JP2973016B2 - Digital radiation image recognition system - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an image recognition apparatus for digital radiation images,
More specifically, the present invention relates to an apparatus capable of recognizing an imaging part, an imaging position, and the like of an object from data obtained by converting a radiation image captured as a radiation dose transmitted through the object into a digital signal.

<Prior Art> A radiographic image such as an X-ray image is widely used for medical purposes, and a method for obtaining the radiographic image as an electric image signal includes, for example, the following.

That is, the radiation that has passed through the human body as a subject is absorbed by a certain kind of phosphor, and then the phosphor is excited by light or heat energy, so that the radiation energy that the phosphor has accumulated by the absorption is absorbed. Radiation is emitted as fluorescent light, and this fluorescent light is detected by a photoelectric conversion element to obtain radiation image information electrically.Then, the radiation image signal is digitized, and then subjected to gradation processing and spatial frequency processing to be applied to a CRT or the like. Some are output and visualized (see Japanese Patent Application Laid-Open No. 63-189853).

By the way, in the processing stage before outputting a digital radiation image signal as a visible image as described above, it is necessary to perform a gradation process and a spatial frequency process as described above to obtain a visible image suitable for image interpretation. However, the density of the region of interest in each of the reconstructed images may change due to a difference in the photographed region and the photographing position of the human body.

Therefore, before performing image processing such as gradation processing or spatial frequency processing, it is required to automatically determine the imaging position from the original digital radiographic image and perform optimal image processing for each position. Various methods and apparatuses have been proposed which enable automatic determination of the photographing position using a density histogram of image data or a signal level distribution in a predetermined direction of the image in the image data.
63-262128, JP-A-63-262132, etc.).

<Problems to be Solved by the Invention> However, in the conventional photographing position determination, the photographing position is limited to a part, and it is possible to determine whether the photographing position is the front or the side. There is a problem that the identification of the photographing position is limited.

In order to solve such a problem, a general image recognition technique is used to extract a plurality of feature amounts and perform a determination based on a combination of the feature amounts. Has become very complicated, and it has been necessary to match a dictionary of feature amounts for each imaging region / posture, requiring a lot of labor, and thus it has been difficult to put it to practical use.

Further, in the conventional image recognition, it is difficult to quantitatively grasp how effective the plurality of extracted feature amounts are for recognition. Therefore, it is difficult to determine which feature amount is to be used for image recognition. Selection is not easy, and it is difficult to apply conventional image recognition technology to digital radiation image recognition from this point, and at present, recognition has been limited to imaging parts and imaging positions. there were.

The present invention has been made in view of the above problems, and does not require a complicated calculation, and of course, does not require an effort such as selection of a feature amount to be extracted or dictionary matching. The purpose is to provide.

<Means for Solving the Problems> Therefore, according to the present invention, as shown in FIG. 1, signal thinning means for reducing at least one of the number of pixels of a digital radiation image and the number of signal quantization steps, A neural network in which the features of the image are learned in advance, an output of the signal thinning means is input, and an identification signal indicating which of the plurality of recognition targets the input image corresponds to is output. Learning result storage means for storing a plurality of learning results obtained by performing learning a plurality of times while changing the contents of a learning data set using the learning data set; and a plurality of outputs of the signal thinning means based on each of the plurality of stored learning results. It is determined which of the recognition targets corresponds to, and the most one of the plurality of identification signals derived from the respective learning results is finally determined. And a discriminating means for outputting the digital radiation image.

Here, the recognition target may be at least one of an imaging part and an imaging position of the subject in the digital radiation image.

Further, the signal thinning means divides the original image into a plurality of small regions, and samples the average value or the representative value of the image signal values in each of these small regions, thereby obtaining the number of pixels of the digital radiation image and the quantum variation step of the signal. Preferably, at least one of the numbers is reduced.

Here, when a plurality of learning results are provided as described above, the first neural network learning is performed by using the unlearned initial value, and in the second and subsequent learning, the previous learning result is set to the initial value. It is preferable to obtain a plurality of learning results by performing learning as.

<Operation> According to the digital radiation image image recognition apparatus having such a configuration, the original image data of the digital radiation image is first subjected to a process of reducing at least one of the number of pixels and the number of signal quantization steps in the signal thinning means. Is done. The digital radiation image whose data amount has been reduced in this way is input to a neural network in which the features of the digital radiation image have been learned in advance. The neural network processes an input signal in accordance with learning performed in advance, and outputs an identification signal indicating which of the plurality of recognition targets the input image corresponds to in advance.

That is, by learning a neural network in advance so that a plurality of recognition targets can be identified when a signal processed by the signal thinning unit is input, a feature amount can be extracted from a digital radiation image, or a feature amount can be extracted. The image can be identified by inputting an image signal whose data amount has been reduced by the signal thinning means to the neural network without requiring labor such as creating a dictionary.

Further, in the neural network, learning is performed a plurality of times while changing the content of the learning data set in advance. However, a plurality of learning results obtained by such learning are provided, and processing is performed using each learning result and output is performed. By taking the majority decision of the identified identification signals, the recognition rate of the digital radiation image can be improved as compared with the case where one recognition result of the neural network based on one learning result is the final result.

Here, as the recognition target, what is required in the digital radiation image is, for example, at least one of the imaging region and the imaging position of the subject. In this case, the neural network uses an identification signal indicating where the imaging region is, Or
An identification signal indicating whether the photographing position is front or side is output.

Further, in order to reduce at least one of the number of pixels of the digital radiographic image and the number of quantization steps of the signal by the signal thinning means, the average value or the representative value of the image signal values of each of the small regions obtained by dividing the original image into a plurality of regions is used. Is sampled, the data amount can be easily reduced and input to the neural network without requiring a complicated calculation formula or a large amount of calculation.

Further, when a plurality of learning results are provided as described above, the first neural network learning is performed using the unlearned initial value, and in the second and subsequent learning, the previous learning result is used as the initial value. Is performed to obtain a plurality of learning results, the learning time can be reduced as compared with the case where the learning is performed with an unlearned value as an initial value.

<Example> An example of the present invention will be described below.

FIG. 2 shows an embodiment of a digital radiation image image recognition apparatus according to the present invention.

Here, the image input unit 1 scans a latent image of a radiation image stored (accumulated) on the stimulable phosphor with a laser beam and reads the latent image. Image scanner to signal, or
For example, a filing system in which a plurality of digital radiation image data are already recorded.

The original digital radiation image data f (x, y) input from the image input unit 1 is first input to the signal thinning unit (signal thinning unit) 2. This signal thinning unit 2
Reduces the number of pixels and the number of gradations (the number of quantization steps) of the original digital image data f (x, y) to reduce the amount of image data used for the photographing posture determination, Is input.

Specifically, the original digital image data f (x,
y) is divided into a plurality of small areas (n pixels × n pixels), and an average value of signal values corresponding to pixels included in each small area is calculated, or a pixel included in each small area is calculated. By obtaining one of the corresponding signal values as a representative value, image data f ′ (x ′, y ′) obtained by thinning out the number of pixels and the number of gradations of the original digital image data f (x, y) is obtained. Set. The original digital image data f (x, y) of 2048 × 2464 pixels and 1024 gradations is thinned out to about 5 × 5 pixels to 10 × 10 pixels and 256 gradations as shown in FIG. Also, if the imaging part and the photographing position are limited to some extent, it can be sufficiently identified. For example, the parts and body positions such as the front and side surfaces of the chest and the front and side surfaces of the head are relatively limited to about two types, respectively. The accuracy is sufficiently high for a rough determination.

Here, without thinning out the number of pixels, the original signal value may be re-quantized with a large step width to reduce only the number of gradations, but this will complicate the neural network structure and learning described later. Because
It is preferable to reduce both the number of pixels and the number of gradations.

The image signal f '(x', y ') whose number of pixels and number of gradations have been reduced by the signal thinning section 2 is input to a neural network (neurocomputer) 3 having a configuration as shown in FIG. An identification signal indicating which of the plurality of recognition targets specified in advance by the operation (determination means 3a) by the network 3 is output.

The plurality of recognition targets are, for example, in the case of medical use in which a subject is a person, a chest front, a chest side, a head front, a head side, and the like.

When the discrimination means 3a outputs an identification signal indicating which of the above-mentioned recognition targets corresponds, the image processing unit 4 receiving this identification signal outputs the original digital image data f () from the image input unit 1. For x, y), various image processing and analysis such as gradation conversion, frequency emphasis, and contour extraction are performed based on the determination result of the imaging region and the imaging position so that a visible image suitable for image interpretation is reproduced. .

In the image output unit 5 to which the image data processed by the image processing unit 4 is input, the image data (digital radiation image signal) after the image processing is recorded again in the filing system or reproduced on the CRT. Or recording on a film by a printer. Here, when filing the image data of the subject (the human body in the present embodiment) after determining the imaging region and the imaging position, the image data such as the patient name and the imaging date is included in the filing system. If the photographing position is recorded in association with the image data, it is convenient for later retrieval, and it is not necessary to repeatedly recognize the photographing site and the photographing position.

By the way, the neural network 3 is described in “Nikkei Electronics No. 427,” August 10, 1987, No. 115.
As shown on page-page 124, etc., a plurality of units corresponding to brain neurons (nerve cells) are connected in a complicated manner by a network imitating the human brain, and the connection between each unit By appropriately determining the form (coupling weight), a pattern recognition function or the like can be embedded (learned).

The neural network 3 in the present embodiment is a third network.
As shown in the figure, the input layer S is constituted by an input layer S, a middle layer A, and an output layer R. The input layer S has the same number of neuron models (units) S ₁ as the number of pixels decimated by the signal decimating section 2, S ₂ ,
... has become from the _{S s,} each neuron model S _1, S _2,
... A signal value of a corresponding pixel is input to S _s . As described above, the number of gray scales is reduced from the original 1024 gray scales to about 256 gray scales (8 bits) before being input to the neural network 3. At this time, the 256 gradations from 0 to 255 are normalized to 0.0 to 1.0 and input.

The neuron models A ₁ , A ₂ ,..., A _a of the intermediate layer A are the neuron models (units) S ₁ , S ₂ ,.
-Can be combined with all of S _s . Although the number a of neuron models in the intermediate layer A is empirically determined, in the case of the present invention, it is preferable that the number of neuron models s in the input layer S and the number r of neuron models in the output layer R be about half of the sum. The intermediate layer A may be a single layer as in this embodiment, but may be two layers, three layers, etc.
・ A multi-layer structure may be used.

The output layer R is a recognition target, for example, when the 4 kinds of the front chest, chest side, the head front-side of the head is composed of four neuron model R ₁ to R _4, the neuron model R ₁ to R ₄ is each neuron model A ₁ , A ₂ ,.
·· _{A a} all and is capable of binding.

Identification signal in the output layer R, for example when identified as a chest front, only the output of the neuron model R ₁ is previously set to be 1, which neuron model R ₁ to R ₄
Image recognition (recognition of the imaging part / posture) based on whether or not the output is 1.

By the way, in order to input the signal value of each pixel after thinning as described above to the neural network 3 and to recognize the image, the characteristics of the digital radiographic image are learned in advance, for example, the image signal of the front of the chest. Needs to be learned in advance so that only the output of the neuron model R1 becomes ₁ when is input.

The learning of the neural network 3 is such that when a digital radiation image signal whose imaging region and imaging position is known in advance is processed by the signal thinning unit 2 and input to the neural network 3, an accurate imaging region and imaging position is obtained. The value of each synapse weight (connection weight) in the neural network 3 is changed from the output layer R while changing the learning data set (thinned-out image signal input to the neural network 3 for learning) so that the corresponding identification signal is output. The learning rate is gradually increased by learning sequentially, and the learning (backpropagation rule) is used to learn the difference in the input data due to the difference in the imaging part and the imaging position of the subject and to identify the imaging part and the imaging position. The result of thinning out the digital radiation image signal by the signal thinning unit 2 is By inputting the work 3,
The purpose of the present invention is to identify an imaging part and an imaging position in a digital radiation image to any of a plurality of predetermined recognition targets.

Here, the learning based on the characteristic of the neural network 3 and the back propagation rule will be further described. Each unit (neuron model) Ui constituting the neural network 3 as shown in FIG. The sum of Qj is converted according to a certain rule, and is defined as Qi.
With Wij (synapse weight). This weight Wij
Represents the strength of the connection between the units. Changing this value will substantially change the network structure without changing the connection state. Learning of the neural network 3 means changing this value.
Wij can be positive, zero, or negative, where zero indicates no binding.

If a unit Ui receives inputs from a plurality of other units Ui, and the sum of the inputs is represented by NET, the sum of the inputs of the unit Ui is Becomes

Each unit Ui applies this input sum NET to a function f and converts it into an output Qi as shown in the following equation.

The function f may be different for each unit Ui, but generally, a threshold function as shown in FIG. 4 or a sigmoid function as shown in FIG. 5 is used.

This sigmoid function is a differentiable pseudo-linear function, Can be represented by The value range is from 0 to 1, approaching 1 as the input value increases and approaching 0 as the input value decreases. When the input is 0
0.5. Adding a threshold θ (bias), In some cases,

In such a neural network 3, when input data is given to the input layer S, this signal is converted in each unit, transmitted to the intermediate layer A, and finally comes out from the output layer R. To obtain a desired output, It is necessary to set the strength of the connection between the units, that is, the weight, to an appropriate value. This weight is set by learning the neural network 3 as follows.

First, all weights are set at random, and input data for learning (data for which a desired output is known in advance) is given to each unit of the input layer. And
At this time, the output value output from each unit of the output layer is compared with the desired output value, and the value of each weight is sequentially corrected from the output layer side so as to reduce the difference (error). This is repeated using a large number of training data until the error converges ("Introductory Neurocomputer" by Toyohiko Kikuchi
January 20, 1990 Ohm Co., Ltd.).

Neural network 3 learned in advance in this way
In contrast, if thinning data of a digital radiological image to be recognized is input, an operation for identifying which imaging part / position is close to the imaging part / position is performed in a wide range of images in real time. In the system using the neural network 3 as in the present embodiment, since it is not necessary to extract a feature amount from a digital radiation image or create a dictionary of the feature amount, a comparison is made between the imaging part and the imaging position. It is possible to easily construct an identifiable system without any limitation.

That is, in recognizing a digital radiation image using the neural network 3, if an image range of a target to be recognized to some extent is determined in advance and the discrimination ability required in the image range is set, the neural network 3
By determining the unit configuration of, and inputting image data as a sample to the neural network 3 and learning it so that a desired answer comes out, a digital radiation image can be recognized with high accuracy. The pre-processing is simple only by the thinning-out processing, and the pre-processing can also simplify the configuration of the neural network 3, resulting in a simplified system.

When the number of recognition targets is increased, it is necessary to input the grayscale information of the image to the neural network 3 more precisely. Therefore, it is necessary to reduce the thinning ratio in the signal thinning unit 2. It is necessary to increase the number of constituent units, and a lot of time is required for learning. Therefore, it is preferable to limit the number of recognition targets to the minimum required number.

By the way, in synapse weight learning of the neural network 3, it is common to perform the body position identification processing in the neural network 3 using only one learning result. However, a plurality of learning results of the synapse weight are stored, If a plurality of recognition results obtained by processing by the neural network 3 networked based on the respective learning results (obtained by calculation by the discriminating means 3a) are used to determine the final position, It is possible to further improve the identification rate.

That is, as shown in FIG. 2, a learning result storage means 3c for storing a result (each synapse weight value) learned by the neural network learning means 3b for learning the neural network 3 is provided. In the discriminating means 3a for inputting a digital radiation image signal in which the number of pixels and the number of gradations are thinned out by the signal thinning-out unit 2 and discriminating a radiographing part and a radiographing position, based on each learning result stored in the learning result storage means 3c Each processing is performed, a majority decision of a plurality of recognition results derived from each learning result is taken, and finally the image processing unit 4 outputs the result.

Here, for example, if five types of learning results (synapse weights) are stored in the learning result storage means 3c,
Neural network 3 using each learning result
In the processing of (determination means 3a), when three types are identified as the front of the chest and the remaining two types are identified as the side of the head, a large number of the front of the chest is output as a final identification result. It is to let.

Usually, in learning the synapse weight of the neural network 3, the random number is often changed by learning using a random number as an unlearned initial value of each synapse weight, but the same processing is performed as described above. When a plurality of learning results to be provided are set, the result of learning using the initial value as a random number is used as the initial value of the next learning, and the learning result of the previous learning is further used as an initial value one after another. For example, the learning time can be greatly reduced when a plurality of learning results are obtained as compared with the case where learning is started using random numbers as initial values.

Further, in learning the neural network 3,
The results of processing a plurality of digital radiographic images, which are the basis of learning, by the signal thinning unit 2 are stored in the thinned image storage unit 6 together with their imaging sites and imaging positions, and the operator operates the console (see FIG. 2 for learning). Data selection means 7
Is equivalent. ), Necessary learning data is read out from the thinned-out image storage unit 6 as needed, and a neural network learning unit is provided based on the read learning data.
If 3b performs learning of the synapse shade in the neural network 3, it is possible to easily adapt learning that meets the demands of the operator while excluding data that hinders learning.

Even if the data for learning is stored in the thinned-out image storage unit 6 as described above, the number of thinned-out images is about 5 × 5 pixels to 10 × 10 pixels, as described above. Even if an image is stored for learning, the storage load on the device is light.

Further, in the present embodiment, both the imaging region and the imaging position are recognized by the neural network 3. However, when one of the region and the posture is fixed, only the other identification is performed by the neural network 3. I just want to be done.

<Effects of the Invention> As described above, according to the digital radiation image image recognition apparatus according to the present invention, the digital radiation image subjected to the signal decimation process is input to the neural network and recognized, so that There is no need to select a feature value to be extracted or match a dictionary of feature values, and a simple image recognition system can be easily constructed.On the other hand, in actual image recognition, a complicated It is possible to recognize an image in real time and at a high identification rate, and it is possible to perform optimal image processing for subsequent image interpretation based on information on an imaging part and an imaging position obtained by the recognition.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the configuration of the present invention, FIG. 2 is a system block diagram showing one embodiment of the present invention, and FIG.
FIGS. 4 and 5 are schematic diagrams showing the configuration of the illustrated neural network, and FIGS. 4 and 5 are diagrams showing examples of change characteristics of each unit of the neural network. 1 image input unit 2 signal thinning unit 3 neural network 3a determination unit 3b neural network learning unit 3c learning result storage unit 4 image processing unit 5 ... Image output unit, 6 ... Thinned image storage unit, 7
...... Learning data selection means

──────────────────────────────────────────────────続 き Continuation of front page (72) Inventor Hiroaki Tajima 1 Konica Corporation, Sakura-cho, Hino-shi, Tokyo (56) References JP-A-63-262137 (JP, A) JP-A-63-262139 ( JP, A) JP-A-2-96283 (JP, A) "Medical image processing" (1982-2-25) Asakura Shoten 112-113 (58) Field surveyed (Int. Cl. ⁶ , DB name) A61B ^6/ 00-6/14

Claims (4)

(57) [Claims] 1. A signal thinning means for reducing at least one of the number of pixels of a digital radiographic image and the number of signal quantization steps, wherein the characteristics of the digital radiographic image are learned in advance and the output of the signal thinning means is input. A neural network that outputs an identification signal indicating to which of a plurality of recognition targets the input image corresponds in advance, and learning is performed a plurality of times by changing the content of a learning data set using the neural network. Learning result storing means for storing a plurality of obtained learning results; and discriminating, based on each of the stored plurality of learning results, that the output of the signal thinning means finally outputs the largest one among a plurality of recognition targets. Means for recognizing an image of a digital radiographic image. 2. The digital radiation image image recognition apparatus according to claim 1, wherein the recognition target is at least one of an imaging region and an imaging position of a subject in the digital radiation image. 3. The signal thinning means divides an original image into a plurality of small areas, and samples the average value or the representative value of the image signal values of each of the small areas to thereby obtain the number of pixels of the digital radiation image and the signal of the signal. The digital radiation image image recognition apparatus according to claim 1, wherein at least one of the quantization step number and the quantization step number is reduced. 4. A plurality of learning results are obtained by performing initial neural network learning using unlearned initial values and performing learning using the previous learning results as initial values in the second and subsequent learning. The digital radiation image image recognition apparatus according to any one of claims 1 to 3, wherein: