JPH09179977A - Automatic processor for intensity level of medical image - Google Patents

Abstract

PROBLEM TO BE SOLVED: To set adequate intensity level conversion characteristics without being affected by a difference in image kind. SOLUTION: The automatic processor for the intensity level of medical image is equipped with plural arrays of parameter calculation parts CL1,..., CLN which calculate parameters for intensity level settings for setting intensity level conversion characteristics on the basis of the density value state of input pixels of the medical image, and each parameter calculation part learns a sample image of a corresponding image kind and sends out the parameters for intensity level setting of the medical image, inputted according to the decision result of the image kind decision part 5, to an intensity level conversion part 7. According to the intensity level conversion characteristics set properly by the intensity level conversion part 7, intensity level conversion for the medical image is performed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a medical image obtained by an image diagnostic apparatus such as a nuclear magnetic resonance tomography apparatus (MRI apparatus), an X-ray CT apparatus, an emission CT apparatus (nuclear medicine apparatus). The present invention relates to a gradation automatic processing device for medical images in which the density value of an input pixel is automatically converted into a density value of an output pixel according to a gradation conversion characteristic (hereinafter, appropriately abbreviated as “gradation automatic processing device”), The present invention relates to a technique for appropriately converting the density value of an input pixel to achieve a medical image display with good contrast regardless of the kind.

[0002]

2. Description of the Related Art An MRI apparatus, which is one of the image diagnostic equipments currently used in hospitals and the like, has a built-in gradation processing apparatus. With this apparatus, an input pixel of a medical image of an affected area of a patient is captured. The density value of is converted into the density value of the output pixel according to the gradation conversion characteristic, and then displayed as a medical image on a monitor or the like. An apparatus for automatically performing such gradation conversion processing is disclosed in, for example, Japanese Patent Laid-Open No. 5-6197.
There is an apparatus described in Japanese Patent No. 3 publication. In this conventional gradation automatic processing apparatus, the apparatus itself has a learning function using a neural network, and a neural network is made to learn a medical image (example image) as a model in advance. , The gradation conversion characteristics are set based on the result of examining the density value status of the input pixel of the medical image (hereinafter referred to as “target image” as appropriate) which is the target to be gradation-converted. The density value of the input pixel of the processed image is automatically converted into the density value of the output pixel according to the tone conversion characteristic.

[0003]

However, in the conventional gradation automatic processing apparatus, even if the model image is preliminarily learned by the neural network, the gradation conversion characteristic is caused by the difference in the image type of the medical image. Is not set properly.

Taking an MRI apparatus as an example, the so-called angio method, IR (inversion recovery) method, S
There are image types such as E (spin echo) method and FE (field echo) method. When there are various image types as described above, an appropriate gradation conversion characteristic is not set for an image type having a small number of learning images due to insufficient learning. For example, 50
Even if it is possible to set an appropriate gradation conversion characteristic for the image type that has learned the five model images, it is appropriate for the image type that has only learned the five model images. It is hard to expect to set different gradation conversion characteristics.

Not only that, even when a sufficient number of model images are learned for each image type, mutual setting between different image types prevents the setting of appropriate gradation conversion characteristics. This is because, in the case of learning using a neural network, not only a specific model image but all model images are reflected in the learning result. 5 for each of the two different image types
When learning 0 model images, to put it simply, it is possible to have an appropriate gradation conversion characteristic for intermediate image types of both image types, On the other hand, an appropriate gradation conversion characteristic cannot be obtained.

In view of the above circumstances, it is an object of the present invention to provide a medical image gradation automatic processing apparatus capable of setting an appropriate gradation conversion characteristic without being affected by a difference in image type.

[0007]

In order to achieve the above-mentioned object, the present invention has the following configuration. That is, in the medical image gradation automatic processing apparatus according to the present invention, the density value of the input pixel of the medical image obtained by the image diagnostic apparatus is automatically converted into the density value of the output pixel according to the gradation conversion characteristic. In a configured device, the image type determination unit that determines the type of the medical image, and the gradation conversion characteristic based on the density value status of the input pixels of the medical image that is input according to the determination result of the image type determination unit. A plurality of columns of parameter calculating means for calculating a gradation setting parameter for setting, and a gradation conversion setting means for setting the gradation conversion characteristic based on the gradation setting parameter are provided, and the parameter Each of the calculating means has a histogram creating means for creating a density histogram of the input pixel and an input of the density histogram to calculate the gradation setting parameter. And the weights in the neural network for the gradation setting trial calculation parameters actually calculated by the neural network for the model image of the corresponding image type and the gradation setting separately determined as appropriate for the model image. The learning means determines the learning based on the deviation from the teacher parameter by causing the neural network to perform learning.

[0008]

The operation for setting the gradation conversion characteristic in the automatic medical image gradation processing apparatus of the present invention will be described. Before actually carrying out the gradation conversion of the medical image, the parameter calculating means for calculating the gradation setting parameter for setting the gradation conversion characteristic must be learned in advance. Each parameter calculation means provided for each image type is made to sufficiently learn the model image of the corresponding image type.

At the time of learning, in the specific parameter calculation means, the histogram creation means first creates a density histogram of the input pixels of the model image belonging to the corresponding image type, and the neural network is modeled on the basis of the density histogram created. Calculate the trial calculation parameters for gradation setting for the image. On the other hand, with respect to the model image, there is prepared a teacher parameter for gradation setting that is separately determined by an expert as appropriate, and the learning means provided in the parameter calculation means is used for the gradation setting for the model image. The weight in the neural network is changed by causing the neural network to perform learning based on the difference between the trial calculation parameter and the gradation setting teacher parameter. The above learning is repeated for a large number of model images of the same image type, and when the weight in the neural network is determined to a value suitable for the corresponding image type, the learning of one parameter calculation means ends. Subsequently, learning of the model image of the corresponding image type different from the other is similarly performed by other parameter calculating means, and when the weights in all the neural networks are all determined, the learning is completed,
It is possible to actually perform gradation conversion on a medical image.

When performing gradation conversion, the image type discriminating means first discriminates the image type of the target image and sends the target image to the parameter calculating means for the image type to which the target image belongs. In the parameter calculation means, the histogram creation means creates a density histogram of the input pixels of the target image and sends it to the neural network, and the weighted sum is calculated based on the weight and the density histogram determined by learning by the neural network and the target image is calculated. The gradation setting parameter is calculated and output to the gradation conversion setting means. On the other hand, the gradation conversion setting means sets the gradation conversion characteristic for the target image based on the calculated gradation setting parameter. The density value of the input pixel of the target image is automatically converted into the density value of the output pixel according to the gradation conversion characteristic set in this way.

[0011]

Next, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a schematic configuration of a gradation automatic processing apparatus according to an embodiment integrated with an MRI apparatus. The MRI apparatus 1 is an apparatus that receives a nuclear magnetic resonance signal emitted from a subject and performs tomographic imaging. The measured image is first held in the image reconstruction unit 2 in a three-dimensional manner, Is designated, the corresponding medical image is sent from the image reconstructing unit 2 to the gradation automatic processing device, and the gradation automatic processing device performs appropriate gradation conversion on the medical image before displaying it on the monitor 3. It is designed to be printed on a film.

The medical image in the form of a digital signal sent from the image reconstructing section 2 is stored in the multi-valued image memory 4 and at the same time sent to the image type discriminating section 5 to discriminate the image type. Based on the discrimination result of the image type discrimination unit 5, the switch S
Send a contact closing control signal to A to SC. When the medical image is determined to be the image type Vn, the contact SAn of the switch SA, the contact SBn of the switch SB, and the contact SCn of the switch SC (where n = 1, 2, ... N) are closed. FIG. 1 shows a case where the medical image is the image type V1. .., CLN of N columns are provided in the subsequent stage of the switch SA. Here, the density value status of the input pixel of the medical image input according to the discrimination result of the image type discrimination unit 5 is provided. Therefore, the gradation setting parameter for setting the gradation conversion characteristic is calculated.

A teacher input unit 6 is provided in front of the switch SB, and a parameter calculation unit to which a model image in the parameter calculation units CL1, CL2, ..., CLN is input at the time of learning by a sample image. A teacher parameter for setting the gradation of the model image is also sent from the teacher input unit 6 via the switch SB. This gradation setting teacher parameter is a parameter that an expert such as a doctor or operator observes the model image in advance and separately determines as appropriate. The configuration of each of the parameter calculation units CL1, CL2, ..., CLN will be described later in detail, but can be regarded as substantially the same.

Parameter calculation units CL1, CL2, ...
The gradation conversion unit 7 is connected to the latter stage of CLN via a switch SC, and the gradation conversion unit 7 sets gradation conversion characteristics in accordance with the gradation setting parameters sent from the respective parameter calculation units. To be done. When the gradation conversion characteristic is set by the gradation conversion unit 7, the density value of the input pixel of the medical image is sequentially converted by the gradation conversion unit 7 into the density value of the output pixel and stored in the output image memory 8. , The density value of the output pixel is D
The analog image is converted by the A converter 9 and then sent to the monitor 3 to display a medical image. Hereinafter, the configuration of each unit will be described more specifically.

The multi-valued image memory 4 stores the density value of the input pixel of the medical image, and the stored density value is in a wide range of, for example, -4000 to +8000. Further, the image type determination unit 5 reads the image type identification data attached to the medical image data together with the information such as the name of the subject, the date of imaging, and the like, and determines the image type. In the case of the MRI apparatus, there are image types such as angio method, IR method, SE method, and FE method.

Before entering the description of the parameter calculation section, the gradation conversion characteristics and gradation setting parameters in the present invention will be described. The density value of the input pixel of the medical image stored in the multi-valued image memory 4 is as described above.
The output pixel has a wide range of 4000 to +8000, while the output pixel has a narrower range of 0 to 255, for example. Therefore, gradation conversion is performed so that the density value of the input pixel of the region of interest in the medical image matches the density value of the output pixel.
The broken line in the graph of FIG. 2 indicates the gradation conversion characteristic,
This polygonal line is defined by the window level L and the window width W.

In this example, the density value of the input pixel at the window level L corresponds to the intermediate density value (brightness) of the output pixel. On the other hand, all the density values of the input pixels that are equal to or larger than the upper limit of the window width W are the upper density values of the output pixels.
55, the density values of the input pixels below the lower limit of the window width W are all 0, which is the lower limit density value (black level) of the output pixels. If the window level L and the window width W are set appropriately (optimum), the range of the density value of the input pixel of the region of interest is well matched to the range of the density value of the output pixel, and the contrast of the medical image is very good. Become. Therefore, in the execution stage, the parameter calculation unit calculates the optimum window level L and the optimum window width W as the gradation setting parameters and sends them to the gradation conversion unit 7.

Next, the structure of the parameter calculation unit CL1 will be described. The parameter calculation unit CL1 is for the image type V1, and the pixel extraction unit 11 extracts pixels belonging to the region of interest from the input pixels of the medical image sent from the multivalued image memory 4. The pixel extracting unit 11 finally extracts a pixel satisfying both of the following two extraction conditions A and B as a pixel (interested pixel) related to a site (interested region) to be observed. Note that FIG. 3 is a block diagram showing a specific configuration of the pixel extraction unit 11.

Extraction condition A: As shown in FIG. 4, the pixel belongs to the contour region G in the imaged subject. The region (region of interest) that the doctor wants to observe at the time of diagnosis is not the uniform flat part in the image, but the contour region G of the lesion or tissue, and thus the pixel located in the contour region G is the pixel of interest. To do.

Extraction condition B: As shown in FIG. 4, the pixel has a density value close to that of the pixel in the central portion C of the image. An example of how to take the central portion C is a square in which the length of the image is 1/4 each of the vertical and horizontal sides, and the center of the image is the center. It may be one. Since the imaging is usually performed so that the region of interest is located in the center of the image, a pixel having a density value close to the density value of the pixel in the central portion C of the image is extracted. With this, inappropriate pixels around the image having a density value far from the central density value are removed.

The extraction of the image satisfying the extraction condition A is as follows. The CPU 16 shown in FIG. 3 performs the differential processing for each input pixel of the medical image stored in the multi-valued image memory 4, and stores the differential value for each pixel in the differential value memory 12. Since the density change is large in the contour part G, the differential value D is higher than the density change on both sides of the contour part G where the density change is small. Therefore, when a pixel whose differential value D is larger than an appropriate constant value (threshold level) is taken out, it becomes a pixel belonging to the contour part G. The extraction memory 13 is a memory having the same matrix configuration as the multi-valued image memory 4, and “1” is assigned to the address corresponding to each pixel satisfying the extraction condition A.
Is stored, and “0” is stored in other addresses.

The extraction of the image satisfying the extraction condition B is as follows. Pixels having the same density value as the pixels located in the central portion C of the medical image stored in the multi-valued image memory 4 are extracted. The average value Q1 and the standard deviation Q2 of the density values of all the pixels located in the central portion C are obtained to determine the density range CA that serves as the extraction reference by the following formula. (Q1-K1 · Q2) ≦ CA ≦ (Q1 + K2 · Q2) K1 and K2 are constants of about 5 to 10, and usually about 7. Pixels having density values in the density range CA are extracted. The extraction memory 14 is a memory having the same matrix configuration as the multi-valued image memory 4, in which “1” is stored in the address corresponding to each pixel satisfying the extraction condition B, and “0” is stored in the other addresses.

An image satisfying both extraction conditions A and B is a pixel for which "1" is stored in both extraction memories 13 and 14. An AND operation of the stored contents of the corresponding addresses of the extraction memories 13 and 14 is performed, the pixel having the operation result of “1” is determined as the pixel of interest, and the density value of each pixel of interest is stored in the pixel of interest memory 15. . The above-mentioned extraction processing and arithmetic processing are mainly executed by the CPU 16. Extraction of the pixel of interest is effective in realizing a medical image display in which the region of interest can be easily observed. In addition, the extraction condition of the pixel of interest is not limited to the above-mentioned aspect, and for example,
There is a mode in which only one of the extraction conditions A and B is used as the extraction condition.

The histogram creating section 17 creates the density histogram shown in FIG. 5 from the density value data of the input pixels extracted as the pixel of interest in the pixel of interest memory 15. In this density histogram, the density value range of the input pixel is, for example, 3
The number of pixels having a density value in each section is taken as the vertical axis, while the 0 axis (30 levels) is arranged in order so that the one having the lower density value is on the origin side. . Then, in order to make the density histogram into a form that can be easily processed by the next neural network 18, as shown in FIG.
Is performed to obtain a standard density histogram.

The following neural network 18 calculates the gradation setting parameter based on the standard density histogram. That is, the neural network 18 calculates the standard window level and standard window width. As shown in the conceptual model of FIG. 7, the neural network 18 is composed of a plurality of neurons (elements having a weighted sum function similar to nerve cells).
The NUs are arranged in layers and each neuron NU is network-connected by a hand called a synapse connection. Here, the input layer 18a, the intermediate layer 18b, and the output layer 18c
It is a layered model. The input layer 18a has the same number (30) of neurons NU as the class number of the standard density histogram shown in FIG. 6, and the intermediate layer 18b has 20 neurons N.
U, the output layer 18c comprises two neurons NU. The number of neurons in the output layer 18c is adjusted to "2" which is the number of output items of standard window level and standard window width. The neural network 18 can be configured by hardware or realized by software.

In order for the neural network 18 to work properly, it is necessary to determine the weight (weighting factor) of the synaptic connection in advance. The learning unit 19 is the neural network 1
The determination of the weight in 8 is performed by the gradation setting trial calculation parameter calculated by the neural network 18 for a large number of model images belonging to the image type V1 and the gradation setting teacher parameter separately determined as appropriate for the model image. This is realized by causing a neural network to perform learning based on deviation.
From the teacher input unit 6, the teacher window level WLT and the teacher window width W are set as the teacher parameters for gradation setting.
Two WTs are sent. The normalization processing unit 20 performs the following normalization processing so that the teacher window level WLT and the teacher window width WWT are in the range of 0 to 1 suitable for learning in the neural network 18.

Standard teacher window level WLt = (teacher window level WLT-minimum density value of model image) /
[(Maximum density value of model image-minimum density value of model image) x
K3] Standard teacher window width WWt = (teacher window level WWt) / [(maximum density value of model image−minimum density value of model image) × K4] where constant K3, K4> 1, minimum density value, maximum The density values are the minimum and maximum density values of the pixels extracted by the pixel extraction unit 11.

The trial calculation parameter for gradation setting calculated and output by the neural network 18 is the trial calculation window level W.
Lm and trial calculation window width WWm. The closer the trial calculation window level WLm is to the standard teacher window level WLt and the closer the trial calculation window width WWm is to the standard teacher window width WWt, the more appropriate the weight in the neural network 18 is (the internal state of the neural network is appropriate). become. That is, the standard teacher window level WLt and the standard teacher window width WW
The learning unit 19 causes the neural network 18 to learn with t as a target value.

The denormalization processing unit 21 inputs the standard optimum window level wl and the standard optimum window width ww outputted by the neural network 18 which has finished learning, and performs conversion of the following equations to optimize the optimum window level WL and the optimum window width. Change to WW. That is, the neural network 18
What has come out from the above is a form suitable for the neural network, and the denormalization process is performed so as to correct it in a form suitable for the gradation converting unit 7.

Optimum window level WL = minimum density value of pixel + standard optimum window level wl × [(maximum density value of pixel−minimum density value of pixel) × K3] optimum window width WW = standard optimum window width ww ×
[(Maximum density value of pixel−minimum density value of pixel) × K4] However, the constants K3 and K4 are the same as K3 and K4 in the formula for the normalization process for the above-mentioned gradation setting teacher parameter in the same parameter calculation unit. Must match. Also,
The minimum density value and the maximum density value are the minimum and maximum density values of the pixels extracted by the pixel extraction unit 11.

The other parameter calculation units CL2 to CLN have the same configuration. However, the extraction condition in the pixel extraction unit, the number of density value divisions in the histogram creation unit, the number of neurons in the neural network, or the like may be appropriately changed according to the corresponding image type, etc., and all parameter calculation units CL1 to CL1. The CLNs do not have to have the exact same configuration.

The gradation conversion unit 7 determines the optimum window level W
The gradation conversion characteristic is set by a look-up table method according to L and the optimum window width WW. The density value of the input pixel of the target image corresponds to the input address, and the density value of the output pixel is stored in the memory cell at each address. Of course, the gradation conversion characteristic may be set as a mathematical expression and the conversion processing may be performed by software. The density value of each input pixel of the target image is gradation-converted and stored in the output image memory 8 one after another.

Next, the learning operation and gradation conversion operation in the gradation automatic processing apparatus of the above embodiment will be described. First,
The learning operation will be described with reference to the flowchart of FIG. 8 showing the flow of learning. [Step S1] The model image of the image type V1 is sent from the MRI apparatus 1 to the multi-valued image memory 4.

[Step S2] The image type determination unit 5 determines the image type V1 and closes the contact SA1 of the switch SA and the contact SB1 of the switch SB.

[Step S3] While the model image is input to the pixel extraction unit 11 of the parameter calculation unit CL1, the teacher window level WLT and the teacher window width WWT of the input model image are input from the teacher input unit 6 to the parameter calculation unit. It is sent to the standardization processing unit 20 of CL1.

[Step S4] The pixel extraction unit 11 extracts a pixel of interest that satisfies both the extraction condition A and the extraction condition B from the pixels of the model image.

[Step S5] The normalization processing unit 20
Using the minimum and maximum density values of the extracted pixels, the teacher window level WLT and the teacher window width WW
The T is converted into the standard teacher window level WLt and the standard teacher window width WWt, and sent to the learning unit 19.

[Step S6] Histogram creation unit 1
Reference numeral 7 creates a density histogram for the extracted pixel of interest and sends the standard density histogram created by the standardization processing to the neural network 18.

[Step S7] In the pre-learning neural net 18, the weight is not an appropriate value, but when the standard density histogram is input, the neuron exerts the weighted sum function, and the trial calculation window level WLm and the trial calculation window are displayed. The width WWm is calculated and sent to the learning unit 19.

[Step S8] The learning section 19 determines the standard teacher window level WLt and the trial calculation window level WL.
A difference (deviation) of m and a difference (deviation) of the standard teacher window level WLt and the trial calculation window width WWm are calculated and sent to the neural network 18 as parameter errors.

[Step S9] The neural network 18
Then, learning of the known back propagation method is executed based on the parameter errors ΔWL and ΔWW. In the case of the back propagation method, the learning signal is obtained in the direction opposite to the above-described parameter trial calculation process. That is, the learning signal in the neuron of the output layer 18c is obtained from the parameter errors ΔWL and ΔWW, and the error of the neuron in the output layer 18b and the learning signal corresponding thereto are obtained from this result,
From this result, the error of the neuron in the output layer 18a and the learning signal corresponding to it are also obtained, and finally the weight in the neural network 18 is changed so that the error is reduced based on these learning signals. The learning of the back propagation method is also described in Japanese Patent Laid-Open No. 61973/1993.

[Step S10] Neural Network 1
When the weight in 8 is changed, the learning of the one model image input this time is finished, and if there are more model images to be learned, the process returns to step S1. If a series of prepared model images have been learned once and no model image remains, the process proceeds to step S10.

[Step S11] If all the sample images of the image type V1 are sequentially learned in accordance with steps S1 to S10 as one learning cycle, and the number of executed learning cycles reaches the predetermined number of cycles. , Step S13
Proceed to. If the scheduled number of cycles has not been reached, the process proceeds to step S12.

[Step S12] The order of the model images of the image type V1 is changed, and the process returns to step S1.

[Step S13] If there is an unlearned image type, the process proceeds to step S14. There are no unlearned image types,
If the learning for all image types has been completed, go to [Step END].

[Step S14] The image type of the model image is changed, and the process returns to step S1.

[Step END] All learning in the parameter calculation section is completed (END), and preparation for executing gradation conversion for an actual medical image is ready. In the learning process above, for one image type,
For example, one learning cycle of preparing 100 model images and continuously learning 100 model images is repeated 1000 to 10,000 times while changing the input order of the model images.

Next, the gradation conversion operation by the gradation automatic processing apparatus which has completed the learning stage will be described with reference to the flow chart of FIG. 9 showing the flow of gradation conversion. [Step T1] The medical image to be subjected to gradation conversion is sent from the MRI apparatus 1 to the multi-valued image memory 4.

[Step T2] When the image type determination unit 5 determines that the image type is V1, for example, a control signal is output from the image type determination unit 5 and the contact SA1 of the switch SA and the contact SC1 of the switch SC are closed.

[Step T3] The medical image is input to the pixel extraction unit 11 of the parameter calculation unit CL1.

[Step T4] The pixel extracting unit 11 extracts a pixel of interest satisfying both the extraction condition A and the extraction condition B from the pixels of the medical image.

[Step T5] Histogram Creation Unit 1
Reference numeral 7 creates a density histogram for the extracted pixel of interest and sends the standard density histogram created by the standardization process to the neural network 18.

[Step T6] In the learned neural net 18, the weight is of course an appropriate value, and when the standard density histogram is input, the neuron NU exerts the weighted sum function, and the standard optimum window level wl and standard The optimum window width ww is sent to the denormalization processing unit 21.

[Step T7] The denormalization processing unit 21 denormalizes the standard optimum window level wl and the standard optimum window width ww to generate the optimum window level WL.
And the optimum window width WW, and then sends it to the gradation converting unit 7 via the switch SC.

[Step T8] The gradation conversion unit 7 sets optimum gradation conversion characteristics in accordance with the received optimum window level WL and optimum window width WW.

[Step T9] The density value of the input pixel of the medical image is converted into the density value of the output pixel according to the set gradation conversion characteristic and stored in the output image memory 8.

[Step T10] The density value of the output pixel of the output image memory 8 is sequentially converted into an analog signal by the DA converter 9 and is sent out to display a medical image with good contrast on the screen of the monitor 3.

The present invention is not limited to the above embodiment, but can be modified as follows, for example. (1) In the above embodiment, the automatic gradation processing device and the MRI device are integrated, but the automatic gradation processing device and the MRI device are completely separate structures, and Another example is a configuration in which a medical image to be converted is supplied by a magnetic disk or the like.

(2) In the above embodiment, the medical device is the MRI.
Although it was a device, it goes without saying that the medical device may be a device of a different type such as an X-ray CT device.

[0060]

According to the automatic medical image gradation processing apparatus of the present invention, since the parameter calculating means having the learning function for the model image is provided for each image type, mutual processing between different image types is possible. It is possible to sufficiently learn the model image of each image type without causing interference, set appropriate gradation conversion characteristics without being influenced by the difference of the image type, and realize sufficient image display with good contrast. it can.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a schematic configuration of an automatic gradation processing apparatus of an embodiment.

FIG. 2 is a graph illustrating a gradation conversion characteristic applied to a medical image.

FIG. 3 is a block diagram showing a configuration of a pixel extraction unit of the device of the embodiment.

FIG. 4 is a plan view schematically showing an example of a medical image.

FIG. 5 is a graph showing a density histogram.

FIG. 6 is a graph showing a standard density histogram.

FIG. 7 is a schematic diagram showing a configuration of a neural network in the apparatus of the embodiment.

FIG. 8 is a flowchart showing a flow of a learning operation in the apparatus of the embodiment.

FIG. 9 is a flowchart showing a flow of a gradation conversion operation in the apparatus of the embodiment.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... MRI apparatus as medical equipment 5 ... Image type discrimination part 7 ... Gradation conversion part 11 ... Pixel extraction part 17 ... Histogram creation part 18 ... Neural network 19 ... Learning part CL1 ... Parameter calculation part CL2 ... Parameter calculation part CLN ... Parameter calculator

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification number Office reference number FI technical display location H04N 1/405 H04N 1/40 B

Claims (1)

[Claims] 1. An automatic gradation processing of a medical image, wherein the density value of an input pixel of a medical image obtained by an image diagnostic device is automatically converted into a density value of an output pixel according to a gradation conversion characteristic. In the apparatus, an image type determining unit that determines the type of the medical image, and to set the gradation conversion characteristic based on the density value status of the input pixels of the medical image that is input according to the determination result of the image type determining unit. A plurality of columns of parameter calculating means for calculating the gradation setting parameters and a gradation conversion setting means for setting the gradation conversion characteristics based on the gradation setting parameters are provided, and each of the parameter calculating means A histogram creating means for creating a density histogram of input pixels, and a neural network for inputting the density histogram to calculate the gradation setting parameter And a weight in the neural network, a gradation setting trial calculation parameter actually calculated by the neural network for the model image of the corresponding image type, and a gradation setting teacher parameter separately determined appropriately for the model image. And a learning unit that determines the learning based on a deviation of the neural network by a neural network.