Estimated dose examination method based on multiple artificial neural networks

Abstract

The present invention relates to an estimated dose inspection method for a plurality of artificial neural networks, which is characterized by calibrating the frequency of a double-centromere chromosome on the basis of the number of chromosomes, the number of chromosome slices, and the frequency of the double-centromere chromosome estimated by a plurality of neural network models, and determining an estimated dose corresponding to the frequency of the calibrated double centromere chromosome, thereby enabling the estimated dose to be obtained with high precision.

Description

Technical Field

[0001] This invention relates to a method for examining and estimating doses by passing chromosome images through multiple sequentially connected artificial neural networks. Background Technology

[0002] Generally, the radiation dose estimate (Gy) is calculated by determining the frequency of exposed dicentric chromosomes and then inputting it into the dose estimate calculation formula.

[0003] Therefore, the first step is to identify dicentric chromosomes. Generally speaking, when exposed to low doses of radiation of 0-1 Gy, many pseudo-dicentric chromosomes will be identified, which leads to a decrease in the accuracy of dicentric chromosome frequency and thus results in the disadvantage of difficulty in accurately estimating radiation dose.

[0004] In view of the above factors, a technology needs to be developed to prevent the identification of pseudodicentric chromosomes, thereby obtaining a more accurate radiation dose estimate. Summary of the Invention

[0005] Technical issues The purpose of this invention is to calibrate the frequency of dicentric chromosomes based on the number of chromosomes, the number of chromosome slices, and the frequency of dicentric chromosomes estimated by multiple neural network models, and to determine the estimated dose corresponding to the calibrated dicentric chromosome frequency.

[0006] The objectives of this invention are not limited to those mentioned above. Other objectives and advantages of this invention not mentioned can be understood through the following description, and can be further clarified through embodiments of the invention. Furthermore, it can be readily seen that the objectives and advantages of this invention can be achieved through the solutions and combinations thereof shown in the claims.

[0007] Technical solution A dose estimation method based on multiple artificial neural networks according to an embodiment of the present invention includes: a step of inputting multiple chromosome images, which are the subjects of dose evaluation, into a first neural network model to determine whether they can be read; a step of inputting the determination result and the chromosome images determined to be readable into a second neural network model to identify the number of chromosomes in the chromosome images; a step of inputting the chromosome images into a third neural network model to identify the number of chromosome slices in the chromosome images; a step of inputting the chromosome images into a fourth neural network model to identify the number of dicentric chromosomes; a step of calculating the frequency of dicentric chromosomes based on the number of dicentric chromosomes; a step of inputting the number of chromosomes, the number of chromosome slices, and the frequency of dicentric chromosomes into a fifth neural network model to calibrate the frequency of dicentric chromosomes; and a step of determining the estimated dose corresponding to the calibrated frequency of dicentric chromosomes.

[0008] According to an embodiment of the present invention, the method for estimating dose based on multiple artificial neural networks further includes: a step of collecting chromosome images for learning; a step of marking whether chromosomes in the chromosome images for learning can be read; and a step of setting the chromosome images for learning as input data of the first neural network model and setting the readability as output data of the first neural network model to guide supervised learning of the first neural network model.

[0009] The first neural network model outputs either of the two categories corresponding to the readable and unreadable chromosome images.

[0010] According to an embodiment of the present invention, the method for estimating dose based on multiple artificial neural networks further includes: a step of collecting chromosome images for learning; a step of marking the positions of individual chromosomes within the chromosome images for learning; and a step of setting the chromosome images for learning as input data of the second neural network model and setting the marked positions as output data of the second neural network model to guide the learning of the second neural network model.

[0011] The second neural network model outputs the position of individual chromosomes within the chromosome image used for reading.

[0012] The step of identifying the number of chromosomes includes calculating the number of bounding boxes corresponding to individual chromosome locations output from the second neural network model.

[0013] According to an embodiment of the present invention, the method for estimating dose based on multiple artificial neural networks further includes: a step of collecting chromosome images for learning; a step of marking the positions of individual chromosome slices within the chromosome images for learning; and a step of setting the chromosome images for learning as input data of the third neural network model and setting the marked positions as output data of the third neural network model to guide the learning of the third neural network model.

[0014] The third neural network model outputs the location within the chromosome image for individual chromosome slice reading.

[0015] The step of identifying the number of chromosome slices includes calculating the number of bounding boxes corresponding to the positions of individual chromosome slices output from the third neural network model.

[0016] The fourth neural network model includes a fourth-first neural network model that outputs the position of individual dicentric chromosomes in a chromosome image for reading, and a fourth-second neural network model that performs a classification action on the dicentric chromosomes corresponding to each position output from the fourth-first neural network model.

[0017] According to an embodiment of the present invention, the dose estimation method based on multiple artificial neural networks further includes the steps of: cropping the bounding box output to the fourth-first neural network model, and inputting the image of the cropped bounding box into the fourth-2 neural network.

[0018] According to an embodiment of the present invention, the estimated dose detection method based on multiple artificial neural networks further includes: a step of collecting a learning chromosome image; a step of marking the position of an individual dicentric chromosome within the learning chromosome image; and a step of setting the learning chromosome image as input data for a 4-1 neural network model and setting the marked position as output data for the 4-1 neural network model to guide the learning of the 4-1 neural network model.

[0019] According to an embodiment of the present invention, the estimated dose detection method based on multiple artificial neural networks further includes: a step of collecting chromosome images for learning; a step of labeling the categories determined based on whether the chromosomes represented in the chromosome images for learning are dicentric chromosomes; and a step of setting the chromosome images for learning as input data of a fourth-second neural network model and setting the labeled categories as output data of the fourth-second neural network model to guide the learning of the fourth-second neural network model.

[0020] The step of identifying the number of dicentric chromosomes includes calculating the number of categories of dicentric chromosomes output from the fourth-second neural network model.

[0021] The step of calculating the frequency of dicentric chromosomes includes dividing the number of dicentric chromosomes by the number of chromosome images read.

[0022] An estimated dose detection method based on multiple artificial neural networks according to an embodiment of the present invention includes the following steps: collecting the number of chromosomes, the number of chromosome slices, and the frequency of dicentric chromosomes; marking the measured frequency of dicentric chromosomes in the chromosome image for reading; and setting the number of chromosomes, the number of chromosome slices, and the number of dicentric chromosomes as input data of the fifth neural network model and setting the measured frequency as output data to guide the learning of the fifth neural network model.

[0023] The step of determining the estimated dose includes substituting the calibrated dicentric chromosome frequency into an equation defining the correlation between the frequency of the dicentric chromosome and the radiation dose to determine the estimated dose.

[0024] Invention Effects The advantage of this invention is that it calibrates the frequency of dicentric chromosomes based on the estimated number of chromosomes, the number of chromosome slices, and the frequency of dicentric chromosomes through multiple neural network models, and determines the estimated dose corresponding to the calibrated frequency of dicentric chromosomes, thereby enabling the estimated dose to be obtained with high accuracy.

[0025] The effects and specific effects of the present invention will be described below while explaining the specific aspects of implementing the invention. Attached Figure Description

[0026] Figure 1 This is a flowchart of a dose estimation and checking method based on multiple artificial neural networks according to an embodiment of the present invention.

[0027] Figure 2 This is a diagram used to illustrate the actions of the first neural network model.

[0028] Figure 3a It is a chromosome image for reading.

[0029] Figure 3b It is impossible to read chromosome images.

[0030] Figure 4 This is a diagram used to illustrate the actions of the second neural network model.

[0031] Figure 5 It is a chromosome image used to display the identified chromosomes and their number.

[0032] Figure 6 This is a diagram used to illustrate the actions of the third neural network model.

[0033] Figure 7 It is a chromosome image used to read the identified chromosome slices and their number.

[0034] Figure 8 This is a diagram used to illustrate the actions of the 4-1 neural network model.

[0035] Figure 9 It is a chromosome image used to display and identify dicentric chromosomes and their number.

[0036] Figure 10 and Figure 11 This is a diagram used to illustrate the actions of the neural network model 4-2.

[0037] Figure 12 This is a diagram used to illustrate the actions of the fifth neural network model.

[0038] Figure 13 This is a graph showing the average error of the estimated dose with or without application according to the present invention. Detailed Implementation

[0039] The technical spirit of the invention will be readily implemented by those skilled in the art through a detailed description of the foregoing objects, features, and advantages with reference to the accompanying drawings. In describing the invention, detailed descriptions of well-known technologies related to the invention are omitted if it is deemed unnecessary to obscure the essence of the invention. Hereinafter, preferred embodiments according to the invention will be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings are used to refer to the same or similar components.

[0040] In this specification, terms such as "first" and "second" are used to describe various components, but these components are by no means limited by these terms. These terms are used only to distinguish one component from another, and unless specifically stated otherwise, a first component may be a second component.

[0041] Furthermore, in this specification, the term "arbitrarily configured in the upper (or lower) part" or "upper (or lower) part" of a component means not only that it is arbitrarily configured in connection with the top (or bottom) of the component, but also that other configurations can be sandwiched between the component and any configuration configured on the upper (or lower) part of the component.

[0042] Additionally, if this list describes a component as "connected," "combined," or "linked" to another component, then the components can be directly connected or linked to each other. However, it must be understood that other components can be "interlocked" between components, or that components can be "connected," "combined," or "linked" through other components. Furthermore, the singular expressions used in this listing contain multiple expressions unless clearly different in context. In this application, terms such as “constituting” or “comprising” should not be construed as necessarily including all the components or steps described in the specification, and should be interpreted as excluding some components or steps or including additional components or steps.

[0043] In addition, "A and / or B" in this specification means A, B or A and B unless otherwise stated otherwise, and "C to D" means above C and below D unless otherwise stated otherwise.

[0044] A dose estimation method based on multiple artificial neural networks according to an embodiment of the present invention may include: step S100 of determining whether multiple chromosome images can be read; step S200 of identifying the number of chromosomes in the chromosome image to be read; step S300 of identifying the number of chromosome slices in the chromosome image to be read; step S400 of identifying dicentric chromosomes in the chromosome image to be read; step S500 of calculating the frequency of dicentric chromosomes based on the number of dicentric chromosomes; step S600 of calibrating the frequency of dicentric chromosomes; and step S700 of determining the estimated dose corresponding to the calibrated dicentric chromosomes.

[0045] However, based on Figure 1 The dose estimation method using multiple artificial neural networks shown is based on one embodiment, and the steps for implementing the invention are not limited to... Figure 1 The embodiments shown can be modified or deleted as needed; Figure 1 Each step shown can be executed by a processor, which may include at least one physical element among application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), a controller, and a micro-controller.

[0046] The following is about Figure 1 Each step S100-S700 shown is explained in detail.

[0047] The processor can input multiple chromosome images as dose evaluation objects into the first neural network model 10 to determine whether they can be read (S100). Specifically, as follows... Figure 2 As shown, the processor can input multiple chromosome images to be dose-evaluated into the first neural network model 10, and the first neural network model 10 can output either of the two categories of readable and unreadable corresponding to the input chromosome images.

[0048] At this point, the chromosome image can be a chromosome image that has not been used in the learning of the first neural network model 10, and the first neural network model 10 can output a readable or unreadable category by receiving the chromosome image.

[0049] For example, the first neural network model 10 can be input as follows: Figure 3a The image shown is of chromosomes. Figure 3a The chromosome image shown clearly shows the chromosome shape represented in chromosome region A, and can be a sharp, clear image. Based on this, the first neural network model 10 can output a readable category. The output readable chromosome image can be named a readable chromosome image.

[0050] Conversely, the first neural network model 10 can be input... Figure 3b The image shows a chromosome. Figure 3b The chromosome image shown can be an image where the chromosome shape represented in chromosome region B is unclear, blurry, or indistinct. Based on this, the first neural network model 10 can output unreadable categories.

[0051] On the other hand, for the aforementioned actions, a learning process for the first neural network model 10 can be performed first. For this purpose, the processor can collect first learning chromosome images. As mentioned earlier, the first learning chromosome images can be images with clearly defined chromosome shapes and sharp focus, or images with indistinct chromosome shapes, blurred focus, or unclear images.

[0052] Subsequently, the processor can label whether the first learning chromosome image has been read or not, based on user commands. To elaborate further, each first learning chromosome image can be labeled with a ground truth (GT) corresponding to the image; that is, whether a first learning chromosome image can be read can be labeled with a different class.

[0053] Subsequently, the processor sets the first learning chromosome image as the input data of the first neural network model 10, and sets the readability / disability as the output data of the first neural network model 10, thereby enabling supervised learning of the first neural network model 10. Specifically, the processor can set the first learning chromosome image as the input data of the first neural network model 10, and can supervise and learn the first neural network model 10 based on the measured values ​​corresponding to the data, that is, based on the training dataset (which uses the readability / disability label as the output data).

[0054] Through this guided learning, the first neural network model 10 can learn the correlation between the first learning chromosome image and whether it can be read, and after the learning is completed, it can also classify whether the chromosome images not used for learning can be read.

[0055] Then, the processor inputs the readable chromosome image into the second neural network model (20) through the first neural network model 10, thereby identifying the number of chromosomes S200 in the readable chromosome image.

[0056] Therefore, such as Figure 4 As shown, the processor can input the chromosome image for reading into the second neural network model 20, and the second neural network model 20 can output the position within the input individual chromosome image for reading.

[0057] On the other hand, a learning process for the second neural network model 20 can be performed first for the aforementioned actions. For this purpose, the processor can collect chromosome images for the second learning. At this time, as mentioned earlier, the chromosome images used for the second learning can be images with clearly defined chromosome shapes and sharp focus.

[0058] Subsequently, referring to Figure 5 The processor can label the bounding boxes C of individual chromosomes within the second learning chromosome image according to user commands. Specifically, the processor can label the measured values ​​corresponding to the second learning chromosome image according to user commands, that is, label the center coordinates, width, and height of the bounding box C of each chromosome represented in the second learning chromosome image.

[0059] Subsequently, the processor can set the second learning chromosome image as the input data of the second neural network model 20, and decrement the marked positions, that is, set the bounding box C as the output data of the second neural network model 20, thereby pre-guiding the learning of the second neural network model 20. Specifically, the processor can set the second learning chromosome image as the input data of the second neural network model 20, and pre-guide the learning of the second neural network model 20 based on the measured values ​​corresponding to the input data, that is, based on the training dataset of the center coordinates, width, and height of the bounding boxes C of individual chromosomes in the second learning chromosome image as the output data.

[0060] Through this guided learning, the second neural network model 20 can learn the correlation between the second learned chromosome image and the chromosome position. After the learning is completed, it can also identify the position of individual chromosomes in unused chromosome images.

[0061] The processor can then calculate the number of bounding boxes C corresponding to the positions of individual chromosomes output from the second neural network model 20. Based on this, the processor identifies the number of individual chromosomes. That is, , as... Figure 5 As shown, the processor can identify 45 individual chromosomes.

[0062] Then, the processor inputs the chromosome image that is determined to be readable into the third neural network model 30 through the first neural network model 10 to identify the number of chromosome slices in the chromosome image S300.

[0063] Therefore, such as Figure 6 As shown, the processor can input the chromosome image for reading into the third neural network model 30, and the third neural network model 30 can output the position of individual chromosome slices within the input chromosome image for reading.

[0064] On the other hand, for the aforementioned actions, the learning process of the third neural network model 30 can be performed first. For this purpose, the processor can collect images of the third learning chromosome. At this time, the images of the third learning chromosome can also be images with clearly defined chromosome shapes and sharp focus.

[0065] Subsequently, referring to Figure 7 The processor can label the bounding boxes (D) of individual chromosome slices within the learning chromosome image according to user commands. Specifically, the processor can label the measured values ​​corresponding to the third learning chromosome image according to user commands, that is, label the center coordinates, width, and height of the bounding box (D) of each chromosome slice represented in the third learning chromosome image.

[0066] Subsequently, the processor sets the third learning chromosome image as the input data of the third neural network model 30, and sets the marked positions, i.e., the bounding boxes D, as the output data of the third neural network model 30, thereby performing pre-guided learning on the third neural network model 30. Specifically, the processor can set the third learning chromosome image as the input data of the third neural network model 30, and perform pre-guided learning on the third neural network model 30 based on the measured values ​​corresponding to the data, i.e., based on the training dataset with the center coordinates, width, and height of the bounding boxes D that divide individual chromosome slices within the third learning chromosome image as output data.

[0067] Through this guided learning, the third neural network model 30 can learn the correlation between the locations of chromosome images and chromosome slices. After the learning is completed, it can also identify the location of individual chromosome slices for unused chromosome images.

[0068] The processor can then calculate the number of bounding boxes C, corresponding to the positions of individual chromosome slices output from the third neural network model 30. Based on this, the processor identifies the number of individual chromosome slices. That is, as... Figure 7 As shown, the processor can identify an individual chromosome.

[0069] Then, the processor reads the chromosome image and inputs it into the fourth neural network model to identify the number of dicentric chromosomes S400. At this time, the action of identifying the number of dicentric chromosomes is the same as the action of identifying the number of chromosomes or the number of chromosome slices mentioned above, and it can be done by only one fourth neural network model.

[0070] However, in order to improve the accuracy of dicentric chromosome identification, the fourth neural network model can include the fourth-first neural network model and the fourth-second neural network model.

[0071] First, the operation of the 4-1 neural network model is explained. The processor can input the readable chromosome image into the 4-1 neural network model 40a through the 1st neural network model 10 to identify the position of the dicentric chromosome in the chromosome image.

[0072] Specifically, such as Figure 8 As shown, the processor can input the chromosome image for reading into the 4-1 neural network model 40a, and the 4-1 neural network model 40a can output the position of individual dicentric chromosomes within the input chromosome image for reading.

[0073] On the other hand, for the aforementioned actions, the learning process of the fourth-first neural network model 40a can be performed first. For this purpose, the processor can collect chromosome images for the fourth-first learning process. At this time, the chromosome images for the fourth-first learning process can also be images with clearly defined chromosome shapes and sharp focus.

[0074] Subsequently, referring to Figure 9 The processor can label the bounding boxes E of individual dicentric chromosomes within the 4-1 learning chromosome image according to user commands. Specifically, the processor can label the measured values ​​corresponding to the 4-1 learning chromosome image according to user commands, that is, label the center coordinates, width, and height of the bounding box E of each dicentric chromosome represented in the 4-1 learning chromosome image.

[0075] Subsequently, the processor sets the 4-1 learning chromosome image as the input data of the 4-1 neural network model 40a, and sets the marked positions, i.e., the bounding boxes E, as the output data of the 4-1 neural network model 40a, thereby performing pre-guided learning of the 4-1 neural network model 40a. Specifically, the processor can set the 4-1 learning chromosome image as the input data of the 4-1 neural network model 40a, and perform pre-guided learning of the 4-1 neural network model 40a based on the measured values ​​corresponding to the input data, i.e., based on the training dataset with the center coordinates, width, and height of the bounding boxes E of individual dicentric chromosomes within the 4-1 learning chromosome image as the output data.

[0076] Through this guided learning, the 4-1 neural network model 40a can learn the correlation between the chromosome images used for learning and the positions of dicentric chromosomes, and after learning, it can also identify the positions of individual dicentric chromosomes for unused chromosome images.

[0077] The processor can then crop the bounding box E output to the 4-1 neural network model 40a. Based on this, an image representing a dicentric chromosome can be generated.

[0078] Next, the operation of the 4-2 neural network model will be explained. The processor can input the image of the cropped bounding box E into the 4-2 neural network model 40b to perform the classification operation of dicentric chromosomes.

[0079] Specifically, such as Figure 10 and 11 As shown, the processor can input the cropped bounding box image E into the 4-2 neural network, and can perform classification operations on the dicentric chromosomes in the input bounding box image E.

[0080] Specifically, such as Figure 10 As shown, the processor can input the cropped bounding box image E into the 4-2 neural network model, and the 4-2 neural network model can output the category corresponding to the input bounding box image E.

[0081] At this point, the cropped bounding box E image can be an image that was not used in the learning of the 4-2 neural network 40b model, and the 4-2 neural network model 40b can output the category of the cropped bounding box E image.

[0082] For example, the 4-2 neural network model 40b can be input as follows: Figure 10 The cropping bounding box E image is shown. Figure 10 The bounding box E image shown can be a dicentric chromosome image (dic). Accordingly, the neural network model 40b of the fourth-second generation can output the category corresponding to the dicentric chromosome.

[0083] Conversely, the 4-2 neural network model 40b can be input as follows: Figure 11 The cropping bounding box E image is shown. Figure 11 The bounding box E shown can be a non-dic chromosome. Accordingly, the neural network model 40b of the fourth-second generation can output the category corresponding to the non-dic chromosome.

[0084] On the other hand, for the aforementioned actions, the learning process of neural network model 40b (4-2) can be performed first. For this purpose, the processor can collect cropped bounding box images. At this time, the bounding box images can be cropped bounding box images.

[0085] Subsequently, the processor can label whether the bounding box contains a dicentric chromosome based on user commands. Specifically, the processor can label the measured values ​​corresponding to the bounding box image according to user commands, that is, the category of dicentric or non-dicentric chromosomes represented in the bounding box image.

[0086] Subsequently, the processor sets the bounding box image as the input data for the 4-2 neural network model 40b, and sets whether it is a dicentric chromosome as the output data for the 4-2 neural network model 40b, thereby enabling pre-guided learning of the 4-2 neural network model 40b. Specifically, the processor can set the bounding box image as the input data for the 4-2 neural network model 40b, and perform pre-guided learning of the 4-2 neural network model 40b based on the measured values ​​corresponding to the input data, that is, based on the training dataset with the category of dicentric or non-dicentric chromosomes in the bounding box image as the output data.

[0087] Through this guided learning, the neural network model 40b can learn the correlation between bounding box images and whether or not a chromosome is dicentric. After learning, it can also identify whether a chromosome is dicentric or non-dicentric even for unused bounding box images.

[0088] The processor can then calculate the number of categories output as dicentric chromosomes in the 4-2 neural network model 40b. This enables the processor to identify the number of dicentric chromosomes. That is, as... Figure 9 As shown, the processor can identify two individual dicentric chromosomes.

[0089] Then, the processor can calculate the frequency S500 of dicentric chromosomes based on the number of identified dicentric chromosomes. At this point, the processor can calculate this by dividing the number of dicentric chromosomes by the number of chromosome images read. For example, if the number of dicentric chromosomes is 30 and the number of chromosome images read is 3000, then the frequency of dicentric chromosomes is 30 / 3000, which is 0.01.

[0090] Then, the processor can input the number of chromosomes, the number of chromosome slices, and the frequency of dicentric chromosomes into the fifth neural network model 50 to calibrate the frequency of dicentric chromosomes S600.

[0091] Specifically, such as Figure 12 As shown, the processor can input the number of chromosomes, the number of chromosome slices, and the frequency of dicentric chromosomes into the fifth neural network model 50, and the fifth neural network model 50 can output the calibrated frequency of dicentric chromosomes.

[0092] On the other hand, for the aforementioned actions, the learning process of the fifth neural network model 50 can be performed first. For this purpose, the processor can collect data on chromosome number, chromosome slice number, and the frequency of dicentric chromosomes.

[0093] Subsequently, the processor can mark the measured frequencies in the chromosome image for reading according to the user's commands. Specifically, the processor can mark the measured values ​​corresponding to the chromosome image for reading, that is, mark the measured frequencies of dicentric chromosomes represented in the chromosome image for reading, according to the user's commands.

[0094] Subsequently, the processor sets the chromosome number, chromosome slice number, and dicentric chromosome frequency as input data for the fifth neural network model 50, and sets the measured frequencies as output data for the fifth neural network model 50, in order to perform pre-guided learning on the fifth neural network model 50. Specifically, the processor can set the chromosome number, chromosome slice number, and dicentric chromosome frequency as input data for the fifth neural network model 50, and perform pre-guided learning on the fifth neural network model 50 based on the measured values ​​corresponding to the data, that is, based on the training dataset with the measured frequencies of dicentric chromosomes in the chromosome images read as output data.

[0095] Through this guided learning, the fifth neural network model 50 can learn the correlation between the number of chromosomes, the number of chromosome slices, and the frequency and measured frequency of dicentric chromosomes. After learning, it can also output the measured frequency for the number of unused chromosomes, the number of chromosome slices, and the frequency of dicentric chromosomes.

[0096] The processor can then determine the estimated dose S700 corresponding to the frequency of the calibrated dicentric chromosome. Specifically, the processor can determine the estimated dose by substituting the frequency of the calibrated dicentric chromosome into an equation that defines the correlation between the frequency of the dicentric chromosome and the radiation dose.

[0097] At this point, the processor can input the frequency of the calibrated dicentric chromosome into the calculation formula given in the 2011 International Atomic Energy Agency manual to finally calculate the estimated dose.

[0098] Accordingly, refer to Figure 13 It can be seen that the error rate of the estimated dose calculated by inputting the calibrated dicentric frequency into the equation through the fifth neural network model 50 is lower than the error rate of the estimated dose calculated by inputting the uncalibrated dicentric chromosome frequency.

[0099] As described above, the present invention has been illustrated with reference to the accompanying drawings. However, the present invention is not limited to the embodiments and drawings disclosed in this specification. Obviously, those skilled in the art can make various modifications within the scope of the technical spirit of the present invention. Even if the effects of the configuration according to the present invention are not explicitly described and explained in the previous description of the embodiments of the present invention, the effects that can be predicted based on this configuration should be recognized.

Claims

1. A method of estimated dose checking based on multiple artificial neural networks, characterized in that, comprising: a step of judging whether a plurality of chromosome images as dose evaluation targets are readable by inputting the chromosome images to a first neural network model; a step of identifying the number of chromosomes in a readable chromosome image judged to be readable by inputting the judgment result and the readable chromosome image to a second neural network model; a step of identifying the number of chromosome sections in the readable chromosome image by inputting the readable chromosome image to a third neural network model; a step of identifying the number of dicentric chromosomes by inputting the readable chromosome image to a fourth neural network model; a step of calculating the frequency of dicentric chromosomes based on the number of dicentric chromosomes; a step of calibrating the frequency of dicentric chromosomes by inputting the number of chromosomes, the number of chromosome sections, and the frequency of dicentric chromosomes to a fifth neural network model; and a step of determining an estimated dose corresponding to the calibrated frequency of dicentric chromosomes.

2. The estimated dose checking method based on a plurality of artificial neural networks according to claim 1, further comprising: a step of collecting learning chromosome images; a step of labeling whether chromosomes in the learning chromosome images are readable; and a step of instructing learning of the first neural network model by setting the learning chromosome images as input data of the first neural network model and setting the label as output data of the first neural network model.

3. The estimated dose checking method based on a plurality of artificial neural networks according to claim 1, wherein the first neural network model outputs either of two classes corresponding to readable and unreadable of a chromosome image.

4. The estimated dose checking method based on a plurality of artificial neural networks according to claim 1, further comprising: a step of collecting learning chromosome images; a step of labeling positions of individual chromosomes in the learning chromosome images; and a step of instructing learning of the second neural network model by setting the learning chromosome images as input data of the second neural network model and setting the labeled positions as output data of the second neural network model.

5. The estimated dose checking method based on a plurality of artificial neural networks according to claim 1, wherein the second neural network model outputs positions of individual chromosomes in the readable chromosome image.

6. The estimated dose checking method based on a plurality of artificial neural networks according to claim 1, wherein the step of identifying the number of chromosomes includes a step of calculating the number of bounding boxes corresponding to the positions of individual chromosomes output from the second neural network model.

7. The estimated dose checking method based on a plurality of artificial neural networks according to claim 1, further comprising: a step of collecting learning chromosome images; a step of labeling positions of individual chromosome sections in the learning chromosome images; and a step of instructing learning of the third neural network model by setting the learning chromosome images as input data of the third neural network model and setting the labeled positions as output data of the third neural network model. ​ ​ a step of guiding learning of the third neural network model by setting the learning chromosome image as input data of the third neural network model and setting the marked position as output data of the third neural network model.

8. The estimated dose checking method based on a plurality of artificial neural networks according to claim 1, wherein the third neural network model outputs positions of individual chromosome sections within the reading chromosome image.

9. The estimated dose checking method based on a plurality of artificial neural networks according to claim 1, wherein the step of identifying the number of chromosome sections includes a step of calculating the number of bounding boxes corresponding to the positions of individual chromosome sections output from the third neural network model.

10. The estimated dose checking method based on a plurality of artificial neural networks according to claim 1, wherein the fourth neural network model includes a fourth-1 neural network model that outputs positions of individual dicentric chromosomes within the reading chromosome image and a fourth-2 neural network model that performs a classification action on dicentric chromosomes corresponding to each position output from the fourth-1 neural network model.

11. The estimated dose checking method based on a plurality of artificial neural networks according to claim 10, wherein the step of cropping the bounding box output to the fourth-1 neural network model and inputting an image of the cropped bounding box to the fourth-2 neural network.

12. The estimated dose checking method based on a plurality of artificial neural networks according to claim 10, wherein further comprising: a step of collecting a learning chromosome image; a step of marking positions of individual dicentric chromosomes within the learning chromosome image; and a step of guiding learning of the fourth-1 neural network model by setting the learning chromosome image as input data of the fourth-1 neural network model and setting the marked positions as output data of the fourth-1 neural network model.

13. The estimated dose checking method based on a plurality of artificial neural networks according to claim 10, wherein further comprising: a step of collecting a learning chromosome image; a step of marking a class determined based on whether a chromosome represented in the learning chromosome image is a dicentric chromosome; and a step of guiding learning of the fourth-2 neural network model by setting the learning chromosome image as input data of the fourth-2 neural network model and setting the marked class as output data of the fourth-2 neural network model.

14. The estimated dose checking method based on a plurality of artificial neural networks according to claim 10, wherein the step of identifying the number of dicentric chromosomes includes a step of calculating the number of classes output to dicentric chromosomes from the fourth-2 neural network model.

15. The estimated dose checking method based on a plurality of artificial neural networks according to claim 1, wherein The step of calculating the frequency of the dicentric chromosome includes a step of dividing the number of dicentric chromosomes by the number of chromosome images read.

16. The estimated dose checking method based on a plurality of artificial neural networks according to claim 1, wherein Further comprising: a step of collecting the number of chromosomes, the number of chromosome sections, and the frequency of dicentric chromosomes; a step of marking the measured frequency of dicentric chromosomes in the chromosome image read; a step of setting the number of chromosomes, the number of chromosome sections, and the number of dicentric chromosomes as input data and the measured frequency as output data to the fifth neural network model and guiding learning of the fifth neural network model.

17. The estimated dose checking method based on a plurality of artificial neural networks according to claim 1, wherein The step of determining the estimated dose includes a step of determining the estimated dose by substituting the calibrated dicentric chromosome frequency into an equation defining a correlation between the frequency of dicentric chromosomes and the radiation dose.

Images