Auxiliary gastric cancer diagnosis method based on artificial intelligence

Abstract

The invention relates to a gastric cancer auxiliary diagnosis method based on artificial intelligence. The method comprises the following steps: acquiring a gastric cancer detection image set; inputting the gastric cancer detection image set into a real-time primary processing sub-model of a gastric cancer auxiliary diagnosis artificial intelligence model, and performing gastric cancer examination state identification and abnormal region identification to obtain gastric cancer examination state information and abnormal region primary identification information; based on the gastric cancer examination state information, combining with a stomach standard digital twinborn model to generate stomach examination area information; generating a to-be-enhanced image of the abnormal region based on the gastric cancer detection image set and the abnormal region preliminary identification information, and generating an enhanced image of the abnormal region; and generating gastric cancer auxiliary diagnosis result information based on the enhanced image and the abnormal region preliminary identification information. By adopting the method, the diagnosis capability of primary hospitals can be improved, the misdiagnosis rate and omission ratio of gastric cancer are reduced, and comprehensive and powerful auxiliary diagnosis capability is provided for gastric cancer detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of image analysis, and particularly relates to an artificial intelligence-based gastric cancer auxiliary diagnosis method. Background Art

[0002] The incidence and mortality rates of gastric cancer are both among the top ten cancers globally. Screening and early detection of lesions are the most effective measures to reduce cancer mortality. Traditional gastric cancer screening mainly involves taking tissue samples from abnormal areas using endoscopes or imaging examinations and making pathological sections, and professional pathologists read and diagnose the pathological images.

[0003] Currently, pathologists are mainly concentrated in tertiary hospitals, and the diagnostic needs of primary hospitals are difficult to be met. In the diagnosis and treatment of gastric cancer, with the continuous development of computer hardware and deep learning technology, the application advantages of artificial intelligence technology in digital pathological images have become more and more significant.

[0004] However, in traditional artificial intelligence-based methods for identifying gastric cancer, high-quality images are required, which is inconsistent with the current status of gastric cancer examination conditions, making it difficult to achieve an ideal training effect and unable to guarantee the accuracy of the artificial intelligence algorithm diagnosis. And the existing technologies have relatively single auxiliary means for gastric cancer diagnosis, and it is difficult to promote them among physicians, making it difficult to meet the diagnostic needs of primary hospitals. Summary of the Invention

[0005] Based on this, it is necessary to provide an artificial intelligence-based gastric cancer auxiliary diagnosis method that can adapt to the current status of endoscopic examination conditions in primary hospitals and has comprehensive gastric cancer auxiliary diagnosis capabilities for the above technical problems.

[0006] The present application provides an artificial intelligence-based gastric cancer auxiliary diagnosis method, including:

[0007] Obtaining original gastric cancer detection images and performing preprocessing to obtain a gastric cancer detection image set;

[0008] Inputting the gastric cancer detection image set into the real-time preliminary processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to perform gastric cancer examination status recognition and abnormal area recognition, and obtaining gastric cancer examination status information and preliminary abnormal area recognition information;

[0009] Based on the gastric cancer examination status information, combining with a standard digital twin model of the stomach to generate gastric examination area information;

[0010] Generating a to-be-enhanced image of the abnormal area based on the gastric cancer detection image set and the preliminary abnormal area recognition information, and inputting the to-be-enhanced image into the abnormal area image enhancement sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to generate an enhanced image of the abnormal area;

[0011] Generate the auxiliary diagnosis result information of gastric cancer based on the enhanced image and the preliminary recognition information of the abnormal area.

[0012] The above artificial intelligence-based auxiliary diagnosis method for gastric cancer can quickly identify the gastric cancer examination status and abnormal areas by inputting the gastric cancer detection image set into the real-time preliminary processing sub-model, accelerating the diagnosis process. By generating the enhanced image of the abnormal area, the details of the abnormal area can be shown more clearly, providing richer visual information for doctors, better meeting the current endoscopic examination conditions in primary hospitals, and helping to improve the accuracy of evaluating the nature of gastric cancer lesions.

[0013] Furthermore, the above artificial intelligence-based auxiliary diagnosis method for gastric cancer can generate the information of the gastric examination area by combining the gastric cancer examination status information and the standard digital twin model of the stomach, enabling a more comprehensive evaluation of the progress of gastric cancer examination, ensuring the integrity and accuracy of the examination, thereby reducing the misdiagnosis rate and missed detection rate of gastric cancer detection, and providing comprehensive and powerful auxiliary diagnosis capabilities for gastric cancer detection. The entire diagnosis process is standardized and intelligentized through the artificial intelligence model, reducing the interference of human factors, improving the repeatability and consistency of the diagnosis results, facilitating promotion among physicians, and thus enhancing the gastric cancer diagnosis ability of primary hospitals. Description of the Drawings

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments or related technologies. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0015] Figure 1 Schematic diagram of the application environment of the solution provided in an embodiment of the present application;

[0016] Figure 2 Schematic diagram of the process of an artificial intelligence-based auxiliary diagnosis method for gastric cancer provided in an embodiment of the present application;

[0017] Figure 3 Schematic diagram of the process of another artificial intelligence-based auxiliary diagnosis method for gastric cancer provided in an embodiment of the present application;

[0018] Figure 4 Schematic diagram of the structure of an artificial intelligence model for auxiliary diagnosis of gastric cancer provided in an embodiment of the present application;

[0019] Figure 5 Schematic diagram of the structure of a real-time preliminary processing sub-model provided in an embodiment of the present application;

[0020] Figure 6Schematic diagram of the structure of a real-time advanced processing sub-model provided by an embodiment of the present application;

[0021] Figure 7 Schematic diagram of the structure of another real-time preliminary processing sub-model provided by an embodiment of the present application;

[0022] Figure 8 Schematic diagram of the structure of another real-time advanced processing sub-model provided by an embodiment of the present application;

[0023] Figure 9 Schematic diagram of the structure of an abnormal area image enhancement sub-model provided by an embodiment of the present application;

[0024] Figure 10 Schematic diagram of the structure of an abnormal area image enhancement sub-model during training provided by an embodiment of the present application;

[0025] Figure 11 Schematic diagram of the process of yet another artificial intelligence-based gastric cancer auxiliary diagnosis method provided by an embodiment of the present application;

[0026] Figure 12 Schematic diagram of the structure of an artificial intelligence-based gastric cancer auxiliary diagnosis device provided by an embodiment of the present application. Detailed implementation manners

[0027] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0028] The artificial intelligence-based gastric cancer auxiliary diagnosis method provided by the embodiments of the present application can be applied to an application environment as shown in Figure 1 . Among them, the computing control platform 101 can communicate with the gastric cancer sensing device 102 and the gastric cancer display device 103 through a communication channel. The gastric cancer sensing device 102 can collect gastric cancer image data that the computing control platform 101 needs to process. The computing control platform 101 can generate a gastric cancer examination analysis and judgment result report for the gastric cancer detection patient based on the gastric cancer image data collected by the gastric cancer sensing device 102, so as to realize gastric cancer auxiliary diagnosis. The gastric cancer display device 103 can display the gastric cancer examination analysis and judgment result report generated by the computing control platform 101.

[0029] Schematically, the computing control platform 101 can be, but is not limited to, a server, a high-performance computing cluster, an edge computing platform, and a hybrid computing platform. The server can be implemented by an independent server or a server cluster composed of multiple servers. The gastric cancer sensing device 102 can be, but is not limited to, a surface-enhanced Raman scattering detection platform (SERS), a chemiluminescence immunoassay device (CLIA), an enzyme-linked immunosorbent assay device (ELISA), a colloidal gold immunochromatography device (GICA), a fluorescence immunochromatography device (FICA), a magnetic resonance imaging device, an intelligent gastroscope capsule, a high-definition electronic gastroscope, an ultra-thin electronic gastroscope, and a magnifying endoscope combined with narrow-band imaging device.

[0030] In an exemplary embodiment, as Figure 2 shown, an artificial intelligence-based gastric cancer auxiliary diagnosis method is provided. Taking the method applied to Figure 1 the computing control platform 101 in

[0031] Step S201, obtain the original gastroscopy detection images and perform preprocessing to obtain a set of gastroscopy detection images.

[0032] Specifically, the computing control platform 101 can obtain the original gastric cancer detection images based on the gastric cancer sensing device, preprocess the original gastric cancer detection images, and form a set of gastric cancer detection images.

[0033] Optionally, preprocessing the original gastric cancer detection images can include, but is not limited to: resizing the original gastric cancer detection images, converting the format of the original gastric cancer detection images, normalizing the original gastric cancer detection images, and enhancing the original gastric cancer detection images. Among them, image enhancement can include, but is not limited to: filtering and denoising, geometric transformation, brightness adjustment, and contrast adjustment.

[0034] Step S202, input the set of gastroscopy detection images into the real-time preliminary processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to perform gastroscopy examination status recognition and abnormal area recognition, and obtain gastroscopy examination status information and preliminary abnormal area recognition information.

[0035] Specifically, the computing control platform 101 can input the formed set of gastric cancer detection images into the real-time preliminary processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to perform gastric cancer examination status recognition and abnormal area recognition, and obtain gastric cancer examination status information and preliminary abnormal area recognition information. Among them, the gastric cancer examination status information can be used for, but is not limited to: identifying unexamined parts, identifying the detection difficulty of the detection area, and identifying the standard degree of gastric cancer detection operations; the preliminary abnormal area recognition information can be used for, but is not limited to: identifying the size of the abnormal area and the preliminary classification of gastric diseases in the abnormal area.

[0036] Schematically, since the stomach is a hollow organ, it is prone to deformation and has many folds. In addition, there are often mucus attachments, mucosal reflections, bubble formations, etc. on the mucosal surface, which can greatly interfere with the microscopic field of view. The gastric cancer examination status information generated by the preliminary processing sub-model of the calculation and control platform 101 may include, but is not limited to, gastroscopy examination difficulty information used to characterize the influence degree of gastric examination environmental factors on the quality of gastroscopy images and diagnostic accuracy. Among them, the gastric examination environmental factors may include, but are not limited to: gastric light conditions, gastric fluid environment, gastric gas environment, gastric residues, and gastric activity factors.

[0037] Step S203: Based on the gastroscopy examination status information, combined with the standard digital twin model of the stomach, generate the gastric examination area information.

[0038] Specifically, the calculation and control platform 101 can obtain the preset standard digital twin model of the stomach from the database. The calculation and control platform 101 can generate the gastric examination area information based on the gastric cancer examination status information generated by the real-time preliminary processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model and combined with the standard digital twin model of the stomach.

[0039] Optionally, the calculation and control platform 101 can generate a three-dimensional standard digital twin model of the stomach containing detailed anatomical structure information of the stomach through processing and modeling of medical image data, and store the three-dimensional standard digital twin model of the stomach in the database. Among them, the medical image data includes, but is not limited to: computed radiography (CR), computed tomography imaging (CT), magnetic resonance imaging (MRI), positron emission tomography imaging (PET), and single photon emission computed tomography imaging (SPECT).

[0040] Furthermore, the calculation and control platform 101 can use spatial positioning technology and image matching algorithms to accurately match the actual position information in the gastric cancer examination status information with the corresponding position in the standard digital twin model of the stomach. And based on the three-dimensional structure information of the standard digital twin model of the stomach, calculate the ratio of the volume of the examined area to the volume of the entire gastric area to generate the gastric examination area information.

[0041] Step S204: Generate the image to be enhanced of the abnormal area based on the gastroscopy detection image set and the preliminary identification information of the abnormal area, and input the image to be enhanced into the abnormal area image enhancement sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to generate the enhanced image of the abnormal area.

[0042] Specifically, the calculation and control platform 101 can generate a mask for segmenting the abnormal region based on the gastric cancer detection image set and the preliminary identification information of the abnormal region, and generate an image to be enhanced based on the mask and the gastric cancer detection image set. The calculation and control platform 101 can input the image to be enhanced into the abnormal region image enhancement sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to generate an enhanced image of the abnormal region.

[0043] Step S205, generate gastric cancer auxiliary diagnosis result information based on the enhanced image and the preliminary identification information of the abnormal region.

[0044] Optionally, the calculation and control platform 101 can use a large number of labeled gastric cancer image data to train a machine learning model or a deep learning model, and input the enhanced image and the preliminary identification information of the abnormal region into the trained model. Through the learned feature patterns and rules, classify and diagnose and predict the lesions.

[0045] In the above artificial intelligence-based gastric cancer auxiliary diagnosis method, by using the real-time preliminary processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to identify the gastric cancer examination status and preliminarily identify the abnormal region in the gastric cancer detection image set, it is possible to quickly identify the gastric cancer examination status and the abnormal region of the stomach, improve the overall diagnosis efficiency, and save the time and energy of doctors. Generating an image to be enhanced based on the preliminary identification information of the abnormal region and generating an enhanced image based on the abnormal region image enhancement sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model helps to more accurately observe and analyze the lesion characteristics, reduce misdiagnosis or missed diagnosis caused by poor image quality, and improve the accuracy of gastric cancer diagnosis.

[0046] Furthermore, in the above artificial intelligence-based gastric cancer auxiliary diagnosis method, generating gastric examination region information based on the gastric cancer examination status information in combination with the standard digital twin model of the stomach can understand the progress of the examination in real time, help doctors grasp the progress of gastric cancer examination in real time, improve the efficiency of gastric cancer examination, optimize the gastric cancer examination process, reduce the missed detection rate of gastric cancer, and further make the gastric cancer diagnosis process more standardized and standardized, and improve the consistency and stability of the medical quality of gastric cancer detection.

[0047] In an optional embodiment, as Figure 3 shown, the gastric cancer auxiliary diagnosis result information includes lesion edge recognition information and abnormal region advanced recognition information. Generating gastric cancer auxiliary diagnosis result information based on the enhanced image and the preliminary identification information of the abnormal region includes:

[0048] Step S305, input the enhanced image and the preliminary identification information of the abnormal region into the real-time advanced processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to perform lesion edge recognition and lesion type recognition, and generate lesion edge recognition information and abnormal region advanced recognition information.

[0049] Specifically, the enhanced image and the preliminary recognition information of the abnormal area can be input into the real-time advanced processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to perform lesion edge recognition and lesion type recognition, generating lesion edge recognition information and advanced recognition information of the abnormal area. Among them, the advanced recognition information of the abnormal area is used to represent the detailed classification information of the abnormal area

[0050] In the gastric cancer auxiliary diagnosis method based on artificial intelligence provided by the embodiments of the present application, through the real-time advanced processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model for lesion edge recognition, the boundary of the lesion area can be determined more accurately, which helps doctors accurately evaluate the size, shape and scope of the lesion, assist doctors in determining the resection range in surgical planning, reduce damage to normal tissues, and improve the success rate of surgery and the prognosis of patients. And based on the real-time advanced processing sub-model, different types of gastric cancer lesions or other gastric diseases can be accurately distinguished, which helps assist doctors in selecting the most appropriate treatment strategy, and thus can improve the pertinence and effectiveness of treatment

[0051] In an alternative embodiment, as Figure 3 shown, a gastric cancer auxiliary diagnosis method based on artificial intelligence is provided, which may include the following steps S301 to step S305. Among them:

[0052] Step S301, obtaining the original gastroscopy detection image and performing preprocessing to obtain the gastroscopy detection image set

[0053] Step S302, inputting the gastroscopy detection image set into the real-time preliminary processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to perform gastroscopy examination status recognition and abnormal area recognition, obtaining the gastroscopy examination status information and the preliminary recognition information of the abnormal area

[0054] Step S303, based on the gastroscopy examination status information, combining with the standard digital twin model of the stomach, generating the stomach examination area information

[0055] Step S304, generating the image to be enhanced of the abnormal area based on the gastroscopy detection image set and the preliminary recognition information of the abnormal area, and inputting the image to be enhanced into the abnormal area image enhancement sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to generate the enhanced image of the abnormal area

[0056] Step S305, inputting the enhanced image and the preliminary recognition information of the abnormal area into the real-time advanced processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to perform lesion edge recognition and lesion type recognition, generating lesion edge recognition information and advanced recognition information of the abnormal area

[0057] In an alternative embodiment, as Figure 4As shown, the gastric cancer auxiliary diagnosis artificial intelligence model may include a real-time preliminary processing sub-model, an abnormal region image enhancement sub-model and a real-time advanced processing sub-model. Among them, the real-time preliminary processing sub-model can generate gastroscopy status information and abnormal region preliminary identification information based on the gastroscopy detection image set; the abnormal region image enhancement sub-model is connected to the real-time preliminary processing sub-model, and the abnormal region image enhancement sub-model can generate an enhanced image of the abnormal region of gastric cancer examination based on the gastroscopy detection image set and the abnormal region preliminary identification information; the real-time advanced processing sub-model is connected to the real-time preliminary processing sub-model and the abnormal region image enhancement sub-model, and the real-time advanced processing sub-model can generate lesion edge identification information and abnormal region advanced identification information based on the enhanced image and the abnormal region preliminary identification information.

[0058] In one of the optional embodiments, the gastric cancer examination status information may include but is not limited to gastric examination area identification information and gastric cancer examination speed information, the abnormal area preliminary identification information may include but is not limited to abnormal area identification information and abnormal area gastric disease preliminary classification information, and the abnormal area advanced identification information may include but is not limited to abnormal morphology subcategory information and infiltration depth subcategory information. Among them, the gastric disease preliminary classification information may include but is not limited to gastric cancer information, gastritis classification information, gastric ulcer information and gastric polyp information, and the abnormal morphology subcategory information may include but is not limited to undifferentiated gastric cancer information, gastric intraepithelial neoplasia information, dysplasia information, chronic atrophic gastritis information and non-atrophic gastritis information.

[0059] Please refer to Figure 5 The real-time preliminary processing sub-model may be an improved YOLO (You Only Look Once) model, and the head network of the improved YOLO model may include a first detection module, a second detection module, and a third detection module in descending order according to the sampling multiple. The first detection module may include but is not limited to a stomach inspection area recognition detection head, the second detection module may include but is not limited to an abnormal area recognition detection head and a stomach disease preliminary classification detection head, and the third detection module may include an abnormal area recognition detection head, a stomach disease preliminary classification detection head, and a stomach cancer inspection speed detection head.

[0060] Optionally, the stomach inspection area recognition detection head, the stomach disease preliminary classification detection head and the stomach cancer inspection speed detection head may be classification detection heads. The abnormal area recognition detection head may be a pre-selection box detection head.

[0061] Please refer to Figure 6, the real-time advanced processing sub-model can be an improved U-NET (U-shaped network) model. The decoder of the improved U-NET model can include an Inception branch and a residual convolution branch. The Inception branch can be used to extract the features of the second abnormal category recognition information, and the residual convolution branch can be used to extract the features of the lesion edge recognition information.

[0062] In the gastric cancer auxiliary diagnosis method based on artificial intelligence provided by the embodiments of the present application, the improved YOLO model can achieve hierarchical detection of various information such as the gastric examination area, gastric abnormal area, gastric disease types, and gastric cancer examination speed through the first detection module, the second detection module, and the third detection module set according to different downsampling multiples in the head network. It can better adapt to information of different scales and modalities, improve the accuracy and comprehensiveness of gastric cancer detection, and enable the model to more accurately locate and identify relevant information of gastric lesions. And by setting targeted detection heads in each module, it can be optimized for specific tasks, enabling the real-time preliminary processing sub-model to obtain different types of information more professionally and efficiently.

[0063] Furthermore, in the gastric cancer auxiliary diagnosis method based on artificial intelligence provided by the embodiments of the present application, the improved U-NET model can, through the Inception branch and the residual convolution branch in the decoder, simultaneously extract the features of the second abnormal category recognition information and the features of the lesion edge recognition information in the enhanced image based on the preliminary recognition information of the abnormal area. Among them, the Inception branch can effectively extract features of different scales and comprehensively understand the category information of the lesion. The residual convolution branch can better capture the detailed features of the lesion edge and improve the accuracy of lesion edge recognition. This combination enables the real-time advanced processing sub-model to excellently identify the lesion type and lesion edge in real time, providing doctors with more detailed and accurate lesion information, which helps to improve the accuracy and reliability of gastric cancer diagnosis.

[0064] In an optional embodiment, as Figure 7 shown, the improved YOLO model can be obtained by replacing the neck network structure of the original YOLOV8 model with a bidirectional feature pyramid network structure with three layers of structure. The output of each layer of the bidirectional feature pyramid network structure with three layers of structure can be respectively connected to the first detection module, the second detection module, and the third detection module. The lateral connection sub-structure of the bidirectional feature pyramid network structure can include a Transformer module.

[0065] Optionally, the repetition number N of the Transformer module can be 2 to 6.

[0066] The expression of the loss function of the improved YOLO model is:

[0067]

[0068] In the formula, L Y is the loss function of the YOLO model, and L EI is the loss function for identifying the gastric inspection area, is the loss function for identifying the abnormal area of the j-th detection module, is the loss function for the preliminary classification and identification of gastric diseases in the j-th detection module, and L V is the loss function for the gastric cancer inspection speed, is the coefficient of the loss function for identifying the gastric inspection area, is the coefficient of the loss function for identifying the abnormal area of the j-th detection module, is the coefficient of the loss function for the preliminary classification and identification of gastric diseases in the j-th detection module, is the coefficient of the loss function for the gastric cancer inspection speed. Among them, L EI and L V adopt a loss function based on linear regression, adopt a loss function based on the intersection over union, adopt a loss function based on cross-entropy.

[0069] Optionally, the loss function for improving the YOLO model may also include Transformer module items.

[0070] In the gastric cancer auxiliary diagnosis method based on artificial intelligence provided by the embodiments of the present application, by replacing the neck network structure of the original YOLOV8 model with a three-layer bidirectional feature pyramid network structure, different levels of feature information can be better fused, effectively integrating the detailed information of the shallow layer and the semantic information of the deep layer, so that the real-time preliminary processing sub-model can more accurately obtain the gastroscopy examination status information and the preliminary identification information of the abnormal area.

[0071] Furthermore, the Transformer module has a powerful self-attention mechanism, which can capture long-range dependencies and context information in the image. In the gastric cancer auxiliary diagnosis method based on artificial intelligence provided by the embodiments of the present application, by setting the Transformer module in the horizontal connection sub-structure of the bidirectional feature pyramid network structure of the real-time preliminary processing sub-model, it can help the real-time preliminary processing sub-model to assist in judging the nature and type of the lesion according to the characteristics of the tissues around the lesion area, and improve the classification and border accuracy of the real-time preliminary processing sub-model.

[0072] In an optional embodiment, such as Figure 8As shown in the figure, the improved U-NET model is obtained by replacing the convolutional module in the original U-NET model with a residual convolutional module, connecting a hybrid local channel attention module before the downsampling layer in the encoder of the original U-NET model, replacing the residual convolutional module with the largest downsampling factor with two Inception standard modules connected through a multi-scale feature pyramid module in the middle, and adding an Inception branch in the decoder.

[0073] Optionally, the Inception module can be obtained by connecting at least two Inception standard modules in series.

[0074] Optionally, the downsampling layer can be a max pooling layer, and the upsampling layer can be an upsampling convolutional layer.

[0075] Optionally, the fully connected layer in the Inception branch can be constructed based on a BP neural network.

[0076] The expression of the loss function of the improved U-NET model is:

[0077] L U-NET =μ IS ·L IS +μ CL ·L CL

[0078]

[0079] In the formula, L U-NET is the loss function of the improved U-NET model, L IS is the loss function term for lesion edge recognition, L CL is the loss function term for the recognition of the second abnormal category, μ IS is the coefficient of the loss function term for lesion edge recognition, μ CL is the coefficient of the loss function term for the second abnormal category recognition, H and W are the height and width of the image respectively, y h,w is the binary label map, is the predicted probability map, N is the number of samples, C is the number of abnormal morphology subcategories, K is the number of infiltration depth subcategories, μ CLC is the weighting coefficient of the abnormal morphology subcategory, y n,c,k is the nth sample, p n,c,k is the probability distribution of the nth sample.

[0080] In the gastric cancer auxiliary diagnosis method based on artificial intelligence provided by the embodiments of the present application, by replacing the convolutional module of the original U-NET model with a residual convolutional module, it can help solve the problem of gradient disappearance in deep networks, so that more abstract and richer features can be extracted, which helps to better capture the subtle features and complex semantic content of lesions and improve the recognition ability of the model for gastric lesions. By connecting a hybrid local channel attention module before the downsampling layer of the encoder of the original U-NET model, the importance of different channels and spatial positions can be automatically recognized, and the interference of irrelevant backgrounds and noises can be suppressed, thereby improving the sensitivity and accuracy of the real-time advanced processing sub-model for gastric lesions. By replacing the residual convolutional module with the largest downsampling multiple with two Inception standard modules connected by a multi-scale feature pyramid module in the middle and adding an Inception branch in the decoder, multi-scale features can be effectively fused to provide more comprehensive information, and then the lesion areas in gastric cancer detection images can be better segmented and classified.

[0081] In an alternative embodiment, please refer to Figure 9 and Figure 10 , the abnormal area image enhancement sub-model is a conditional Laplacian pyramid variational autoencoder generative adversarial network model conditional on the first abnormal type recognition information. The conditional Laplacian pyramid variational autoencoder generative adversarial network model includes a first autoencoder, a second autoencoder, a third autoencoder, a first generator, a second generator, a third generator, a first discriminator, a second discriminator, a third discriminator, a first classifier, a second classifier, and a third classifier.

[0082] Optionally, the first autoencoder, the second autoencoder, the third autoencoder, the first generator, the second generator, and the third generator can be neural networks conditional on the first abnormal type recognition information.

[0083] The inputs of the first autoencoder, the second autoencoder, and the third autoencoder are the gastric cancer detection image set, which includes a gastric cancer detection patch image subset. The outputs of the first autoencoder, the second autoencoder, and the third autoencoder are respectively the inputs of the first generator, the second generator, and the third generator. The input of the second generator also includes the first reconstructed image, and the input of the third generator also includes the second reconstructed image. Among them, the first reconstructed image is obtained by upsampling the output image of the first generator, and the second reconstructed image is obtained by upsampling the added image of the first reconstructed image and the output image of the second generator.

[0084] When training the conditional Laplacian pyramid variational autoencoder generative adversarial network model, the inputs of the first discriminator and the first classifier include the output of the first generator and the two downsampled atlases of the gastric cancer high-definition enhanced training atlas. The inputs of the second discriminator and the second classifier include the output of the second generator conditioned on the residual atlas of the two downsampled atlases and the first real difference image set. The inputs of the third discriminator and the third classifier include the output of the third generator conditioned on the residual atlas of the one-downsampled atlas and the second real difference image set. Among them, the first real difference image set is the difference between the residual atlas of the two downsampled atlases and the one-downsampled atlas, and the second real difference image set is the difference between the residual atlas of the one-downsampled atlas and the gastric cancer high-definition enhanced training atlas.

[0085] Optionally, the sharpness of the training data W1, the training data W2, the training data W3, and the real data Y can increase in sequence.

[0086] The expression of the loss function of the conditional Laplacian pyramid variational autoencoder generative adversarial network model is:

[0087]

[0088] In the formula, L C·LAP·VEA-GAN is the loss function of the conditional Laplacian pyramid variational autoencoder generative adversarial network model, N I is the number of pyramid layers of the conditional Laplacian pyramid variational autoencoder generative adversarial network model, is the reconstruction loss function term of the i-th layer, is the coefficient of the reconstruction loss function term of the i-th layer, is the encoder loss function term of the i-th layer, is the coefficient of the encoder loss function term of the i-th layer, is the generator loss function term of the i-th layer, is the coefficient of the generator loss function term of the i-th layer, is the discriminator loss function term of the i-th layer, is the coefficient of the discriminator loss function term of the i-th layer, is the classifier loss function term of the i-th layer, is the coefficient of the classifier loss function term of the i-th layer.

[0089] In the gastric cancer assisted diagnosis method based on artificial intelligence provided by the embodiments of the present application, by introducing an abnormal region image enhancement sub-model based on a conditional Laplacian pyramid variational autoencoder generative adversarial network model, the generation of high-quality gastric cancer detection images is realized. The characteristics of the abnormal region image enhancement sub-model, such as multi-scale generation, conditional generation, multi-level loss function design, multi-discriminator and multi-classifier design, significantly improve the quality of the generated images and the training stability of the model, providing strong support for the assisted diagnosis of gastric cancer and data augmentation.

[0090] In an alternative embodiment, please refer to Figure 11 , and based on the gastric cancer examination status information, combined with the standard gastric digital twin model, generate gastric examination area information, including:

[0091] Step S1103, input the gastric cancer examination status information into the standard gastric digital twin model to generate and update the gastric cancer examination identification digital twin model.

[0092] Specifically, the gastric cancer examination status information can be input into the standard gastric digital twin model to generate and update the gastric cancer examination identification digital twin model, and the gastric cancer examination identification digital twin model can be used to display the gastric detection area corresponding to the original gastric cancer detection image.

[0093] Step S1104, generate gastric examination area information based on the gastric area corresponding to the original gastric cancer detection image in the gastric cancer examination identification digital twin model.

[0094] In the gastric cancer assisted diagnosis method based on artificial intelligence provided by the embodiments of the present application, by generating the gastric cancer examination identification digital twin model, the examined area and unexamined area of the gastric cancer detection can be intuitively displayed, which helps the doctor to understand the coverage of the gastroscopy examination, avoid missing important areas, and ensure the comprehensiveness and integrity of the examination.

[0095] In an alternative embodiment, as Figure 11 shown, a gastric cancer assisted diagnosis method based on artificial intelligence is provided, which may include the following steps S1101 to step S1106. Among them:

[0096] Step S1101, obtain the original gastric cancer detection image and perform preprocessing to obtain a gastric cancer detection image set.

[0097] Step S1102, input the gastric cancer detection image set into the real-time preliminary processing sub-model of the gastric cancer assisted diagnosis artificial intelligence model to perform gastric cancer examination status recognition and abnormal area recognition, and obtain gastric cancer examination status information and preliminary abnormal area recognition information.

[0098] Step S1103: Input the gastric cancer examination status information into the standard gastric digital twin model to generate and update the gastric cancer examination identification digital twin model.

[0099] Step S1104: Generate gastric examination area information based on the gastric area corresponding to the original gastric cancer detection image in the gastric cancer examination identification digital twin model.

[0100] Step S1105: Generate the image to be enhanced for the abnormal area based on the gastric cancer detection image set and the preliminary identification information of the abnormal area, and input the image to be enhanced into the abnormal area image enhancement sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to generate the enhanced image of the abnormal area.

[0101] Step S1106: Input the enhanced image and the preliminary identification information of the abnormal area into the real-time advanced processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to perform lesion edge recognition and lesion type recognition, and generate lesion edge recognition information and advanced recognition information of the abnormal area.

[0102] In the gastric cancer auxiliary diagnosis method based on artificial intelligence provided by the embodiments of the present application, through the real-time preliminary processing sub-model, the gastric cancer examination status and the abnormal area can be quickly recognized, the time for manual judgment can be reduced, and the diagnosis efficiency can be improved; through the abnormal area image enhancement sub-model, high-quality enhanced images can be generated to assist doctors in observing the lesion details more clearly and improving the diagnosis accuracy; the gastric examination area information generated based on the gastric cancer examination identification digital twin model can assist doctors in mastering the examination site in real time, optimizing the examination process and methods; through the real-time advanced processing sub-model, more detailed and accurate lesion information can be provided to assist doctors in more accurately evaluating the nature, scope and severity of gastric lesions and assisting in formulating personalized treatment plans.

[0103] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least a part of other steps or steps in other steps.

[0104] Based on the same inventive concept, an embodiment of the present application further provides an artificial intelligence-based gastric cancer auxiliary diagnosis device for implementing the above-mentioned artificial intelligence-based gastric cancer auxiliary diagnosis method. The solution provided by this device for solving problems is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the artificial intelligence-based gastric cancer auxiliary diagnosis device provided below can refer to the limitations on the artificial intelligence-based gastric cancer auxiliary diagnosis method in the above text, and will not be repeated here.

[0105] In an exemplary embodiment, as Figure 12 shown, an artificial intelligence-based gastric cancer auxiliary diagnosis device 1200 is provided, including:

[0106] A gastric cancer detection image acquisition module 1201, which can be used to acquire the original gastric cancer detection images and perform preprocessing to obtain a gastric cancer detection image set.

[0107] A preliminary gastric cancer image processing module 1202, which can be used to input the gastric cancer detection image set into the real-time preliminary processing sub-model of the artificial intelligence model for gastric cancer auxiliary diagnosis to identify the gastric cancer examination status and abnormal areas, and obtain the gastric cancer examination status information and preliminary abnormal area identification information.

[0108] An examination area information generation module 1203, which can be used to generate gastric examination area information based on the gastric cancer examination status information and in combination with the standard digital twin model of the stomach.

[0109] An abnormal area image enhancement module 1204, which can be used to generate an image to be enhanced for the abnormal area based on the gastric cancer detection image set and the preliminary abnormal area identification information, and input the image to be enhanced into the abnormal area image enhancement sub-model of the artificial intelligence model for gastric cancer auxiliary diagnosis to generate an enhanced image of the abnormal area.

[0110] An auxiliary diagnosis result generation module 1205, which can be used to generate gastric cancer auxiliary diagnosis result information based on the enhanced image and the preliminary abnormal area identification information.

[0111] In an alternative embodiment of the present application, the auxiliary diagnosis result generation module 1205 can also be used to input the enhanced image and the preliminary abnormal area identification information into the real-time advanced processing sub-model of the artificial intelligence model for gastric cancer auxiliary diagnosis to identify the lesion edge and lesion type, and generate the lesion edge identification information and the advanced abnormal area identification information.

[0112] In an alternative embodiment of the present application, the inspection area information generation module 1203 may also be configured to input the gastroscopy inspection status information into the gastric standard digital twin model to generate and update the gastroscopy inspection identification digital twin model, which is used to display the areas that have been inspected by the gastroscopy and the areas that have not been inspected by the gastroscopy; and generate the gastric inspection area information based on the ratio of the areas that have been inspected by the gastroscopy to the areas that have not been inspected by the gastroscopy in the gastroscopy inspection identification digital twin model.

[0113] In an embodiment of the present application, a computer device is provided, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of a power supply safety management method as described above are implemented.

[0114] In an embodiment of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0115] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can refer to the partial descriptions of the method embodiments. The device embodiments described above are only illustrative. The components described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present disclosure solution. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0116] The above embodiments only represent several implementation manners of the embodiments of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the application embodiments. It should be noted that for those of ordinary skill in the art, without departing from the concept of the embodiments of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the embodiments of the present application.

Claims

1. An artificial intelligence-based auxiliary diagnosis method for gastric cancer, characterized in that: The method comprises: Obtaining original gastric cancer detection images and performing preprocessing to obtain a gastric cancer detection image set; Inputting the gastric cancer detection image set into the real-time preliminary processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to perform gastric cancer examination status recognition and abnormal area recognition, and obtaining gastric cancer examination status information and abnormal area preliminary recognition information; Based on the gastric cancer examination status information and in combination with the standard digital twin model of the stomach, stomach examination area information is generated; Generate an image to be enhanced of the abnormal region based on the gastric cancer detection image set and the preliminary identification information of the abnormal region, input the image to be enhanced into the abnormal region image enhancement sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model, and generate an enhanced image of the abnormal region; Gastric cancer auxiliary diagnosis result information is generated based on the enhanced image and the preliminary identification information of the abnormal area.

2. The method according to claim 1, characterized in that The gastric cancer auxiliary diagnosis result information includes lesion edge recognition information and abnormal region advanced recognition information, and the gastric cancer auxiliary diagnosis result information is generated based on the enhanced image and the abnormal region preliminary recognition information, including: The enhanced image and the preliminary identification information of the abnormal area are input into the real-time advanced processing sub-model of the artificial intelligence model for auxiliary diagnosis of gastric cancer to perform lesion edge identification and lesion type identification, and generate the lesion edge identification information and the advanced identification information of the abnormal area.

3. The method according to claim 2, characterized in that The gastric cancer examination status information includes gastric examination area identification information and gastric cancer examination speed information, the abnormal area preliminary identification information includes abnormal area identification information and gastric disease preliminary classification information of the abnormal area, and the abnormal area advanced identification information includes abnormal morphology subcategory information and infiltration depth subcategory information; The real-time preliminary processing sub-model is an improved YOLO model, and the head network of the improved YOLO model includes a first detection module, a second detection module and a third detection module in descending order according to the sampling multiple; wherein the first detection module includes a stomach inspection area recognition detection head, the second detection module includes an abnormal area recognition detection head and a stomach disease preliminary classification detection head, and the third detection module includes an abnormal area recognition detection head, a stomach disease preliminary classification detection head and a stomach cancer inspection speed detection head; The real-time advanced processing sub-model is an improved U-NET model, and the decoder of the improved U-NET model includes an Inception branch and a residual convolution branch. The Inception branch is used to extract the features of the second abnormal category identification information, and the residual convolution branch is used to extract the features of the lesion edge identification information.

4. The method according to claim 3, characterized in that The improved YOLO model is obtained by replacing the neck network structure of the original YOLOV8 model with a bidirectional feature pyramid network structure with three structural layers, wherein the output of each layer of the bidirectional feature pyramid network structure with three structural layers is respectively connected to the first detection module, the second detection module and the third detection module, and the lateral connection substructure of the bidirectional feature pyramid network structure includes a Transformer module; The expression of the loss function of the improved YOLO model is: Where, L Y is the loss function of the YOLO model, L EI Loss function for stomach inspection region identification, is the abnormal region identification loss function of the j-th detection module, The loss function for the initial classification and recognition of gastric diseases in the jth detection module, L V is the gastric cancer inspection speed loss function, Identify the loss function coefficients for the stomach inspection region, is the loss function coefficient for abnormal region identification of the jth detection module, The loss function coefficient of the initial classification and recognition of gastric diseases in the jth detection module, is the gastric cancer inspection speed loss function coefficient; where L EI and L V Using a loss function based on linear regression, Using the loss function based on intersection-over-union ratio, A cross-entropy based loss function is used.

5. The method according to claim 3, characterized in that: The improved U-NET model is obtained by replacing the convolution module with a residual convolution module in the original U-NET model, connecting a mixed local channel attention module before the downsampling layer in the encoder of the original U-NET model, replacing the residual convolution module with the largest downsampling multiple with two Inception standard modules connected in the middle through a multi-scale feature pyramid module, and adding the Inception branch in the decoder; The expression of the loss function of the improved U-NET model is: L U-NET =μ IS ·L IS +m CL ·L CL Where, L U-NET To improve the loss function of the U-NET model, L IS is the loss function term for lesion edge identification, L CL Identify the loss function term for the second abnormal category, μ IS is the coefficient of the lesion edge identification loss function, μ CL Identify the loss function coefficient for the second abnormal category. H and W are the height and width of the image respectively. h,w is a binary label map, is the predicted probability map, N is the number of samples, C is the number of abnormal morphology subcategories, K is the number of infiltration depth subcategories, μ CLC is the weighting coefficient of the abnormal morphology subcategory, y n,c,k is the nth sample, p n,c,k is the probability distribution of the nth sample.

6. The method according to claim 1, characterized in that; The abnormal region image enhancement sub-model is a conditional Laplacian pyramid variational autoencoder generative adversarial network model based on the first abnormal category identification information, and the conditional Laplacian pyramid variational autoencoder generative adversarial network model includes a first autoencoder, a second autoencoder, a third autoencoder, a first generator, a second generator, a third generator, a first discriminator, a second discriminator, a third discriminator, a first classifier, a second classifier and a third classifier; The inputs of the first autoencoder, the second autoencoder and the third autoencoder are the gastric cancer detection image set, the gastric cancer detection image set includes a gastric cancer detection patch image subset, the outputs of the first autoencoder, the second autoencoder and the third autoencoder are the inputs of the first generator, the second generator and the third generator respectively, the input of the second generator also includes the first reconstructed image, and the input of the third generator also includes the second reconstructed image; wherein the first reconstructed image is obtained by upsampling the output image of the first generator, and the second reconstructed image is obtained by upsampling the added image of the first reconstructed image and the output image of the second generator; When training the conditional Laplacian pyramid variational autoencoder to generate an adversarial network model, the inputs of the first discriminator and the first classifier include the output of the first generator and the twice downsampled atlas of the gastric cancer high-definition enhanced training atlas, the inputs of the second discriminator and the second classifier include the output of the second generator conditioned on the residual atlas of the twice downsampled atlas and the first real difference image set, and the inputs of the third discriminator and the third classifier include the output of the third generator conditioned on the residual atlas of the once downsampled atlas and the second real difference image set; wherein the first real difference image set is the difference between the residual atlas of the twice downsampled atlas and the once downsampled atlas, and the second real difference image set is the difference between the residual atlas of the once downsampled atlas and the gastric cancer high-definition enhanced training atlas; The expression of the loss function of the conditional Laplacian pyramid variational autoencoder generative adversarial network model is: Where, L C·LAP·VEA-GAN The loss function for the conditional Laplacian pyramid variational autoencoder to generate the adversarial network model, N I The number of pyramid layers for the conditional Laplacian pyramid variational autoencoder to generate the adversarial network model, is the reconstruction loss function term of the i-th layer, is the coefficient of the reconstruction loss function term of the i-th layer, is the encoder loss function term of the i-th layer, is the coefficient of the encoder loss function term of the i-th layer, is the generator loss function term of the i-th layer, is the coefficient of the generator loss function term of the i-th layer, is the discriminator loss function term of the i-th layer, is the coefficient of the discriminator loss function term of the i-th layer, is the classifier loss function term of the i-th layer, is the coefficient of the classifier loss function term of the i-th layer.

7. The method according to any one of claims 1 to 6, characterized in that: The stomach inspection area information is generated based on the stomach cancer inspection status information and combined with the stomach standard digital twin model, including: Inputting the gastric cancer examination status information into the stomach standard digital twin model, generating and updating a gastric cancer examination identification digital twin model, wherein the gastric cancer examination identification digital twin model is used to display the stomach area corresponding to the gastric cancer detection original image; The stomach inspection area information is generated based on the gastric cancer area corresponding to the gastric cancer detection original image in the gastric cancer inspection identification digital twin model.

8. An artificial intelligence-based gastric cancer auxiliary diagnosis device, characterized in that: The device comprises: A gastric cancer detection image acquisition module is used to acquire original gastric cancer detection images and perform preprocessing to obtain a gastric cancer detection image set; A gastric cancer image preliminary processing module is used to input the gastric cancer detection image set into the real-time preliminary processing sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to perform gastric cancer examination status recognition and abnormal area recognition, and obtain gastric cancer examination status information and abnormal area preliminary recognition information; An inspection area information generation module, used to generate stomach inspection area information based on the gastric cancer inspection status information and in combination with a stomach standard digital twin model; An abnormal region image enhancement module is used to generate an image to be enhanced in the abnormal region based on the gastric cancer detection image set and the preliminary identification information of the abnormal region, and input the image to be enhanced into the abnormal region image enhancement sub-model of the gastric cancer auxiliary diagnosis artificial intelligence model to generate an enhanced image of the abnormal region; The auxiliary diagnosis result generation module is used to generate auxiliary diagnosis result information of gastric cancer based on the enhanced image and the preliminary identification information of the abnormal area.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images