A digestive tract endoscopy auxiliary system based on ensemble learning
Through integrated learning technology and combined with data acquisition, identification and early warning modules, the automatic increase in image acquisition in medium and high-risk areas of endoscopy is achieved, solving the problems of missed diagnosis and misdiagnosis in the existing technology, and improving the inspection efficiency and accuracy.
Patent Information
- Application Number
- CN202310041105.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-01-12
- Publication Date
- 2024-12-10
- Estimated Expiration
- 2043-01-12
AI Technical Summary
The existing endoscopy technology relies on manual diagnosis by doctors, which is prone to missed diagnosis and misdiagnosis due to insufficient image acquisition, and cannot automatically increase the number of image acquisitions in high-risk areas, and lacks real-time early warning and efficient assistance.
Adopt the digestive tract endoscopy auxiliary system based on integrated learning, including data acquisition module, identification module, early warning module and image acquisition control module. Through the abnormal analysis model, the suspected target in the image is judged in real time and the warning signal is issued to increase the images in high-risk areas. Number of collected.
The targeted nature of endoscopy equipment for image acquisition in high-risk areas is improved, the efficiency of doctors' analysis is improved, and the generation of reinforced abnormality analysis models through integrated learning is improved, which improves the accuracy of model recognition.
Smart Images
Figure CN115965617B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of inspection auxiliary equipment, and specifically to a digestive tract endoscopy inspection auxiliary system based on integrated learning. The auxiliary system regulates the number of image acquisitions in different areas of the digestive tract, and its working process does not belong to a diagnostic method. Background Art
[0002] With the development of society and the change of people's eating habits, the incidence of digestive tract diseases in my country has increased year by year. Studies have shown that the detection rate of adenoma in colonoscopy is negatively correlated with the incidence and mortality of colorectal cancer. With the help of digestive tract endoscopy technology (Endoscopy), such as gastroscopy, colonoscopy, wireless capsule endoscopy (WCE), etc., the digestive tract mucosal lesions can be visually examined, which is an important means of detecting digestive tract diseases.
[0003] At present, the intraoperative diagnosis of clinical endoscopic examination mainly relies on manual diagnosis by doctors. However, manual diagnosis by doctors is often prone to missed diagnosis and misdiagnosis due to insufficient number of images of the lesion area collected, resulting in low examination accuracy and efficiency. Although some studies have also focused on real-time lesion detection during surgery to help doctors with intraoperative examinations, such as: the patent application with publication number "CN113990456A" provides "a deep learning-based early gastrointestinal cancer graphic analysis screening system", which requires doctors to manually intercept and analyze suspicious areas to obtain lesion detection data; the patent application with publication number "CN114569043A" provides "an artificial intelligence-based endoscopic assisted examination method and device", which requires doctors to manually select different models for suspicious areas to detect different lesions.
[0004] The above-mentioned existing technologies rely on the subjective awareness of doctors and have low efficiency. They are unable to automatically increase the number of images collected in high-risk areas or remind doctors of the appearance of high-risk areas, making it difficult to provide doctors with real-time warnings and efficient assistance during surgery. Summary of the invention
[0005] The purpose of the present invention is to provide a digestive tract endoscopy auxiliary system based on integrated learning; the auxiliary system regulates the number of image acquisitions in different areas of the digestive tract, and its working process does not belong to a diagnostic method.
[0006] In a first aspect, the present invention provides a gastrointestinal endoscopy auxiliary system based on integrated learning, which includes a data acquisition module, an identification module, an early warning module and an image acquisition control module. The data acquisition module is connected to the endoscopy device, receives the gastrointestinal images collected by the endoscopy device in real time, and transmits the gastrointestinal images to the identification module. The identification module uses an abnormal analysis model to determine in real time whether there is a suspected target in the gastrointestinal image. The early warning module is used to send a warning signal when the identification module detects a suspected target, so that the endoscopy device increases the number of images collected in the area where the suspected target is found. The abnormal analysis model includes a backbone feature extraction module, a network layer module and a prediction module connected in sequence. The image acquisition control module is used to control the sampling frequency of the endoscopy device.
[0007] Preferably, the backbone feature extraction module adopts the ResNet50vd_dcn network.
[0008] Preferably, the network layer module includes a path aggregation network (i.e., PANet network, PathAggregation Network), a Position Embedding module, and a Transformer Encoder module. The Position Embedding module divides the image input to the anomaly analysis model into multiple patches and corresponding position information; the feature information output by the Position Embedding module and the backbone feature extraction module are combined and input into the Transformer Encoder module; the Transformer Encoder module re-encodes the input feature information and synthesizes a feature map with position information; the feature map output by the Transformer Encoder module and the feature information output by the lower and middle layers of the backbone feature extraction module are input into the path aggregation network; the path aggregation network fuses the input features.
[0009] Preferably, the prediction module uses the classification regression layer of YOLOV5 to predict the image features processed by the network layer module.
[0010] In a second aspect, the present invention provides a method for assisting gastrointestinal endoscopy, which uses the aforementioned gastrointestinal endoscopy assisting system.
[0011] The digestive tract endoscopy auxiliary method comprises the following steps:
[0012] S1. Obtain multiple labeled digestive tract endoscopy images and divide them into dataset D and dataset T. The dataset D is used to train the primary abnormality model, and the dataset T is used to test the final enhanced abnormality analysis model. Use the bootstrap sampling method to randomly sample from the dataset D to obtain a training set for constructing the primary abnormality analysis model. Each training set corresponds to a test set from the dataset D. Use the bootstrap sampling method to obtain N training sets.
[0013] S2. Use the N different training sets obtained in step S1 to train the anomaly analysis models respectively; and obtain N different primary anomaly analysis models, namely base models.
[0014] S3. Using the weighted average method, the N base models obtained in step S2 are integrated into an enhanced anomaly analysis model and optimized.
[0015] S4. Use an endoscopic examination device to collect images of the digestive tract of the subject; the digestive tract images obtained during the collection process are input into the enhanced abnormality analysis model H(x) in real time; the enhanced abnormality analysis model H(x) determines whether there is a suspected target in the digestive tract image; when the enhanced abnormality analysis model H(x) detects a suspected target in the real-time returned digestive tract image, the endoscopic examination device increases the number of image collections of the area where the suspected target exists.
[0016] Preferably, the specific process of step S1 is as follows:
[0017] S11. Divide the labeled digestive tract endoscopy image set into a data set D and a data set T. The data set D is used for training the primary abnormality analysis model, and the data set T is used for testing and optimizing the final enhanced abnormality analysis model.
[0018] S12. Sample the dataset D by using the self-service sampling method. Each time, randomly select a sample from the dataset D and copy it into the sampling dataset D' which is initially an empty set. The selected sample is still retained in the dataset D. Assuming that the dataset has m samples in total, the sampling process is repeated m times to obtain a sampling dataset D' which also contains m samples. Due to the sampling with replacement, some samples in the dataset D will not be sampled into the sampling dataset D'. The probability of not being sampled is (1-1 / m) m The sample data in the data set D that are not sampled in the sampled data set D' is used as the test set of the primary anomaly analysis model. The ratio of the test set and the training set of the primary anomaly analysis model is determined by m.
[0019] S13. Repeat the steps in step S12 N times to obtain N different training sets and their corresponding test sets.
[0020] Preferably, the specific process of step S2 is as follows:
[0021] S21. Use the N training sets obtained in step S1 to perform model training respectively, and obtain N different primary anomaly analysis models; use the test sets corresponding to each training set to test the N models respectively, and obtain the confidence scores of the N models, as well as the accuracy and recall rate of each model under different target sizes.
[0022] S22, adjust the parameters of N different primary anomaly analysis models according to the corresponding accuracy and recall rate, and continue to train the model until an optimal model is obtained for each training set, which is used as the primary anomaly analysis model h i (x).
[0023] Preferably, the specific process of step S3 is as follows:
[0024] S31. Extract the overall AP of N primary anomaly analysis models and the AP of small targets s , Precision AP under medium targets m , Precision AP under large targets l , the overall recall rate AR; the i-th primary anomaly analysis model h i The above five indices of (x) are respectively denoted as i=1,2,...,N; set the indexes respectively Weight Initial value of . Construct N primary anomaly analysis models h i The weight w of (x) i The expression is as follows:
[0025]
[0026] S32, according to N primary anomaly analysis models h i (x) and its corresponding weight w i , the enhanced anomaly analysis model H(x) is constructed as follows:
[0027]
[0028] S33. Use the data set T to evaluate the performance of the enhanced anomaly analysis model H(x), and adjust the indicators of each base model according to the evaluation results. And reconstruct the enhanced anomaly analysis model H(x).
[0029] S34. Repeat step S33 until the optimal abnormality analysis model H(x) is obtained.
[0030] As a preference, the indicator Weight The initial values of are set to 0.4, 0.2, 0.15, 0.15, 0.1 respectively; j=1,2,3,4,5.
[0031] Preferably, when the enhanced anomaly analysis model H(x) detects a suspected target, the early warning module emits a prompt sound and / or turns on a warning light.
[0032] The present invention has the following beneficial effects:
[0033] 1. The present invention performs real-time analysis on the digestive tract images obtained while the endoscopic examination equipment collects the digestive tract images, thereby increasing the number of images collected in the digestive tract areas with a higher risk of abnormalities, improving the pertinence of the image collection by the endoscopic examination equipment, and helping to improve the efficiency of doctors' analysis of digestive tract images.
[0034] 2. The present invention adopts a method combining ensemble learning and deep learning to generate an enhanced anomaly analysis model, which can improve the accuracy of model recognition and thus more specifically increase the number of local image sampling. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 A flowchart of the present invention for assisting digestive tract endoscopy;
[0036] Figure 2 A schematic diagram of the process of constructing an abnormality analysis model in the present invention;
[0037] Figure 3 It is a structural block diagram of the abnormal analysis model in the present invention;
[0038] Figure 4 This is a flow chart of step S2 in the present invention. DETAILED DESCRIPTION
[0039] The present invention will be further described below in conjunction with the accompanying drawings.
[0040] A digestive tract endoscopy auxiliary system based on integrated learning includes a data acquisition module, a recognition module, an early warning module and an image acquisition control module. The data acquisition module is connected to an endoscopy device, and is used to receive the digestive tract images collected by the endoscopy device in real time, and transmit the digestive tract images to the recognition module. The recognition module determines in real time whether there is a suspected target in the digestive tract image through an abnormal analysis model. The image acquisition control module is used to control the sampling frequency of the endoscopy device.
[0041] The warning module is used to send out a warning signal when the identification module detects a suspected target, so as to manually or automatically increase the number of images collected by the endoscopic examination equipment in the area where the suspected target is found. Collecting different numbers of digestive tract images for different areas will help doctors more accurately judge the digestive tract condition of the subject in subsequent analysis.
[0042] The anomaly analysis model uses Figure 3 The structure contains the following modules:
[0043] Backbone feature extraction module (i.e. backbone): ResNet50vd_dcn network is selected as the backbone feature extraction module to extract image features. The network consists of 5 layers of convolution blocks (i.e. Conv). The first layer of convolution blocks includes a 7×7 convolution, a normalization process and an activation process of the ReLu function; the second layer of convolution blocks consists of a 3×3 maximum pooling layer and 3 residual blocks (Res Block); the third layer of convolution blocks consists of 4 residual blocks; the fourth layer of convolution blocks consists of 6 residual blocks; the fifth layer of convolution blocks consists of 4 residual blocks. Each residual block is a residual block that replaces the last layer of convolution with a deformable convolution based on the residual block structure of ResNet-D.
[0044] Network layer module (ie Neck): Select the path aggregation network (PANet network) as the model feature fusion neck. The path aggregation network first performs top-down feature fusion and then performs bottom-up feature fusion, which can better combine the extracted high-level features with the image position. Innovatively, the present invention combines the Transformer global attention mechanism and adds a Position Embedding module and a Transformer Encoder module to the network layer module for improvement. The image input to the anomaly analysis model is divided into multiple Patches and corresponding position information in the Position Embedding module, and is input into the Transformer Encoder module together with the last layer of feature information obtained by the backbone feature extraction module. After encoding by this module, a new feature map containing position information is formed and input into PANet. After the high-level (high-level feature information is processed by combining the Position Embedding module and the Transformer Encoder module and then input into PANet), middle-level, and low-level feature maps are sent to PANet, the high-level feature information is combined with the low-level feature information through upsampling, and then the position information in the low-level feature map is combined with the higher feature map through bottom-up convolution and connection (Concat) operations. In this way, the extracted high-level feature map can be better combined with the position of the image, thereby achieving the purpose of optimizing the model detection effect.
[0045] Prediction module (i.e. Head): Use the classification regression layer of YOLOV5 to predict the image features processed by the network layer module, generate bounding boxes and predicted categories. The prediction module performs a convolution operation on the three branches output from the network layer module, including normalization and activation functions, and then performs another convolution to obtain a vector containing the predicted category, feature map length, and feature map width. Then, the output vectors in all batches are traversed, and the sigmoid operation is used to map the position information prediction [x, y] to the range of [0, 1]. At the same time, the confidence of the prediction and its category score are obtained, and the decoder is used to decode and obtain the predicted box. Finally, the boxes with confidence greater than a certain value are retained, and the remaining boxes are screened through NMS maximum suppression to obtain the final predicted box.
[0046] This embodiment is only used to adjust the number of images collected in different areas of the digestive tract according to the specific conditions of different areas when collecting digestive tract images, so that the acquired digestive tract image set is more representative. It cannot determine whether the person being examined suffers from a digestive tract disease. Therefore, the working process of this embodiment does not belong to a diagnostic method, but is only an auxiliary method for digestive tract endoscopy.
[0047] like Figure 1 and 2 As shown, the process of constructing and using the digestive tract endoscopy auxiliary system based on ensemble learning is as follows:
[0048] S1. Obtain batches of labeled digestive tract endoscopy lesion image data, divide them into training sets and test sets, and sample the training sets by self-service sampling to obtain N different sample data sets. N is the preset number of training sets. The label is the lesion area in the image.
[0049] This step mainly obtains data sets for training different primary anomaly analysis models and data sets for testing and adjusting enhanced anomaly analysis models. The specific process of step S1 is as follows:
[0050] S11. Obtain a dataset containing annotated gastrointestinal endoscopic lesion images. Assume that the dataset contains n sample data. Divide m of the datasets into dataset D for training and testing the primary abnormality analysis model, and divide the remaining nm sample data into dataset T for testing and improving the performance of the enhanced abnormality analysis model.
[0051] S12. Sample the dataset D by using the self-service sampling method. Each time, randomly select a sample from the dataset D and copy it into the sampling dataset D' which is initially an empty set. Then put the sample back into the dataset D so that the sample still has a probability of being sampled in the next sampling. Repeat this sampling process m times to obtain a sampling dataset D' containing m samples. Therefore, some samples in the dataset D appear multiple times in the sampling dataset D', while some samples are not sampled. The probability that a sample is never sampled in m samplings is (1-1 / m) m Since the collection of medical image datasets may be difficult and small in size, the self-service sampling method can more effectively divide the training set and test set of the model; the sampling dataset D' is used as the training set of the primary anomaly analysis model; the sample data in the dataset D that is not sampled in the sampling dataset D' (i.e., D-D') is used as the test set of the primary anomaly analysis model, and they are used together for the training of the primary anomaly analysis model; by controlling the sampling number m, the ratio of the model training set and the test set can be determined.
[0052] S13. Repeat the steps in step S12 N times to obtain N different training sets and their corresponding test sets.
[0053] S2. Use the N different training sets obtained in step S1 to build N different primary anomaly analysis models. The specific process is as follows:
[0054] S21. Use the N training sets obtained in step S1 to train the primary anomaly analysis model to obtain N different models; use the test sets corresponding to each training set to test the N models to obtain N model confidence scores, as well as the accuracy and recall rate of each model under different target sizes.
[0055] S22, adjust the parameters of N primary anomaly analysis models according to the corresponding accuracy and recall rate, and continue to train the model until an optimal model and corresponding performance indicators are obtained for each training set, which is used as the primary anomaly analysis model h i (x).
[0056] The function of step S2 is to use the multiple independent data sets obtained by the self-service sampling method in step S1 to train multiple optimal primary abnormality analysis models on different data sets, and obtain the recognition capabilities of these primary abnormality analysis models for lesions of different sizes, types and characteristics. The primary abnormality analysis model used in training often adopts a single-stage target detection model framework, which can achieve a better balance between precision and accuracy.
[0057] The purpose of executing steps S1 and S2 is to train multiple primary anomaly analysis models, such as Figure 4 shown.
[0058] S3. Use the weighted average method to integrate the N primary anomaly analysis models (specifically, N different deep learning real-time anomaly analysis models) obtained in step S2 into an enhanced anomaly analysis model with strong detection capabilities for diverse features, and continuously optimize it to obtain the optimized anomaly analysis model. The specific process is as follows:
[0059] S31, adjust each primary anomaly analysis model h according to the accuracy and recall of different models obtained in S2 (including the accuracy and recall under different target types and different target sizes) i The weight w of (x) i (in ). N primary anomaly analysis models h i The weight w of (x) i The acquisition process is as follows:
[0060] S311: Count the indicators obtained in step S2 to obtain the overall accuracy AP and the accuracy AP under small targets. s , Precision AP under medium targets m , Precision AP under large targets l , the overall recall rate AR; the i-th primary anomaly analysis model h i The above five indices of (x) are respectively denoted as i=1,2,...,n; Weight The initial values of are set to 0.4, 0.2, 0.15, 0.15, 0.1 respectively; j=1,2,3,4,5.
[0061] S312, construct N primary anomaly analysis models h i The weight w of (x) i The expression is as follows:
[0062]
[0063] S32, according to N primary anomaly analysis models h i (x) and its corresponding weight w i , the enhanced anomaly analysis model H(x) is constructed as follows:
[0064]
[0065] S33, use the data set T to evaluate the performance of the enhanced anomaly analysis model H(x), and compare the N primary anomaly analysis models h(x) based on the obtained accuracy and recall. i The weight w of (x) iRe-adjust them separately, specifically by manually adjusting each primary anomaly analysis model h i (x) indicator The corresponding weight To achieve the weight w i adjustments.
[0066] S34. Repeat step S33 until an optimal strong learning model is obtained as the final enhanced anomaly analysis model H(x).
[0067] The purpose of step S3 is to use the ensemble learning method to continuously iterate and optimize the precision and recall of the model, reduce the false positive rate and missed positive rate, and thus improve the model performance.
[0068] S4. In the process of using the endoscopic examination equipment to continuously collect the digestive tract images of the subject; the collected digestive tract images are input into the enhanced abnormality analysis model H(x) in real time, and the enhanced abnormality analysis model H(x) determines whether there is an area with suspected targets; when the enhanced abnormality analysis model H(x) detects a suspected target in the digestive tract images transmitted back in real time, the early warning module emits a prompt sound and a prompt light to the doctor, and at the same time, the endoscopic examination equipment increases the frequency of image acquisition; after hearing the prompt sound, the doctor stops the endoscope in the current detection area, further making the number of digestive tract images collected in the area where the enhanced abnormality analysis model H(x) detects suspected targets greater than the number of digestive tract images collected in other areas, so as to improve the accuracy of the doctor's judgment on the condition.
[0069] The specific process of step S4 is as follows:
[0070] S41. During the endoscopic examination, the enhanced abnormal analysis model H(x) detects the collected endoscopic images in real time. When a suspected target is detected, the type and position of the suspected target are displayed in real time, the probability of it being an abnormal target is determined, and the early warning module is triggered.
[0071] S42. The early warning module emits a prompt sound, controls the screen of the endoscope inspection instrument to flash, and controls the warning light to light up red.
[0072] S43. The doctor further observes the current inspection area, increases the number of images collected in the current inspection area (the doctor can choose whether to conduct a more detailed inspection in the area and take photos for evidence), and then closes the current prompt of the warning module until the warning module is triggered again.
[0073] In general, step S4 is a specific method of using the early warning module, so as to better help doctors during digestive tract examinations or operations.
[0074] It should be understood that the technical content disclosed in this embodiment can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.
[0075] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.
[0076] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0077] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.
[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for assisting digestive tract endoscopy, characterized in that: The following steps are involved: S1. Obtain multiple labeled digestive tract endoscopy images and divide them into data set D and data set T; construct N training sets based on data set D, each training set corresponds to a test set from data set D; N is the preset number of sample training sets; The specific process of step S1 is as follows: S11, dividing the labeled digestive tract endoscopy image set into a data set D and a data set T; S12. Sample the dataset D by using the self-service sampling method. Each time, randomly select a sample from the dataset D and copy it into the sampling dataset D' which is initially an empty set. The sample is still retained in the dataset D. The sampling process is repeated m times to obtain a sampling dataset D' containing m samples. m is the number of samples in the dataset D. The sample data in the dataset D that is not sampled into the sampling dataset D' is used as the test set. S13, repeat the steps in step S12 N times to obtain N different training sets and their corresponding test sets; S2. Use the N different training sets obtained in step S1 to train the anomaly analysis models respectively; obtain N different primary anomaly analysis models; the anomaly analysis model comprises a backbone feature extraction module, a network layer module and a prediction module connected in sequence; the backbone feature extraction module is composed of 5 layers of convolution blocks; the prediction module uses the classification regression layer of YOLOV5 to predict the image features processed by the network layer module; The network layer module includes a path aggregation network, a Position Embedding module and a Transformer Encoder module; the Position Embedding module divides the image input to the anomaly analysis model into multiple blocks and corresponding position information; the information output by the Position Embedding module is combined with the feature information output by the backbone feature extraction module and then input into the Transformer Encoder module; the Transformer Encoder module re-encodes the input feature information to synthesize a feature map with position information; the feature map output by the Transformer Encoder module and the feature information output by the lower and middle layers of the backbone feature extraction module are all input into the path aggregation network; the path aggregation network fuses the input features; The specific process of step S2 is as follows: S21, using the N training sets obtained in step S1 to respectively train the anomaly analysis model pairs, and obtain N different primary anomaly analysis models; using the test sets corresponding to the training sets to respectively test the N primary anomaly analysis models, and respectively obtain the confidence scores of the N primary anomaly analysis models, and the accuracy and recall rate of each primary anomaly analysis model under different target sizes; S22, adjust the parameters of N primary anomaly analysis models according to the corresponding accuracy and recall rate, and continue to train the model until an optimal model is obtained for each training set, which is used as the primary anomaly analysis model h i (x); S3, using a weighted average method to integrate the N primary anomaly analysis models obtained in step S2 into an enhanced anomaly analysis model and optimize it; The specific process of step S3 is as follows: S31. Extract the overall AP of N primary anomaly analysis models and the AP of small targets s , Precision AP under medium targets m , Precision AP under large targets l , the overall recall rate AR; the i-th primary anomaly analysis model h i The above five indicators of (x) are respectively recorded as Set indicators separately Weight Initial value of; Construct N primary anomaly analysis models h i The weight w of (x) i The expression is as follows: S32, according to N primary anomaly analysis models h i (x) and its corresponding weight w i , the enhanced anomaly analysis model H(x) is constructed as follows: S33, use the data set T to evaluate the performance of the anomaly analysis model H(x), and adjust the indicators of each primary anomaly analysis model according to the evaluation results And reconstruct the enhanced anomaly analysis model H(x); S34, repeat step S33 until the optimal enhanced anomaly analysis model H(x) is obtained; S4. Use the endoscopic examination equipment to collect images of the digestive tract of the subject; the digestive tract images obtained during the collection process are input into the enhanced abnormality analysis model H(x) in real time; the enhanced abnormality analysis model H(x) determines whether there is a suspected target in the digestive tract image; when the enhanced abnormality analysis model H(x) detects a suspected target in the real-time returned digestive tract image, the endoscopic examination equipment increases the number of image collections in the area where the suspected target exists.
2. A digestive tract endoscopy auxiliary method according to claim 1, characterized in that: index Weight The initial values of are set to 0.4, 0.2, 0.15, 0.15, 0.1 respectively; j=1,2,3,4,5.
3. A digestive tract endoscopy auxiliary method according to claim 1, characterized in that: When the enhanced anomaly analysis model H(x) detects a suspected target, the early warning module emits a prompt sound and / or turns on a warning light.
4. A digestive tract endoscopy auxiliary system based on ensemble learning, characterized in that: Used to execute the gastrointestinal endoscopy auxiliary method as described in claim 1; the gastrointestinal endoscopy auxiliary system includes a data acquisition module, an identification module, an early warning module and an image acquisition control module; the data acquisition module is connected to the endoscopic inspection device, receives the gastrointestinal images collected by the endoscopic inspection device in real time, and transmits the gastrointestinal images to the identification module; the identification module uses the abnormal analysis model to determine in real time whether there is a suspected target in the gastrointestinal image; the early warning module is used to send a warning signal when the identification module detects a suspected target, so that the endoscopic inspection device increases the number of images collected in the area where the suspected target is found; the abnormal analysis model includes a backbone feature extraction module, a network layer module and a prediction module connected in sequence; the image acquisition control module is used to control the sampling frequency of the endoscopic inspection device.
5. The digestive tract endoscopy auxiliary system based on ensemble learning according to claim 4 is characterized in that: The five-layer convolution block of the backbone feature extraction module is as follows: the first layer of convolution blocks includes 7×7 convolution processing, normalization processing and ReLu function activation processing; the second layer of convolution blocks consists of a 3×3 maximum pooling layer and 3 residual blocks; the third layer of convolution blocks consists of 4 residual blocks; the fourth layer of convolution blocks consists of 6 residual blocks; the fifth layer of convolution blocks consists of 4 residual blocks; each residual block is a residual block in which the last layer of convolution is replaced by a deformable convolution based on the residual block structure of ResNet-D.