Digestive endoscopy image auxiliary diagnosis system based on artificial intelligence

Abstract

The invention relates to the technical field of digestive endoscopy image auxiliary diagnosis, in particular to a digestive endoscopy image auxiliary diagnosis system based on artificial intelligence, which comprises an image preprocessing module, a tiny focus detection module, a difficult focus recognition module, a focus classification module, a three-dimensional reconstruction module, a focus change tracking module, an auxiliary decision module and a privacy protection module. The image preprocessing module is used for image enhancement and noise removal. By automatically analyzing and identifying the endoscopic images, the subjective influence of human factors is reduced, the objectivity and accuracy of diagnosis are improved, a large number of endoscopic images can be quickly processed, lesions can be automatically positioned, features can be extracted, classification and diagnosis can be carried out, the system can capture tiny or hidden lesions, missed diagnosis is reduced, and the diagnosis efficiency is improved. Three-dimensional reconstruction and virtual reality interaction of a focus can be realized, visual space structure information is provided for a doctor, the focus can be observed from different angles, and the diagnosis accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of digestive endoscopy image - assisted diagnosis, and specifically to an artificial - intelligence - based digestive endoscopy image - assisted diagnosis system. Background Technique

[0002] Digestive endoscopy images are images of the internal structure of the digestive tract obtained through digestive endoscopes (such as gastroscopes, colonoscopes, etc.). These images are used for the diagnosis and treatment of digestive tract diseases, such as gastritis, gastric ulcer, colon cancer, etc. Digestive endoscopy images usually have high resolution and can clearly show the details of the digestive tract mucosa, which are important bases for doctors to make diagnoses and treatments.

[0003] However, generally, traditional digestive endoscopy image diagnosis methods mainly rely on doctors' experience and naked - eye observation. Doctors observe the inside of the digestive tract through the endoscope, and based on the characteristics of the image such as color, texture, and shape, combined with their own professional knowledge and experience, they identify and diagnose lesions. However, this method has the following disadvantages:

[0004] The diagnostic results are greatly affected by doctors' personal experience and knowledge level. Different doctors may give different diagnostic opinions on the same lesion. Doctors need to view the endoscopy images frame by frame to find lesions and classify and diagnose them. This process is time - consuming and laborious, with low efficiency. For small or hidden lesions, it may be difficult to detect them with the naked eye, resulting in missed diagnoses.

[0005] In summary, it is necessary to propose an artificial - intelligence - based digestive endoscopy image - assisted diagnosis system to solve the above problems. Summary of the Invention

[0006] The purpose of the present invention is to provide an artificial - intelligence - based digestive endoscopy image - assisted diagnosis system to solve the problems raised in the above background technique.

[0007] To achieve the above purpose, the present invention provides the following technical solutions:

[0008] An artificial - intelligence - based digestive endoscopy image - assisted diagnosis system includes an image pre - processing module, a small - lesion detection module, a difficult - lesion recognition module, a lesion classification module, a three - dimensional reconstruction module, a lesion change tracking module, an auxiliary decision - making module, and a privacy protection module;

[0009] The image pre - processing module is used for image enhancement and noise removal;

[0010] The small - lesion detection module is used for object detection and lesion feature extraction based on deep learning;

[0011] The difficult - lesion recognition module is used for multi - scale feature fusion and introducing an attention mechanism;

[0012] The lesion classification module is used for fine-grained classification and multi-classifier fusion;

[0013] The 3D reconstruction module is used for 3D reconstruction and virtual reality interaction;

[0014] The lesion change tracking module is used for time series analysis and change detection;

[0015] The auxiliary decision-making module is used for generating diagnostic suggestions and querying knowledge graphs;

[0016] The privacy protection module is used for data encryption and access control.

[0017] Preferably, the image preprocessing module further includes an image enhancement unit and a noise removal unit;

[0018] The image enhancement unit enhances the image contrast through an adaptive histogram equalization algorithm to improve the visibility of image details;

[0019] The noise removal unit uses a Gaussian filtering algorithm to remove random noise in the image, improve the image quality, and lay a foundation for subsequent processing.

[0020] Preferably, the micro-lesion detection module further includes an object detection unit and a lesion feature extraction unit;

[0021] The object detection unit uses the YOLO algorithm to quickly locate micro-lesions in the image to improve the detection efficiency;

[0022] The lesion feature extraction unit uses a convolutional neural network to extract the texture and shape features of the lesion to provide key information for classification.

[0023] Preferably, the difficult lesion recognition module further includes a multi-scale feature fusion unit and an attention mechanism unit;

[0024] The multi-scale feature fusion unit combines pyramid pooling to capture lesion features at different scales and improve the recognition rate of difficult lesions;

[0025] The attention mechanism unit introduces channel attention and spatial attention mechanisms to focus on the key regions of the lesion and ignore the irrelevant background.

[0026] Preferably, the lesion classification module further includes a fine-grained classification unit and a multi-classifier fusion unit;

[0027] The fine-grained classification unit uses a fine-grained classification method to finely classify the lesion and distinguish different types of lesions;

[0028] The multi-classifier fusion unit integrates the results of multiple classifiers through an ensemble random forest to improve the classification accuracy.

[0029] Preferably, the three-dimensional reconstruction module further includes a three-dimensional reconstruction unit and a virtual reality interaction unit;

[0030] The three-dimensional reconstruction unit performs three-dimensional reconstruction of the lesion using the ray casting method to provide intuitive spatial structure information;

[0031] The virtual reality interaction unit combines VR technology to realize the interactive operation between the doctor and the three-dimensional lesion model, facilitating the doctor to observe the lesion from different angles.

[0032] Preferably, the implementation steps of the three-dimensional reconstruction unit are as follows:

[0033] S5.1. Perform preprocessing of two-dimensional endoscopic images and extraction of the lesion area. Perform denoising and enhancement preprocessing operations on the two-dimensional endoscopic images, use threshold segmentation to extract the lesion area, introduce the U-Net deep learning algorithm for automatic recognition and precise segmentation of the lesion area, learn the complex features of the lesion, and improve the accuracy and robustness of the segmentation;

[0034] S5.2. Stack the two-dimensional lesion areas into three-dimensional volume data. The extracted two-dimensional lesion areas are stacked into three-dimensional volume data according to the shooting order and position information of the endoscopic images. Use deep learning algorithms for automated processing of image registration and calibration. Learn the transformation relationship between images by training neural networks to achieve precise alignment and stacking of the lesion areas, reducing manual intervention and errors;

[0035] S5.3. Render the three-dimensional volume data by the ray casting method. Simulate the propagation process of light in three-dimensional space, calculate the intersection points of each ray with the three-dimensional volume data, and dynamically adjust the sampling rate and step size of the ray according to the density and complexity of the three-dimensional volume data. Increase the sampling points in the dense lesion area to improve the rendering accuracy;

[0036] S5.4. Use GPU parallel computing to improve the calculation speed of the ray casting method. Utilize the parallel processing ability of the GPU to calculate the intersection points of multiple rays simultaneously, shorten the rendering time, reduce the sampling points in the sparse lesion area to improve the rendering efficiency, and generate the final three-dimensional image according to the color and transparency information of the intersection points;

[0037] S5.5. Smooth the generated three-dimensional lesion model to remove burrs and noise on the surface, improve the visual effect and accuracy of the model, map the texture information in the two-dimensional endoscopic image to the surface of the three-dimensional lesion model, enhance the realism and detail performance of the model, and provide interactive editing tools that allow doctors to manually adjust the three-dimensional lesion model, including correcting the shape and adjusting the color, to meet clinical needs.

[0038] Preferably, the lesion change tracking module further includes a time series analysis unit and a change detection unit;

[0039] The time series analysis unit uses a long short-term memory network to perform time series analysis on consecutive endoscopic images and track the changing trends of lesions;

[0040] The change detection unit detects the size and morphological changes of lesions through an image difference algorithm, providing a basis for formulating treatment plans.

[0041] Preferably, the auxiliary decision-making module further includes a diagnosis recommendation generation unit and a knowledge graph query unit;

[0042] The diagnosis recommendation generation unit generates diagnosis recommendations based on the classification and change information of lesions, combined with clinical guidelines, to assist doctors in decision-making;

[0043] The knowledge graph query unit constructs a knowledge graph in the field of digestive endoscopy, provides relevant medical knowledge query services, and supports doctors' learning and research.

[0044] Preferably, the privacy protection module further includes a data encryption unit and an access control unit;

[0045] The data encryption unit uses the Advanced Encryption Standard to encrypt and store image data and patient information to ensure data security;

[0046] The access control unit manages the permissions of system users based on the role-based access control model to prevent unauthorized access and data leakage.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows: By automatically analyzing and identifying endoscopic images, the present invention reduces the subjective influence of human factors, improves the objectivity and accuracy of diagnosis, can quickly process a large number of endoscopic images, automatically locate lesions, extract features, classify and diagnose, improves the diagnostic efficiency, through technologies such as deep learning and multi-scale feature fusion, the system can capture small or hidden lesions, reduce the occurrence of missed diagnoses, can realize three-dimensional reconstruction and virtual reality interaction of lesions, provide intuitive spatial structure information for doctors, facilitate observing lesions from different angles, improve the accuracy of diagnosis, at the same time, the present invention also combines clinical guidelines and medical knowledge bases to generate diagnosis recommendations, provides comprehensive decision-making support for doctors, improves the accuracy and scientific nature of diagnosis, encrypts and stores image data and patient information and performs access control to ensure data security and protect patient privacy. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Shows the topology diagram of the artificial intelligence-based digestive endoscopy image auxiliary diagnosis system of the present invention;

[0049] Figure 2 Shows the flowchart of the artificial intelligence-based digestive endoscopy image auxiliary diagnosis method of the present invention. Detailed implementation manners

[0050] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0051] Embodiment 1

[0052] Please refer to Figure 1 , the present invention provides a digestive endoscopy image-assisted diagnosis system based on artificial intelligence, including an image preprocessing module, a micro-lesion detection module, a difficult-lesion recognition module, a lesion classification module, a three-dimensional reconstruction module, a lesion change tracking module, an auxiliary decision-making module, and a privacy protection module;

[0053] Among them, the image preprocessing module of the present invention is used for image enhancement and noise removal, the micro-lesion detection module is used for object detection based on deep learning and lesion feature extraction, the difficult-lesion recognition module is used for multi-scale feature fusion and introducing an attention mechanism, the lesion classification module is used for fine-grained classification and multi-classifier fusion, the three-dimensional reconstruction module is used for three-dimensional reconstruction and virtual reality interaction, the lesion change tracking module is used for time series analysis and change detection, the auxiliary decision-making module is used for generating diagnostic suggestions and querying a knowledge graph, and the privacy protection module is used for data encryption and access control.

[0054] In this embodiment, it should also be noted that the image preprocessing module of the present invention further includes an image enhancement unit and a noise removal unit. The image enhancement unit enhances the image contrast through an adaptive histogram equalization algorithm to improve the visibility of image details. The noise removal unit uses a Gaussian filtering algorithm to remove random noise in the image and improve the image quality, laying a foundation for subsequent processing.

[0055] In this embodiment, it should also be noted that the micro-lesion detection module of the present invention further includes an object detection unit and a lesion feature extraction unit. The object detection unit uses the YOLO algorithm to quickly locate micro-lesions in the image to improve the detection efficiency. The lesion feature extraction unit uses a convolutional neural network to extract the texture and shape features of the lesions, providing key information for classification.

[0056] In this embodiment, it should also be noted that the difficult lesion recognition module of the present invention further includes a multi-scale feature fusion unit and an attention mechanism unit. The multi-scale feature fusion unit combines pyramid pooling to capture lesion features at different scales, improving the recognition rate of difficult lesions. The attention mechanism unit introduces channel attention and spatial attention mechanisms, focusing on the key areas of the lesions and ignoring irrelevant backgrounds.

[0057] In this embodiment, it should also be noted that the lesion classification module of the present invention further includes a fine-grained classification unit and a multi-classifier fusion unit. The fine-grained classification unit uses a fine-grained classification method to finely classify the lesions and distinguish different types of lesions. The multi-classifier fusion unit improves the classification accuracy by integrating the results of multiple classifiers through a random forest.

[0058] In this embodiment, it should also be noted that the 3D reconstruction module of the present invention further includes a 3D reconstruction unit and a virtual reality interaction unit. The 3D reconstruction unit uses the ray casting method to perform 3D reconstruction on the lesions, providing intuitive spatial structure information. The virtual reality interaction unit combines VR technology to realize the interactive operation between the doctor and the 3D lesion model, facilitating the doctor to observe the lesions from different angles.

[0059] In this embodiment, it should also be noted that the implementation steps of the 3D reconstruction unit of the present invention are as follows:

[0060] S5.1. Perform preprocessing of two-dimensional endoscopic images and extraction of lesion regions. Perform denoising and enhancement preprocessing operations on the two-dimensional endoscopic images, use threshold segmentation to extract the lesion regions, and introduce the U-Net deep learning algorithm for automatic recognition and precise segmentation of the lesion regions, learning the complex features of the lesions to improve the accuracy and robustness of the segmentation.

[0061] S5.2. Stack the two-dimensional lesion regions into three-dimensional volume data. The extracted two-dimensional lesion regions are stacked into three-dimensional volume data according to the shooting order and position information of the endoscopic images. Use a deep learning algorithm to perform automated processing of image registration and calibration, learn the transformation relationship between images through training a neural network, and achieve precise alignment and stacking of the lesion regions, reducing manual intervention and errors.

[0062] S5.3. Render the three-dimensional volume data through the ray casting method, simulate the propagation process of light in three-dimensional space, calculate the intersection points of each ray with the three-dimensional volume data, and dynamically adjust the sampling rate and step size of the rays according to the density and complexity of the three-dimensional volume data, increasing the sampling points in the lesion-dense areas to improve the rendering accuracy.

[0063] S5.4. Improve the calculation speed of the ray casting method by using GPU parallel computing. Utilize the parallel processing ability of the GPU to calculate the intersection points of multiple rays simultaneously, shorten the rendering time, reduce the sampling points in the sparse lesion area, improve the rendering efficiency, and generate the final three-dimensional image based on the color and transparency information of the intersection points;

[0064] S5.5. Smooth the generated three-dimensional lesion model to remove burrs and noise on the surface, improve the visual effect and accuracy of the model, map the texture information in the two-dimensional endoscope image to the surface of the three-dimensional lesion model, enhance the realism and detail performance of the model, and provide an interactive editing tool that allows doctors to manually adjust the three-dimensional lesion model, including correcting the shape and adjusting the color, to meet clinical needs.

[0065] In this embodiment, it should also be noted that the lesion change tracking module of the present invention further includes a time series analysis unit and a change detection unit. The time series analysis unit uses a long short-term memory network to perform time series analysis on continuous endoscope images to track the change trend of the lesion. The change detection unit detects the size and morphological changes of the lesion through an image difference algorithm, providing a basis for formulating treatment plans.

[0066] In this embodiment, it should also be noted that the auxiliary decision-making module of the present invention further includes a diagnosis recommendation generation unit and a knowledge graph query unit. The diagnosis recommendation generation unit generates diagnosis recommendations based on the classification and change information of the lesion, combined with clinical guidelines, to assist doctors in making decisions. The knowledge graph query unit constructs a knowledge graph in the field of digestive endoscopy to provide relevant medical knowledge query services to support doctors' learning and research.

[0067] In this embodiment, it should also be noted that the privacy protection module of the present invention further includes a data encryption unit and an access control unit. The data encryption unit uses the Advanced Encryption Standard to encrypt and store image data and patient information to ensure data security. The access control unit manages the permissions of system users based on the role-based access control model to prevent unauthorized access and data leakage.

[0068] Embodiment 2

[0069] Please refer to Figure 2 , in the actual application process of the artificial intelligence-based digestive endoscopy image auxiliary diagnosis system, it specifically includes the following steps:

[0070] S1: Image preprocessing

[0071] S1.1. Conduct a preliminary inspection on the obtained digestive endoscopy images to ensure that the image quality meets the requirements for subsequent processing;

[0072] S1.2. Apply the adaptive histogram equalization algorithm to enhance the contrast of the image to highlight the details of the lesions;

[0073] S1.3. Use the Gaussian filtering algorithm to denoise the image and reduce the influence of random noise on subsequent analysis;

[0074] S2: Micro-lesion detection

[0075] S2.1. Adopt the YOLO algorithm to quickly scan the preprocessed image and locate the possible micro-lesion areas;

[0076] S2.2. Crop the located lesion areas to extract the region of interest (ROI);

[0077] S2.3. Use a convolutional neural network to extract features from the ROI, including key features of texture and shape, as input for subsequent classification;

[0078] S3: Difficult lesion identification

[0079] S3.1. Combine the pyramid pooling technique to perform multi-scale fusion on the extracted lesion features to capture lesion information at different scales;

[0080] S3.2. Introduce channel attention and spatial attention mechanisms to perform weighted processing on the fused features and focus on the key areas of the lesions;

[0081] S3.3. Identify difficult lesions based on the weighted features to improve the recognition accuracy and robustness;

[0082] In this step, specifically, in the identification of difficult lesions, the sizes and shapes of the lesions are diverse, and it is difficult to comprehensively capture the lesion information by single-scale feature extraction. Therefore, this unit uses the Pyramid Pooling Module (PPM) to achieve multi-scale feature fusion, as shown in Equation (1):

[0083] F multiscale = PPM(F input ) (1);

[0084] In the formula, F input represents the input feature map, PPM represents the pyramid pooling operation, and F multiscale represents the fused multi-scale features. The pyramid pooling module performs pooling operations on the input feature map at different scales to obtain multiple-scale feature representations, and then splices or fuses these feature representations to obtain the feature map F multiscale, in this way, even if the sizes and shapes of the lesions are different, their key features can be captured at multiple scales, improving the recognition rate. To further improve the recognition accuracy of difficult lesions, this unit introduces channel attention and spatial attention mechanisms, as shown in Equation (2):

[0085] F attention = σ(W 2 ·δ(W 1 ·F multiscale ))·F multiscale (2);

[0086] In the formula, F multiscale represents the multi-scale feature map, W 1 and W 2 represent the weights of the fully connected layers, δ represents the ReLU activation function, σ represents the Sigmoid activation function, F attention represents the feature map after adding the attention mechanism. The global average pooling is performed on the multi-scale feature map F multiscale to obtain the global features of each channel. Then, through two layers of fully connected layers and activation functions, the weights of each channel are learned. Finally, the weights are multiplied by the original feature map to obtain the feature map F attention after adding the channel attention mechanism. The spatial attention mechanism is similar, but it operates in the spatial dimension. By introducing the attention mechanism, it is possible to focus on the key regions of the lesions, ignore the irrelevant background, and improve the recognition accuracy;

[0087] S4: Lesion classification

[0088] S4.1. Adopt a fine-grained classification method to perform fine classification on the recognized lesions to distinguish different types of lesions;

[0089] S4.2. Train multiple classifiers, integrate the classification results of each classifier, and fuse them through the random forest algorithm;

[0090] S4.3. Determine the final type of the lesion according to the fused classification results, providing a basis for subsequent diagnosis and treatment;

[0091] S5: 3D reconstruction and virtual reality interaction

[0092] S5.1. Perform preprocessing of 2D endoscopic images and extraction of lesion regions, and use the U-Net deep learning algorithm for automatic recognition and accurate segmentation of lesion regions;

[0093] S5.2. Stack the 2D lesion regions into 3D volume data, and use deep learning algorithms for automatic processing of image registration and calibration;

[0094] S5.3. Render the three-dimensional volume data by ray casting method, and dynamically adjust the sampling rate and step length of the ray according to the density and complexity of the three-dimensional volume data;

[0095] S5.4. Use GPU parallel computing to improve the calculation speed of the ray casting method and generate the final three-dimensional image;

[0096] S5.5. Smooth the generated three-dimensional lesion model and perform texture mapping, and provide interactive editing tools for doctors to use;

[0097] S6: Lesion change tracking

[0098] S6.1. Perform time series analysis on consecutive endoscopic images, and use long short-term memory network to track the change trend of the lesion;

[0099] S6.2. Use the image difference algorithm to compare adjacent images in the time series and detect the size and morphological changes of the lesion;

[0100] S6.3. Generate a lesion change report according to the change detection results, and provide a basis for doctors to formulate treatment plans;

[0101] S7: Auxiliary decision-making and knowledge graph query

[0102] S7.1. Based on the classification and change information of the lesion, combined with clinical guidelines and medical knowledge bases, generate diagnostic suggestions;

[0103] S7.2. Construct a knowledge graph in the field of digestive endoscopy, provide medical knowledge query services, and support doctors' learning and research;

[0104] S7.3. Integrate the diagnostic suggestions and medical knowledge into the auxiliary decision-making system to provide comprehensive decision-making support for doctors;

[0105] S8: Privacy protection

[0106] S8.1. Encrypt and store the image data and patient information, and use the Advanced Encryption Standard to ensure data security;

[0107] S8.2. Establish an access control mechanism, and manage the permissions of system users based on the role-based access control model;

[0108] S8.3. Regularly conduct security audits and vulnerability scans on the system to ensure the security and stability of the system;

[0109] In summary, the artificial intelligence-based digestive endoscopy image-assisted diagnosis system can overcome the shortcomings of traditional diagnostic methods, improve the accuracy, efficiency and scientific nature of diagnosis, and provide strong support for the diagnosis and treatment of digestive tract diseases.

[0110] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. The digestive endoscopy image-assisted diagnosis system based on artificial intelligence is characterized by: It includes image preprocessing module, micro-lesion detection module, difficult lesion recognition module, lesion classification module, 3D reconstruction module, lesion change tracking module, decision support module and privacy protection module; The image preprocessing module is used for image enhancement and noise removal; The micro-lesion detection module is used for target detection and lesion feature extraction based on deep learning; The difficult lesion recognition module is used for multi-scale feature fusion and introduces an attention mechanism; The lesion classification module is used for fine-grained classification and multi-classifier fusion; The three-dimensional reconstruction module is used for three-dimensional reconstruction and virtual reality interaction; The lesion change tracking module is used for time series analysis and change detection; The auxiliary decision module is used for diagnosis suggestion generation and knowledge graph query; The privacy protection module is used for data encryption and access control.

2. The AI-based digestive endoscopy image-assisted diagnosis system according to claim 1, characterized in that: The image preprocessing module also includes an image enhancement unit and a noise removal unit; The image enhancement unit performs image contrast enhancement by an adaptive histogram equalization algorithm to improve visibility of image details; The noise removal unit removes random noise in the image using a Gaussian filtering algorithm.

3. The AI-based digestive endoscopy image-assisted diagnosis system according to claim 2 is characterized in that: The micro-lesion detection module also includes a target detection unit and a lesion feature extraction unit; The target detection unit uses the YOLO algorithm to quickly locate tiny lesions in the image; The lesion feature extraction unit uses a convolutional neural network to extract the texture and shape features of the lesion.

4. The AI-based digestive endoscopy image-assisted diagnosis system according to claim 3 is characterized by: The difficult lesion recognition module also includes a multi-scale feature fusion unit and an attention mechanism unit; The multi-scale feature fusion unit combines pyramid pooling to capture lesion features at different scales; The attention mechanism unit introduces channel attention and spatial attention mechanisms to focus on the key areas of the lesion and ignore irrelevant background.

5. The AI-based digestive endoscopy image-assisted diagnosis system according to claim 4 is characterized in that: The lesion classification module also includes a fine-grained classification unit and a multi-classifier fusion unit; The fine-grained classification unit adopts a fine-grained classification method to finely classify the lesions and distinguish different types of lesions; The multi-classifier fusion unit fuses the results of multiple classifiers by integrating random forests.

6. The AI-based digestive endoscopy image-assisted diagnosis system according to claim 5 is characterized by: The three-dimensional reconstruction module also includes a three-dimensional reconstruction unit and a virtual reality interaction unit; The three-dimensional reconstruction unit uses a ray projection method to perform three-dimensional reconstruction of the lesion to provide intuitive spatial structure information; The virtual reality interaction unit combines VR technology to realize the interactive operation between the doctor and the three-dimensional lesion model.

7. The AI-based digestive endoscopy image-assisted diagnosis system according to claim 6 is characterized in that: The implementation steps of the 3D reconstruction unit are as follows: S5.

1. Perform 2D endoscopic image preprocessing and lesion area extraction. Perform denoising and enhancement preprocessing operations on 2D endoscopic images, extract lesion areas using threshold segmentation, introduce U-Net deep learning algorithm for automatic identification and precise segmentation of lesion areas, and learn the complex characteristics of lesions. S5.

2. Stack the two-dimensional lesion areas into three-dimensional volume data. The extracted two-dimensional lesion areas are stacked into three-dimensional volume data according to the shooting order and position information of the endoscopic images. A deep learning algorithm is used to automate the image registration and calibration. The transformation relationship between images is learned by training a neural network to achieve accurate alignment and stacking of lesion areas. S5.

3. Render the 3D volume data by ray casting, simulate the propagation of light in 3D space, calculate the intersection of each ray with the 3D volume data, dynamically adjust the sampling rate and step size of the ray according to the density and complexity of the 3D volume data, and increase the sampling points in the lesion-dense area; S5.

4. Use GPU parallel computing to improve the calculation speed of the ray casting method. Utilize the parallel processing capability of GPU to simultaneously calculate the intersection points of multiple rays, reduce the sampling points in sparse lesion areas, and generate the final three-dimensional image based on the color and transparency information of the intersection points; S5.

5. Smooth the generated three-dimensional lesion model to remove surface burrs and noise, map the texture information in the two-dimensional endoscopic image to the surface of the three-dimensional lesion model, and provide interactive editing tools to allow doctors to manually adjust the three-dimensional lesion model, including correcting the shape and adjusting the color.

8. The AI-based digestive endoscopy image-assisted diagnosis system according to claim 7, characterized in that: The lesion change tracking module also includes a time series analysis unit and a change detection unit; The time series analysis unit uses a long short-term memory network to perform time series analysis on continuous endoscopic images to track the changing trend of lesions; The change detection unit detects the size and shape changes of the lesion through an image difference algorithm, providing a basis for formulating a treatment plan.

9. The AI-based digestive endoscopy image-assisted diagnosis system according to claim 8, characterized in that: The auxiliary decision module also includes a diagnosis suggestion generation unit and a knowledge graph query unit; The diagnostic suggestion generating unit generates diagnostic suggestions based on the classification and change information of the lesions and in combination with clinical guidelines to assist doctors in making decisions; The knowledge graph query unit constructs a knowledge graph in the field of digestive endoscopy, provides relevant medical knowledge query services, and supports doctors' learning and research.

10. The AI-based digestive endoscopy image-assisted diagnosis system according to claim 9, characterized in that: The privacy protection module also includes a data encryption unit and an access control unit; The data encryption unit uses the advanced encryption standard to encrypt and store image data and patient information to ensure data security; The access control unit performs authority management on system users based on a role-based access control model to prevent unauthorized access and data leakage.

Images