Recognition of realistic synthetic images generated using generative adversarial networks based on deep neural networks

Abstract

The present invention is entitled "Deep Neural Network-Based Recognition of Realistic Synthetic Images Generated Using Generative Adversarial Networks". The present invention provides a technique for deep neural network (DNN) recognition of realistic synthetic images generated using generative adversarial networks (GANs). According to one embodiment, a system is described that may include a memory storing computer executable components and a processor executing the computer executable components stored in the memory. The computer executable components may include a first extraction component that extracts a subset of synthetic images that are classified as non-real rather than real, wherein the subset of synthetic images is generated using a GAN model. The computer executable component may also include a training component that uses a subset of synthetic images and real images to train a DNN model to classify synthetic images generated using a GAN model as real or non-real.

Description

Technical Field

[0001] The present patent application relates generally to artificial intelligence image analysis, and more specifically to computer-implemented techniques for employing a deep neural network (DNN) to recognize realistic synthetic images generated using a generative adversarial network (GAN). Summary of the invention

[0002] The following presents a summary of the invention to provide a basic understanding of one or more embodiments of the invention. The summary of the invention is not intended to identify key or important elements, nor is it intended to delineate any scope of a specific embodiment or any scope of the claims. Its sole purpose is to present concepts in a simplified form as a prelude to a more detailed description presented later. In one or more embodiments described herein, a system, computer-implemented method, apparatus, and / or computer program product that employs a DNN to recognize realistic synthetic images generated using a GAN is provided.

[0003] According to one embodiment, a system may include a memory storing computer executable components and a processor executing the computer executable components stored in the memory. The computer executable components may include a first extraction component that extracts a subset of synthetic images classified as non-real rather than real, wherein the subset of synthetic images is generated using a GAN model. The computer executable component may also include a training component that uses a subset of synthetic images and real images to train a DNN network model to classify synthetic images generated using the GAN model as real or non-real. In one or more embodiments, the DNN includes a very deep convolutional neural network.

[0004] In various implementations, the computer executable components further include: a composite image generator that generates a first composite image using a GAN model, and an annotation component that marks the first composite image as real or not real. With respect to these implementations, the first extraction component extracts a subset of composite images from the first composite image in response to marking a corresponding composite image included in the subset as not real. In some embodiments, the annotation component marks the first composite image based on receiving manual input that classifies the first composite image as real or not real.

[0005] In one or more embodiments, the synthetic image generator also uses the GAN model to generate a second synthetic image, and the computer executable component also includes an inference component that uses the DNN model to classify the second synthetic image as real or non-real. With respect to these embodiments, the subset of synthetic images includes a first subset of synthetic images, and the computer executable component also includes a second extraction component that extracts a second subset of synthetic images classified by the inference component as real rather than non-real from the second synthetic image. In some specific implementations, the computer executable component also includes a target model training component that uses the second subset of synthetic images to train a machine learning model to perform artificial intelligence analysis on the real image. For example, in various embodiments, the first synthetic image, the second synthetic image, and the real image include medical images, and the target model training component uses the second subset of synthetic images to train the machine learning model to diagnose the medical conditions reflected in the real image.

[0006] In some embodiments, elements described in connection with the system may be embodied in a different form, such as a computer-implemented method, a computer program product, or another form. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 A block diagram of an exemplary non-limiting system that facilitates developing a deep neural network (DNN) model that facilitates recognizing realistic synthetic images generated using a GAN is shown in accordance with one or more embodiments of the disclosed subject matter.

[0008] Figure 2 Exemplary realistic and non-realistic synthetic medical images generated using a GAN model according to one or more embodiments of the disclosed subject matter are shown.

[0009] Figure 3 A flow chart of an exemplary non-limiting process for developing a DNN model that facilitates recognition of realistic synthetic images generated using a GAN, according to one or more embodiments of the disclosed subject matter, is shown.

[0010] Figure 4 A block diagram of an exemplary non-limiting system that facilitates employing a DNN model to recognize realistic synthetic images generated using a GAN model is shown in accordance with one or more embodiments of the disclosed subject matter.

[0011] Figure 5 A flow chart illustrating an exemplary non-limiting process for employing a DNN model to recognize realistic synthetic images generated using a GAN model, according to one or more embodiments of the disclosed subject matter.

[0012] Figure 6A block diagram of an exemplary non-limiting system that facilitates using realistic synthetic images to train a machine learning model to make inferences about real images is shown in accordance with one or more embodiments of the disclosed subject matter.

[0013] Figure 7 A flow chart of an exemplary non-limiting computer-implemented method for developing a DNN model that facilitates recognition of realistic synthetic images generated using a GAN in accordance with one or more embodiments of the disclosed subject matter is provided.

[0014] Figure 8 A flowchart is provided for another exemplary non-limiting computer-implemented method that facilitates evaluating and scoping a data-driven deep learning model in accordance with one or more embodiments of the disclosed subject matter.

[0015] Fig. 9 A flow chart of an exemplary non-limiting computer-implemented method for employing a DNN model to recognize realistic synthetic images generated using a GAN model in accordance with one or more embodiments of the disclosed subject matter is provided.

[0016] Fig.10 A flow chart of an exemplary non-limiting computer-implemented method for training a machine learning model using realistic synthetic images to make inferences about real images is provided in accordance with one or more embodiments of the disclosed subject matter.

[0017] Fig.11 A block diagram of an exemplary non-limiting operating environment is shown in which one or more embodiments described herein may be facilitated. DETAILED DESCRIPTION

[0018] The following specific embodiments are merely exemplary and are not intended to limit the application or use of the embodiments and / or the embodiments. In addition, it is not intended to be bound by any express or implied information set forth in the aforementioned "background technology" or "content of the invention" section or "specific embodiments" section.

[0019] The present disclosure provides systems, computer-implemented methods, apparatus, and / or computer program products that facilitate the use of a deep neural network (DNN) model to automatically identify and extract realistic synthetic images rather than non-realistic synthetic images, which is trained on real images and previously identified synthetic images that are classified as non-real. The synthetic images may include images generated using a generative adversarial network (GAN) model. The identified and extracted realistic synthetic images can be combined with real images to form a comprehensive training data set, which can be used to train a target machine learning model to perform various automated image analysis functions. For example, in various embodiments, the synthetic images may include synthetic medical images generated using a GAN model, and the target machine learning model may include a medical image analysis model that is configured to automatically evaluate and diagnose medical conditions reflected in real (e.g., non-synthetic) medical images.

[0020] In this regard, artificial intelligence (AI) and machine learning (ML) are rapidly developing technology fields that affect many industries. Advances in machine learning techniques (such as deep neural networks (DNNs)) have recently demonstrated impressive performance in various AI fields, sometimes even exceeding that of humans, including computer vision, speech, natural language processing (NPL), bioinformatics, drug design, medical image analysis, etc. These achievements are possible due to significant improvements in computing power and the availability of large-scale annotated data sets, resulting in better inference performance than traditional models. However, one of the fundamental problems of data-driven machine learning methods (such as DNNs) is that the final model inference ability is limited by the scope of the training data used to develop the model. With respect to the field of medical imaging, due to various rules and privacy restrictions associated with accessing and using patient data, it may be difficult to obtain sufficient medical images for model training that provide a comprehensive representation of the target medical condition in different patient groups. Therefore, techniques for generating synthetic medical images have been developed to increase the number and distribution of available training images.

[0021] One technique for generating synthetic medical images involves the use of a generative adversarial network (GAN) model. The GAN model is a powerful latent variable model that can be used to learn complex real-world data distributions. In computer vision, GAN has become one of the main methods for generating realistic-looking samples after learning the training image data distribution. However, generating high-quality synthetic images is very challenging. For example, in the field of healthcare, each patient and their medical condition is different from the rest of the patients in the group. Many variables play a vital role in the appearance of the image, such as medical history, family history, daily activities, or work environment. During GAN model training, the algorithm's sole responsibility is to mimic the input data distribution so that the output data distribution is similar to the input data distribution.

[0022] Different types of GAN models have been explored to increase the realistic quality and resolution of the output synthetic images. However, GANs can be difficult to train in practice, and despite improvements, it has been observed that optimization does not always lead to convergence. A common failure mode involves the generator crashing, producing only a single sample or a small series of very similar samples. Although some GAN models have demonstrated local convergence for training for absolutely continuous data and generator distributions, even if convergence is achieved, the generator may still produce unrealistic synthetic images. These unrealistic synthetic images can contaminate the training dataset and significantly reduce the performance of the final machine learning model trained on it.

[0023] The disclosed subject matter provides techniques for automatically identifying and removing these unrealistic-looking GAN model-generated synthetic images from a training data set, particularly in the context of a GAN model that has demonstrated convergence. In one or more embodiments, the disclosed techniques involve training a DNN to filter out unrealistic-looking synthetic images. In some embodiments, the DNN model may include a very complex convolutional DNN model that is a binary classifier. The trained model may provide for classifying synthetic images as real images (e.g., that look realistic) or unreal images (e.g., that look unrealistic). For example, when applied to medical images, the trained DNN model may mimic a radiologist who classifies realistic-looking and unrealistic synthetic medical images. In some embodiments where a DNN model is trained to classify synthetic medical images in this manner, the DNN model is referred to herein as a virtual radiologist network (VRN).

[0024] This problem can be modeled as a fine-grained classification problem. In this regard, the DNN model can be trained using real (e.g., non-synthetic) images and non-real synthetic images that have been previously classified and labeled as non-real. In some embodiments, AI techniques can be used to automatically identify and select non-real synthetic training images from a set of synthetic images generated by a GAN model. For example, one or more supervised, semi-supervised, or unsupervised learning machine learning models can be used to identify and extract non-real synthetic images from real synthetic images. In other embodiments, non-real synthetic training images can be manually labeled. For example, in a specific implementation in which the synthetic images include medical images, radiologists and / or non-domain experts can manually review the synthetic images and label them as real or non-real.

[0025] Once there are enough synthetic training images for non-realistic class training, the DNN model can be trained using real images from the training set of the real image class. The trained DNN model can then be used to separate real and non-realistic synthetic images by running in inference mode. All classified real synthetic images can then be used to augment the training dataset of the clinical model to improve its performance.

[0026] Various embodiments of the disclosed subject matter are directed to filtering realistic synthetic medical images from non-realistic synthetic medical images. These may include various types of synthetic medical images associated with various different capture modalities, including (but not limited to) X-ray, computed tomography (CT), mammography (MG), magnetic resonance imaging (MRI), etc. However, the disclosed techniques are not limited to medical images. In this regard, the disclosed techniques may be applied to automatically classify synthetic images based on substantially any type of GAN model as realistic or non-realistic.

[0027] One or more embodiments are now described with reference to the accompanying drawings, wherein the same reference numerals are used to represent the same elements throughout. In the following description, for the purpose of explanation, many specific details are set forth in order to provide a more thorough understanding of one or more embodiments. However, it is apparent that in various cases, one or more embodiments may be practiced without these specific details.

[0028] Now turning to the attached figure, Figure 1 A block diagram of an exemplary non-limiting system 100 is shown that facilitates the development of a DNN model for recognizing realistic synthetic images generated using a GAN in accordance with one or more embodiments of the disclosed subject matter. Embodiments of the system described herein may include one or more machine-executable components embodied in one or more machines (e.g., embodied in one or more computer-readable storage media associated with one or more machines). Such components, when executed by one or more machines (e.g., processors, computers, computing devices, virtual machines, etc.), may cause the one or more machines to perform the operations described.

[0029] For example, system 100 includes a synthetic image generation module 102 and a DNN development module 112. The synthetic image generation module 102 and the DNN development module 112 can each be and include machine-executable components stored in a memory (not shown) associated with one or more machines (not shown). The memory can be further operably coupled to at least one processor (not shown) so that at least one processor can execute the components (e.g., the synthetic image generation module 102, the DNN development module 112, and components associated therewith) to perform the described operations. Examples of the memory and processors and other suitable computer or computing-based elements can be referenced. Fig.11Find and can be combined to achieve about Figure 1 Or use with one or more systems or components shown and described in other figures disclosed herein.

[0030] System 100 and / or one or more components of system 100 or other systems disclosed herein can be used to solve problems that are highly technical in nature, are not abstract, and cannot be performed by humans as a series of mental behaviors using hardware and / or software. In this regard, system 100 and / or components of system 100 can use a GAN model (e.g., GAN model 108) to generate synthetic images and further train a DNN model (e.g., DNN model 126) to automatically identify and extract realistic synthetic images, rather than synthetic images that appear unrealistic. The real synthetic images selected by these DNN models can be used for further clinical model training. Previous solutions involve requiring domain experts (e.g., radiologists in the context of medical images) or non-domain experts to manually view and separate real or non-real synthetic images based on visual observations. Then, they select real images for clinical model training. However, this is highly labor-intensive and therefore not a scalable solution because it is necessary to view the synthetic images again when generating synthetic images for different clinical model training. Technically, the trained GAN model can generate countable and infinite synthetic images, which can be used to enhance the training dataset for clinical AI model training.

[0031] Therefore, the trained DNN model can be used to appropriately filter these countable infinite synthetic images and select only real images for model training.

[0032] Thus, system 100 and / or components of system 100 (as well as other systems described herein) can facilitate improving the quality and quantity of training data sets by augmenting the training data sets with realistic-looking synthetic images in an efficient, automated manner. Thus, system 100 and / or components of system 100 are used to improve the performance accuracy and specificity of an end-goal model trained using realistic synthetic images. In this regard, system 100 and / or components of system 100 or other systems described herein can be used to solve new problems arising from advances in technology, computer networks, the Internet, and the like, specifically advances in AI solutions in which realistic synthetic images can be used to enhance model training and performance.

[0033] In this regard, the synthetic image generation module 102 may provide for generating a synthetic image. The synthetic image generation module 102 may include a synthetic image generator component 106, a database or data storage device including latent random variable data 104, and a GAN model 108. The synthetic image generator component 106 may be configured to apply the GAN model 108 to the latent random variable data 104 to generate a synthetic image. For example, in the illustrated embodiment, the synthetic image generator component 106 may use the GAN model 108 and the latent random variable data 104 to generate a first synthetic image 110. In various exemplary embodiments, the first synthetic image 110 may include a synthetic medical image.

[0034] A GAN model, such as GAN model 108, may be configured to generate realistic-looking synthetic images after learning a training image data distribution.

[0035] GANs typically involve two neural networks, a generative network and a discriminator network. The generator produces samples (such as synthetic images) from a latent code (e.g., latent random variable data 104). Ideally, the distribution of synthetic images generated from the latent code should be indistinguishable from the training distribution. Since it is usually impossible to design a function that tells what is going on, a discriminator network is trained to evaluate. Because neural networks are distinguishable, gradients can be used to steer the two networks in the right direction. The generator network is of primary interest. The discriminator network is an adaptive loss function that acts as a relatively weak classifier. Once the generator network has been trained, the discriminator network is discarded. In this regard, the GAN model 108 may include a trained generator network that is configured to generate a model of synthetic images.

[0036] During GAN model training, the algorithm's sole responsibility is to mimic the input data distribution so that the output data distribution is similar to the input data distribution. However, GANs can be difficult to train in practice, and despite improvements, it has been observed that optimizations do not always lead to convergence. A common failure mode involves the generator crashing, producing only a single sample or a small series of very similar samples. Although some GAN models have demonstrated local convergence for training for absolutely continuous data and generator distributions, even if convergence is achieved, the generator will likely produce unrealistic synthetic images.

[0037] For example, Figure 2 Exemplary realistic and non-realistic synthetic medical images generated using a GAN model (e.g., GAN model 108) in accordance with one or more embodiments of the disclosed subject matter are shown. Figure 2The synthetic images shown are generated by a converged GAN model. The synthetic images correspond to real chest x-rays, respectively. Image 201 provides an exemplary real-looking synthetic image generated by the GAN model, and images 202-204 show unrealistic synthetic images generated by the same GAN model. By comparison with image 201, it can be seen that image 202 has many irregularities, such as severely damaged lungs, ribs, and heart features.

[0038] Images 203 and 204 may appear very real at first glance, but they also have inevitable problems, such as missing rib parts, two heart features, and so on.

[0039] Reference again Figure 1 , the DNN development module 112 can receive and process the first synthetic image 110 to identify and separate non-real synthetic images from real synthetic images. The DNN development module 112 can also use the non-real synthetic image 120 to train the DNN model 126 to automatically classify the synthetic image as real or non-real. In the illustrated embodiment, the DNN development module 112 can include an annotation component 114, a first extraction component 116, and a training component 122.

[0040] The annotation component 114 can facilitate annotating or labeling the first composite image 110 as real or not real. In some embodiments, the annotation component 114 can label the first composite image as real or not real based on receiving user input that identifies or classifies the corresponding image as real or not real. For example, in some embodiments, the first composite image 110 can be manually viewed and annotated by a domain expert (e.g., a radiologist) or a non-domain expert. Through these embodiments, the annotation component 114 can apply a label (e.g., metadata) or otherwise associate a label with the corresponding first composite image to classify the corresponding first composite image as real or not real based on the manual input.

[0041] In other embodiments, the annotation component 114 can employ AI and one or more machine learning techniques to label the corresponding first composite image 110 as real or not real. With respect to these embodiments, the annotation component 114 can include providing for using one or more machine learning models to automatically (e.g., without human input) classify the first composite image 110 as real or not real.

[0042] For example, the one or more machine learning category models may include one or more AI / machine learning classification models configured to identify distinguishing features associated with the synthetic image that can be used to classify the synthetic image as real or not real. Using these embodiments, the annotation component 114 can employ one or more machine learning classification models to classify the corresponding image of the first synthetic image as real or not real, and then label the corresponding image according to the classification determined by using the one or more machine learning classification models.

[0043] In this regard, one or more machine learning classification models can employ various classification (explicitly and / or implicitly trained) schemes and / or systems (e.g., support vector machines, neural networks, expert systems, Bayesian confidence networks, fuzzy logic, data fusion engines, etc.) combined with classifying synthetic images as real or non-real. For example, one or more machine learning classification models can map an input attribute vector x = (xl, x2, x4, x4, xn) to ensure that the input belongs to a real or non-real category, such as through f(x) = confidence(class). This classification can employ probability and / or statistical-based analysis (e.g., taking into account analysis tools and costs) to predict or infer the actions that the user expects to be automatically performed. A support vector machine (SVM) is an example of a classifier that can be used. An SVM operates by finding a hypersurface in the space of possible inputs, where the hypersurface attempts to separate trigger criteria from non-trigger events. Intuitively, this makes the classification correct for testing data that is close to but different from the training data. Other directed and undirected model classification methods that may be used by the annotation component 114 to classify and label the first composite image 110 as real or not real may include, but are not limited to: Naive Bayes, Bayesian networks, decision trees, neural networks, fuzzy logic models, and probabilistic classification models, which provide different independence patterns that may be employed.

[0044] Classification as used herein also includes statistical regression for developing a priority model. Regardless of the manner in which the first synthetic image 110 is classified and labeled as real or non-real (e.g., using manual annotation input and / or automatically using one or more AI-based classification techniques), the first extraction component 116 can further filter the synthetic images to separate real synthetic images from non-real synthetic images. In various embodiments, the first extraction component can identify and extract all (or one or more in some specific implementations) of the first synthetic images 110 classified / labeled as non-real and collect them into a training dataset. In the illustrated embodiment, the extracted non-real synthetic images are shown in a separate database that includes (only) non-real synthetic images 120. The first extraction component 116 can also identify and extract all (or one or more in some specific implementations) of the first synthetic images classified / labeled as real and add them to the target model dataset 118. For example, the target model dataset can include all "good" images that will be used to train a target model, such as a clinical model designed to diagnose medical conditions reflected in real (non-synthetic) images.

[0045] The training component 122 can also use the non-real synthetic image 120 and the real image 124 to train the DNN model 126 to classify the synthetic image as real or non-real. Unlike the discriminator network used to train the GAN model 108 (which is a weak classifier), the DNN model 126 can include a very deep convolutional neural network as a binary classifier. In this regard, the training component 122 can train the DNN model 126 to distinguish between the non-real synthetic image that uses the non-real synthetic image 120 to represent the "non-real" category and the real image (e.g., non-synthetic) used to represent the real category. For example, in a specific implementation in which the non-real synthetic image is a medical image of a specific type (e.g., a chest x-ray), the real image 124 can include a real medical image of the same specific type. As described in more detail below, once the training of the DNN model 126 is completed, the DNN model 126 can be used to automatically classify the synthetic image generated using the GAN model 108 as real or non-real.

[0046] Figure 3 A flowchart of an exemplary non-limiting process 300 for developing a DNN model (e.g., DNN model 126) that facilitates recognition of realistic synthetic images generated using a GAN is shown according to one or more embodiments of the disclosed subject matter. In one or more embodiments, process 300 provides an exemplary process that can be performed by system 100 to generate DNN model 126. For the sake of brevity, repeated descriptions of similar elements employed in various embodiments are omitted.

[0047] At 302, a first synthetic image 110 may be generated based on the latent random variable data 104 using the GAN model 108 (e.g., via the synthetic image generator component 106). At 304, a corresponding first synthetic image of the first synthetic image 110 may be labeled as real or non-real (e.g., via the annotation component 114). For example, in some embodiments, the corresponding first synthetic image may be manually reviewed and annotated (e.g., by a radiologist or non-domain expert). In other embodiments, one or more machine learning techniques may be applied to automatically classify the synthetic images as real or non-real. At 306, each (or one or more in some implementations) of the first synthetic images may be filtered based on the classification of the first synthetic image as real or non-real. For example, at 306, if the synthetic image is classified / labeled as real, then at 308, the real synthetic image may be extracted (e.g., via the first extraction component 116) and added to the target model dataset 118.

[0048] However, if at 306, the first synthetic image is classified as non-real, then the non-real image synthetic image can be extracted and added to the training data set of non-real image synthetic images 120. Once there are enough non-real synthetic images 120 (e.g., based on a defined minimum amount or another deterministic criterion), then at 312, the DNN model can be trained (e.g., via the training component 122). In this regard, the DNN model 126 can be trained using the real images 124 (e.g., non-synthetic images) and the non-real synthetic images 120 as input to classify the synthetic images as real or non-real.

[0049] Figure 4 A block diagram of an exemplary non-limiting system 400 that facilitates employing a DNN model to recognize realistic synthetic images generated using a GAN model is shown, in accordance with one or more embodiments of the disclosed subject matter.

[0050] System 400 includes at least some components that are the same or similar to system 100. In one or more embodiments, system 100 may include system 400, and vice versa. For the sake of brevity, repeated descriptions of similar elements employed in various embodiments are omitted.

[0051] Similar to system 100, system 400 may include synthetic image generation module 102 and target model dataset 118. System 400 also includes a DNN application module 404, which may include an inference component 406 and a second extraction component 408. DNN application module 404 may be and include machine-executable components stored in a memory (not shown) associated with one or more machines (not shown). The memory may be further operably coupled to at least one processor (not shown) so that at least one processor can execute components (e.g., DNN application module 404 and components associated therewith) to perform the described operations. Examples of the memory and processors and other suitable computer or computing-based elements may refer to Fig.11 Find and can be combined to achieve about Figure 4 Or use with one or more systems or components shown and described in other figures disclosed herein.

[0052] The DNN application module 404 can be configured to apply the DNN model 126 after training and developing it (e.g., by the DNN development module 112) to separate real synthetic images from non-real synthetic images by running in an inference mode. In this regard, after the DNN model 126 has been trained, the synthetic image generation module 102 can use the synthetic image generator component 106, the GAN model 108, and the latent random variable data 104 to generate a second synthetic image. The inference component 406 can further employ the DNN model 126 to classify each (or one or more in some implementations) of the second synthetic images as real or non-real. The second extraction component 408 can further extract the images classified as real (e.g., Figure 4 4. The second extraction component 408 may further extract all (or in some implementations, one or more) of the second synthetic images classified as true synthetic images 410 (as depicted in FIG. 4 ). These true synthetic images may be further added to the target model dataset 118. In some embodiments, the second extraction component 408 may discard the second synthetic images 402 that are classified as non-true.

[0053] Figure 5 A flowchart of an exemplary non-limiting process for employing a DNN model to identify realistic synthetic images generated using a GAN model according to one or more embodiments of the disclosed subject matter is shown. In one or more embodiments, process 500 provides an exemplary process that can be performed by system 400 to generate a target model dataset (e.g., target model dataset 118) that includes realistic synthetic images and excludes non-realistic synthetic images. For the sake of brevity, repeated descriptions of similar elements employed in various embodiments are omitted.

[0054] At 502, the second synthetic images 402 can be generated based on the latent random variable data 104 using the GAN model 108 (e.g., via the synthetic image generator component 106). At 504, using the DNN model 126, the respective second synthetic images 402 can be classified as real or not real (e.g., by the inference component 406). At 506, each (or one or more in some implementations) of the second synthetic images can be filtered based on the classification of the second synthetic images as real or not real. For example, at 506, if the second synthetic image is classified as real, then at 508, the real synthetic image can be extracted (e.g., via the second extraction component 408) and added to the target model dataset 118. However, if at 506, the second synthetic image is classified as not real, then the not real image can be discarded (e.g., added to the trash 512).

[0055] Figure 6 A block diagram of an exemplary non-limiting system 600 is shown that facilitates training a machine learning model using a realistic synthetic image to make inferences about a real image in accordance with one or more embodiments of the disclosed subject matter. System 600 includes the same or similar components as system 400, with the addition of a target model training module 602 and a second real image 608. In one or more embodiments, system 100 may include system 600, and vice versa. For the sake of brevity, repeated descriptions of similar elements employed in various embodiments are omitted.

[0056] The target model training module 602 can be and include machine-executable components stored in a memory (not shown) associated with one or more machines (not shown). The memory can be further operably coupled to at least one processor (not shown) so that at least one processor can execute components (e.g., the target model training module 602 and components associated therewith) to perform the described operations. Examples of the memory and processor and other suitable computer or computing-based elements can be referenced. Fig.11 Find and can be combined to achieve about Figure 6 Or use with one or more systems or components shown and described in other figures disclosed herein.

[0057] The target model training module 602 may include a target model training component 604 and a target model 606. The target model training component 604 is configured to apply the target model dataset 118 including the real synthetic image 410 to train the target model 606. For example, in one embodiment, the target model 606 may include a machine learning model that is configured to recognize and / or diagnose a medical condition reflected in a real medical image (e.g., in a clinical context). With respect to these embodiments, the target model dataset 118 may include a real synthetic medical image, and the target model training component 604 may train the target model 606 to recognize and / or diagnose a medical condition in the real medical image based on a representation of the medical condition reflected in the real synthetic image 410. In some embodiments, the target model dataset 118 also includes real images. In this regard, in scenarios where real medical images of training samples are unavailable or limited, the real synthetic images 410 may be used to increase the size and distribution of the training images. Therefore, the final performance of the (trained) target model 606 may be improved.

[0058] Figure 7 A flowchart of an exemplary non-limiting computer-implemented method 800 according to one or more embodiments of the disclosed subject matter is provided, which facilitates DNN recognition of realistic synthetic images generated using GANs. For the sake of brevity, repeated descriptions of similar elements employed in various embodiments are omitted.

[0059] At 702, a system (e.g., system 100, etc.) operably coupled to a processor may extract a subset of synthetic images (e.g., non-realistic synthetic images 120) that are classified as non-realistic rather than real (e.g., using first extraction component 116), wherein the subset of synthetic images is generated using a GAN model (e.g., GAN model 108) (e.g., by synthetic image generation module 102). At 702, the system may train (e.g., using training component 122) a DNN model (e.g., DNN model 126) with the subset of synthetic images and real images (e.g., real images 124) to classify synthetic images generated using a generative adversarial network model as real or non-realistic.

[0060] Figure 8 A flowchart of another exemplary non-limiting computer-implemented method 800 according to one or more embodiments of the disclosed subject matter is provided, which facilitates DNN recognition of realistic synthetic images generated using GANs. For the sake of brevity, repeated descriptions of similar elements employed in various embodiments are omitted.

[0061] At 802, a system (e.g., system 100, etc.) operably coupled to a processor may generate (e.g., using synthetic image generation module 102) a first synthetic image (e.g., first synthetic image 110) using a GAN model (e.g., GAN model 108). At 804, the system may classify the first synthetic image as real or non-real using artificial intelligence. At 806, the system may extract (e.g., using first extraction component 116) a subset of the first synthetic images (e.g., non-real synthetic images 120) that are classified as non-real and not real (e.g., using first extraction component 116). At 808, the system may train (e.g., using training component 122) a DNN model (e.g., DNN model 126) using the subset of the first synthetic images and a real image (e.g., real image 124) to classify a second synthetic image generated using the GAN model as real or non-real.

[0062] Fig. 9 A flowchart of another exemplary non-limiting computer-implemented method 900 according to one or more embodiments of the disclosed subject matter is provided, which facilitates DNN recognition of realistic synthetic images generated using GANs. For the sake of brevity, repeated descriptions of similar elements employed in various embodiments are omitted.

[0063] At 902, a system (e.g., system 100 and / or system 400) operably coupled to a processor may extract (e.g., using first extraction component 116) a first subset of synthetic images (e.g., non-realistic synthetic images 120) from a first synthetic image (e.g., first synthetic image 110) based on the first subset of synthetic images being classified as non-realistic and not real, wherein the first synthetic image is generated using a GAN model (e.g., GAN model 108). At 904, the system may employ the first subset of synthetic images and real images (e.g., real images 124) to train (e.g., using training component 122) a DNN model (e.g., DNN model 126) to classify a second synthetic image generated using the GAN model as real or non-realistic. At 906, the system may employ the DNN model to classify the second synthetic image as real or non-realistic (e.g., using inference component 406). At 908 , the system can extract (eg, using second extraction component 408 ) a second subset of synthetic images (eg, real synthetic images 410 ) from the second synthetic images based on the second subset of synthetic images being classified as real rather than non-real.

[0064] Fig.10A flowchart of another exemplary non-limiting computer-implemented method 1000 according to one or more embodiments of the disclosed subject matter is provided, which facilitates DNN recognition of realistic synthetic images generated using GANs. For the sake of brevity, repeated descriptions of similar elements employed in various embodiments are omitted.

[0065] At 1002, a system (e.g., system 100, system 400, and / or system 600) operably coupled to a processor may employ a first real image (e.g., real image 124) and a first synthetic image classified as non-real (e.g., non-real synthetic image 120) to train (e.g., using training component 122) a DNN model (e.g., DNN model 126) to classify a second synthetic image as real or non-real, wherein the first synthetic image and the second synthetic image include synthetic images generated using a GAN model (e.g., GAN model 108). At 1004, the system may employ the DNN model to classify the second synthetic image as real or non-real (e.g., using inference component 406). At 1006, the system may extract (e.g., using second extraction component 408) a subset of the second synthetic images classified as real (e.g., real synthetic images 410). At 1008, the system can employ a subset of the second synthetic image to train (e.g., using target model training component 604) a target machine learning model (e.g., target model 606) to perform AI analysis on the second real image.

[0066] One or more embodiments may be systems, methods and / or computer program products at any possible level of technical detail of integration. The computer program product may include a computer-readable storage medium (or multiple media) having computer-readable program instructions thereon for causing a processor to perform aspects of the present invention.

[0067] A computer-readable storage medium may be a tangible device that can hold and store instructions for use by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of computer-readable storage media includes the following items: a portable computer floppy disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device (such as a punch card or a raised structure in a groove on which instructions are recorded), and any suitable combination of the foregoing items. As used herein, a computer-readable storage medium should not be understood to be a transient signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagated by a waveguide or other transmission medium (e.g., a light pulse passing through a fiber optic cable), or an electrical signal transmitted by a wire.

[0068] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to a corresponding computing / processing device, or downloaded to an external computer or external storage device via a network (e.g., the Internet, a local area network, a wide area network, and / or a wireless network). The network can include copper transmission cables, optical transmission fibers, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network, and forwards the computer-readable program instructions for storage in a computer-readable storage medium in the corresponding computing / processing device.

[0069] The computer-readable program instructions for performing the operation of the present invention can be assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, configuration data of integrated circuits, or source code or object code written in any combination of one or more programming languages ​​(including object-oriented programming languages, such as Smalltalk, C++, etc.) and process programming languages ​​(such as "C" programming language or similar programming languages). Computer-readable program instructions can be executed completely on the user's computer, partially on the user's computer, executed as an independent software package, partially on the user's computer and partially on a remote computer, or completely on a remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network (including local area network (LAN) or wide area network (WAN)), or can be connected to an external computer (for example, by using the Internet of an Internet service provider). In some embodiments, the electronic circuit including, for example, a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA) can execute computer-readable program instructions to personalize the electronic circuit by utilizing the state information of the computer-readable program instructions, so as to perform aspects of the present invention.

[0070] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present invention. It is understood that each frame of the flowchart illustration and / or block diagram, and the combination of frames in the flowchart illustration and / or block diagram can be implemented by computer-readable program instructions.

[0071] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device create a device for implementing the function / action specified in one or more boxes of the flow chart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium, which can guide the computer, programmable data processing device, and / or other equipment to function in a specific manner, so that the computer-readable storage medium with instructions stored therein includes an article of manufacture, which includes instructions for implementing various aspects of the function / action specified in one or more boxes of the flow chart and / or block diagram.

[0072] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be executed on the computer, other programmable apparatus, or other device to produce a computer-implemented process, such that the instructions executed on the computer, other programmable apparatus, or other device implement the functions / actions specified in one or more boxes of the flowchart and / or block diagram.

[0073] The flow chart and block diagram in the accompanying drawings illustrate the possible specific implementation architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention.In this respect, each frame in the flow chart or block diagram can represent a module, a fragment or a part of an instruction, which includes one or more executable instructions for realizing a specified logical function.In some alternative specific implementations, the function indicated in the frame may not occur in the order indicated in the figure.For example, depending on the function involved, two frames shown in succession can actually be executed substantially simultaneously, or these frames may be executed in reverse order sometimes.It should also be noted that each frame of the block diagram and / or flow chart illustration and the frame combination in the block diagram and / or flow chart illustration can be implemented by a system based on special-purpose hardware, and these systems perform a specified function or action or perform a combination of special-purpose hardware and computer instructions.

[0074] Combination Fig.11 , the systems and processes described below may be embodied in hardware, such as a single integrated circuit (IC) chip, multiple ICs, application specific integrated circuits (ASICs), etc. In addition, the order in which some or all of the process blocks appear in each process should not be considered limiting. Instead, it should be understood that some process blocks can be performed in a variety of orders, not all of which can be explicitly shown herein.

[0075] refer to Fig.11 , an exemplary environment 1100 for implementing various aspects of the claimed subject matter includes a computer 1102. The computer 1102 includes a processing unit 1104, a system memory 1106, a codec 1135, and a system bus 1108. The system bus 1108 couples system components including, but not limited to, the system memory 1106 to the processing unit 1104. The processing unit 1104 can be any of a variety of available processors. Dual microprocessors and other multi-processor architectures can also be used as the processing unit 1104.

[0076] The system bus 1108 may be any of several types of bus structures including a memory bus or memory controller, a peripheral bus or external bus, or a local bus using a variety of available bus architectures including, but not limited to, Industry Standard Architecture (ISA), Micro Channel Architecture (MSA), Extended ISA (EISA), Intelligent Drive Electronics (IDE), VESA Local Bus (VLB), Peripheral Component Interconnect (PCI), card bus, Universal Serial Bus (USB), Advanced Graphics Port (AGP), Personal Computer Memory Card International Association bus (PCMCIA), FireWire (IEEE 1394), and Small Computer System Interface (SCSI).

[0077] In various embodiments, the system memory 1106 includes a volatile memory 1110 and a non-volatile memory 1112, which may employ one or more of the disclosed memory architectures. A basic input / output system (BIOS), which contains the basic routines for transferring information between elements within the computer 1102, such as during startup, is stored in the non-volatile memory 1112. In addition, according to the present invention, the codec 1135 may include at least one of an encoder or a decoder, wherein at least one of the encoder or the decoder may be composed of hardware, software, or a combination of hardware and software.

[0078] Although the codec 1135 is shown as a separate component, the codec 1135 can be included in the non-volatile memory 1112. By way of example and not limitation, the non-volatile memory 1112 may include a read-only memory (ROM), a programmable ROM (PROM), an electrically programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), a flash memory, a 3D flash memory, or a resistive memory, such as a resistive random access memory (RRAM). In at least some embodiments, the non-volatile memory 1112 can employ one or more disclosed memory devices. In addition, the non-volatile memory 1112 can be a computer memory (e.g., physically integrated with the computer 1102 or its motherboard) or a removable memory. Examples of suitable removable memories that can be used to implement the disclosed embodiments may include a secure digital (SD) card, a compact flash (CF) card, a universal serial bus (USB) memory stick, and the like. The volatile memory 1110 includes a random access memory (RAM) that acts as an external cache memory, and one or more disclosed memory devices may also be employed in various embodiments. By way of example and not limitation, RAM can be provided in a variety of forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and enhanced SDRAM (ESDRAM).

[0079] The computer 1102 may also include removable / non-removable, volatile / nonvolatile computer storage media. Fig.11 For example, a disk storage device 1114 is shown. The disk storage device 1114 includes, but is not limited to, devices such as disk drives, solid-state disks (SSDs), flash memory cards, or memory sticks. In addition, the disk storage device 1114 can include storage media alone or in combination with other storage media, including but not limited to optical disk drives, such as optical disk ROM devices (CD-ROMs), CD recordable drives (CD-R drives), CD rewritable drives (CD-RW drives), or digital versatile disk ROM drives (DVD-ROMs). In order to facilitate the connection of the disk storage device 1114 to the system bus 1108, a removable or non-removable interface, such as an interface 1116, is generally used. It should be understood that the disk storage device 1114 can store information related to the user. Such information can be stored at a server or provided to an application running on a server or user device. In one embodiment, the user can be notified (for example, through one or more output devices 1136) of the type of information stored in the disk storage device 1114 or transmitted to the server or application. The user may be provided an opportunity to opt-in or opt-out of the collection or sharing of such information by the server or application (eg, via input from one or more input devices 1128 ).

[0080] It should be understood that Fig.11 Software is described that acts as an intermediary between a user and the basic computer resources described in a suitable operating environment 1100. Such software includes an operating system 1118. The operating system 1118, which may be stored on disk storage 1114, is used to control and allocate the resources of the computer 1102. Application programs 1120 utilize the operating system 1118 through program modules 1124 to manage resources, as well as program data 1126 stored on system memory 1106 or disk storage 1114, such as boot / shutdown transaction tables, etc. It should be understood that the claimed subject matter can be implemented with various operating systems or combinations of operating systems.

[0081] The user enters commands or information into the computer 1102 through input devices 1128. Input devices 1128 include, but are not limited to, pointing devices such as a mouse, trackball, stylus, touchpad, keyboard, microphone, joystick, gamepad, satellite dish, scanner, TV tuner card, digital camera, digital video camera, webcam, etc. These and other input devices are connected to the processing unit 1104 through the system bus 1108 via interface ports 1130. Interface ports 1130 include, for example, serial ports, parallel ports, game ports, and universal serial bus (USB). Output devices 1136 use some of the same types of ports as input devices 1128. Thus, for example, a USB port can be used to provide input to the computer 1102 and output information from the computer 1102 to output devices 1136. Output adapters 1134 are provided to illustrate that there are some output devices 1136 such as monitors, speakers, and printers, as well as other output devices 1136 that require special adapters. By way of example and not limitation, output adapters 1134 include video and sound cards that provide a means of connection between output devices 1136 and system bus 1108. It should be noted that other devices or systems of devices provide input and output capabilities, such as remote computer 1138.

[0082] The computer 1102 can operate in a networked environment using logical connections to one or more remote computers, such as remote computers 1138. The remote computer 1138 can be a personal computer, server, router, network PC, workstation, microprocessor-based device, peer device, smart phone, tablet computer or other network node, and typically includes many of the elements described relative to the computer 1102. For the purpose of simplicity, only the memory storage device 1140 with the remote computer 1138 is shown. The remote computer 1138 is logically connected to the computer 1102 through a network interface 1142, and then connected via a communication connection 1144.

[0083] The network interface 1142 includes wired or wireless communication networks, such as local area networks (LANs) and wide area networks (WANs) and cellular networks. LAN technologies include fiber distributed data interface (FDDI), copper distributed data interface (CDDI), Ethernet, token ring, etc. WAN technologies include, but are not limited to, point-to-point links, circuit switching networks such as integrated services digital networks (ISDN) and variations thereon, packet switching networks, and digital subscriber lines (DSL).

[0084] The communication connection 1144 refers to the hardware / software used to connect the network interface 1142 to the bus 1108. Although the communication connection 1144 is shown within the computer 1102 for clarity of illustration, it can also be external to the computer 1102. For example purposes only, the hardware / software required to connect to the network interface 1142 includes internal and external technologies, such as modems, including conventional telephone grade modems, cable modems and DSL modems, ISDN adapters, as well as wired and wireless Ethernet cards, hubs and routers.

[0085] Although the subject matter is described above in the general context of computer executable instructions of a computer program product running on one or more computers, it will be appreciated by those skilled in the art that the present disclosure may also or may be implemented in combination with other program modules. Typically, program modules include routines, programs, components, data structures, etc. that perform specific tasks and / or implement specific abstract data types. In addition, it will be appreciated by those skilled in the art that the computer implementation method of the present invention may be practiced with other computer system configurations, including single-processor or multi-processor computer systems, small computing devices, large computers, and computers, handheld computing devices (e.g., PDAs, phones), microprocessor-based or programmable consumer or industrial electronic devices, etc. The illustrated aspects may also be practiced in a distributed computing environment, in which tasks are performed by a remote processing device linked by a communication network. However, some (if not all) aspects of the present disclosure may be practiced on a stand-alone computer. In a distributed computing environment, program modules may be located in local and remote memory storage devices.

[0086] As used in this application, the terms "component", "system", "platform", "interface", etc. may refer to and / or may include computer-related entities, or entities related to an operating machine having one or more specific functions. The entities disclosed herein may be hardware, a combination of hardware and software, software, or software in execution. For example, a component may be, but is not limited to, a process, a processor, an object, an executable file, a thread of execution, a program, and / or a computer running on a processor. By way of example, both an application running on a server and a server may be a component. One or more components may reside in a process and / or a thread of execution, and a component may be located on a computer and / or distributed between two or more computers. In another example, a corresponding component may be executed from various computer-readable media having various data structures stored thereon. A component may communicate via a local and / or remote process, such as according to a signal having one or more data packets (e.g., data from a component interacting with another component in a local system, a distributed system, and / or interacting with other systems via a signal across a network (such as the Internet). As another example, a component may be a device having a specific function provided by a mechanical part operated by an electrical or electronic circuit, which is operated by a software or firmware application executed by a processor. In this case, the processor may be internal or external to the device and may execute at least a portion of a software or firmware application. As yet another example, the component may be a device that provides a specific functionality through electronic components without mechanical parts, where the electronic components may include a processor or other device to execute software or firmware that at least partially gives the electronic components functionality. In one aspect, the component may emulate the electronic component via, for example, a virtual machine within a cloud computing system.

[0087] In addition, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or". That is, unless otherwise specified or clear from the context, "X employs A or B" is intended to mean any natural inclusive permutation. That is, if X employs A; X employs B; or X employs both A and B, then "X employs A or B" is satisfied in any of the foregoing cases. In addition, the articles "one" and "an" used in this specification and the drawings should generally be interpreted to mean "one or more" unless otherwise specified or clear from the context to refer to the singular form. As used herein, the terms "example" and / or "exemplary" are used to indicate use as an example, instance, or illustration, and are intended to be non-limiting. For the avoidance of doubt, the subject matter disclosed herein is not limited to such examples. In addition, any aspect or design described herein as "example" and / or "exemplary" is not necessarily to be construed as preferred or advantageous over other aspects or designs, nor is it meant to exclude equivalent exemplary structures and techniques known to those of ordinary skill in the art.

[0088] As used in this subject specification, the term "processor" may refer to substantially any computing processing unit or device, including but not limited to a single-core processor; a single processor with software multithreaded execution capability; a multi-core processor; a multi-core processor with software multithreaded execution capability; a multi-core processor with hardware multithreading technology; a parallel platform; and a parallel platform with distributed shared memory. In addition, a processor may refer to an integrated circuit, an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a field programmable gate array (FPGA), a programmable logic controller (PLC), a complex programmable logic device (CPLD), discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. In addition, the processor may utilize nanoscale architectures (such as, but not limited to, transistors, switches, and gates based on molecules and quantum dots) in order to optimize space usage or enhance the performance of user devices. The processor may also be implemented as a combination of computing processing units. In the present disclosure, terms such as "storage", "storage device", "data storage", "data storage device", "database", and substantially any other information storage component related to the operation and function of the component are used to refer to a "memory component", an entity embodied in the "memory", or a component including a memory. It should be understood that the memory and / or memory components described herein may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. As an illustration and not limitation, non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable ROM (EEPROM), flash memory, or non-volatile random access memory (RAM) (e.g., ferroelectric RAM (FeRAM)). Volatile memory may include RAM, for example, which may be used as an external cache memory. As an illustration and not limitation, RAM may be in a variety of forms, such as synchronous RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), direct Rambus RAM (DRRAM), direct Rambus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM). In addition, the memory components of the system or computer-implemented method disclosed herein are intended to include, but are not limited to, these and any other suitable types of memory.

[0089] What has been described above includes only examples of systems and computer-implemented methods. Of course, it is not possible to describe every conceivable combination of components or computer-implemented methods for the purposes of describing the present disclosure, but one of ordinary skill in the art will recognize that many further combinations and permutations of the present disclosure are possible. In addition, to the extent that the terms "including," "having," "having," and the like are used in the detailed description, claims, appendices, and drawings, such terms are intended to be inclusive in a manner similar to the term "comprising," as when used as a transitional word in a claim, "comprising" is interpreted. Descriptions of various embodiments have been given for illustrative purposes, but these descriptions are not intended to be exhaustive or limited to the disclosed embodiments.

[0090] Without departing from the scope and spirit of the described embodiments, many modifications and variations will be apparent to those of ordinary skill in the art. The terminology used herein is selected to best explain the principles of the embodiments, practical applications or technical improvements over technologies found on the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A system for identifying realistic synthetic images generated using a generative adversarial network, the system comprising: a memory storing computer executable components; and a processor that executes the computer executable components stored in the memory, wherein the computer executable components include: a first extraction component that extracts a subset of synthetic images classified as non-real rather than real from a first synthetic image, wherein the first synthetic image is generated using a generative adversarial network model; as well as A training component that uses a subset of the synthetic images and real images to train a deep neural network model to classify a second synthetic image generated using the generative adversarial network model as real or non-real.

2. The system of claim 1, wherein the computer executable component further comprises: a synthetic image generator, the synthetic image generator using the generative adversarial network model to generate a first synthetic image; as well as An annotation component that marks the first composite image as authentic or not authentic, wherein the first extraction component extracts the subset of composite images from the first composite image in response to marking the corresponding composite images included in the subset as not authentic. 3 . The system of claim 2 , wherein the annotation component labels the first composite image based on receiving manual input classifying the first composite image as real or not real.

4. The system of claim 2, wherein the synthetic image generator generates a second synthetic image using the generative adversarial network model, and wherein the computer executable components further comprise: An inference component employing the deep neural network model to classify the second synthetic image as real or not real.

5. The system of claim 4, wherein the subset of composite images comprises a first subset of composite images, and wherein the computer executable component further comprises: A second extraction component extracts from the second synthetic image a second subset of synthetic images classified by the inference component as real rather than not real.

6. The system of claim 5, wherein the real image comprises a first real image and wherein the computer executable component further comprises: A target model training component is provided, wherein the target model training component uses a second subset of the synthetic images to train a machine learning model to perform artificial intelligence analysis on a second real image.

7. A system according to claim 6, wherein the first synthetic image, the second synthetic image, the first real image and the second real image include medical images, and wherein the target model training component uses a second subset of the synthetic images to train the machine learning model to diagnose the medical condition reflected in the second real image.

8. The system of claim 6, wherein the real image comprises a first real image and wherein the second extraction component combines the second subset of the synthetic images with the second real image to form a target model training dataset, and wherein the computer executable component further comprises: A target model training component, wherein the target model training component uses the target model training data set to train a machine learning model to perform artificial intelligence analysis on a third real image.

9. The system of claim 8, wherein the first synthetic image, the second synthetic image, the first real image, the second real image, and the third real image comprise medical images, and wherein the target model training component employs the target model training dataset to train the machine learning model to diagnose a medical condition reflected in the third real image.

10. The system of claim 1, wherein the subset of synthetic images and the real images comprise medical images.

11. The system of claim 1 , wherein the computer executable component further comprises: An inference component employing the deep neural network model to classify the synthetic image as real or not real.

12. The system of claim 11, wherein the subset of synthetic images comprises a first subset, wherein the real images comprise a first real image, and wherein the computer executable component further comprises: a second extraction component that extracts from the second synthetic image a second subset of synthetic images classified by the inference component as real rather than not real; as well as A target model training component is provided, wherein the target model training component uses a second subset of the synthetic images to train a machine learning model to perform artificial intelligence analysis on a second real image.

13. A method for identifying realistic synthetic images generated using a generative adversarial network, the method comprising: extracting, by a system operably coupled to a processor, a subset of synthetic images classified as non-real and not real from a first synthetic image, wherein the first synthetic image was generated using a generative adversarial network model; as well as A subset of the synthetic images and real images are employed by the system to train a deep neural network model to classify a second synthetic image generated using the generative adversarial network model as real or not real.

14. The method of claim 13, wherein the deep neural network model comprises a convolutional neural network.

15. The method according to claim 13, further comprising: generating, by the system, a first synthetic image using the generative adversarial network model; as well as The first composite image is labeled, by the system, as authentic or not authentic, wherein the extracting includes extracting a subset of the composite images from the first composite image in response to labeling a corresponding composite image included in the subset as not authentic.

16. The method of claim 15, wherein the labeling is based on receiving manual input classifying the first composite image as real or not real.

17. The method according to claim 15, further comprising: The first composite image is classified by the system as real or not real using artificial intelligence, wherein the labeling is based on the classification.

18. The method according to claim 13, further comprising: The deep neural network model is employed by the system to classify the synthetic image as real or not real.

19. The method of claim 18, wherein the subset of synthetic images comprises a first subset, wherein the real image comprises a first real image, and wherein the method further comprises: extracting, by the system, from the second synthetic image, a second subset of synthetic images classified by the deep neural network model as real rather than not real; as well as A second subset of the synthetic images is employed by the system to train a machine learning model to perform artificial intelligence analysis on a second real image.

20. A machine-readable storage medium comprising executable instructions that, when executed by a processor, facilitate performance of operations comprising: extracting a subset of synthetic images classified as non-real and not real from a first synthetic image, wherein the first synthetic image is generated using a generative adversarial network model; as well as A deep neural network model is trained using the subset of the synthetic images and real images to classify a second synthetic image generated using the generative adversarial network model as real or not real.

Images