Specific image feature mining method and system for cervical cancer cell nuclei
By constructing a cell nuclear segmentation model based on Xception+FPN+Unet, the specific image features in the cervical cell sample images were extracted, and the problem of insufficient interpretability and accuracy of the cell classification task in cervical cancer detection was solved, and more stable and Reliable cell classification.
Patent Information
- Application Number
- CN202410408566.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-07
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-04-07
AI Technical Summary
The interpretability and accuracy of cell classification tasks in existing cervical cancer detection is poor, especially due to the low classification accuracy caused by the black box nature of deep learning models and the scale invariance of convolutional neural networks.
The nuclear segmentation model based on Xception+FPN+Unet structure was constructed. Through the data-labeled nuclear material training model, the nuclear specific image characteristics in the cervical cell sample images were extracted, including the nuclear coordinates, outline information, elliptical approximation, color depth and Nuclear membrane anomaly, used to determine cell types.
The stability and accuracy of cell classification results are improved, the interpretability and reliability of classification results are enhanced, and the problem of insufficient interpretability and accuracy of cell classification tasks in the prior art is solved.
Smart Images

Figure CN118333957B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and in particular to a method and system for mining specific image features of cervical cancer cell nuclei. Background Art
[0002] Cervical cancer, also known as cervical cancer, is one of the most common malignant tumors in women. Its incidence rate ranks second only to breast cancer. In recent years, the incidence rate of cervical cancer has increased and tends to be younger. For cervical cancer, early detection and treatment are the key. Regular gynecological examinations such as cervical cytology and HPV examinations are important means to prevent cervical cancer.
[0003] At present, the commonly used cervical cytology test is TCT cervical cancer screening, which uses a liquid-based thin-layer cell kit to collect exfoliated cells from the cervical os, uses a fully automatic thin-layer cell slide making machine to slide, and performs cytological classification based on the morphology of the cell nucleus to diagnose whether there are symptoms of cancer. In TCT cervical cancer screening, the diagnosis is mainly made by confirming ASCUS (atypical squamous epithelial cells) and non-hollow LSIL (low-grade squamous intraepithelial lesion-superficial type). Among them, the characteristics and morphological characteristics of ASCUS cells under the microscope mainly include: cells with low nuclear-cytoplasmic ratio, increased nuclear-cytoplasmic ratio, nuclear changes, cell nuclei that are 2.5 to 3 times the nuclei of middle-layer squamous cells, and irregular nuclear membranes; the characteristics and morphological characteristics of non-hollow LSIL cells under the microscope mainly include: atypical superficial squamous epithelial cells with high nuclear-cytoplasmic ratio, cell nuclei that are more than 3 times the nuclei of middle-layer squamous cells, and irregular nuclear membranes.
[0004] The cell classification task in the existing TCT-assisted diagnosis and analysis system is used to distinguish normal negative cells from ASCUS cells and non-hollow LSIL cells. This is mainly accomplished by collecting data sets and training deep learning convolutional neural network models. Since deep learning models are black box models with poor interpretability, although they can achieve certain classification effects by fitting data sets, they still cannot provide a reasonable visual explanation for why they are classified in this way, whether they are consistent with the logic of pathologists in diagnosing such cells, and why false positives and false negatives occur. In addition, convolutional neural networks theoretically have only weak scale invariance. By studying the characteristics and morphological features of ASCUS and non-hollow LSIL cells under the microscope, it can be seen that the nuclei of these two types of positive cells vary in size. Therefore, the accuracy of using convolutional neural networks to classify such cells is poor. Summary of the invention
[0005] The purpose of the present invention is to provide a method and system for mining specific image features of cervical cancer cell nuclei, so as to solve the problems of poor interpretability and accuracy of cell classification tasks in existing cervical cancer detection.
[0006] In order to solve the above technical problems, the present invention provides a method for mining specific image features of cervical cancer cell nuclei, comprising:
[0007] Obtain images of cervical cell samples;
[0008] Construct a cell nucleus segmentation model;
[0009] Using the cell nucleus segmentation model, the cell nucleus region in the cervical cell sample image is analyzed and processed to extract specific image features;
[0010] The cell category is determined based on the extracted specific image features.
[0011] Optionally, in the method for mining specific image features of cervical cancer cell nuclei, the method for constructing a cell nucleus segmentation model includes:
[0012] Build a cell nucleus segmentation model based on Xception+FPN+Unet structure;
[0013] Collect data-annotated nuclear materials;
[0014] The constructed cell nucleus segmentation model is trained using the cell nucleus material with data annotation.
[0015] Optionally, in the method for mining specific image features for cervical cancer cell nuclei, the method of using a cell nucleus segmentation model to analyze and process the cell nucleus region in the cervical cell sample image to extract specific image features includes:
[0016] The cervical cell sample image is divided into image blocks according to a preset size with the cell as the center;
[0017] The cell nucleus segmentation model automatically locates the coordinates and contour information of the cell nucleus based on the image block;
[0018] The specific image features of the cell nucleus are calculated based on the cell nucleus coordinates and contour information.
[0019] Optionally, in the method for mining specific image features for cervical cancer cell nuclei, the method by which the cell nucleus segmentation model automatically locates the cell nucleus coordinates and contour information according to the image block includes: locating the cell nucleus coordinates and contour information by using the cv2.findContours method.
[0020] Optionally, in the method for mining specific image features of cervical cancer cell nuclei, the method for calculating specific image features of cell nuclei according to cell nucleus coordinates and contour information includes:
[0021] According to the coordinates and contour information of the cell nucleus, an image of the cell nucleus contour area is captured;
[0022] Count the number of non-zero pixels in the image of the cell nucleus contour area to obtain the mask area of the cell nucleus;
[0023] Obtain an ellipse that is similar to the cell nucleus, and use the area and contour overlap between the cell nucleus and the approximate ellipse to measure the ellipse approximation of the cell nucleus;
[0024] Count the color distribution of the cell nucleus to obtain the color depth of the cell nucleus;
[0025] The nuclear membrane heteromorphism of the cell nucleus was calculated based on the coordinates and contour information of the cell nucleus.
[0026] Optionally, in the method for mining specific image features of cervical cancer cell nuclei, the method of obtaining an ellipse similar to the cell nucleus and measuring the ellipse approximation of the cell nucleus by using the area and contour overlap between the cell nucleus and the approximate ellipse includes:
[0027] Obtain the contour points of multiple cell nuclei;
[0028] Use the contour points of the cell nucleus to generate an ellipse that is similar to the cell nucleus;
[0029] If the contour point of the cell nucleus falls on the contour line of the approximate ellipse, the contour point is judged as a matching contour point;
[0030] The contour coincidence is obtained by using the proportion of the number of matching contour points in the total number of contour points of the cell nucleus;
[0031] The area overlap was obtained using the ratio of the masked area of the cell nucleus to the area of the approximate ellipse;
[0032] The average of the contour overlap and the area overlap was taken as the ellipse approximation of the cell nucleus.
[0033] Optionally, in the method for mining specific image features of cervical cancer cell nuclei, the method for counting the color distribution of cell nuclei to obtain the color depth of cell nuclei includes:
[0034] Convert the captured cell nucleus contour area image to RGB color space;
[0035] Count the color distribution of each color channel;
[0036] According to the color distribution, calculate the color probability of each color channel;
[0037] The color depth of the cell nucleus is calculated using the color probability, where the calculation formula is:
[0038] CD=Pilog2P
[0039] Among them, CD is the color depth of the cell nucleus, P i is the color probability of color channel i.
[0040] Optionally, in the method for mining specific image features for cervical cancer cell nuclei, the method for calculating the nuclear membrane heteromorphism of the cell nucleus according to the cell nucleus coordinates and contour information includes:
[0041] Use Gaussian smoothing to soften the nucleus contour, where the Gaussian smoothing formula is:
[0042]
[0043] Among them, (x, y) is the point coordinate of the cell nucleus contour; σ is an adjustable parameter, the larger the σ value, the higher the smoothness;
[0044] Get the convex points and concave points on the outline of the cell nucleus, and calculate the depth between adjacent convex points and concave points, where the calculation formula is:
[0045]
[0046] Where d is the depth, (x1, y1) is the coordinate of the concave point in the concave area on the nucleus contour, and (y2, y2) is the coordinate of the convex point on the nucleus contour closest to the concave point obtained by the convex hull algorithm;
[0047] Take each concave point and the adjacent convex points on both sides as a point set, and calculate the angle between the concave point and the convex point in the point set. The calculation formula is:
[0048]
[0049] Among them, θ is the angle, a is the vector formed by the concave point and the convex point on one side, and b is the vector formed by the concave point and the convex point on the other side;
[0050] According to the preset rules, the depth value and angle value are used to obtain the nuclear membrane heteromorphism of the cell nucleus.
[0051] Optionally, in the method for mining specific image features for cervical cancer cell nuclei, the method for determining cell types based on the extracted specific image features includes:
[0052] According to the pathological diagnosis language, the thresholds for nuclear area, nuclear ellipse approximation, nuclear staining depth, and nuclear membrane heteromorphism were set.
[0053] If the mask area of the cell nucleus is greater than the cell nucleus area threshold, the ellipse approximation is greater than the cell nucleus ellipse approximation threshold, the color depth is greater than the cell nucleus staining depth threshold, and the nuclear membrane heteromorphism is greater than the nuclear membrane heteromorphism threshold, then the cell corresponding to the cell nucleus is judged to be an ASCUS cell; otherwise, it is a non-hollow LSIL cell.
[0054] In order to solve the above technical problems, the present invention further provides a specific image feature mining system for cervical cancer cell nuclei, which is used to implement the specific image feature mining method for cervical cancer cell nuclei as described in any one of the above items, and the specific image feature mining system for cervical cancer cell nuclei comprises:
[0055] An image acquisition module, used for acquiring cervical cell sample images;
[0056] A feature extraction module, including a cell nucleus segmentation model, is used to analyze and process the cell nucleus region in the cervical cell sample image to extract specific image features;
[0057] A category judgment module is used to judge the cell category based on the extracted specific image features;
[0058] The result output module is used to output the judgment results of cell categories.
[0059] The method and system for mining specific image features of cervical cancer cell nuclei provided by the present invention include: obtaining cervical cell sample images; constructing a cell nucleus segmentation model; using the cell nucleus segmentation model to analyze and process the cell nucleus region in the cervical cell sample image to extract specific image features; and judging the cell category based on the extracted specific image features. By constructing a cell nucleus segmentation model to extract specific image features from the cell nucleus to classify the cells, the classification results will not be affected by the size and position of the cells in the image, thereby improving the stability and accuracy of the cell classification results; at the same time, since the classification relies on specific image features, the cell classification results are more interpretable and reliable, solving the problem of poor interpretability and accuracy of cell classification tasks in existing cervical cancer detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 A flow chart of the specific image feature mining method for cervical cancer cell nuclei provided in this embodiment;
[0061] Figure 2 A structural block diagram of the cell nucleus segmentation model provided in this embodiment;
[0062] Figure 3 A schematic diagram of the structure of the specific image feature mining system for cervical cancer cell nuclei provided in this embodiment. DETAILED DESCRIPTION
[0063] The following is a further detailed description of the method and system for mining specific image features of cervical cancer cell nuclei proposed by the present invention in combination with the accompanying drawings and specific embodiments. It should be noted that the accompanying drawings are all in a very simplified form and use non-precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. In addition, the structure shown in the accompanying drawings is often a part of the actual structure. In particular, the emphasis that each drawing needs to show is different, and sometimes different proportions are used.
[0064] It should be noted that the terms "first", "second", etc. in the specification, claims and drawings of the present invention are used to distinguish similar objects in order to describe the embodiments of the present invention, rather than to describe a specific order or sequence. It should be understood that the structures used in this way can be interchanged under appropriate circumstances. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0065] This embodiment provides a method for mining specific image features of cervical cancer cell nuclei. Figure 1 As shown, including:
[0066] S1, obtaining cervical cell sample images;
[0067] S2, construct the cell nucleus segmentation model;
[0068] S3, using the cell nucleus segmentation model, analyzing and processing the cell nucleus region in the cervical cell sample image to extract specific image features;
[0069] S4, judging the cell category based on the extracted specific image features.
[0070] The specific image feature mining method for cervical cancer cell nuclei provided in this embodiment extracts specific image features from cell nuclei to classify cells by constructing a cell nucleus segmentation model. The classification results will not be affected by the size and position of the cells in the image, thereby improving the stability and accuracy of the cell classification results. At the same time, since the classification relies on specific image features, the cell classification results are more interpretable and reliable, thereby solving the problem of poor interpretability and accuracy of cell classification tasks in existing cervical cancer detection.
[0071] It should be noted that the order of the method steps proposed in this embodiment can be adjusted according to actual needs. For example, a cell nucleus segmentation model can be constructed first, and then a cervical cell sample image can be obtained. Also, in practical applications, other steps can be added between each step to optimize or increase functions, and this application does not limit this. The adjustment and addition of the order of steps without violating the purpose of this application should also fall within the scope of protection of this application.
[0072] Furthermore, in this embodiment, step S1, the method for obtaining a cervical cell sample image mainly includes: obtaining a cervical cell sample image using a TCT cervical cancer screening method. The specific operating steps of TCT cervical cancer screening are well known to those skilled in the art and will not be described in detail in this application.
[0073] Furthermore, in this embodiment, step S2, the method of constructing a cell nucleus segmentation model includes:
[0074] S21, build a cell nucleus segmentation model based on Xception+FPN+Unet structure.
[0075] Specifically, Figure 2 As shown in the figure, the cell nucleus segmentation model mainly includes the separation convolution structure in the Xception network, the feature pyramid structure of FPN and the Unet network structure.
[0076] The separation convolution structure in the Xception network includes multiple sequentially arranged convolution layers and batch normalization layers, separation convolution layers and batch normalization layers. Its function is to achieve channel separation, so that the cervical cell sample image is separated into different channels through different convolution layers, and then enters the feature pyramid structure of FPN corresponding to different channels for processing, thereby improving the efficiency of the cell nucleus segmentation model. Figure 2 The middle part is the left side of the dotted frame.
[0077] The feature pyramid structure of FPN includes a feature extraction structure corresponding to the number of channels of the separated convolution structure in the Xception network. The feature extraction structure mainly includes a convolution layer and a batch normalization layer, a convolution layer, and an upsampling layer. By fusing the results extracted by each feature extraction structure, the network performance can be effectively improved. Figure 2 The dotted frame part is in the middle.
[0078] In addition, the Unet network structure includes convolutional layers, batch normalization layers, upsampling layers, relu activation and sigmoid activation processes, etc., so as to predict the fused features output by the feature pyramid structure of FPN and obtain the predicted mask of the cell nucleus in the cervical cell sample image. Figure 2The part in the middle is on the right side of the dotted frame.
[0079] exist Figure 2 In the cell nucleus segmentation model structure shown in the figure, Conv1 represents a 1×1 convolution operation, Conv3 represents a 3×3 convolution operation, Upsample represents an upsampling operation, BN represents a batch normalization operation, ⊕ represents a concat operation, relu represents an activation function, Scale:1, Scale:2, Scale:4, and Scale:8 represent upsampling scaling factor multiples, and the operation S represents a Sigmoid activation function.
[0080] S22, collect data-annotated nuclear materials.
[0081] Specifically, in the present embodiment, the collected cell nucleus materials are mainly cell nucleus images that have been data-annotated. In order to improve the performance of the cell nucleus segmentation model, in the present embodiment, the cell nucleus materials are images obtained by segmenting the cells according to a preset size with the cells as the center. For example, an image with a size of 1024×1024 pixels is obtained with the cells as the center, so that training can be performed on situations where cells have different proportions and positions in the image, thereby avoiding the weak scale invariance in existing convolutional neural networks and improving the performance of the cell nucleus segmentation model.
[0082] In practical applications, the cell nucleus material can be annotated manually, or with the help of relevant algorithm programs, or a combination of the two. The more accurate the data annotation, the better the training effect of the cell nucleus segmentation model.
[0083] Of course, during the actual use of the cell nucleus segmentation model, the output results can be used as training data for iteration, so that the cell nucleus segmentation model has more and more training data, thereby making its prediction effect more efficient and accurate.
[0084] S23, using the data-labeled cell nucleus materials to train the constructed cell nucleus segmentation model.
[0085] The specific training method, process and confirmation of training effect are well known to those skilled in the art and will not be elaborated in this application.
[0086] Furthermore, in this embodiment, in step S3, the method of analyzing and processing the cell nucleus region in the cervical cell sample image using the cell nucleus segmentation model to extract specific image features includes:
[0087] S31, dividing the cervical cell sample image into image blocks according to a preset size with the cells as the center.
[0088] Specifically, in this embodiment, the size of the image block should be consistent with the image size used in training, for example, it can be 1024×1024 pixels. Of course, in other embodiments, image blocks of other sizes can also be selected, as long as they are consistent with the image size used in training, so as to ensure that the image blocks can be accurately analyzed to obtain accurate cell classification results.
[0089] S32, the cell nucleus segmentation model automatically locates the cell nucleus coordinates and contour information according to the image block.
[0090] Specifically, in this embodiment, the coordinates and contour information of the cell nucleus can be located by the cv2.findContours method. The cv2.findContours function is a common method for performing image contour search, and its specific implementation process is well known to those skilled in the art, and this application will not repeat it in detail.
[0091] S33, calculating specific image features of the cell nucleus according to the cell nucleus coordinates and contour information.
[0092] Specifically, in this embodiment, the method for calculating the specific image features of the cell nucleus includes:
[0093] S331, capturing a cell nucleus contour area image according to the cell nucleus coordinates and contour information.
[0094] S332, counting the number of non-zero pixels in the cell nucleus contour area image to obtain the mask area of the cell nucleus.
[0095] Since in this embodiment, the cv2.findContours function is used to search for the nucleus contour, the pixels in the nucleus contour area image are black and white pixels, wherein the black pixels (non-zero pixels) represent the nucleus area. By counting the number of non-zero pixels, the area size of the current nucleus can be known.
[0096] S333, obtaining an ellipse that is similar to the cell nucleus, and using the area and contour overlap between the cell nucleus and the approximate ellipse to measure the ellipse approximation of the cell nucleus.
[0097] In a specific embodiment, first, multiple contour points of the cell nucleus are obtained. The more contour points there are, the more accurate the calculated ellipse approximation is, but it will also take up more computing resources. Therefore, in practical applications, it is necessary to comprehensively consider accuracy and computing resources to determine the number range of contour points. Then, the contour points of the cell nucleus are used to generate an ellipse similar to the cell nucleus. Specifically, the approximate ellipse can be automatically generated according to the contour point fitting with the help of a related algorithm model. Next, if the contour point of the cell nucleus falls on the contour line of the approximate ellipse, the contour point is judged to be a matching contour point; the proportion of the number of matching contour points in the total number of contour points of the cell nucleus is used to obtain the contour overlap, that is, the contour overlap = the number of matching contour points / the total number of contour points. Afterwards, the ratio of the mask area of the cell nucleus to the area of the approximate ellipse is used to obtain the area overlap, that is, the area overlap = the size of the mask area of the cell nucleus / the size of the area of the approximate ellipse. Finally, the average value of the contour overlap and the area overlap is taken as the ellipse approximation of the cell nucleus, that is, the ellipse approximation of the cell nucleus = (contour overlap + area overlap) / 2.
[0098] Of course, in other embodiments, weights may be set for the contour overlap and the area overlap according to actual conditions, and the ellipse approximation of the cell nucleus may be obtained by using a weighted average calculation method.
[0099] S334, counting the color distribution of the cell nucleus to obtain the color depth of the cell nucleus.
[0100] In a specific embodiment, first, the intercepted image of the cell nucleus contour area is converted to the RGB color space. Since the previous contour calculation obtains a mask image, in order to calculate the color depth here, a color picture must be obtained for accurate calculation. The conversion method of the image space is well known to those skilled in the art, and this application will not repeat it. Then, the color distribution of each color channel is counted, that is, the color distribution of the R channel, the G channel, and the B channel is counted. Then, according to the color distribution, the color probability of each color channel is calculated. Finally, the color depth of the cell nucleus is calculated using the color probability, where the calculation formula is:
[0101] CDPilog2P
[0102] Among them, CD is the color depth of the cell nucleus, P i is the color probability of color channel i.
[0103] S335, calculating the nuclear membrane heteromorphism of the cell nucleus according to the cell nucleus coordinates and contour information.
[0104] In a specific embodiment, first, Gaussian smoothing is used to soften the cell nucleus contour line, thereby effectively preventing the interference of the cell nucleus contour jaggedness on the calculation, wherein the Gaussian smoothing formula is:
[0105]
[0106] Where (x, y) is the coordinate of the point of the cell nucleus contour; σ is an adjustable parameter. The larger the σ value, the higher the smoothness. In practical applications, the σ value can be set according to actual needs.
[0107] Then, the convex points and concave points on the outline of the cell nucleus are obtained, and the depths between adjacent convex points and concave points are calculated, where the calculation formula is:
[0108]
[0109] Where d is the depth, (x1, y1) is the coordinate of the concave point in the concave area of the nucleus contour, and (x2, y2) is the coordinate of the convex point on the nucleus contour closest to the concave point obtained using the convex hull algorithm. The convex hull algorithm is a method for calculating the minimum convex polygon in a point set. The convex hull can be regarded as the minimum convex polygon that encloses the point set. In practical applications, concave points are usually defined as the difference between the contour and the convex hull. These differences can be regarded as concave areas, which are usually due to the irregular shape of the nucleus or concave areas.
[0110] Next, each concave point and the adjacent convex points on both sides are taken as a point set, and the angle between the concave point and the convex point in the point set is calculated, where the calculation formula is:
[0111]
[0112] Among them, θ is the angle, a is the vector formed by the concave point and the convex point on one side, and b is the vector formed by the concave point and the convex point on the other side. a and b can be obtained by using the coordinates of the concave point and the convex point. In practical applications, for each point set, the depth of the concave point and the two convex points needs to be calculated to obtain two depth values. The specific calculation method is the same as the above depth calculation method.
[0113] In this embodiment, the Euclidean distance is used to calculate the depth. Of course, in other embodiments, other methods may be selected to calculate the depth, such as Manhattan distance, Chebyshev distance, etc.
[0114] Finally, according to the preset rules, the nuclear membrane heteromorphism of the cell nucleus is obtained using the depth value and the angle value. The preset rules can be set according to the actual situation. By setting the range of the depth value and the angle value, the nuclear membrane heteromorphism is quantified, so as to facilitate the classification of cell nuclei with different nuclear membrane heteromorphism.
[0115] Furthermore, in this embodiment, in step S4, the method for determining the cell type based on the extracted specific image features includes:
[0116] S41, according to the pathological diagnosis language, set the cell nuclear area threshold, cell nuclear ellipse approximation threshold, cell nuclear staining depth threshold and cell nuclear membrane heteromorphism threshold.
[0117] As can be seen from the background technology, ASCUS cells and non-hollow LSIL cells have different characteristics, so these obvious characteristics can be used to set thresholds to accurately distinguish the two types of cells. In the method provided in this embodiment, each feature is separated from the cell nucleus size and is not affected by the image size. Therefore, a unified threshold can be set for a feature for judgment, which greatly simplifies the computational complexity of cell classification.
[0118] S42, if the mask area of the cell nucleus is greater than the cell nucleus area threshold, the ellipse approximation is greater than the cell nucleus ellipse approximation threshold, the color depth is greater than the cell nucleus staining depth threshold, and the nuclear membrane heteromorphism is greater than the nuclear membrane heteromorphism threshold, then the cell corresponding to the cell nucleus is judged to be an ASCUS cell; otherwise, it is a non-hollow LSIL cell.
[0119] The specific image feature mining method for cervical cancer cell nuclei provided in this embodiment, in the feature extraction stage, the selected feature extraction method is insensitive to the scale change of the image, for example, a method based on local features such as the depth of nuclear staining, the approximation of the ellipse of the nuclear cell, and the heteromorphism of the nuclear contour is used, so that the stability of the features at different scales can be maintained. At the same time, the specific image feature mining method for cervical cancer cell nuclei provided in this embodiment, in the feature extraction stage, the selected feature extraction method can extract similar features at different positions in the image, for example, a method based on local features such as the area of the nuclear cell, the depth of nuclear staining, the approximation of the ellipse of the nuclear cell, and the heteromorphism of the nuclear contour is used, which is not affected by the image position and rotation angle, so that the stability of the features at different positions and rotation angles can be maintained, so that the features are invariant to the translation and rotation of the image.
[0120] In addition, the specific image feature mining method for cervical cancer cell nuclei provided in this embodiment can also assist the existing convolutional neural image classification model. While achieving higher accuracy, it can also take into account the interpretability of the features, making it easier for pathologists and other non-related field personnel to understand the principles behind the classification and improve the acceptance of AI for auxiliary diagnosis.
[0121] Specifically, the cell nucleus segmentation model provided in this embodiment can obtain the segmentation result of the cell nucleus, and then can extract a series of morphological features (specific image features) that can better describe the cell, such as the size of the cell nucleus, ellipse approximation, color depth, etc. These specific image features are fused with the features extracted by the neural network image classification model to form a more comprehensive feature vector. The fused feature vector is input into the cell classification model, and the cell classification model determines whether the cell belongs to ASCUS or non-hollow LSIL cells based on the features contained in the feature vector and the pathological diagnosis language.
[0122] This embodiment also provides a specific image feature mining system for cervical cancer cell nuclei, which is used to implement the specific image feature mining method for cervical cancer cell nuclei as described above. Figure 3 As shown, the specific image feature mining system for cervical cancer cell nuclei includes: an image acquisition module for acquiring cervical cell sample images; a feature extraction module, including a cell nucleus segmentation model, for analyzing and processing the cell nucleus area in the cervical cell sample image to extract specific image features; a category judgment module, for judging the cell category based on the extracted specific image features; and a result output module, for outputting the judgment result of the cell category.
[0123] The specific image feature mining system for cervical cancer cell nuclei provided in this embodiment extracts specific image features from cell nuclei through the cell nucleus segmentation model in the feature extraction module to classify cells. The classification results will not be affected by the size and position of the cells in the image, thereby improving the stability and accuracy of the cell classification results. At the same time, since the classification relies on specific image features, the cell classification results are more interpretable and reliable, thus solving the problem of poor interpretability and accuracy of cell classification tasks in existing cervical cancer detection.
[0124] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other. In addition, the different parts between the various embodiments can also be used in combination with each other, and the present invention is not limited to this.
[0125] The method and system for mining specific image features for cervical cancer cell nuclei provided in this embodiment include: obtaining cervical cell sample images; constructing a cell nucleus segmentation model; using the cell nucleus segmentation model to analyze and process the cell nucleus region in the cervical cell sample image to extract specific image features; and judging the cell category based on the extracted specific image features. By constructing a cell nucleus segmentation model to extract specific image features from the cell nucleus to classify cells, the classification results will not be affected by the size and position of the cells in the image, thereby improving the stability and accuracy of the cell classification results; at the same time, since the classification relies on specific image features, the cell classification results are more interpretable and reliable, solving the problem of poor interpretability and accuracy of cell classification tasks in existing cervical cancer detection.
[0126] The above description is only a description of the preferred embodiments of the present invention, and is not intended to limit the scope of the present invention. Any changes or modifications made by a person skilled in the art in the field of the present invention based on the above disclosure shall fall within the scope of protection of the claims.
Claims
1. A method for mining specific image features of cervical cancer cell nuclei, characterized in that: include: Obtain images of cervical cell samples; Construct a cell nucleus segmentation model; Using the cell nucleus segmentation model, the cell nucleus region in the cervical cell sample image is analyzed and processed to extract specific image features, and the cervical cell sample image is segmented into image blocks of a preset size with the cell as the center; The cell nucleus segmentation model automatically locates the coordinates and contour information of the cell nucleus based on the image block; Calculate specific image features of the cell nucleus based on the cell nucleus coordinates and contour information; According to the coordinates and contour information of the cell nucleus, an image of the cell nucleus contour area is captured; Count the number of non-zero pixels in the image of the cell nucleus contour area to obtain the mask area of the cell nucleus; Obtain an ellipse that is similar to the cell nucleus, and use the area and contour overlap between the cell nucleus and the approximate ellipse to measure the ellipse approximation of the cell nucleus; Count the color distribution of the cell nucleus to obtain the color depth of the cell nucleus; According to the coordinates and contour information of the cell nucleus, the nuclear membrane heteromorphism of the cell nucleus is calculated; The cell category is determined based on the extracted specific image features.
2. The method for mining specific image features of cervical cancer cell nuclei according to claim 1, characterized in that: The method for constructing a cell nucleus segmentation model comprises: Build a cell nucleus segmentation model based on Xception+FPN+Unet structure; Collect data-annotated nuclear materials; The constructed cell nucleus segmentation model is trained using the cell nucleus material with data annotation.
3. The method for mining specific image features of cervical cancer cell nuclei according to claim 1, characterized in that: The method for automatically locating the coordinates and contour information of the cell nucleus according to the image block by the cell nucleus segmentation model includes: locating the coordinates and contour information of the cell nucleus by using the cv2.findContours method.
4. The method for mining specific image features of cervical cancer cell nuclei according to claim 1, characterized in that: The method of obtaining an ellipse similar to the cell nucleus and measuring the ellipse approximation of the cell nucleus by using the area and contour overlap between the cell nucleus and the approximate ellipse includes: Obtain the contour points of multiple cell nuclei; Use the contour points of the cell nucleus to generate an ellipse that is similar to the cell nucleus; If the contour point of the cell nucleus falls on the contour line of the approximate ellipse, the contour point is judged as a matching contour point; The contour coincidence is obtained by using the proportion of the number of matching contour points in the total number of contour points of the cell nucleus; The area overlap was obtained using the ratio of the masked area of the cell nucleus to the area of the approximate ellipse; The average of the contour overlap and the area overlap was taken as the ellipse approximation of the cell nucleus.
5. The method for mining specific image features of cervical cancer cell nuclei according to claim 1, characterized in that: The method of counting the color distribution of the cell nucleus to obtain the color depth of the cell nucleus includes: Convert the captured cell nucleus contour area image to RGB color space; Count the color distribution of each color channel; According to the color distribution, calculate the color probability of each color channel; The color depth of the cell nucleus is calculated using the color probability, where the calculation formula is: CD=∑P i log2P i Among them, CD is the color depth of the cell nucleus, P i is the color probability of color channel i.
6. The method for mining specific image features for cervical cancer cell nuclei according to claim 1, characterized in that: The method for calculating the nuclear membrane heteromorphism of the cell nucleus according to the cell nucleus coordinates and contour information comprises: Use Gaussian smoothing to soften the nucleus contour, where the Gaussian smoothing formula is: Among them, (x, y) is the point coordinate of the cell nucleus contour; σ is an adjustable parameter, the larger the σ value, the higher the smoothness; Get the convex points and concave points on the outline of the cell nucleus, and calculate the depth between adjacent convex points and concave points, where the calculation formula is: Where d is the depth, (x1, y1) is the coordinate of the concave point in the concave area on the nucleus contour, and (x2, y2) is the coordinate of the convex point on the nucleus contour closest to the concave point obtained using the convex hull algorithm; Take each concave point and the adjacent convex points on both sides as a point set, and calculate the angle between the concave point and the convex point in the point set. The calculation formula is: Among them, θ is the angle, a is the vector formed by the concave point and the convex point on one side, and b is the vector formed by the concave point and the convex point on the other side; According to the preset rules, the depth value and angle value are used to obtain the nuclear membrane heteromorphism of the cell nucleus.
7. The method for mining specific image features for cervical cancer cell nuclei according to claim 1, characterized in that: The method for determining the cell type based on the extracted specific image features includes: According to the pathological diagnosis language, the thresholds for nuclear area, nuclear ellipse approximation, nuclear staining depth, and nuclear membrane heteromorphism were set. If the mask area of the cell nucleus is greater than the cell nucleus area threshold, the ellipse approximation is greater than the cell nucleus ellipse approximation threshold, the color depth is greater than the cell nucleus staining depth threshold, and the nuclear membrane heteromorphism is greater than the nuclear membrane heteromorphism threshold, then the cell corresponding to the cell nucleus is judged to be an ASCUS cell; otherwise, it is a non-hollow LSIL cell.
8. A system for mining specific image features of cervical cancer cell nuclei, used to implement the method for mining specific image features of cervical cancer cell nuclei as claimed in any one of claims 1 to 7, characterized in that: The specific image feature mining system for cervical cancer cell nuclei includes: An image acquisition module, used for acquiring cervical cell sample images; A feature extraction module, including a cell nucleus segmentation model, is used to analyze and process the cell nucleus region in the cervical cell sample image to extract specific image features; A category judgment module is used to judge the cell category based on the extracted specific image features; The result output module is used to output the judgment results of cell categories.
