An automatic screening method and system for cervical cell neoplasia
By dividing cervical cell tumor slice images into image blocks and using multi-model inference and aggregation of positive probabilities, the problem of high manual annotation requirements and insufficient identification accuracy in existing technologies is solved, thus achieving efficient and accurate cervical cell tumor screening.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN MODERN PATHOLOGY ENGINEERING RESEARCH INSTITUTE CO LTD
- Filing Date
- 2026-03-11
- Publication Date
- 2026-06-19
AI Technical Summary
Existing technologies for cervical cell tumor screening suffer from problems such as high demand for manual data annotation, poor algorithm flexibility, and insufficient identification accuracy.
The pathological slide images to be screened are divided into multiple image blocks. The feature vector of each image block is obtained through a feature extraction model. The positive probability of the image blocks is inferred and aggregated using multiple models. The K image blocks with the highest positive probability are selected for the final positive and negative determination.
It improves the accuracy and efficiency of cervical cell tumor screening, reduces reliance on manual annotation, and enhances the flexibility and accuracy of the algorithm.
Smart Images
Figure CN122244859A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary field of machine learning and medical imaging, and more specifically, to an automated screening method and system for cervical cell tumors. Background Technology
[0002] TCT (Liquid-based cytology) cervical cancer screening has become a widely used early screening method globally, primarily for detecting cervical cancer and its precancerous lesions, especially those caused by high-risk HPV (human papillomavirus) infection. This method involves collecting cell samples from the cervix and performing cytological examination using liquid-based technology, enabling more precise detection of abnormal cell changes. However, due to a shortage of pathologists, the collected TCT slides require extensive screening and interpretation by a large number of professionals, and various human factors can lead to false negatives (failure to detect lesions) and false positives (misdiagnosis). Therefore, developing a rapid and accurate screening method is of paramount importance.
[0003] In recent years, the rapid development of image processing and artificial intelligence (AI) technologies has provided enormous potential for TCT cervical cell carcinoma screening, especially in improving screening efficiency and accuracy. Based on digital microscopic images and trained with large-scale data, AI algorithms can effectively learn to distinguish the characteristics of negative and positive cells, automating the TCT cervical cell carcinoma screening task. For example, by marking rectangular boxes on positive cells for target detection training, the algorithm can identify and locate positive cells, thereby determining the positive or negative status of the sample. Another method involves classifying and training the algorithm by labeling instances in microscopic images with positive or negative markers. This allows for automatic identification of the presence of positive instances in the image through a traversal approach, and based on this, the positive or negative status of the entire slide is determined.
[0004] However, the above methods share a common challenge: the development phase requires extensive manual data annotation. This process not only places enormous demands on human and material resources, but annotation-based algorithms are also typically task-specific, lacking flexibility and adjustability. Furthermore, directly identifying the presence of positive instances in images using AI models is not accurate enough. Summary of the Invention
[0005] This invention addresses the technical problems existing in the prior art by providing an automated screening method and system for cervical cell tumors, which can overcome the inaccuracy of existing methods in screening for cervical cell tumors.
[0006] According to a first aspect of the present invention, an automated screening method for cervical cell tumors is provided, comprising: The foreground region of the pathological slide image to be screened is divided into multiple image blocks; The feature vector of each image block is extracted using the first model; The feature vector of each image block is input into the second model, and the positive probability of each image block is inferred through the second model. Based on the positive probability of each image block, the K image blocks with the highest positive probability in the pathological slide images to be screened are selected, where K is a positive integer. The feature vectors of K image blocks are input into the third model. The feature vectors of the K image blocks are aggregated by the third model to obtain aggregated features. Based on the aggregated features, the positive probability of the pathological slide image to be screened is output. Automatic screening for cervical cell tumors is performed based on the positive probability of the pathological slide images to be screened and the k image blocks with the highest positive probability.
[0007] According to a second aspect of the present invention, an automated cervical cell tumor screening system is provided, comprising: The segmentation module is used to divide the foreground region of the pathological slide image to be screened into multiple image blocks; The feature extraction module is used to extract the feature vector of each of the image blocks using the first model; The inference module is used to input the feature vector of each image block into the second model, infer the positive probability of each image block through the second model, and select the K image blocks with the highest positive probability in the pathological slide images to be screened based on the positive probability of each image block, where K is a positive integer. The output module is used to input the feature vectors of K image blocks into the third model, aggregate the feature vectors of the K image blocks through the third model to obtain aggregated features, and output the positive probability of the pathological slide image to be screened based on the aggregated features. The screening module is used to automatically screen for cervical cell tumors based on the positive probability of the pathological slide images to be screened and the k image blocks with the highest positive probability.
[0008] This invention provides an automated screening method and system for cervical cell tumors. The method divides the foreground region of a pathological slide image to be screened into multiple image blocks, extracts the feature vector of each image block, and inputs the feature vector of each image block into a second model. The second model infers the positive probability of each image block. Based on the positive probability of each image block, the K image blocks with the highest positive probability (K is a positive integer) in the pathological slide image to be screened are selected. The feature vectors of the K image blocks are then input into a third model, which outputs the positive probability of the pathological slide image to be screened. This invention selects the K image blocks with the highest positive probability through a second model, then focuses on these K image blocks, and infers the positive or negative status of the entire image block based on the image blocks with the highest positive probability, resulting in more accurate inference results. Attached Figure Description
[0009] Figure 1 A flowchart of an automated cervical cell tumor screening method provided in one embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an automated cervical cell tumor screening system according to an embodiment of the present invention; Figure 3 A schematic diagram of the hardware structure of a possible electronic device provided by the present invention; Figure 4 This is a schematic diagram of the hardware structure of a possible computer-readable storage medium provided by the present invention. Detailed Implementation
[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined with each other to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.
[0011] Figure 1 A flowchart of an automated cervical cell tumor screening method according to an embodiment of the present invention is shown, as follows: Figure 1 As shown, the method includes the following steps: Step 1: Divide the foreground region of the pathological slide image to be screened into multiple image blocks.
[0012] Understandably, the foreground portion of the pathological slide image (WSI) to be screened at 10x magnification is traversed using patches (image patches) (the foreground region is obtained by a dedicated TCT foreground segmentation model), resulting in multiple patches of the pathological slide image to be screened. The patch size is 512×512, and the sliding overlay is set to 128.
[0013] Step 2: Extract the feature vector of each image block using the first model.
[0014] Understandably, each patch of the pathological slide image to be screened is input into the first model, and the feature vector of each patch is extracted by the first model.
[0015] The first model is a feature extraction model, and the training process of the feature extraction model includes: TCT cervical cell data from different hospitals are collected. Image patches are traversed in the foreground region of each pathological slide image at 10x magnification to obtain all image patches as the first training set. In this embodiment of the invention, the first training set includes more than 1 million patches for training.
[0016] Based on the first training set, the feature extraction model is trained unsupervised to obtain the trained feature extraction model. Specifically, based on the first training set, the first model is trained using an unsupervised training method, wherein a 512×512×3 image block is input, which is divided into 1024 16×16×3 tokens and fed into the feature extraction model DINOV3, and finally outputs a 1024-dimensional feature vector.
[0017] The feature vector of each patch in the pathological slide image to be screened is extracted based on the trained feature extraction model.
[0018] Step 3: Input the feature vector of each image block into the second model, and infer the positive probability of each image block through the second model. Based on the positive probability of each image block, select the K image blocks with the highest positive probability in the pathological slide images to be screened, where K is a positive integer.
[0019] Understandably, the feature vector of each patch of the pathological slide image to be screened is input into the second model, which infers the positive probability of each patch. Then, the K patches with the highest positive probability are selected from all patches, and the size of K is determined according to specific needs.
[0020] In one embodiment of the present invention, the training of the second model includes: Collect multiple pathological slide images and label the positive and negative values of each pathological slide image; The foreground region of each pathological slide image is divided into multiple image blocks. Each image block shares a positive or negative label, which is the same as the positive or negative label of the pathological slide image to which it belongs. The feature vector of each image patch is extracted based on the first model; Construct a second training set, which includes feature vectors of all image patches with positive and negative labels; The second model is trained based on the second training set.
[0021] Specifically, the training process for the second model includes: First, a dataset is constructed, consisting of weakly supervised samples at the WSI level (digital slices with positive and negative labels), approximately 1000+ samples in this example. Then, each WSI is traversed through patches (512×512 patches of the foreground region at 10x magnification). For all patches under the same WSI, a common positive / negative label is used, identical to the positive / negative label of the respective WSI. The final result is a compressed package containing 1000+ patches with weakly supervised labels.
[0022] For each patch in the compressed package, the first model is used to extract the feature vector of each patch, and a second training set is constructed. The second training set includes the feature vectors of all image patches with positive and negative labels. The second model is then trained based on the second training set.
[0023] The second model is trained based on the second training set, including: a) Input the feature vector of each image patch into the second model, and infer the positive probability of each image patch through the second model.
[0024] Specifically, the feature vector of each patch is input into the second model, and the second model infers the positive probability of each patch.
[0025] The positive probability of each image patch is expressed as:
[0026] in, It is the first i The first pathological section image j Feature vectors of image patches These are the model parameters for the second model.
[0027] b. Based on the positive probability of each image patch in each pathological slide image, select the K image patches with the highest positive probability, continue to train the second model based on the K image patches with the highest positive probability in each pathological slide image, and calculate the loss value based on the first loss function.
[0028] Understandably, based on the positive probability of each image patch, for each WSI, the K image patches with the highest positive probability are selected. Then, the second model is trained again based on the K image patches with the highest positive probability in each pathological slide image, and the loss value is calculated.
[0029] The loss value is calculated using a first loss function, which is the cross-entropy loss function. The expression for the first loss function is as follows:
[0030] in, It is the loss value. It is the positive or negative label of the i-th pathological slide image. It represents the positive probability of the k-th image patch among the K image patches with the highest positive probability in the i-th pathological slide image. and It is the category weight.
[0031] c. Based on the loss value, adjust the model parameters of the second model and return to step a for iterative training until the loss function value meets the accuracy condition or the number of iterations reaches the maximum.
[0032] The process involves adjusting the model parameters of the second model based on the calculated loss value, and then returning to step a for the next round of iterative training. In each round of training, all patches of all WSIs are reordered for inference, and the K patches with the highest positive probabilities in each WSI are selected for training. This process is repeated until convergence is achieved and the second model is obtained.
[0033] Based on the trained second model, the feature vector of each image block of the pathological slide image to be screened is input into the second model, and the positive probability of each image block is inferred by the second model, and the K image blocks with the highest positive probability are selected.
[0034] Step 4: Input the feature vectors of K image blocks into the third model. The third model aggregates the feature vectors of the K image blocks to obtain aggregated features. Based on the aggregated features, output the positive probability of the pathological slide image to be screened.
[0035] Understandably, the K image blocks with the highest positive probability of the pathological slide images to be screened are input into the third model, and the third model outputs the positive probability of the pathological slide images to be screened.
[0036] The training process for the third model includes: Dataset Construction: The data for multi-instance learning primarily comes from the second model's inference on all WSIs. For each WSI, features (k, 1024) of the k patches with the highest positive probabilities are obtained through inference; in this example, k is set to 100. To ensure sample reliability, false positive and false negative samples inferred by the second model are removed. Finally, the features of the qualified samples are saved as an h5 format file (labels consistent with those used above). A single patch feature can be represented as... :
[0037] in, This represents the k-th image patch. This represents the function expression for the first model.
[0038] After feature extraction, the attention weight of each patch is calculated using a third model, and then the features of K patches are aggregated for classification prediction.
[0039] Calculate the attention weights for each patch:
[0040] in, , and Here is the learnable parameter matrix, where L is the 512-layer hidden dimension and D is the 1024-layer output dimension. , For learnable parameters, for transpose, This is the feature vector of the k-th image patch with the highest positive probability. This represents the feature vector of the j-th image patch among the K image patches with the highest positive probability. and This represents the activation function. This represents element-wise multiplication. The weight is the weight of the k-th patch.
[0041] Calculate the aggregated features based on the feature vector and attention weights of each image patch: .
[0042] Based on the aggregated features, the positive probability of the entire image is inferred, thus enabling the determination of the positive or negative status of the entire image.
[0043] The second loss function of the third model during training is expressed as follows:
[0044] in, For cross-entropy loss, For clustering constraint loss, These are the weighting coefficients;
[0045] Where N is the number of clusters, This represents the t instances with the highest attention weight in the m-th cluster.
[0046] Based on the loss value of each iteration, the model parameters of the third model are adjusted, and the training is repeated until convergence to obtain the third model.
[0047] After training to obtain the third model, the K image blocks with the highest positive probability of the pathological slide image to be screened are input into the third model. The feature vectors of the K image blocks are aggregated by the third model to obtain aggregated features. Based on the aggregated features of the pathological slide image to be screened, the positive probability of the pathological slide image to be screened is output.
[0048] Step 5: Automatically screen for cervical cell tumors based on the positive probability of the pathological slide images to be screened and the k image blocks with the highest positive probability.
[0049] Specifically, if the positive probability of the pathological slide image to be screened is greater than or equal to the preset probability value, then the pathological slide image to be screened is determined to be positive; and the coordinates of the K image blocks with the highest positive probability in the pathological slide image to be screened are recorded in order to locate the positive cells.
[0050] After receiving a large amount of cervical cell data, the system batch-compiles the positive and negative results and generates a statistical table, with positive samples listed separately for special emphasis. For positive samples, the coordinates of the k patches with the highest positive probability are automatically recorded and displayed (k is set to 10 in this example). Users can quickly analyze the condition using the k patches with the highest probability.
[0051] Figure 2 An automated cervical cell tumor screening system according to an embodiment of the present invention is shown, comprising: The segmentation module 201 is used to divide the foreground region of the pathological slide image to be screened into multiple image blocks; Feature extraction module 202 is used to extract feature vectors for each of the image blocks using the first model; The inference module 203 is used to input the feature vector of each image block into the second model, infer the positive probability of each image block through the second model, and filter out the K image blocks with the highest positive probability in the pathological slide images to be screened based on the positive probability of each image block, where K is a positive integer. Output module 204 is used to input the feature vectors of K image blocks into the third model, aggregate the feature vectors of K image blocks through the third model to obtain aggregated features, and output the positive probability of the pathological slide image to be screened based on the aggregated features. The screening module 205 is used to automatically screen for cervical cell tumors based on the positive probability of the pathological slide images to be screened and the k image blocks with the highest positive probability. Display module 206 is used to display the boundary lines of the K image blocks with the highest positive probability in the pathological slide images to be screened and the positive probability of the K image blocks on the WSI platform.
[0052] It is understood that the cervical cell tumor automatic screening system provided by the present invention corresponds to the cervical cell tumor automatic screening method provided in the foregoing embodiments. The relevant technical features of the cervical cell tumor automatic screening system can be referred to the relevant technical features of the cervical cell tumor automatic screening method, and will not be repeated here.
[0053] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating an embodiment of the electronic device provided in this invention. For example... Figure 3 As shown, an embodiment of the present invention provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program 311, it implements an automatic screening method for cervical cell tumors.
[0054] Please see Figure 4 , Figure 4 This is a schematic diagram illustrating an embodiment of a computer-readable storage medium provided by the present invention. (See diagram below.) Figure 4 As shown, this embodiment provides a computer-readable storage medium 400 on which a computer program 411 is stored. When the computer program 411 is executed by a processor, it implements an automatic screening method for cervical cell tumors.
[0055] This invention provides an automated screening method and system for cervical cell tumors. The method divides the foreground region of a pathological slide image to be screened into multiple image blocks, extracts the feature vector of each image block, and inputs the feature vector of each image block into a second model. The second model infers the positive probability of each image block. Based on the positive probability of each image block, the K image blocks with the highest positive probability (K is a positive integer) in the pathological slide image to be screened are selected. The feature vectors of the K image blocks are then input into a third model, which outputs the positive probability of the pathological slide image to be screened. This invention selects the K image blocks with the highest positive probability through a second model, then focuses on these K image blocks, and infers the positive or negative status of the entire image block based on the image blocks with the highest positive probability, resulting in more accurate inference results.
[0056] 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.
[0057] Those skilled in the art will understand that embodiments of the present invention can be provided as 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.
[0058] 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, and 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 illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0059] 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.
[0060] 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.
[0061] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0062] 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 the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. An automated screening method for cervical cell tumors, characterized in that, include: The foreground region of the pathological slide image to be screened is divided into multiple image blocks; The feature vector of each image block is extracted using the first model; The feature vector of each image block is input into the second model, and the positive probability of each image block is inferred through the second model. Based on the positive probability of each image block, the K image blocks with the highest positive probability in the pathological slide images to be screened are selected, where K is a positive integer. The feature vectors of K image blocks are input into the third model. The feature vectors of the K image blocks are aggregated by the third model to obtain aggregated features. Based on the aggregated features, the positive probability of the pathological slide image to be screened is output. Automatic screening for cervical cell tumors is performed based on the positive probability of the pathological slide images to be screened and the k image blocks with the highest positive probability.
2. The automated cervical cell tumor screening method according to claim 1, characterized in that, The first model is a feature extraction model, and the training process of the feature extraction model includes: We collected TCT cervical cell data from different hospitals, and traversed the foreground region of each pathological slide image at a set magnification to obtain all image blocks as the first training set. Based on the first training set, the feature extraction model is trained unsupervised to obtain the trained feature extraction model.
3. The automated cervical cell tumor screening method according to claim 1, characterized in that, The training of the second model includes: Collect multiple pathological slide images and label the positive and negative values of each pathological slide image; The foreground region of each pathological slide image is divided into multiple image blocks. Each image block shares a positive or negative label, which is the same as the positive or negative label of the pathological slide image to which it belongs. The feature vector of each image patch is extracted based on the first model; Construct a second training set, which includes feature vectors of all image patches with positive and negative labels; The second model is trained based on the second training set.
4. The automated cervical cell tumor screening method according to claim 3, characterized in that, Based on the second training set, the second model is trained, including: a) Input the feature vector of each image patch into the second model, and infer the positive probability of each image patch through the second model; b. Based on the positive probability of each image patch in each pathological slide image, select the K image patches with the highest positive probability, continue to train the second model based on the K image patches with the highest positive probability in each pathological slide image, and calculate the loss value based on the first loss function. c. Based on the loss value, adjust the model parameters of the second model and return to step a for iterative training until the loss value meets the accuracy condition or the number of iterations reaches the maximum.
5. The automated cervical cell tumor screening method according to claim 4, characterized in that, The positive probability of each image patch is expressed as: in, It is the first i The first pathological section image j Feature vectors of image patches These are the model parameters of the second model. This is the function expression for the second model.
6. The automated cervical cell tumor screening method according to claim 4, characterized in that, The expression for the first loss function is: in, It is the loss value. It is the positive or negative label of the i-th pathological slide image. It represents the positive probability of the k-th image patch among the K image patches with the highest positive probability in the i-th pathological slide image. and It is the category weight.
7. The automated cervical cell tumor screening method according to claim 1, characterized in that, The step of inputting the feature vectors of K image patches into a third model, and then aggregating the feature vectors of the K image patches through the third model to obtain aggregated features includes: The attention weights for each image patch are calculated using the third model: in, , and Here is the learnable parameter matrix, where L is the 512-layer hidden dimension and D is the 1024-layer output dimension. , For learnable parameters, for transpose, This is the feature vector of the k-th image patch with the highest positive probability. This represents the feature vector of the j-th image patch among the K image patches with the highest positive probability. and Indicates the activation function; in: in, This represents the k-th image patch. The function expression representing the first model; Calculate the aggregated features based on the feature vector and attention weights of each image patch: .
8. The automated cervical cell tumor screening method according to claim 1 or 7, characterized in that, The second loss function of the third model during training is expressed as follows: in, For cross-entropy loss, For clustering constraint loss, These are the weighting coefficients; Where N is the number of clusters, This represents the t image patches with the highest attention weights in the m-th cluster.
9. The automated cervical cell tumor screening method according to claim 1, characterized in that, The automatic screening for cervical cell tumors based on the positive probability of the pathological slide images to be screened and the k image blocks with the highest positive probabilities includes: If the positive probability of the pathological slide image to be screened is greater than or equal to the preset probability value, then the pathological slide image to be screened is determined to be positive. Record the coordinates of the K image blocks with the highest positive probability in the pathological slide images to be screened.
10. An automated cervical cell tumor screening system, characterized in that, include: The segmentation module is used to divide the foreground region of the pathological slide image to be screened into multiple image blocks; The extraction module is used to extract the feature vector of each image patch using the first model; The inference module is used to input the feature vector of each image block into the second model, infer the positive probability of each image block through the second model, and select the K image blocks with the highest positive probability in the pathological slide images to be screened based on the positive probability of each image block, where K is a positive integer. The output module is used to input the feature vectors of K image blocks into the third model, aggregate the feature vectors of the K image blocks through the third model to obtain aggregated features, and output the positive probability of the pathological slide image to be screened based on the aggregated features. The screening module is used to automatically screen for cervical cell tumors based on the positive probability of the pathological slide images to be screened and the k image blocks with the highest positive probability.