Method for identifying adenomatous polyp in colonoscope screening

Abstract

The invention relates to the technical field of medical image recognition, in particular to an adenomatous polyp recognition method in colonoscope screening. Acquiring a colonoscope image, and acquiring colonoscope image definition of the colonoscope image; screening the colonoscope image based on the colonoscope image definition to obtain a primarily screened image; a boundary recognizer is adopted to recognize the preliminary screening image, and a boundary recognition result is obtained; based on the boundary recognition result, screening the primary screening image, obtaining a secondary screening image, and recognizing the secondary screening image by adopting a structure recognizer to obtain a structure recognition result; and obtaining an adenomatous polyp recognition result based on the colonoscope image definition, the boundary recognition result and the structure recognition result. By implementing the method, efficient and accurate identification of the adenomatous polyp in the colonoscope image can be realized.

Description

Technical Field

[0001] This invention relates to the field of medical image recognition technology, and in particular to a method for identifying adenomatous polyps in colonoscopy screening. Background Technology

[0002] Current techniques for screening adenomatous polyps via colonoscopy largely rely on conventional image analysis, identifying polyp types by observing static morphological characteristics and combining them with single parameters such as diameter. This technique has significant drawbacks: adenomatous and hyperplastic polyps have similar static characteristics, making differentiation difficult with a single parameter and prone to misdiagnosis; dynamic interferences such as intestinal peristalsis distort static parameters; direct use of low-quality images leads to unstable identification; single models have limitations and low identification efficiency; relying solely on a single feature or parameter, ignoring multidimensional factors, results in insufficient reliability and fails to meet the needs of precise clinical screening. Summary of the Invention

[0003] This invention addresses the problem that existing technologies lack sufficient reliability in identifying adenomatous polyps, making it difficult to meet the needs of accurate clinical screening. It provides a method for identifying adenomatous polyps in colonoscopy screening.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0005] In a first aspect, the present invention provides a method for identifying adenomatous polyps in colonoscopy screening, comprising: acquiring colonoscopy images and acquiring the colonoscopy image clarity; screening the colonoscopy images based on the colonoscopy image clarity to obtain preliminary screening images; using a boundary recognizer to identify the preliminary screening images and obtaining boundary recognition results; screening the preliminary screening images based on the boundary recognition results to obtain secondary screening images; using a structure recognizer to identify the secondary screening images and obtaining structure recognition results; and obtaining adenomatous polyp identification results based on the colonoscopy image clarity, the boundary recognition results, and the structure recognition results.

[0006] Optionally, acquiring colonoscopy images and obtaining the colonoscopy image sharpness of the colonoscopy images includes: acquiring an original colonoscopy image and partitioning the colonoscopy image to obtain colonoscopy images; obtaining the colonoscopy image sharpness of the colonoscopy images, wherein the colonoscopy image sharpness is obtained by weighted calculation of Laplace variance and image variance.

[0007] The process of filtering colonoscopy images based on their clarity to obtain initial screening images includes: acquiring a set of historical colonoscopy images; acquiring the clarity of historical colonoscopy images based on the set of historical colonoscopy images and acquiring a clarity threshold; and filtering the colonoscopy images based on the clarity threshold to obtain images with a clarity greater than or equal to the clarity threshold as initial screening images.

[0008] The construction of the boundary recognizer includes: labeling colonoscopy images in the historical colonoscopy image set, with the labeling content being the boundary irregularity of adenomatous polyps, to obtain a sample boundary recognition result set; performing extraction with replacement on the historical colonoscopy image set and the sample boundary recognition result set to obtain N training datasets; constructing boundary recognition branches, using the N training datasets to train the boundary recognition branches respectively until convergence, to obtain N boundary recognition branches; and integrating the N boundary recognition branches to obtain the boundary recognizer.

[0009] Optionally, a boundary recognizer is used to recognize the initial screening image and obtain boundary recognition results, including: obtaining a branch selection coefficient based on the clarity of the colonoscopy image and the clarity of the historical colonoscopy image; obtaining a branch selection number based on the branch selection coefficient and a preset branch selection number; selecting a boundary recognition branch in the boundary recognizer; recognizing the initial screening image; and obtaining boundary recognition results, wherein the branch selection number is greater than or equal to 5.

[0010] Optionally, based on the boundary recognition result, the initial screening images are filtered to obtain secondary screening images. A structure recognizer is used to recognize the secondary screening images to obtain structure recognition results. This includes: filtering the initial screening images based on the boundary recognition result, removing initial screening images whose boundary recognition results are less than or equal to a preset boundary recognition result threshold, and obtaining secondary screening images; magnifying the secondary screening images to obtain magnified secondary screening images; and using the structure recognizer to recognize the magnified secondary screening images to obtain structure recognition results.

[0011] Optionally, the construction of the structure recognizer includes: magnifying the colonoscopy images in the historical colonoscopy image set to obtain magnified historical colonoscopy images; annotating the magnified historical colonoscopy images with the annotation content being the morphology of glandular openings to obtain a sample structure recognition result set; constructing a structure recognition branch, and training the structure recognition branch separately until convergence using the magnified historical colonoscopy images and the sample structure recognition result set to obtain the structure recognizer.

[0012] Optionally, based on the colonoscopy image clarity, the boundary recognition result, and the structure recognition result, the adenomatous polyp identification result is obtained, including: based on the historical colonoscopy image clarity, the sample boundary recognition result set, and the sample structure recognition result set, obtaining clarity similarity, boundary similarity, and structure similarity; based on the clarity similarity, the boundary similarity, and the structure similarity, obtaining clarity weight, boundary weight, and image weight; and based on the clarity weight, the boundary weight, and the image weight, performing a weighted summation of the colonoscopy image clarity, the boundary recognition result, and the structure recognition result to obtain the adenomatous polyp identification result.

[0013] Secondly, the present invention provides a system for identifying adenomatous polyps in colonoscopy screening, comprising:

[0014] An image clarity acquisition module is used to acquire colonoscopy images and obtain the colonoscopy image clarity of the colonoscopy images;

[0015] The initial screening image acquisition module is used to screen the colonoscopy images based on their clarity to acquire initial screening images;

[0016] The boundary recognition result acquisition module is used to identify the preliminary screening image using a boundary recognizer and acquire the boundary recognition result;

[0017] The structure recognition result acquisition module is used to filter the initial screening image based on the boundary recognition result, obtain the secondary screening image, and use a structure recognizer to recognize the secondary screening image to obtain the structure recognition result;

[0018] The final result acquisition module is used to obtain the adenomatous polyp identification result based on the clarity of the colonoscopy image, the boundary recognition result, and the structure recognition result.

[0019] By implementing this invention, it is possible to acquire colonoscopy images and obtain the clarity of the colonoscopy images. The clarity of the colonoscopy images, as a key initial evaluation parameter, can reflect the degree of interference in the image in advance, avoid subsequent analysis using blurry and distorted images, reduce invalid identification work from the source, and lay the foundation for the accuracy of subsequent screening and identification.

[0020] By implementing this invention, colonoscopy images can be screened based on their clarity to obtain preliminary screening images. This preliminary screening significantly reduces the amount of data required for subsequent processing, focuses on high-quality images, improves the efficiency of the entire recognition process, and avoids the risk of misjudgment caused by low-quality images.

[0021] By implementing this invention, a boundary recognizer can be used to identify the initial screening image and obtain boundary recognition results. Accurate boundary recognition results provide a key basis for subsequent rescreening, which helps to further distinguish different types of polyps and reduce misjudgments caused by unclear boundary features.

[0022] By implementing this invention, it is possible to filter the initial screening images based on the boundary recognition results, obtain secondary screening images, and use a structure recognizer to recognize the secondary screening images to obtain structure recognition results. The structure recognizer is specifically trained for the morphology of glandular openings and can accurately identify the unique glandular opening structures of adenomatous polyps, effectively distinguishing adenomatous polyps from hyperplastic polyps, and making up for the shortcomings of conventional single parameters in distinguishing them.

[0023] By implementing this invention, it is possible to obtain adenomatous polyp identification results based on the clarity of the colonoscopy image, the boundary recognition results, and the structure recognition results. By integrating multi-dimensional information, it is possible to comprehensively evaluate whether the polyps in the image are adenomatous polyps, significantly reduce the misdiagnosis rate, and improve the reliability of identification.

[0024] In summary, by implementing this invention, efficient and accurate identification of adenomatous polyps in colonoscopy images can be achieved. Attached Figure Description

[0025] Figure 1 A schematic flowchart of a method for identifying adenomatous polyps in colonoscopy screening provided by the present invention;

[0026] Figure 2 This is a schematic diagram of the structure of an adenomatous polyp identification system for colonoscopy screening provided by the present invention.

[0027] In the attached diagram, the components represented by each number are as follows:

[0028] Image clarity acquisition module 11, preliminary screening image acquisition module 12, boundary recognition result acquisition module 13, structure recognition result acquisition module 14, and final result acquisition module 15. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0031] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0032] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for identifying adenomatous polyps during colonoscopy screening, including:

[0033] S100: Acquire colonoscopy images and obtain the colonoscopy image clarity of the colonoscopy images;

[0034] S200: Based on the clarity of the colonoscopy images, the colonoscopy images are screened to obtain preliminary screening images;

[0035] S300: Use a boundary recognizer to identify the initial screening image and obtain the boundary recognition result;

[0036] S400: Based on the boundary recognition result, the initial screening image is filtered to obtain the secondary screening image. A structure recognizer is used to recognize the secondary screening image to obtain the structure recognition result.

[0037] S500: Based on the clarity of the colonoscopy image, the boundary recognition result, and the structure recognition result, obtain the adenomatous polyp recognition result.

[0038] In step S100 of this application embodiment, acquiring a colonoscopy image and acquiring the colonoscopy image clarity includes:

[0039] Acquire raw colonoscopy images and partition the colonoscopy images to obtain colonoscopy images;

[0040] The colonoscopy image sharpness is obtained by weighted calculation of Laplace variance and image variance.

[0041] In this embodiment, step S100 is the basic data preprocessing step in the adenomatous polyp identification process. The core objective is to provide high-quality, quantifiable image data for subsequent accurate screening and identification.

[0042] To achieve the above objectives, it is first necessary to acquire the original colonoscopy image and then partition the colonoscopy image to obtain the colonoscopy image.

[0043] Specifically, the process involves first acquiring unprocessed raw colonoscopy images from the colonoscopy equipment. Then, based on intestinal anatomy, such as the ascending colon, transverse colon, and descending colon, or according to the pixel distribution pattern, the raw colonoscopy images are divided into multiple independent local regions, such as N×M small regions divided by a grid. Each region serves as an independent analysis unit, ultimately resulting in a set of partitioned colonoscopy images.

[0044] Next, it is necessary to obtain the colonoscopy image clarity.

[0045] Specifically, for each partitioned local image, the Laplacian variance (denoted as V1) and image variance (denoted as V2) need to be calculated separately. The Laplacian variance measures the edge sharpness of the image—the sharper the edges, the greater the pixel grayscale variation calculated by the Laplacian operator, and the higher the variance value; conversely, the image is blurry, and the variance value is lower. The image variance measures the overall grayscale distribution variation of the image—the higher the overall image contrast, the more dispersed the grayscale distribution, and the higher the variance value; if there is slime coverage or blurring, the grayscale distribution is more uniform, and the variance value is lower.

[0046] Then, weighting coefficients for the Laplacian variance and image variance are set. These weights can be calibrated based on historical colonoscopy image data. For example, by comparing clear and blurry historical images, the contribution ratios of the Laplacian variance and image variance to sharpness can be determined and denoted as W1 and W2, respectively, where W1 + W2 = 1. Then, a weighted calculation is performed using the formula "Sharpness value = W1 × V1 + W2 × V2" to obtain the sharpness value of each colonoscopy image in the colonoscopy image set. Furthermore, if the sharpness of the entire image is required, the sharpness values ​​of all local colonoscopy images can be averaged to obtain the overall sharpness of the colonoscopy image.

[0047] In step S200 of this application embodiment, the colonoscopy images are screened based on their clarity to obtain preliminary screening images, including:

[0048] Obtain a set of historical colonoscopy images; based on the set of historical colonoscopy images, obtain the image clarity of the historical colonoscopy images and obtain a clarity threshold.

[0049] Based on the clarity threshold, the colonoscopy images are screened, and images with a clarity greater than or equal to the clarity threshold are selected as initial screening images.

[0050] In this embodiment, step S200 is the initial screening and filtering step in the adenomatous polyp identification process. The core objective is to remove low-quality blurry images caused by dynamic interference such as intestinal peristalsis, camera movement, and mucus coverage in advance, thereby reducing invalid data interference for subsequent accurate identification.

[0051] To achieve the above objectives, it is first necessary to obtain a set of historical colonoscopy images, and based on the set of historical colonoscopy images, obtain the clarity of the historical colonoscopy images and obtain a clarity threshold.

[0052] Specifically, it is necessary to collect a set of labeled historical colonoscopy images, such as images labeled "clear and polyps can be accurately identified" or "blurred and polyps cannot be identified". Then, following the Laplacian variance and image variance weighting method in step S100, the sharpness value of each image in the historical colonoscopy image set is calculated one by one to obtain the historical colonoscopy image sharpness dataset.

[0053] Then, through statistical analysis, a sharpness threshold is determined. For example, a distribution curve of the sharpness values ​​of historical colonoscopy images can be plotted to find the boundary between the "sharp image sharpness interval" and the "blurred image sharpness interval". For instance, if 95% of the sharp images in the historical colonoscopy image sharpness dataset have a sharpness ≥ 80, and 95% of the blurred images have a sharpness < 80, then the sharpness threshold is set to 80.

[0054] Next, the colonoscopy images need to be screened based on the clarity threshold, and images with a clarity greater than or equal to the clarity threshold are selected as initial screening images.

[0055] That is, the sharpness value of the colonoscopy image to be screened calculated in step S100 is called, which can be the overall sharpness of a single image or the sharpness of a section, and then the sharpness value of each colonoscopy image to be screened is compared with the "sharpness threshold".

[0056] Images with a sharpness value greater than or equal to the sharpness threshold are retained, while blurry images with a sharpness value less than the sharpness threshold are removed. The retained images are the "initial screening images" and can proceed to the boundary recognition step S300.

[0057] Furthermore, if the number of initially screened images is less than the preset number of initial screening images, colonoscopy images are reacquired, and the clarity of the colonoscopy images is assessed for further screening. For example, the preset number of initial screening images can be set to 10.

[0058] In step S300 of this application embodiment, a boundary recognizer is used to recognize the preliminary screening image and obtain the boundary recognition result, including:

[0059] Based on the clarity of the colonoscopy image and the clarity of the historical colonoscopy image, obtain the branch selection coefficient;

[0060] Based on the branch selection coefficient and the preset branch selection number, the branch selection number is obtained, and a boundary recognition branch is selected in the boundary recognizer to recognize the initial screened image and obtain the boundary recognition result, wherein the branch selection number is greater than or equal to 5.

[0061] In this embodiment, step S300 is the precise boundary extraction step in the adenomatous polyp identification process. The core objective is to improve the accuracy and robustness of polyp boundary identification by dynamically selecting boundary identification branches for clear images after initial screening.

[0062] To achieve the above objective, it is first necessary to obtain the branch selection coefficient based on the clarity of the colonoscopy image and the clarity of the historical colonoscopy image.

[0063] First, it is necessary to extract the sharpness value (denoted as C) of the current initial screening image in step S100 and the sharpness distribution characteristics of the historical image set in step S200, such as the mean μ and standard deviation σ of the historical sharpness.

[0064] Then, based on the preset algorithm, the difference between the clarity of the current colonoscopy image and the clarity of the historical images is converted into a branch selection coefficient K. K is a positive number, and the larger the value, the more branches are needed. Normalization calculation can be used, i.e., coefficient K = (C-μ) / σ + reference value. The reference value is an integer greater than or equal to 5. Finally, the value of K is rounded up to an integer. The maximum value of K is the total number of boundary recognition branches.

[0065] Based on the branch selection coefficient and the preset branch selection number, the branch selection number is obtained. Boundary recognition branches are then selected in the boundary recognizer to identify the initially screened image and obtain the boundary recognition result. Specifically, based on the K value calculated in the above steps, the top N boundary recognition branches are selected from the branch pool of the boundary recognizer according to their branch performance, or N boundary recognition branches are randomly selected.

[0066] The initially screened image is then input into N selected boundary recognition branches. Each branch outputs its own boundary prediction result, such as a pixel-level mask of the polyp region and a set of boundary coordinate points. The results of the N branches are then fused. For example, a "majority vote" is used for the pixel-level mask, where pixels marked as boundaries by more than half of the branches are considered the final boundaries. A "weighted average" is used for the coordinate point set, with the specific weights configured based on the historical accuracy of each branch. The fused boundary recognition result is then output, such as the complete polyp boundary contour and boundary confidence score, providing a foundation for subsequent structure recognition.

[0067] In step S300 of this application embodiment, the construction of the boundary identifier includes:

[0068] The colonoscopy images in the historical colonoscopy image set are labeled with the irregularity of the boundary of adenomatous polyps, and a sample boundary recognition result set is obtained.

[0069] The historical colonoscopy image set and the sample boundary recognition result set are extracted with replacement to obtain N training datasets.

[0070] Construct boundary recognition branches by training each of the N training datasets until convergence, thereby obtaining N boundary recognition branches.

[0071] The N boundary recognition branches are integrated to obtain a boundary recognizer.

[0072] In this embodiment of the application, the boundary recognizer is a core model for accurately identifying the boundary of adenomatous polyps. To obtain the boundary recognizer, it is first necessary to annotate the colonoscopy images in the historical colonoscopy image set, with the annotation content being the boundary irregularity of the adenomatous polyps, and then obtain a sample boundary recognition result set.

[0073] Specifically, a professional physician needs to select images containing adenomatous polyps from a historical colonoscopy image set. The physician or annotation team then annotates the boundaries of the polyps. The annotation includes not only the boundary contour (such as pixel-level mask), but also the quantification of boundary irregularity. This can be achieved by calculating the deviation value between the boundary contour and the fitted ellipse, the rate of change of boundary curvature, and other indicators, thus converting the degree of irregularity into a quantifiable numerical value or level.

[0074] Finally, the annotation results are integrated to form a "sample boundary recognition result set" that corresponds one-to-one with historical colonoscopy images, which includes boundary contour data and irregularity labels.

[0075] Furthermore, the historical colonoscopy image set and the sample boundary recognition result set need to be sampled with replacement to obtain N training datasets. First, the number of training sets N needs to be determined; the value of N is usually set according to the model complexity, such as N=10. Then, the "historical colonoscopy image set" and the corresponding "sample boundary recognition result set" are sampled with replacement: each time, samples with the same number of samples as the original dataset are randomly drawn from the original dataset to generate one training dataset. This process is repeated N times to obtain N independent but partially overlapping training datasets, each dataset containing colonoscopy image data and corresponding boundary annotations.

[0076] Furthermore, it is necessary to construct boundary recognition branches. Using the N training datasets, the boundary recognition branches are trained separately until convergence, resulting in N boundary recognition branches. The N boundary recognition branches are then integrated to obtain a boundary recognizer.

[0077] Considering the purpose of the boundary recognition branch, it can be built using a convolutional neural network, and the boundary recognizer can be built using a multi-branch convolutional neural network integrated architecture.

[0078] Each boundary recognition branch includes an input layer, a feature extraction layer, and a boundary prediction layer.

[0079] The feature extraction layer comprises two convolutional blocks. Specifically, the first convolutional block contains two convolutional layers with 3×3 kernels, numbered 32 and 64 respectively, using ReLU activation and a max-pooling layer. The second convolutional block contains two convolutional layers with 3×3 kernels, numbered 128 and 256 respectively, using ReLU activation and a max-pooling layer.

[0080] The boundary prediction layer consists of a fully connected layer 1 and an output layer. Fully connected layer 1 has 512 neurons, with ReLU activation. The output layer contains two sub-outputs: a "boundary contour mask" and a "boundary irregularity score." The irregularity score is a single numerical value ranging from 0 to 10, quantifying the degree of irregularity. Both output parameters use the Sigmoid activation function.

[0081] The boundary recognizer integrates the selected branch outputs using a weighted average fusion method. For the boundary contour mask, weights are assigned based on the branch's historical accuracy (IoU value on the validation set), while for the boundary irregularity, weights are assigned inversely proportional to the branch prediction error (MSE value). The final output is the fused boundary recognition result, including a "boundary contour integrity score" (0-10 points, 10 points for integrity) and a "boundary irregularity metric" (0-10 points, higher values ​​indicate greater irregularity).

[0082] For the boundary recognizer parameter settings, the optimizer uses the Adam optimizer with a learning rate of 0.001 and a decay rate of 0.0001. The loss function is a joint loss function, including "boundary contour loss" (Intersection over Union loss IoULoss) and "boundary irregularity loss" (Mean Squared Error MSELoss), with a weight ratio of 7:3. Each batch of input consists of 32 colonoscopy images. L2 regularization with a coefficient of 0.0005 is added to the fully connected layer to prevent overfitting.

[0083] For training the boundary recognizer, at least 10,000 sample boundary recognition results are obtained from a historical colonoscopy image set and divided into a training set and a validation set in an 8:2 ratio. The maximum number of training rounds is set to 100. The model is considered converged when the joint loss function value of the validation set is below 0.05 for three consecutive rounds, and the mean intersection-over-union (IoU) of the boundary contour is ≥0.85 and the mean absolute error (MAE) of the boundary irregularity is ≤0.5. The boundary recognizer is then obtained.

[0084] In step S400 of this application embodiment, based on the boundary recognition result, the initial screening image is filtered to obtain a secondary screening image. A structure recognizer is then used to recognize the secondary screening image to obtain a structure recognition result, including:

[0085] Based on the boundary recognition results, the initial screening images are filtered out, and the initial screening images with boundary recognition results less than or equal to a preset boundary recognition result threshold are removed to obtain the secondary screening images.

[0086] The rescreening image is magnified to obtain an magnified rescreening image;

[0087] The structure recognizer is used to identify the magnified and screened image to obtain the structure recognition result.

[0088] In this embodiment of the application, step S400 is the precise structural analysis step in the adenomatous polyp identification process. The core objective is to further focus on high-value images based on the initial screening images, enhance the fine structural features through magnification processing, and ultimately achieve accurate identification of polyps.

[0089] To achieve the above objectives, the initial screening images need to be filtered based on the boundary recognition results. Initial screening images with boundary recognition results less than or equal to a preset boundary recognition result threshold are removed to obtain secondary screening images.

[0090] The boundary recognition results output in step S300 are extracted, including the aforementioned boundary contour integrity score and boundary irregularity quantification value. Then, a preset boundary recognition result threshold is set. For example, based on historical data calibration, the preset boundary recognition result threshold is: boundary contour integrity score ≥ 7 points, and boundary irregularity quantification value ≥ 3 points.

[0091] Compare the boundary recognition results of the initial screening images with the boundary recognition result threshold: retain the images that meet the boundary recognition result threshold, and remove the images that do not meet the boundary recognition result threshold. The retained images are the "secondary screening images".

[0092] Furthermore, when the number of rescreened images is less than the preset number of rescreened images, colonoscopy images are reacquired, and the clarity boundary recognition results of the colonoscopy images are obtained for further filtering. For example, the preset number of rescreened images can be set to 10.

[0093] Furthermore, the rescreening image needs to be magnified to obtain an magnified rescreening image. Specifically, the polyp region in the rescreening image can be located, for example, based on the boundary contour in the boundary recognition result, the local area containing the polyp can be cropped to reduce background interference; then, a bicubic interpolation algorithm is used to magnify the cropped polyp region, taking into account both the smoothness and detail preservation of the magnified image, for example, magnifying the original 256×256 pixel polyp region to 512×512 pixels; the pixel values ​​of the magnified image are normalized, such as mapping the pixel values ​​to the range of 0-255, to ensure consistent image brightness and contrast, and the "magnified rescreening image" is output.

[0094] Furthermore, the structure recognizer is required to identify the magnified and screened image to obtain the structure recognition result.

[0095] The image to be enlarged and screened will be input into the structure recognizer, which will then output the structure recognition result through multi-layer feature extraction.

[0096] In step S400 of this application embodiment, the construction of the structure identifier includes:

[0097] The colonoscopy images in the historical colonoscopy image set are magnified to obtain magnified historical colonoscopy images;

[0098] The magnified historical colonoscopy images were annotated with the morphology of glandular openings to obtain a sample structure recognition result set;

[0099] A structure recognition branch is constructed. The structure recognition branch is trained using the magnified historical colonoscopy images and the sample structure recognition result set until convergence, thereby obtaining the structure recognizer.

[0100] First, the colonoscopy images in the historical colonoscopy image set need to be magnified to obtain magnified historical colonoscopy images. This involves filtering images containing adenomatous polyps from the historical colonoscopy image set, and then cropping out a local area containing only the polyps based on the marked polyp boundaries in the historical colonoscopy images to eliminate interference from the intestinal background. The polyp boundary marking method here is the same as the boundary contour acquisition method for the sample boundary recognition result set in step S300. Then, using the same magnification method as the rescreening images in step S400, the cropped polyp area is magnified to a preset size to ensure that the magnified historical image is the same size as the subsequent magnified rescreening images to be identified, and the "magnified historical colonoscopy image" is output.

[0101] Next, the magnified historical colonoscopy images need to be annotated, with the annotation focusing on the morphology of the glandular openings, to obtain a sample structure recognition result set. First, annotation standards for glandular opening morphology are established, referencing clinical classifications such as regular, irregular, villi-like, and disappearing types. Then, endoscopists or a professional annotation team annotate the morphology of the glandular openings within the polyp area frame by frame, based on the magnified historical colonoscopy images. The annotation content includes: the type of glandular opening, its distribution density, and the degree of morphological abnormality (quantitative scoring, 0-10 points). The annotation results are then correlated with the corresponding magnified historical colonoscopy images to form a "sample structure recognition result set."

[0102] Then, a structure recognition branch is constructed. The structure recognition branch is trained using the magnified historical colonoscopy images and the sample structure recognition result set until convergence, thereby obtaining the structure recognizer.

[0103] Considering the task type of the structure recognizer, the structure recognition branch can be built using a convolutional neural network architecture with an attention mechanism.

[0104] The structure recognition branch includes an input layer, a feature extraction layer, and a classification and regression layer. The input layer receives magnified and re-screened images.

[0105] The feature extraction layer consists of a first convolutional block, a second convolutional block, and an attention mechanism layer. The first convolutional block contains two convolutional layers with 3×3 kernels (32 and 64 kernels respectively), using ReLU activation and a max-pooling layer. The second convolutional block also contains two convolutional layers with 3×3 kernels (128 and 256 kernels respectively), using ReLU activation and a max-pooling layer. The attention mechanism layer is a spatial attention layer added after the second convolutional block, which learns a weight matrix to highlight features in the region where the glandular openings are located.

[0106] The classification and regression layer consists of a fully connected layer 1 and an output layer. Fully connected layer 1 contains 512 neurons, with ReLU activation. The output layer contains two sub-outputs: one is "glandular opening type," activated using the Softmax activation function; the other is "glandular opening morphological abnormality score" (quantitative value, range 0-10), activated using the Sigmoid activation function.

[0107] For the parameter settings of the structure recognition branch, the optimizer uses the Adam optimizer with a learning rate of 0.001. The loss function is a joint loss function, including "glandular opening type loss" (cross-entropy loss) and "morphological abnormality scoring loss" (mean squared error loss), with a weight ratio of 6:4. Each batch inputs 16 magnified historical colonoscopy images. L2 regularization with a coefficient of 0.0005 is added to the fully connected layer to prevent overfitting.

[0108] For the training of the structure recognition branch, the training samples are the magnified historical colonoscopy images and the sample structure recognition result set. At least 8000 magnified historical colonoscopy images are prepared and divided into training and validation sets in an 8:2 ratio. The maximum number of training rounds is set to 80, and an early stopping mechanism is used: if the validation set loss does not decrease for 8 consecutive rounds, training is terminated early. The model is considered converged when the joint loss function value of the validation set is below 0.06 for 3 consecutive rounds, the glandular opening type recognition accuracy is ≥0.89, and the mean absolute error of the morphological abnormality score is ≤0.7. The structure recognizer is then obtained.

[0109] In step S500 of this application embodiment, the adenomatous polyp identification result is obtained based on the clarity of the colonoscopy image, the boundary recognition result, and the structure recognition result, including:

[0110] Based on the clarity of the historical colonoscopy images, the sample boundary recognition result set, and the sample structure recognition result set, the clarity similarity, boundary similarity, and structural similarity are obtained.

[0111] Based on the sharpness similarity, the boundary similarity, and the structural similarity, sharpness weight, boundary weight, and image weight are obtained;

[0112] Based on the clarity weight, the boundary weight, and the image weight, the clarity of the colonoscopy image, the boundary recognition result, and the structure recognition result are weighted and summed to obtain the adenomatous polyp recognition result.

[0113] In this embodiment of the application, step S500 is the final result determination stage of the adenomatous polyp identification process. The core objective is to integrate the key data from the preceding stages and output accurate and reliable identification results through similarity calculation, weight allocation, and weighted fusion logic, thereby solving the problems that a single parameter is difficult to distinguish between adenomatous polyps and hyperplastic polyps, and that dynamic interference causes instability of static parameters.

[0114] To achieve the above objectives, it is first necessary to obtain the similarity in clarity, boundary similarity, and structural similarity based on the clarity of the historical colonoscopy images, the sample boundary recognition result set, and the sample structure recognition result set.

[0115] For clarity similarity, the clarity value of the current colonoscopy image in step S100 can be extracted and denoted as Ccurrent, and the clarity values ​​of all adenomatous polyp samples in the historical colonoscopy image set can be denoted as Chistory1, Chistory2...Chistoryn;

[0116] Then, the similarity is calculated using either "cosine similarity" or "Euclidean distance normalization". For example, the sharpness similarity = 1 - |C_current - averageC_history| / (maxC_history - minC_history), where averageC_history is the mean sharpness of all adenomatous polyp samples in the historical colonoscopy image set, and max / min is the maximum / minimum sharpness of all adenomatous polyp samples in the historical colonoscopy image set, ensuring that the result is mapped to the range of 0-1.

[0117] For boundary similarity, the boundary recognition results of the current colonoscopy image in step S300 can be extracted, such as the boundary irregularity quantification value B_current, and the boundary irregularity of historical adenomatous polyp samples in the sample boundary recognition result set, B_history1...B_historyn, etc.

[0118] Using the same calculation logic as for clarity similarity, the difference between the current B and the average historical B is compared to obtain the boundary similarity. The boundary similarity is in the range of 0-1. The closer the boundary irregularity is to the historical adenomatous polyp sample, the higher the similarity.

[0119] For structural similarity, the structural recognition results of the current colonoscopy image in step S400 can be extracted, such as the abnormal morphological score of the glandular opening S_current, and the abnormal morphological scores of historical adenomatous polyp samples in the sample structural recognition result set, S_history1…S_historyn.

[0120] Similarly, by comparing the current S with the historical average S, the structural similarity is obtained. The structural similarity ranges from 0 to 1. The closer the morphology of the glandular opening is to the historical adenomatous polyp sample, the higher the similarity.

[0121] Furthermore, based on the sharpness similarity, the boundary similarity, and the structural similarity, it is necessary to obtain the sharpness weight, boundary weight, and image weight. Specifically, based on the principle that "the higher the similarity, the greater the weight," the three similarities need to be normalized to obtain the weights for each item.

[0122] For example, let the sharpness similarity be S_clear, the boundary similarity be S_edge, and the structural similarity be S_structure; the denominator for calculating the weights is: total similarity = S_clear + S_edge + S_structure; the weights for each item are calculated as follows: sharpness weight W_clear = S_clear / total similarity; boundary weight W_edge = S_edge / total similarity; image weight W_structure = S_structure / total similarity. Here, since structural recognition is based on the core features of the magnified image, the "image weight" is the weight corresponding to the structural features.

[0123] Furthermore, based on the clarity weight, the boundary weight, and the image weight, the colonoscopy image clarity, the boundary recognition result, and the structure recognition result need to be weighted and summed to obtain the adenomatous polyp recognition result.

[0124] Specifically, the image sharpness, boundary recognition results, and structure recognition results of the current colonoscopy image need to be extracted and normalized to the range of 0-1 to ensure consistent magnitude: normalized colonoscopy image sharpness value C_standard = (C_current - minC_history) / (maxC_history - minC_history); normalized boundary recognition result B_standard = (B_current - minB_history) / (maxB_history - minB_history); normalized structure recognition result S_standard = (S_current - minS_history) / (maxS_history - minS_history).

[0125] Then calculate the weighted total score: Identification score = W_clear × C_standard + W_edge × B_standard + W_terminal × S_standard; then set the scoring threshold, which can be calibrated based on historical data, such as setting it to 0.7. If the identification score is ≥0.7, it is judged as "adenomatous polyp", and if the identification score is <0.7, it is judged as "non-adenomatous polyp".

[0126] The final identification result is output, which includes a classification conclusion on whether it is an adenomatous polyp and the corresponding identification score.

[0127] Example 2, as Figure 2 As shown, based on the same inventive concept as the method for identifying adenomatous polyps in colonoscopy screening provided in Embodiment 1, this embodiment of the invention also provides a system for identifying adenomatous polyps in colonoscopy screening, comprising:

[0128] Image clarity acquisition module 11 is used to acquire colonoscopy images and acquire the colonoscopy image clarity of the colonoscopy images;

[0129] The initial screening image acquisition module 12 is used to screen the colonoscopy image based on the clarity of the colonoscopy image to acquire an initial screening image;

[0130] The boundary recognition result acquisition module 13 is used to identify the preliminary screening image using a boundary recognizer and acquire the boundary recognition result;

[0131] The structure recognition result acquisition module 14 is used to filter the initial screening image based on the boundary recognition result, obtain the secondary screening image, and use a structure recognizer to recognize the secondary screening image to obtain the structure recognition result.

[0132] The final result acquisition module 15 is used to acquire the adenomatous polyp identification result based on the clarity of the colonoscopy image, the boundary recognition result, and the structure recognition result.

[0133] Furthermore, the image sharpness acquisition module 11 includes the following execution steps:

[0134] Acquire raw colonoscopy images and partition the colonoscopy images to obtain colonoscopy images;

[0135] The colonoscopy image sharpness is obtained by weighted calculation of Laplace variance and image variance.

[0136] Furthermore, the initial screening image acquisition module 12 includes the following execution steps:

[0137] Obtain a set of historical colonoscopy images; based on the set of historical colonoscopy images, obtain the image clarity of the historical colonoscopy images and obtain a clarity threshold.

[0138] Based on the clarity threshold, the colonoscopy images are screened, and images with a clarity greater than or equal to the clarity threshold are selected as initial screening images.

[0139] Furthermore, the boundary recognition result acquisition module 13 includes the following execution steps:

[0140] Based on the clarity of the colonoscopy image and the clarity of the historical colonoscopy image, obtain the branch selection coefficient;

[0141] Based on the branch selection coefficient and the preset branch selection number, the branch selection number is obtained, and a boundary recognition branch is selected in the boundary recognizer to recognize the initial screened image and obtain the boundary recognition result, wherein the branch selection number is greater than or equal to 5.

[0142] The construction of the boundary identifier includes:

[0143] The colonoscopy images in the historical colonoscopy image set are labeled with the irregularity of the boundary of adenomatous polyps, and a sample boundary recognition result set is obtained.

[0144] The historical colonoscopy image set and the sample boundary recognition result set are extracted with replacement to obtain N training datasets.

[0145] Construct boundary recognition branches by training each of the N training datasets until convergence, thereby obtaining N boundary recognition branches.

[0146] The N boundary recognition branches are integrated to obtain a boundary recognizer.

[0147] Furthermore, the structure recognition result acquisition module 14 includes the following execution steps:

[0148] Based on the boundary recognition results, the initial screening images are filtered out, and the initial screening images with boundary recognition results less than or equal to a preset boundary recognition result threshold are removed to obtain the secondary screening images.

[0149] The rescreening image is magnified to obtain an magnified rescreening image;

[0150] The structure recognizer is used to identify the magnified and screened image to obtain the structure recognition result.

[0151] The construction of the structure identifier includes:

[0152] The colonoscopy images in the historical colonoscopy image set are magnified to obtain magnified historical colonoscopy images;

[0153] The magnified historical colonoscopy images were annotated with the morphology of glandular openings to obtain a sample structure recognition result set;

[0154] A structure recognition branch is constructed. The structure recognition branch is trained using the magnified historical colonoscopy images and the sample structure recognition result set until convergence, thereby obtaining the structure recognizer.

[0155] Furthermore, the final result acquisition module 15 includes the following execution steps:

[0156] Based on the clarity of the historical colonoscopy images, the sample boundary recognition result set, and the sample structure recognition result set, the clarity similarity, boundary similarity, and structural similarity are obtained.

[0157] Based on the sharpness similarity, the boundary similarity, and the structural similarity, sharpness weight, boundary weight, and image weight are obtained;

[0158] Based on the clarity weight, the boundary weight, and the image weight, the clarity of the colonoscopy image, the boundary recognition result, and the structure recognition result are weighted and summed to obtain the adenomatous polyp recognition result.

[0159] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0160] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0161] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0162] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0163] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0164] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0165] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for identifying adenomatous polyps during colonoscopy screening, characterized in that, include: Acquire colonoscopy images and obtain the colonoscopy image clarity of the colonoscopy images; Based on the clarity of the colonoscopy images, the colonoscopy images are screened to obtain preliminary screening images; A boundary recognizer is used to identify the initial screening image to obtain the boundary recognition result; Based on the boundary recognition results, the initial screening images are filtered to obtain secondary screening images. A structure recognizer is then used to recognize the secondary screening images to obtain structure recognition results. Based on the clarity of the colonoscopy image, the boundary recognition result, and the structure recognition result, the adenomatous polyp recognition result is obtained.

2. The method for identifying adenomatous polyps in colonoscopy screening according to claim 1, characterized in that, Acquire colonoscopy images, and obtain the colonoscopy image clarity of the colonoscopy images, including: Acquire raw colonoscopy images and partition the colonoscopy images to obtain colonoscopy images; The colonoscopy image sharpness is obtained by weighted calculation of Laplace variance and image variance.

3. The method for identifying adenomatous polyps in colonoscopy screening according to claim 1, characterized in that, Based on the clarity of the colonoscopy images, the colonoscopy images are screened to obtain preliminary screening images, including: Obtain a set of historical colonoscopy images; based on the set of historical colonoscopy images, obtain the image clarity of the historical colonoscopy images and obtain a clarity threshold. Based on the clarity threshold, the colonoscopy images are screened, and images with a clarity greater than or equal to the clarity threshold are selected as initial screening images.

4. The method for identifying adenomatous polyps in colonoscopy screening according to claim 3, characterized in that, The construction of the boundary recognizer includes: The colonoscopy images in the historical colonoscopy image set are labeled with the irregularity of the boundary of adenomatous polyps, and a sample boundary recognition result set is obtained. The historical colonoscopy image set and the sample boundary recognition result set are extracted with replacement to obtain N training datasets. Construct boundary recognition branches by training each of the N training datasets until convergence, thereby obtaining N boundary recognition branches. The N boundary recognition branches are integrated to obtain a boundary recognizer.

5. The method for identifying adenomatous polyps in colonoscopy screening according to claim 1, characterized in that, The initial screening image is identified using a boundary recognizer to obtain boundary recognition results, including: Based on the clarity of the colonoscopy image and the clarity of the historical colonoscopy image, obtain the branch selection coefficient; Based on the branch selection coefficient and the preset branch selection number, the branch selection number is obtained, and a boundary recognition branch is selected in the boundary recognizer to recognize the initial screened image and obtain the boundary recognition result, wherein the branch selection number is greater than or equal to 5.

6. The method for identifying adenomatous polyps in colonoscopy screening according to claim 1, characterized in that, Based on the boundary recognition results, the initial screening images are filtered to obtain secondary screening images. A structure recognizer is then used to recognize the secondary screening images to obtain structure recognition results, including: Based on the boundary recognition results, the initial screening images are filtered out, and the initial screening images with boundary recognition results less than or equal to a preset boundary recognition result threshold are removed to obtain the secondary screening images. The rescreening image is magnified to obtain an magnified rescreening image; The structure recognizer is used to identify the magnified and screened image to obtain the structure recognition result.

7. The method for identifying adenomatous polyps in colonoscopy screening according to claim 1, characterized in that, The construction of the structure identifier includes: The colonoscopy images in the historical colonoscopy image set are magnified to obtain magnified historical colonoscopy images; The magnified historical colonoscopy images were annotated with the morphology of glandular openings to obtain a sample structure recognition result set; A structure recognition branch is constructed. The structure recognition branch is trained using the magnified historical colonoscopy images and the sample structure recognition result set until convergence, thereby obtaining the structure recognizer.

8. The method for identifying adenomatous polyps in colonoscopy screening according to claim 1, characterized in that, Based on the clarity of the colonoscopy image, the boundary recognition result, and the structure recognition result, the adenomatous polyp identification result is obtained, including: Based on the clarity of the historical colonoscopy images, the sample boundary recognition result set, and the sample structure recognition result set, the clarity similarity, boundary similarity, and structural similarity are obtained. Based on the sharpness similarity, the boundary similarity, and the structural similarity, sharpness weight, boundary weight, and image weight are obtained; Based on the clarity weight, the boundary weight, and the image weight, the clarity of the colonoscopy image, the boundary recognition result, and the structure recognition result are weighted and summed to obtain the adenomatous polyp recognition result.

9. A system for identifying adenomatous polyps in colonoscopy screening, characterized in that, A method for identifying adenomatous polyps in colonoscopy screening as described in any one of claims 1-8, comprising: An image clarity acquisition module is used to acquire colonoscopy images and obtain the colonoscopy image clarity of the colonoscopy images; The initial screening image acquisition module is used to screen the colonoscopy images based on their clarity to acquire initial screening images; The boundary recognition result acquisition module is used to identify the preliminary screening image using a boundary recognizer and acquire the boundary recognition result; The structure recognition result acquisition module is used to filter the initial screening image based on the boundary recognition result, obtain the secondary screening image, and use a structure recognizer to recognize the secondary screening image to obtain the structure recognition result; The final result acquisition module is used to obtain the adenomatous polyp identification result based on the clarity of the colonoscopy image, the boundary recognition result, and the structure recognition result.

Images