A deep learning-based early screening system for cervical cancer

Abstract

The application provides a cervical cancer early screening system based on deep learning, fuses a probability density feature vector and deep features to obtain a classification result, adds positioning position information on this basis, multiplies the obtained positioning weight value with the classification result to obtain a final screening result, and can accurately output a prediction result and effectively reduce the occurrence of false positives.

Description

Technical Field

[0001] This invention belongs to the field of image processing, and in particular relates to a deep learning-based early cervical cancer screening system. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Cervical cancer is one of the most common gynecological malignancies. There are many causes of cervical cancer, such as viral infections, number of childbirths, smoking, malnutrition, and poor hygiene. The number of new cervical cancer cases is high each year, and the disease can occur in adult women of all ages, thus early diagnosis and treatment are crucial. TCT (ThinPrep Cytologic Test) is an effective means of prevention and early diagnosis, helping to detect lesions early. However, due to the large number of cells tested, a severe shortage of pathologists, and the fact that tens of thousands of cells on a TCT smear require individual identification by pathologists, the screening efficiency is relatively low.

[0004] Therefore, how to quickly and accurately test TCT smears to assist doctors in screening for cervical cancer is a problem that needs to be solved. Summary of the Invention

[0005] To overcome the shortcomings of the existing technology, this invention provides a deep learning-based early cervical cancer screening system. It fuses the probability density feature vector and depth features of local images to obtain a classification result. Based on this, it adds location information and multiplies the obtained location weights with the classification result to obtain the final screening result. It can accurately and quickly output prediction results and effectively reduce the occurrence of false positives.

[0006] To achieve the above objectives, a first aspect of the present invention provides a deep learning-based early cervical cancer screening system, comprising:

[0007] Acquisition module: Acquires and segments the TCT digital slices to be processed, obtaining multiple local images;

[0008] Processing module: Inputs local images into the trained screening model to obtain screening results;

[0009] The filtering model includes a first branch, a second branch, and a third branch. The first branch is used to calculate the probability density histogram of a local image and use the calculated probability density histogram as the probability density feature vector of the local image. The second branch is used to extract feature vectors from the local image. The third branch is used to locate the local image and obtain the corresponding weights. The probability density feature vector is concatenated with the feature vector, and the result of the concatenation operation is input into a fully connected layer to obtain a classification result. The classification result is multiplied by the weights to obtain the final filtering result.

[0010] A third aspect of the present invention provides a computer device, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, the following steps are performed:

[0011] The TCT digital slice to be processed is obtained and segmented to obtain multiple local images;

[0012] The local image is input into the trained screening model to obtain the screening results;

[0013] The filtering model includes a first branch, a second branch, and a third branch. The first branch is used to calculate the probability density histogram of a local image and use the calculated probability density histogram as the probability density feature vector of the local image. The second branch is used to extract feature vectors from the local image. The third branch is used to locate the local image and obtain the corresponding weights. The probability density feature vector is concatenated with the feature vector, and the result of the concatenation operation is input into a fully connected layer to obtain a classification result. The classification result is multiplied by the weights to obtain the final filtering result.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the following steps:

[0015] The TCT digital slice to be processed is obtained and segmented to obtain multiple local images;

[0016] The local image is input into the trained screening model to obtain the screening results;

[0017] The filtering model includes a first branch, a second branch, and a third branch. The first branch is used to calculate the probability density histogram of a local image and use the calculated probability density histogram as the probability density feature vector of the local image. The second branch is used to extract feature vectors from the local image. The third branch is used to locate the local image and obtain the corresponding weights. The probability density feature vector is concatenated with the feature vector, and the result of the concatenation operation is input into a fully connected layer to obtain a classification result. The classification result is multiplied by the weights to obtain the final filtering result.

[0018] The above one or more technical solutions have the following beneficial effects:

[0019] In this invention, the probability density feature vector and depth features of a local image are fused to obtain a classification result. Based on this, location information is added, and the obtained location weights are multiplied with the classification result to obtain the final screening result. This method can accurately output the prediction result and effectively reduce the occurrence of false positives.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a flowchart of a deep learning-based early cervical cancer screening process according to Embodiment 1 of the present invention.

[0023] Figure 2 This is a schematic diagram of the position markings in Embodiment 1 of the present invention;

[0024] Figure 3 This is a schematic diagram of the screening model structure in Embodiment 1 of the present invention;

[0025] Figure 4 This is a flowchart of the LBP branch in Embodiment 1 of the present invention. Detailed Implementation

[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0028] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0029] Example 1

[0030] This embodiment discloses a deep learning-based early cervical cancer screening system, including:

[0031] Acquisition module: Acquires and segments the TCT digital slices to be processed, obtaining multiple local images;

[0032] Processing module: Inputs local images into the trained screening model to obtain screening results;

[0033] The filtering model includes a first branch, a second branch, and a third branch. The first branch is used to calculate the probability density histogram of a local image and use the calculated probability density histogram as the probability density feature vector of the local image. The second branch is used to extract feature vectors from the local image. The third branch is used to locate the local image and obtain the corresponding weights. The probability density feature vectors are concatenated with the feature vectors, and the result of the concatenation operation is input into a fully connected layer to obtain a classification result. The obtained classification result is multiplied by the weights to obtain the final filtering result.

[0034] The following combination Figure 1 This embodiment provides a detailed description of a deep learning-based early cervical cancer screening system, specifically including:

[0035] Sample construction: Obtain labeled TCT digital slices (Whole Slide Image, WSI), labeled with the location and type of positive cells in the slice.

[0036] Specifically, (1) Data cropping: The TCT digital slice is cropped into multiple local images according to a preset magnification and size. (2) Generation of classification labels: The cropped local images are labeled according to whether they contain positive cells. Local images containing positive cells are labeled as positive samples, and local images without positive cells are labeled as negative samples. (3) Generation of location labels: such as Figure 2 As shown, the positive sample image is binarized, with regions containing positive cells marked as 1 and the rest marked as 0, while all negative samples are marked as 0. This labeling is then scaled to the size required for locating the branch.

[0037] Model building: Constructing a screening model, which consists of three branches, such as... Figure 3As shown.

[0038] like Figure 4 As shown, specifically, the first branch, the LBP branch, converts a local image of size 512×512×3 to grayscale, performs a closing operation on the grayscale image, and then obtains its Local Binary Patterns (LBP). Subsequently, it calculates the probability density histogram of the LBP image and uses it as the probability density feature vector corresponding to the local image. The probability density feature vector is then concatenated with the feature vector extracted by the feature module. The concatenated result is then passed through a fully connected layer to output the classification result.

[0039] The second branch, the feature extraction module, consists of multiple convolutional layers connected in sequence. It extracts features from a local image of size 512×512×3, generating a feature map of size 32×32×64. Then, it performs global average pooling on the generated feature map to obtain a feature vector of size 1×1×64.

[0040] The feature map of size 32×32×64 extracted by the feature extraction module is input into the third branch, namely the localization branch. In the localization branch, the feature map is first processed by a 1×1 convolutional layer for channel integration, and then processed by the Sigmond activation function to obtain a feature map of size 32×32×1. Subsequently, global average pooling is performed on the feature map with 1 channel to obtain a weight. The obtained weight is multiplied by the classification result to obtain the final output.

[0041] Model training: The localization branch output uses the location labels generated in the data processing module as the desired output, and the localization loss L... loc And classification loss L cls Both models use binary classification with cross-entropy, and the overall loss L of the model is... total The weighted sum of the localization loss and classification loss is calculated using the following formula:

[0042]

[0043] L total =αL loc +βL cls

[0044] in, The model predicts the probability that a sample is a positive example. y is the sample label, which is 1 if the sample is a positive example and 0 otherwise. α and β are weight parameters.

[0045] Model testing: Input the cropped local image into the model to obtain the output positive image probability P. Set the positive image confidence threshold T. If P>T, the image is judged as a positive image; otherwise, it is a negative image.

[0046] Manual cervical cancer screening is inefficient and fails to meet the requirements of early diagnosis and treatment. Using deep learning-based cervical cancer screening can significantly reduce labor costs and efficiently complete the diagnostic process. Current cervical cancer screening algorithms suffer from a high false positive rate. This invention integrates LBP features and deep features, while also adding location information of positive cells, enabling accurate prediction results and effectively reducing false positives.

[0047] Example 2

[0048] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it performs the following steps:

[0049] The TCT digital slice to be processed is obtained and segmented to obtain multiple local images;

[0050] The local image is input into the trained screening model to obtain the screening results;

[0051] The filtering model includes a first branch, a second branch, and a third branch. The first branch is used to calculate the probability density histogram of a local image and use the calculated probability density histogram as the probability density feature vector of the local image. The second branch is used to extract feature vectors from the local image. The third branch is used to locate the local image and obtain the corresponding weights. The probability density feature vectors are concatenated with the feature vectors, and the result of the concatenation operation is input into a fully connected layer to obtain a classification result. The obtained classification result is multiplied by the weights to obtain the final filtering result.

[0052] Example 4

[0053] The purpose of this embodiment is to provide a computer-readable storage medium.

[0054] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0055] The TCT digital slice to be processed is obtained and segmented to obtain multiple local images;

[0056] The local image is input into the trained screening model to obtain the screening results;

[0057] The filtering model includes a first branch, a second branch, and a third branch. The first branch is used to calculate the probability density histogram of a local image and use the calculated probability density histogram as the probability density feature vector of the local image. The second branch is used to extract feature vectors from the local image. The third branch is used to locate the local image and obtain the corresponding weights. The probability density feature vectors are concatenated with the feature vectors, and the result of the concatenation operation is input into a fully connected layer to obtain a classification result. The obtained classification result is multiplied by the weights to obtain the final filtering result.

[0058] The steps and methods involved in the devices in Embodiments 2 and 3 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0059] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0060] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A deep learning-based early cervical cancer screening system, characterized in that, include: Acquisition module: Acquires and segments the TCT digital slices to be processed, obtaining multiple local images; Processing module: Inputs local images into the trained screening model to obtain screening results; The filtering model includes a first branch, a second branch, and a third branch. The first branch calculates the probability density histogram of a local image and uses the calculated probability density histogram as the probability density feature vector of the local image. The second branch extracts feature vectors from the local image. The third branch locates the local image and obtains the corresponding weights. The probability density feature vector is concatenated with the feature vector, and the result of the concatenation operation is input into a fully connected layer to obtain a classification result. The classification result is multiplied by the weights to obtain the final filtering result. In the processing module, the specific operation of the third branch on the local image is as follows: The feature map obtained by the second branch from the local image feature extraction is then channel-integrated through a 1×1 convolution. The feature map after channel integration is sequentially processed by the Sigmond activation function and global average pooling to obtain the weights corresponding to the local image. The construction of training samples for the screening model is as follows: the TCT digital slices are cropped to obtain multiple local images. The local images are labeled according to whether they contain positive cells. Local images containing positive cells are labeled as positive samples, and local images without positive cells are labeled as negative samples. The positive samples are binarized, with regions containing positive cells labeled as 1 and other regions labeled as 0. All negative samples are labeled as 0. The screening model is trained as follows: the third branch uses the position labels generated in the training samples as the expected output and uses the total loss function to train the screening model.

2. The cervical cancer early screening system based on deep learning as described in claim 1, characterized in that, In the processing module, the specific operation of the first branch on the local image is as follows: Perform grayscale conversion on a local image, and then perform a closing operation on the grayscale-converted image. Calculate the probability density histogram of the image after the closing operation.

3. The cervical cancer early screening system based on deep learning as described in claim 1, characterized in that, In the processing module, the specific operation of the second branch on the local image is as follows: Convolutional layers are used to extract features from local images to obtain feature maps; Perform global average pooling on the obtained feature map to obtain the feature vector.

4. The cervical cancer early screening system based on deep learning as described in claim 1, characterized in that, In the processing module, the screening model is tested by inputting the cropped local image into the screening model to obtain the probability of a positive image. If the probability of a positive image is greater than the set confidence threshold for a positive image, then the input local image is a positive image.

5. The cervical cancer early screening system based on deep learning as described in claim 1, characterized in that, The total loss function is a weighted sum of the localization loss and the classification loss, both of which use binary cross-entropy.

6. A computer device, characterized in that, include: The computer device includes a processor, memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the following steps are performed: The TCT digital slice to be processed is obtained and segmented to obtain multiple local images; The local image is input into the trained screening model to obtain the screening results; The filtering model includes a first branch, a second branch, and a third branch. The first branch calculates the probability density histogram of a local image and uses the calculated probability density histogram as the probability density feature vector of the local image. The second branch extracts feature vectors from the local image. The third branch locates the local image and obtains the corresponding weights. The probability density feature vector is concatenated with the feature vector, and the result of the concatenation operation is input into a fully connected layer to obtain a classification result. The classification result is multiplied by the weights to obtain the final filtering result. The specific operations of the third branch on the local image are as follows: The feature map obtained by the second branch from the local image feature extraction is then channel-integrated through a 1×1 convolution. The feature map after channel integration is sequentially processed by the Sigmond activation function and global average pooling to obtain the weights corresponding to the local image. The construction of training samples for the screening model is as follows: the TCT digital slices are cropped to obtain multiple local images. The local images are labeled according to whether they contain positive cells. Local images containing positive cells are labeled as positive samples, and local images without positive cells are labeled as negative samples. The positive samples are binarized, with regions containing positive cells labeled as 1 and other regions labeled as 0. All negative samples are labeled as 0. The screening model is trained as follows: the third branch uses the position labels generated in the training samples as the expected output and uses the total loss function to train the screening model.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, performs the following steps: The TCT digital slice to be processed is obtained and segmented to obtain multiple local images; The local image is input into the trained screening model to obtain the screening results; The filtering model includes a first branch, a second branch, and a third branch. The first branch calculates the probability density histogram of a local image and uses the calculated probability density histogram as the probability density feature vector of the local image. The second branch extracts feature vectors from the local image. The third branch locates the local image and obtains the corresponding weights. The probability density feature vector is concatenated with the feature vector, and the result of the concatenation operation is input into a fully connected layer to obtain a classification result. The classification result is multiplied by the weights to obtain the final filtering result. The specific operations of the third branch on the local image are as follows: The feature map obtained by the second branch from the local image feature extraction is then channel-integrated through a 1×1 convolution. The feature map after channel integration is sequentially processed by the Sigmond activation function and global average pooling to obtain the weights corresponding to the local image. The construction of training samples for the screening model is as follows: the TCT digital slices are cropped to obtain multiple local images. The local images are labeled according to whether they contain positive cells. Local images containing positive cells are labeled as positive samples, and local images without positive cells are labeled as negative samples. The positive samples are binarized, with regions containing positive cells labeled as 1 and other regions labeled as 0. All negative samples are labeled as 0. The screening model is trained as follows: the third branch uses the position labels generated in the training samples as the expected output and uses the total loss function to train the screening model.

Images