Stomach image intelligent detection system and method based on YOLOv8

Abstract

The invention relates to the technical field of medical image analysis, in particular to a stomach image intelligent detection system and method based on YOLOv8, and aims to realize efficient and accurate auxiliary detection of early gastric cancer (EGC). By introducing an improved YOLOv8 (EGC-YOLOv8) network model, the system can accurately identify and analyze minimal lesion characteristics in gastroscope images, such as mucous membrane color change, texture abnormity, minimal upheaval or depression and the like, and the accuracy of auxiliary detection is remarkably improved. Meanwhile, the system adopts a video frame-by-frame analysis technology to ensure that each frame of image is subjected to fine detection, so that the problem that missed diagnosis or misdiagnosis is easy to occur in a traditional detection means is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and in particular to a gastric image intelligent detection system and method based on YOLOv8. Background Art

[0002] Gastric cancer, as a common malignant tumor, its early detection and treatment are crucial to improving the survival rate of patients. However, traditional gastric cancer screening methods such as gastroscopy and tissue biopsy are not only cumbersome, but may also cause discomfort to patients. In addition, due to the uneven distribution of medical resources, it is difficult to achieve large-scale popularization and timely screening. In addition, the early symptoms of gastric cancer are often not obvious and easy to be ignored, causing many patients to miss the best time for treatment. Therefore, the development of an efficient, non-invasive, and easy-to-popularize intelligent detection system for early gastric cancer has become the key to improving the level of gastric cancer prevention and treatment.

[0003] Limitations and challenges of existing technologies: The application of traditional image recognition technology in the early detection of gastric cancer is limited by image quality, lighting conditions and subtle differences in lesion characteristics, resulting in low detection accuracy and prone to missed or misdiagnosis; analysis systems based on a single data source (such as gastroscopy images) are difficult to fully capture lesion information, especially the lack of recognition capabilities for tiny lesions; existing intelligent detection systems mostly rely on manual annotation and review by experts, which is not only time-consuming and labor-intensive, but also limited by the scarcity of expert resources, making it difficult to achieve large-scale and efficient screening; in addition, patients' psychological fear and physical discomfort of gastroscopy are also one of the important factors hindering the popularization of early gastric cancer screening. Summary of the invention

[0004] The purpose of the present invention is to provide a gastric image intelligent detection system and method based on YOLOv8, aiming to solve the problem of low detection accuracy of traditional detection methods.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a gastric image intelligent detection method based on YOLOv8, comprising the following steps:

[0006] Establish database;

[0007] Based on the YOLOv8 model, the EGC detection technology was developed and packaged to obtain the detection system;

[0008] Set the gastroscope video range to import into the detection system, and analyze and mark key frames frame by frame;

[0009] Perform in-depth analysis on key frames to locate suspected EGC lesions and obtain coordinate information;

[0010] The detection system identifies and locates suspected lesions and transmits the identification results back to the host computer for the doctor's reference.

[0011] Among them, in "Establishing a Database", the database includes 2902 EGC images and gastroscopic clinical examination videos of 50 EGC patients, and the video frame rate is 30FPS.

[0012] Among them, in "setting the gastroscope video range to import the detection system, and analyzing and marking key frames frame by frame", it includes:

[0013] The video dataset of possible or confirmed early gastric cancer lesions obtained during the consultation is edited and imported into the detection system;

[0014] The detection system performs frame-by-frame analysis based on the imported video data to obtain the video area where the suspected lesions exist, and performs frame sampling detection in the video area to obtain reliable detection results, and saves the video with the suspected lesions.

[0015] Among them, "deep analysis of key frames to locate suspected EGC lesions and obtain coordinate information" includes: deep analysis of key frames, using the detection system to locate the suspected EGC lesions from the gastroscopy video, and obtaining the coordinate information of these lesions relative to the video frame.

[0016] Among them, in "the detection system identifies and locates suspected lesions, and transmits the identification results back to the host computer for the doctor's reference", it includes: the detection system identifies the detected suspected EGC lesions, and transmits the coordinate information of the lesions, the identification results and the on-site gastroscopy images back to the host computer in real time for the doctor's reference.

[0017] In the second aspect, a gastric image intelligent detection system based on YOLOv8 adopts the gastric image intelligent detection method based on YOLOv8 described in the first aspect, including Ghostconv, C3 module, Horblock module and CARAFE upsampling operator.

[0018] The gastric image intelligent detection method based on YOLOv8 of the present invention comprises the following steps: establishing a database; improving based on the YOLOv8 model, developing EGC detection technology and encapsulating it to obtain a detection system; setting the gastroscope video range to import into the detection system, and analyzing and marking key frames frame by frame; deeply analyzing the key frames to locate the suspected EGC lesion position and obtain coordinate information; the system automatically classifies the suspected lesions, evaluates the malignancy, and transmits it back to the host computer for reference by the doctor. The present invention can achieve efficient, accurate and real-time detection of early gastric cancer. By introducing the improved YOLOv8 (EGC-YOLOv8) network model, the system can accurately identify and analyze the characteristics of microlesions in gastroscope images, such as changes in mucosal color, abnormal texture, and micro-protrusions or depressions, etc., significantly improving the accuracy of auxiliary detection. At the same time, the system adopts video frame-by-frame analysis technology to ensure that each frame of the image is finely detected, effectively overcoming the problem of easy missed diagnosis or misdiagnosis in traditional detection methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 This is a schematic diagram of the C3 structure.

[0021] Figure 2 This is a schematic diagram of the C2f structure.

[0022] Figure 3 It is the detection interface of the early gastric cancer detection system.

[0023] Figure 4 It is the result of early gastric cancer detection.

[0024] Figure 5 This is an example of an image from the EGC dataset.

[0025] Figure 6 It is the comparison of PR curves of the two models and the corresponding AP values.

[0026] Figure 7 The following are examples of test results of the two models on the self-built dataset: (a) Both models successfully detect EGC, and the confidence of the EGC-YOLOv8 model is higher; (b) YOLOv8 detects the presence of FN; (c) YOLOv8 detects the presence of FN and FP; (d) YOLOv8 detects the presence of FP; (e) EGC-YOLOv8 detects the presence of FP.

[0027] Figure 8 It is a flow chart of the gastric image intelligent detection method based on YOLOv8 provided by the present invention. DETAILED DESCRIPTION

[0028] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0029] See also Figures 1 to 8 In a first aspect, the present invention provides a gastric image intelligent detection method based on YOLOv8, comprising the following steps:

[0030] S1 establishes a database;

[0031] The database includes 2902 EGC images and gastroscopic clinical examination videos of 50 EGC patients, and the video frame rate is 30FPS.

[0032] Specifically, we continuously improve and supplement our own database with a large amount of medical image data provided by multiple hospitals. In order to apply the data set to the deep learning model, we carefully select the data set and annotate it by professional doctors.

[0033] To ensure the accuracy of the data, all EGC image samples were manually inspected and annotated by experienced endoscopists, and the relevant lesions were confirmed as EGC by clinical biopsy to ensure the high quality of the dataset and the rigor of the experiment. Figure 5 Some EGC images in the dataset are shown. The materials collected for constructing the EGC dataset include 2902 EGC images and 50 gastroscopy clinical examination videos of EGC patients, all with a video frame rate of 30FPS. In terms of data annotation, a labeling working group was formed, consisting of a director-level expert, several professional endoscopists with rich professional experience, and myself. Specifically, several professional endoscopists and myself are responsible for cooperating to complete the labeling of EGC, and then the experts review and errata until all the labeled data meet the experimental requirements to ensure the quality of the annotation. In order to be compatible with the YOLO model, the image annotation information is converted into the YOLO dataset format.

[0034] S2 improves the YOLOv8 model, develops and encapsulates EGC detection technology, and obtains a detection system;

[0035] Clip the video dataset of early gastric cancer lesions that may exist or have been diagnosed during the visit and import it into the detection system; the detection system performs frame-by-frame analysis on the imported video data to obtain the video area where suspected lesions exist, and performs frame extraction detection in the video area to obtain reliable detection results, and saves the videos with suspected lesions.

[0036] Specifically, based on the data provided by the professional medical team and the lesion area, improve the YOLOv8 model and continuously test it to achieve better detection effects. After the technology development is completed, package the system and deliver it to the medical team for third-party verification.

[0037] Introduce Ghostconv and C3 modules: The time complexity of the traditional convolutional layer is O(H×W×Cin×Cout×k2), where H, W, and Cin represent the height, width, and number of channels of the input feature map respectively, Cout is the number of output channels, and k is the convolutional kernel size. In contrast, GhostConv adopts a two-step strategy: First, it calculates a set of fewer intrinsic feature maps, and its time complexity is reduced to O(H×W×Cin×(r·Cout)×k2), where r is the ratio of the intrinsic feature maps (0 < r < 1); subsequently, through a low-complexity linear transformation, the remaining phantom feature maps are derived from these intrinsic feature maps, and the time complexity of this step is approximately O(H×W×(1 - r)·Cout×Cintermediate), where Cintermediate is the dimension of the intermediate features and is much smaller than the number of output channels Cout. Given the advantages of GhostConv in reducing model parameters and improving computational efficiency, when designing the model for gastric cancer video stream analysis, we choose to replace some standard convolutional modules with GhostConv modules. The aim is to reduce model parameters, improve the operation rate, and at the same time maintain or enhance the auxiliary detection accuracy. In the YOLO series backbone networks, the C3 module uses 3 convolutional blocks and stacks n BottleNecks for feature extraction and control by referring to the ideas of CSPNet's shunt and residual structures, as Figure 1 (where CBS represents a specific convolutional block, S represents the stride, and K represents the kernel). The C2f module in YOLOv8 is designed by referring to the ideas of the C3 module and ELAN, and can provide richer gradient flow information, as Figure 2(Where CBS stands for a specific convolutional block, S stands for stride, and K stands for kernel size). The C2f and C3 modules have their own performance advantages in the model. However, in complex image feature processing tasks such as gastric cancer video streams, the C3 module is more suitable because it can promote rapid convergence of the model and maintain stable performance. Therefore, in the backbone network, some C2f modules are selectively replaced with C3 modules to meet the model performance requirements of complex tasks such as EGC video streams, thereby accelerating the training process and ensuring the robustness of the model.

[0038] Embedding Horblock module

[0039] In image processing technology, traditional convolution operations mainly focus on local features and ignore the long-distance dependencies between pixels, which limits the performance of the model in processing global interactive image tasks. To make up for this shortcoming, we introduce the Horblock module. As an efficient high-order spatial interaction component. This module deeply analyzes the image spatial relationship through the recursive gate convolution mechanism, and its mathematical expression is: Fn = σ (Wn · Fn-1 + bn) where Fn represents the feature map after the nth iteration, Wn and bn represent the convolution kernel and bias term of the nth iteration respectively, and σ represents the activation function, usually nonlinear functions such as ReLU. The core of its improvement lies in integrating the Horblock module into the backbone network of YOLOv8. By iteratively capturing the long-distance dependencies in the image, in each iteration, the model uses the current convolution kernel to extract features and integrates the output through a recursive gated unit to generate global high-order spatial features. This improvement enhances the model's understanding of the global structure of the image and provides more accurate and reliable results for image detection. Although the Horblock module increases the ability to handle long-distance dependencies, its time complexity increase relative to traditional convolution is controllable, and an effective balance between performance and computational overhead can be achieved by adjusting the number of iterations.

[0040] Introduce the CARAFE upsampling operator:

[0041] In the YOLOv8 framework, the feature fusion of the neck network is crucial to improving the model performance. In order to further improve the detection accuracy and performance of the model, we conducted an in-depth study on the YOLOv8 neck network. In the original YOLOv8 framework, traditional upsampling methods such as bilinear interpolation, nearest neighbor interpolation and other simple methods are usually used. Although these methods are computationally efficient and simple to implement, they often fail to fully capture and utilize contextual information in a larger range, which has certain limitations when processing EGC images with complex backgrounds and subtle features. To overcome this limitation, we introduced the advanced CARAFE upsampling operator. The uniqueness of the CARAFE operator lies in the use of a content-aware feature recombination mechanism, which can dynamically aggregate contextual information in a larger receptive field, not just in the neighboring pixels. This means that CARAFE can not only consider the information of neighboring pixels, but also make full use of the information of other related areas far away from the current pixel position. As a result, more accurate and rich feature representations can be generated when sampling EGC image information to reduce the problem of lesion information loss, thereby improving the detection performance of the model.

[0042] S3 sets the gastroscope video range to import into the detection system, and analyzes and marks key frames frame by frame;

[0043] Specifically, the expert team set the gastroscopy video range for early gastric cancer (EGC) detection based on the video data of the consultation outside the self-built database and imported it into the detection system. After the system is started, the EGC-YOLOv8 model begins to analyze the video frame by frame. The system first calculates the video range and marks the key frames, and designs an efficient scanning path to ensure that each key frame of the suspicious lesion part in the video is carefully detected.

[0044] S4 performs in-depth analysis on the key frames to locate the suspected EGC lesion and obtain coordinate information;

[0045] The key frames are deeply analyzed, and the detection system is used to locate the positions of suspected EGC lesions from the gastroscopy video and obtain the coordinate information of these lesions relative to the video frame.

[0046] Specifically, during the video analysis process, the EGC-YOLOv8 model will perform in-depth analysis on the selected key frames. Using the trained EGC-YOLOv8 network, the positions of suspected EGC lesions are located from the gastroscopy video, and the coordinate information of these lesions relative to the video frame is obtained.

[0047] The S5 detection system identifies and locates suspected lesions and transmits the identification results back to the host computer for the doctor's reference.

[0048] The detection system identifies the detected suspected EGC lesions and transmits the coordinate information, identification results, and on-site gastroscope images of the lesions back to the host computer in real time for doctors' reference.

[0049] Specifically, the system automatically classifies the detected suspected EGC lesions, evaluates their malignancy degree, and transmits the coordinate information, classification results, and on-site gastroscope images of the lesions back to the host computer in real time for doctors' reference. Doctors can quickly make a diagnosis and formulate corresponding treatment plans based on the detailed information provided by the system. For multiple detected lesions, the system will also provide suggestions on the priority processing order for doctors according to the malignancy degree and location information.

[0050] In a second aspect, a gastric image intelligent detection system based on YOLOv8 adopts the gastric image intelligent detection method described in the first aspect, including Ghostconv, C3 module, Horblock module, and CARAFE upsampling operator.

[0051] In an implementation, Ghostconv and C3 module are introduced. In the CNN structure, in order to capture rich image feature information, it is usually necessary to stack multiple convolutional layers. However, while this method extracts useful features, it also introduces a large amount of redundant information, resulting in the model requiring more parameters and computing resources to achieve ideal performance. To address this issue, in 2020, Kai Ha et al. proposed a lightweight convolution method called Ghostconv, which generates phantom feature maps with a similar structure but different content from the original feature map through simple linear mapping. To enhance the feature representation ability without significantly increasing the computational cost. Specifically, the time complexity of the traditional convolutional layer is O(H×W×Cin×Cout×k2), where H, W, and Cin represent the height, width, and number of channels of the input feature map respectively, Cout is the number of output channels, and k is the convolutional kernel size. In contrast, GhostConv adopts a two-step strategy: First, it calculates a set of fewer intrinsic feature maps, and its time complexity is reduced to O(H×W×Cin×(r·Cout)×k2), where r is the ratio of the intrinsic feature maps (0 < r < 1); subsequently, through a low-complexity linear transformation, the remaining phantom feature maps are derived from these intrinsic feature maps, and the time complexity of this step is approximately O(H×W×(1 - r)·Cout×Cintermediate), where Cintermediate is the dimension of the intermediate features and is much smaller than the number of output channels Cout. Given the advantages of GhostConv in reducing model parameters and improving computational efficiency, when designing the model for gastric cancer video stream analysis, we choose to replace some standard convolutional modules with GhostConv modules. Aimed at reducing model parameters, improving the operation rate, and maintaining or enhancing the detection accuracy.

[0052] In the YOLO series backbone network, the C3 module uses 3 convolutional blocks and stacks n BottleNecks to extract and control features by borrowing the ideas of the shunt and residual structure of CSPNet. Figure 1 The C2f module in YOLOv8 is designed with reference to the C3 module and the idea of ​​ELAN, and can provide richer gradient flow information, such as Figure 2 . The C2f and C3 modules have their own advantages in model performance. However, in complex image feature processing tasks such as gastric cancer video streams, the C3 module is more suitable because it can promote rapid convergence of the model and maintain stable performance. This is mainly due to the jump connection design used in the C3 module, which not only allows information to flow directly between different layers of the network, but also effectively avoids the common gradient vanishing or gradient exploding problems in deep network training. Therefore, in the backbone network, we selectively replaced some C2f modules with C3 modules to adapt to the model performance requirements of complex tasks such as EGC video streams, thereby accelerating the training process and ensuring the robustness of the model.

[0053] Embedding Horblock module,In image processing technology, traditional convolution operations mainly focus on local features, while ignoring the long-distance dependencies between pixels, which limits the performance of the model in processing global interactive image tasks. To make up for this deficiency, we introduce the Horblock module. As an efficient high-order spatial interaction component. This module deeply analyzes the spatial relationship of the image through the recursive gate convolution mechanism, and its mathematical expression is:

[0054] F n =σ(W n ·F n-1 +b n)

[0055] Among them, Fn represents the feature map after the nth iteration, Wn and bn represent the convolution kernel and bias term of the nth iteration respectively, and σ represents the activation function, usually a nonlinear function such as ReLU is selected.

[0056] The core of the improvement is to integrate the Horblock module into the backbone network of YOLOv8. The long-distance dependencies in the image are captured by iteration. In each iteration, the model uses the current convolution kernel to extract features and integrates the output through a recursive gating unit to generate global high-order spatial features. This improvement enhances the model's understanding of the global structure of the image and provides more accurate and reliable results for image detection. Although the Horblock module increases the ability to handle long-distance dependencies, its time complexity is controllable compared to the increase in traditional convolution, and an effective balance between performance and computational overhead can be achieved by adjusting the number of iterations.

[0057] Introducing the CARAFE upsampling operator, in the YOLOv8 framework, the feature fusion of the neck network is crucial to improving the model performance. In order to further improve the detection accuracy and performance of the model, the YOLOv8 neck network is studied in depth. In the original YOLOv8 framework, traditional upsampling methods such as bilinear interpolation, nearest neighbor interpolation and other simple methods are usually used. Although these methods are computationally efficient and simple to implement, they often cannot fully capture and utilize contextual information in a larger range, which has certain limitations when processing EGC images with complex backgrounds and subtle features. To overcome this limitation, the advanced CARAFE upsampling operator is introduced. The uniqueness of the CARAFE operator lies in the use of a content-aware feature recombination mechanism, which can dynamically aggregate contextual information in a larger receptive field, not just in the neighboring pixels. This means that CARAFE can not only consider the information of neighboring pixels, but also make full use of the information of other related areas far away from the current pixel position. Therefore, when sampling EGC image information, a more accurate and rich feature representation can be generated to reduce the problem of lesion information loss, thereby improving the detection performance of the model.

[0058] To ensure an objective and comprehensive evaluation of the model performance, a series of standard object detection-related evaluation indicators are adopted, including precision, recall, average precision (AP) and FPS. These indicators are calculated based on the actual value of the actual category and the value predicted by the model, and the sample detection results are counted and summarized. There are four key elements involved: True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN).

[0059] Precision: represents the ratio of true positive examples to the total number of examples predicted as positive examples. It measures the accuracy of positive example prediction. The formula is as shown in formula (1).

[0060] Recall: represents the ratio of true positive examples to the total number of examples that are actually positive. It measures the ability of the model to identify all relevant examples. The formula is as shown in formula (2).

[0061] AP: used to measure the accuracy of the model at different recall rates. The formula is as follows: (3), where p(r) is the precision value at the recall rate r.

[0062] FPS: indicates the number of frames processed per second by the model, which is an indicator of the model's inference speed. The higher the FPS value, the faster the model processing speed. In this experiment, the calculation of FPS involves three key time stages: preprocessing time, inference time, and postprocessing time. The time units are all milliseconds. The sum of the time of these three stages constitutes the total time required to process one frame of image. The formula is as follows (4):

[0063]

[0064] The collected EGC images may have problems such as noise, artifacts, and occlusion. In order to reduce the impact of these adverse factors on the research, image filtering technology is used to reduce these interferences. At the same time, in order to enhance the generalization ability of the model, the training set images were subjected to data enhancement processing, including rotation, flipping, translation, etc. In terms of data annotation, a labeling working group was formed with the participation of a director-level expert, several professional endoscopists with rich professional experience, and myself. Specifically, several professional endoscopists and myself are responsible for cooperating to complete the labeling of EGC, and then the experts review and errata until all the labeled data meet the experimental requirements to ensure the quality of the labeling. In order to be compatible with the YOLO model, the image annotation information is converted into the YOLO dataset format. Finally, it is divided into training set, validation set, and test set according to the ratio of 7:1:2. The specific number of image allocations is shown in Table 1.

[0065] Table 1 Data image number

[0066]

[0067] During the model training process, in order to improve memory usage efficiency and speed up training, the method of batching the dataset images was adopted. By dynamically adjusting the learning rate, the parameter update speed in the gradient descent algorithm is finely regulated, and the optimal solution of the parameters is approached quickly and accurately during the training process, thereby improving the training efficiency. The specific training configuration is as follows: The entire dataset is iteratively trained for 300 rounds to ensure the stability and accuracy of the model. During the training process, we set the initial learning rate to 0.01 and used a weight decay of 0.0005 to prevent the model from overfitting and improve its generalization ability. The specific hardware configuration environment is shown in Table 2.

[0068] Table 2 Experimental environment configuration

[0069]

[0070] The two models were trained and tested on the self-built EGC dataset. In order to intuitively and quantitatively evaluate the performance of the models, the PR curve was drawn and the AP value was calculated. The results are shown in the figure below. Figure 6 As shown in the figure. The AP value of YOLOv8 is 91.5%, while the AP value of EGC-YOLOv8 is increased to 94.6%. The area enclosed by the PR curve of the latter is larger, which means that the model has achieved a better balance between precision and recall. Compared with the original model, the AP value of EGC-YOLOv8 is improved by 3.1%. This significant improvement fully proves that the improved model has a better performance in target detection accuracy. To further verify the applicability of the proposed model, experimental verification is carried out on a self-built dataset. Figure 7 The paper shows a comparison of some prediction results of two different models when detecting EGC lesions. Through comparative analysis, it can be clearly observed that compared with the original model, EGC-YOLOv8 shows higher performance in detection accuracy.

[0071] Compared with the prior art, the present invention has the following beneficial effects:

[0072] (1) The present invention can achieve efficient, accurate and real-time detection of early gastric cancer. By introducing the improved YOLOv8 (EGC-YOLOv8) network model, the system can accurately identify and analyze the characteristics of microlesions in gastroscopy images, such as changes in mucosal color, abnormal texture, and micro-protrusions or depressions, significantly improving the detection accuracy. At the same time, the system uses video frame-by-frame analysis technology to ensure that each frame of the image is carefully detected, effectively overcoming the problem of missed diagnosis or misdiagnosis in traditional detection methods.

[0073] (2) This system integrates multi-source data and conducts comprehensive analysis and judgment through deep learning algorithms, which improves the comprehensiveness and accuracy of diagnosis. Compared with the analysis system of a single data source, the present invention can capture lesion information more comprehensively, especially the ability to identify small lesions is significantly enhanced. In addition, the system also introduces automated preprocessing and intelligent labeling technology, which greatly reduces manual intervention, improves detection efficiency, and makes large-scale gastric cancer screening possible.

[0074] (3) The present invention uses a user-friendly interface design, so patients can complete preliminary screening without having to go to the hospital in person, which greatly reduces the psychological burden on patients. At the same time, the system's real-time feedback function provides doctors with detailed lesion information and diagnostic suggestions, which helps doctors make a quick diagnosis and formulate corresponding treatment plans. This innovation not only promotes the optimal allocation of medical resources, but also wins precious time for the discovery and treatment of early gastric cancer, and improves the survival rate and quality of life of patients.

[0075] The above disclosure is only a preferred embodiment of the gastric image intelligent detection system and method based on YOLOv8 of the present invention. Of course, it cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiments and equivalent changes made according to the claims of the present invention are still within the scope of the invention.

Claims

1. An intelligent gastric image detection method based on YOLOv8, characterized by: The following steps are involved: Establish database; Based on the YOLOv8 model, the EGC detection technology was developed and packaged to obtain the detection system; Set the gastroscope video range to import into the detection system, and analyze and mark key frames frame by frame; Perform in-depth analysis on key frames to locate suspected EGC lesions and obtain coordinate information; The detection system identifies and locates suspected lesions and transmits the identification results back to the host computer for the doctor's reference.

2. The gastric image intelligent detection method based on YOLOv8 according to claim 1, characterized in that: In "Establishing a Database", the database includes 2902 EGC images and 50 gastroscopic clinical examination videos of EGC patients, and the video frame rate is 30FPS.

3. The gastric image intelligent detection method based on YOLOv8 as claimed in claim 1, characterized in that: In "Setting the gastroscope video range to import into the detection system, and analyzing and marking key frames frame by frame", it includes: The video dataset of possible or confirmed early gastric cancer lesions obtained during the consultation is edited and imported into the detection system; The detection system performs frame-by-frame analysis based on the imported video data to obtain the video area where the suspected lesions exist, and performs frame sampling detection in the video area to obtain reliable detection results, and saves the video with the suspected lesions.

4. The gastric image intelligent detection method based on YOLOv8 according to claim 1, characterized in that: In "performing in-depth analysis on key frames to locate suspected EGC lesions and obtain coordinate information", it includes: performing in-depth analysis on key frames, using a detection system to locate suspected EGC lesions from gastroscopy videos, and obtaining coordinate information of these lesions relative to video frames.

5. The gastric image intelligent detection method based on YOLOv8 according to claim 1, characterized in that: "The detection system identifies and locates suspected lesions, and transmits the identification results back to the host computer for the doctor's reference" includes: the detection system identifies the detected suspected EGC lesions, and transmits the coordinate information of the lesions, the identification results and the on-site gastroscopy images back to the host computer in real time for the doctor's reference.

6. A gastric image intelligent detection system based on YOLOv8, using the gastric image intelligent detection method based on YOLOv8 according to any one of claims 1 to 5, characterized in that: Including Ghostconv, C3 module, Horblock module and CARAFE upsampling operator.

Images