Cervical image generation system and method using generative AI model

Abstract

The present application relates to a cervical image generation system and method using a generative AI model, and more particularly, to a cervical image generation system and method using a generative AI model for generating a virtual positive cervical image using a diffusion rate model and a transfer learning technique. The cervical image generation system using a generative AI model according to an embodiment of the present application includes a data collection unit that acquires cervical image data to collect, a data learning unit that learns a generative AI model based on the cervical image data collected by the data collection unit, and an inference unit that inferences input data using a cervical image generation model derived from the data learning unit to generate and output a virtual cervical image.

Description

Technical Field

[0001] This invention relates to a cervical image generation system and method using a generative AI model, and more specifically, to a cervical image generation system and method using a generative AI model to generate virtual positive cervical images using a diffusion rate model and transfer learning techniques. Background Technology

[0002] Unless otherwise stated in this specification, the content described in the background section is not prior art as covered by the claims of this application, and should not be regarded as prior art simply because it is recorded herein.

[0003] A woman's uterus is broadly divided into the uterine body and the cervix. The uterine body (body) accounts for about 75% of the uterus, while the cervix (neck) is the part that connects to the vagina. Cancer that occurs in the cervix is ​​called "cervical cancer."

[0004] According to statistics from the Korea Health Insurance Review and Evaluation Service, over 50,000 women seek medical attention for cervical cancer annually over the past five years. Of these, approximately 25% are around 40 years old, followed by women around 30 and 50 years old, each accounting for over 20%. The number of patients is substantial. Cervical cancer is one of the leading causes of cancer death among women worldwide.

[0005] Recently, according to the cancer registry statistics of the National Cancer Center of Korea in 2020, the incidence of cervical cancer has been declining, with the incidence rate per 100,000 people decreasing from 16.7 in 2009-2013 to 14.2 in 2014-2018.

[0006] This trend has led to a continuous decrease in the amount of positive image data required for AI-based cervical cancer grading models to learn, resulting in a data imbalance between negative and positive image categories. Therefore, acquiring positive cervical cancer image data is crucial for continuously improving the performance of AI-based cervical cancer grading models.

[0007] Against this backdrop, to improve the performance of AI-based cervical cancer grading models and enhance the diagnostic accuracy of cervical cancer, it is necessary to continuously expand the learning data by generating virtual cervical images. Given the declining incidence of cervical cancer in South Korea, the acquisition of positive cervical cancer image data is even more crucial, making virtual image generation technology particularly effective.

[0008] Regarding artificial intelligence-based cervical cancer diagnosis, the applicant previously submitted and obtained authorization for "Artificial Intelligence-Based Cervical Cancer Screening Service System" (Authorized Patent Publication No. 10-2462975, Publication Date: November 8, 2022), which proposed a method for improving the accuracy, objectivity, and speed of cervical cancer screening by utilizing an artificial intelligence learning model, thereby solving the aforementioned problems.

[0009] However, in order to more effectively utilize the AI-based cervical cancer screening service system of the aforementioned prior patent and improve the performance of the AI-based cervical cancer grading model for screening services, it is necessary to supplement the decreasing positive image data to address the data imbalance between negative and positive image categories, and to expand the learning data by generating virtual cervical images.

[0010] In this regard, Patent Publication No. 10-2041402 (published on November 7, 2019) describes a prior art concerning a "cervical learning data generation system and a cervical learning data classification method". However, the above-mentioned prior art does not involve the positive cervical image generation method that utilizes diffusion probability models and transfer learning techniques to generate virtual positive cervical images as expected by the present invention. Instead, it only involves a cervical image data classification method for constructing an AI learning system related to the cervix. From this perspective, it cannot solve the problem proposed by the present invention.

[0011] Prior art literature

[0012] Patent documents

[0013] (Patent Document 1) Patent Publication No. 10-2462975 (Published on November 8, 2022)

[0014] (Patent Document 2) Granted Patent Publication No. 10-2041402 (Published on November 7, 2019) Summary of the Invention

[0015] Technical problems to be solved

[0016] The present invention aims to improve the aforementioned existing problems and provides a cervical image generation system and method using a generative AI model to solve the data imbalance problem between negative and positive image categories required for building an AI-based cervical cancer grading classification model. The invention applies AI diffusion probability models and transfer learning techniques to these image data to generate virtual positive cervical images, thereby continuously improving the performance of the AI-based cervical cancer grading classification model.

[0017] This invention is not limited to the above-mentioned technical issues, and other technical issues can be derived from the following description.

[0018] Technical methods for solving problems

[0019] A cervical image generation system utilizing a generative AI model according to an embodiment of the present invention includes: a data acquisition unit for acquiring cervical image data; a data learning unit for learning a generative AI model based on the cervical image data acquired by the data acquisition unit; and an inference unit for inferring from the input data using the cervical image generation model derived from the data learning unit to generate a virtual cervical image for output.

[0020] According to a preferred embodiment of the present invention, the generative AI model is a diffusion probability model.

[0021] According to a preferred embodiment of the present invention, the data acquisition unit includes: a camera capturing unit for acquiring cervical images; a cervical image interpretation unit for medical personnel to interpret the captured cervical images; and a database server storage unit for receiving and storing the interpreted image data via communication means.

[0022] According to a preferred embodiment of the present invention, the data learning unit includes: a training dataset classification unit, which classifies image data received from the database server storage unit into a training dataset; a training dataset preprocessing unit, which preprocesses the training dataset; and a diffusion probability model learning unit, which learns the diffusion probability model based on the preprocessed data.

[0023] According to a preferred embodiment of the present invention, the inference unit includes: an inference input unit that receives input of an input image or prompt word; an inference preprocessing unit that preprocesses the input data; a cervical image generation model inference unit that applies the preprocessed data to a cervical image generation model learned by the diffusion probability model; and an output unit that outputs a virtual cervical image generated by the generation model inference unit.

[0024] According to a preferred embodiment of the present invention, it further includes: a database for learning a cervical cancer grading classification model based on artificial intelligence, which stores the virtual positive cervical images generated and output by the inference unit after being interpreted by medical personnel.

[0025] A cervical image generation method using a generative AI model according to a preferred embodiment of the present invention includes: a data acquisition step, acquiring cervical image data; a data learning step, learning a generative AI model based on the cervical image data acquired in the data acquisition step; and an inference step, using the cervical image generation model derived from the data learning step to infer the input data, so as to generate a virtual cervical image for output.

[0026] According to a preferred embodiment of the present invention, the generative AI model is a diffusion probability model.

[0027] According to a preferred embodiment of the present invention, the data acquisition step includes: a capture step, in which a cervical image is captured using a camera; a cervical image interpretation step, in which medical personnel interpret the captured cervical image; and a database server storage step, in which the interpreted image data is received and stored via communication means.

[0028] According to a preferred embodiment of the present invention, the data learning step includes: a training dataset classification step, receiving image data from the database server storage unit and classifying it into a training dataset; a training dataset preprocessing step, preprocessing the training dataset; and a diffusion probability model learning step, learning the diffusion probability model based on the preprocessed data.

[0029] According to a preferred embodiment of the present invention, the reasoning step includes: a reasoning input step, in which an input image or prompt word is input into a reasoning input unit; a reasoning preprocessing step, which preprocesses the input data; a cervical image generation model reasoning step, in which the preprocessed data is applied to a cervical image generation model learned by the diffusion probability model; and an output step, in which the output unit outputs the virtual cervical image generated in the generation model reasoning step.

[0030] According to a preferred embodiment of the present invention, the method further includes: an interpretation step, in which medical personnel interpret the virtual positive cervical image generated and output in the reasoning step to determine whether it is a real positive image; and a storage step, in which if the interpretation step determines that the image is the same as a real positive image, the image data is stored in a database for learning a cervical cancer grading classification model based on artificial intelligence.

[0031] According to a preferred embodiment of the present invention, the method further includes: a relearning step, which uses negative and positive image data stored in the AI-based cervical cancer grading model learning database to relearn the AI-based cervical cancer grading model.

[0032] Invention Effects

[0033] According to the present invention, the beneficial effect of the present invention is that it can solve the problem of data imbalance between negative and positive image categories required for constructing an artificial intelligence-based cervical cancer grading classification model.

[0034] According to the present invention, the beneficial effect of the present invention is that, for negative and positive image data, virtual positive cervical images can be generated by using a diffusion probability model and transfer learning technology in a generative AI model, and the AI-based cervical cancer grading classification model can be relearned, thereby continuously improving the performance of the AI-based cervical cancer grading classification model.

[0035] It should be understood that the effects of the present invention are not limited to the above-described effects, but also include all effects that can be inferred from the inventive structure described in the specific description or claims of the present invention. Attached Figure Description

[0036] Figure 1 This is a general block diagram of the cervical image generation system utilizing a generative AI model according to the present invention.

[0037] Figure 2 yes Figure 1 A detailed flowchart of the data learning department.

[0038] Figure 3 yes Figure 1 A detailed block diagram of the reasoning department.

[0039] Figure 4 This is a flowchart of a cervical image generation system utilizing a generative AI model according to Embodiments 1 and 2 of the present invention.

[0040] Figure 5a This is a photograph of an example of virtual positive image generation according to Embodiment 1 of the present invention, which utilizes a generative AI model to generate cervical images.

[0041] Figure 5b This is a photograph of an example of virtual positive image generation using a generative AI model-based cervical image generation system and method according to Embodiment 2 of the present invention.

[0042] Explanation of reference numerals in the attached figures

[0043] 100: Data Acquisition Department

[0044] 110: Camera Shooting Department (Image Acquisition Department)

[0045] 120: Cervical Image Interpretation Department

[0046] 130: Database Server Storage Department

[0047] 131: Production Technology Selection Department

[0048] 132: Selection Department for Unconditional Image Generation Techniques

[0049] 133: Image Conversion Technology Selection Department

[0050] 134: Conditional Image Generation Technology Selection Department

[0051] 143: Condition Setting Department

[0052] 144: Condition Setting Department

[0053] 200: Data Learning Department

[0054] 210: Classification Department of Training Dataset

[0055] 220: Training Dataset Preprocessing Department

[0056] 230: Generative AI Model Learning Department (Diffusion Probability Model)

[0057] 240: Cervical Image Generation Model

[0058] 300: Reasoning Department

[0059] 301: Inference Generation Technology Selection Department

[0060] 302: Selection Department for Unconditional Image Reasoning Generation Technology

[0061] 303: Image Conversion Technology Inference and Selection Department

[0062] 304: Conditional Image Reasoning Generation Technology Selection Department

[0063] 310: Inference Input Section (Input Image or Prompt)

[0064] 311: Image Input Unit

[0065] 312: Image Quantity Setting Section

[0066] 320: Inference Preprocessing Department

[0067] 330: Cervical Image Generation Model Inference Department

[0068] 340; Virtual positive cervical image generation and output unit

[0069] 400: Medical personnel interpretation of virtual generated images (positive)

[0070] 500: Database used for classification learning in an AI-based cervical cancer grading model.

[0071] 600: Relearning Department of Cervical Cancer Grading Model Based on Artificial Intelligence

[0072] 1000: A Cervical Image Generation System Utilizing Generative AI Models Detailed Implementation

[0073] The cervical image generation system and method utilizing a generative AI model according to a preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0074] For ease of explanation and accuracy, the constituent elements are omitted or shown in a simplified manner in the following drawings. The sizes of the constituent elements do not represent actual dimensions. In this specification, the same reference numerals refer to the same constituent elements, and in individual drawings, reference numerals are omitted for the same structures. Furthermore, to avoid unnecessarily obscuring the essence of the invention, detailed descriptions of known functions and configurations are omitted herein.

[0075] The following will refer to Figure 1 The preferred embodiments of the present invention will be described up to Figure 5, as follows: Figure 1 As shown, the cervical image generation system 1000 utilizing a generative artificial intelligence (AI) model of the present invention includes: a data acquisition unit 100, which acquires cervical image data; a data learning unit 200, which learns a generative AI model based on the cervical image data acquired by the data acquisition unit 100; and an inference unit 300, which uses the cervical image generation model derived from the data learning unit 200 to infer the input data to generate a virtual cervical image for output.

[0076] In this invention, when using a generative AI model to generate virtual cervical images, the generative AI model employs a diffusion probability model. Furthermore, this invention utilizes transfer learning techniques to create a cervical image generation model. Although the embodiments of this invention use a diffusion probability model, generative AI models described later, such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), or Autoregressive Models, can also be used; they are not excluded.

[0077] The overall generation system possessing the above-mentioned features of the present invention is summarized as follows:

[0078] Specifically, professional medical personnel use the interpretation unit 120 to judge the cervical images captured by the camera capturing unit 110, determining whether they are negative or positive and marking the lesions, and then store them in the database server storage unit 130. The training data stored in the database server storage unit 130 is classified in the training dataset classification unit 210 and preprocessed in the training dataset preprocessing unit 220. Based on the preprocessed dataset, a diffusion probability model is learned in the generative AI model learning unit 230. Subsequently, an inference process is performed in the inference unit 300, which receives data input about the input image or prompt words from the inference input unit 310, and then preprocesses it as inference data in the preprocessing unit 320. The preprocessed inference data is input to the learned cervical image generation model inference unit 330, and a virtual image is generated through the output unit 300.

[0079] More specifically, see Figure 1 To explain, firstly, the image data acquisition unit 100 uses the camera capturing unit 110 to capture and acquire an image of the patient's cervix. Professional medical personnel receive the acquired cervical image through various channels such as communication methods, and the cervical image interpretation unit 120 determines whether the acquired cervical image is negative or positive. If positive, the lesion is marked. Then, the positive images marked with lesions in the interpretation results are stored in the database server storage unit 130.

[0080] As described above, the data acquisition unit 100 according to a preferred embodiment of the present invention may include: a camera capturing unit 110 for capturing cervical images; a cervical image interpretation unit 120 for medical personnel to interpret the captured cervical images; and a database server storage unit 130 for storing the interpreted image data through communication means.

[0081] Furthermore, according to a preferred embodiment of the present invention, such as Figure 1 As shown, the data learning unit 200 learns based on the cervical image data acquired by the data acquisition unit 100. That is, the data learning unit 200 may include: a training dataset classification unit 210, which classifies the image data received by the database server storage unit 130 into a training dataset; a training dataset preprocessing unit 220, which preprocesses the training dataset; and a diffusion probability model learning unit 230, which learns a generative AI model (diffusion probability model) based on the preprocessed data.

[0082] In the preferred embodiments of the present invention described above, see Figure 2 The technique for generating positive cervical images is explained.

[0083] In other words, such as Figure 2As shown, the present invention mainly provides three cervical image generation techniques 131. (1) Generating a desired number of virtual cervical images through unconditional image generation technique 132; (2) Generating virtual cervical images through image conversion technique 133, such as converting negative cervical images into positive cervical images through cross-domain conversion; (3) Generating conditional cervical images according to different diseases and conditions through conditional image generation technique 134.

[0084] In other words, as described above, cervical images are captured using a cervical magnification imaging device (camera), then interpreted by experts, and stored in a database along with characteristics such as whether the image is negative or positive, and whether lesions are present. The images stored in the database are then preprocessed using the three techniques described earlier to create training data suitable for different purposes. Subsequently, the preprocessed cervical image dataset is used as training data input into a diffusion probability model, a type of generative AI model, for learning. The input image or prompt is preprocessed according to the learned cervical image generation model and used as input to the model. Then, a virtual cervical image is generated through the inference process of the inference unit.

[0085] Furthermore, according to a preferred embodiment of the present invention, such as Figure 1 As shown, in order to generate a virtual positive cervical image using the cervical image generation model 240 learned in the data learning unit 200, the input data can be transmitted to the inference unit 300 to generate the virtual positive cervical image. That is, the inference unit 300 may include an inference input unit 310; an inference preprocessing unit 320 that preprocesses the input data; a cervical image generation model inference unit 330 that applies the preprocessed data to the cervical image generation model learned by the diffusion probability model; and an output unit 340 that outputs the virtual cervical image generated therefrom.

[0086] Furthermore, according to a preferred embodiment of the present invention, such as Figure 4 As shown, the system may include a database 500 for learning an AI-based cervical cancer grading classification model. This database stores virtual positive cervical images generated by the inference unit 300 after they have been interpreted by medical personnel. In other words, if the generated virtual positive cervical image is assessed by professional medical personnel as being at the same level as a real positive image, the image data can be stored in the AI-based cervical cancer grading classification model learning database 500. This improves the performance of the AI-based cervical cancer grading classification model, allowing it to be more effectively applied in practical AI-based cervical cancer screening service systems.

[0087] Next, a method for generating cervical images using a generative AI model according to an embodiment of the present invention will be described.

[0088] The cervical image generation method using a generative AI model according to the present invention includes: a data acquisition step S100, acquiring cervical image data; a data learning step S200, learning a generative AI model based on the cervical image data acquired in the data acquisition step; and an inference step S300, using the cervical image generation model derived from the data learning unit 200 to infer the input data to generate a virtual cervical image for output.

[0089] The data acquisition step S100 includes: a capture acquisition step S110, in which a cervical image is captured using a camera capture unit 110; a cervical image interpretation step S120, in which medical personnel interpret the captured cervical image; and a database server storage step S130, in which the interpreted image data is received and stored via communication means.

[0090] The data learning step S200 includes: a training dataset classification step S210, which receives image data from the database server storage unit 130 and classifies it into a training dataset; a training dataset preprocessing step S220, which preprocesses the training dataset; and a diffusion probability model learning step S230, which learns the diffusion probability model based on the preprocessed data.

[0091] Furthermore, the reasoning step S300 includes: a reasoning input step S310, in which an input image or prompt word is input into the reasoning input unit 310; a reasoning preprocessing step S320, in which the input data is preprocessed; a cervical image generation model reasoning step S330, in which the preprocessed data is applied to a cervical image generation model learned by the diffusion probability model; and an output step S340, in which the output unit 340 outputs the virtual cervical image generated in the generation model reasoning step S330.

[0092] According to a preferred embodiment of the present invention, the system may further include a judgment step S400, in which a medical professional judges whether the virtual positive cervical image generated and output in the inference unit 300 is a real positive image; and a storage step S500, in which if the judgment step S400 determines that the image is the same as a real positive image, the image data is stored in a database 500 for learning a cervical cancer grading classification model based on artificial intelligence.

[0093] According to a preferred embodiment of the present invention, the method further includes: a relearning step S600, which involves relearning the artificial intelligence-based cervical cancer grading model based on negative and positive image data stored in the artificial intelligence-based cervical cancer grading model learning database 500.

[0094] Embodiment 1 of the present invention as described above is as follows: Figures 2 to 4 The first example selected image conversion technology 133, which generates a positive image based on a negative cervical image; the second example selected conditional image generation technology 134, which generates a virtual positive cervical image based on a positive cervical image.

[0095] See Figure 4 Examples 1 and 2 will be described as follows: Figure 4 As shown, the upper process corresponding to Embodiment 1 selects image conversion technology 133 to classify negative cervical images in the classification unit 210, thereby performing the image generation process; while the lower process corresponding to Embodiment 2 selects conditional image generation technology 134 to classify positive cervical images in the classification unit 210, thereby performing the image generation process.

[0096] More specifically, see Figure 4 The image conversion technology 133 of Embodiment 1 requires input in three pairs when constructing the training dataset. Specifically, the image before conversion is designated as a negative cervical image; the prompt word is designated as "positive cervix"; and the converted image is designated as a positive cervical image. After preprocessing in the training dataset preprocessing unit 220 of the image conversion technology 133, it is learned by the generative AI model learning unit 230. After completing the diffusion probability model learning, the image before conversion is designated as a negative image and the prompt word is designated as "positive cervix" to input the inference data through the inference input unit 310. After input, the inference data is preprocessed in the inference preprocessing unit 320. Using the preprocessed inference data as input to the diffusion probability model, inference is performed in the cervical image generation model inference unit 330, and a virtual positive cervical image can be generated and output by the output unit 340.

[0097] also, Figure 4The conditional image generation technology 134 of Embodiment 2 acquires positive images from the database server storage unit 130 and classifies them to be used as images in the training dataset, specifying the prompt word as "positive cervix". After preprocessing them into training data in the training dataset preprocessing unit 220, a diffusion probability model is learned by the generative AI model learning unit 230. After completing the learning of the diffusion probability model, in order to input inference data, the prompt word "positive cervix" is specified by the inference input unit 310, and the number of images to be generated is specified for input. Then, the inference data is preprocessed in the inference preprocessing unit 320. The preprocessed inference data is used as input to the diffusion probability model in the cervical image generation model inference unit 330, and a virtual positive cervical image can be generated and output by the output unit 340.

[0098] Furthermore, after completing the processes of Embodiments 1 and 2 above, in the interpretation step S400, medical personnel use the interpretation unit 400 to determine whether the image possesses the characteristics of a true positive image. This interpretation step S400 includes the interpretation process of professional medical personnel. If the image possesses the characteristics of a positive image, the image data is stored in the artificial intelligence-based cervical cancer grading classification model learning database S500. Based on the stored positive images, and through the relearning step S600 of the artificial intelligence-based cervical cancer grading classification model relearning unit 600, performance can be improved.

[0099] Meanwhile, in this invention, the "unconditional image generation technology" is not limited to a single embodiment, but is only used as explanatory material to help understand the invention.

[0100] also, Figure 5a An example photograph illustrating the generation of a virtual positive image according to Embodiment 1 of the present invention is shown. Figure 5b An example photograph illustrating the generation of a virtual positive image according to Embodiment 2 of the present invention is shown.

[0101] Below, to aid in understanding the preferred embodiments of the present invention described above, we will focus on detailing diffusion probability models and transfer learning among the generative AI model types related to the present invention, and then briefly describe other types of generative AI models.

[0102] <Explanation of Generative AI Models>

[0103] Machine learning models can be broadly categorized into classification models and generative models based on the functions they perform. Classification models typically involve the task of transforming high-dimensional input data into low-dimensional labeled data, while generative models involve the task of generating high-dimensional data from relatively low-dimensional data.

[0104] The generative AI models that generate the images include Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), Autoregressive Models, and Diffusion Models, which will be briefly described below.

[0105] First, the generative AI model used in this invention utilizes a diffusion probability model and learns through transfer learning.

[0106] <Explanation of the Diffusion Probabilistic Model>

[0107] The idea of ​​the diffusion probability model originates from Langevin dynamics, which describes the phenomenon that molecules gathered in space are uniformly distributed throughout the space through diffusion. The following diagram illustrates this.

[0108]

[0109] The motion of the individual molecules follows a Gaussian distribution, which can be represented as a normal distribution following the mean and standard deviation. If the motion of each molecule in each time interval can be calculated, the diffusion process can be reversed.

[0110]

[0111] By applying this method directly to the pixel values ​​of an image, just as the motion of molecules is regarded as noise following a normal distribution, noise values ​​following a normal distribution can be added to the pixel values ​​of the image.

[0112]

[0113] x0 is the original image. The image generated by adding noise to all pixel values ​​of this original image is x1. Using the same method as the molecular diffusion process, noise is added to each pixel value in the image every time interval t. After T iterations, the resulting image is a completely noisy image. Taking molecules as an example, in the diffusion process, x0 represents the state of aggregation, x... T This represents a state of completely uniform distribution. Furthermore, as mentioned earlier, this process can be reversed if the noise added at each time interval can be calculated. For molecules, it is possible to return from a state of uniform distribution throughout space to a state of aggregation in one space.

[0114]

[0115] For the image, if the noise added to the image pixel values ​​every t time interval can be calculated, then the noisy image x can be transformed. T Restore the input image x0. This can be illustrated by the following formula.

[0116]

[0117] The input image corresponds to the rightmost x0, and the noise image corresponds to x. T The process is divided into forward transformation and reverse transformation. Forward transformation converts the input image into a noisy image, while reverse transformation returns the input data from the noisy image. Therefore, by using the method of generating input data from a noisy image, the reverse transformation step can be parameterized in the diffusion probability model and learned using a deep learning model. This allows the input image to be generated from the noisy image through the learned reverse transformation. Thus, the diffusion probability model can be called a parameterized Markov chain, which is learned to generate an image after a finite amount of time. It possesses the Markov property that the current state (t) depends on the previous state (t-1). The goal of the diffusion probability model is to ensure that the noisy image, after undergoing forward transformation and then reverse transformation, has a probability distribution similar to the input image. To achieve this, the learning process involves updating the mean and standard deviation, which are parameters of the noise generation probability distribution, in the reverse transformation step. Furthermore, in the Gaussian probability distribution of the actual noise generation process, different values ​​are sampled each time, so the generated image will continuously change. And the image generated by the reverse transformation will change according to the learned data.

[0118] <Instructions for learning about diffusion probability models>

[0119] In this invention, the generative AI model uses a neural network similar to neurons in the human brain to learn patterns and features from existing data. The model can then generate new data that matches the learned patterns and features. For example, if the generative AI model is trained on a set of cervical images, it can generate virtual cervical images that resemble the trained cervical images.

[0120] As described in Embodiments 1 and 2 above, the generative AI model of the present invention provides three techniques: unconditional image generation technique 132, conditional image generation technique 133, and image transformation technique 134. That is, although the present invention does not present unconditional image generation technique 132 as a separate embodiment, unconditional image generation technique 132 only uses a set of cervical images to generate virtual cervical images. Furthermore, the image transformation technique 134 used in Embodiment 1 of the present invention generates cervical images through methods such as style transfer, image restoration, cross-domain transformation, and positional transformation. Additionally, the conditional image generation technique 133 used in Embodiment 2 of the present invention generates cervical images based on specific cervical features.

[0121] The difference between the unconditional image generation technology 132 and the conditional image generation technology 133 is as follows: For example, the unconditional image generation technology 132 can generate images without data, thus generating a large number of cervical images; while the conditional image generation technology 133 can generate cervical images for specific situations, such as positive cervical images, negative cervical images, and cervical images where white lesions are clearly visible.

[0122] <Notes on transfer learning>

[0123] Furthermore, this invention employs transfer learning technology to learn a diffusion probability model. Transfer learning refers to applying a model learned for a specific purpose to other tasks. It is a technique that utilizes model knowledge learned from different datasets to handle new tasks or datasets. Its advantage lies in achieving high performance even with small datasets and creating models specifically tailored to the task and domain to be performed. Transfer learning consists of two phases: a pre-learning phase and a fine-tuning phase. In the pre-learning phase, the model learns general features using a large dataset. For example, when learning an image recognition model, a large dataset like ImageNet is used to learn general features of various images (e.g., edges, textures, etc.). In the fine-tuning phase, the pre-learned model is adjusted to suit the specific task. This is a process of adjusting the model to fit a specific domain using a small dataset. That is, adjusting the pre-learned model to be suitable for generating cervical images. In this process, high performance is achieved with minimal data and resources by using a cervical dataset and relearning the last few layers of the model.

[0124] <Detailed instructions on learning diffusion probability models>

[0125] like Figure 2 As shown, the first unconditional image generation technique 132 acquires cervical images from the database server storage unit 130 and preprocesses them into training data. The diffusion probability model learning unit 230 learns the basic algorithm of the diffusion probability model based on the preprocessed cervical image set.

[0126] In addition, such as Figure 2 As shown, the second image conversion technique 133 constructs the three elements—the image before conversion, the conditional prompts, and the converted image—into pairs and preprocesses them into training data. The diffusion probability model learning unit 230 learns the diffusion probability model based on the preprocessed dataset using an image-to-image algorithm.

[0127] In addition, such as Figure 2 As shown, the third conditional image generation technique 134 uses conditions containing specific scenes or cases as prompt words, then constructs pairs of these two elements—the prompt words and images that meet the conditions—and preprocesses them into training data. The diffusion probability model learning unit 230 learns a diffusion probability model based on the preprocessed dataset using a text-to-image algorithm.

[0128] <Detailed Explanation of the Reasoning for the Diffusion Probability Model>

[0129] like Figure 3 As shown, the first unconditional image generation technique 302 can specify the number of images to be generated in a diffusion probability model based on cervical image set learning, thereby generating virtual cervical images.

[0130] Furthermore, the second image conversion technique 303 is a technique for converting an input image into another image with specific attributes or styles. It generates a virtual image based on the input image 311 and the prompt word 310. For example, it converts a negative cervical image into a virtual positive cervical image based on the prompt word "positive cervix." Examples of conversion techniques include style conversion, image restoration, cross-domain conversion, and position conversion. Style conversion converts a given image into a specific style. Specific styles include ethnicity, doctor, etc. If cervical features differ by ethnicity, or if the cervical imaging method differs by doctor, a style conversion technique can be used to generate a cervical image. Image restoration restores damaged or incomplete images. If a cervical image captured using a camera is noisy or damaged, it can be converted into a complete image. Cross-domain conversion refers to converting an image from one domain to another. For example, converting a negative cervical image into a positive cervical image, or vice versa. Position conversion changes the position or appearance of a specific object. For example, converting an image where cervical deformity bands are not visible into an image where they are visible. Alternatively, convert an image where the external cervical os is not visible to an image where the external cervical os is visible.

[0131] Furthermore, a third conditional image generation technique 304 can generate virtual negative cervical images by learning a diffusion probability model based on negative cervical images. Alternatively, it can generate virtual positive cervical images by learning a diffusion probability model based on positive cervical images. This generation technique is called conditional image generation. The condition is called a cue word 310, which is given in text form. For example, in the conditional image generation model, if "cervical negative" is entered as the cue word and the number of images to be generated 312 is specified as N, then N negative cervical images can be generated. Further, various images can be generated based on all visual features of the cervix (e.g., it can learn to generate positive cervical images, cervical images containing specific lesions, cervical images where the external os is not visible, etc.).

[0132] As described above, virtual cervical images can be generated in three main ways. For each technology's input method, unconditional image generation technology 302 uses the number of images to be generated 312; image conversion technology 303 uses the image before conversion 311 and a prompt word 310; and conditional image generation technology 304 can use the prompt word 310 and the number of images to be generated 312 as input.

[0133] This paper will also explain generative adversarial networks (GANs), variational autoencoders (VAEs), and autoregressive models in generative AI models.

[0134] <Explanation of Generative Adversarial Networks (GANs)>

[0135] Generative Adversarial Networks (GANs) are models consisting of two neural networks: a generator and a discriminator. These models generate data by having the generator and discriminator compete against each other. The generator creates new data, while the discriminator determines whether the data is true or false.

[0136]

[0137] The generator receives the input data z and learns to generate data similar to real data, while the discriminator learns to distinguish between real data and fake data generated by the generator. The generator learns the distribution of the input data and reproduces that distribution, making it indistinguishable from the original data distribution. The discriminator distinguishes between real and fake data and estimates the probability of each. The ultimate goal of a Generative Adversarial Network (GAN) is to generate data that closely approximates the distribution of real data. The generator strives to generate fake data that is highly similar to real data, making it impossible for the discriminator to identify it as fake. Through this process, the performance of both the generator and the discriminator is gradually improved, eventually rendering the discriminator unable to distinguish between real and fake data.

[0138] <Instructions on Variational Autoencoders (VAE)>

[0139] The variational autoencoder is a generative AI model that learns a compressed representation of training data into a probability distribution. It generates new sample data by creating variant forms of the learned representation. The basic concept of an autoencoder is as follows: the process of reasoning from raw data into latent variables suitable for a specific purpose is called encoding, and the process of restoring the encoded data back to its original form is called decoding. A neural network consisting of an encoder and a decoder is called an autoencoder. Most autoencoders learn discrete latent space models; in contrast, variational autoencoders learn continuous latent variable models. Instead of modeling a single encoded vector about the latent space, variational autoencoders model two different vectors: the mean vector and the standard deviation vector.

[0140]

[0141] The input dog photo above is connected to an encoder network. The encoder compresses the input image and transforms it into a low-dimensional latent space (latent vector). The key here is that the encoder doesn't simply transform it into a single vector, but rather estimates two parameters: the mean and the variance. These parameters define the probability distribution in the latent space. The latent vector is sampled based on the mean and variance obtained from the encoder. Since this is a random vector following a normal distribution, random sampling is incorporated into the process. It is precisely because of this randomness that the variational autoencoder can function as a generative AI model rather than a simple compression model. The sampled latent vector is passed to a decoder network to reconstruct an image similar to the original image. The decoder receives the latent vector and again reconstructs it into high-dimensional image data. Ultimately, the decoder's output is very similar to the original input image. However, because the sampling of the latent vector involves randomness, it is not exactly the same as the original image.

[0142] <Explanation of Autoregressive Models>

[0143] The term autoregressive refers to a model that regresses on itself. This means predicting the next pixel value based on the previous pixel value. When generating an image, an autoregressive model goes through a process of setting up the pixel sequence and then making predictions based on previous pixels.

[0144]

[0145] The image is represented in a two-dimensional arrangement. The autoregressive model unfolds the two-dimensional image into a one-dimensional sequence and processes each pixel sequentially. After generating the first pixel, the model predicts and generates the second pixel based on that pixel. This process involves calculating the next pixel value using information from previous pixels. To do this, the model estimates the probability distribution of the next pixel using information from previously generated pixels and samples the new pixel value according to the distribution. This process is repeated to generate all pixels to produce the final image. Alternatively, recurrent neural networks or convolutional neural networks can be used to generate images on a pixel-by-pixel basis.

[0146] According to the present invention as described above, the data imbalance problem between negative and positive image categories required for constructing an AI-based cervical cancer grading classification model can be overcome, thereby effectively performing AI-based cervical cancer screening services. Furthermore, according to the present invention, for negative and positive image data, by using a diffusion probability model and transfer learning techniques in a generative AI model to generate virtual positive cervical images and then relearning, the performance of the AI-based cervical cancer grading classification model can be continuously improved.

[0147] Although preferred embodiments of the invention have been described with reference to the accompanying drawings, the embodiments described in this specification and the configurations shown in the drawings are merely preferred embodiments of the invention and do not represent all the technical ideas of the invention. Therefore, it should be understood that various equivalent solutions and modifications may exist to replace the above embodiments at the time of filing this application. Thus, the described embodiments should be understood as exemplary in all respects, not limiting, and the scope of the invention is defined by the claims listed below, not by the specific description, and modifications or variations in the meaning and scope of the claims and their equivalents should be interpreted as being included within the scope of the invention.

Claims

1. A cervical image generation system utilizing a generative AI model, characterized in that: include: The data acquisition department collects cervical image data. The data learning unit, based on cervical image data acquired by the data acquisition unit, learns a generative AI model; and The inference unit uses a cervical image generation model derived from the data learning unit to infer the input data, generate a virtual cervical image, and output it.

2. The cervical image generation system utilizing a generative AI model according to claim 1, characterized in that, The generative AI model is a diffusion probability model.

3. The cervical image generation system utilizing a generative AI model according to claim 2, characterized in that, The data acquisition unit includes: The camera unit captures images of the cervix to obtain data. A cervical image interpretation unit, which allows medical personnel to interpret the captured cervical images; and The database server storage unit receives and stores the interpreted image data via communication means.

4. The cervical image generation system utilizing a generative AI model according to claim 2, characterized in that, The data learning unit includes: The training dataset classification unit classifies the image data received from the database server storage unit into a training dataset. A training dataset preprocessing unit, which preprocesses the training dataset; and The diffusion probability model learning unit learns the diffusion probability model based on the preprocessed data.

5. The cervical image generation system utilizing a generative AI model according to claim 2, characterized in that, The reasoning unit includes: The inference input unit receives input images or prompts. The inference preprocessing unit preprocesses the input data; The cervical image generation model inference unit applies the preprocessed data to the cervical image generation model learned by the diffusion probability model; and The output unit outputs a virtual cervical image generated by the generative model inference unit.

6. The cervical image generation system utilizing a generative AI model according to any one of claims 1 to 5, characterized in that, Also includes: The database for learning the cervical cancer grading model based on artificial intelligence stores virtual positive cervical images generated and output by the inference unit after they have been interpreted by medical personnel.

7. A method for generating cervical images using a generative AI model, characterized in that, include: The data acquisition steps involve obtaining cervical image data for data collection. The data learning step involves learning a generative AI model based on the cervical image data acquired in the data acquisition step; and The reasoning step uses the cervical image generation model derived in the data learning step to reason about the input data in order to generate a virtual cervical image for output.

8. The cervical image generation method using a generative AI model according to claim 7, characterized in that, The generative AI model is a diffusion probability model.

9. The cervical image generation method using a generative AI model according to claim 8, characterized in that, The data acquisition steps include: The imaging and acquisition steps involve using a camera to capture images of the cervix. The cervical image interpretation process involves medical personnel interpreting the captured cervical images; and The database server storage step involves receiving and storing the interpreted image data via communication means.

10. The cervical image generation method using a generative AI model according to claim 8, characterized in that, The data learning steps include: The training dataset classification step involves receiving image data from the database server storage and classifying it into a training dataset. The training dataset preprocessing step involves preprocessing the training dataset; and The diffusion probability model learning step involves learning the diffusion probability model based on the preprocessed data.

11. The cervical image generation method using a generative AI model according to claim 8, characterized in that, The reasoning steps include: The reasoning input step involves inputting the input image or prompt words into the reasoning input unit; The inference preprocessing step involves preprocessing the input data. The cervical image generation model inference step involves applying the preprocessed data to the cervical image generation model learned by the diffusion probability model; and The output step involves the output unit outputting the virtual cervical image generated in the generative model inference step.

12. The method for generating cervical images using a generative AI model according to any one of claims 7 to 11, characterized in that, Also includes: The interpretation step involves medical personnel interpreting the virtual positive cervical image generated and output in the reasoning step to determine whether it is a real positive image; and In the storage step, if the image is determined to be the same as a real positive image in the judgment step, the image data is stored in the database for learning the cervical cancer grading classification model based on artificial intelligence.

13. The cervical image generation method using a generative AI model according to claim 12, characterized in that, Also includes: The relearning step involves using negative and positive image data stored in the AI-based cervical cancer grading model learning database as a basis to relearn the AI-based cervical cancer grading model.

Images