration, it extracts the abnormal shadow candidate C without omission even though noise may be mixed in, and extracts the abnormal shadow candidate C based on a plurality of feature amounts. By extracting the abnormal shadow T from the candidates, abnormal shadow extraction can be performed with less omission, and patterns that are not abnormal shadows are rarely extracted as abnormal shadows. That is, concrete shadows can be detected with high accuracy.

Moreover, this means that the display means 104 and the input means 1
05, and further includes the second candidate extraction means 10.
3 has a learning function that changes the calculation for extracting the cumulative shadow T from the candidate C. Therefore, for example, this abnormal shadow detection device can be observed at the facility where it is installed or the visible image displayed on the display means 104 of this device. Detection can be performed with even higher precision depending on the observer or the like.

In this embodiment, the learning function is not limited to a specific learning function. If a neural network with a learning function is provided, the accuracy of detecting abnormal shadows can be further improved due to the excellent learning function of this neural network.

Hereinafter, the above embodiment will be described in more detail with reference to the drawings. Here, an example will be described in which the stimulable phosphor sheet described above is used to extract an ulcer shadow from an X-ray image of the chest of a human body.

FIG. 8 is a diagram schematically showing an example of a chest X-ray image accumulated and recorded on a stimulable phosphor sheet.

In the center is an X-ray image of the lungs I] 5 (hereinafter referred to as the ligament 115).

57, 58, and so on), a skin part 1 part 6 is formed on both sides thereof, and a direct X-ray part 117 is formed on both sides thereof, in which the X-rays are directly irradiated onto the sheet H4 without passing through the subject.

FIG. 9 is a perspective view showing an example of an X-ray image reading device and a computer system which is an example of the abnormal shadow detecting device of the present invention.

A stimulable phosphor sheet LL4 having an X-ray image recorded thereon as shown in FIG. 8 is set at a predetermined position of the X-ray image reading device 120. The stimulable phosphor sheet 114 set at a predetermined position is conveyed (sub-scanned) in the direction of arrow Y by a sheet conveying means 122 such as an endless belt driven by a motor 121. On the other hand, the laser light source 123
The light beam {24 emitted from sheet] 14 and is main-scanned in the direction of arrow X, which is substantially perpendicular to the sub-scanning direction (direction of arrow Y). Excitation light 124 of sheet 114
From the irradiated area, stimulated luminescence light 129 is emitted in an amount corresponding to the accumulated and recorded X-ray image information, and this stimulated luminescence light 129 is guided by a light guide 130 and is passed through a photomultiplier (photoelectron). It is photoelectrically detected by a multiplier tube) 181. The light guide 130 is made by cutting a light-guiding material such as an acrylic plate into a rectangular shape, and the linear input end surface 30a extends along the main scanning line on the stimulable phosphor seam 14. The light-receiving surface of the photomultiplier 131 is bonded to the annularly formed exit end surface 130b. The stimulated luminescent light 129 that has entered the light guide 130 from the incident end surface 130a is transmitted through the light guide 1
30 through repeated total reflection, and the exit end surface 130
The light is emitted from the photomultiplier 13{ and is received by the photomultiplier 13{, and is converted into an electrical signal by the photomultiplier 131 or stimulated luminescence light 129 representing an X-ray image.

The analog output signal SO output from the photomultiplier 131 is logarithmically amplified by the logarithmic amplifier 132, and
The image data S1 is digitized by a /D converter 133, and image data S1 as an electrical signal is obtained.

59 60 The obtained image signal S1 is transmitted through the commutan stem 140
is input. This computer system 140 constitutes an example of the abnormal shadow detection device of the present invention,
A main unit 14l with a built-in CPU and internal memory, a drive unit 142 into which a floppy disk as an auxiliary memory is inserted and driven, a keyboard 143 for an operator to input instructions, etc. necessary for this computer system 140, and necessary information. CR to display
It consists of a T display 144.

Image data S input into the computer system 140
1, abnormal shadows on X-ray images are detected. Functionally, the computer system 140 performs calculations corresponding to the blocks 101 to 103 shown in FIG. 7. Below, blocks 101 to 10 shown in FIG.
3 is considered to be a block representing the functions of the computer system 140. In addition, the CR of the computer system 140
T-display 144 and keyboard 143 are examples of display means 104 and input means 105, respectively, of the present invention.

Image data S1 as a digital electrical signal representing an X-ray image as shown in FIG. 8 is input to the first candidate extracting means J-01 shown in FIG.

Here, an example will be described in which a tumor shadow, which typically occurs in the lungs of a human body in a substantially spherical shape, is detected as an abnormal shadow. On the visible image based on the image data S1, this tumor appears as a whitish (low a degree) approximately circular pattern compared to the surrounding area.

First candidate extraction means FIG. 10 shows the first candidate extraction means 101 used in
In order to explain an example of the large space filter for extracting the tumor shadow, it is a diagram virtually drawn on an X-ray image, centering on a predetermined pixel Po on the image. It is determined whether the predetermined pixel Po is a pixel within the tumor shadow. By scanning an X-ray image using a filter like the one shown here,
A tumor shadow on the X-ray image is extracted.

FIG. 11 shows the above-mentioned predetermined pixel P. The 1st O, centered on
Figure 6 shows an example of the profile of an X-ray image in the direction in which line segments L1 and L5 extend (X direction). . Here, it is assumed that the predetermined pixel Po is located at the center of the cheek of the rib shadow {the ulcer shadow located very close to 06] 07.

The tumor shadow 107 typically appears as a laterally symmetrical cheek profile, but it may not be symmetrical in some cases, such as when the tumor shadow 107 is located very close to the rib shadow 106 as in this example. In such cases, this tumor shadow J0
It is important to be able to extract 7. Note that the broken line 10g in FIG. 11 is an example of a profile when there is no tumor.

As shown in FIG. 10, a plurality of (eight lines in this case) line segments L1 (1 = 1..2... .
Circle R+ of l+ r2+ r3 (j = I., 2
．． Assume 3). Predetermined pixel P. The image data of f
. Each pixel P+1 located at each intersection of each line segment L1 and each circle R (FIG. 5 shows PII+ p, ■+ P 13
Symbols are shown for +P 51+P 52+P 53. ) is assumed to be fl,.

? Here, image data fo of a predetermined pixel Po and each pixel PI
, image data f. The difference Δ1 from

Δl,−f l,1f. ...(24)(f
-1.2,... 8; j = 1, 2.3>Next, for each line segment LI, the maximum value of the difference Δ1 determined by equation (24) is determined. That is, to give an example regarding line segments L1 and L5, for line segment L1, pixel PI1+''
+2+ The maximum value of each difference ΔII"'fl■-fo Δ12""fl2 fQ Δ+:+=f+3fo corresponding to P13 is found. In this example, ffill
As shown in the figure, Δ,3<Δ1■<Δ,,ku0, and therefore Δ1 is the maximum value. Further, for the line segment L5, the maximum value Δ,3 of the differences Δ51−f51 fo Δ52=f52 fo Δ53−f53 fO corresponding to the pixels P 51 + P 52 + P53 is determined.

63 In this way, the predetermined pixel P. By calculating the maximum value of the difference between each line segment L and a plurality of pixels, it is possible to deal with tumor shadows of various sizes.

Next, two line segments extending in opposite directions from a predetermined pixel Po are set as a set, that is, line segment L1, line segment L5, line segment L
2 and line segment L6, line segment L3, line segment L'rs, and line segment L4
and line segment LB as one set, and the average value of the two maximum values for each set (respectively hL5. M26. M37
．． M48) is required.

Line segment L1 and line For the pair with minute L5, the average value M,, is 2 It is required as.

In this way, by treating two line segments extending in opposite directions from a predetermined pixel Po as a set, the 11th
As shown in the figure, the tumor shadow 107 is located near the rib shadow 106, for example, and the image data of the tumor shadow 107 is 64? It is possible to reliably detect Ilili tumor shadows even if the images are asymmetric.

As above, the average values M15, M26, M37,
Once M4B is determined, these average values M, , M2
6, M37, M4B, a predetermined pixel P is determined as follows. A characteristic value C used for determining whether or not a pixel is within a tumor shadow is determined.

FIG. 12 is a diagram for explaining how to obtain this characteristic value Cl. The horizontal axis is the average value M obtained as above,
5, M26, M37, M48
, the vertical axis is each evaluation value CI I+ corresponding to these average values.
C26, C3■C48.

Average value M 1 5 , M 2 6 , M 3 7
, M 4 If B is smaller than a certain value M1, the evaluation value is zero; if it is larger than a certain value M2, the evaluation value is 10, Ml ~ M
Between 2 and 0.0 to 1.0 depending on the size of the value.
The value between is the evaluation value. In this way, each average value M
,,, Evaluation value CI5i C26C37+C4 corresponding to M26M 3 7 and M 4 B, respectively
B is calculated, and these evaluation values CI5C 26+ C
31+ Sum of C4B 65 66 C 1 = C I, + C 26+ C 37+
C 48 (26) is taken as the characteristic value C1. That is, this characteristic value C1 has a minimum value of 0, or any value between 0 and a maximum value of 4.0.

This characteristic value 01 is compared with a predetermined threshold Thl, and C
Depending on whether 1≧Thl or Cl<TI11,
Predetermined pixel P. It is determined whether each pixel is within the tumor shadow.

By scanning the X-ray image using the real space filter, that is, each pixel on the X-ray image is converted to the predetermined pixel P.

The tumor shadow on the X-ray image is extracted by determining whether each pixel is a pixel within the tumor shadow. Note that in the above-described scanning of the filter, there is a possibility that objects other than tumor shadows similar to the tumor shadow are also extracted, and therefore, herein, the patterns extracted by the above-mentioned scanning are referred to as tumor shadow candidates.

The filter for extracting tumor shadow candidates in the first candidate extraction means 1 [11 (see FIG. 12) is not limited to the above filter. Other examples of filters that can be used by the first candidate extraction means 101 will be described below.

Each pixel P++ (i = 1.2..., 8
．． gradient fl of image data f1 of j=12,3)
, is required.

Here, gradient 1 is the xy coordinates of a certain pixel P on an X-ray image (m, n), and the coordinates of pixels P' and P' adjacent to the pixel P in the X and y directions ( m+i,n
) , (m, n+1), and their pixels P, P'
The image data of P′ are respectively f (m, n) and 4
(n++1n) , f (m, n+1), then f (m, n) = ( f (m+1..n) −
f (m.n)f (m, n+1) − f
(m, n) ) − (27).

FIG. 13 is a diagram showing the gradient 1 and the calculation method described below.

After the gradient f1 is determined, the lengths of these gradient f1 vectors are aligned to zero. That is,
When the magnitude of gradient Vf++ is Vf+11, normalized gradient 67 68 ? f++/lf++l is determined.

Next, the precept of this normalized gradient r1/1f,1 in the direction of the line segment L1 is determined. That is, each pixel P. from a predetermined pixel P. When the unit vector toward is 01, f++/lf1■*e. (However, * represents the inner product)
is required.

After that, each line segment L+ (i-
1, 2. ..., 8), each maximum value {Vf. /lf1
1l*e1)v (i-↓,2,...,8) is calculated, and the sum value Σ {f ++/l is obtained by adding each of these maximum values {Vf+1/l Vf+11 *e+ )M V f ++ l *e+
) M is required. This added value Σ {f1+/lVf *e11 y is set as a feature RC2, and this feature In C 2 is compared with a predetermined threshold Th2 to determine whether C2≧Th2 or not.
<Th2, it is determined whether the predetermined pixel Po is a pixel within the tumor shadow.

This filter has a gradient f1 of magnitude f. By normalizing 1 and focusing only on its direction (the degree of difference in direction from line segment L1), we can obtain a feature C that has a large value due to its circular shape, regardless of the contrast with the surroundings.
2 is obtained, and thereby the ulcer shadow can be extracted with high accuracy.

Next, further different examples of filters that can be used in the first candidate extracting means 101 (see FIG. 12) will be explained.

Qo and Q mouth (1 = 1.2,...
..., 8j = 1.2.3) (However, in Figure 10, Qo and Q++・Q12・Q+3・
Only Q51, Q52, and Q53 are shown) are pixels P, respectively. and each pixel P B (i =
], 2, ..., 8; j=1. , 2.3).

6 70 Each area Q. and Q+1 (] =1,2,...
. . , 8; j −1.2.3 'I for each region Q.

, Q++, the average value Qo, Q . (i = 1,2s・-・
, 8; j =]. 2.3) is required. In addition,
Here, for simplicity, each region Qo, Q ++ (f
=1.2s...-,8; j =1,2.3) and the same symbol is used for the average value of the image data in each area.

Next, the average value Qo of the central area and the average value Qo of each peripheral area. each difference Δ1) (i = 1.2,...
・,8． j=1.2.3) is Δ1+=Q++ Q. ...(28), and the maximum value Δ1 of the difference Δ11 is determined for each line segment L. That is, to give an example regarding the line segment LIL5, for the line segment L1, the maximum value Δ1 among Δ1, Δ12, Δ, and 3, and the line segment L! For Δ,1,Δ,2,
The maximum value Δ5 of Δ, 3 is determined.

Next, a first characteristic value U representing the maximum value Δ+ (.i=1 to 8) and a second characteristic value ■ representing the variation in the maximum value Δ+ (1=1 to 8) are determined. For this purpose, first, each characteristic value U1~U4, V1~V
4 is required.

U1-(Δ1+Δ2+Δ5+Δ6)/4...(29)
U2=(Δ2+Δ3+Δ6+Δt)/4...(30
)U3-(Δ3+Δ4+ΔT+Δ8)/4...(3l
)U4-(Δ4+Δ5+ΔB+Δwork)/4...(32
)v1=U. /U3...(
33) Vz −Uz / Ua
・=(34)V3=U3/Ul
...(35) Va=Ua/U2
...(36) Here, for example, <29)
To explain the case of calculating the characteristic value U1 according to the formula, two adjacent areas (Δ1 and Δ2, or Δ5 and Δ6
) means smoothing, and the areas on opposite sides of the pixel Po (Δ1+Δ2 and Δ5+Δ6)
Similar to the first filter described above, the purpose of adding the sums is to enable the detection of ulcer shadows even if the image data is asymmetrical as shown in FIG.

7] For example, to explain the case where the characteristic value V is determined according to equation (33), the characteristic value U1 and the characteristic value U3 are characteristic values determined in directions orthogonal to each other, and therefore the tumor shadow shown in FIG. If 107 is circular, V1
+1.0 and deviates from the circle, that is, pixel P. If it is in a linear shadow like a rib shadow, Vl is 1,
It will deviate from 0.

The first characteristic value U representing the maximum value Δ1 (i=1 to 8) of the above difference is the maximum value U-MAX of U1 to U4.
(Ui.U2.U3,U4) ・=
(37) is adopted, and the maximum value of the above difference Δ+(1=1~
The second characteristic value V representing the variation in 8) is 1
~ Maximum value of V4 V = MAX (V1, vz , v3, v,) ・・
・(38) is adopted. When the first and second characteristic values U and V are obtained in this way, a predetermined pixel P is determined. Characteristic value C for determining whether or not is a pixel within the tumor shadow
3, the ratio 72 U C3 = ・(39) ■ of these first and second characteristic values is adopted, and this characteristic value 03 is the predetermined threshold value T1]3
It is determined whether each pixel Po is a pixel within the tumor shadow, depending on whether C3≧Th3 or C3<Th3.

As shown in the example above, the first candidate extraction means 101 of the computer system 40 shown in FIG. Extraction takes place.

In each of the above filter examples, as shown in FIG. However, it is needless to say that the number of line segments does not have to be eight, and may be, for example, 16. Also, regarding the distance from the predetermined pixel Po, r1, r2+
Although calculations were performed for three distances of r3, this is not limited to three distances; if the size of the tumor shadow to be extracted is approximately constant, one distance may be used, or In order to more accurately extract tumor shadows of various sizes,
The calculation may be performed by continuously changing the distance from r1 to r3. Furthermore, the first candidate extracting means 101 is not limited to the above-mentioned filters, and may use other filters.

However, here we do not extract only patterns that are definitely considered to be tumor shadows, but rather extract some noise (fFf
It is preferable to extract patterns that are not tumor shadows as tumor shadows, or to extract patterns that are not tumor shadows without leaking tumor shadows even if patterns that are not tumor shadows are included. Therefore, this first candidate extraction means 101 The filter used in
It is preferable that it is in line with that purpose.

Note that when an X-ray image is scanned using each of the above-mentioned filters, intersections of linear shadows, such as areas with dense blood vessels, may be extracted as tumor shadow candidates. Therefore, in the first candidate extracting means 101 in this embodiment, after extracting tumor shadow candidates, areas with dense blood vessels etc. that are not tumor shadows among the extracted tumor shadow candidates are extracted from the tumor shadow candidates as follows. Excluded.

FIGS. 14A and 1.4B are diagrams showing X-ray images of a true tumor shadow and a region densely populated with blood vessels, respectively, once extracted as tumor shadow candidates. Dashed line 1 in each figure
The area surrounded by 09 is the area A once extracted as a tumor shadow candidate, and each graph is a profile in the X direction and y direction within each area A (plotted from image data S1)
It is.

The tumor shadow (Fig. 14.A) has a relatively flat profile with a valley near the center in both the The profile has fine fluctuations in one direction (the X direction in FIG. 14A) and is relatively flat in the other direction (the y direction).

Therefore, here, by utilizing this difference in profiles, regions where blood vessels and the like are densely extracted once as tumor shadow candidates are excluded from tumor shadow candidates. That is, the pixels lined up in the X direction are expressed as m (m = 1.2,...), and the pixels lined up in the y direction are expressed as n (n-1.2...), Let f (m.n) be the image data of the pixel represented by (m, n). At this time, as shown in the following equation,
The average value of the squares of the primary difference values of the image data in the area A is calculated.

Zx one ΣΣ ( f (m+1.n) − f (
m, n) ) 2/ N(ffi.n) et al.
...(4o)Z, -ΣΣ {
f (IN, n+1) − f (m, n) ]
2/N(m.n)j;^
−(41) (where Σ Σ represents the addition of (lI.I1) et al.^ of the first-order difference value within region A, and N represents the number of pixels within region A) Next, as a tumor shadow sleeve, As the feature fflct for determining whether to keep it alive or exclude it from tumor shadow candidates, the smaller value of the above z8 and Zy is set as min(2., 2,)
, the larger value is set as max (1, Z,), max (ZX IZy) is calculated, this feature amount C4 is compared with a predetermined threshold Th4, and when C4≧Th4, it is selected as an ulcer shadow candidate. It is kept alive and excluded from the ulcer shadow candidates when Ca<Th4.

Note that the feature amount C4 is not limited to that calculated by equation (42), and for example, Cz-l Z x Z y l
...(44) etc. may be used. Also, in the above example x
, the first-order difference f (In+l, n) − in two directions of y
f (m.n), f (m, n+1) - f
(m, n), but for example in the diagonal direction (X direction 1
It is also possible to calculate the difference in a direction (direction that is not orthogonal to any of the y-directions).

77 78 In the first candidate extracting means 101 in the computer system 40 of FIG. 9, tumor shadow candidates are extracted by scanning the X-ray image using the real space filter as described above, and the expression (42) is It is determined whether or not to continue as a tumor shadow candidate based on the feature amount C4 of The resulting binary image data is obtained. Thereafter, the tumor shadow candidates that have been determined using the feature ffi C a of equation (42) are extracted by the first candidate extracting means 1.
01 is referred to as the extracted tumor shadow candidate C.

Feature amount calculation means The tumor shadow candidate C extracted by the above-mentioned amorous candidate extraction means 101 is processed by the feature amount calculation means 1 of the computer system 40.
02 is input. The feature amount calculating means 102 calculates a plurality of feature amounts F1, F2, ..., Fns for each candidate C based on the image data S1 in the vicinity of the tumor shadow candidate C, that is, in this embodiment, Area Fl1 of each tumor shadow candidate C obtained as follows Shape F2 of each tumor shadow candidate C
And three feature quantities F1 to F3 of the contrast F3 of each tumor shadow candidate C are determined.

FIG. 15 is a diagram showing an example of tumor shadow candidates C extracted by the first candidate extraction means 101. The central region 7 is the region extracted as the tumor shadow candidate C.

Among the tumor shadow candidates C extracted by the first candidate extraction means 101, if the area is small (for example, 25 mm2 or less), the tangent of the blood vessel (in the direction perpendicular to the X-ray image (see Figure 8) There is a high possibility that the shadow is a blood vessel shadow extending in the irradiation direction), and there is a strong suspicion that this is the case, especially if contrast F3, which will be described later, is high. Further, the size of a tumor shadow that is normally seen is approximately 10 m to 40 mm in diameter, and a tumor shadow candidate C that is too large beyond this range is often not a tumor shadow.

Therefore, in this embodiment, the area F1 of the region 7A is first determined, and the area F1 is adopted as one of a plurality of feature quantities that reflect the probability that the tumor shadow candidate C is a IIiRm shadow.

In addition, in this example, a tumor shadow that typically appears as a circle on an X-ray image is extracted, and therefore, a quantity representing the extent (shape) of this circle is adopted as another feature quantity. be done. Here, the center of gravity O of the area A7 is determined based on binary image data in which a value of 1 is assigned to the area 7A and a value of O is assigned to the surrounding area, and a large number of line segments extending from the center of gravity O to the periphery (see FIG. So Jlx~f!8 of 8
book) is imagined and each line segment j! 1 (1=L2,...
..., 8), each distance d + (1 = 1, 2, ..., 8) from the center of gravity 0 to the periphery 7a of the area 7A is found, and the dispersion F2 is found according to the following formula. , this variance F2 is employed as a feature amount representing the circularity (shape) of the tumor shadow candidate C.

Furthermore, the average value Ayl of the image data S1 in the region 7A of the tumor shadow candidate C and the band-shaped region 7 of a constant width around the region 7A
The average Ay2 of the image data S1 of A' is calculated using the following formula 81 F3 −Av+ AV2 −(40)
Accordingly, the contrast F3 is determined, and this contrast F3 is further employed as a different feature quantity. Since a tumor shadow appears on an X-ray image as a projected image of a roughly spherical tumor within the lung, its contrast F3 is often close to a certain predetermined value, and if it deviates from that value and becomes a low contrast of 1. Since a tumor shadow candidate with too high contrast or a tumor shadow candidate with too high contrast is often not a tumor shadow, this contrast F3 can be employed as a feature amount.

In this embodiment, two or three features ffiFx to F3 are obtained, but depending on the type of abnormal shadow to be extracted, the accuracy of extraction, etc. Of course, in place of F3, various other feature quantities may be employed.

Second candidate extraction means Three features ffiFx~F obtained as above
82 in the computer system 40 of FIG. 9 is input to the second extracting means 103. The second candidate extracting means 103 in this embodiment employs a neural network described below to extract the first Among the tumor shadow candidates C extracted by the candidate extracting means 1, a candidate F with a high probability of being a tumor shadow (this may be referred to as a tumor shadow F) is extracted.

FIG. 16 is a diagram showing an example of a neural network equipped with an error backpropagation learning (pack propagation) function employed by the second candidate extraction means 103. Error backpropagation learning (pack propagation) is, as mentioned above, by comparing the output of the neural network with the correct answer (teacher signal), and sequentially adjusting the connection weights (synaptic connection weights) from the output side to the input side. It refers to a “learning” algorithm that makes corrections.

As shown in the figure, the first layer (input layer) and second layer (hidden layer) of this neural network. The third layer (output layer) is composed of three units, three units, and one unit, respectively. The first unit of the k-th layer is u7, and all inputs to unit 7 are X? % full power”/7 s u〒from 7”1
The weight of viscosity to W7 'H''hL, each unit 1・u:
The first layer (input layer), second layer (hidden layer), and third layer (output layer) of the same characteristic function are composed of three units, three units, and one unit, respectively. The unit “i number 1 of the kth layer” is LI7, all inputs to this unit 7 are X7, and all outputs are y
↑, the weight of the connection from LJ7 to k+l is Wi1"1, and each unit h, has the same characteristic function 1. At this time, the input X7 and output Y7 of each unit 7 are expressed as Xi-ΣWi- ': ・ 7-1 ... (4
8) 83 84 y7 = f (X:) ... (49)
becomes. However, each unit u}(1−
1.2.3) each input F1 + F2 + F3
is not weighted and remains as it is for each unit u: (i=
1, 2, 3) l: Input. Three input features i
F. , F2F3 are transmitted to the final output y? while being weighted by the weight W77'' of each connection.

Here, we will explain how to determine the weight w::'' of each connection above. First, the weight w k k of each connection is determined by a random number.
An initial value of +l is given. At this time, the entrance F
Even if 1 to F3 fluctuate to the maximum, will the output be y? It is preferable to limit the range of the random number so that the value is within the range of 0 to 1 or a value close to this.

Next, a large number of X-ray images are prepared in which the presence or absence of IMI tumor shadows and the location of Ihli tumor shadows, if any, are known. Tumor shadow candidates C are extracted as described above, and each tumor shadow Regarding candidate C, the above three feature quantities F1 and F2
, F3 are obtained. These three characteristics 1tF1, F2
, F3 are input to the neural network shown in FIG. 18, and the output y of each unit ui is monitored.

Once each output yi is determined, the final output 3/I
'The input features ffiF1, F2 . A value of 1 if the tumor shadow candidate C corresponding to F3 (as described above, it is known whether or not this tumor shadow candidate C is a tumor shadow) is a tumor shadow, and a value of 0 if it is not a tumor shadow. The teacher signal d with
The squared error between In order to minimize this squared error E, the weight W 7 7''1 of each connection is modified as follows.

To minimize the squared error E, this E should be W 7 : ”
Since it is a function of 1, the weight W〒1+1 of each connection is modified as 85 86 . Here, from equation (47), η is a coefficient called a learning coefficient.

f′ (x)=f (x) (1-f (x) ) ...(56) here, Because it is, f'(X:) Mo y (Co- y+) (57) becomes.

In equation (53), put l(-2, (55) (57) the expression and From equation (48) Substituting into equation (53), we get 1ΣW1 ◆y ...(4g)' Because it is, Equation (52) is One (y1 d)・y1 (1 y1) 0y ...(58) Substituting this equation (58) into equation (51) yields.

here, From formula (50), ? 1　■　-W+　, η・(y? d)・y1 (1 y+) 0y ...(59) becomes.

According to this equation (59), when this equation (54) is transformed using equation (49), the weight of each connection is corrected.

Next, since 87 88, substitute the formula (48) (49) into this (en) formula = (y: d) ・y1 (1 yy+), and get ・ydō ・(1−y,)・W, 0 y...(64>) Substituting this equation (64) into equation (51), setting k=1, w:.i-W,?-η・(ymard)・y1Here, (
From formula 5B), f' (x d) ippa (1 y+) ... (62) (1 y+) ●y, ◆ (1 y+) ◆y, so this formula (62) and (55 ) By substituting equation (57) into equation (61) * W,, ... (65), we get 1 (i-1 2, 3) or (65) with W corrected in equation (59). W1 (1,j-1.2.3) is modified.

◆y, ● (I Y+) ●W, 1... (6
3> Place k-1 in equation (53) and replace equation (63) with equation (53) ΔWI1+(t+1)-α・ΔW+ ,+1 (t)
+, 89 90 However, Δw k k + l (t) is the modified connection weight w k k '- 1 in learning t times l'
represents the amount of correction obtained by subtracting the weight Wkk+1 of the connection before correction from Wkk+1. Further, α is a coefficient called an inertia term.

For example, α-0, 9η-0 as the inertia term α and the learning coefficient η.
．． Modification (learning) of the weight Wkk+1 of each connection using 25
For example, 200,000 times, and then the weight of each connection W
7 7 ” l is fixed to the final value. At the end of this learning, the value close to 1 for tumor shadows among tumor shadow candidates C, and the value close to 0 for tumor shadow candidates C that are not tumor shadows. is now output as the output y?.

After the learning is completed, a tumor shadow candidate C, which is unknown whether it is a tumor shadow or not, is extracted and has a characteristic value F1 + F.
2, F3 is determined and input to the neural network shown in Figure 18, and the output y obtained thereby is
? is a signal indicating the degree of probability that the tumor shadow C is a 1114 tumor shadow. Since this signal is obtained after learning as described above, it accurately represents the probability that the tumor shadow candidate is a tumor shadow.

Note that the neural network described above is not limited to a three-layer structure, and may of course have even more layers. Furthermore, the number of units in each layer is not limited to the number in the above embodiment, but may be any number depending on the number of input characteristic values, the required accuracy of the signal representing the probability of being an ulcer shadow, etc. Of course, each layer can be constructed from units.

When the signal (output y1) is obtained as described above, this signal (output y1) is compared with a predetermined threshold value 91 92 Th5 (for example, 0.5) to determine whether yi≧Th5?'/: <Th5, the feature amounts F1, F2, F3 input to the neural network
It is determined whether the tumor shadow candidates corresponding to the respective tumor shadows are tumor shadows or not tumor shadows.

Here, the threshold value Th5 may be fixed, but it does not need to be fixed; for example, the value can be changed depending on the anatomical information of the subject in the X-ray image. You can. An example of how to obtain anatomical information is described below.

The llrf tumor shadow of 11i1i has the property that it appears more in the hatched lung field 115b on both sides than in the II, li hilum 115a near the center shown in FIG. Therefore, here, the first sister part 115 or the first sister part 115a and the lung field part 1.5b are divided as follows, and the information thereof is outputted from the output V? When comparing the threshold value Th5 with the threshold value Th5, the threshold value T
Used to determine the value of 115.

18 is a diagram showing a histogram of image data S1 obtained from the X-ray image shown in FIG. 17. FIG. The horizontal axis is image data S
Image data S representing a value of 1 and the vertical axis having that value]
represents the frequency of appearance.

The peak 1{7 on the right side of this histogram corresponds to the direct X-ray section {17 shown in FIG. The center peak 115 corresponds to the lung part 115 shown in FIG. 8, and the left peak 116 corresponds to the skin part 11B shown in FIG. For simplicity, here
The same numbers are used for each part in FIG. 8 and each mountain in FIG. 17 that correspond to each other.

Here, based on the empirical rule that the area of the lung field 115b relative to the area of the lung 115 is approximately constant, a predetermined ratio (for example, 50%) of the total area of the mountain 115 is determined from the high concentration side (
17), a set of pixels corresponding to image data within this area is defined as a lung field 115b.

In this way, the four lung parts 115a and the lung field part 115b are divided, and when determining whether or not a tumor shadow candidate in the lung field part 115b is a tumor shadow, the threshold value Th5 is set to be low (for example, 0 , 4), and a high value (for example, 0.6) is set when determining a tumor shadow candidate in the hilum 115a.
is set to

93 94 In addition, rib 'f'l-i is also extracted here.

FIG. 18 is an enlarged view of a part of the rib shadow (not shown in FIG. 8) in the chest X-ray image shown in FIG. 8. Note that FIG. 13 also shows a diagram of the filter shown in FIG. 10, which will be described later.

The chest X-ray image shows two rib shadows 1 as shown in Figure 18.
15c appears, and this crossed area 115d may be extracted as a tumor shadow candidate. Moreover, the image data within this intersection region 1.15d is almost flat, and unlike the case where a region where blood vessel shadows are gathered is extracted as a tumor shadow candidate, the image data within this intersection region 1.15d is Therefore, when extracting the rib shadow and calculating the tumor shadow from the tumor shadow candidate, the tumor shadow candidate is not excluded from the intersection area 11 of the rib shadow.
5d, the candidate is excluded from the tumor shadow.

An example of a method for extracting the rib shadow 115C is "Identification of Rib Images in Indirect Chest X-ray Photographs" (Institute of Electronics and Communication Engineers, 1972 {October 26, 1972, Imaging Engineering 95 Academic Research Group Material, Material No. IT72-24). (19γ2
-10)) is adopted. In this method, a chest X-ray image is scanned using a filter that is sensitive to lines, line figures are extracted, and the ribs and +1 shadows are carved based on the position of the line figure on the X-ray image, the direction in which the line figure extends, etc. A corresponding line is extracted, and a rib shadow is extracted by approximating the rib boundary line with a quadratic function.

Note that the threshold value Th5 is fixed and the lung field 1
15b or hilum 115a as a feature f
fiF. ．． F2, . . . , Fn may be input to the neural network described above.

Further, the above anatomical information may be used when extracting abnormal shadow candidates.

For example, when determining whether a predetermined pixel Po is a pixel within a tumor shadow as described above using FIG. In determining whether the pixel is within the L. ，
For L3, L5, and L7, we use the information of r1 and r2, for L2 and L6, we use the information of r1 to 3, and for L4 and L8, we use only the information of r1, so that this rib shadow can be This can be avoided without affecting the judgment.

As described above, in the second candidate extraction means 103 of the computer system 14 (1 (see FIG. 9)), the plurality of features ffiF1,
F2,...,Fn are connected to neural network (
(see FIG. 16) and first candidate extraction means 1 (see FIG. 16).
A signal bar representing the probability that the tumor shadow candidate C extracted in is a tumor shadow is obtained, and this signal bar is subjected to threshold processing to determine the candidate C.
After it is determined that the tumor shadow T is a tumor shadow T, data representing the tumor shadow T (for example, information on the position and shape of the tumor shadow T, etc.) is displayed in the display means 101 of the present invention together with the image data s1.
A CRT display 144 (see FIG. 9) is an example of
A visible image based on the image data s1 is displayed on the screen of the CRT display 144, and the tumor shadow T is clearly indicated on the image by, for example, an arrow.

An observer who observes the visible image can obtain information T that distinguishes whether the tumor shadow T is a true tumor shadow or whether something that is not actually a tumor shadow is displayed as a tumor shadow. A keyboard 143 is an example of the input means 1f)5.
Enter from. This information T is stored in the computer system 14
0 is input to the second candidate extraction means 103. The second candidate extracting means 103 receives this information T, and if this information T is information indicating that it is a true ulcer shadow, the teacher signal d (see formula (50)) is set to d=1, and the tumor If the information is that it is not a shadow, the error backpropagation learning of the neural network described above is performed as d − 'O, and the connection weight W 7 7 + 1 is set so that a more accurate judgment will be made next time. changes.

As described above, according to the present invention, while actually using the S-copy shadow detection device of the present invention, further knowledge is accumulated so that the device is more suitable for the facility where the device is installed and the observer.

97 98 Note that in the second embodiment, a neural network is used as the second candidate extraction means 1.03, but the second candidate extraction means 103 of the present invention is limited to one equipped with a neural network. Rather, instead of or in conjunction with a neural network, a plurality of features ffiF1. A determination may be made based on F2 and F3.

As in the case of using the neural network described above, a large number of X-ray images in which the presence or absence of tumor shadows and the position of tumor shadows are known are prepared, tumor shadow candidates C are extracted, and each tumor shadow is The three features faFz+Fz and F3 described above for candidate C are found. These features ffiF1. The set of F2, F3 and these feature members F1,
Information as to whether or not the ulcer shadow candidate C corresponding to F2 and F3 is a tumor shadow is stored in the second candidate sleeve extraction means 103 as a database.

Next, the characteristic values F1'', F;, F; extracted from the tumor shadow candidate C, which is unknown whether or not it is an IMf tumor shadow, and the many characteristic values F1, F1, which have already been input as a database.
F2, F3 (1-1.2, ---, n)
Using, the distance a, is a{-(F, F:)2+(Fz F2)20(F
3F;)2...It is determined according to the equation (64). These distances a(1-1.,2
．． ... + n) minimum kneeling a. Whether or not the tumor shadow candidate C having the characteristic value FT + F''; , F 3 corresponding to is a tumor shadow, and its distance a.

Characteristic value F1 + F2 input based on the size, etc.
The degree of probability that the ulcer shadow candidate C having +F2 is an ulcer shadow is determined.

When this NN method is used, the second candidate extracting means 103 receives the information Tl., which indicates whether or not it is a true tumor shadow inputted from the keyboard 143. :^(Next, this99 100? Multiple feature values FIF2 obtained for tumor shadow candidate C,
..., Fn and the above information T are added to the database, and this is considered to be the learning function of the second candidate extracting means 103.

Further, as in the above embodiment, a large number of X-ray images in which the presence or absence of a tumor shadow and the position of the tumor shadow are known are prepared, tumor shadow candidates C are extracted, and each tumor shadow candidate C
The three features described above for ffiFx , Fz !
After F3 is obtained, these feature quantities F■, F2, F3
and these feature quantities F. , F2, l corresponding to F3
Based on the information as to whether or not the tumor shadow candidate C is a tumor shadow, a threshold value for classifying whether or not it is a tumor shadow is determined for each feature value F1 + F2 + F3 to extract a second candidate. A tumor shadow candidate C, which is unknown whether or not it is a tumor shadow, is stored in the means 103 and the feature ffiFt is extracted.
. may also be determined.

When the second candidate extraction means 103 performs such a simple threshold processing, the information T inputted from the keyboard 43 regarding whether or not it is a true tumor shadow is used to determine this threshold value. This is used as the basis for statistical processing, and is considered to be the learning function of the second candidate extraction means 103.

In this way, the second candidate extracting means 103 extracts a plurality of features ff by any one of the above-mentioned methods, various different known methods, and various combinations of these methods.
iFt. Tumor shadow f based on Fz,...,Fn
A judgment is made as to whether +DC is an ulcer shadow or not.

Note that although the above embodiment is an example of extracting a tumor shadow that typically appears as a circle, this is not limited to extraction of a circular tumor shadow, nor is it limited to chest X-ray images. Furthermore, the present invention is not limited to systems using stimulable phosphors, and can be widely used when detecting abnormal shadows on radiographic images based on image data representing radiographic images of a subject.

101 102

[Brief explanation of drawings]

Figure 1 shows the structure of the present invention. ! I (Block diagram for explaining the basic printing process, Figure 2A. Figure 2B is a diagram showing the X-ray images of the right shoulder and left shoulder, respectively. Figures 3A and 3B are the standard pattern and Figure 4 is a diagram showing an example of a neural network, Figure 5 is a schematic diagram of an example of an X-ray imaging device, and Figure 6 is a diagram showing an example of an X-ray imaging device.
A perspective view showing an example of a line image reading device and an example of a computer system including an example of the present invention; FIG. 7 is a block diagram clearly showing the configuration of the abnormal shadow detection device of the present invention; FIG. 8 is a Figure 9, which schematically represents an example of a chest X-ray image stored and recorded on a stimulable phosphor sheet, shows an example of an X-ray image reading device and an embodiment of the abnormal shadow detection device of the present invention. FIG. 10 is a perspective view of the computer system used in the first candidate extraction means 1.
11 In order to explain an example of a spatial filter for extracting a 1m shadow, FIG. A diagram showing an example of the profile of an X-ray image in the extending direction (X direction) of line segments Ll and L5 in FIG. FIG. 13 is a diagram showing vectors such as gradient f1 of image data r1. FIGS. 14A and 14B are Figure 15 is a diagram showing the X-ray images of areas with dense areas such as true ulcer shadows and blood vessels extracted as candidates for tumor shadows, and their profiles in the X and y directions. FIG. 16 is a diagram showing an example of an extracted tumor shadow candidate. FIG. 18 is a diagram showing the histogram of the image data obtained from both the X-ray images shown in Figure 3, and Figure 18 is a diagram showing the costolunar shadow in the chest X-ray image shown in Figure 3. FIG. 1... X-ray imaging device 2... X-ray source 5... Subject part 6... Direct X-ray part 11. 11'...
・Stormable phosphor sheet 19. 19'... Stimulated luminescent light 21, 2]'... FoI multiplier 26.
26'... Logarithmic amplifier 27 27'... A/D converter 40... Computer system 100... Pre-reading means 100'... Main reading means 101... First candidate extracting means 102...Feature melon calculation means 103...Second candidate calculation means 104...Display means 105...Input means 10
6... Rib Tsukuda shadow 107... Ulcer shadow 1 1 5
-= Illi part 1 1 5a - 11
+li hilum 1 15 b... lung field 116...
・Skin part 117...Direct x1i1 part 12[1...
・X-ray image reading device 123...Laser light source {2B
... Rotating polygon mirror 129 ... Stimulated luminescence light 130.
・Light guide 131 ・Photo multiplier 140 ... computer system 105] 06 Figure 2A Figure 3A■ Houka Gorge

Claims (5)

[Claims] (1) Recognizing the irradiation field based on radiation image data, determining image reading conditions for main reading based on pre-read image data,
In a radiation image processing device that uses a neural network to perform data processing such as determining image processing conditions and detecting abnormal shadows, after implementing the trained neural network in the device,
A radiation image processing apparatus comprising: input means for inputting correction information to the neural network for correcting and relearning data processing by the neural network. (2) A first image signal representing the radiation image obtained by irradiating the stimulable phosphor sheet on which the radiation image is recorded with excitation light and reading the stimulated luminescence light emitted from the stimulable phosphor sheet. based on the reading conditions for obtaining a second image signal representing the radiation image by irradiating the stimulable phosphor sheet with excitation light again and reading the stimulated luminescence light emitted from the stimulable phosphor sheet. and/or in a radiation image reading condition and/or image processing condition determination device for determining image processing conditions for performing image processing on the obtained second image signal, storing a standard pattern of a radiation image; storage means; signal conversion means for converting the radiographic image carried by the first image signal into the standard pattern and obtaining a converted image signal carrying the converted radiographic image; and the converted image signal. condition determining means comprising a neural network that receives as input the reading condition and/or the image processing condition as output;
or display means for displaying a radiation image read and/or image-processed based on the image processing conditions as a visible image, and correction for converting the visible image displayed by the display means into a more appropriate visible image. an input means for inputting information into the condition determining means, the condition determining means comprising:
Based on the correction information input from the input means,
A radiographic image reading condition and/or image processing condition determining device characterized by having a learning function that changes the calculation of a neural network that receives the converted image signal as input and outputs the reading condition and/or the image processing condition. . (3) In a radiation image processing condition determination device for determining image processing conditions for performing image processing on an image signal based on an image signal representing a radiation image, a storage means for storing a standard pattern of a radiation image. and a signal conversion means for converting the radiographic image carried by the image signal into the standard pattern to obtain a converted image signal carrying the converted radiographic image, and the image processing using the converted image signal as input. Condition determining means comprising a neural network outputting conditions; and/or the reading conditions outputted by the condition determining means;
or display means for displaying a radiation image read and/or image-processed based on the image processing conditions as a visible image, and correction for converting the visible image displayed by the display means into a more appropriate visible image. an input means for inputting information into the condition determining means, the condition determining means comprising:
Based on the correction information input from the input means,
A radiation image processing apparatus comprising a learning function for changing calculations of a neural network that receives the converted image signal as an input and outputs the reading conditions and/or the image processing conditions. (4) In an abnormal shadow detection device that detects an abnormal shadow on the radiation image based on image data representing a radiation image of a subject, a first candidate extraction that extracts candidates for the abnormal shadow based on the image data. means, feature amount calculation means for calculating a plurality of feature amounts for each candidate based on the image data in the vicinity of the candidate extracted by the first candidate extraction means; a second candidate extraction means for extracting a candidate with a high probability of being the abnormal shadow from among the candidates based on the plurality of feature amounts; based on the image data and data representing the candidate with a high probability; The second candidate extracting means comprises a display means for displaying a visible image showing the candidate, and an input means for inputting information as to whether the candidate shown in the visible image is the abnormal shadow. is characterized by comprising a learning function for changing the calculation for extracting the highly probable candidate from the candidates extracted by the first candidate extracting means, based on the information input from the input means. Abnormal shadow detection device. (5) The abnormal shadow detection apparatus according to claim 4, wherein the second candidate extracting means includes a neutral network having an error backpropagation learning function using the information as a teacher signal.