Judgment evaluation device, method, and program

Abstract

To provide a determination evaluation device, method and program for evaluating reliability of determination results obtained by a determination model.SOLUTION: A determination evaluation device 1 for evaluating a determination result for input data based on a determination model is provided, the device comprising an uncertainty computation unit 111 for computing uncertainty of the determination result, an analysis / evaluation unit 109 for evaluating the determination result on the basis of the uncertainty, and an evaluation result display unit 112 for displaying a screen based on the evaluation of the determination result.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a technique for evaluating the judgment ability of an artificial intelligence to analyze an image and classify each pixel into a class. [Background technology]

[0002] In various fields, artificial intelligence is used to divide captured images into multiple regions. For example, in fields such as medicine and materials science, artificial intelligence is useful for coloring images of a mixture of multiple materials into different colors for each material. To enable judgment using artificial intelligence, it is necessary to build a judgment model by learning from training data consisting of pre-prepared images and correct labels that color each region of the image.

[0003] However, preparing highly accurate training data requires accurate manual labeling of images, which is a time-consuming process. This often results in a shortage of training data. If the training data is insufficient, there will be areas where the judgment model has not sufficiently learned the training data. In such cases, the accuracy of the judgment results for images belonging to such areas will decrease. In relation to this problem, Patent Document 1 discloses a technology for exploring measures to improve the value of a predetermined index while keeping the uncertainty, which represents the reliability of the estimation result, within an acceptable range. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2020-30500 A Summary of the Invention [Problem to be solved by the invention]

[0005] If the results of the judgment process by artificial intelligence are reliable, it will be easy to create training data based on those judgment results, reducing the manual work required to assign correct labels to input data.

[0006] However, the technology of Patent Document 1 is a technology for searching for measures as described above, and is not suitable for evaluating the reliability of the results of determination processing by artificial intelligence.

[0007] One objective of the present disclosure is to provide a technique for evaluating the reliability of the results of a determination made by a determination model. [Means for solving the problem]

[0008] A judgment evaluation device according to one aspect of the present disclosure is a judgment evaluation device that evaluates a judgment result for input data based on a judgment model, and includes an uncertainty calculation unit that calculates the uncertainty of the judgment result, and an analysis and evaluation unit that evaluates the judgment result based on the uncertainty.

[0009] An evaluation and judgment method according to one aspect of the present disclosure is a judgment and evaluation method for evaluating a judgment result for input data based on a judgment model, which calculates the uncertainty of the judgment result and evaluates the judgment result based on the uncertainty.

[0010] An evaluation and judgment program according to one aspect of the present disclosure is a judgment and evaluation program that evaluates judgment results for input data based on a judgment model, and causes a computer to calculate the uncertainty of the judgment result and evaluate the judgment result based on the uncertainty. [Effects of the Invention]

[0011] According to one aspect of the present disclosure, it is possible to evaluate the reliability of the results of a judgment made by a judgment model. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a block diagram showing an example of the configuration of a judgment evaluation device according to an embodiment of the present invention; [Figure 2] 1 shows an example of learning data according to the present embodiment. [Figure 3] 1 illustrates an example of a machine learning model according to the present embodiment. [Figure 4] 10 shows an example of test data according to the present embodiment. [Figure 5] 10 shows an example of determination result data according to the present embodiment. [Figure 6] 10 shows an example of uncertainty data according to the present embodiment. [Figure 7] 10 shows an example of an evaluation rule according to the present embodiment. [Figure 8] An example of an evaluation result according to the present embodiment will be shown. [Figure 9] FIG. 2 is a block diagram showing details of an analysis and evaluation unit according to the present embodiment. [Figure 10] 10 is a flowchart illustrating an example of processing by an uncertainty calculation unit according to the present embodiment. [Figure 11] 10 is a flowchart showing an example of processing by an analysis and evaluation unit according to the present embodiment. [Figure 12] 12 is a flowchart showing an example of processing when the condition of rule ID "1001" in FIG. 7 is applied in step S206 shown in FIG. [Figure 13] FIG. 2 is a diagram showing an example of an evaluation result display screen 1 according to the present embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0014] (Embodiment 1) FIG. 1 is a block diagram showing an example of the configuration of a judgment evaluation device 1 according to this embodiment.

[0015] The judgment evaluation device 1 is a device that evaluates the uncertainty of the boundary of a segment (i.e., the boundary between different classes) when segmentation (i.e., classification of each pixel into classes) is performed on an input image using a machine learning model 103. Uncertainty is an index that indicates the low reliability of the boundary of a segment calculated by the machine learning model 103. In other words, a boundary of a segment with high uncertainty has low reliability, and a boundary of a segment with low uncertainty has high reliability.

[0016] The determination evaluation device 1 includes a machine learning unit 101, a determination unit 110, an uncertainty calculation unit 111, an analysis and evaluation unit 109, and an evaluation result display unit 112. The determination evaluation device 1 also stores learning data 102, a machine learning model 103, test data 104, determination result data 105, uncertainty data 106, evaluation rules 107, and evaluation results 108.

[0017] The training data 102 is data used for training the machine learning model 103. Fig. 2 shows an example of the training data 102 according to this embodiment. As shown in Fig. 2, the training data 102 is data in which a training data ID 201 and a training data content 202 are recorded in association with each other.

[0018] The learning data ID 201 is information for identifying the learning data content 202. The learning data content 202 has the pixel value of each pixel that constitutes an input image for learning, and the correct class (correct label) into which each pixel that constitutes the input image is classified.

[0019] The machine learning unit 101 uses learning data 102 to perform machine learning on a machine learning model 103 that determines a class (assigns a label) for each pixel of an input image. The machine learning may be deep learning. Examples of the machine learning model 103 may include a CNN (Convolutional Neural Network), SegNet, U-Net, and PSPNet. FIG. 3 shows an example of the machine learning model 103 according to this embodiment. The machine learning model 103 is represented by data that associates a layer 301, a layer type 302, and a layer parameter 303.

[0020] The layer 301 is information for identifying a layer of the machine learning model 103.

[0021] The layer type 302 is information indicating the type of the layer 301. Examples of layer types include an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer.

[0022] The layer parameters are information indicating the parameters of the layer 301 (for example, the weight of each neuron).

[0023] The test data 104 is data used to evaluate the machine learning model 103. Fig. 4 shows an example of the test data 104 according to this embodiment. As shown in Fig. 4, the test data 104 has, as items, a test data ID 401 and test data content 402.

[0024] The test data ID 401 is information for identifying the test data content 402 .

[0025] As test data content 402, the pixel values ​​of each pixel of the input image for evaluation and judgment are recorded.

[0026] The judgment unit 110 evaluates and judges the machine learning model 103 using the test data 104, and generates judgment result data 105. Fig. 5 shows an example of the judgment result data 105 according to this embodiment. As shown in Fig. 5, the judgment result data 105 is data in which a judgment result ID 501, a test data ID 502, and a judgment result content 503 are recorded in association with each other for each entry.

[0027] The judgment result ID 501 is information for identifying the judgment result in the entry.

[0028] The test data ID 502 is information for identifying the test data used for the evaluation of the entry, and corresponds to the test data ID 401 shown in FIG.

[0029] The determination result content 503 indicates the content of the determination result for the entry. The content of the determination result indicates the probability of classification of each pixel constituting the test data content 402 (test image) of the test data ID 502 into each class by the machine learning model 103. In other words, the determination result content 503 is data indicating the probability that each pixel constituting the test image is classified into each class. For example, the determination result for a pixel near the boundary between the first segment and the second segment includes the probability that the pixel belongs to the first segment (first class) and the probability that the pixel belongs to the second segment (second class), and it is considered that the sum of the probability that the pixel belongs to the first segment and the probability that the pixel belongs to the second segment is 1.

[0030] For example, the uncertainty calculation unit 111 calculates the uncertainty of the machine learning model 103 using the test data 104, and generates the uncertainty data 106. Fig. 6 shows an example of the uncertainty data 106 according to this embodiment. As shown in Fig. 6, the uncertainty data 106 is data in which an uncertainty ID 601, a determination result ID 602, and test data 603 are recorded in association with each other for each entry.

[0031] The uncertainty ID 601 is information for identifying the entry for which uncertainty calculation was performed using test data.

[0032] The judgment result ID 602 ​​is information for identifying the judgment result corresponding to the entry, and corresponds to the judgment result ID 501 shown in FIG.

[0033] The test data ID 603 is information for identifying the test data used to calculate the uncertainty in the entry, and corresponds to the test data ID 401 shown in FIG. 4 and the test data ID 502 shown in FIG.

[0034] The uncertainty content 604 indicates the uncertainty value calculated for the entry. That is, the uncertainty content 604 records the uncertainty value of each pixel calculated using the test data 104 of the test data ID 603.

[0035] The uncertainty can be calculated, for example, by the following process. First, multiple machine learning models (hereinafter referred to as Monte Carlo dropout models) are generated in which some of the multiple nodes constituting the machine learning model 103 are intentionally dropped out using the Monte Carlo method. Next, an input image is input to each Monte Carlo dropout model, and each judgment result (class classification of each pixel) is obtained. Then, the variance of the multiple judgment results thus obtained for each pixel is calculated and used as the uncertainty value for that pixel. In other words, the uncertainty can be said to be the variance of the class of each pixel obtained from different Monte Carlo dropout models. In other words, the uncertainty can be said to be a value indicating how much the segmentation boundaries obtained by each Monte Carlo dropout model vary depending on each dropout model (i.e., the reliability of the segmentation boundaries).

[0036] That is, the uncertainty calculation unit 111 may use the variance value of the class of each pixel calculated using Monte Carlo dropout as the uncertainty, which allows the uncertainty calculation unit 111 to calculate the uncertainty using test data without a correct answer label.

[0037] The evaluation rules 107 are rules used to evaluate uncertainty. Fig. 7 shows an example of the evaluation rules 107 according to this embodiment. As shown in Fig. 7, the evaluation rules 107 record, for each rule, a rule ID 701, a condition 702, and a display template 703 in association with each other.

[0038] The rule ID 701 is information for identifying the rule, and identifies a set of the condition 702 and the display template 703 of the rule.

[0039] Condition 702 is information indicating the condition (evaluation method) used to evaluate the uncertainty. For example, if the condition is satisfied, the uncertainty is evaluated as high.

[0040] The display template 703 is a template that defines the content to be displayed when the condition 702 is satisfied. For example, the display content is defined so that a message that there is high uncertainty and a recommended way to deal with the situation are displayed.

[0041] That is, when condition 702 is satisfied, the content of the uncertainty is displayed based on display template 703 associated with condition 702. By looking at the content of the uncertainty displayed based on display template 703, the user can learn what kind of uncertainty may occur and how to deal with it to further reduce the uncertainty.

[0042] The analysis and evaluation unit 109 performs analysis and evaluation using the evaluation rules 107, test data 104, judgment result data 105, and uncertainty data 106, and generates an evaluation result 108. Details of the analysis and evaluation process will be described later. Fig. 8 shows an example of the evaluation result 108 according to this embodiment. The evaluation result 108 is, for example, data in which an evaluation result ID 801, a rule ID 802, an uncertainty ID 803, a judgment result ID 804, a test data ID 805, and an evaluation content 806 are recorded in association with each other for each entry of the analysis and evaluation.

[0043] The evaluation result ID 801 is information for identifying the analysis evaluation of the entry.

[0044] The rule ID 802 is information for identifying the rule used in the analysis and evaluation of the entry, and corresponds to the rule ID 701 shown in FIG.

[0045] The uncertainty ID 803 is information for identifying the uncertainty data used in the analysis and evaluation of the entry, and corresponds to the uncertainty ID 601 shown in FIG.

[0046] The judgment result ID 804 is information for identifying the judgment result used in the analysis and evaluation of the entry, and corresponds to the judgment result ID 501 shown in FIG.

[0047] The test data ID 805 is information for identifying the test data used in the analysis and evaluation of the entry, and corresponds to the test data ID 502 shown in FIG.

[0048] The evaluation content 806 indicates the content of the evaluation obtained by applying the evaluation rule indicated by the rule ID 802 to the test data indicated by the test data ID 805, the judgment result indicated by the judgment result ID 804, and the uncertainty indicated by the uncertainty ID 803. For example, if the condition 702 of the evaluation rule for an entry with a certain rule ID 701 shown in FIG. 7 is met, the display content defined in the display template 703 of that entry is recorded in the evaluation content 806.

[0049] The analysis and evaluation unit 109 will be described in detail later (see FIG. 9).

[0050] The evaluation result display unit 112 generates an evaluation result display screen 1000 (see FIG. 13) based on the test data 104, the judgment result data 105, the uncertainty data 106, and the evaluation result 108, and displays it on a monitor or the like.

[0051] As described above, the determination evaluation device 1 is a device that evaluates a determination result for input data based on a determination model, and includes an uncertainty calculation unit 111 that calculates the uncertainty of the determination result, and an analysis and evaluation unit 109 that evaluates the determination result based on the uncertainty. This allows the determination evaluation device 1 to evaluate how reliable the result of determination by the determination model is.

[0052] In this embodiment, the input data is an input image, and the above-mentioned judgment is a process of classifying the input image into regions of multiple classes. The analysis and evaluation unit 109 may calculate the classification uncertainty for each pixel of the input image and evaluate the judgment result based on the distribution of classes on the input image and the distribution of uncertainty on the input image. This makes it possible to evaluate the reliability of the judgment result by the judgment model based on the distribution of uncertainty at each position in the image in the judgment result of dividing the image into multiple regions.

[0053] FIG. 9 is a block diagram showing details of the analysis and evaluation unit 109 according to this embodiment.

[0054] The analysis and evaluation unit 109 includes an area division unit 901, a color histogram generation unit 902, a histogram matching calculation unit 903, a boundary discrimination unit 904, a boundary uncertainty width measurement unit 905, a mode extraction unit 906, a high uncertainty pixel extraction unit 907, a mask processing unit 908, an average calculation unit 909, a low frequency component extraction unit 910, and a control unit 911.

[0055] The area dividing unit 901 divides the image of the test data 104 into a plurality of areas, for example, in a grid pattern.

[0056] The color histogram generating unit 902 generates a color histogram for each area divided by the area division.

[0057] A histogram matching calculation unit 903 calculates the matching of color histograms between areas generated in the color histogram generation process.

[0058] The boundary determining unit 904 determines the boundary between classes in the image based on the determination result data 105 .

[0059] The boundary uncertainty measurement unit 905 measures the uncertainty range at the boundary part of the class determined by the boundary part determination unit 904 based on the uncertainty data 106. The uncertainty range at the boundary part indicates the degree of uncertainty at the boundary part, that is, the level of uncertainty at the boundary. A method for calculating the uncertainty range will be described later.

[0060] The mode extractor 906 extracts the mode from among the ranges of uncertainty measured by the boundary uncertainty measurer 905 .

[0061] The high uncertainty pixel extractor 907 extracts pixels with high uncertainty (for example, pixels with uncertainty values ​​equal to or greater than a predetermined threshold) based on the uncertainty data 106.

[0062] The mask processing unit 908 masks the data of the pixels extracted in the high uncertainty pixel extraction 907 in the determination result data 105 .

[0063] The average calculation unit 909 calculates the average value of the uncertainty values ​​in the uncertainty data 106 .

[0064] The low frequency component extractor 910 extracts low frequency components (for example, frequency components below a predetermined threshold) of the uncertainty values ​​in the uncertainty data 106.

[0065] The control unit 911 evaluates the results of the histogram coincidence calculation 903, the result of the mode extraction 906, the result of the mask processing 908, the result of the average calculation 909, and the result of the low-frequency component extraction 910 by applying each rule defined in the evaluation rules 107, and outputs the evaluation result 108.

[0066] FIG. 10 is a flowchart showing an example of the processing performed by the uncertainty calculation unit 111 according to this embodiment.

[0067] The uncertainty calculation unit 111 reads the machine learning model 103 (S101).

[0068] The uncertainty calculation unit 111 repeats steps S102 to S107 for the number of entries included in the test data 104 (S102).

[0069] The uncertainty calculation unit 111 reads the test data of the next entry from the test data 104 (S103).

[0070] The uncertainty calculation unit 111 repeats step S105 a predetermined number of times (number of MCD determinations) for the number of Monte Carlo dropout models (S104).

[0071] The uncertainty calculation unit 111 selects one of the multiple Monte Carlo dropout models generated from the machine learning model read in step S101, and classifies each pixel of the test data (i.e., the input image) read in step S103 into a class using the selected Monte Carlo dropout model (S105).

[0072] The uncertainty calculation unit 111 repeats step S105 while switching the Monte Carlo dropout model a number of times equal to the number of Monte Carlo dropout models (the number of MCD determinations), and then proceeds to step S107 (S106).

[0073] The uncertainty calculation unit 111 calculates the variance of the determination results obtained from the different Monte Carlo dropout models in step S105 for each pixel, and stores the variance in the uncertainty data 106 (S107).

[0074] The uncertainty calculation unit 111 repeats steps S103 to S107 as many times as the number of entries included in the test data 104 (that is, the number of test data) (S108), and then ends this process.

[0075] Through the above processing, uncertainty data is obtained.

[0076] FIG. 11 is a flowchart showing an example of the processing of the analysis and evaluation unit 109 according to this embodiment.

[0077] The analysis and evaluation unit 109 repeats steps S202 to S209 for the number of entries included in the uncertainty data 106 (S201).

[0078] The analysis and evaluation unit 109 reads the next entry of the uncertainty data 106 (S202).

[0079] The analysis and evaluation unit 109 reads out the test data 104 and the judgment result data 105 corresponding to the entry of the uncertainty data 106 read out in step S201 (S203).

[0080] The analysis and evaluation unit 109 repeats steps S205 to S208 for the number of entries included in the evaluation rule 107 (S204).

[0081] The analysis and evaluation unit 109 reads the next entry of the evaluation rules 107 (S205).

[0082] The analysis and evaluation unit 109 applies the conditions of the evaluation rule 107 of the entry read in step S205 to the uncertainty data 106 read in step S202 and the test data 104 and judgment result data 105 read in step S203 (S206).

[0083] The analysis and evaluation unit 109 determines whether the uncertainty data 106, the test data 104, and the judgment result data 105 satisfy the condition 702 of the evaluation rule 107 (S207).

[0084] If the condition 702 is not satisfied (S207: No), the analysis and evaluation unit 109 advances the process to step S209.

[0085] If the condition 702 is satisfied (S207: YES), the analysis and evaluation unit 109 substitutes variables into the display content defined in the display template 703 of the evaluation rule 107, stores the result in the evaluation result 108 (S208), and proceeds to step S209.

[0086] The analysis and evaluation unit 109 repeats steps S205 to S208 for the number of entries included in the evaluation rule 107 (S209), and then proceeds to step S210.

[0087] The analysis and evaluation unit 109 repeats steps S205 to S208 for the number of entries in the uncertainty data 106 (S210), and then ends this process.

[0088] Through the above processing, an evaluation result 108 is obtained.

[0089] FIG. 12 is a flowchart showing, as an example of the process of step S206 shown in FIG. 11, the process when the condition 702 of the rule ID "1001" in FIG. 7 is applied.

[0090] The analysis and evaluation unit 109 repeats steps S302 to S305 for each column of the uncertainty content 604 of the entry of the uncertainty data to be processed (S301).

[0091] The analysis and evaluation unit 109 repeats steps S303 to S304 (S302) the number of times corresponding to the number of local peaks of uncertainty present in the column selected in step S301. A local peak of uncertainty is a point that indicates a value greater than the uncertainty values ​​on either side.

[0092] For the local peak of uncertainty selected in step S302, the analysis and evaluation unit 109 determines whether the class of the pixel on the left side of the local peak that has a value equal to one-fourth of the local peak value (hereinafter referred to as the left pixel) is different from the class of the pixel on the right side of the local peak that has a value equal to one-fourth of the local peak value (hereinafter referred to as the right pixel) (S303). If the class of the pixel on the left side of the local peak is different from the class of the pixel on the right side, it can be said that the vicinity of the local peak is the boundary of a segment.

[0093] If the class of the pixel on the left side of the local peak and the class of the pixel on the right side of the local peak are not different (S303: NO), the analysis and evaluation unit 109 proceeds to step S305.

[0094] If the class of the pixel on the left side of the local peak is different from the class of the pixel on the right side (S303: YES), the analysis / evaluation unit 109 measures the half-width of this local peak (S304) and proceeds to step S305.

[0095] The analysis and evaluation unit 109 repeats steps S303 to S304 for the number of local peaks of uncertainty, and then proceeds to step S306 (S305).

[0096] The analysis and evaluation unit 109 repeats steps S302 to S305 as many times as the number of columns of uncertainty data, and then proceeds to step S307 (S306).

[0097] The analysis and evaluation unit 109 creates a histogram of the multiple half-widths measured in step S304 (S307).

[0098] The analysis and evaluation unit 109 acquires the most frequent value of the half width based on the histogram created in step S307 (S308).

[0099] The analysis and evaluation unit 109 returns whether the half-width mode calculated in step S308 is greater than a predetermined threshold value ("2" in the example of the evaluation rule with rule ID "1001") as the condition determination result for this evaluation rule (S309), and then this process ends.

[0100] A large half-width means that the fluctuation of the segment boundary is large (i.e., the boundary uncertainty is high), and a small half-width means that the fluctuation of the segment boundary is small (i.e., the boundary uncertainty is low). The level of reliability required for this boundary varies depending on the user, the application field, etc. If the user requires high reliability for the boundary, the user can set the threshold value in step S309 to a small value, and if the user does not require such high reliability for the boundary, the user can set the threshold value in step S309 to a large value.

[0101] As described above, when applying the condition 702 of the rule ID 701 "1001" shown in Figure 7, the analysis and evaluation unit 109 may perform the following process. That is, the analysis and evaluation unit 109 calculates the half-width of the local peaks at locations where there are regions of different classes in the class distribution on both sides of the local peak of uncertainty in the uncertainty distribution over the entire input image, calculates the mode of the half-width, and if the mode exceeds a predetermined threshold, evaluates the image assessment result as having high boundary uncertainty. This makes it possible to evaluate that there is a high degree of boundary uncertainty, i.e., that there is a risk that the resolution of the input image is low.

[0102] When applying the condition 702 of the rule ID 701 "1002" shown in FIG. 7, the analysis and evaluation unit 109 may perform the following process. That is, if the average value of uncertainty in the input image is smaller than a predetermined threshold, the analysis and evaluation unit 109 evaluates the judgment result of the input image as having low uncertainty. This makes it possible to evaluate that the input image as a whole has low uncertainty. It becomes possible to add good images with low uncertainty to the training data. When adding good images with low uncertainty to the training data, if there are any uncertain parts in the image, these parts may be partially corrected.

[0103] FIG. 13 is a diagram showing an example of an evaluation result display screen 1000 according to this embodiment.

[0104] The evaluation result display screen 1000 may be generated by the evaluation result display unit 112 and displayed on a monitor or the like. The evaluation result display screen 1000 has a machine learning model input area 1301, a test data input area 1302, an evaluation rule input area 1303, a diagnosis execution button 1311, a test data selection area 1321, an input image display area 1322, a judgment result display area 1323, an uncertainty display area 1324, and an evaluation result display area 1325.

[0105] The user inputs the machine learning model 103 to be evaluated in the machine learning model input area 1301. For example, the user may select the file of the machine learning model 103 in the machine learning model input area 1301.

[0106] The user inputs the test data 104 to be used for evaluation in the test data input area 1302. For example, the user may select a file of the test data 104 in the test data input area 1302.

[0107] The user inputs the evaluation rules 107 to be used for evaluation in the evaluation rule input area 1303. For example, the user may select a file of the evaluation rules 107 in the evaluation rule input area 1303.

[0108] When the user presses the diagnosis execution button 1311, the determination evaluation device 1 executes the uncertainty evaluation process described above.

[0109] The user selects an input image from the test data selection area 1321, from among the images included in the test data 104, for which the user wishes to check the evaluation results.

[0110] The evaluation result display unit 112 displays the input image selected in the test data selection area 1321 in the input image display area 1322 .

[0111] The evaluation result display unit 112 displays, in the determination result display area 1323, a determination result image that indicates the determination result for the input image displayed in the input image display area 1322. The determination result image may be an image that indicates the class classification (segmentation result) for the input image.

[0112] The evaluation result display unit 112 displays, in the uncertainty display area 1324, an uncertainty image for the determination result image displayed in the determination result display area 1323. The uncertainty image may be an image that indicates the level of uncertainty of the segment boundaries. For example, the uncertainty image may be expressed by thickening the lines indicating the segment boundaries according to the level of uncertainty. Alternatively, the uncertainty image may be expressed by using a heat map with a higher color according to the level of uncertainty.

[0113] The evaluation result display unit 112 displays the evaluation result 108 in the evaluation result display area 1325. For example, when the most frequent value of the half width shown in step S309 of Fig. 12 is greater than a predetermined threshold, the evaluation result display unit 112 may display information in the evaluation result display area 1325 indicating that the most frequent value of the half width of the boundary uncertainty is greater than a predetermined threshold.

[0114] As described above, the evaluation result display unit 112 may display the input image, an image of the distribution of classes on the input image as the judgment result, and an image of the distribution of uncertainty on the input image, thereby allowing the user to visually recognize the uncertainty of the segment boundaries for the input image.

[0115] For input images evaluated as having low uncertainty, the evaluation result display unit 112 may recommend correcting the judgment results for parts of the input image with high uncertainty, creating correct labels, and adding the correct labels to the training data. This allows the machine learning model 103 to be retrained using the added training data, thereby reducing the uncertainty of the machine learning model 103 for the input image (i.e., improving reliability).

[0116] Each unit of the judgment evaluation device 1 in the above-described embodiment can also be realized by causing a computer including a processor and a memory to execute a software program that defines the processing procedure of each unit.

[0117] Furthermore, the above-described embodiments are merely examples for the purpose of disclosure, and are not intended to limit the scope of the present invention to the above-described embodiments. Those skilled in the art can implement the present invention in various other forms without departing from the scope of the present invention. [Explanation of symbols]

[0118] 1: Judgment result device 101: Machine learning unit 102: Learning data 103: Machine learning model 104: Test data 105: Judgment result data 106: Uncertainty data 107: Evaluation rule 108: Evaluation result 109: Analysis and evaluation unit 110: Judgment unit 111: Uncertainty calculation unit 112: Evaluation result display unit 201: Learning data ID 202: Learning data content 301: Layer 302: Layer type 303: Layer parameter 401: Test data ID 402: Test data content 501: Judgment result ID 502: Test data ID 503: Judgment result content 601: Uncertainty ID 602: Judgment result ID 603: Test data ID 604: Uncertainty content 701: Rule ID 702: Condition 703: Display template 801: Evaluation result ID 802: Rule ID 803: Uncertainty ID 804: Judgment result ID 805: Test data ID 806: Evaluation content 901: Area division 902: Color histogram generation 903: Histogram matching calculation 904: Boundary discrimination 905: Boundary uncertainty width measurement 905: Boundary uncertainty measurement 906: Mode extraction 907: High uncertainty pixel extraction 908: Mask processing 909: Average calculation 910: Low frequency component extraction 911: Control unit 1000: Evaluation result display screen 1301: Machine learning model input area 1302: Test data input area 1303: Evaluation rule input area 1311: Diagnosis execution button 1321: Test data selection area 1322: Input image display area 1323: Judgment result display area 1324: Uncertainty display area 1325: Evaluation result display area

Claims

1. A judgment evaluation device for evaluating a judgment result for input data based on a judgment model, an uncertainty calculation unit that calculates the uncertainty of the determination result; an analysis and evaluation unit that evaluates the determination result based on the uncertainty; and the input data is an image, The determination is a process of classifying the image into regions of a plurality of classes, the analysis and evaluation unit calculates classification uncertainty for each pixel, calculates half-widths of local peaks of uncertainty in the distribution of uncertainty on the image across the entire image at locations where there are regions of different classes in the distribution of classes on the image on both sides of the local peaks of uncertainty in the distribution of uncertainty on the image, calculates the mode of the half-widths, and if the mode exceeds a predetermined threshold, evaluates the judgment result of the image as having high boundary uncertainty. Judgment evaluation device.

2. If the average value of the uncertainty in the image is smaller than a predetermined threshold, the analysis and evaluation unit evaluates the judgment result of the image as having low uncertainty. The judgment evaluation device according to claim 1 .

3. further comprising an evaluation result display unit that displays a screen based on the evaluation of the judgment result; The evaluation result display unit recommends that, for an image evaluated as having low uncertainty, the judgment result of a part of the image having high uncertainty be corrected to create a correct label and add the correct label to the training data. The judgment evaluation device according to claim 2 .

4. further comprising an evaluation result display unit that displays a screen based on the evaluation of the judgment result; the evaluation result display unit displays an image that is the input data, an image of a distribution of classes on the image that is the judgment result, and an image of a distribution of uncertainty on the image. The judgment evaluation device according to claim 1 .

5. the uncertainty calculation unit sets the variance value of the class of each pixel calculated using Monte Carlo dropout as uncertainty; The judgment evaluation device according to claim 1 .

6. A judgment evaluation method for evaluating a judgment result for input data based on a judgment model, comprising: The computer Calculating the uncertainty of the determination result; evaluating the determination result based on the uncertainty; It is a way of doing things, the input data is an image, The determination is a process of classifying the image into regions of a plurality of classes, The computer Calculating classification uncertainty for each pixel, calculating the half-width of local peaks of uncertainty in the distribution of uncertainty across the entire image at locations where there are regions of different classes in the class distribution on both sides of the local peaks of uncertainty in the distribution of uncertainty, calculating the mode of the half-width, and if the mode exceeds a predetermined threshold, evaluating the judgment result of the image as having high boundary uncertainty. Judgment evaluation method.

7. A judgment evaluation program for evaluating a judgment result for input data based on a judgment model, Calculating the uncertainty of the determination result; evaluating the determination result based on the uncertainty; It is a program that makes a computer do things, the input data is an image, The determination is a process of classifying the image into regions of a plurality of classes, The computer, Calculating classification uncertainty for each pixel, calculating the half-width of local peaks of uncertainty in the distribution of uncertainty across the entire image at locations where there are regions of different classes in the class distribution on both sides of the local peaks of uncertainty in the distribution of uncertainty, calculating the mode of the half-width, and if the mode exceeds a predetermined threshold, evaluating the judgment result of the image as having high boundary uncertainty. To make it happen, Judgment evaluation program.

Images