Image processing device, image processing method, and program for image processing device

Abstract

This image processing device 1 includes a processor 10. The processor 10 stores one or more first machine learning models in a setting space having a first setting condition and a second setting condition. The one or more first machine learning models are disposed at a prescribed density or below in the setting space. The processor: receives an identification image, selects one or more second machine learning models in a prescribed range from the identification image in the setting space or in the order of increasing distance from the identification image in the setting space, and identifies the identification image using the second machine learning model.

Description

Image processing device, image processing method, and image processing program

[0001] The present invention relates to an image processing device that performs image classification processing using a machine learning model, an image processing method that performs image classification processing using a machine learning model, and an image processing program that performs image classification processing using a machine learning model.

[0002] Development is underway of an image processing method that uses an image processing device to input endoscopic images taken with an endoscope or pathological images taken with a microscope of a pathological specimen, and identifies lesions contained in the images using AI with a machine-learned neural network model (hereinafter, "model"). For example, in an endoscopic image identification process using AI, images for identification are created using various parameter sets.

[0003] An image parameter set is a combination of setting conditions such as parameters at the time of shooting and image processing parameters. Therefore, in order to ensure high classification accuracy in AI image classification processing, it is preferable to create a dedicated model using dedicated training data for each parameter set, which is a combination of multiple parameters.

[0004] However, creating many models requires the preparation of a huge amount of training data and long learning times, and storing many models also places a heavy load on the hardware.

[0005] Japanese Patent Application Laid-Open No. 2021-149640 discloses a data classification method for synthesizing data output from multiple models.

[0006] Japanese Patent Application Laid-Open No. 2021-149640

[0007] An embodiment of the present invention aims to provide an image processing device that minimizes the number of machine learning models it incorporates and outputs highly accurate judgment results, an image processing method that minimizes the number of machine learning models it incorporates and outputs highly accurate judgment results, and an image processing program that minimizes the number of machine learning models it incorporates and outputs highly accurate judgment results.

[0008] An image processing device according to an embodiment of the present invention includes a processor, wherein in a setting space having first setting conditions and second setting conditions, the processor stores at least one first machine learning model and a plurality of first thresholds, the number of which is less than the total number of combinations of the first setting conditions and the second setting conditions, and the one or more first machine learning models and at least one of the plurality of first thresholds are arranged in the setting space at a predetermined density or less, and the processor receives an identification image input as an identification target, and selects at least one second machine learning model and a plurality of thresholds from among the one or more first machine learning models and at least one of the plurality of first thresholds, the second machine learning models and at least one of the plurality of first thresholds being within a predetermined range from the identification image in the setting space or in order of shortest distance from the identification image, and identifies the identification image using the one or more second machine learning models and at least one of the plurality of thresholds.

[0009] In an image processing method according to an embodiment of the present invention, a processor stores, in a setting space having first setting conditions and second setting conditions, one or more first machine learning models and at least one of a plurality of first thresholds, the number of which is less than the total number of combinations of the first setting conditions and the second setting conditions, and the one or more first machine learning models and at least one of the plurality of first thresholds are arranged in the setting space at a density equal to or less than a predetermined density, the processor receives an identification image input as an identification target, and selects, from among the one or more first machine learning models and at least one of the plurality of first thresholds, one or more second machine learning models and at least one of the plurality of first thresholds that are within a predetermined range from the identification image in the setting space or in order of shortest distance from the identification image, and the processor identifies the identification image using the one or more second machine learning models and at least one of the plurality of thresholds.

[0010] In an image processing program according to an embodiment of the present invention, a processor of the image processing device stores, in a setting space having first setting conditions and second setting conditions, one or more first machine learning models and at least one of a plurality of first thresholds, the number of which is less than the total number of combinations of the first setting conditions and the second setting conditions, and the one or more first machine learning models and at least one of the plurality of first thresholds are arranged in the setting space at a density equal to or less than a predetermined density, the processor receives an identification image input as an identification target, and selects at least one of one or more second machine learning models and at least one of the plurality of first thresholds from among the one or more first machine learning models and at least one of the plurality of first thresholds, the second machine learning models and at least one of the plurality of first thresholds being within a predetermined range from the identification image in the setting space or in order of shortest distance from the identification image, and the processor is configured to identify the identification image using the one or more second machine learning models and at least one of the plurality of thresholds.

[0011] According to an embodiment of the present invention, it is possible to provide an image processing device that minimizes the number of machine learning models it incorporates and outputs highly accurate judgment results, an image processing method that minimizes the number of machine learning models it incorporates and outputs highly accurate judgment results, and an image processing program that minimizes the number of machine learning models it incorporates and outputs highly accurate judgment results.

[0012] 1 is a configuration diagram of an endoscope system including an image processing device of an embodiment; FIG. 2 is a flowchart of model creation processing of the image processing device of an embodiment; FIG. 3 is a flowchart of identification processing of the image processing device of a first embodiment; FIG. 4 is a layout diagram of a setting space of image parameters of the image processing device of an embodiment; FIG. 5 is a layout diagram of a setting space of image parameters of the image processing device of an embodiment; FIG. 6 is an endoscopic image of the image processing device of an embodiment; FIG. 7 is an endoscopic image of the image processing device of an embodiment; FIG. 8 is a layout diagram of a setting space of image parameters of the image processing device of an embodiment; FIG. 9 is a layout diagram of a setting space of image parameters of the image processing device of an embodiment; FIG. 10 is a flowchart of identification processing of the image processing device of a second embodiment; FIG. 11 is a flowchart of identification processing of the image processing device of a third embodiment; FIG. 12 is a layout diagram of a setting space of image parameters of the third embodiment; FIG. 13 is a layout diagram of a setting space of image parameters of the image processing device of a second modified example; FIG. 14 is a layout diagram of a setting space of image parameters of the image processing device of a second modified example;

[0013] First Embodiment An image processing device 1 of this embodiment will be described below with reference to the drawings. As shown in Fig. 1, an image captured by an endoscope 2 is processed by a video processor (endoscopic processor) 3 and displayed on a monitor 4. A user detects a lesion site based on the endoscopic image displayed on the monitor 4 while operating the endoscope 2. The image processing device 1 is an AI system including a machine-learned model that assists the user in detecting a lesion site.

[0014] The image processing device 1 is a computer including a processor 10 and a storage unit (memory) 20. The processor 10, which is made up of a CPU, loads a program and a machine-learned model (hereinafter also referred to as a "model") stored in the storage unit 20, and constitutes a plurality of functional units, each of which performs a predetermined function. That is, the processor 10 includes a selection unit 11, a calculation unit 12, an identification unit 13, a synthesis unit 14, and a determination unit 15. The processor 10 may include the storage unit 20.

[0015] At least one of the functional units of the processor 10 may be configured by a dedicated hardware circuit. The programs and models may be stored in the server 31 via a network. The image processing device 1 and the video processor 3 may be connected via a network.

[0016] <Creating Machine Learning Models> In the image processing device 1, before the start of the classification process, a plurality of trained models are created in advance, and the trained models are stored, for example, in the storage unit 20. A method for creating trained models will be described below with reference to the flowchart in FIG.

[0017] <Step S10> Setting Parameter Setting Space An endoscopic image is acquired based on a combination of a plurality of image parameters (setting conditions) (hereinafter referred to as a "parameter set").

[0018] Each parameter (setting condition) is classified into a predetermined number of parameter levels in the setting space. If a parameter does not match the parameter level of the setting, the parameter is processed, for example, by rounding off. For example, a parameter with a color tone R level of "1.3" is processed as a parameter level of "1" in the setting space.

[0019] Examples of parameters are shown below. Models: 100 series, 200 series, 1500 series Observation light: WLI (white light), NBI (narrow band special light), RDI (red light) Structure emphasis processing: A processing (A1 to A8), B processing (B1 to B8) TXI (structure emphasis processing): TXI1 (HIGH to LOW), TXI2 (HIGH to LOW) Color tone: R (R-8 to R8), B (B-8 to B8), saturation (-8 to +8)

[0020] Color conversion parameters (bias level, contrast level, gamma level), geometric conversion parameters (sharpening processing level, blur correction level, distortion correction level), etc. may also be set as image parameters.

[0021] Note that the subject data and detection target data may be set as image parameters. Subject: Upper digestive tract (pharynx, esophagus, stomach, duodenum), lower digestive tract (small intestine, large intestine, rectum) Detection target: Cancer (adenocarcinoma, squamous cell carcinoma, small cell carcinoma), inflammation, blood vessels, bleeding points, nerves, fat, ureters, or urethra

[0022] From the plurality of image parameters, K image parameters are selected for classification processing. There is no particular limit to the number K of image parameters to be selected. However, by prioritizing industrially important image parameters, the number K of image parameters to be selected is preferably, for example, 2 or more, and more preferably 3 or more. The number K of image parameters is preferably 5 or less.

[0023] Then, when K image parameters are selected, a K-dimensional parameter setting space having K parameter axes is set.

[0024] For simplicity's sake, the following description will be given taking as an example a two-dimensional setting space (setting plane) having two parameter axes, as shown in Figure 4. The first setting condition, structure enhancement processing A processing, has eight levels (A1 to A8). The second setting condition, color tone R, has seven levels (-3 to 3). Therefore, the maximum (total) number N of parameter sets (first parameter sets) that are combinations of the first and second setting conditions and can be arranged in this setting space is 56.

[0025] In this embodiment, the first setting condition has eight levels and the second setting condition has seven levels, but the number of levels can be set appropriately depending on each parameter. Furthermore, the upper and lower limits of each parameter value are determined based on the specifications of the endoscope or endoscope processor.

[0026] Hereinafter, for example, a parameter set in which the A processing is at level 5 and the color tone R is at level 0 will be referred to as parameter set (A5:R0).

[0027] <Step S11> Selection of Second Parameter Sets The selector 11 selects a plurality of second parameter sets (neighboring parameter sets) for creating a model from a plurality of first parameter sets arranged in the setting space.

[0028] 4, nine second parameter sets A to I (M=0.16N) are arranged in a space in which a total of 56 first parameter sets (N=56) can be arranged. The selection of the second parameter set may be performed automatically by the selection unit 11 based on predetermined conditions, or may be performed by the user.

[0029] The plurality of second parameter sets are selected so that the arrangement density in the setting space is equal to or less than a predetermined density.

[0030] The upper limit of the number M of second parameter sets is preferably 50% or less of the number N of first parameter sets, and particularly preferably 30% or less of the number N of first parameter sets. The smaller the number of second parameter sets, the less effort is required to prepare models corresponding to the second parameter sets. Furthermore, if the number M of second parameter sets is equal to or less than the upper limit, the image processing device 1 has high efficiency and a small hardware load. The lower limit of the number M of second parameter sets is preferably 5% or more of the number N of first parameter sets, and particularly preferably 10% or more of the number N of first parameter sets. If the number M of second parameter sets is equal to or greater than the lower limit, the image processing device 1 has high recognition reliability.

[0031] The plurality of second parameter sets are arranged in a dispersed state in the setting space. For example, a virtual frame (also called a "kernel") having a size of 30% of the size of the setting space is set. Regardless of where this frame is positioned in the setting space, the number M of second parameter sets is 40% or less, and preferably 20% or less, of the number N of first parameter sets within the frame. However, the setting space may have agglomerated regions where the second parameter sets are partially adjacent to each other or where the plurality of second parameter sets are arranged closer to each other than in the surrounding regions.

[0032] The shape of the frame is not particularly limited. When the setting space is two-dimensional, the shape of the frame is, for example, a circle, an ellipse, a rectangle, a square, or a polygon. When the setting space is three-dimensional, the shape of the frame is, for example, a sphere, a hemisphere, a cube, or a cone.

[0033] It is preferable that the plurality of second parameter sets distributed in the setting space have at least some of the parameter levels spaced at equal intervals in the setting space.

[0034] For example, on the hue R axis, the hue R level of the second parameter set A is (R+3), the hue R level of the second parameter set D is (R0), and the hue R level of the second parameter set G is (R-3). The interval between the level of the second parameter set A and the level of the second parameter set D and the interval between the level of the second parameter set D and the level of the second parameter set G are the same, 2 levels.

[0035] The procedure for selecting the plurality of second parameter sets preferably involves first selecting a representative parameter set, and then selecting a plurality of parameter sets that are a predetermined distance from the representative parameter set. When selecting more second parameter sets, the procedure involves selecting a plurality of parameter sets that are a predetermined distance from the already selected parameter set.

[0036] The representative parameter set may be, for example, a parameter set obtained during the product development stage and set in advance, or a parameter set that is expected to be used most frequently in the user's environment. The frequency of use may be calculated based on the parameter sets applied to multiple images collected as training data.

[0037] The second parameter set is distributed approximately evenly in the setting space, so that the reliability of the classification process is guaranteed regardless of where the third parameter set of the classification image is located in the setting space, as will be described later.

[0038] For example, the second parameter set may be arranged uniformly or approximately uniformly throughout the entire setting space, which has the advantage that output can be performed with a certain degree of accuracy for any identification image.

[0039] <Step S12> Model Learning M trained models (first machine learning models) corresponding to the selected M second parameter sets are created. That is, the M trained models are created by learning using a plurality of images (teacher data), for example, by deep learning.

[0040] The correct answer value in the training data stores the positive probability (also called the "classification result" in the case of a value output by an AI model) for identifying positive areas (lesion areas) or negative areas (normal areas) contained in the image.

[0041] For example, a large number of images judged to be normal (positive probability P = 0) and images of cancerous lesions (positive probability P = 1) are prepared as training data and input into the neural network to perform model learning. Data augmented images may also be used as training data.

[0042] A model optimal for certain conditions may be created by training using only images acquired under certain conditions and their data augmentation images as training data. Alternatively, the training data may be limited to certain conditions, and only images acquired under multiple conditions close to the certain conditions and their data augmentation images may be used as training data. In other words, by training only images under any set conditions, a model optimal for the set conditions can be created.

[0043] <Step S13> Model Storage The created M first machine learning models are stored in the storage unit 20, for example.

[0044] That is, in a setting space having first and second setting conditions, the image processing device 1 stores a plurality of first machine learning models, the number of which is less than the total number of combinations of the first and second setting conditions, and the first plurality of machine learning models are arranged in the setting space at a density equal to or less than a predetermined density.

[0045] As long as the processor 10 of the image processing device 1 is in a state where it can use a predetermined model. The storage unit 20 may be a storage unit of the processor 10, the external storage device 30, or the server 31. Furthermore, model learning may be performed using a computer separate from the image processing device 1.

[0046] Furthermore, a model selected by the user from among models created by a third party and stored in the external storage device 30, the server 31, or the like may be transferred to the processor 10 and used in the identification process. For example, if there are already multiple trained models and the parameters of the multiple trained models can be arranged at a predetermined density or less in the setting space, these already created models may be used.

[0047] <Classification Processing> The classification processing in the image processing device 1 will be described with reference to the flowchart in FIG.

[0048] <Step S20> Input of image for identification The image of the subject captured by the endoscope 2 is processed by the video processor 3 and displayed on the monitor 4 as an endoscopic image. The processor 10 of the image processing device 1 also receives the endoscopic image. The image received by the processor 10 may be the same as or different from the image displayed on the monitor 4. For example, a white light image may be displayed on the monitor 4, and a special light image may be input to the processor 10.

[0049] The processor 10 may acquire the imaging conditions from the endoscope 2 or the video processor 3. The imaging conditions may be manually input by a person to the processor 10. The imaging conditions correspond to the above-mentioned parameters.

[0050] For example, 30 endoscopic images per second are output to the monitor 4 and displayed as a moving image on the monitor 4. The transfer rate of the endoscopic images input to the processor 10 may be lower than the transfer rate of the endoscopic images input to the monitor 4, for example, 1 image per second, in accordance with the processing capacity of the processor 10.

[0051] The parameter set data (third parameter set, classification parameter set) for an endoscopic image to be classified (image data, classification image, third image) input from the video processor 3 to the image processing device 1 may be analyzed. Furthermore, when the imaging conditions are acquired as described above, parameters may be identified from the imaging conditions and linked to the endoscopic image. For example, in the example shown in FIG. 5, the third parameter set for the input endoscopic image is X(A3, B+2) in the setting space.

[0052] The image processing device 1 calculates the distances between the parameter information (third parameter set) held by the video processor 3 and a plurality of machine learning models (fourth parameter set) in the setting space.

[0053] Some of the conditions of the third parameter set may be input by the user, or may be calculated by the image processing device 1 from the endoscopic image.

[0054] <Step S21> Distance Calculation The distance L between the third parameter set X and each of the plurality of second parameter sets A to I in the setting space is calculated. For example, the Manhattan distance L(X-A) between the third parameter set X and the second parameter set A is calculated to be "4". Similarly, L(X-B), L(X-C), L(X-D), L(X-E), L(X-F), L(X-G), L(X-H), and L(X-I) are calculated.

[0055] The distance L in the space may be any of the Euclidean distance, Mahalanobis distance, Manhattan distance, and Chebyshev distance.

[0056] For example, the Euclidean distance L between parameter set A (w, x, y, z) and parameter set B (w + Δw, x + Δx, y + Δy, z + Δz) in a four-dimensional space is L ２ = Δw ２ +Δx ２ +Δy ２ +Δz ２ When the Mahalanobis distance is used as the distance L, the calculation is performed using the coordinates of the parameter set to be compared instead of the mean vector.

[0057] The calculation method of the distance L may be changed depending on the dimension of the setting space. For example, a multidimensional setting space may be regarded as a collection of multiple lower-dimensional setting spaces, and the average value of the distances calculated in each lower-dimensional setting space may be used as the distance L.

[0058] <Step S22> Selection of Fourth Parameter Set In the setting space, K fourth parameter sets are selected from the second parameter sets in ascending order of distance L from the third parameter set, where K is an integer equal to or greater than 1.

[0059] For example, in the setting space of Fig. 5, four second parameter sets A, B, D, and E are selected as the fourth parameter set in order of shortest distance L. Images corresponding to the fourth parameter set are called second images. In other words, the second images are selected from the first images corresponding to the second parameter set.

[0060] When model synthesis is performed, the number K of fourth parameter sets is preferably 2 or more, and particularly preferably 3 or more. The number K of fourth parameter sets is preferably 20 or less, and more preferably 10 or less. If the number K of fourth parameter sets is equal to or greater than the above range, the image processing device 1 has high classification reliability. If the number K of fourth parameter sets is equal to or less than the above range, the image processing device 1 is highly efficient and has a small hardware load.

[0061] In addition, if there are multiple second parameter sets with the same distance L, only one of the second parameter sets may be selected as the fourth parameter set, or more than a predetermined number K of fourth parameter sets may be selected.

[0062] For example, the third parameter set of the endoscopic image input in the setting space shown in Fig. 5 is (A7, B+2). In the image processing device 1 set to select two fourth parameter sets, two second parameter sets B and C, two second parameter sets C and F, or three second parameter sets B, C, and F are selected.

[0063] It goes without saying that if any second parameter set exists at the same position as the third parameter set in the setting space, only the second parameter set at the same position as the third parameter set is selected as the fourth parameter set, and in this case, there is no need to calculate the following synthesis coefficients.

[0064] <Step S23> Calculation of Synthesis Coefficient The calculation unit 12 calculates a synthesis coefficient k (model synthesis coefficient) which is a weighting coefficient corresponding to the distance L of the fourth parameter set. Specifically, each synthesis coefficient k is calculated so that the weight w of the fourth parameter set having a shorter distance L is larger. For example, the weighting in the synthesis coefficient k is inversely proportional to the distance L.

[0065] For example, the combination coefficient kA-kE of the models (second machine learning models) A, B, D, and E corresponding to the fourth parameter set A is calculated as follows using the Manhattan distance L:

[0066] Manhattan distance L L(X-A) = 3, L(X-B) = 3, L(X-D) = 4, L(X-E) = 4 Weight w wA = 1 / 3, wB = 1 / 3, wD = 1 / 4, wE = 1 / 4 Weight sum wALL wALL = wA + wB + wD + wE = (1 / 3) + (1 / 3) + (1 / 4) + (1 / 4) = 14 / 12 Combining coefficient k kA = wA / wLL = (1 / 3) / (14 / 12) = 4 / 14 kB = wB / wLL = (1 / 3) / (14 / 12) = 4 / 14 kD = wD / wLL = (1 / 4) / (14 / 12) = 3 / 14 kE=wE / wLL=(1 / 4) / (14 / 12)=3 / 14

[0067] The weight w in the synthesis coefficient k may be inversely proportional to the distance L, either logarithmically or exponentially.

[0068] In the above example, one distance L is applied to one parameter set to calculate the weight w. In contrast, when a multidimensional setting space is regarded as a collection of multiple lower-dimensional setting spaces and weights are calculated from multiple distances calculated in each of the lower-dimensional setting spaces, a weight may be calculated from each of the multiple distances and the weight w may be calculated by multiplying these weights.

[0069] <Step S24> Model Composition The composition unit 14 of the image processing device 1 is a model composition unit that combines a plurality of models (second machine learning models) A, B, D, and E corresponding to the fourth parameter sets A, B, D, and E in accordance with a composition coefficient k for weighting each of the models. Because the model composition is a weighted average using the composition coefficient k, it is a low-load process that can be completed in a short time.

[0070] A model may consist of multiple layers, and the synthesis coefficients may differ for each layer. For example, the synthesis unit 14 may weight and synthesize only the output layers and intermediate layers adjacent to the output layers of each of the multiple models according to the synthesis coefficient k. For example, if a model corresponding to the fourth parameter set E is used as a representative model, the input layer and intermediate layers adjacent to the input layer may use the values ​​of the representative model as is, and the output layer and intermediate layers adjacent to the output layer may be weighted and synthesized according to the synthesis coefficient k. In this method, the input layer and intermediate layers adjacent to the input layer using the representative model can extract common features from the input image, and the output layer and intermediate layers adjacent to the output layer using the synthesis model can output the extracted common features while taking into account the influence of the parameter set. Furthermore, since a common value is used on the input layer side, models can be developed at low cost.

[0071] Furthermore, the representative model and the composite model may be composited. In this case, the composite coefficient may be changed for each layer. For example, in the input layer, the composite coefficient of the representative model may be set to 0.9, and the composite coefficient of the composite model may be set to 0.1. In the intermediate layer, the composite coefficient of the representative model may be decreased and the composite coefficient of the composite model may be increased as the model transitions from the input layer side to the output layer side. In the output layer, the composite coefficient of the representative model may be set to 0.1, and the composite coefficient of the composite model may be set to 0.9. In the above method, the representative model and the composite model can be used together smoothly.

[0072] <Step S25> Classification Processing The classification unit 13 performs classification processing on the input classification image (third image) using a synthetic model. The probability that the classification image contains a lesion site is output as a positive probability P (0≦P≦1) (S25A). A positive probability P=0 means the possibility is 0%, and a positive probability P=1 means the possibility is 100%. The determination unit 15 then determines that the image contains a lesion site if the positive probability P is equal to or greater than a predetermined determination threshold T (S25B). For example, if the threshold T is 0.5, a commonly used standard threshold, and the positive probability P is 0.8, it is determined that the image contains a lesion site.

[0073] The classification is performed for each pixel of the image, for each region including a plurality of pixels, or for the entire image including the lesion site as one region.

[0074] <Step S26> Output The processor 10 outputs the determination result, which is an estimated value estimated from the positive probability and the threshold value.

[0075] For example, for the endoscopic image of Fig. 6, the lesion area is highlighted and superimposed on the endoscopic image and displayed on the monitor 4 as shown in Fig. 7. In the case of a detection type in which identification is performed in rectangular frame units, rather than a segmentation type in which identification is performed in pixel units of an image, the lesion area is displayed as a rectangular frame.

[0076] The identification unit 13 may output the positive probability P without performing the determination process S25B. For example, in the segmentation type, a probability map may be created and the positive probability may be displayed in shades of color or superimposed on the endoscopic image.

[0077] As long as the judgment result can be notified to the user, a classification type may be used in which the judgment result is displayed as text indicating the probability that the image contains a lesion area in an area separate from the endoscopic image display area on the monitor 4. If the judgment result is equal to or greater than a predetermined probability, an alarm may be sounded.

[0078] <Step S27> The process from step S20 onwards is repeated until an instruction is received from the user. That is, the operation of the image processing device 1 continues in real time until the endoscopic examination is completed.

[0079] 8A to 8D, a specific example of the classification process by the image processing device 1 will be described. In Figures 8A to 8D, models of nine second parameter sets (first machine learning models) are stored in the image processing device 1 from 61 first parameter sets that can be arranged in the setting space.

[0080] For example, in the case shown in Figure 8*, an image of parameter set U observed by a doctor at medical institution U is subjected to classification processing using a model (second machine learning model) corresponding to parameter sets A, C, and D.

[0081] On the other hand, for an image of parameter set V observed with an endoscope by a doctor at medical institution V, a classification process is performed using models corresponding to parameter sets B, C, and E.

[0082] For example, in the case shown in Figure 8A, for an image with parameter set X observed with an endoscope of model number X, classification processing is performed using a model (second machine learning model) corresponding to parameter sets A, C, and D.

[0083] On the other hand, for an image of parameter set Y observed with an endoscope of model number Y, classification processing is performed using models corresponding to parameter sets B, C, and E.

[0084] In order to further improve the classification accuracy, a setting space may be provided for each model or each organ, etc. For example, the setting space shown in Fig. 8B corresponds to an endoscopic image of the esophagus, and when the endoscopic image has parameter set X, classification processing is performed using models corresponding to parameter sets A, C, and D.

[0085] The setting space shown in Figure 8C corresponds to a white light image of the stomach, and when the endoscopic image is parameter set Y, the identification process is performed using a model corresponding to parameter sets B, C, and E, or a model corresponding to parameter sets B, E, and F.

[0086] The setting space shown in Figure 8D corresponds to NBI light of the stomach, and four fourth parameter sets are selected. When the endoscopic image is parameter set Z, the classification process is performed using models corresponding to parameter sets E, F, H, and I.

[0087] As described above, the processor 10 receives an identification image input as an identification target. The processor 10 selects a plurality of second machine learning models from a plurality of first machine learning models in order of whether they are within a predetermined range from the identification image in the setting space or in order of the shortest distance from the identification image. The processor 10 creates a composite model by combining the selected plurality of second machine learning models based on their respective combination coefficients. The processor 10 identifies the identification image using the composite model. That is, the processor 10 identifies the identification image using the selected plurality of second machine learning models.

[0088] In other words, the storage unit 20 of the image processing device 1 stores a plurality of first machine learning models, each corresponding to a first image acquired using a plurality of second parameter sets selected from a maximum number of a plurality of first parameter sets that can be arranged in an image parameter setting space having a plurality of parameter axes. The calculation unit 12 calculates a distance in the setting space between the plurality of second parameter sets and a third parameter set of an identification image input as an identification target. The selection unit 11 selects a plurality of fourth parameter sets from the plurality of second parameter sets, within a predetermined range or in order of shortest distance. The classification unit 13 classifies the identification image using the plurality of second machine learning models corresponding to each second image acquired using the plurality of fourth parameter sets.

[0089] The calculation unit 12 of the image processing device 1 further calculates model combination coefficients for each of the second machine learning models corresponding to the second images acquired using the selected fourth parameter sets. The combination unit 14 creates a combined model by combining the second machine learning models using the respective model combination coefficients. The identification unit 13 identifies the identification image using the combined model.

[0090] In machine learning (including deep learning), if there is a difference in the setting conditions (parameter set) between the training data and the data for judgment, the accuracy of classification will decrease.

[0091] An image parameter set is a combination of setting conditions such as parameters at the time of shooting and image processing parameters. Therefore, in order to ensure high classification accuracy in AI image classification processing, it is preferable to create a dedicated model using dedicated training data for each parameter set, which is a combination of multiple parameters.

[0092] However, creating many models requires the preparation of a huge amount of training data and long learning times, and storing many models also places a heavy load on the hardware.

[0093] It is possible to train a single model using training data for various parameter sets, but a model trained using training data for various parameter sets is not optimized for any of the parameter sets, resulting in generally low accuracy.

[0094] The image processing device 1 ensures the reliability of classification even when it is unable to use a model corresponding to the same parameters as the input image for classification. Because there is no need to create (train) and store many models, the image processing device 1 can minimize the number of machine learning models installed and output highly accurate determination results.

[0095] Second Embodiment Image Processing Device 1A The classification process in the image processing device 1A of this embodiment will be described with reference to the flowchart in Fig. 9. Since the image processing device 1A is similar to the image processing device 1, components with the same functions are assigned the same reference numerals and descriptions thereof will be omitted.

[0096] In the image processing device 1A, the processor 10 selects, from among a plurality of first machine learning models, a plurality of second machine learning models that are within a predetermined range from the classification image in the setting space. The processor 10 classifies the classification image using each of the selected second machine learning models, and outputs a plurality of positive probabilities (classification results). The processor 10 combines the output positive probabilities (classification results) based on their respective combination coefficients.

[0097] In other words, in the image processing device 1A, the selection unit 11 selects multiple fourth parameter sets. The calculation unit 12 calculates each positive probability combination coefficient k corresponding to the second image acquired using the selected multiple fourth parameter sets. The identification unit 13 outputs multiple positive probabilities using multiple machine learning models corresponding to the selected multiple fourth parameter sets, and combines the multiple positive probabilities using the respective combination coefficients k.

[0098] The method for creating the machine learning model (steps S10 to S13) and steps S30 to S33 are the same as those of the image processing device 1 shown in FIGS. 2 and 3, and therefore a description thereof will be omitted.

[0099] <Step S34> Classification Processing The image processing apparatus 1A performs classification processing (AI processing) using a plurality of models corresponding to the images of the plurality of fourth parameter sets (S34A).

[0100] In the example of Figure 5, the classification unit 13 performs classification processing (AI processing) of the classification image of the third parameter set X four times using four second models A, B, D, and E corresponding to the fourth parameter sets A, B, D, and E, respectively (S34A).

[0101] The synthesis unit 14 is a positive probability synthesis unit that synthesizes the positive probabilities PA, PB, PD, and PE output by the four second models A, B, D, and E using respective synthesis coefficients (positive probability synthesis coefficients) kA, kB, kD, and kE (S34B). The positive probability synthesis coefficients are weighted according to the distance L, and the calculation method therefor is substantially the same as the model synthesis coefficient calculation method already described.

[0102] Below, an example will be described in which the positive probability P of a region in an image is as follows:

[0103] Model A (kA = 4 / 14): PA = 0.8 Model B (kB = 4 / 14): PB = 0.7 Model D (kD = 3 / 14): PD = 0.6 Model E (kE = 3 / 14): PE = 0.5 PT = PA x kA + PB x kB + PD x kD + PE x kE = 0.66

[0104] The determination unit 15 determines that the region corresponds to a lesion site because the combined positive probability PT is equal to or greater than a predetermined threshold value (for example, 0.5) (S34C).

[0105] Steps S35 and S36 are the same as steps S26 and S27 of the image processing device 1 shown in FIG. 3, and therefore a description thereof will be omitted.

[0106] The image processing device 1A, like the image processing device 1, ensures the reliability of classification even when it is not possible to use a model corresponding to the same parameters as the parameters of the input classification image. Because there is no need to create (learn) and store many models, the image processing device 1A can minimize the number of machine learning models installed and output highly accurate determination results.

[0107] Third Embodiment: Image Processing Device 1B The classification process in the image processing device 1B of this embodiment will be described with reference to the flowchart in Fig. 10. Since the image processing device 1B is similar to the image processing devices 1 and 1A, components with the same functions are designated by the same reference numerals and will not be described again.

[0108] The processor 10 of the image processing device 1B stores, in a setting space having first setting conditions and second setting conditions, a plurality of first thresholds that are fewer than the total number of combinations of the first setting conditions and the second setting conditions, and the plurality of first thresholds are arranged at a predetermined density or less in the setting space. The processor 10 classifies the identification image using a single machine learning model that is a predetermined representative model, and outputs a positive probability (identification result). The processor 10 selects, from the plurality of first thresholds, a plurality of second thresholds that are within a predetermined range from the identification image in the setting space or in order of shortest distance from the identification image. The processor 10 determines the output positive probability (identification result) based on a combined threshold obtained by combining the plurality of second thresholds.

[0109] In other words, the image processing device 1B calculates a threshold synthesis coefficient from the second thresholds of the multiple fourth parameter sets in accordance with the distance between the third parameter set and each of the multiple selected fourth parameter sets in the setting space, and identifies the identification image using the threshold synthesis coefficient.

[0110] 11, the storage unit 20 of the image processing device 1B stores only one representative model E corresponding to an image of one preset parameter set E. The storage unit 20 stores first thresholds TA to TI corresponding to each of the selected second parameter sets A to I. The representative model E may be, for example, a preset parameter set or a frequently used parameter set. The first thresholds TA to TI are acquired in advance.

[0111] The selection unit 11 of the image processing device 1B selects a plurality of fourth parameter sets A, B, D, and E. The calculation unit 12 calculates threshold synthesis coefficients kA, kB, kD, and kE for the selected plurality of fourth parameter sets A, B, D, and E. The threshold synthesis coefficients are weighted according to the distance L, and the calculation method thereof is substantially the same as the model synthesis coefficient calculation method already described.

[0112] The synthesis unit 14 is a threshold synthesis unit that synthesizes thresholds (second thresholds) TA, TB, TD, TE of the plurality of fourth parameter sets A, B, D, E based on respective threshold synthesis coefficients kA, kB, kD, kE to create a synthesized threshold TT. The classification unit 13 processes the classification image using a predetermined representative model and classifies the positive probability using the synthesized threshold TT.

[0113] Steps S40 to S42 are the same as those in the image processing device 1 shown in Fig. 3, and therefore will not be described here. A second threshold value T is set for each of the fourth parameter sets selected in step S42.

[0114] As already explained, first thresholds TA to TI are set in advance for each of the second parameter sets A to I of the image processing device 1B. Of the first thresholds TA to TI, the second thresholds TA, TB, TD, and TE of the fourth parameter set are used in the following calculations.

[0115] The second thresholds TA, TB, TD, TE of the selected fourth parameter set A, B, D, E are illustrated.

[0116] Parameter set A (A1, R+3): threshold value TA = 0.8 Parameter set B (A5, R+3): threshold value TB = 0.7 Parameter set D (A1, R0): threshold value TD = 0.6 Parameter set E (A5, R0): threshold value TE = 0.5

[0117] <Step S43> Calculation of composite threshold The calculation unit 12 calculates a composite threshold TT from the distances LA, LB, LD, and LE between the third parameter set X in the setting space and the fourth parameter sets A, B, D, and E. The composite threshold TT is weighted so that the shorter the distance L, the greater the weight of the threshold of the fourth parameter set.

[0118] Synthesis threshold TT=0.8×(4 / 14)+0.7×(4 / 14)+0.6×(3 / 14)+0.5×(3 / 14)=0.66

[0119] <Step S44> Classification The classification unit 13 performs AI processing on the classification image using the representative model E and outputs the positive probability P E (S44A). For example, if the positive probability P E is 0.7, the determination unit 15 determines that the combined positive probability P E is equal to or greater than the combined threshold T T (0.66), and therefore determines that the region corresponds to a lesion (S44B).

[0120] <Step S45> Output The processor 10 displays the estimated value estimated from the positive probability and the combined threshold as the judgment result on the monitor 4, superimposed on the endoscopic image.

[0121] Like the image processing device 1, the image processing device 1B ensures the reliability of classification even when it is not possible to use a model corresponding to the same parameters as the parameters of the input classification image. Because only one model is created, the image processing device 1B can minimize the number of machine learning models installed and output highly accurate determination results.

[0122] <Modifications> Image processing devices 1C to 1G of modifications are similar to the image processing device 1 and the like, so components with the same functions are given the same reference numerals and descriptions thereof are omitted.

[0123] <Modification 1> Image processing device 1C As shown in Fig. 5, in the image processing device 1 or the like, a plurality of fourth parameter sets A, B, D, and E are selected from the second parameter sets A to I in the setting space. In contrast, in the image processing device 1C of this modification, one fourth parameter set (for example, parameter set B) having the shortest distance L from the third parameter set is selected. Then, the selected one fourth parameter set B is used to perform the classification process.

[0124] Compared to the image processing device 1, the image processing device 1C has lower recognition reliability, but is more efficient and places a smaller load on hardware.

[0125] <Modification 2> Image Processing Device 1D In the image processing device 1D, the method of selecting the fourth parameter set by the selector 11 differs from the selection method of the image processing device 1 and the like.

[0126] The selector 11 of the image processing device 1D selects, from the plurality of second parameter sets A to I, all of the plurality of fourth parameter sets that are within a predetermined range from the third parameter set of the identification image.

[0127] For example, the selection unit 11 selects, as the fourth parameter set, the second parameter set whose distance L from the third parameter set is equal to or less than 4. In this case, the number of fourth parameter sets selected for the image processing device 1D is four, which is the same as the number of fourth parameter sets for the image processing device 1, etc.

[0128] <Modification 3> Image processing device 1E In the image processing device 1, the second parameter sets are arranged to be approximately evenly dispersed in the setting space ( FIG. 4 ). In contrast, as shown in FIG. 12 , in the image processing device 1E of this modification, the multiple second parameter sets A to I are arranged more densely in the setting space in an area close to the position of a preset representative parameter set E than in an area far from the position of the representative parameter set E.

[0129] The representative parameter set E is set in advance by the user and is, for example, the most frequently used parameter set.

[0130] That is, for example, more second parameter sets may be placed in positions in the setting space that are expected to be used frequently than in other positions. In this case, there is an advantage that highly accurate output can be achieved for most classification targets. Figure 13 shows an example of such placement. In this case, of parameter sets A to I, parameter sets C, D, E, F, and G are expected to be used frequently. In other words, placement in this manner is used when the classification target is expected to have parameters from parameter sets C, D, E, F, and G or parameters in the vicinity of these.

[0131] The image processing device 1E is more likely to be able to select a fourth parameter set that is a shorter distance L from the third parameter set of the classification image in the setting space than the image processing device 1. Therefore, the image processing device 1E can achieve classification processing that is more reliable than the image processing device 1.

[0132] <Modification 4> Image Processing Device 1F An image processing device 1F of this modification is a combination of the functions of the image processing device 1 and the functions of the image processing device 1B.

[0133] The processor 10 of the image processing device 1F stores, in a setting space having first setting conditions and second setting conditions, one or more first machine learning models and multiple first thresholds, the number of which is less than the total number of combinations of the first setting conditions and the second setting conditions. The one or more first machine learning models and the multiple first thresholds are arranged in the setting space at a density equal to or less than a predetermined density. The processor 10 receives an identification image input as an identification target. The processor 10 selects, from the one or more first machine learning models and the multiple first thresholds, one or more second machine learning models and multiple second thresholds that are within a predetermined range from the identification image in the setting space or in order of shortest distance from the identification image. The processor 10 identifies the identification image using the one or more second machine learning models and the multiple second thresholds.

[0134] For example, the memory unit 20 of the image processing device 1F stores machine learning models A, B, D, E and thresholds TA, kB, kD, kE corresponding to each second image acquired in multiple fourth parameter sets A, B, D, E.

[0135] In the image processing device 1F, similar to the image processing device 1, an AI process is performed on the image for classification using a composite model that combines machine learning models A, B, D, and E, and a positive probability is output.

[0136] The positive probability determination process uses a composite threshold value obtained by combining the threshold values ​​TA, kB, kD, and kE, as in the image processing device 1B.

[0137] <Modification 5> Image Processing Device 1G An image processing device 1G of this modification is a combination of the functions of the image processing device 1A and the functions of the image processing device 1B.

[0138] The memory unit 20 of the image processing device 1G stores machine learning models A, B, D, E and threshold synthesis coefficients kA, kB, kD, kE corresponding to each second image acquired in multiple fourth parameter sets A, B, D, E.

[0139] In the image processing device 1G, the processor 10 calculates the positive probability of each of the classification images using machine learning models A, B, D, and E, in the same way as in the image processing device 1A, and calculates a combined positive probability.

[0140] For the positive probability determination process, as with the image processing device 1B, the machine learning models A, B, D, and E are combined based on their respective threshold combination coefficients kA, kB, kD, and kE, and the combined positive probability is determined using a combination threshold TT.

[0141] In the image processing devices 1 and 1A to 1G described above, the image for identification is a medical image, and the identification unit 13 identifies a lesion site.

[0142] When a machine learning model is used in the above embodiment and modified examples, it can also be applied to the technology of an automatic switching model.

[0143] For example, after inserting an endoscope through the mouth and advancing it through the pharynx, upper esophagus, lower esophagus, and stomach, the image for identification changes as it is returned to the mouth. This is because the mucous membranes in each area have different colors, or because the observation light is changed from white light mode to special light mode when observing the stomach.

[0144] For example, the preferred model differs depending on whether the image for identification is an image of the lower esophagus or an image of the stomach. However, by using the present invention, an appropriate model or synthetic model can be applied even if the observation site changes.

[0145] That is, the identification image (third image) is an endoscopic image acquired continuously, and the identification unit 13 automatically selects a machine learning model and performs identification processing in accordance with changes in multiple parameters of the endoscopic image.

[0146] The identification image may be a medical image such as a pathological image that has been acquired in advance and stored in, for example, the external storage device 30. In other words, the image processing device does not need to be directly connected to the video processor 3, but may be indirectly connected via the external storage device 30 or the like.

[0147] Alternatively, the image may be an endoscopic image, and the identifying unit 13 may identify the position of the tip of the endoscope, which is the position inside the body where the identification image was taken.

[0148] For example, an insertion navigation system has been disclosed that forms a three-dimensional image of the bronchi of the lungs from three-dimensional image data of a subject obtained by a CT device or the like, determines a path to a destination point along the lumen on the three-dimensional image, and further generates and displays a virtual endoscope image of the lumen based on the three-dimensional image data.The image processing device can identify the position where the identification image was taken (the tip position of the endoscope inside the body) by selecting a virtual endoscope image that matches the endoscopic image of the bronchoscope inserted into the lungs.

[0149] The endoscope of the embodiment may be a flexible endoscope with a flexible insertion portion or a rigid endoscope with a rigid insertion portion, and may be used for medical or industrial purposes.

[0150] In an image processing method according to an embodiment, a processor stores at least one first machine learning model and a plurality of first thresholds in a setting space having first setting conditions and second setting conditions, the number of which is less than the total number of combinations of the first setting conditions and the second setting conditions. The one or more first machine learning models and at least one of the plurality of first thresholds are arranged at a predetermined density or less in the setting space. The processor receives an identification image input as an identification target. The processor selects at least one second machine learning model and a plurality of second thresholds from among the one or more first machine learning models and at least one of the plurality of first thresholds, in order of whether the second machine learning model and the plurality of second thresholds are located within a predetermined range from the identification image in the setting space or in order of the shortest distance from the identification image. The processor identifies the identification image using the one or more second machine learning models and at least one of the plurality of second thresholds.

[0151] In an embodiment, an image processing program is configured such that a processor of an image processing device stores, in a setting space having first setting conditions and second setting conditions, at least one first machine learning model and at least one of a plurality of first thresholds, the number of which is less than the total number of combinations of the first setting conditions and the second setting conditions. The one or more first machine learning models and at least one of the plurality of first thresholds are arranged at a predetermined density or less in the setting space. The processor receives an identification image input as an identification target. The processor is configured to select at least one second machine learning model and at least one of a plurality of second thresholds from among the one or more first machine learning models and at least one of the plurality of first thresholds, either within a predetermined range from the identification image in the setting space or in order of shortest distance from the identification image. The processor is configured to identify the identification image using at least one of the one or more second machine learning models and at least one of the plurality of second thresholds.

[0152] The program can also be distributed by storing it on a non-transitory computer-readable storage medium such as a magnetic disk (e.g., a floppy disk, a hard disk), an optical disk (e.g., a CD-ROM, a DVD), a magneto-optical disk (MO), or a semiconductor memory. The storage medium may be in any storage format as long as it can store the program and is computer-readable.

[0153] In addition, an OS (operating system), database management software, network software, or other MW (middleware) running on a computer may execute some of the processes required to realize the above-described embodiments based on instructions from a program installed on the computer from a storage medium.

[0154] The storage medium is not limited to a medium independent of a computer, but also includes a storage medium that stores or temporarily stores a program downloaded from a LAN, the Internet, etc. Furthermore, the storage medium is not limited to one medium, and the storage medium of the present invention also includes a case where each of the above processes is executed from multiple media, and the media configuration may be any configuration.

[0155] In addition, the computer in each embodiment executes each process in each of the above embodiments based on a program stored in a storage medium, and may be configured as either a single device such as a personal computer, or a system in which multiple devices are connected to a network.

[0156] Furthermore, the term "computer" is not limited to personal computers but also includes arithmetic processing units and microcomputers included in information processing equipment, and is a general term for devices and equipment that can realize the functions of the present invention by means of a program.

[0157] The present invention is not limited to the above-described embodiments and modifications, and various changes and modifications can be made without departing from the spirit and scope of the present invention.

[0158] REFERENCE SIGNS LIST 1, 1A-1G... Image processing device 2... Endoscope 3... Video processor 4... Monitor 10... Processor 11... Selection unit 12... Calculation unit 13... Identification unit 14... Synthesis unit 15... Determination unit 20... Storage unit 30... External storage device 31... Server

Claims

1. a storage unit that stores a setting space having a first setting condition and a second setting condition, wherein at least one or more first machine learning models and a plurality of first thresholds are stored, the number of which is less than the total number of combinations of the first setting condition and the second setting condition, and the one or more first machine learning models and at least one of the plurality of first thresholds are arranged at a density equal to or less than a predetermined density; a selection unit that receives an identification image input as an identification target, and selects, from among the one or more first machine learning models and the plurality of first thresholds, at least one second machine learning model and a plurality of second thresholds that are located within a predetermined range from the identification image in the setting space or in order of shortest distance from the identification image; and an identification unit that identifies the identification image using at least one of the one or more second machine learning models and the plurality of second thresholds.

2. A compound comprising a synthesis portion, the selection unit selects a plurality of second machine learning models; the combining unit combines the plurality of second machine learning models based on their respective combination coefficients to create a combined model; The image processing device according to claim 1 , wherein the classifying unit classifies the image for classification using the synthetic model.

3. A compound comprising a synthesis part, the selection unit selects a plurality of second machine learning models; the classification unit classifies the image for classification using each of the plurality of second machine learning models, and outputs a plurality of classification results; The image processing apparatus according to claim 1 , wherein the synthesis unit synthesizes the plurality of output classification results based on respective synthesis coefficients.

4. A compound comprising a synthesis part, the identification unit uses a predetermined machine learning model to identify the image for identification, and outputs an identification result; the selection unit selects a plurality of second thresholds in order of whether they are within a predetermined range from the identification image in the setting space or in order of whether they have a shortest distance from the identification image; The image processing apparatus according to claim 1 , wherein the combining unit determines the output classification result based on a combined threshold value obtained by combining the plurality of second threshold values.

5. The image processing device described in Claim 2, characterized in that the synthesis unit calculates the synthesis coefficient according to the distance between the identification image in the setting space and at least one of the one or more second machine learning models and the plurality of second thresholds.

6. The image processing device according to claim 2, further comprising a calculation unit that calculates the distance between the identification image and the one or more second machine learning models in the setting space from parameter information held by an endoscope or an endoscope processor.

7. The image processing device according to claim 1 , wherein the machine learning models are arranged evenly in terms of at least one of the first setting condition and the second setting condition.

8. The storage unit stores a plurality of machine learning models, The image processing device according to claim 1 , wherein the placement density in the setting space of at least one of the plurality of machine learning models and the plurality of thresholds is 50% or less.

9. The image processing device according to claim 1, wherein the placement density in the setting space of at least one of the plurality of first machine learning models and the plurality of first thresholds is higher the closer it is to the position of a preset representative parameter set.

10. The image processing device according to claim 1 , wherein the first machine learning model is generated to be optimal for an arbitrary set condition by learning only images under the arbitrary set condition.

11. The image processing device described in Claim 1, characterized in that the selection unit selects one second machine learning model with the shortest distance from the identification image in the setting space.

12. the identification image is a medical image, The image processing device according to claim 1 , wherein the identifying unit identifies a lesion site.

13. the identification image is an endoscopic image, The image processing device according to claim 1 , wherein the identification image is output from an endoscope processor.

14. the identification images are endoscopic images acquired successively, The image processing device described in claim 1, characterized in that the image processing device automatically selects at least one of the one or more second machine learning models and multiple second thresholds in response to changes in the first setting conditions and the second setting conditions of the endoscopic image, and performs classification processing.

15. The machine learning model comprises a plurality of layers, The image processing device according to claim 2 , wherein the blending coefficients are different for each of the plurality of layers.

16. An image processing method using an image processing device including a storage unit, a selection unit, and a classification unit, the storage unit stores, in a setting space having a first setting condition and a second setting condition, at least one of one or more first machine learning models and a plurality of first thresholds, the number of which is less than a total number of combinations of the first setting condition and the second setting condition, and the one or more first machine learning models and at least one of the plurality of first thresholds are arranged in the setting space at a density equal to or less than a predetermined density; the selection unit receives an identification image input as an identification target, and selects, from among the one or more first machine learning models and the plurality of first thresholds, at least one of one or more second machine learning models and a plurality of second thresholds in order of being within a predetermined range from the identification image in the setting space or being closest to the identification image; An image processing method, characterized in that the classification unit classifies the image for classification using at least one of the one or more second machine learning models and the plurality of second thresholds.

17. A program for an image processing device including a selection unit, an identification unit, and a memory unit that stores, in a setting space having a first setting condition and a second setting condition, at least one first machine learning model and at least one of a plurality of first thresholds that are less than the total number of combinations of the first setting condition and the second setting condition, and the one or more first machine learning models and at least one of the plurality of first thresholds are arranged at a predetermined density or less in the setting space; the selection unit receives an identification image input as an identification target, and selects at least one of one or more second machine learning models and a plurality of second thresholds from among at least one of the one or more first machine learning models and the plurality of first thresholds, in order of being within a predetermined range from the identification image in the setting space or being closest to the identification image; The program for an image processing device, characterized in that the identification unit is configured to identify the identification image using at least one of the one or more second machine learning models and the plurality of second thresholds.