False Positive Identification Method for Polyp Detection Based on Cross-Level Incomplete Connection Mechanism

Abstract

The present invention discloses a method for identifying false positives in polyp detection based on a cross-level incomplete connection mechanism, belonging to the technical field of colonoscopy polyp detection. It includes constructing a spatial multi-path mechanism to locate the characteristic regions of polyps shown on the feature map. The feature fusion main body fuses lesion features of different sizes using feature maps of three scales to form sets of positive sample data and negative sample data. After training the positive samples of the same batch and the same scale, the cosine similarity loss is calculated. The present invention locates the characteristic regions of polyps on the feature map through the spatial multi-path mechanism, forms sets of positive sample data and negative sample data, calculates the cosine similarity loss after training the positive samples of the same batch, and at the same time combines a single-stage model to perform multi-scale sampling on the negative sample data, and combines the maximum likelihood estimation to select the negative sample with the highest confidence score, improving the problem of high false alarm rates in detecting colon polyps using existing R-CNN methods and active learning-based methods.

Description

Technical Field

[0001] The present invention relates to a method for identifying false positives in polyp detection, in particular to a method for identifying false positives in polyp detection based on a cross-level incomplete connection mechanism, belonging to the technical field of colonoscopy polyp detection. Background Art

[0002] Adenomatous polyps detected during colonoscopy are common origins of colorectal cancer. Timely detection and resection of adenomatous polyps can reduce the risk of developing colorectal cancer. However, in current colonoscopies, complex environments such as air bubbles, lens reflections, residues, and fold shadows in the intestinal environment can all be recognized as polyp-like features, resulting in false positives for adenomatous polyps and affecting the identification of true adenomatous polyps.

[0003] With the development of computer technology, deep learning detection models have been widely used in polyp detection methods. Among them, the R-CNN (Region-CNN) method first finds false positive regions, adds the detection samples to the training set, and then retrains the overall network detection system. There is also a method using active learning to process false positives detected by the CADe system. Both of these methods first find potential false positive regions and then perform image processing and identification on these regions. The process is very complex and requires a large computational cost. Therefore, the present invention provides a method that can use a multi-path spatial attention method to pay more attention to the background of the region, reduce false positive false alarms, and at the same time use the cross-level incomplete connection mechanism after multi-level connection of the model network to improve the computational efficiency of the model and save computational costs. Summary of the Invention

[0004] The main purpose of the present invention is to solve the deficiencies of the R-CNN method and the method based on active learning in detecting colon polyps, which are prone to false positive false alarms, have a complex process, and a high computational cost, and to provide a method for identifying false positives in polyp detection based on a cross-level incomplete connection mechanism.

[0005] The object of the present invention can be achieved by adopting the following technical solutions:

[0006] A method for identifying false positives in polyp detection based on a cross-level incomplete connection mechanism includes the following steps:

[0007] S1. Construct a spatial multi-path mechanism to locate the feature regions showing polyps on the feature map. The feature fusion main body uses feature maps of three scales to fuse different-sized lesion features to form positive sample data and negative sample data sets;

[0008] S2. Calculate the cosine similarity loss after training with positive samples of the same batch and the same scale. After multi-scale sampling of the negative sample dataset based on the single-stage model, use the maximum likelihood estimation loss to select the negative sample with the highest confidence score, output the confidence score, class probability, and regression boundary of the image, and circle the negative sample as a suspicious target area. Minimize the image box with a high score and optimize all image boxes into negative sample areas as the computational graph for the next stage;

[0009] S3. Use CSPNet to divide the gradient flow on the computational graph, enable the image gradient flow to propagate on different network paths, increase the number of gradient flow paths through splitting and merging methods, and use the cross-stage method to reduce the use of explicit feature map copies for splicing.

[0010] As a further solution of the present invention, in S1, the mapping function of the spatial multi-path mechanism is related to the convolution kernel size of the convolutional neural network. The mapping function formula of the spatial multi-path mechanism is:

[0011] ;

[0012] In the formula: Denoted as the mapping function of the spatial multi-path mechanism, Denoted as the sigmoid function of the activation function of the convolutional neural network. The scale of the model is denoted as , and it satisfies: , Denoted as the independent variable of the sigmoid function , Denoted as the convolution operation with a convolution kernel size of given the input X, where The convolution kernel corresponds to the first receptive field in the convolutional neural network, The convolution kernel corresponds to the second receptive field in the convolutional neural network, The convolution kernel corresponds to the third receptive field in the neural network. The area size of the first receptive field is denoted as , the area size of the second receptive field is denoted as , the area size of the third receptive field is denoted as , , and Satisfy the relationship: .

[0013] As a further solution of the present invention, when training negative samples in S2, when calculating the credibility score, select the top 50 anchor boxes from 150 anchor boxes at three scales to score the negative anchor boxes, calculate the explicit maximum likelihood loss on the sampling mechanism at different scales, and calculate the cosine similarity loss on the sampling mechanism at each same scale.

[0014] As a further solution of the present invention, the feature fusion main body described in S1 includes three upsampling modules and two downsampling modules.

[0015] As a further solution of the present invention, the maximum likelihood loss is calculated by using the cross-entropy loss function in combination with the size of the target image region, the serial number of the anchor point of the anchor box, and the predicted classification score of the anchor point. The formula of the maximum likelihood loss function is:

[0016] ;

[0017] In the formula: is denoted as the maximum likelihood function value, is denoted as the area of the target image region, is denoted as the serial number of the anchor point of the anchor box, is denoted as the data set of the serial numbers of the anchor points of the anchor box, , is denoted as the binary cross-entropy loss function, is denoted as the predicted classification score of the i-th anchor point, is denoted as the standardized value of.

[0018] As a further solution of the present invention, the cosine similarity loss is obtained by substituting the feature vectors of the positive sample and the negative sample as variables of the cosine similarity loss function into the cosine similarity loss function. The calculation formula of the cosine similarity loss value is:

[0019] ;

[0020] In the formula: is denoted as the cosine similarity loss value, is denoted as the highest predicted classification score in the positive sample, is denoted as the highest predicted classification score of the negative sample paired with in the negative sample. j is denoted as the minimum value of , and i is denoted as the starting value of the scoring value of each scale image. is denoted as the scoring value of each image, and respectively represent the feature vectors of the positive sample and the negative sample with a score of i.

[0021] Advantageous technical effects of the present invention: According to the polyp detection false positive identification method based on the cross-level incomplete connection mechanism of the present invention, the characteristic regions of polyps on the feature map are located through the spatial multi-path mechanism, and the lesion characteristics of different scales are fused to form a positive sample data set and a negative sample data set. After training the positive samples of the same batch, the cosine similarity loss is calculated. At the same time, a single-stage model is combined to perform multi-scale sampling on the negative sample data, and the negative sample with the highest confidence score is selected by combining the maximum likelihood estimation, improving the problem of high false positive rate in detecting colon polyps using the existing R-CNN method and the method based on active learning. At the same time, the created cross-level incomplete connection mechanism can increase the number of image gradient flows, thereby reducing the disadvantage of using explicit feature map copies for splicing, reducing the number of calculation parameters, saving the calculation cost, and improving the calculation efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flowchart of the polyp detection false positive identification method based on the cross-level incomplete connection mechanism of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] To make the technical solutions of the present invention clearer and more definite for those skilled in the art, the present invention will be further described in detail below with reference to the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0024] As Figure 1 shown, the polyp detection false positive identification method based on the cross-level incomplete connection mechanism provided in this embodiment includes the following steps:

[0025] S1. Construct a spatial multi-path mechanism to locate the characteristic regions of polyps shown on the feature map. The feature fusion main body fuses the lesion characteristics of different sizes using three scales of feature maps to form a positive sample data and a negative sample data set;

[0026] S2. After training the positive samples of the same batch and the same scale, calculate the cosine similarity loss. Based on the single-stage model, perform multi-scale sampling on the negative sample data set, and then use the maximum likelihood estimation loss to select the negative sample with the highest confidence score, output the confidence score, class probability, and regression boundary of the image, and circle the negative sample as the suspicious target area. Minimize the image box with a high score and optimize all image boxes into negative sample areas as the computational graph for the next stage;

[0027] S3. Use CSPNet to divide the gradient flow on the computational graph, make the image gradient flow propagate on different network paths, increase the number of gradient flow paths through the method of splitting and merging, and use the cross-stage method to reduce the use of explicit feature map copies for splicing.

[0028] As a further solution of the present invention, the mapping function of the spatial multi-path mechanism in S1 is related to the convolution kernel size of the convolutional neural network, and the mapping function formula of the spatial multi-path mechanism is:

[0029] ;

[0030] In the formula: is denoted as the mapping function of the spatial multi-path mechanism, is denoted as the sigmoid function of the activation function of the convolutional neural network, and the scale of the model is denoted as , and it satisfies: , is denoted as the sigmoid function independent variable, is denoted as the convolution operation with a convolution kernel size of when the given input is X, where the convolution kernel corresponds to the first receptive field in the convolutional neural network, the convolution kernel corresponds to the second receptive field in the convolutional neural network, the convolution kernel corresponds to the third receptive field in the neural network. The area size of the first receptive field is denoted as , the area size of the second receptive field is denoted as , the area size of the third receptive field is denoted as , , and satisfy the relational formula: .

[0031] When training negative samples in S2, when calculating the confidence score, the top 50 anchor boxes are selected from 150 anchor boxes at three scales to score the negative anchor boxes, and the explicit maximum likelihood loss is calculated on the sampling mechanisms at different scales, and the cosine similarity loss is calculated on the sampling mechanisms at each same scale.

[0032] The feature fusion main body described in S1 includes three upsampling modules and two downsampling modules.

[0033] The maximum likelihood loss is calculated using the cross-entropy loss function in combination with the target image region size, the serial number of the anchor box anchor point, and the predicted classification score of the anchor point. The maximum likelihood loss function formula is:

[0034] ;

[0035] In the formula: is denoted as the maximum likelihood function value, is denoted as the area of the target image region, is denoted as the serial number of the anchor box anchor point, is denoted as the data set of the serial numbers of the anchor box anchor points, , is denoted as the binary cross-entropy loss function, is denoted as the predicted classification score of the i-th anchor point, is denoted as the normalized value of.

[0036] The cosine similarity loss takes the feature vectors of the positive and negative samples as variables of the cosine similarity loss function and substitutes them into the cosine similarity loss function to obtain the cosine similarity loss value. The calculation formula of the cosine similarity loss value is:

[0037] ;

[0038] In the formula: is denoted as the cosine similarity loss value, is denoted as the highest predicted classification score in the positive samples, is denoted as the highest predicted classification score of the negative sample paired with the negative sample, j is denoted as the minimum value of, i is denoted as the starting value of the scoring value of each scale image, is denoted as the scoring value of each image, and respectively represent the feature vectors of the positive and negative samples with a score of i. Embodiment

[0039] The colon polyp dataset collected and labeled from the colorectal and anorectal surgery department of the hospital is used as the first dataset, and the public datasets of CVC-Clinic2015 and Etis-Larib are used as the second dataset: The first dataset contains 1720 colon polyp images of 175 patients. The images in the first dataset are divided into a training set and a test set according to a ratio of 4:1. Combining the actual application scenarios of colonoscopes, the clarity, brightness, and deformation of the images are transformed to expand the first dataset into 3582 images. 1000 ordinary images without colon polyp annotation information are added to the training set to construct a new training set. The original images vary in size from 612×524 to 1280×720, so the original images are all adjusted to 512×512; The CVC-Clinic2015 data in the second dataset contains 612 well-defined standard polyp pictures, extracted from 29 different classifications, each classification consists of 6 to 26 frames, and each classification contains images of a type of polyp at different angles, distances, and perspectives. The clear boundary lines are marked on each polyp. The Etis-Larib data contains 196 polyp images with a resolution of 1255×966, including 34 classifications, and each classification has 44 different polyp images.

[0040] Ablation experiment on the first data set: An ablation experiment is performed on the first data set by using the spatial multipath mechanism and the inspection function combining the maximum likelihood loss and the cosine loss.

[0041] Ablation experiment on the second data set: The model is trained using the data packets on CVC-Clinic2015, and then the trained model is evaluated using the Etis-Larib data packets.

[0042] Set the evaluation criteria:

[0043] Use the percentage of true-positive cases among true-positive and false-positive pathologies as the first evaluation criterion, and use the true-positive cases among true-positive and false-negative cases as the second evaluation criterion. Denote the first evaluation criterion as and denote the second evaluation criterion as The comprehensive evaluation index is related to both the first and second evaluation criteria. The formula for the comprehensive evaluation index is:

[0044] ;

[0045] In the formula: is denoted as the comprehensive evaluation index.

[0046] Use Pytorch to train the model, set the training batch size to 8, randomly rotate the image angle and scale the image. The training consists of 2000 segments, and each segment is iterated 574 times.

[0047] The present invention locates the characteristic regions of polyps on the feature map through the spatial multipath mechanism, fuses the lesion characteristics of different scales, forms a positive sample data set and a negative sample data set, calculates the cosine similarity loss after training the positive samples in the same batch, and at the same time combines the single-stage model to perform multi-scale sampling on the negative sample data, and combines the maximum likelihood estimation to select the negative sample with the highest confidence score, improving the problem of high false positive rate in detecting colon polyps using the existing R-CNN method and the active learning-based method. At the same time, the created cross-level incomplete connection mechanism can increase the number of image gradient flows, thereby reducing the disadvantage of using explicit feature map copies for splicing, reducing the number of calculation parameters, saving the calculation cost, and improving the calculation efficiency of the model.

[0048] The above is only a further embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the scope disclosed by the present invention, according to the technical solution and its concept of the present invention, makes equivalent substitutions or changes, all belong to the protection scope of the present invention.

Claims

1. A false positive identification method for polyp detection based on a cross-level incomplete connection mechanism, characterized in that It includes the following steps: S1. Construct a spatial multi-path mechanism to locate the feature region showing polyps on the feature map. The feature fusion main body fuses lesion features of different sizes using feature maps of three scales to form sets of positive sample data and negative sample data; S2. After training positive samples of the same batch and the same scale, calculate the cosine similarity loss. After multi-scale sampling of the negative sample dataset based on the single-stage model, use the maximum likelihood estimation loss to select the negative sample with the highest confidence score, output the confidence score, class probability, and regression boundary of the image, and circle the negative sample as a suspicious target region. Minimize the image box with a high score and optimize all image boxes into negative sample regions as the computational graph for the next stage; S3. Use CSPNet to divide the gradient flow on the computational graph, make the image gradient flow propagate on different network paths, increase the number of gradient flow paths through splitting and merging methods, and use the cross-stage method to reduce the use of stitching of display feature map copies.

2. The false positive identification method for polyp detection based on a cross-level incomplete connection mechanism according to claim 1, wherein, The mapping function of the spatial multi-path mechanism in S1 is related to the convolution kernel size of the convolutional neural network. The formula for the mapping function of the spatial multi-path mechanism is: ； In the formula: Denoted as the mapping function of the spatial multipath mechanism, Denoted as the sigmoid function of the convolutional neural network activation function, and the scale of the model is denoted as and satisfies: , Denoted as the sigmoid function The independent variable of Denoted as the convolution operation with a convolution kernel size of when the given input is X, where The convolution kernel of corresponds to the first receptive field in the convolutional neural network, The convolution kernel of corresponds to the second receptive field in the convolutional neural network, The convolution kernel of corresponds to the third receptive field in the neural network. The area size of the first receptive field is denoted as The area size of the second receptive field is denoted as The area size of the third receptive field is denoted as , , and Satisfy the relational expression: .

3. The false positive identification method for polyp detection based on a cross-level incomplete connection mechanism according to claim 1, wherein When training negative samples in S2, when calculating the credibility score, select the top 50 anchor boxes from 150 anchor boxes at three scales to score the negative anchor boxes, calculate the explicit maximum likelihood loss on the sampling mechanisms of different scales, and calculate the cosine similarity loss on the sampling mechanisms of each same scale.

4. The polyp detection false positive identification method based on a cross-level incomplete connection mechanism according to claim 1, characterized in that The feature fusion main body described in S1 includes three upsampling modules and two downsampling modules.

5. The false positive identification method for polyp detection based on a cross-level incomplete connection mechanism according to claim 3, characterized in that, The maximum likelihood loss is calculated using the cross-entropy loss function in combination with the size of the target image region, the serial number of the anchor point of the anchor box, and the predicted classification score of the anchor point. The formula for the maximum likelihood loss function is: ； In the formula: Denoted as the maximum likelihood function value, Denoted as the area of the target image region, Denoted as the serial number of the anchor point of the anchor box, Denoted as the data set of the serial numbers of the anchor points of the anchor box, , Denoted as the binary cross-entropy loss function, Denoted as the predicted classification score of the i-th anchor point, Denoted as The standardized value of.

6. The false positive identification method for polyp detection based on a cross-level incomplete connection mechanism according to claim 3, characterized in that The cosine similarity loss substitutes the feature vectors of positive and negative samples as variables of the cosine similarity loss function into the cosine similarity loss function to obtain the cosine similarity loss value. The calculation formula for the cosine similarity loss value is: ； Wherein: Denoted as the cosine similarity loss value, denoted as the highest predicted classification score in the positive samples, Denoted as in the negative samples The highest predicted classification score of the negative sample paired with, j is denoted as The minimum value of, i is denoted as the starting value of the scoring value of each scale image, Denoted as the scoring value of each graph, And Respectively represent the feature vectors of the positive samples and negative samples with a score of i.

Images