A polyp detection method based on convolutional neural network

Abstract

The present application relates to a kind of polyp detection technical field, the present application discloses a kind of polyp detection method based on convolutional neural network, including collecting patient intestinal tract detection original image, polyp detection model is constructed, polyp detection dataset is constructed, generates synthetic image, verifies classification loss, extracts transition area image feature, the cross-entropy loss between the prediction probability distribution of model and image real distribution is calculated, with the contrast loss of synthetic image and target image and the feature classification loss of transition area and the weighted sum of cross-entropy loss are used as the total loss function of polyp training model, according to model total loss function selection prediction result, in a plurality of perspective multi-distance mode polyp image is collected, in the transition area of polyp and intestinal tract background is introduced in polyp classification process to further detect and identify polyp image, by multi-level multi-region multi-feature polyp dataset is fused and classified, the accuracy of polyp detection can be improved, and false positive of polyp detection is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of polyp detection, and particularly relates to a polyp detection method based on a convolutional neural network. BACKGROUND

[0002] Polyp detection is commonly used for diagnosing colorectal cancer, which is one of the most common malignant tumors of the digestive system in the world. Most colorectal cancers originate from tumor polyps. Polyp detection based on colonoscopy is a key task in medical image computing. However, due to the complex environment of the intestinal tract, bubbles, lens reflections, residues, and shadows may all show polyp characteristics. The differentiation degree of polyps is difficult to distinguish only by colonoscopy images under white light. Therefore, a precise and objective computer-aided classification system is necessary. Polyp classification is a routine computer vision task that can automatically identify the pathological type of polyps in images or videos, helping doctors make correct judgments in the diagnosis and treatment process, thereby saving the doctor's energy and reducing the possibility of misdiagnosis and missed diagnosis. Polyp classification is a routine computer vision task that can automatically identify the pathological type of polyps in images or videos, helping doctors make correct judgments in the diagnosis and treatment process, thereby saving the doctor's energy and reducing the possibility of misdiagnosis and missed diagnosis. At present, deep learning detection models are widely used in polyp detection. However, in clinical practice, the complex intestinal environment often causes a large number of foreign bodies, liquids, and blurs in the lens field of view, and it is difficult to make accurate result evaluation only according to a single colonoscopy image. This is also the reason why the accuracy and sensitivity of existing methods are not satisfactory. Research has found that a high-efficiency polyp detection framework proposes a multi-branch spatial attention mechanism to make the model pay more attention to the polyp lesion area. At the same time, a multi-scale sampling strategy is used in combination with maximum likelihood loss to reduce false positives by learning from suspicious areas in the background. Further, a cosine similarity loss is proposed to improve the discrimination ability of positive and negative images. However, in the polyp classification process, only real images are processed for features, and image acquisition is often affected by the intestinal environment, often leading to deviations between the collected images and the actual situation. Therefore, it is necessary to develop a new polyp recognition and classification method to reduce false positives in polyp detection and reduce the probability of misdiagnosis and missed diagnosis. SUMMARY

[0003] The present application aims to provide a polyp detection method based on a convolutional neural network. By using a simulator to synthesize polyp images and incorporating the synthesized polyp images and real polyp images into polyp training, a more generalized polyp classification model is obtained. In the polyp classification process, the transition area between the polyp and the intestinal background is introduced to further detect and recognize polyp images, which can effectively solve the problems in the background technology.

[0004] To achieve the above object, the present application provides the following technical scheme:

[0005] A polyp detection method based on convolutional neural network, comprising the following steps:

[0006] Step 1: Collect the original image of the patient's intestinal tract for detection;

[0007] Step 2: Mark the salient region of the suspected polyp target in the image data, and take the intercepted suspected polyp image as the target image based on the identified frame size in the image data set;

[0008] Step 3: Establish a polyp detection data set, including a normal image data set and a polyp data set, the normal image data set including normal examination images with suspicious targets in past case examinations, and the polyp data set including examination images of the target to be detected;

[0009] Step 4: Polyp detection data set processing, randomly selecting a pair of random images of the same category as the target image from the normal image data set, mapping the random images to a synthetic image of the target image using a convolutional neural network model and a multi-branch spatial attention mechanism, and calculating the contrast loss of the synthetic image and the target image;

[0010] Step 5: According to the verification result of the synthetic image and the target image, input the image set that meets the requirements of relevance and similarity into the classification network, and classify and evaluate the synthetic image and the target image to verify the classification loss of the synthetic image and the target image;

[0011] Step 6: Add the transition region (Ω) between the target region and the background as an evaluation factor in the classification evaluation process, select feature values from the transition region (Ω), compare the differences of the feature values, select the color level (c) and the blood vessel distribution density (m) and the smoothness (s) of the transition region (Ω) as the feature values, define the extracted feature map as domain R, and divide domain R into domain Ω1 containing transition region features and domain R1 outside the transition region, respectively expand Ω1 and R1 into feature vectors with dimensions c×m×s And Add an embedding layer to project it into an i-dimensional feature vector, where i represents the number of sample categories in the data set, and the final feature vector is regarded as the prediction vector of the polyp training model for the transition region; the classification loss evaluation uses maximum likelihood loss and cosine similarity loss to reduce the prediction error rate, and the overall classification loss

[0012] represents the cosine similarity loss, k represents the K nearest neighbor classification coefficient, and i and j represent the coordinates of a certain point on the feature vector;

[0013] Step 7: Calculate the cross-entropy loss between the prediction probability distribution of the polyp training model and the real distribution of the image;

[0014] The eighth step: taking the contrast loss Ls of the synthetic image and the target image and the feature classification loss L of the transition area as a weighted sum of cross-entropy loss, and taking the weighted sum as a total loss function of the polyp training model, and the smaller the total loss function value is, the higher the model training accuracy is cs The eighth step: taking the contrast loss Ls of the synthetic image and the target image and the feature classification loss L of the transition area as a weighted sum of cross-entropy loss, and taking the weighted sum as a total loss function of the polyp training model, and the smaller the total loss function value is, the higher the model training accuracy is

[0015] The ninth step: setting a polyp training model total loss function threshold, and directly outputting a polyp prediction result for a classification result with a total loss function less than the total loss function threshold, and returning to generate a synthetic image path to re-detect a suspected polyp image for a classification model with a total loss function exceeding the total loss function threshold.

[0016] As a further scheme of the present application: in the first step, the patient intestinal detection original image is collected by using a colonoscopy technology means, the collected patient intestinal detection original image is a high-resolution image, a low-resolution image is processed by using a super-resolution technology, the resolution of the image is improved, and the influence of shaking on the resolution of the image during image collection is reduced.

[0017] As a further scheme of the present application: in the third step, the data set is established based on the automatic polyp detection and identification function of the colonoscopy, multi-angle and multi-directional image data of the detection area are collected, the sampling angle a is adjusted based on the suspected target area, the suspected target area image is sampled at one sampling angle and one sampling distance, the sampling port is domain A, the detection area is domain B, the domain A has feature vectors a (a1, a2, a3,..., an), the domain B has feature vectors b (b1, b2, b3,..., bn), and the distance between the points is The distance between the domain A and the domain B The boundary curve function of the domain A records the sampling angle a and the sampling distance D.

[0018] As a further scheme of the present application: in the fourth step, the synthetic image is generated by setting a background image, rendering the polyp under the background image to obtain the synthetic image, and setting the background image by using an environment simulator to set the polyp environment, including setting the humidity, temperature and blood vessel density of the polyp environment.

[0019] As a further scheme of the present application: in the fourth step, the required contrast loss function in the calculation of the contrast loss of the synthetic image and the target image is N represents the number of samples, used for verifying the correlation loss of the synthetic image and the target image, used for verifying the matching degree of the synthetic image and the target image, Dw is the Euclidean distance of the feature X1 of the synthetic image and the feature X2 of the target image, that is Represent the influence factor in the optimization process, d is the dimension of the feature X1 of the synthetic image and the feature X2 of the target image, C S , C T Represent the covariance matrix of the feature X1 of the synthetic image and the feature X2 of the target image in d dimension, F is an input initial constant, verify the contrast loss of the synthetic image and the target image, which can reduce the false positive probability of polyp in sample classification.

[0020] As a further scheme of the application: the classification evaluation standard in the sixth step is T r Represent the accuracy, F L Represent the error rate, M i Represent the false negative rate, N T Is the number of samples correctly classified in the detection image, N is the total number of detection image samples, N fp Is the number of samples in the detection sample that are correctly classified as polyp images, N p-po Is the number of samples in the detection image that are classified as polyp images by the model, N fn Is the number of samples in the detection image that are classified as normal images, N po Is the number of samples in the detection image that are polyp images.

[0021] As a further scheme of the application: the polyp training in the ninth step is a multiple cycle convolutional neural network model, which can improve the accuracy of polyp prediction saliency detection results, and integrate the saliency information of the transition region into the final polyp saliency prediction, which can highlight the saliency features of the whole image and accurately locate the boundary region of the suspected target, and can obtain accurate result prediction.

[0022] Compared with the prior art, the application has the advantages of:

[0023] The polyp image is collected in a multi-view and multi-distance manner, the simulator generates a synthetic polyp image, and the synthetic polyp image and the real image are simultaneously trained, so that the synthetic image is closer to the real image, a more generalized polyp classification model is obtained, and the transition region between the polyp and the intestinal background is introduced in the polyp classification process to further detect and identify the polyp image, the polyp data set is fused and classified through multiple levels, multiple regions and multiple features, the accuracy of polyp detection is improved, the false positive of polyp detection is reduced, and the probability of misdiagnosis and missed diagnosis is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0025] Figure 1 A flowchart of a polyp detection method based on a convolutional neural network. DETAILED DESCRIPTION

[0026] In order to make the technical problems to be solved by the present application, technical solutions and beneficial effects more clearly, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0027] Embodiment 1:

[0028] Please refer to Figure 1 In the embodiments of the present application, a polyp detection method based on a convolutional neural network comprises the following steps:

[0029] Step 1: Collecting patient intestinal detection original images, using colonoscopy technology to collect the original images to be detected, the collected patient intestinal detection original images to be detected are high-resolution images, using super-resolution technology to process the collected low-resolution images, improving the resolution of the images, and reducing the influence of image shaking on the resolution of the images when collecting images;

[0030] Step 2: Marking the significant area of the suspected polyp target in the image data, taking the size of the identification frame in the image data set as the benchmark, and taking the intercepted suspected polyp image as the target image. Considering the influence of the patient's abdominal environment on the polyp, the polyp is prone to be confused with the background and is not easy to be distinguished. First, according to the previous marking result, the collected patient intestinal detection original image is pixel-level cropped, which can improve the data set classification accuracy. Then, since the collected patient intestinal detection original image has a high image resolution, and the polyp occupies a small proportion, the patient intestinal detection original image is further cropped by marking frame, so that the polyp is in a suitable field of view ratio, thereby obtaining a polyp image representing a polyp individual;

[0031] Third step: Establish polyp detection dataset, including normal image dataset and polyp dataset, normal image dataset includes normal examination image set with suspicious target in past case examination, polyp dataset includes examination image set of target to be detected, since the establishment of dataset is based on automatic polyp detection recognition function of colonoscopy, the target image collected on the same polyp sample has similarity, only certain difference in spatial perspective and distance, therefore, in the division of polyp detection dataset, the sampling angle a is adjusted based on the region where the suspicious target is located, the sampling distance is adjusted under one sampling angle, the sampling port is domain A, the region to be detected is domain B, domain A has feature vector a (a1, a2, a3, …, an), domain B has feature vector b (b1, b2, b3, …, bn), the distance between point to point is Distance between domain A and domain B The boundary curve function of domain A records the sampling angle a and the sampling distance D, based on the perspective and distance, the dataset is grouped in multiple dimensions, which can ensure that the difference between images in the same group will not be too large, and to a certain extent, the abnormal image data is eliminated;

[0032] Fourth step: Polyp detection dataset processing, randomly select a pair of random images of the same category as the target image from the normal image dataset, use convolutional neural network model and multi-branch spatial attention mechanism to map the random image to the synthetic image of the target image, set the background image, render the polyp under the background image to get the synthetic image, set the background image by using the environment simulator to set the polyp environment, according to the influence degree of polyp environment on polyp, select the humidity, temperature and blood vessel density of polyp environment as background setting factors, which can inhibit the interference of background on polyp to a certain extent, calculate the contrast loss of synthetic image and target image;

[0033] Fifth step: According to the verification result of synthetic image and target image, input the image set which meets the requirements of correlation and similarity into the classification network, classify and evaluate the synthetic image and the target image, and verify the classification loss of the synthetic image and the target image;

[0034] Sixth step: Add the transition region (Ω) between the target region and the background as an evaluation factor in the classification evaluation process, select the feature value from the transition region (Ω), compare the difference of the feature value, select the color level (c) and the blood vessel distribution density (m) and the smoothness (s) of the transition region (Ω) as the feature value, define the extracted feature map as domain R, domain R is divided into domain Ω1 containing transition region features and domain R1 outside the transition region, expand Ω1 and R1 into feature vectors with dimensions c x m x s respectively And And add an embedding layer to project it into i-dimensional feature vectors, i represents the number of sample categories in the data set, and the final feature vector is regarded as the prediction vector of the polyp training model for the transition area; the classification loss evaluation uses maximum likelihood loss and cosine similarity loss to reduce the prediction error rate, and the overall classification loss Represent the cosine similarity loss, k represents the K nearest neighbor classification coefficient, and i, j represent the coordinates of a certain point on the feature vector;

[0035] Step 7: Calculate the cross-entropy loss between the prediction probability distribution of the polyp training model and the real distribution of the image. The polyp training model result includes two cases of negative and positive, so the probability of prediction as positive is defined as p, and the negative probability is 1-p. At this time, the cross-entropy loss function of a certain sample w is y w 1 when the sample is positive, and 0 when it is negative;

[0036] Step 8: Take the weighted sum of the contrast loss Ls of the synthetic image and the target image and the feature classification loss L cs And the cross-entropy loss as the total loss function of the polyp training model, the smaller the total loss function value, the higher the model training accuracy

[0037] Step 9: Set the total loss function threshold of the polyp training model. The classification result of the total loss function less than the total loss function threshold can be directly output as the polyp prediction result, and the classification model exceeding the total loss function threshold is returned to generate a synthetic image path to detect the suspected polyp image again.

[0038] By adopting the above technical scheme: the polyp image is collected in a multi-view and multi-distance manner, the simulator generates a synthetic polyp image, and the synthetic polyp image and the real image are simultaneously trained, so that the synthetic image is closer to the real image, a more generalized polyp classification model is obtained, and the transition area between the polyp and the intestinal background is introduced in the polyp classification process to further detect and identify the polyp image. Through multi-level, multi-region and multi-feature fusion classification of the polyp data set, the accuracy of polyp detection can be improved, the false positive of polyp detection is reduced, and the probability of misdiagnosis and missed diagnosis is reduced.

[0039] Embodiment 2

[0040] Please refer to Figure 1 In the embodiment of the application, a polyp detection method based on a convolutional neural network comprises the following steps:

[0041] The first step is to collect the original image of the patient's intestinal tract for detection, and use colonoscopy technology to collect the original image for detection. The collected original image of the patient's intestinal tract for detection is a high-resolution image. The collected low-resolution image is processed by using super-resolution technology to improve the resolution of the image and reduce the influence of image shaking on the resolution of the image during image collection.

[0042] The second step is to mark the significant area of the suspected polyp target in the image data. The size of the identified frame in the image data set is used as a reference, and the intercepted suspected polyp image is used as the target image.

[0043] The third step is to establish a polyp detection data set, including a normal image data set and a polyp data set. The normal image data set includes normal examination images with suspicious targets in previous case examinations, and the polyp data set includes examination images of the target to be detected.

[0044] The fourth step is to process the polyp detection data set. A pair of random images of the same category as the target image is randomly selected from the normal image data set. The convolutional neural network model and the multi-branch spatial attention mechanism are used to map the random images to the synthetic images of the target image. The contrast loss between the synthetic images and the target image is calculated. The contrast loss function required in the calculation of the contrast loss between the synthetic images and the target image is N represents the number of samples,

[0045] The correlation loss between the synthetic images and the target image is verified, The matching degree between the synthetic images and the target image is verified. Dw is the Euclidean distance between the feature X1 of the synthetic image and the feature X2 of the target image, i.e. represents the influence factor in the optimization process, d is the dimension of the feature X1 of the synthetic image and the feature X2 of the target image, C S , C T represents the covariance matrix of the feature X1 of the synthetic image and the feature X2 of the target image in d dimensions, F is an input initial constant, and the contrast loss between the synthetic images and the target image is verified. This can reduce the false positive probability of polyps during sample classification.

[0046] The fifth step is to input the image set that meets the requirements of correlation and similarity into the classification network according to the verification results of the synthetic images and the target images, and to classify and evaluate the synthetic images and the target images. The classification loss between the synthetic images and the target images is verified.

[0047] Step 6: The transition region (Ω) between the target region and the background is added as an evaluation factor in the classification evaluation process. Feature values are selected from the transition region (Ω), and the differences in the feature values are compared. The color level (c) and the blood vessel distribution density (m) and the smoothness (s) of the transition region (Ω) are selected as feature values. The extracted feature map is defined as domain R, which is divided into domain Ω1 containing transition region features and domain R1 outside the transition region. Ω1 and R1 are respectively expanded into feature vectors with dimensions c x m x s With An embedding layer is added to project it into an i-dimensional feature vector, where i represents the number of sample categories in the data set. The final feature vector is considered as the prediction vector of the polyp training model for the transition region. The classification loss evaluation uses maximum likelihood loss and cosine similarity loss to reduce the prediction error rate. The overall classification loss represents the cosine similarity loss, k represents the K-nearest neighbor classification coefficient, and i and j represent the coordinates of a certain point on the feature vector. The classification evaluation standard is T r represents the accuracy, F L represents the error rate, M i represents the false negative rate, N T is the number of samples in the detection image that are correctly classified, N is the total number of detection image samples, N fp is the number of samples in the detection image that are correctly classified, N is the total number of detection image samples, N p-po is the number of samples in the detection image that are correctly classified, N is the total number of detection image samples, N fn is the number of samples in the detection image that are correctly classified, N is the total number of detection image samples, N po is the number of samples in the detection image that are correctly classified, N is the total number of detection image samples, N

[0048] Step 7: Calculate the cross-entropy loss between the prediction probability distribution of the polyp training model and the true distribution of the image.

[0049] Step 8: Take the weighted sum of the contrast loss Ls between the synthetic image and the target image and the feature classification loss L cs of the transition region and the cross-entropy loss as the total loss function of the polyp training model. The smaller the total loss function value, the higher the model training accuracy.

[0050] The ninth step is to set a polyp training model total loss function threshold, and the classification result of the total loss function less than the total loss function threshold can directly output the polyp prediction result, and the classification model exceeding the total loss function threshold returns to generate a synthetic image path to detect the suspected polyp image again, the polyp training is a multiple loop convolutional neural network model, which can improve the accuracy of polyp prediction saliency detection result, and integrate the saliency information of the transition region into the final polyp saliency prediction, which can highlight the saliency features of the whole image and accurately locate the boundary region of the suspected target, and can obtain accurate result prediction.

[0051] The classification effect of the above method is verified by the result verification theory, and since the above method adds extra transition region feature learning, the image may be over-fitted due to excessive parameters, therefore, the formula is introduced to reduce the correlation between feature maps, which can prevent over-fitting of feature maps, is an input value, is an output value, (x, y) is the coordinate position of the target image pixel, i, j is the feature map channel serial number, N is the total channel number, k, alpha and beta are self-defined hyperparameters, and the prediction result is verified by using the visual analysis theory, so that the prediction distribution change before and after adding the transition region feature analysis can be more intuitively observed, and as the number of normal images in the data set increases, the error rate or accuracy of the polyp training result will change to a certain extent, therefore, in the polyp training process, the normal image data set and the polyp image data set should be balanced as much as possible, so as to prevent the phenomenon that the classification performance of the polyp training is unstable.

[0052] The working principle of the present application is that the polyp data set is classified by multi-level, multi-region and multi-feature fusion, and the feature learning parameters are reduced in dimension, so that the phenomenon of feature map over-fitting does not occur when the image data in the data set is complicated, the operation ability of data classification processing and detection recognition is guaranteed, and the efficiency of polyp detection is improved.

[0053] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1.A polyp detection method based on a convolutional neural network, characterized by, The method comprises the following steps: Step 1: Collecting original images of intestinal tract detection of patients; Step 2: Marking the salient region of the suspected polyp target in the image data, and taking the intercepted suspected polyp image as a target image based on the size of the identified frame in the image data set; Step 3: Establishing a polyp detection data set, including a normal image data set and a polyp data set, wherein the normal image data set includes normal examination images with suspected targets in previous case examinations, and the polyp data set includes examination images of the target to be detected; Step 4: Polyp detection data set processing, randomly selecting a pair of random images of the same category as the target image from the normal image data set, mapping the random images to a synthetic image of the target image by using a convolutional neural network model and a multi-branch spatial attention mechanism, and calculating the contrast loss of the synthetic image and the target image; Step 5: According to the verification result of the synthetic image and the target image, inputting the image set meeting the requirements of correlation and similarity into the classification network, and classifying and evaluating the synthetic image and the target image to verify the classification loss of the synthetic image and the target image; The sixth step is to add the transition region Ω between the target region and the background as an evaluation factor in the classification evaluation process, select characteristic values from the transition region Ω, compare the difference of the characteristic values, select the color level c, the blood vessel distribution density m and the smoothness s of the transition region Ω as the characteristic values, define the extracted characteristic map as the domain R, divide the domain R into the domain Ω1 containing the transition region characteristics and the domain R1 outside the transition region, respectively expand Ω1 and R1 into characteristic vectors with the dimension of c×m×s With and add an embedding layer to project it into an i-dimensional characteristic vector, i represents the number of sample categories in the data set, and the finally obtained characteristic vector is regarded as the prediction vector of the polyp training model for the transition region; the classification loss evaluation uses the maximum likelihood loss and the cosine similarity loss to reduce the prediction error rate, and the overall classification loss where k represents the cosine similarity loss, i, j represent the coordinates of a certain point on the feature vector. Step 7: Calculating the cross-entropy loss between the prediction probability distribution of the polyp training model and the real distribution of the image; Step 8: Use the contrast loss Ls between the synthesized image and the target image and the feature classification loss L of the transition region. cs The weighted sum of cross-entropy loss is used as the total loss function for training the polyp model. The smaller the total loss function value, the higher the training accuracy of the model. Step 9: Setting a total loss function threshold of the polyp training model, and directly outputting the polyp prediction result for the classification result with a total loss function less than the total loss function threshold, and returning the path of the generated synthetic image for re-detection of the suspected polyp image for the classification model exceeding the total loss function threshold. 2.The polyp detection method based on convolutional neural network according to claim 1, characterized in that, In the first step, the original image of intestinal tract detection of patients is collected by using colonoscopy technology, and the collected original image of intestinal tract detection of patients is a high-resolution image. The low-resolution image is processed by using super-resolution technology to improve the resolution of the image and reduce the influence of shaking on the resolution of the image during image collection. 3.The polyp detection method based on convolutional neural network according to claim 1, characterized in that, The third step is to establish a data set based on the automatic polyp detection and identification function of colonoscopy to collect multi-angle and multi-direction image data in the detection area, take the suspected target area as the reference, adjust the sampling angle a, sample the suspected target area image under one sampling angle, the sampling port is domain A, the area to be detected is domain B, domain A has feature vectors a(a1, a2, a3,..., an), domain B has feature vectors b(b1, b2, b3,..., bn), and the distance between point to point is The distance between domain A and domain B The boundary curve function of domain A records the sampling angle a and the sampling distance D. 4.The polyp detection method based on convolutional neural network according to claim 1, characterized in that, In the fourth step, the synthetic image is generated by setting a background image and rendering a polyp under the background image to obtain the synthetic image. The background image is set by using an environment simulator to set the polyp environment, including setting the humidity, temperature and blood vessel density of the polyp environment. 5.The polyp detection method based on convolutional neural network according to claim 1, characterized in that, The contrast loss function required in calculating the contrast loss of the synthetic image and the target image in the fourth step is N represents the number of samples, The correlation loss of the synthetic image and the target image is verified, The matching degree of the synthetic image and the target image is verified, Dw is the Euclidean distance between the feature X1 of the synthetic image and the feature X2 of the target image, that is, γ represents the influence factor in the optimization process, d is the dimension of the feature X1 of the synthetic image and the feature X2 of the target image, C S , C T represents the covariance matrix of the feature X1 of the synthetic image and the feature X2 of the target image in d dimensions, F is an input initial constant, the contrast loss of the synthetic image and the target image is verified, and the probability of false positive of polyps in sample classification can be reduced. 6.The polyp detection method based on convolutional neural network according to claim 1, characterized in that, The classification evaluation criteria in the sixth step are T r represents the accuracy, F L represents the error rate, M i represents the false negative rate, N T is the number of samples in the detection image that are correctly classified, N is the total number of detection image samples, N fp is the number of samples in the detection image that are correctly classified, N is the total number of detection image samples, N p-po is the number of samples in the detection image that are correctly classified, N is the total number of detection image samples, N fn is the number of samples in the detection image that are correctly classified, N is the total number of detection image samples, N po is the number of samples in the detection image that are correctly classified, N is the total number of detection image samples, N 7.The polyp detection method based on convolutional neural network according to claim 1, characterized in that, In the ninth step, the polyp training is a multiple cycle convolutional neural network model, which can improve the accuracy of the polyp prediction saliency detection result, integrate the saliency information of the transition region into the final polyp saliency prediction, highlight the saliency features of the entire image, accurately locate the boundary region of the suspected target, and obtain accurate result prediction.

Images