Gastroscope celiac disease virtual dyeing auxiliary diagnosis method and device

Abstract

The invention discloses a gastroscope celiac disease virtual staining auxiliary diagnosis method and device, relates to the technical field of medical assistance, and mainly aims to improve the accuracy of disease probability evaluation and improve the accuracy of a diagnosis result. According to the main technical scheme, the gastroscope celiac disease virtual dyeing auxiliary diagnosis method comprises the following steps that firstly, an original white light gastroscope image, collected through gastroscopy, of a target patient is obtained; 2, processing the original white light gastroscope image by using a pre-trained cyclic generative adversarial network model to generate a corresponding virtual indigo carmine dyeing image so as to enhance the visibility of a focus in the image; 3, inputting the original white light gastroscope image and the virtual indigo carmine dyeing image into a pre-trained image classification model for analysis, and outputting the disease probability of the target patient; and 4, according to the disease probability, determining a stomach disease diagnosis result of the target patient.

Description

Technical Field

[0001] The present invention relates to the field of medical assistance technology, and in particular to a virtual staining-assisted diagnosis method and device for celiac disease under gastroscopy. Background Art

[0002] Gastroscopy is one of the important means of diagnosing gastric diseases. Due to the limited contrast and detail of traditional white light gastroscopy images, doctors may face challenges in identifying lesion tissues. In order to improve image quality, endoscopic staining methods such as indigo carmine staining are often used clinically to enhance the visibility of the lesion area. However, the actual staining process requires additional operating time and cost, and may increase patient discomfort and infection risks.

[0003] With the development of deep learning and artificial intelligence technology, virtual staining technology has gradually attracted attention. By using models such as generative adversarial networks, the original white-light gastroscopy images can be converted into images with staining effects, aiming to improve the visualization of the lesion area. However, the existing virtual staining methods still have shortcomings in the quality of generated images and the realism of the staining effect. The generated virtual indigo carmine-stained images may lack prominent representation of important disease features, resulting in doctors still being limited in the diagnosis process. At the same time, existing methods usually only focus on the conversion of images, lacking quantitative analysis of the lesion area and evaluation of the disease probability. This makes it impossible for doctors to directly obtain diagnostic auxiliary information from virtual indigo carmine-stained images, affecting the efficiency and accuracy of clinical decision-making. Summary of the invention

[0004] In view of this, the present invention provides a virtual staining-assisted diagnosis method and device for celiac disease under gastroscopy, the main purpose of which is to improve the accuracy of disease probability assessment and improve the accuracy of diagnosis results.

[0005] In order to achieve the above object, the present invention mainly provides the following technical solutions:

[0006] In one aspect, the present invention provides a method for endoscopic virtual staining-assisted diagnosis of celiac disease, comprising the following steps:

[0007] Step 1, obtaining an original white-light gastroscopy image of a target patient collected through gastroscopy;

[0008] Step 2: Using a pre-trained recurrent generative adversarial network model, the original white-light gastroscopy image is processed to generate a corresponding virtual indigo carmine-stained image to enhance the visibility of the lesion in the image;

[0009] Step 3: inputting the original white-light gastroscopy image and the virtual indigo carmine-stained image into a pre-trained image classification model for analysis, and outputting the disease probability of the target patient;

[0010] Step 4: Determine the gastric disease diagnosis result of the target patient based on the disease probability.

[0011] The purpose of the present invention and the solution to its technical problems can be further achieved by adopting the following technical measures.

[0012] Optionally, the step 1 includes:

[0013] (1) obtaining a gastroscopic video stream collected by a gastroscopy, wherein the video stream contains multiple gastroscopic image frames per second;

[0014] (2) determining a valid area of ​​each frame of the gastroscopic image, and determining a horizontal circumscribed rectangle of the valid area;

[0015] (3) The gastroscopic image is cropped according to the horizontal circumscribed rectangle to obtain an original white-light gastroscopic image of appropriate size.

[0016] Optionally, the training method of the cyclic generative adversarial network model includes:

[0017] (1) Collect a large number of pairs of original white-light gastroscopy images and real indigo carmine-stained images to form a training dataset;

[0018] (2) using the training data set and adopting a training method of a cyclic generative adversarial network to train a generator and a discriminator;

[0019] (3) The generator receives the original white light gastroscopy image as input and outputs a virtual indigo carmine stained image;

[0020] (4) The discriminator receives the real indigo carmine dyed image and the virtual indigo carmine dyed image as input, and outputs a true or false discrimination result;

[0021] (5) By optimizing the adversarial loss function, update the parameters of the generator and discriminator until the model converges.

[0022] Optionally, the training method of the image classification model includes:

[0023] (1) Collect a dataset of gastroscopy images annotated with disease conditions, including original white-light gastroscopy images and corresponding virtual indigo carmine-stained images;

[0024] (2) establishing image classification models for the original white light image and the virtual indigo carmine-dyed image, respectively, wherein the input data of the classification models are the original white light image and the virtual indigo carmine-dyed image, respectively;

[0025] (3) Use deep convolutional neural networks to train input data, optimize model parameters, and improve classification accuracy;

[0026] (4) the image classification model outputs the disease probability of the target patients respectively;

[0027] (5) Weighting the target patient's disease probability output by the image classification model to obtain a final predicted target patient's disease probability.

[0028] Optionally, in the step 1, the original white light gastroscopy image is also preprocessed, specifically as follows:

[0029] (1) Normalize the original white-light gastroscopy image and adjust the pixel value range;

[0030] (2) Perform image data enhancement processing, including rotation, flipping, and cropping operations, to increase data diversity and quantity.

[0031] Optionally, in the step 1, the original white-light gastroscopy images of the target patient are identified to determine the images belonging to the duodenal bulb and descending part in the original white-light gastroscopy image set.

[0032] Optionally, the step three is specifically as follows:

[0033] (1) extracting features from the original white light image and the virtual indigo carmine dyed image respectively using the image classification model;

[0034] (2) Classify the extracted features and predict the disease probability of each image;

[0035] (3) Through the weighted average method, the results of the original white light image classification network and the virtual indigo carmine staining image classification network are fused to obtain the final disease probability.

[0036] Optionally, the step 4 is specifically as follows:

[0037] (1) When the image classification model outputs the analysis results, the analysis results are converted into disease probabilities;

[0038] (2) Setting a disease probability threshold, when the disease probability is higher than the threshold, determining that the target patient may suffer from the corresponding gastric disease.

[0039] On the other hand, the present invention also provides a virtual staining auxiliary diagnosis device for celiac disease under gastroscopy, the device comprising:

[0040] A data acquisition module, used for acquiring original white-light gastroscopy images of target patients;

[0041] A virtual dyeing generation module, used to process the original white-light gastroscopy image using a pre-trained recurrent generative adversarial network model to generate a virtual indigo carmine dyeing image;

[0042] An image classification module is used to input the original white-light gastroscopy image and the virtual indigo carmine-stained image into a pre-trained image classification model for analysis and output of a disease probability;

[0043] The result output module is used to determine the gastric disease diagnosis result of the target patient according to the disease probability and generate a report.

[0044] By means of the above technical solution, the present invention has at least the following advantages:

[0045] The present invention uses a generative adversarial network to achieve virtual staining of gastroscopic images: The present invention applies a generative adversarial network to virtual indigo carmine staining of gastroscopic images for the first time, and establishes a mapping model from the original white-light gastroscopic image to the virtual stained image. By training the generator and the discriminator, the generated virtual stained image has high fidelity and clinical practicality, enhances the visibility of features in the image, and facilitates doctors to identify the lesion area.

[0046] Computer-aided diagnosis by integrating virtual stained images with original images: This invention combines virtual stained images with original white light images and inputs them into the image classification network, comprehensively utilizing multimodal information to improve the accuracy of disease diagnosis. The images are analyzed through a deep learning model, and the probability of disease is output, providing doctors with an objective diagnostic reference.

[0047] Joint model based on generative adversarial network and image classification network: This paper innovatively constructs a joint model, including a virtual staining generation module and an image classification module. This model can generate high-quality virtual staining images, and can directly classify the images and output the diagnosis results, thus realizing the complete process from virtual staining to auxiliary diagnosis.

[0048] This method enhances the visualization of the lesion area, making it easier for doctors to identify abnormal tissues and improve the accuracy of diagnosis. The current diagnostic accuracy of the model has reached 85%. Through virtual staining technology, additional operations on patients during traditional endoscopic staining are avoided, reducing patient discomfort. At the same time, this method improves diagnostic efficiency, reduces reliance on doctors' personal experience, saves examination time and medical costs, and improves the utilization of medical resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flowchart of a virtual staining-assisted diagnosis method for celiac disease under gastroscopy provided by an embodiment of the present invention;

[0050] Figure 2 This is a white light image of a gastroscope;

[0051] Figure 3 Dyeing images with virtual indigo carmine;

[0052] Figure 4Output images for celiac disease image classification and disease probability;

[0053] Figure 5 A schematic diagram of the structure of a virtual staining-assisted diagnosis device for celiac disease under gastroscopy provided in an embodiment of the present invention.

[0054] The reference numerals in the drawings of the specification include: a data acquisition module 1, a virtual staining generation module 2, an image classification module 3, and a result output module 4. DETAILED DESCRIPTION

[0055] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the present invention application are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures, or characteristics in one or more embodiments may be combined in any suitable form.

[0056] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0057] like Figure 1 As shown, on the one hand, an embodiment of the present invention provides a virtual staining-assisted diagnosis method for celiac disease under gastroscopy, comprising the following steps:

[0058] Step 1: Obtain the original white-light gastroscopy image of the target patient collected through gastroscopy.

[0059] The original white-light gastroscope image is a set of digital images collected by a gastroscope device, which can reflect the patient's stomach and the inner wall information of the duodenal bulb and descending part. Each target patient corresponds to an original white-light gastroscope image set. Specifically, the target patient can be examined by an electronic gastroscope probe to collect a gastroscope video containing stomach and duodenum information. Then, the gastroscope video is frame-processed and screened to generate an original white-light gastroscope image set.

[0060] In this embodiment, the step of obtaining the original white-light gastroscopic image specifically includes: collecting the gastroscopic video stream of the target patient through a gastroscopic device, wherein the video stream contains multiple frames of gastroscopic images per second; performing frame processing on the gastroscopic video stream to extract the gastroscopic image of each frame; converting the video data into a static image sequence to facilitate subsequent image processing and analysis; for each frame of the gastroscopic image, determining the area containing the stomach and the bulb and descending part of the duodenum; and performing a preprocessing process on the extracted image.

[0061] As an optional embodiment, please refer to Figure 2-Figure 4 ,in, Figure 2The original white-light gastroscopy images of the target patients collected through gastroscopy are displayed, which clearly show the inner wall structure and mucosal surface characteristics of the stomach and duodenal bulb and descending part; in these white-light images, the color, texture and morphological characteristics of the gastric mucosa are presented, but due to the limitations of the contrast and resolution of white-light imaging, some tiny lesions such as villous atrophy, fissure or groove-like appearance, and mucosal granularity are not easy to be identified. Figure 3 is corresponding to Figure 2 The virtual indigo carmine-stained image generated from each white light image in the image is processed by a pre-trained cyclic generative adversarial network model (CycleGAN) to simulate the actual staining effect and enhance the subtle structure and texture features of the gastric mucosa. In the virtual indigo carmine-stained image, the lesion area forms a clear contrast with the normal tissue, and the pathological features such as villous atrophy, fissure or groove appearance, and mucosal granularity are highlighted, which is convenient for doctors to observe and analyze. Figure 4 This is a schematic diagram of celiac disease classification and disease probability output. The original white-light gastroscopy image is shown on the left, the corresponding virtual indigo carmine-stained image is shown on the right, and the analysis results based on the image classification network are shown below, including the probability values ​​of each lesion category. The comprehensive celiac disease probability is calculated to help doctors make accurate clinical decisions.

[0062] Among them, the image preprocessing process includes: cropping the original gastroscopic image according to the horizontal circumscribed rectangle to obtain a qualified image containing a valid area; the cropped image not only retains important diagnostic information, but also reduces background noise, which is convenient for subsequent processing; the cropped qualified image is uniformly adjusted to a fixed size (such as 512×512 pixels) to meet the input requirements of the subsequent model; the pixel values ​​of the qualified image are normalized to improve the stability of model training.

[0063] The image virtual dyeing process is implemented by a pre-trained cyclic generative adversarial network model. For example, the CycleGAN cyclic generative adversarial network is selected for training. First, two generators and two discriminators are constructed. One of the generators converts the white-light gastroscopy image into a virtual indigo carmine dyed image, and the other generator converts the virtual indigo carmine dyed image back to a white-light image. The discriminator is used to distinguish between the real white-light image and the generated white-light image, and the real dyed image and the generated virtual indigo carmine dyed image. During the training process, the generator and the discriminator compete with each other through adversarial training. The generator tries to generate images that can deceive the discriminator, and the discriminator is committed to improving the ability to distinguish between real and generated images. At the same time, a cycle consistency constraint is added to ensure that the image can be restored to the original input after bidirectional conversion, maintaining the consistency of image content and structure. Internal details also include using a residual network or encoder-decoder structure in the generator to extract and transform the feature representation of the image and retain important details and texture information; using a convolutional neural network in the discriminator to discriminate the input image and improve the accuracy of the discrimination. Through this architecture, CycleGAN learns the complex nonlinear mapping relationship between white light images and stained images, and achieves high-quality virtual staining processing of gastroscopy images.

[0064] Step 2: Use a pre-trained recurrent generative adversarial network model to process the original white-light gastroscopy image to generate a corresponding virtual indigo carmine-stained image.

[0065] In this embodiment, a pre-trained generative adversarial network (CycleGAN) model is used to process the original white-light gastroscopy image obtained in step 1 to generate a corresponding virtual indigo carmine-stained image. Figure 2 This is a virtual indigo carmine-stained image. The purpose of virtual staining is to enhance the visibility of lesions and important features in the image, helping doctors make diagnoses without actual staining, thereby improving diagnostic efficiency and accuracy.

[0066] Among them, the cyclic generative adversarial network model is used as a virtual staining model. The cyclic generative adversarial network model can learn the mapping between two domains without paired training data. In this case, the two domains are the original white-light gastroscopy image domain and the indigo carmine gastroscopy image domain. The cyclic generative adversarial network model needs to prepare data for training in advance. A large number of unpaired original white-light gastroscopy images and indigo carmine gastroscopy images are collected to construct a training data set; the images should cover various gastric conditions, including normal tissues and different types of lesions (such as villous atrophy, fissure or groove appearance, mucosal granularity, etc.) to improve the generalization ability of the model. After the data collection is completed, the collected data is preprocessed, which specifically includes: all images are uniformly adjusted to a fixed size (for example, 512×512 pixels) to standardize the input size of the neural network; data enhancement techniques such as rotation, flipping, cropping, and color jittering are applied to increase the diversity of training data and prevent overfitting. The model training loss function includes adversarial loss, cycle consistency loss, and identity loss. After the training is completed, the preprocessed original white-light gastroscopy images are input into the trained generator to generate virtual indigo carmine-stained images.

[0067] Step three, inputting the gastroscope white light image into a pre-trained gastroscope white light image classification model for analysis, and outputting the image classification result.

[0068] In this step, the original white-light gastroscopy image obtained in step one is analyzed using a pre-trained gastroscopy white-light image classification model, and the classification result of the image is output. The classification result is used to assess whether the target patient is at risk of celiac gastric disease. A deep convolutional neural network with excellent performance is selected as the classification model for gastroscopy white-light images. In this embodiment, the PVT-2 (Pyramid Vision Transformer v2) network is used, which combines the global feature modeling capability of the Transformer and the local feature extraction capability of the convolutional neural network, and is suitable for processing complex texture and structural information in medical images. Before network training, a large number of gastroscopy white-light images annotated with the disease conditions are collected to construct a training data set. The data should cover normal tissues and various pathological conditions (such as villous atrophy, fissure or groove appearance, mucosal granularity, etc.); adjust all images to a uniform size (for example, 512×512 pixels), and apply rotation, flipping, cropping, brightness and contrast adjustment techniques to increase data diversity and prevent overfitting; use the cross-entropy loss function to measure the difference between the model predicted category and the true category; use the Adam optimizer to set appropriate learning rate and momentum parameters; during training, use the back-propagation algorithm to update the model parameters; stop training when the classification accuracy on the validation set reaches the expected level or the loss no longer decreases significantly; save the trained model parameters and structure for subsequent reasoning and application; for each input image, output the classification result and the corresponding disease probability.

[0069] Among them, image classification can be achieved through a pre-trained image classification model, which can be two PVT-2 deep neural networks. First, the input image is divided into image blocks of fixed size, similar to the sequence segmentation in Transformer. Then, PVT-2 adopts a pyramid-like multi-scale structure to process feature maps of different resolutions layer by layer. At each stage, the improved Transformer encoder is used to calculate the self-attention mechanism on the image block to capture the global and local feature relationships. PVT-2 introduces a sequence downsampling strategy to gradually reduce the size of the feature map while maintaining the richness of the features. Position encoding and multi-head self-attention mechanism are also used internally to enhance the model's ability to express spatial position information and multi-scale features. Finally, the extracted features are classified through the fully connected layer and Softmax classifier, and the probability distribution of the category to which the image belongs is output.

[0070] Step 4: input the virtual indigo carmine dyed image into a pre-trained virtual indigo carmine dyed image classification model for analysis, and output the image classification result.

[0071] In this step, the virtual indigo carmine stained image generated in step 2 is analyzed using a pre-trained virtual indigo carmine stained image classification model, and the classification result of the image is output. The classification result is used to assess whether the target patient is at risk of gastric diseases such as celiac disease. By classifying the virtual indigo carmine stained images, the accuracy and reliability of the diagnosis can be further improved. In this embodiment, the PVT-2 network is used as the virtual indigo carmine stained image classification model. Before training the virtual indigo carmine stained image classification model, a large number of virtual indigo carmine stained images annotated with the disease conditions are collected, including normal tissues and various lesions (such as villous atrophy, fissure or groove appearance, mucosal granularity, etc.). These data can be virtual indigo carmine stained images generated by a cyclic generative adversarial network model from actual stained images. All images are resized to a uniform size (e.g., 512×512 pixels), and techniques such as rotation, flipping, cropping, brightness and contrast adjustment are applied to increase data diversity and prevent overfitting; the cross-entropy loss function is used to measure the difference between the model predicted category and the true category; the Adam optimizer is used to set appropriate learning rate and momentum parameters; during training, the back-propagation algorithm is used to update the model parameters; training is stopped when the classification accuracy on the validation set reaches the expected level or the loss no longer decreases significantly; the trained model parameters and structure are saved for subsequent reasoning and application; for each input virtual indigo carmine-dyed image, the classification result, i.e., the corresponding probability, is output.

[0072] Step 5: Determine the probability of the target patient suffering from celiac disease according to the classification results of the virtual indigo carmine stained image and the white light image.

[0073] In this step, based on the classification results of the white light gastroscopy image and the virtual indigo carmine image obtained in step three and step four, the information of the two is comprehensively analyzed to determine the probability of celiac disease in the target patient. By fusing the classification results of the two images, the accuracy and reliability of the diagnosis can be improved, providing support for subsequent diagnostic decisions. In order to comprehensively utilize the classification information of the white light image and the virtual indigo carmine image, it is necessary to design a fusion strategy to merge the classification results of the two and obtain the probability of celiac disease in the target patient. Figure 4 Output images for celiac disease diagnosis and disease probability. The specific fusion strategy includes: setting weights W1 and W2 to satisfy W1+W2=1. These weights can be determined based on the performance of the model, clinical experience, or the effect on the validation set; for each lesion category of interest (such as celiac disease-related lesions), calculate the comprehensive probability:

[0074] p_disease=w_whitelight*p_whitelight+wdye*p_dye

[0075] The probability of celiac disease in target patients was determined by comprehensively analyzing the classification results of white-light gastroscopy images and virtual indigo carmine-stained images using a weighted fusion method.

[0076] Step six: output the diagnosis result according to the probability of celiac disease in the target patient.

[0077] In this step, the final diagnosis result is generated based on the probability of celiac disease of the target patient calculated in step 5, and is output to the doctor or relevant medical personnel in an appropriate form. This step includes comparing the probability of disease with the preset threshold, drawing a diagnosis conclusion, and generating a detailed diagnosis report to assist clinical decision-making.

[0078] Specifically, it includes: according to clinical experience and model performance, a probability threshold T is set to determine whether the patient may have celiac disease. The commonly used threshold is 0.5, but it can be adjusted according to specific circumstances; when the model outputs the probability result, if the probability is greater than or equal to the threshold T, it is considered that the target patient is more likely to have celiac disease, and further clinical examination and confirmation are recommended; if the probability is less than the threshold T, the diagnosis is that the target patient is less likely to have celiac disease; for the diagnosis result, representative white light and virtual indigo carmine staining images are selected, and the classification results and probability of disease are attached to assist doctors in making judgments.

[0079] like Figure 5 As shown, on the other hand, an embodiment of the present invention provides a virtual staining auxiliary diagnosis device for celiac disease under gastroscopy, comprising:

[0080] Data acquisition module 1, used to acquire original white light gastroscopy images of target patients;

[0081] Specifically, the data acquisition module acquires a gastroscopic video collected by gastroscopic detection, wherein the gastroscopic video includes multiple frames of gastroscopic images; determines a valid area of ​​each frame of the gastroscopic image, and determines a horizontal circumscribed rectangle of the valid area; crops the gastroscopic image according to the horizontal circumscribed rectangle to obtain a qualified image; calculates the clarity of each frame of the qualified image, and determines the qualified image with a clarity greater than a preset threshold as the endoscopic image set;

[0082] A virtual dyeing generation module 2 is used to process the original white light gastroscopy image using a pre-trained recurrent generative adversarial network model to generate a virtual indigo carmine dyeing image;

[0083] Specifically, the virtual staining generation module uses a recurrent generative adversarial network model to process the original white light image, simulating the actual staining effect and enhancing the subtle structure and texture features of the gastric mucosa;

[0084] Image classification module 3, used for inputting the original white light gastroscopy image and the virtual indigo carmine staining image into a pre-trained image classification model, analyzing them, and outputting the disease probability;

[0085] Specifically, the image classification module uses the gastroscope white light image classification model to analyze the gastroscope white light image and output the image classification result, and uses the virtual indigo carmine staining image classification model to analyze the virtual indigo carmine staining image and output the image classification result;

[0086] A result output module 4 is used to determine the gastric disease diagnosis result of the target patient according to the disease probability and generate a report;

[0087] Specifically, the result output module 4 determines the probability of the target patient suffering from celiac disease according to the classification results of the white light gastroscopy image and the classification results of the virtual indigo carmine staining image, and gives a clear diagnosis result.

[0088] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A virtual staining-assisted diagnosis method for celiac disease under gastroscopy, characterized in that: The steps include: Step 1, obtaining an original white-light gastroscopy image of a target patient collected through gastroscopy; Step 2: Using a pre-trained recurrent generative adversarial network model, the original white-light gastroscopy image is processed to generate a corresponding virtual indigo carmine-stained image to enhance the visibility of the lesion in the image; Step 3: inputting the original white-light gastroscopy image and the virtual indigo carmine-stained image into a pre-trained image classification model for analysis, and outputting the disease probability of the target patient; Step 4: Determine the gastric disease diagnosis result of the target patient based on the disease probability.

2. The virtual staining-assisted diagnosis method for celiac disease under gastroscopy according to claim 1, characterized in that: The step one comprises: (1) obtaining a gastroscopic video stream collected by a gastroscopy, wherein the video stream contains multiple gastroscopic image frames per second; (2) determining a valid area of ​​each frame of the gastroscopic image, and determining a horizontal circumscribed rectangle of the valid area; (3) The gastroscopic image is cropped according to the horizontal circumscribed rectangle to obtain an original white-light gastroscopic image of appropriate size.

3. The virtual staining-assisted diagnosis method for celiac disease under gastroscopy according to claim 1, characterized in that: The training method of the cyclic generative adversarial network model includes: (1) Collect a large number of pairs of original white-light gastroscopy images and real indigo carmine-stained images to form a training dataset; (2) using the training data set and adopting a training method of a cyclic generative adversarial network to train a generator and a discriminator; (3) The generator receives the original white light gastroscopy image as input and outputs a virtual indigo carmine stained image; (4) The discriminator receives the real indigo carmine dyed image and the virtual indigo carmine dyed image as input, and outputs a true or false discrimination result; (5) By optimizing the adversarial loss function, update the parameters of the generator and discriminator until the model converges.

4. The virtual staining-assisted diagnosis method for celiac disease under gastroscopy according to claim 1, characterized in that: The training method of the image classification model includes: (1) Collect a dataset of gastroscopy images annotated with disease conditions, including original white-light gastroscopy images and corresponding virtual indigo carmine-stained images; (2) establishing image classification models for the original white light image and the virtual indigo carmine-dyed image, respectively, wherein the input data of the classification models are the original white light image and the virtual indigo carmine-dyed image, respectively; (3) Use deep convolutional neural networks to train input data, optimize model parameters, and improve classification accuracy; (4) the image classification model outputs the disease probability of the target patients respectively; (5) Weighting the target patient's disease probability output by the image classification model to obtain a final predicted target patient's disease probability.

5. The method for virtual staining-assisted diagnosis of celiac disease under gastroscopy according to claim 1, characterized in that: In the step 1, the original white light gastroscopy image is also preprocessed, which is as follows: (1) Normalize the original white-light gastroscopy image and adjust the pixel value range; (2) Perform image data enhancement processing, including rotation, flipping, and cropping operations, to increase data diversity and quantity.

6. The virtual staining-assisted diagnosis method for celiac disease under gastroscopy according to claim 1, characterized in that: In the step 1, the original white-light gastroscopy images of the target patient are identified to determine the images belonging to the duodenal bulb and descending part in the original white-light gastroscopy image set.

7. The virtual staining-assisted diagnosis method for celiac disease under gastroscopy according to claim 4, characterized in that: The step three is as follows: (1) extracting features from the original white light image and the virtual indigo carmine dyed image respectively using the image classification model; (2) Classify the extracted features and predict the disease probability of each image; (3) Through the weighted average method, the results of the original white light image classification network and the virtual indigo carmine staining image classification network are fused to obtain the final disease probability.

8. The virtual staining-assisted diagnosis method for celiac disease under gastroscopy according to claim 1, characterized in that: The step 4 is specifically as follows: (1) When the image classification model outputs the analysis results, the analysis results are converted into disease probabilities; (2) Setting a disease probability threshold, when the disease probability is higher than the threshold, determining that the target patient may suffer from the corresponding gastric disease.

9. A virtual staining auxiliary diagnosis device for celiac disease under gastroscopy, characterized in that: include: A data acquisition module, used for acquiring original white-light gastroscopy images of target patients; A virtual dyeing generation module, used to process the original white-light gastroscopy image using a pre-trained recurrent generative adversarial network model to generate a virtual indigo carmine dyeing image; An image classification module is used to input the original white-light gastroscopy image and the virtual indigo carmine-stained image into a pre-trained image classification model for analysis and output of a disease probability; The result output module is used to determine the gastric disease diagnosis result of the target patient according to the disease probability and generate a report.

Images