Method and system for constructing auxiliary prediction model of benign and malignant endometrial diseases

Abstract

The invention relates to an endometrial benign and malignant disease auxiliary prediction model construction method and system. The method comprises the following steps: acquiring first historical image information of a non-target part and second historical image information of a target part; preprocessing is carried out respectively, and a corresponding pre-training sample data set and a training sample data set are constructed respectively; constructing a convolutional neural network model, and inputting the pre-training sample data set into the convolutional neural network model for comparative learning pre-training; and inputting the training sample data set into the pre-trained convolutional neural network model for training until the training is completed, thereby obtaining the endometrial benign and malignant disease auxiliary prediction model. By utilizing a contrast learning technology, the recognition capability of a convolutional neural network model on tiny lesion features is enhanced, the generalization capability of the model is improved, the sensitivity and specificity of diagnosis are remarkably improved, the misdiagnosis rate is reduced, the recognition precision of the model is greatly improved, and misdiagnosis and unnecessary biopsy possibilities are reduced; and perception deviation and visual fatigue of an endoscope physician can be relieved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical auxiliary diagnosis, and in particular to a method and system for constructing an auxiliary prediction model for benign and malignant endometrial diseases. Background Art

[0002] Endometrial cancer (EC) is one of the most common gynecological malignancies worldwide. Early diagnosis and treatment are crucial to improving the survival rate of patients. Atypical endometrial hyperplasia (AEH) is a precancerous lesion with a high risk of progression to endometrial cancer. Traditional hysteroscopy and diagnostic curettage methods are prone to misdiagnosis when identifying atypical endometrial hyperplasia (AEH) and endometrial cancer (EC). Even experienced gynecological endoscopists may suffer from visual fatigue due to long hours of work, thereby increasing the chance of misdiagnosis or missed diagnosis, which brings great trouble to the actual clinical work of doctors. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a method and system for constructing an auxiliary prediction model for benign and malignant endometrial diseases in view of the deficiencies of the above-mentioned prior art.

[0004] The technical solution of the present invention to solve the above technical problem is as follows: a method for constructing an auxiliary prediction model for benign and malignant endometrial diseases, comprising the following steps:

[0005] Acquire first historical image information of a non-target part and second historical image information of a target part;

[0006] Preprocessing the first historical image information and the second historical image information respectively, and constructing corresponding pre-training sample data sets and training sample data sets respectively;

[0007] Constructing a convolutional neural network model, and inputting the pre-training sample data set into the convolutional neural network model for comparative learning pre-training;

[0008] The training sample data set is input into the pre-trained convolutional neural network model for training until the training is completed, thereby obtaining an auxiliary prediction model for benign and malignant endometrial diseases.

[0009] The beneficial effects of the present invention are as follows: the method for constructing an auxiliary prediction model for benign and malignant endometrial diseases of the present invention performs comparative learning pre-training on a convolutional neural network model by using a pre-training sample data set constructed using the first historical image information of the pre-processed non-target part, and utilizes the comparative learning technology to enhance the recognition ability of the convolutional neural network model for subtle lesion features and improve the generalization ability of the model, especially in cases where malignant and benign lesions are difficult to distinguish, significantly improves the sensitivity and specificity of diagnosis, reduces the misdiagnosis rate, and enhances the performance of the convolutional neural network model under conditions of limited labeled data, and can accurately distinguish between atypical endometrial hyperplasia (AEH), endometrial cancer (EC) and benign lesions. The convolutional neural network model is trained by using a training sample data set constructed using the second historical image information of the pre-processed target part, which greatly improves the recognition accuracy of the model, reduces the possibility of misdiagnosis and unnecessary biopsy, and helps alleviate the perceptual bias and visual fatigue commonly encountered by endoscopists.

[0010] On the basis of the above technical solution, the present invention can also be improved as follows:

[0011] Further: the preprocessing of the first historical image information and the second historical image information specifically includes the following steps:

[0012] The first historical image information and the second historical image information are respectively subjected to standardization processing and normalization processing, wherein the standardization processing includes scaling processing, translation processing, rotation processing and flipping processing, so as to obtain images with uniform format and size.

[0013] The beneficial effect of the above further scheme is: by standardizing and normalizing the first historical image information and the second historical image information, overfitting can be prevented, and the image can be normalized and adjusted to uniform pixels, which facilitates the convolutional neural network model to perform recognition and extraction processing, thereby improving the performance of the model.

[0014] Further: the step of inputting the pre-training sample data set into the convolutional neural network model for comparative learning pre-training specifically includes the following steps:

[0015] The convolutional neural network model performs feature extraction processing on the first historical image information in the pre-training sample data set, and maps the extracted pre-training features into a low-dimensional space;

[0016] In the low-dimensional space, a contrast loss function is constructed, and the cosine similarity between the feature vectors corresponding to the extracted pre-trained features is calculated according to the contrast loss function. The calculation formula of the contrast loss function is:

[0017]

[0018] Among them, zi and z j represents the positive eigenvector, sim(z i ,z j ) represents the opposite eigenvector z i and z j The cosine similarity between them, τ is the temperature parameter, l i,j Denotes the opposite eigenvector z i and z j The loss function between , N is the number of samples;

[0019] The distribution concentration of the loss function is adjusted based on the cosine similarity until the change value of the contrast loss function is within a preset range, thereby completing the pre-training.

[0020] The beneficial effect of the above further scheme is: by constructing a contrast loss function, the cosine similarity between feature vectors can be accurately calculated, and the cross entropy is adjusted and scaled using the temperature parameter so that the change value of the contrast loss function tends to be stable, thereby greatly improving the feature extraction ability of the convolutional neural network model, so that the convolutional neural network model can distinguish subtle differences in medical images.

[0021] Further: the step of inputting the training sample data set into the pre-trained convolutional neural network model for training until the training is completed to obtain an auxiliary prediction model for benign and malignant endometrial diseases specifically includes the following steps:

[0022] The pre-trained convolutional neural network model performs classification and recognition processing on the second historical image information in the training sample data set to obtain case classification information;

[0023] A focal loss function is constructed based on the case classification information and the corresponding true label value. The specific formula is:

[0024] focal loss(p t )=(1-p t ) 0·1 CE(p,y) (2)

[0025] CE(p,y)= -log (p t ) (3)

[0026]

[0027] Among them, p is the probability that the sample is correctly predicted and classified by the convolutional neural network model, y is the true label value of the sample, and p t is the predicted label value of the model, CE(p,y) is the cross entropy loss value between the true label value and the predicted label value, focal loss(p t) is the focus loss value;

[0028] The model parameter with the smallest loss value of the focal loss function is used as the target parameter of the convolutional neural network model, and the training is completed to obtain an auxiliary prediction model for benign and malignant endometrial diseases.

[0029] The beneficial effect of the above further scheme is: the second historical image information in the training sample data set is classified and identified through the convolutional neural network model to identify the case classification information, and then a focus loss function is constructed based on the case classification information identified by the convolutional neural network model and the corresponding true label value, and the focus loss value between the two is calculated. By introducing an adjustment factor on the basis of the cross-entropy loss function, the sensitivity of the auxiliary prediction model for benign and malignant endometrial diseases to endometrial images is greatly increased, thereby ensuring the recognition accuracy of the auxiliary prediction model for benign and malignant endometrial diseases.

[0030] Further: the method for constructing an auxiliary prediction model for benign and malignant endometrial diseases also includes the following steps:

[0031] The auxiliary prediction model for benign and malignant endometrial diseases obtained after training is evaluated to obtain an evaluation result. The specific calculation formula is:

[0032]

[0033] Among them, TP, TN, FP, and FN represent true positive examples, true negative examples, false positive examples, and false negative examples, respectively, PPV represents positive predictive value, and NPV represents negative predictive value.

[0034] The beneficial effect of the above further scheme is that by calculating the accuracy, sensitivity, specificity, positive predictive value, negative predictive value and F1 score of the convolutional neural network model, the classification efficiency and classification accuracy of the trained convolutional neural network model can be characterized, so as to facilitate the calculation of the prediction score of each lesion according to the prediction results.

[0035] The present invention also provides a system for constructing an auxiliary prediction model for benign and malignant endometrial diseases, including an acquisition module, a preprocessing module, and a model training and recognition module;

[0036] The acquisition module is used to acquire the first historical image information of the non-target part and the second historical image information of the target part;

[0037] The preprocessing module is used to preprocess the first historical image information and the second historical image information respectively, and construct corresponding pre-training sample data sets and training sample data sets respectively;

[0038] The model training recognition module is used to construct a convolutional neural network model and input the pre-training sample data set into the convolutional neural network model for comparative learning pre-training;

[0039] The model training identification module is also used to input the training sample data set into the pre-trained convolutional neural network model for training until the training is completed to obtain an auxiliary prediction model for benign and malignant endometrial diseases.

[0040] On the basis of the above technical solution, the present invention can also be improved as follows:

[0041] Further: the model training recognition module inputs the pre-training sample data set into the convolutional neural network model for comparative learning pre-training, and the specific implementation is:

[0042] The convolutional neural network model performs feature extraction processing on the first historical image information in the pre-training sample data set, and maps the extracted pre-training features into a low-dimensional space;

[0043] In the low-dimensional space, a contrast loss function is constructed, and the cosine similarity between the feature vectors corresponding to the extracted pre-trained features is calculated according to the contrast loss function. The calculation formula of the contrast loss function is:

[0044]

[0045] Among them, z i and z j represents the positive eigenvector, sim(z i ,z j ) represents the opposite eigenvector z i and z j The cosine similarity between them, τ is the temperature parameter, l i,j Denotes the opposite eigenvector z i and z j The loss function between , N is the number of samples;

[0046] The distribution concentration of the loss function is adjusted based on the cosine similarity until the change value of the contrast loss function is within a preset range, thereby completing the pre-training.

[0047] The beneficial effect of the above further scheme is: by constructing a contrast loss function, the cosine similarity between feature vectors can be accurately calculated, and the cross entropy is adjusted and scaled using the temperature parameter so that the change value of the contrast loss function tends to be stable, thereby greatly improving the feature extraction ability of the convolutional neural network model, so that the convolutional neural network model can distinguish subtle differences in medical images.

[0048] Further: the model training recognition module inputs the training sample data set into the pre-trained convolutional neural network model for training until the training is completed, and the specific implementation of the auxiliary prediction model for benign and malignant endometrial diseases is obtained as follows:

[0049] The pre-trained convolutional neural network model performs classification and recognition processing on the second historical image information in the training sample data set to obtain case classification information;

[0050] A focal loss function is constructed based on the case classification information and the corresponding true label value. The specific formula is:

[0051] focal loss(p t )=(1-p t ) 0·1 CE(p,y) (2)

[0052] CE(p,y)= -log (p t ) (3)

[0053]

[0054] Among them, p is the probability that the sample is correctly predicted and classified by the convolutional neural network model, y is the true label value of the sample, and p t is the predicted label value of the model, CE(p,y) is the cross entropy loss value between the true label value and the predicted label value, focal loss(p t ) is the focus loss value;

[0055] The model parameter with the smallest loss value of the focal loss function is used as the target parameter of the convolutional neural network model, and the training is completed to obtain an auxiliary prediction model for benign and malignant endometrial diseases.

[0056] The beneficial effect of the above further scheme is: the second historical image information in the training sample data set is classified and identified through the convolutional neural network model to identify the case classification information, and then a focus loss function is constructed based on the case classification information identified by the convolutional neural network model and the corresponding true label value, and the focus loss value between the two is calculated. By introducing an adjustment factor on the basis of the cross-entropy loss function, the sensitivity of the auxiliary prediction model for benign and malignant endometrial diseases to endometrial images is greatly increased, thereby ensuring the recognition accuracy of the auxiliary prediction model for benign and malignant endometrial diseases.

[0057] The present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the method for constructing an auxiliary prediction model for benign and malignant endometrial diseases is implemented.

[0058] The present invention also provides an auxiliary prediction model construction device for benign and malignant endometrial diseases, characterized in that it includes a communication interface, a memory, a communication bus and a processor, wherein the processor, the communication interface and the memory communicate with each other through the communication bus;

[0059] The memory is used to store computer programs;

[0060] The processor is used to implement the steps of the method for constructing an auxiliary prediction model for benign and malignant endometrial diseases when executing the program stored in the memory. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic diagram of a flow chart of a method for constructing an auxiliary prediction model for benign and malignant endometrial diseases according to an embodiment of the present invention;

[0062] Figure 2 It is a schematic diagram showing the comparison of the accuracy of the auxiliary prediction model for benign and malignant endometrial diseases and experts with different experience levels according to an embodiment of the present invention;

[0063] Figure 3 A schematic diagram showing the visualization of the prediction of atypical endometrial hyperplasia and endometrial cancer by the auxiliary prediction model for benign and malignant endometrial diseases according to an embodiment of the present invention;

[0064] Figure 4 The schematic diagram is a structural diagram of a system for constructing an auxiliary prediction model for benign and malignant endometrial diseases according to an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0066] like Figure 1 As shown, a method for constructing an auxiliary prediction model for benign and malignant endometrial diseases comprises the following steps:

[0067] S1: Acquire first historical image information of a non-target part and second historical image information of a target part;

[0068] S2: Preprocessing the first historical image information and the second historical image information respectively, and constructing corresponding pre-training sample data sets and training sample data sets respectively;

[0069] S3: constructing a convolutional neural network model, and inputting the pre-training sample data set into the convolutional neural network model for comparative learning pre-training;

[0070] S4: Inputting the training sample data set into the pre-trained convolutional neural network model for training until the convolutional neural network model meets preset conditions and the training is completed to obtain an auxiliary prediction model for benign and malignant endometrial diseases.

[0071] The method for constructing an auxiliary prediction model for benign and malignant endometrial diseases of the present invention performs comparative learning pre-training on a convolutional neural network model by using a pre-training sample data set constructed using first historical image information of pre-processed non-target parts, and utilizes comparative learning technology to enhance the recognition ability of the convolutional neural network model for subtle lesion features and improve the generalization ability of the model, especially in cases where malignant and benign lesions are difficult to distinguish, thereby significantly improving the sensitivity and specificity of diagnosis, reducing the misdiagnosis rate, and enhancing the performance of the convolutional neural network model under conditions of limited labeled data, and being able to accurately distinguish between atypical endometrial hyperplasia (AEH), endometrial carcinoma (EC) and benign lesions. The convolutional neural network model is trained by using a training sample data set constructed using second historical image information of pre-processed target parts, thereby greatly improving the recognition accuracy of the model, reducing the possibility of misdiagnosis and unnecessary biopsy, and helping to alleviate the perceptual bias and visual fatigue commonly encountered by endoscopists.

[0072] In one or more embodiments of the present invention, in order to enhance the learning ability of the convolutional neural network model, contrastive learning pre-training is adopted. Specifically, the first historical image information of the non-target part is obtained and pre-processed to obtain a pre-training sample data set. Here, the pre-training sample data set includes four data sets: CP-CHILD17, PolypGen18–20, IPCL21 and Hyper Kvasir22. The CP-CHILD data set contains pediatric colonoscopy images, which are divided into CP-CHILD-A (8,000 images) and CP-CHILD-B (1,500 images); the PolypGen data set includes 8,037 frames of images from six hospitals for polyp segmentation and detection; the IPCL21 data set contains ME-NBI video frames from 114 patients, mainly focusing on magnified endoscopy sequences; Hyper Kvasir22 is the largest public dataset of large intestinal images, containing 110,079 images and 373 videos. These diverse data sets ensure powerful and discriminative feature learning, thereby improving the performance of the convolutional neural network model.

[0073] The training sample data set of the present invention was collected in three tertiary hospitals, including 1394 cases in total, and the total number of images was 55874 PNG format hysteroscopic images. The number of cases and images in the training set and test set is shown in Table 1. The images were taken by one of three high-resolution devices (Olympus OTV-S190, Japan; Karl Storz 26105FA or 26120BA, Germany). All images were confirmed by two experts, WW.W. and WM. The control group categories (benign lesions) included endometrial polyps, submucosal uterine leiomyoma, atypical endometrial hyperplasia, and normal uterine cavity.

[0074] Table 1. Baseline characteristics of the training and test datasets

[0075]

[0076] 1. AEH / EC: atypical endometrial hyperplasia / endometrial carcinoma

[0077] The training dataset was collected from the Hubei Maternal and Child Hospital (MCH) between January 2008 and December 2017 using Olympus OTV-S190 (Japan) and Karl Storz 26105FA or 26120BA (Germany). The internal test dataset consisted of hysteroscopic images collected at MCH using the same equipment between January 2018 and June 2019. The external test dataset consisted of data collected from Tongji Hospital of Huazhong University of Science and Technology (TJH) and the Second Affiliated Hospital of Zhengzhou University (ZZSH) between January 2019 and December 2019. The AEH / EC category includes cases of endometrial atypical hyperplasia and endometrial cancer. The external test dataset was mainly collected using the Olympus OTV-S190 (Japan) device, and there was no case overlap between the training and test sets.

[0078] In the embodiments of the present invention, two additional test datasets are also included: the first contains 3419 images from 23 AEH / EC cases and 62 control cases diagnosed at MCH between January 2018 and June 2019, and the second contains 2809 images from 16 AEH / EC cases and 89 control cases diagnosed at TJH / ZZSH between January 2019 and December 2019. The former is used as an internal test dataset, and the latter is used as an external test dataset.

[0079] The case classification information corresponding to the training sample data set includes atypical endometrial hyperplasia (AEH), endometrial cancer (EC) and benign lesions.

[0080] In one or more embodiments of the present invention, the preprocessing of the first historical image information and the second historical image information specifically includes the following steps:

[0081] S21: performing standardization and normalization processing on the first historical image information and the second historical image information respectively, wherein the standardization processing includes scaling processing, translation processing, rotation processing and flipping processing, so as to obtain images with uniform format and size.

[0082] By standardizing and normalizing the first historical image information and the second historical image information, overfitting can be prevented, and the images are normalized and adjusted to uniform pixels, which facilitates the convolutional neural network model to perform recognition and extraction processing, thereby improving the performance of the model. In an embodiment of the present invention, all sample data are adjusted to 224×224 pixels so as to be input into the convolutional neural network for recognition and analysis.

[0083] In addition, in the embodiment of the present invention, an "oversampling" technique is used to compensate for the impact of data imbalance in the training data set.

[0084] In one or more embodiments of the present invention, the step of inputting the pre-training sample data set into the convolutional neural network model for contrastive learning pre-training specifically includes the following steps:

[0085] S31: the convolutional neural network model performs feature extraction processing on the first historical image information in the pre-training sample data set, and maps the extracted pre-training features into a low-dimensional space;

[0086] Here, the present invention adopts ResNet-50 as the backbone network. ResNet-50 is a CNN with a depth of 50 layers, which is famous for its residual connection, which helps to alleviate the gradient vanishing problem and supports the training of deeper networks. In the present invention, a projection head is added to the ResNet-50 network. The projection head consists of three fully connected layers, which is designed to map the features extracted by ResNet-50 to a lower dimensional space, in which contrastive learning is applied for pre-training.

[0087] S32: In the low-dimensional space, a contrast loss function is constructed, and the cosine similarity between the feature vectors corresponding to the extracted pre-trained features is calculated according to the contrast loss function. The calculation formula of the contrast loss function is:

[0088]

[0089] Among them, z i and z j represents the positive eigenvector, sim(z i ,z j) represents the opposite eigenvector z i and z j The cosine similarity between them, τ is the temperature parameter, l i,j Denotes the opposite eigenvector z i and z j The loss function between , N is the number of samples;

[0090] The contrastive loss function scales the cross entropy by a temperature parameter and uses temperature scaling to adjust the sharpness of the resulting distribution. This contrastive learning pre-training method significantly improves the feature extraction capability of the convolutional neural network model, enabling the convolutional neural network model to distinguish subtle differences in medical images, such as the difference between atypical endometrial hyperplasia (AEH) and endometrial cancer (EC). The application of contrastive learning pre-training is crucial to improving the overall performance of the convolutional neural network model.

[0091] S33: Based on the cosine similarity, the distribution concentration of the loss function is adjusted until the change value of the contrast loss function is within a preset range, and the training is completed to obtain an auxiliary prediction model for benign and malignant endometrial diseases.

[0092] By constructing a contrast loss function, the cosine similarity between feature vectors can be accurately calculated, and the cross entropy is adjusted and scaled using the temperature parameter so that the change value of the contrast loss function tends to be stable, thereby greatly improving the feature extraction ability of the convolutional neural network model, and enabling the auxiliary prediction model for benign and malignant endometrial diseases to distinguish subtle differences in medical images.

[0093] In the pre-training stage, the embodiments of the present invention adopt contrastive training learning to enable the convolutional neural network model to extract more meaningful and reliable features. Contrastive training learning is a self-supervised learning method that improves feature learning by optimizing the differences of the same data points, enhancing the similarity between positive images (the image itself and its pre-processed images), and minimizing the differences of different data points, reducing the similarity between negative images (images and other pre-processed images). This technology shows significant potential in medical imaging tasks, especially in improving diagnostic accuracy and model robustness.

[0094] In one or more embodiments of the present invention, the step of inputting the training sample data set into the pre-trained convolutional neural network model for training until the training is completed to obtain an auxiliary prediction model for benign and malignant endometrial diseases specifically includes the following steps:

[0095] S41: the pre-trained convolutional neural network model performs classification and recognition processing on the second historical image information in the training sample data set to obtain case classification information. Here, a binary classification recognition algorithm is adopted, and the classification information obtained is atypical endometrial hyperplasia (AEH) or endometrial cancer (EC), and benign lesions;

[0096] S42: Constructing a focal loss function based on the case classification information and the corresponding true label value, the specific formula is:

[0097] focal loss(p t )=(1-p t ) 0·1 CE(p,y) (2)

[0098] CE(p,y)= -log (p t ) (3)

[0099]

[0100] Wherein, p is the probability that the sample is correctly predicted and classified by the convolutional neural network model, y is the true label value of the sample, and in this embodiment, when the sample is atypical endometrial hyperplasia (AEH) and endometrial cancer (EC), y=1, and in other cases y=0, p t is the predicted label value of the model, CE(p,y) is the cross entropy loss value between the true label value and the predicted label value, focal loss(p t ) is the focus loss value;

[0101] In order to overcome the data imbalance problem caused by the small number of malignant cases (such as AEH / EC), the convolutional neural network model proposed in the present invention adopts the focal loss method, and increases the sensitivity to misclassified AEH / EC cases by introducing an adjustment factor in the cross entropy loss function (Formula (2) (3)).

[0102] S43: The model parameter with the smallest loss value of the focal loss function is used as the target parameter of the convolutional neural network model to complete the training and obtain an auxiliary prediction model for benign and malignant endometrial diseases.

[0103] In practice, the training method of the present invention combines multiple key deep learning libraries, such as PyTorch, torchvision, timem, etc. In order to prevent overfitting and accelerate network training, the present invention "freezes" the first 10 layers in the convolutional neural network model by setting the learning rate to zero. In the present invention, a stochastic gradient descent (SGD) optimizer is used, the learning rate is set to 0.01, the momentum is 0.9, the learning rate is decayed every 10 epochs (the decay rate is 0.1), and the number of training rounds is 30.

[0104] The second historical image information in the training sample data set is classified and recognized by a convolutional neural network model to identify case classification information, and then a focal loss function is constructed based on the case classification information recognized by the convolutional neural network model and the corresponding true label value, and the focal loss value between the two is calculated. By introducing an adjustment factor on the basis of the cross-entropy loss function, the sensitivity of the auxiliary prediction model for benign and malignant endometrial diseases to endometrial images is greatly increased, thereby ensuring the recognition accuracy of the auxiliary prediction model for benign and malignant endometrial diseases.

[0105] In an embodiment of the present invention, after the training is completed, the target endometrial image information can be input into the endometrial benign and malignant disease auxiliary prediction model obtained after the training to obtain the corresponding endometrial benign and malignant disease type.

[0106] In one or more embodiments of the present invention, the method for constructing an auxiliary prediction model for benign and malignant endometrial diseases further comprises the following steps:

[0107] S5: Evaluate the auxiliary prediction model for benign and malignant endometrial diseases obtained after training to obtain an evaluation result. The specific calculation formula is:

[0108]

[0109]

[0110] Among them, TP, TN, FP, and FN represent true positive examples, true negative examples, false positive examples, and false negative examples, respectively, PPV represents positive predictive value, and NPV represents negative predictive value.

[0111] By calculating the accuracy, sensitivity, specificity, positive predictive value, negative predictive value and F1 score of the auxiliary prediction model for benign and malignant endometrial diseases, the classification efficiency and classification accuracy of the auxiliary prediction model for benign and malignant endometrial diseases obtained after training can be characterized, making it convenient to calculate the prediction score of each lesion based on the prediction results.

[0112] In order to verify the recognition accuracy of the auxiliary prediction model for benign and malignant endometrial diseases obtained after contrastive learning pre-training in the present invention, in this embodiment, the recognition results of the auxiliary prediction model for benign and malignant endometrial diseases without contrastive learning pre-training and the auxiliary prediction model for benign and malignant endometrial diseases after contrastive learning pre-training were tested, and the results are as follows:

[0113] Test sample dataset collected for Hubei Maternal and Child Hospital (MCH):

[0114] For the auxiliary prediction model for benign and malignant endometrial diseases without contrast learning (CL), the AUC (area under the ROC curve and the coordinate axis) value was 0.969 (95% CI: 0.928-0.999), the accuracy was 91.8% (95% CI: 85.9-97.6%), the sensitivity was 96.8% (95% CI: 91.7-100%), the specificity was 78.3% (95% CI: 60.9-92.9%), and the F1 score was 0.945 (95% CI: 0.902-0.979). In contrast, the performance of the auxiliary prediction model for benign and malignant endometrial diseases using contrastive learning (CL) was slightly improved, with an AUC value of 0.979 (95% CI: 0.942-1.000), an accuracy of 94.1% (95% CI: 89.1-99.1%), a sensitivity of 95.2% (95% CI: 89.5-100%), a specificity of 91.3% (95% CI: 78.2-100%), and an F1 score of 0.959 (95% CI: 0.920-0.992).

[0115] Test sample dataset collected for Tongji Hospital of Huazhong University of Science and Technology (TJH) / The Second Affiliated Hospital of Zhengzhou University (ZZSH):

[0116] For the auxiliary prediction model for benign and malignant endometrial diseases without contrast learning (CL), the AUC (area under the ROC curve and the coordinate axis) value was 0.891 (95% CI: 0.810-0.964), the accuracy was 89.5% (95% CI: 83.7-95.4%), the sensitivity was 94.4% (95% CI: 89.4-98.8%), the specificity was 62.5% (95% CI: 40.0-86.7%), and the F1 score was 0.939 (95% CI: 0.898-0.973). In contrast, the auxiliary prediction model for benign and malignant endometrial diseases using contrastive learning (CL) significantly improved the performance, with an AUC value of 0.975 (95% CI: 0.942-0.998), an accuracy of 93.3% (95% CI: 88.6-98.1%), a sensitivity of 92.1% (95% CI: 86.4-96.8%), a specificity of 100% (95% CI: 100-100%), and an F1 score of 0.959 (95% CI: 0.925-0.988).

[0117] The present invention compares the recognition results of the pre-trained and trained endometrial benign and malignant disease auxiliary prediction model with the judgment results of the endoscopist, and the results are as follows:

[0118] The test sample dataset collected for Hubei Maternal and Child Hospital (MCH):

[0119] When comparing the average performance of junior, intermediate and senior endoscopists (4 people in each group) with the auxiliary prediction model for endometrial benign and malignant diseases, the deep learning auxiliary prediction model for endometrial benign and malignant diseases consistently outperformed manual experts. When contrastive learning (CL) was used, the AUC value (0.979 vs. 0.952), accuracy (94.1% vs. 87.0%), sensitivity (95.2% vs. 92.3%) and F1 score (0.959 vs. 0.770) of the auxiliary prediction model for endometrial benign and malignant diseases were all better than the average results of senior endoscopists. These results highlight the advantages of the auxiliary prediction model for endometrial benign and malignant diseases in diagnostic accuracy and consistency, especially when contrastive learning (CL) was used.

[0120] Test sample dataset collected for Tongji Hospital of Huazhong University of Science and Technology (TJH) / The Second Affiliated Hospital of Zhengzhou University (ZZSH):

[0121] Compared with the average performance of experts from Tongji Hospital (TJH) of Huazhong University of Science and Technology (4 people in each group), the auxiliary prediction model for benign and malignant endometrial diseases, especially when using contrastive learning (CL), consistently showed excellent diagnostic accuracy. After integrating contrastive learning (CL), the AUC value of the auxiliary prediction model for benign and malignant endometrial diseases was 0.975, significantly higher than the expert average of 0.862. Similarly, the auxiliary prediction model for benign and malignant endometrial diseases surpassed the experts in accuracy (93.3% vs. 80.2%) and sensitivity (92.1% vs. 71.9%). In addition, the F1 score (0.959) of the auxiliary prediction model for benign and malignant endometrial diseases far exceeded that of the experts (0.530), which highlights the consistency and precision of the auxiliary prediction model for benign and malignant endometrial diseases in diagnostic performance.

[0122] The comparison results of the auxiliary prediction model for benign and malignant endometrial diseases and endoscopists in other evaluation indicators such as sensitivity, specificity, negative predictive value and Kappa coefficient are also listed in Tables 2 and 3.

[0123] Table 2. Comparison of diagnostic performance of endoscopists and the auxiliary prediction model for endometrial benign and malignant diseases for each patient in the MCH test dataset

[0124]

[0125] Table 3. Comparison of diagnostic performance of endoscopists and the auxiliary prediction model for endometrial benign and malignant diseases for each patient in the TJH / ZZSH test dataset

[0126]

[0127] For details on the comparison of the accuracy of the auxiliary prediction model for benign and malignant endometrial diseases and experts with different experience levels, see Figure 2 ,in, Figure 2 (a) is a schematic diagram of the accuracy of the test results of the auxiliary prediction model for benign and malignant endometrial diseases of the present invention and experts with different experience levels based on the test sample data set collected by Hubei Maternity and Child Hospital (MCH), Figure 2 (b) is a schematic diagram of the test results accuracy of the auxiliary prediction model for benign and malignant endometrial diseases of the present invention and experts with different experience levels based on the test sample data sets collected by Tongji Hospital of Huazhong University of Science and Technology (TJH) / The Second Affiliated Hospital of Zhengzhou University (ZZSH).

[0128] It can be seen that the auxiliary prediction model for benign and malignant endometrial diseases, especially when combined with contrastive learning (CL), can surpass endoscopists of different experience levels on the test data sets of different medical centers and accurately identify patients with AEH / EC. The auxiliary prediction model for benign and malignant endometrial diseases performs superiorly in sensitivity and specificity, especially in distinguishing AEH / EC from other benign lesions (such as polyps, uterine fibroids, endometrial hyperplasia without atypical hyperplasia, and normal uterine cavity). By integrating the auxiliary prediction model for benign and malignant endometrial diseases with the hysteroscopic system, the diagnostic process can be accelerated to ensure balanced performance regardless of the experience of the endoscopist. In addition, as a reliable auxiliary tool, the auxiliary prediction model for benign and malignant endometrial diseases reduces the possibility of misdiagnosis and unnecessary biopsy, helping to alleviate the perceptual bias and visual fatigue commonly encountered by endoscopists.

[0129] In an embodiment of the present invention, the effectiveness of contrastive learning comes from its ability to enhance feature extraction, which helps the model to better distinguish samples, especially when the labeled data is limited and unbalanced. The present invention uses colonoscopy datasets, which are similar to hysteroscopy datasets, for pre-training convolutional neural network models. Therefore, the auxiliary prediction model for benign and malignant endometrial diseases can more effectively capture the subtle differences between AEH / EC and benign lesions. This has led to a significant improvement in diagnostic accuracy and robustness, and the effect is still good even on external datasets from different medical centers.

[0130] The present invention uses the Grad-CAM algorithm to identify the key areas of the endometrial benign and malignant disease auxiliary prediction model for predicting AEH / EC, such as Figure 3 As shown in the figure, the first column is a hysteroscopic image, the second column is a visualized heat map of the prediction of atypical endometrial hyperplasia (AEH) and endometrial cancer (EC) by the auxiliary prediction model for benign and malignant endometrial diseases without contrast learning (CL), and the third column is a visualized heat map of the prediction of atypical endometrial hyperplasia (AEH) and endometrial cancer (EC) by the auxiliary prediction model for benign and malignant endometrial diseases after contrast learning (CL). The heat map in this figure highlights the areas in the hysteroscopic image that may contain important morphological and vascular features. These features, such as obvious deformation of the endometrial cavity, local necrosis, friable consistency, and atypical blood vessels, are associated with atypical endometrial hyperplasia (AEH) and endometrial cancer (EC) with different pathological patterns. Compared with the model without contrast learning, the model combined with contrast learning focuses on these key areas more effectively. This shows that contrast learning significantly improves the ability of the auxiliary prediction model for benign and malignant endometrial diseases in identifying and distinguishing AEH / EC, and enhances its sensitivity to these key pathological features.

[0131] like Figure 4As shown, the present invention also provides a system for constructing an auxiliary prediction model for benign and malignant endometrial diseases, including an acquisition module, a preprocessing module, and a model training and recognition module;

[0132] The acquisition module is used to acquire the first historical image information of the non-target part and the second historical image information of the target part;

[0133] The preprocessing module is used to preprocess the first historical image information and the second historical image information respectively, and construct corresponding pre-training sample data sets and training sample data sets respectively;

[0134] The model training recognition module is used to construct a convolutional neural network model and input the pre-training sample data set into the convolutional neural network model for comparative learning pre-training;

[0135] The model training recognition module is also used to input the training sample data set into the pre-trained convolutional neural network model for training until the convolutional neural network model meets preset conditions and completes the training.

[0136] In one or more embodiments of the present invention, the model training recognition module inputs the pre-training sample data set into the convolutional neural network model for contrastive learning pre-training in the following specific implementations:

[0137] The convolutional neural network model performs feature extraction processing on the first historical image information in the pre-training sample data set, and maps the extracted pre-training features into a low-dimensional space;

[0138] In the low-dimensional space, a contrast loss function is constructed, and the cosine similarity between the feature vectors corresponding to the extracted pre-trained features is calculated according to the contrast loss function. The calculation formula of the contrast loss function is:

[0139]

[0140] Among them, z i and z j represents the positive eigenvector, sim(z i ,z j ) represents the opposite eigenvector z i and z j The cosine similarity between them, τ is the temperature parameter, l i,j represents the opposite eigenvector z i and z j The loss function between , N is the number of samples;

[0141] The distribution concentration of the loss function is adjusted based on the cosine similarity until the change value of the contrast loss function is within a preset range, thereby completing the pre-training.

[0142] By constructing a contrast loss function, the cosine similarity between feature vectors can be accurately calculated, and the cross entropy is adjusted and scaled using the temperature parameter so that the change value of the contrast loss function tends to be stable, thereby greatly improving the feature extraction ability of the convolutional neural network model, so that the convolutional neural network model can distinguish subtle differences in medical images.

[0143] In one or more embodiments of the present invention, the model training recognition module inputs the training sample data set into the pre-trained convolutional neural network model for training in a specific implementation as follows:

[0144] The pre-trained convolutional neural network model performs classification and recognition processing on the second historical image information in the training sample data set to obtain case classification information;

[0145] A focal loss function is constructed based on the case classification information and the corresponding true label value. The specific formula is:

[0146] focal loss(p t )=(1-p t ) 0·1 CE(p,y) (2)

[0147] CE(p,y)= -log (p t ) (3)

[0148]

[0149] Among them, p is the probability that the sample is correctly predicted and classified by the convolutional neural network model, y is the true label value of the sample, and p t is the predicted label value of the model, CE(p,y) is the cross entropy loss value between the true label value and the predicted label value, focal loss(p t ) is the focus loss value;

[0150] The model parameter with the smallest loss value of the focal loss function is used as the target parameter of the convolutional neural network model to complete the training and form an auxiliary prediction model for benign and malignant endometrial diseases.

[0151] The second historical image information in the training sample data set is classified and recognized by a convolutional neural network model to identify case classification information, and then a focus loss function is constructed based on the case classification information recognized by the convolutional neural network model and the corresponding true label value, and the focus loss value between the two is calculated. By introducing an adjustment factor on the basis of the cross entropy loss function, the sensitivity of the convolutional neural network model to endometrial images is greatly increased, thereby ensuring the recognition accuracy of the convolutional neural network model.

[0152] The present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the method for constructing an auxiliary prediction model for benign and malignant endometrial diseases is implemented.

[0153] The present invention also provides an auxiliary prediction model construction device for benign and malignant endometrial diseases, characterized in that it includes a communication interface, a memory, a communication bus and a processor, wherein the processor, the communication interface and the memory communicate with each other through the communication bus;

[0154] The memory is used to store computer programs;

[0155] The processor is used to implement the steps of the method for constructing an auxiliary prediction model for benign and malignant endometrial diseases when executing the program stored in the memory.

[0156] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing an auxiliary prediction model for benign and malignant endometrial diseases, characterized in that: The steps include: Acquire first historical image information of a non-target part and second historical image information of a target part; Preprocessing the first historical image information and the second historical image information respectively, and constructing corresponding pre-training sample data sets and training sample data sets respectively; Constructing a convolutional neural network model, and inputting the pre-training sample data set into the convolutional neural network model for comparative learning pre-training; The training sample data set is input into the pre-trained convolutional neural network model for training until the training is completed, thereby obtaining an auxiliary prediction model for benign and malignant endometrial diseases.

2. The method for constructing an auxiliary prediction model for benign and malignant endometrial diseases according to claim 1, characterized in that The preprocessing of the first historical image information and the second historical image information specifically includes the following steps: The first historical image information and the second historical image information are respectively subjected to standardization processing and normalization processing, wherein the standardization processing includes scaling processing, translation processing, rotation processing and flipping processing, so as to obtain images with uniform format and size.

3. The method for constructing an auxiliary prediction model for benign and malignant endometrial diseases according to claim 1, characterized in that The step of inputting the pre-training sample data set into the convolutional neural network model for contrastive learning pre-training specifically includes the following steps: The convolutional neural network model performs feature extraction processing on the first historical image information in the pre-training sample data set, and maps the extracted pre-training features into a low-dimensional space; In the low-dimensional space, a contrast loss function is constructed, and the cosine similarity between the feature vectors corresponding to the extracted pre-trained features is calculated according to the contrast loss function. The calculation formula of the contrast loss function is: Among them, z i and z j represents the positive eigenvector, sim(z i ,z j ) represents the opposite eigenvector z i and z j The cosine similarity between them, τ is the temperature parameter, l i,j Denotes the opposite eigenvector z i and z j The loss function between , N is the number of samples; The distribution concentration of the loss function is adjusted based on the cosine similarity until the change value of the contrast loss function is within a preset range, and the training is completed to obtain an auxiliary prediction model for benign and malignant endometrial diseases.

4. The method for constructing an auxiliary prediction model for benign and malignant endometrial diseases according to claim 1, characterized in that The step of inputting the training sample data set into the pre-trained convolutional neural network model for training until the training is completed to obtain an auxiliary prediction model for benign and malignant endometrial diseases specifically includes the following steps: The pre-trained convolutional neural network model performs classification and recognition processing on the second historical image information in the training sample data set to obtain case classification information; A focal loss function is constructed based on the case classification information and the corresponding true label value. The specific formula is: focal loss(p t )=(1-p t ) 0·1 CE(p,y) (2) CE(p,y)= -log (p t ) (3) Among them, p is the probability that the sample is correctly predicted and classified by the convolutional neural network model, y is the true label value of the sample, and p t is the predicted label value of the model, CE(p,y) is the cross entropy loss value between the true label value and the predicted label value, focal loss(p t ) is the focus loss value; The model parameter with the smallest loss value of the focal loss function is used as the target parameter of the convolutional neural network model to complete the training and form an auxiliary prediction model for benign and malignant endometrial diseases.

5. The method for constructing an auxiliary prediction model for benign and malignant endometrial diseases according to any one of claims 1 to 4, characterized in that: The following steps are also included: The auxiliary prediction model for benign and malignant endometrial diseases obtained after training is evaluated to obtain an evaluation result. The specific calculation formula is: Among them, TP, TN, FP, and FN represent true positive examples, true negative examples, false positive examples, and false negative examples, respectively, PPV represents positive predictive value, and NPV represents negative predictive value.

6. A system for constructing an auxiliary prediction model for benign and malignant endometrial diseases based on deep learning, characterized in that: It includes acquisition module, preprocessing module and model training and recognition module; The acquisition module is used to acquire the first historical image information of the non-target part and the second historical image information of the target part; The preprocessing module is used to preprocess the first historical image information and the second historical image information respectively, and construct corresponding pre-training sample data sets and training sample data sets respectively; The model training recognition module is used to construct a convolutional neural network model and input the pre-training sample data set into the convolutional neural network model for comparative learning pre-training; The model training identification module is also used to input the training sample data set into the pre-trained convolutional neural network model for training until the training is completed to obtain an auxiliary prediction model for benign and malignant endometrial diseases.

7. The system for constructing an auxiliary prediction model for benign and malignant endometrial diseases according to claim 6, characterized in that : The specific implementation of the model training recognition module inputting the pre-training sample data set into the convolutional neural network model for comparative learning pre-training is: The convolutional neural network model performs feature extraction processing on the first historical image information in the pre-training sample data set, and maps the extracted pre-training features into a low-dimensional space; In the low-dimensional space, a contrast loss function is constructed, and the cosine similarity between the feature vectors corresponding to the extracted pre-trained features is calculated according to the contrast loss function. The calculation formula of the contrast loss function is: Among them, z i and z j represents the positive eigenvector, sim(z i ,z j ) represents the opposite eigenvector z i and z j The cosine similarity between them, τ is the temperature parameter, l i,j Denotes the opposite eigenvector z i and z j The loss function between , N is the number of samples; The distribution concentration of the loss function is adjusted based on the cosine similarity until the change value of the contrast loss function is within a preset range, thereby completing the pre-training.

8. The system for constructing an auxiliary prediction model for benign and malignant endometrial diseases according to claim 6, characterized in that : The model training recognition module inputs the training sample data set into the pre-trained convolutional neural network model for training until the training is completed, and the specific implementation of the auxiliary prediction model for benign and malignant endometrial diseases is obtained as follows: The pre-trained convolutional neural network model performs classification and recognition processing on the second historical image information in the training sample data set to obtain case classification information; A focal loss function is constructed based on the case classification information and the corresponding true label value. The specific formula is: focal loss(p t )=(1-p t ) 0·1 CE(p,y) (2) CE(p,y)= -log (p t ) (3) Among them, p is the probability that the sample is correctly predicted and classified by the convolutional neural network model, y is the true label value of the sample, and p t is the predicted label value of the model, CE(p,y) is the cross entropy loss value between the true label value and the predicted label value, focal loss(p t ) is the focus loss value; The model parameter with the smallest loss value of the focal loss function is used as the target parameter of the convolutional neural network model, and the training is completed to obtain an auxiliary prediction model for benign and malignant endometrial diseases.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for constructing an auxiliary prediction model for benign and malignant endometrial diseases according to any one of claims 1 to 5 is implemented.

10. A device for constructing an auxiliary prediction model for benign and malignant endometrial diseases, characterized in that: It includes a communication interface, a memory, a communication bus and a processor, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; The memory is used to store computer programs; The processor is used to implement the steps of the method for constructing an auxiliary prediction model for benign and malignant endometrial diseases as described in any one of claims 1 to 5 when executing the program stored in the memory.

Images