JPH0458943A - Image recognizing device for digital radiation image - Google Patents

Abstract

PURPOSE:To eliminate the need of a complicated calculation, and to recognize a wide image with a high recognition rate by providing a neural network for inputting an output of a signal thinning-out means in which a feature is learned in advance, and outputting a discriminating signal for showing which of plural recognition objects the image corresponds to. CONSTITUTION:Digital radiation image data inputted from an image input part 1 is inputted to a signal thinning-out part 2 of the number of picture elements and the number of gradations, and thereafter, inputted to a neural. network 3, and by an operation (discriminating means 3a), a discriminating signal for showing which of recognition objects specified in advance the data corresponds to is outputted. Subsequently, it reaches an image output part 5 from an image processing part 4. In that case, in order to recognize the image, a feature of the image is learned. That is, a learning result storage means 3c for storing a result leaned by a neural net learning means 3b is provided. Subsequently, a result of processing by the signal thinning-out part 2 is stored in a thinned-out image storage part 6, and by an image data selecting means 7, arbitrary learning data is read outs and the learning means 3b executes learning of a synapse weight in the neural network 3.

Description

[Detailed Description of the Invention] <Industrial Application Field> The present invention relates to an image recognition device for digital radiation images, and more specifically, the present invention relates to an image recognition device for digital radiation images. The present invention relates to a device that can recognize the body part to be imaged and the body position to be imaged.

<Prior Art> Radiographic images such as X-ray images are often used for medical purposes, and examples of methods for obtaining radiographic images as electrical image signals include the following.

In other words, the radiation that has passed through the human body as an object is absorbed by a certain type of phosphor, and then this phosphor is excited with light or thermal energy, thereby removing the radiation energy that the phosphor has accumulated due to the absorption. It emits fluorescence and detects this fluorescence with a photoelectric conversion element to electrically obtain radiographic image information.The radiographic image signal is digitized and then subjected to gradation processing and spatial frequency processing to be displayed on a CRT, etc. There are some that are designed to be output and visualized (see Japanese Unexamined Patent Application Publication No. 189853/1983, etc.).

By the way, in the processing stage before outputting the digital radiation image signal as a visible image as described above, it is necessary to perform gradation processing and spatial frequency processing as described above to make a visible image suitable for interpretation. However, due to differences in the part of the human body to be photographed or the position to be photographed, the density of the region of interest in the part may change in each reproduced image.

Therefore, before performing image processing such as gradation processing and spatial frequency processing, it is required to automatically determine the imaging position from the original digital radiation image and apply the optimal image processing for each body position. Various methods and devices have been proposed that can automatically determine the imaging position using the density histogram of image data or the signal level distribution along a predetermined image direction in the image data (Japanese Patent Laid-Open No. 6
3-262128, JP-A-63-262132, etc.).

<Problems to be Solved by the Invention> However, in conventional imaging body position determination, it is only possible to determine whether the imaging position is front or side after limiting the imaging area to one area. There was a problem with the ability to identify shooting positions to be limited.

In order to solve this problem, even if you use in-depth image recognition technology to extract multiple feature quantities and perform discrimination based on a combination of these feature quantities, the processing flow and the calculations used for determination The equations become very complex, and it is necessary to match a dictionary of feature values for each body part and body position, which requires a lot of effort.
It was difficult to put it into practical use.

Furthermore, in conventional image recognition, it is difficult to quantitatively understand how effective the multiple extracted features are for recognition, so it is difficult to determine which features to extract and use for image recognition. Selection cannot be made easily, and from this point of view, it is difficult to apply conventional image recognition technology to the recognition of digital radiological images.Currently, recognition is limited to the area to be imaged and the body position to be imaged. Met.

The present invention has been made in view of the above-mentioned problems, and is an image recognition device for digital radiation images that does not require complicated calculations and also eliminates the labor of selecting feature quantities to be extracted and dictionary matching. The purpose is to provide

<Means for Solving the Problems> Therefore, in the present invention, as shown in FIG. a neural network that is trained in advance on image features, inputs the output of a signal thinning means, and outputs an identification signal indicating which of a plurality of pre-identified recognition targets the input image corresponds to; An image recognition device for radiation images was constructed.

Here, the recognition target may be at least one of the photographed region and the photographed body position of the subject in the digital radiation image.

In addition, the signal thinning means divides the original image into a plurality of small regions, and samples the average value or representative value of the image signal values in each of these small regions, thereby determining the number of pixels of the digital radiation image and quantizing the signal. It is preferable to reduce at least one of the numbers.

Furthermore, neural screw 1 to work are provided with a plurality of learning results obtained by learning multiple times while changing the contents of the learning data set in advance, and a plurality of neural screws processed and output based on each of these plural learning results are provided. It is preferable that the configuration is such that the identification signal with the largest number among the identification signals is finally output.

Here, if multiple learning results are provided as described above, the first neural network learning is performed using the unlearned initial values, and the previous learning results are used as the initial values for the second and subsequent learnings. It is preferable to configure the system so that a plurality of learning results can be obtained by performing learning as follows.

<Function> According to the image recognition device for a digital radiation image 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. be done. Digital radiological images whose data amount has been reduced in this way are
The features of the digital radiation image are input into a pre-trained neural network. The neural network processes the input signal according to learning that has been performed in advance, and outputs an identification signal indicating which of a plurality of recognition targets specified in advance the input image corresponds to.

In other words, by training a neural network in advance so that it can identify multiple recognition targets when a signal processed by the signal thinning means is input, it is possible to extract a feature quantity from a digital radiation image or Neural Screw 1~ without the need for effort such as creating a dictionary.
If an image signal whose data amount has been reduced by the signal thinning means is input to the workpiece, the image can be identified.

Here, what is required in the digital radiation image as the recognition target is, for example, at least one of the imaged part and the imaged body position of the subject, and in this case, the neural network uses an identification signal indicating where the imaged part is, Alternatively, an identification signal indicating whether the photographing position is frontal or sideward is output.

In addition, in order to reduce at least one of the number of pixels of the digital radiation image and the number of signal quantization steps using the signal thinning means, it is necessary to By sampling the values, the amount of data can be easily reduced and input to the neural network without requiring complicated calculation formulas or a large amount of calculation.

On the other hand, in a neural network, learning is performed multiple times while changing the contents of the training data set in advance, but it is provided with multiple learning results obtained through such learning, and is processed and output using each learning result. By taking a majority vote of the identified identification signals, it is possible to improve the recognition rate of digital radiation images, compared to the case where one recognition result of a neural network based on one learning result is used as the final result.

In addition, when providing multiple learning results as described above, the first neural network training is performed using unlearned initial values, and the second and subsequent learning uses the previous learning results as initial values. By performing this to obtain a plurality of learning results, the learning time can be shortened compared to the case where each learning is performed using an unlearned value as an initial value.

<Example> The present invention will be explained in detail below.

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

Here, the image input unit 1 includes a digital X-ray imaging device configured to scan and read a latent image of a radiation image stored (accumulated) on a stimulable phosphor, or a digital This may be an image scanner that converts it into a signal, or a filing system that already records a plurality of digital radiation image data.

The original digital radiation image data f (x, y) input from the image input section 1 is first input to the signal thinning section (signal thinning means) 2.

This signal thinning unit 2 thins out 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 determining the shooting position, and performs a neural Make sure that only the large grayscale features of the image are input to the network.

Specifically, the original digital image data f (x
, y) into multiple small areas (n pixels x n pixels) and calculate the average value of the signal values corresponding to each pixel included in each small area, or By determining one of the signal values corresponding to as a representative value, the original digital image data f (x,
The image data f'(
x', y'). Furthermore, 2048 x 2464
Even if the original digital image data f (x, y) with pixels and 1024 gradations is thinned out to about 5 x 5 pixels to 10XIO pixels and 256 gradations as shown in Fig. 3,
If the body parts and body positions to be photographed are limited to a certain extent, they can be sufficiently identified. For discrimination, the accuracy is sufficiently high.

Here, it is possible to reduce only the number of gradations by requantizing the original signal value with a large step size without thinning out the number of pixels, but this will complicate the neural network structure and learning described later. Therefore, 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 gradations have been reduced in the signal thinning unit 2 is input to the neural network (neurocomputer) 3 having the configuration as shown in FIG. An identification signal indicating whether the recognition object corresponds to one of the plurality of recognition targets specified in advance by the calculation by the neural network 3 (discrimination means 3a) is output.

For example, in the case of medical use where the subject is a person, the plurality of recognition targets include the front of the chest, the side of the chest, the front of the head, and the side of the head.

When the discrimination means 3a outputs an identification signal indicating which of the above-mentioned recognition targets corresponds, the image processing section 4 that receives this identification signal converts the original digital image data f from the image input section 1. For (x, y), various image processing and analyzes such as gradation conversion and frequency enhancement 2 contour extraction are performed based on the determination results of the imaged area and imaged body position so that it is reproduced as a visible image suitable for interpretation. I do.

The image output section 5, which receives the image data processed by the image processing section 4, records the image data (digital radiation image signal) after image processing again in the filing system or reproduces it on a CRT. or recorded on film by a printer. Here, when filing the image data in the filing system after determining the imaged part and imaged position of the subject (in this example, a human body), the imaged area is If the image data and the imaged body position are recorded in correspondence with the image data, it will be convenient for later retrieval, and there will be no need to repeatedly recognize the imaged body part and imaged body position.

By the way, the neural screws 1 to 3 are described in "Nikkei Electronics No. 427 J August 10, 1987.
As introduced on pages 115 to 124 of issue 8, it is a network that imitates the human brain and is made up of multiple units that correspond to the brain's neurons (nerve cells) and are connected together in a complex manner. By appropriately determining the connection form (coupling load) between each unit I, it is possible to embed (learn) a pattern recognition function or the like.

The neural network 3 in this embodiment is composed of an input layer S, an intermediate layer A, and an output layer R, as shown in FIG. Same number of neuron models (units) Sl
, S2...S6, and each neuron model S1. S2...The signal value of the pixel corresponding to S6 is manually entered. As mentioned above, the number of gradations is the original 1024024.
The gradation is reduced to approximately 56 gradations (8 pins) before being input to the neural networks 1 to 3. However, when inputting the signal value to the input layer S, the 0 to
The 256 gradations of 255 are normalized to 0.0 to 1.0 and input.

The neuron models A, A2゜...A, of the intermediate layer A are the two neuron models (units) S, , S of the input layer S.
2: Can be combined with all of S8. The intermediate layer A
The number a of neuron models in is determined empirically, but in the case of the present invention, the number S of neuron models in the input layer S and the number S of neuron models in the output layer R
The number r of neuron models is preferably about half of the total.

Further, the intermediate layer A may be one layer as in this embodiment, but
2 layers.

It may also have a multi-layer structure with three layers...

The output layer R recognizes the recognition target, for example, the front of the chest, the side of the chest,
When there are four types of head front and head side, it is composed of four neuron models R1 to R4, and these neuron models R1 to R4 are connected to each neuron model A,
, A2... Can be combined with all Aa.

For example, when the front chest is identified, the identification signal in the output layer R may be only the output of the neuron model R or 1.
The neuron model R7~R with
Depending on whether the output of 4 is 1 or not, image recognition (recognition of the body part and body position to be photographed) can be performed.

By the way, in order to input the signal values of each pixel after thinning to the neural network 3 and recognize the image as described above, the characteristics of the digital radiation image are learned in advance, and for example, the image signal of the front chest is It is necessary to learn in advance so that when input, only the output of the two-neelon model R1 becomes 1.

The learning of the neural network 3 is such that when a digital radiation image signal whose imaging region and imaging position are known in advance is processed by the signal thinning section 2 and input to the neural networks 1 to 3, the correct imaging region and imaging position are determined. A learning data set (
Neural network 3 while changing the thinned image signal input to neural network 3 for learning
The recognition rate is gradually increased by sequentially learning the value of each synaptic weight (connection weight) from the output layer R side, and by this learning (back propagation law), the recognition rate is By having the neural network 3 input the results of thinning out the digital radiographic image signals for which you want to identify the region to be imaged and the body position to be imaged by learning the differences in the input data, the region to be imaged and the body position to be imaged in the digital radiographic image can be determined in advance. This allows any one of a plurality of predetermined recognition targets to be identified.

Here, to further explain the characteristics of the neural network 3 and the learning based on the pack propagation law,
(neuron model for each unit constituting the neural network 3 as shown in Fig. 3) Ui converts the sum of inputs Qj from other units according to a certain rule and sets it as Qi, but the connection with other units Each section has a variable weight WiJ (synaptic way l-). This weight WIJ is used to express the strength of the connection between each unit, and changing this value will substantially change the structure of the network without changing the connection state.

Learning of the neural network 3 means changing this value, and the weight WiJ takes a positive value of 2 and a negative value, and zero represents the absence of a connection.

When input is received from a certain unit Ui or from multiple other units 1 = Ui, if the sum of the inputs is expressed as NET, then the sum of the inputs of the unit Ui is NETi = ΣWiJ
-Qj yells.

Each unit Ui applies this input summation NET to the function f and converts it into an output Q1 as shown in the following equation.

Qi=f (NETi) f (ΣWIJ-Qj) The above function f may be different for each unit Ui, but is generally a threshold numerical value as shown in FIG. 4 or a sigmoid function as shown in FIG. Use.

] 7 This sigmoid function is a differentiable pseudo-linear function and can be expressed as follows. The value range is θ to 1, which approaches 1 as the input value becomes larger and approaches 0 as it becomes smaller. When the input is 0, it becomes 0.5. In some cases, a threshold value θ (bias) is added.

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. In order to obtain the desired output, It is necessary to set the strength of the connection between each unit, that is, the weight, to an appropriate value.

This weight setting is performed by making the neural network 3 learn as follows.

First, set all the weights randomly,
Input data for learning (data whose desired output is known in advance) is given to each unit in the input layer. Then, at this time, the value of each weight is sequentially corrected from the output layer side so as to reduce the difference (error) between the output value output from each unit of the output layer and the desired output value. This is then repeated using a large amount of learning data until the error converges ("Introductory Neurocomputer J, Toyohiko Kikuchi, 1
January 20, 990, published by Ohmsha Co., Ltd.)
.

If the neural network 3, which has been trained in this way, is input with thinned-out data of the digital radiation image that you want to recognize, a wide range of calculations can be performed to identify which part of the image is being imaged or which body position is being imaged. However, in a system using a neural network 3 as in this example, it requires labor to extract the characteristic quantities from the digital radiation image and to create a dictionary of the characteristic quantities. Therefore, it is possible to easily construct a system that can identify the body part and the body position to be photographed without being relatively limited.

That is, in recognizing digital radiation images using the neural network 3, once an image range of the target to be recognized is determined to some extent in advance and the discrimination ability required for this image range is set, the rest is done by the units of the neural network 3. By determining the configuration and inputting sample image data to Neural Networks Work 3 and having it learn to come up with the desired answer, it is possible to recognize digital radiation images with high accuracy. The preprocessing or thinning process is simple, and the configuration of the neural network 3 can be simplified by such preprocessing, resulting in a simplified system.

Note that when increasing the number of recognition targets, it is necessary to input the grayscale information of the image into the neural network 3 in more detail, so it is necessary to reduce the thinning ratio in the signal thinning section 2. Since it is necessary to increase the number of constituent units and a lot of time is required for learning, it is preferable to limit the number of recognition targets to the minimum required number.

By the way, in the synaptic weight learning of the neural network 3, it is common to have the neural network 3 perform body position identification processing using only one learning result, but it is possible to store a plurality of learning results of synaptic weights, In order to make the final body position identification based on a majority vote of a plurality of recognition results obtained by processing the neural network 3 based on each learning result (obtained by calculation by the discrimination means 3a), It becomes possible to further improve the identification rate.

That is, as shown in FIG.
A learning result storage means 3c is provided for storing the results (each synapse weight value) learned by the neural net learning means 3b for learning the signal, and the signal thinning section 2 thins out the number of pixels and gradations. The discriminating means 3a inputs digital radiographic image signals and discriminates the radiographed region and the radiographed body position. A majority vote is taken from a plurality of recognition results to be finally output to the image processing section 4.

Here, for example, 5 types of learning results (synaptic weights)
If the learning results are stored in the learning result storage means 3c, the neural network 3 (
In the processing of the discriminating means 3a), if the three types are identified as the front of the chest and the remaining two types are identified as the side of the head, the final identification result is output as the front of the chest, which is the most common. It is something.

In addition, normally, when learning the synaptic weights of the neural network 3, a random number is often used as the unlearned initial value of each synaptic weight, and the random number is changed by learning. When setting multiple learning results to be provided, use the learning results with the initial value as a random number as the initial value for the next learning, and then perform further learning one after another using the previous learning result as the initial value. Compared to the case where learning is started using random numbers as initial values, the learning time can be significantly shortened when obtaining multiple learning results.

Furthermore, in the learning of the neural network 3, the results of processing a plurality of digital radiation images, which are the basis for learning, in the signal thinning unit 2 are stored in the thinned image storage unit 6 in advance, along with the imaging region and imaging position, and the operator reads out as much arbitrary learning data as necessary from the thinned-out image storage section 6 through a console operation (corresponding to the learning data selection means 7 in FIG. 2), and performs neural processing based on the read learning data. Online learning method 3
If the synaptic weights are learned in the neural network 3, it is possible to easily perform learning that meets the operator's requirements while eliminating data that would impede learning.

Incidentally, even if the learning data is stored in the thinned-out image storage unit 6 as described above, the thinned-out image is sufficient to have about 5×5 pixels to 10XIO pixels as described above, so it is not necessary to store multiple images. Even if the information is stored for learning purposes, the memory load on the device is light.

Furthermore, in this embodiment, the neural network 3 recognizes both the body part and the body position to be photographed. However, when either body part or body position is fixed, the neural network 3 only identifies the other. As you can see, it's fine.

<Effects of the Invention> As explained above, according to the image recognition device for digital radiation images according to the present invention, the digital radiation images that have been subjected to signal distortion processing are input to a neural network for recognition. , it is possible to easily build a simple image recognition system without having to select the features to be extracted in advance or to match a dictionary of features, while in actual image recognition, complex calculations are not required. A wide range of images can be recognized in real time with a high identification rate, and image processing optimal for subsequent image interpretation can be performed based on information about the imaging site and imaging position obtained through such recognition.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing the configuration of the present invention, Fig. 2 is a system block diagram showing an embodiment of the invention, and Fig. 3 is a block diagram showing the configuration of the present invention.
4 and 5 are diagrams each showing an example of change characteristics of each unit of the neural network. 1... Image input unit 2... Signal thinning section 3...
Neural network 3a...Discrimination means 3b.
...Neural network learning means 3C...Learning result storage means 4...Image processing unit 5...Image output unit 6...Thinned image storage unit 7...Learning data selection means

Claims (4)

[Claims] (1) Signal thinning means for reducing at least one of the number of pixels and the number of signal quantization steps of a digital radiographic image; the characteristics of the digital radiographic image are learned in advance; the output of the signal thinning means is input; What is claimed is: 1. An image recognition device for a digital radiation image, comprising: a neural network that outputs an identification signal indicating which of a plurality of pre-specified recognition targets the image corresponds to; (2) The image recognition apparatus for digital radiographic images according to claim 1, wherein the recognition target is at least one of a photographed region and a photographed body position of a subject in a digital radiographic image. (3) The signal thinning means divides the original image into a plurality of small regions, and samples the average value or representative value of the image signal values in each of the small regions, thereby increasing the number of pixels of the digital radiation image and the signal quantum. 3. The image recognition apparatus for digital radiographic images according to claim 1, wherein the apparatus is configured to reduce at least one of the number of conversion steps. (4) The neural network has a plurality of learning results obtained by learning multiple times while changing the contents of the learning data set in advance, and a plurality of identification signals processed and output based on each of the plurality of learning results. 4. The image recognition apparatus for digital radiation images according to claim 1, wherein the apparatus is configured to finally output the largest number of images. (5) The configuration is configured so that the first neural network learning is performed using unlearned initial values, and in the second and subsequent learnings, multiple learning results are obtained by performing learning using the previous learning results as initial values. The image recognition device for digital radiation images according to claim 4, characterized in that: