JP2010277231A - Data processing apparatus, method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a data processing apparatus, a method, and a program capable of calculating reliability of the amount of learning obtained by a regression function. <P>SOLUTION: The apparatus includes an input unit 14 for inputting data, a storage unit 15 for storing the inputted data, and a control unit 11. In the data processing apparatus 10, the control unit 11 calculates a first estimated value of the newest input data based on past data which is inputted in the past and stored currently and the newest input data, and calculates reliability of the calculated first estimated value based on the data amount of the past data or the feature amount of the past data. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a data processing apparatus, method, and program.

Typical learning systems include classification problems and regression problems.
In the classification problem, an estimated value (classification result) for the input data is calculated by a learning classifier that has been learned based on past data. For example, when the input data is a radiographic image, a classification result of the chest, abdomen, and the like to be classified with respect to the input radiographic image is calculated. That is, in the classification problem, part recognition is performed by the classifier.

In the regression problem, one estimated value (learning amount) is calculated with respect to input data by a regression function learned based on past data. When the input data is, for example, a radiographic image, a learning amount such as density or contrast to be adjusted with respect to the input radiographic image is calculated.
The input radiation image is adjusted in density, contrast, or the like based on the calculated learning amount, and the adjusted radiation image is displayed on the display unit or the like.

  Here, according to the learning system, the data processing in the classification problem and the regression problem is performed based on the input data (past data) input in the past as described above. That is, a classification result or a learning amount (hereinafter collectively referred to as “learning result”) is calculated by learning from past data. Therefore, when the amount of past data is insufficient, the reliability of the obtained learning result is low.

  According to Patent Document 1, a technique is disclosed that includes a standard model prepared in advance and adapts parameters of the standard model with a small amount of past data (see Patent Document 1).

Japanese Patent Laid-Open No. 8-95592

  However, according to the technique of Patent Document 1, when a small amount of past data is within a local range, the reliability of the learning result obtained by the adapted standard model is the learning result obtained by the default standard model. Lower than.

  An object of the present invention is to provide a data processing device, method, and program capable of calculating the reliability of the first estimated value (learning amount) of the adjustment amount obtained by the regression function.

According to the present invention,
An input section for inputting data;
A storage unit for storing the input data;
A control unit,
The control unit calculates a first estimated value for the latest input data based on past data and the latest input data that have been input and are currently stored, and the calculated first estimated value Is provided based on the data amount of the past data or the feature amount of the past data.

Moreover, according to the present invention,
A process of inputting data by an input unit;
Storing the input data by a storage unit;
A step of controlling by the control unit,
The controlling step includes a step of calculating a first estimated value of the latest input data based on past data and the latest input data that have been input and are currently stored, and the calculated first estimate And a step of calculating a reliability of the value based on the data amount of the past data or the feature amount of the past data.

Moreover, according to the present invention,
Computer
Input means for inputting data,
Storage means for storing the input data;
A program that functions as a control means,
The control means calculates the first estimated value of the latest input data based on the past data and the latest input data that have been input and are currently stored, and the reliability of the calculated first estimated value. A program for calculating the degree based on the data amount of the past data or the feature amount of the past data is provided.

  According to the present invention, based on the reliability of the first estimated value (learning amount) of the adjustment amount obtained by the regression function, the learning amount by the regression function or the standard function can be optimized to calculate one learning amount. it can. Therefore, even if a small amount of past data is within a local range, a learning amount with higher reliability than the learning amount obtained by the standard function can be calculated.

It is a functional block diagram of a data processor. It is a flowchart which shows the outline | summary of a data process. It is a flowchart which shows a learning amount calculation process. It is a conceptual diagram of a learning amount calculation process. It is a flowchart which shows a reliability calculation process. It is a figure which shows the relationship between an image number reliability coefficient and the number of images. It is a figure which shows the relationship between a feature quantity reliability coefficient, an average distance, and a distance threshold value. It is a flowchart which shows a distance threshold value determination process. It is a conceptual diagram of distance threshold value determination processing. It is a flowchart which shows another data processing. It is a flowchart which shows a classification result calculation process. It is a flowchart which shows a classification process. It is a figure which shows an output value determination table.

The configuration and operation of the data processing apparatus in this embodiment will be described in detail with reference to the drawings. In addition, this embodiment is only an example of embodiment in this invention, and this invention is not limited to this.
Hereinafter, the case where input data is a radiation image will be described as an example. For example, audio data can be applied other than the radiographic image.

FIG. 1 shows a functional configuration of the data processing apparatus 10.
The data processing apparatus 10 includes a control unit 11, an operation unit 12, a display unit 13, an input unit 14, a storage unit 15, and the like.

The control unit 11 includes a CPU (Central Processing Unit), a RAM (Random Access Memory), and the like.
The control unit 11 performs various calculations in cooperation with a program stored in the storage unit 15 and performs centralized control of operations of the units (12 to 15).

The operation unit 12 includes a keyboard, a mouse, and the like.
The operation unit 12 generates an operation signal by a user operation, and outputs the generated operation signal to the control unit 11.

  The display unit 13 includes a display and the like, and displays various operation screens and radiation images according to display control of the control unit 11.

  Although the display unit 13 displays a radiographic image here, the display target is not limited to the radiographic image, and a general digital image can also be displayed. Similarly, in the following description, the present invention is not limited to the “radiation image”.

The input unit 14 includes a communication interface and the like.
The input unit 14 inputs a radiation image from a radiation imaging apparatus installed on the same network.

The storage unit 15 includes a nonvolatile semiconductor memory such as a ROM (Read Only Memory).
The storage unit 15 stores system programs and various processing programs, and stores data obtained by various processes. Further, the storage unit 15 stores the image input by the input unit 14.

With reference to FIG. 2, data processing in this embodiment will be described.
Here, data processing in the regression problem will be described.

The control part 11 inputs the radiographic image output from a radiography apparatus by the input part 14 (step S1).

The control unit 11 calculates a feature amount of the input radiation image (input data) (step S2).
The feature amount is a value calculated based on the size (number of pixels) of the shooting target, the shape of the center line of the shooting target, the shape of the density histogram, the distribution of primary differential values in the main scanning or sub-scanning direction, and the like. is there.

  The control unit 11 calculates (1) a learning amount SQ1 based on a regression function and (2) a learning amount SQ2 based on a standard function (step S3).

(1) “Learning amount SQ1 by regression function” is a first adjustment amount calculated for the latest input data based on input data (past data) that has been input in the past and is currently stored. Estimated value.
The estimated amount of adjustment is an estimated value of density correction value or contrast value to be adjusted when input data is displayed on the display unit 13.

  (2) “Learning amount SQ2 by standard function” is a second estimated value of the adjustment amount calculated for the latest input data based on a function that outputs a standard learning amount.

With reference to FIG. 3, a learning amount calculation process using a regression function will be described.
The control unit 11 extracts feature values of past data and feature values of input data from the storage unit 15 (step S31).

The feature amount of the past data is calculated and stored in the storage unit 15 when input in the past.
The adjustment amount of the past data is a density correction value or a contrast value adjusted through the operation unit 12 when the image processing condition is determined, and is stored in the storage unit 15 in association with the feature amount.

  The feature amount of the input data is calculated at the time of input (see step S2) and stored in the storage unit 15.

  The control unit 11 calculates the degree (distance d) that the feature amount of the past data approximates the feature amount of the input data (step S32).

  It can be said that the past data with a smaller distance d is more similar to the input data. Specifically, when the distance d between the feature amounts of the two radiographic images is small, it can be said that the two radiographic images are images of the same part (for example, the abdomen).

  The control unit 11 extracts k past data set in advance in ascending order of the calculated distance d (step S33).

  The control unit 11 weights the adjustment amount of the k past data extracted by 1 / k, and weights the adjustment amount of the other past data by 0 (step S34).

  The control unit 11 calculates the average of past data adjustment amounts weighted by 1 / k or 0 as the learning amount SQ1 (step S35), and ends the learning amount calculation process.

  In the learning amount calculation process using the standard function, the control unit 11 calculates the average of the adjustment amounts of all past data as the learning amount SQ2.

FIG. 4 shows a conceptual diagram of the learning amount calculation process.
X1 is the feature quantity of the input data.
A1 to A5 are adjustment amounts of k pieces of past data (here, k = 5) extracted in order of the distance d from X1.

  The learning amount SQ1 is an adjustment amount obtained by weighting A1 to A5 by multiplying the inverse of k and weighting the other adjustment amounts by 0.

That is, the learning amount SQ1 when k = 5 is calculated by SQ1 = (A1 + A2 + A3 + A4 + A5) × 1/5 + (adjustment amount of remaining past data) × 0.
In addition, SQ2 = (A1 + A2 + A3 + A4 + A5) × 1/5 + (adjustment amount of remaining standard data) × 0.

  Returning to FIG. 2, the control unit 11 calculates the reliability of the learning amount (hereinafter, simply “reliability”) based on the regression function calculated in step S3 (step S4).

The reliability calculation process will be described with reference to FIG.
The reliability is a value indicating the reliability of the learning amount calculated by the regression function.
The reliability is calculated by the following formula 1.
<Formula 1>
(Reliability A) = (Data amount reliability coefficient a1) × (Feature amount reliability coefficient a2)

  Although details will be described below, the data amount reliability coefficient a1 is a coefficient that increases as past data increases. Here, since data = radiation image, it will be described as “image number reliability coefficient a1”.

  The feature quantity reliability coefficient a2 is a coefficient that increases as the average of the degree of approximation between the feature quantity of the past data and the feature quantity of the input data decreases.

  Hereinafter, the calculation process of the reliability will be described together with the calculation process of the image number reliability coefficient a1 and the feature quantity reliability coefficient a2.

The control unit 11 extracts the number of past data (number of images) n from the storage unit 15 (step S41).
The control unit 11 determines whether or not the image number n is larger than a predetermined image number threshold value N (for example, N = 100) (step S42).

  If n> N is not satisfied (step S42; N), the control unit 11 sets the image number reliability coefficient a1 = n / N (step S43).

  When n> N (step S42; Y), the control unit 11 sets the image number reliability coefficient a1 = 1 (step S44).

FIG. 6A shows the relationship between the image number reliability coefficient a1 and the image number n.
The vertical axis represents the image number reliability coefficient a1, and the horizontal axis represents the number n of past data images.
The image number reliability coefficient a1 increases in proportion to the image number n (a1 = n / N), and becomes a constant value after the stored image number n exceeds a predetermined threshold N (a1 = 1). ).

Returning to FIG. 5, the control unit 11 calculates an average value (average distance dA) of the degree of approximation between the feature amount of the past data and the feature amount of the input data (see step S2) (step S45).
The average distance dA is an average value of k distances d of past data similar to the input data.

  The control unit 11 determines a threshold value (distance threshold value D) for the degree of approximation (step S46).

The distance threshold value determination process will be described with reference to FIG.
The control unit 11 selects any one of the past data (step S461).
After calculating the distance to all past data other than the selected one data, the control unit 11 calculates the feature quantity of the selected one data and the feature quantity of k other data whose distance is close to the feature quantity. The average distance dAk1 is calculated (step S462).

  The control unit 11 selects all n pieces of past data, and determines whether or not the average distance dA is calculated for each selected data (step S463).

  When all past data has not been selected (step S463; N), the control unit 11 selects other data (step S464), and proceeds to step S462.

  When all the past data is selected (step S463; Y), the control unit 11 rearranges the data of the calculated average distances dAk1 to dAkn for the past data (n) in ascending order of values (step S465).

The control unit 11 determines a count number for determining the distance threshold D from a predetermined cumulative percentage point and the number n of past data, and counts the number of data in ascending order (step S466).
The control unit 11 determines the average distance of the data that has reached the determined count as the distance threshold D (step S467), and ends the distance threshold determination process.

FIG. 8 shows a conceptual diagram of the distance threshold value determination process.
The vertical axis represents the count number, and the horizontal axis represents the average distance dA. Further, the area U indicates cumulative%. The cumulative percentage of the total area U is 100%.
An average distance of data of a predetermined cumulative percentage point (for example, 97%) is determined as the distance threshold value D.

  Returning to FIG. 5, the control unit 11 determines whether or not the average distance dA between the feature amount of the k pieces of past data and the feature amount of the input data is larger than the distance threshold D (step S47).

When dA> D is not satisfied (step S47; N), the control unit 11 sets the feature amount reliability coefficient a2 = 1 (step S48).
When dA> D is satisfied (step S47; Y), the control unit 11 sets the feature amount reliability coefficient a2 = 1−1 / D × (dA−D) (step S49).

FIG. 6B shows the relationship between the feature amount reliability coefficient a2, the average distance dA, and the distance threshold value D.
The vertical axis represents the feature amount reliability coefficient a2, and the horizontal axis represents the average distance dA.
The feature amount reliability coefficient a2 is a constant value (a2 = 1) when the average distance dA is within the range of the distance threshold D, and becomes smaller in proportion to the average distance dA after the average distance dA exceeds the distance threshold D. (A2 = 1−1 / D × (dA−D)).

  The control unit 11 calculates the reliability A by the reliability A = a1 × a2 (step S50), and ends the reliability calculation process.

Returning to FIG. 2, the control unit 11 calculates a third estimated value (learning amount) of the adjustment amount based on the reliability A (step S <b> 5).
The learning amount based on the reliability A is calculated by the following equation 2.
<Formula 2>
(Learning amount SQA based on reliability A) = (Learning amount SQ1 based on regression function) × (Reliability A) + (Learning amount SQ2 based on standard function) × (1−Reliability A)

  The control unit 11 adjusts the density or contrast of the input data (= radiation image) based on the calculated learning amount SQA, displays the adjusted radiation image on the display unit 13 (step S6), and performs data processing. finish.

With reference to FIG. 9, another data process will be described.
Here, data processing in the classification problem will be described.

  Steps S11 and S12 are the same as the processing described with reference to FIG.

  The control unit 11 calculates (1) a learning result SR1 by the learning classifier and (2) a learning result SR2 by the standard function (step S13).

(1) “Learning result SR1 by learning classifier” is a classification result calculated for the latest input data based on input data (past data) that has been input in the past and currently stored. 1 Estimated value.
The estimated value of the classification result is an estimated value of a part where the input data should be classified.

  (2) “Learning result SR2 by standard function” is a second estimated value of the classification result calculated for the latest input data based on a function that outputs a standard classification result.

The learning result calculation process by the learning classifier will be described with reference to FIG.
The control unit 11 extracts feature values of past data and feature values of input data from the storage unit 15 (step S131).

The feature amount of the past data is calculated and stored in the storage unit 15 when input in the past.
The past data classification result is part information input via the operation unit 12 when the image processing condition is determined, and is stored in the storage unit 15 in association with the feature amount.

  The feature amount of the input data is calculated at the time of input and stored in the storage unit 15.

  The control unit 11 calculates the degree (distance d) that the feature amount of the past data approximates the feature amount of the input data (step S132).

  It can be said that the past data with a smaller distance d is more similar to the input data. Specifically, when the distance d between the feature amounts of the two radiographic images is small, it can be said that the two radiographic images are images of the same part (for example, the abdomen).

  The control unit 11 extracts k past data set in advance from the smallest in the calculated distance d (step S133).

  The control unit 11 calculates the classification result having the largest number among the extracted classification results of the k past data as a learning result (step S134).

  When “past data” is set as “standard data” in the above processing, a learning result SR2 by a standard function is calculated. Since the other processes are the same as the learning result calculation process (FIG. 12) by the learning classifier, the description of the learning result calculation process by the standard function is omitted.

  Returning to FIG. 9, the control unit 11 calculates the reliability (reliability) of the learning result by the learning classifier calculated in step S13 (step S14).

Since the process of step S14 is the same as the process of step S4, description here is abbreviate | omitted.
The control unit 11 performs a classification process based on the reliability A (step S15).

With reference to FIG. 11, the classification process by the learning classifier will be described.
The control unit 11 determines whether or not the reliability A is 1/3 or more (step S151).

  When A ≧ 1/3 is not satisfied (step S151; N), the control unit 11 selects the learning result SR2 calculated by the standard function (step S152), and ends the classification process by the learning classifier.

  When A ≧ 1/3 (step S151; Y), the control unit 11 determines whether or not the reliability A is 2/3 or more (step S153).

  When A ≧ 2/3 (step S153; Y), the control unit 11 selects the learning result SR1 calculated by the learning classifier (step S154), and ends the classification process by the learning classifier.

  When A ≧ 2/3 is not satisfied (step S153; N), the control unit 11 selects a learning result calculated by either the standard function or the learning classifier based on the output value determination table stored in advance. (Step S155), the classification process by the learning classifier is terminated.

FIG. 12 shows an output value determination table T.
C1 to C3 shown in the output value determination table T are learning results calculated by a learning classifier or a standard function.

  Here, a learning result calculated as a special image by a learning classifier or a standard function is C1, a learning result calculated as a general image is C2, and a learning result calculated as a more general image is C3. To do.

The “special image” for which C1 is calculated is, for example, mammography.
The “general image” for which C2 is calculated is, for example, an image of a limb bone.
The “more general image” for which C3 is calculated is, for example, an image of the chest.

  According to the output value determination table T, when the learning result by the learning classifier or the standard function is the same, the same learning result is selected.

  For example, when the learning result calculated by the learning classifier is C1 and the learning result calculated by the standard function is C1, C1 is selected.

  When learning results calculated by the learning classifier or the standard function are different, they are selected in the order of C3, C2, and C1. That is, the learning result C3 calculated for a more general image is preferentially selected, and is selected in the order of C2 and C1.

For example, when the learning result by the learning classifier is C1, and the learning result calculated by the standard learner is C2, C2 is selected.
Further, when the learning result calculated by the learning classifier is C3 and the learning result calculated by the standard function is C1, C3 is selected.

  Returning to FIG. 9, the control unit 11 adjusts the density or contrast of the input data (= radiation image) based on the selected learning results (C1 to C3), and displays the adjusted radiation image on the display unit 13. (Step S16), and the other data processing ends.

  As described above, according to the present embodiment, in the regression problem, the learning amount SQ1 based on the regression function can be calculated, and the reliability A of the calculated learning amount SQ1 can be calculated.

  The reliability A can be calculated by the product of the image number reliability coefficient a1 and the feature amount reliability coefficient a2.

  The image number reliability coefficient a1 is set to a1 = n / N when the image number n is smaller than the image number threshold value N, and is set to a1 = 1 when the image number n is larger than the image number threshold value N. be able to.

  The feature amount reliability coefficient a2 is set to a2 = 1−1 / D (dA−D) when the average distance dA is larger than the distance threshold D, and a2 when the average distance dA is smaller than the distance threshold D. = 1 can be set.

  The distance threshold D calculates the average distance dA with other data for all the past data, sorts the past data in ascending order of the average distance dA, and counts the number of data of the rearranged past data. Thus, the average distance dA at the time when the ratio becomes a predetermined ratio can be determined as the distance threshold value D.

  Further, the learning amount SQ1 optimized based on the learning amount SQ1 based on the regression function, the learning amount SQ2 based on the standard function, and the reliability A can be calculated.

  Further, the optimized learning amount SQA can be calculated by SQA = (learning amount by regression function) × A + (learning amount by standard function) × (1−reliability A).

  Further, in the classification problem, either learning result (C1 to C3) calculated by the learning classifier or learning result (C1 to C3) calculated by the standard function is selected based on the reliability A. Thus, one learning result can be calculated.

  When the input data is a radiographic image, the density correction value or contrast value of the radiographic image can be calculated as a learning amount or a learning result (SQ1, SQ2, C1 to C3).

  Further, when the data processing apparatus 10 is configured to include the display unit 13, the display unit 13 displays a radiation image in which the density or contrast is adjusted based on the learning amount or the learning result (SQ1, SQ2, C1 to C3) can do.

DESCRIPTION OF SYMBOLS 10 Data processor 11 Control part 12 Operation part 13 Display part 14 Input part 15 Storage part

Claims (19)

An input section for inputting data;
A storage unit for storing the input data;
A control unit,
The control unit calculates a first estimated value for the latest input data based on past data and the latest input data that have been input and are currently stored, and the calculated first estimated value A data processing apparatus that calculates the reliability of the data based on the data amount of the past data or the feature amount of the past data and the feature amount of the input data.   The control unit is a product of a data amount reliability coefficient calculated based on the data amount of the past data and a feature amount reliability coefficient calculated based on the feature amount of the past data and the feature amount of the input data, The data processing apparatus according to claim 1, wherein the reliability is calculated.   When the data amount of the past data is smaller than a predetermined data amount threshold, the control unit sets a value obtained by dividing the data amount of the past data by the data amount threshold as the data amount reliability coefficient, and the past data The data processing apparatus according to claim 2, wherein when the data amount is greater than the data amount threshold, 1 is used as the data amount reliability coefficient.   The control unit extracts k past data from the past data, calculates an average distance between the extracted feature values of the k past data and the feature amount of the input data, and calculates the calculated average When the distance is larger than a predetermined distance threshold value, a value obtained by multiplying the difference between the average distance and the distance threshold value by the reciprocal of the distance threshold value and subtracting 1 is used as the feature amount reliability coefficient, The data processing apparatus according to claim 2 or 3, wherein 1 is used as the feature amount reliability coefficient when the distance is smaller than a distance threshold.   The control unit calculates an average distance between a feature amount of one past data and a feature amount of the k pieces of past data for all data of the past data, and associates the correspondence with the one past data. The past data is rearranged in ascending order of the attached average distance, the number of data of the past data rearranged is counted, and the ratio becomes a predetermined ratio with respect to the total number of data of the past data. The data processing apparatus according to claim 4, wherein an average distance of past data at the time point is set as the distance threshold.   The control unit, based on the first estimated value calculated based on the past data and the input data, a standard function and a second estimated value calculated based on the input data, and the calculated reliability, The data processing apparatus according to any one of claims 1 to 5, wherein a third estimated value is calculated. The estimated value is an estimated value of the adjustment amount,
The control unit calculates the third estimated value by adding a value obtained by multiplying the first estimated value by the reliability and a value obtained by multiplying the second estimated value by the reliability obtained by subtracting 1. Item 7. The data processing device according to item 6. The estimated value is an estimated value of the classification result,
The control unit selects any one of the first estimated value or a second estimated value calculated based on a standard function and the input data based on the reliability. A data processing apparatus according to claim 1. The storage unit stores a table for determining one of the first estimated value and the second estimated value based on the reliability,
The data processing apparatus according to claim 8, wherein the control unit selects one of the first estimated value and the second estimated value based on the stored table. The estimated value is an estimated value of the adjustment amount,
The input data is a radiographic image;
The data processing apparatus according to claim 6 or 7, wherein the first estimated value, the second estimated value, and the third estimated value are density correction values or contrast values of the radiation image. The estimated value is an estimated value of the adjustment amount,
A display unit for displaying the input data;
The control unit may display the input data on the display unit based on any one of the first estimated value, the second estimated value, and the third estimated value. The data processing apparatus described. A process of inputting data by an input unit;
Storing the input data by a storage unit;
A step of controlling by the control unit,
The controlling step includes a step of calculating a first estimated value of the latest input data based on past data and the latest input data that have been input and are currently stored, and the calculated first estimate Calculating the reliability of the value based on the data amount of the past data or the feature amount of the past data and the feature amount of the input data.   The controlling step is based on a first estimated value calculated based on the past data and the input data, a standard function, a second estimated value calculated based on the input data, and the calculated reliability. The data processing method according to claim 12, further comprising a step of calculating a third estimated value.   13. The step of controlling includes a step of selecting any one of the first estimated value or a second estimated value calculated based on a standard function and the input data based on the reliability. The data processing method described in 1. A step of displaying the input data by a display unit;
The step of controlling includes a step of displaying the input data on the display unit based on any one of the first estimated value, the second estimated value, and the third estimated value. 14. The data processing method according to 13. Computer
Input means for inputting data,
Storage means for storing the input data;
A program that functions as a control means,
The control means calculates the first estimated value of the latest input data based on the past data and the latest input data that have been input and are currently stored, and the reliability of the calculated first estimated value. A program for calculating a degree based on a data amount of the past data or a feature amount of the past data and a feature amount of the input data.   The control means, based on the first estimated value calculated based on the past data and the input data, a standard function and a second estimated value calculated based on the input data, and the calculated reliability, The program according to claim 16, wherein the third estimated value is calculated.   The program according to claim 16, wherein the control unit selects one of the first estimated value or a second estimated value calculated based on a standard function and the input data based on the reliability. . A program for functioning as a display means for displaying the input data,
The control unit according to claim 17, wherein the control unit displays the input data on the display unit based on any one of the first estimated value, the second estimated value, and the third estimated value. program.

Images