Prediction model training method and device for predicting prognostic risk of cesarean section scar pregnancy based on ultrasonic image

Abstract

The invention discloses a prediction model training method and device for predicting the prognostic risk of cesarean scar pregnancy based on an ultrasonic image. The method comprises the following steps: acquiring a multi-stage risk model and a three-graph feature fusion risk prediction model; obtaining prediction image groups for training, wherein each prediction image group for training comprises a GS image, an RMT image and a CDFI image; respectively extracting fusion features of each prediction image group for training; and training the multi-stage risk model and the three-graph feature fusion risk prediction model through each fusion feature, thereby obtaining the trained multi-stage risk model and the trained three-graph feature fusion risk prediction model. According to the two trained models, the ultrasound image analysis technology is directly used for predicting disease prognosis, ultrasound doctors do not need to conduct ultrasound typing on lesions, measure indexes such as residual muscular layer thickness and grade the richness degree of lesion blood flow, and more convenience and rapidness are achieved.

Description

Technical Field

[0001] The present application relates to the technical field of image recognition, and particularly to a method for training a prediction model for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images, a device for training a prediction model for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images, and a method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images. Background Art

[0002] Re-pregnancy after cesarean section has become a relatively common phenomenon. However, due to poor healing of the uterine myometrium at the cesarean scar site and the formation of scar diverticula, when re-pregnant, the gestational sac implants in the cesarean diverticulum at the lower segment of the anterior uterine wall or is in close contact with the diverticulum, that is, cesarean scar pregnancy (CSP) occurs. As a special type of pregnancy, CSP brings new challenges to clinical decision-making.

[0003] The Chinese patent "CN117912681A A prediction model, system, and terminal device for massive hemorrhage during cesarean scar pregnancy" provides a prediction model for massive hemorrhage during cesarean scar pregnancy. It screens that the myometrial thickness at the scar and the mean diameter of the gestational sac are two independent risk factors for massive hemorrhage during cesarean scar pregnancy after cesarean section. The AUC of the Nomogram model constructed to predict massive hemorrhage during cesarean scar pregnancy is 0.87 (95% CI 0.77 - 0.97), the diagnostic sensitivity is 88.0%, and the specificity is 81.7%.

[0004] Therefore, it is desirable to have a technical solution to overcome or at least mitigate at least one of the above defects of the prior art. Summary of the Invention

[0005] The purpose of the present application is to provide a method for training a prediction model for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images to overcome or at least mitigate at least one of the above defects of the prior art.

[0006] To achieve the above purpose, the present application provides a method for training a prediction model for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images. The method for training a prediction model for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images includes:

[0007] Obtain a multi-stage risk model and a three-image feature fusion risk prediction model;

[0008] Obtain a training prediction image group. One training prediction image group includes a GS image, an RMT image, and a CDFI image;

[0009] Extract the fusion features of each training prediction image group respectively;

[0010] Train the multi - stage risk model and the three - image - feature fusion risk prediction model respectively through each of the fusion features, so as to obtain the trained multi - stage risk model and the trained three - image - feature fusion risk prediction model.

[0011] Optionally, the three - image - feature fusion risk prediction model includes an image feature extraction module. The image feature extraction module includes a ResNet - 50 module initialized with weights obtained from the natural scene ImageNet - 1k data, and the first three - layer parameter modules of ResNet - 50 are frozen.

[0012] Optionally, the image feature extraction module is used to extract the image features of the GS image, the RMT image, and the CDFI image respectively, and the parameter sharing is performed among the feature extraction branches for extracting the image features of the GS image, the RMT image, and the CDFI image.

[0013] Optionally, the loss function of the three - image - feature fusion risk prediction model includes:

[0014]

[0015] Weight i = n / (m i ×C);

[0016] where x represents the output of the model, y represents the true label, C represents the number of classes, n represents the number of samples, and m i is the total number of samples belonging to the class of the i - th class sample.

[0017] Optionally, the multi - stage risk model includes a first risk model and a second risk model. Among them, the first risk model, the second risk model, and the three - image - feature fusion risk prediction model have the same structure.

[0018] Optionally, the binary classification model designs a training method of transfer learning and data reuse.

[0019] Optionally, both the first risk model and the second risk model are binary classification models. Among them,

[0020] When the output of the first risk model is the first classification result, the fusion feature used to obtain this first classification result is input into the second classification model.

[0021] This application also provides a prediction model training device for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images. The prediction model training device for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images includes:

[0022] Multi-stage risk model acquisition module, which is used to acquire a multi-stage risk model;

[0023] Three-image feature fusion risk prediction model acquisition module, which is used to acquire a three-image feature fusion risk prediction model;

[0024] Training data acquisition module, which is used to acquire a group of training prediction images. One group of training prediction images includes one GS image, one RMT image, and one CDFI image;

[0025] Fusion feature acquisition module, which is used to extract the fusion features of each group of training prediction images respectively;

[0026] Training module, which is used to train the multi-stage risk model and the three-image feature fusion risk prediction model respectively through each of the fusion features, so as to obtain a trained multi-stage risk model and a trained three-image feature fusion risk prediction model.

[0027] The present application also provides a prediction method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images. The prediction method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images includes:

[0028] Acquire a trained multi-stage risk model and a trained three-image feature fusion risk prediction model trained by the above-mentioned prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images;

[0029] Acquire an image to be recognized;

[0030] Input the image to be recognized into the trained multi-stage risk model to obtain a first classification result;

[0031] Input the image to be recognized into the trained three-image feature fusion risk prediction model to obtain a second classification result.

[0032] The prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images in the present application directly uses ultrasound image analysis technology to predict the disease prognosis through two trained models, without the need for ultrasound doctors to perform ultrasound typing on the lesions, measure indicators such as the remaining myometrial thickness, and grade the blood flow richness of the lesions, etc. It is more convenient, faster, and more user-friendly for beginners. At the same time, our definition of clinical prognosis is also more diversified, adding sub-classifications of drug treatment and different surgical methods. Description of the Drawings

[0033] Figure 1It is a schematic flowchart of a method for training a prediction model for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images according to an embodiment of the present application.

[0034] Figure 2 It is a schematic structural diagram of an electronic device according to an embodiment of the present application.

[0035] Figure 3 It is a schematic diagram of ROI annotation of CSP ultrasonic images according to an embodiment of the present application.

[0036] Figure 4 It is a schematic structural diagram of a three - image feature fusion risk prediction model according to an embodiment of the present application.

[0037] Figure 5 It is a schematic structural diagram of a multi - stage risk model according to an embodiment of the present application. Detailed implementation manners

[0038] To make the purpose, technical solutions, and advantages of the implementation of the present application clearer, the technical solutions in the embodiments of the present application will be described in more detail below with reference to the accompanying drawings in the embodiments of the present application. In the drawings, the same or similar reference numerals denote the same or similar elements or elements with the same or similar functions from beginning to end. The described embodiments are some, but not all, of the embodiments of the present application. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present application and should not be construed as limiting the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application. The embodiments of the present application will be described in detail below with reference to the drawings.

[0039] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as limiting the protection scope of the present application.

[0040] As Figure 1 shown, the method for training a prediction model for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images includes:

[0041] Step 1: Obtain a multi - stage risk model and a three - image feature fusion risk prediction model;

[0042] Step 2: Obtain the prediction image groups for training. One prediction image group for training includes one GS image, one RMT image, and one CDFI image;

[0043] Step 3: Extract the fusion features of each of the prediction image groups for training;

[0044] Step 4: Train the multi-stage risk model and the three-image feature fusion risk prediction model respectively through each of the fusion features, so as to obtain the trained multi-stage risk model and the trained three-image feature fusion risk prediction model.

[0045] The prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images in this application directly uses ultrasonic image analysis technology to predict the disease prognosis through the two trained models, without the need for ultrasonic doctors to perform ultrasonic typing on lesions, measure indicators such as the remaining myometrial thickness, grade the blood flow richness of lesions, etc., which is more convenient, faster, and more user-friendly for beginners. At the same time, our definition of clinical prognosis is also more diversified, adding detailed classifications of drug treatment and different surgical methods.

[0046] In this embodiment, the prediction image groups for training in this application are obtained in the following manner:

[0047] This application retrospectively collected a total of 585 cases clinically diagnosed with CSP in a certain hospital in Beijing from 2006 to 2024, excluding those with incomplete clinical data, those who did not undergo preoperative ultrasonic examination, those with poor quality preoperative ultrasonic images, those who underwent transabdominal ultrasonic examination, and those who were diagnosed as not conforming to scar pregnancy through retrospective analysis. Finally, 256 CSP patients were included, their clinical data were collected, and they were divided into three groups according to the clinical treatment prognosis outcome: Group A - low-risk group (simple methotrexate drug treatment; simple uterine curettage with a smooth operation and a blood loss < 200 ml), Group B - medium-risk group (uterine artery embolization + uterine curettage; uterine curettage + methotrexate treatment or secondary uterine curettage, etc.), and Group C - high-risk group (hysterectomy; blood loss ≥ 200 ml during uterine curettage or lesion resection or massive bleeding after surgery; the original surgery was not smooth and the surgical method was changed midway, such as uterine curettage converted to laparotomy or laparoscopy, etc.). There are 100, 118, and 38 patients in Groups A, B, and C respectively.

[0048] Based on the latest research results and doctors' experience, the present invention selects the preoperative transvaginal ultrasonic images of the cases as the basis for risk prediction. Under the guidance of experts in a certain hospital in Beijing, based on the clinical diagnosis principle, a complete and predictable prediction data paradigm is designed, that is, a training sample needs to consist of a complete gestational sac cross-sectional image (GS image), a cross-sectional image of the remaining myometrial thickness (RMT image), a CDFI cross-sectional image at the scar (CDFI image), and its corresponding clinical treatment prognosis outcome group.

[0049] The image selection criteria are as follows: (1) GS images need to select the largest longitudinal section of the lesion, and RMT and CDFI images need to show the thinnest part of the lower uterine myometrium and take into account the gestational sac; (2) Try to avoid selecting images with measurement marks, arrows, letters, etc.; (3) Excessive pressure on the probe will compress the lesion area and cause morphological changes, so such images should be avoided as much as possible; (4) Do not over-enlarge the image, and the image should include part of the uterine cavity and cervix. The collected ultrasound images of each case were named "number-group-date-image category GS / RMT / CDFI". In order to facilitate the model to focus on pathological features, the patient privacy information, instrument equipment information, etc. in the original image were cropped and erased.

[0050] The images are imported into the Label Studio annotation platform. The annotation team annotates the region of interest (ROI) for each image. The review team can use the platform to perform quality control on the annotation. The ROI annotation operation standards for each section image are as follows: Figure 3 ): (1) GS section annotation method: Use a polygon to outline along the outer edge of the gestational sac, appropriately enlarge the image when outlining, and select "1GS-MAX" in the annotation label; (2) RMT section annotation method: First, define the niche, whose size and position reflect the diverticulum (imaginary) of the case. Use the niche as the central grid and define the ROI. The ROI is equivalent to a nine-square grid with the niche as the central grid. All gestational sacs or lesions, the lower myometrium of the anterior uterine wall, part of the uterine cavity and the cervical canal should be included in it, and select "2RMT" in the annotation label; (3) The CDFI section annotation method is the same as the RMT section, and select "3CDFI" in the annotation label.

[0051] Then, in order to further adapt to the input form of the deep learning model, the image is finally processed to 224x224 size through padding and resizing. In order to make full use of limited data and conduct fair evaluation, the three categories A, B, and C are divided into training sets and test sets with a ratio of 85:15, 101:17, and 33:5, respectively, for the training and verification of artificial intelligence models.

[0052] In the field of CSP, the research on prognosis risk prediction based on ultrasound images is not sufficient, no effective prediction paradigm has been formed, and no relevant prediction model has been established. Therefore, the present invention innovatively establishes a CSP prognosis risk prediction paradigm based on GS graph, RMT graph, and CDFI graph, and trains a reliable prediction model based on the collected relevant case samples.

[0053] There are usually two ways to establish traditional deep learning image classification models. One is to train the model's capabilities from scratch based on existing data, and the other is to obtain basic feature extraction capabilities from open-source, larger-scale, similar or related basic databases, and then train and learn through the data of the current scenario. The latter can often accelerate model training and improve the model's generalization performance. In the field of medical imaging, due to factors such as data security and domain differences, it is difficult to form an effective general basic database, especially for CSP data. Considering comprehensively, in order to enable the model to obtain knowledge from open-source data as much as possible and improve the model's robustness, the model trained with natural scene pictures is still used as the basis of the present invention. In addition, for the task design of multi-image joint classification, this design selects the "post-fusion" strategy. After obtaining the features of each image respectively, the features are concatenated and classified based on the fully connected layer.

[0054] In this embodiment, in order to enrich the distribution of case data and improve the robustness of the model's capabilities, the image data input module is designed as a data augmentation structure composed of random range cropping, random probability flipping, etc.

[0055] In this embodiment, the three-image feature fusion risk prediction model includes an image feature extraction module. The image feature extraction module includes a ResNet-50 module initialized with weights obtained from natural scene ImageNet-1k data, and the first three-layer parameter modules of ResNet-50 are frozen. Specifically, the image feature extraction module is based on ResNet-50 initialized with weights obtained from natural scene ImageNet-1k data as the cornerstone. In addition, in order to adapt to the CSP sample image features while avoiding the model falling into the local optimization dilemma and overfitting, the present invention selectively freezes the first three-layer parameters of ResNet-50 and only opens the last two layers for effective learning.

[0056] In this embodiment, since the GS image, RMT image, and CDFI image of a case are all ultrasonic images, the distribution differences between them are relatively small. And in order to give full play to the influence of sample features on the model and reduce the model parameters, parameter sharing is carried out between the feature extraction branches of the three images of the present invention.

[0057] In addition, in the field of CSP, due to the lack of a unified risk grading goal, the establishment of a large-scale open-source and shared case database, and the low incidence of CSP, the data collected currently shows the characteristics of few samples and sample imbalance. Specifically, there are two points: First, the total number of cases is only 256, far lower than the data scale of thousands or even tens of thousands commonly seen in the field of deep learning; Second, the ratio of the three risk categories is approximately 2.63:3.11:1. Compared with medium and low risks, the data of the C-high risk group is particularly scarce, and in actual clinical research, high-risk patients are often the focus of doctors' attention. Against this background, the present invention designs a loss balance optimization strategy.

[0058] In deep learning, for multi-classification tasks, CrossEntropyLoss is generally used as the optimization loss object to train the model, and its formula is as follows:

[0059]

[0060] Weight i = n / (m i × C);

[0061] Among them, x represents the output of the model, y represents the true label, C represents the number of classes, n represents the number of samples, and m i is the total number of samples belonging to the class of the i-th sample.

[0062] However, through preliminary experimental demonstration, when the model is trained using the data collected by the present invention, the model output shows a consistent prediction bias, manifested as almost all samples including the A-low risk group, B-medium risk group, and C-high risk group being predicted as the B-medium risk in the evaluation session. The reason is that the largest number of B-medium risk samples causes the model to fall into a local optimal trap. Subsequently, attempts are made to make the model jump out of the local optimal point by adjusting training strides and other means, but no obvious improvement is obtained. Therefore, the present invention solves the above problems by designing a weighted loss function. The actual operation is to assign a higher weight to the minority class and a lower weight to the majority class during the loss calculation process, and increase the misclassification cost of the minority class to make the model more sensitive to the minority class. Since the prediction task of the present invention is a three-classification task, the analogy weight calculation formula is as follows:

[0063] Weight i = n / (m i × C);

[0064] Among them, n is the total number of samples in the dataset, m i is the number of samples in the i-th class, C is the total number of classes, and Weight i is the weight size corresponding to the i-th class.

[0065] In this embodiment, the multi-stage risk model includes a first risk model and a second risk model. Among them, the first risk model, the second risk model and the three-graph feature fusion risk prediction model have the same structure.

[0066] In this embodiment, the present invention also proposes a multi-stage risk model. This design concept is mainly based on the following viewpoints: First, whether from clinical experience or model experiment results, among the samples of high, medium and low risks, the B-medium risk group is prone to be confused with the A-low risk group or the C-high risk group, and the decision boundary is relatively fuzzy, while the feature differences between the A-low risk group and the C-high risk group are relatively large, and the probability of prediction confusion is relatively small; Second, in clinical practice, it is particularly important to predict C-high risk patients for timely treatment. Therefore, the present invention decomposes this risk prediction task as follows: First, a one-stage model is used to divide the samples into high-risk and non-high-risk to detect the C-high risk group; then, for the non-high-risk samples detected in the first stage, a second-stage detailed prediction is performed to complete the classification of the A-low risk group and the B-medium risk group.

[0067] In the database, the scales of the A-low risk group and the B-medium risk group are quite similar and account for a large proportion. At the same time, as mentioned above, the prediction of categories A and B is prone to confusion. Therefore, an independent binary classification task learning can be performed based on the data of groups A and B, focusing on the classification of low and medium risks. The obtained model is named CSP-AB, which can be used as the model in the second stage of the multi-stage risk model of this design.

[0068] In addition, to fully exploit the case data collected by the present invention, this two-stage risk prediction model also designs a training method of transfer learning and data reuse.

[0069] Specifically, the binary classification model is constructed in the following way:

[0070] 1. Based on the two groups of data of the A-low risk group and the B-medium risk group with a large proportion and similar scales, an independent binary classification task learning is performed, focusing on the classification of low and medium risks. The obtained model is named CSP-AB, which can be used as the model in the second stage of this design.

[0071] 2. To benefit from the classification capabilities of A and B and strengthen the generalization performance of the model, based on the method of transfer learning, when training the one-stage model CSP-C that divides high-risk (group C) and non-high-risk (groups A and B), the model weights of CSP-AB are used as the initial learning weights to obtain a primary understanding of the CSP data features.

[0072] 3. Data reuse refers to solving the problems of data scarcity and imbalance in this scenario. During the actual training process of the one-stage model CSP-C, the data of the high-risk group is reused 6 times.

[0073] From the perspective of transfer learning, the model for the downstream task can acquire basic feature extraction capabilities from the pre-training of similar data. Therefore, in this design, if the one-stage model for high-risk and non-high-risk classification starts learning based on CSP-AB, even if CSP-AB lacks the learning of the C-high-risk group, it can still benefit from the classification capabilities of A and B to a certain extent, strengthen the generalization performance of the model, and obtain a primary understanding of the CSP data features. Then, the non-high-risk group composed of data from groups A and B and the C-high-risk group are used for binary classification task training, and the resulting model is named CSP-C, which is the first-stage model of this design. It should be noted that since the data ratio of the non-high-risk group (i.e., groups A and B) to the high-risk group (i.e., group C) is extremely unbalanced: 218:38, during the actual training of the one-stage model, to balance the influence of categories, the data of the high-risk group is reused 6 times.

[0074] Whether it is CSP-C or CSP-AB, except for the output part, the basic structure of the model refers to the design of the three-graph feature fusion risk prediction model of the present invention.

[0075] In this embodiment, both the first risk model and the second risk model are binary classification models, where,

[0076] When the output of the first risk model is the first classification result, the fusion feature used to obtain this first classification result is input into the second classification model.

[0077] The model flow of the two sets of solutions of the present invention is as Figure 4 、 5 shown. In the three-graph feature risk prediction model, the three ultrasound subgraphs of the case sample respectively extract independent features through the convolutional neural network branches, then perform feature splicing, and then import them into the fully connected layer for class prediction. The model uses the weight balance loss calculated based on the sample distribution as the optimization object for supervised training.

[0078] In the multi-stage risk model, the CSP-C model for predicting and detecting the C-high-risk group and the CSP-AB model for distinguishing the A-low-risk group and the B-medium-risk group are respectively trained, where the CSP-C model is obtained by transfer learning based on CSP-AB as a pre-training model; in the actual prediction link, the case picture first passes through the CSP-C model in the first stage for prediction. If the output is the high-risk group - C, it is predicted that the prognosis risk of this patient is high risk. If the output is non-high risk, the case picture needs to pass through the second-stage model CSP-AB for further prediction to complete the prediction of the A-low-risk group and the B-medium-risk group.

[0079] The advantages of this application are as follows:

[0080] 1. A comprehensive CSP prognosis risk prediction paradigm is established, and two risk prediction schemes are designed, filling the research gap in the field of predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images.

[0081] 2. Based on the existing CSP case ultrasound data collected, a fusion prediction model with clear and efficient processes is trained. By methods such as parameter sharing, feature fusion, and transfer learning, the performance of the model is simply and efficiently improved, and the application of artificial intelligence models in the clinical field of CSP is promoted.

[0082] This application also provides a prediction model training device for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images. The prediction model training device for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images includes a multi-stage risk model acquisition module, a three-image feature fusion risk prediction model acquisition module, a training data acquisition module, a fusion feature acquisition module, and a training module. Among them,

[0083] The multi-stage risk model acquisition module is used to acquire a multi-stage risk model;

[0084] The three-image feature fusion risk prediction model acquisition module is used to acquire a three-image feature fusion risk prediction model;

[0085] The training data acquisition module is used to acquire a group of training prediction images. One group of training prediction images includes one GS image, one RMT image, and one CDFI image;

[0086] The fusion feature acquisition module is used to extract the fusion features of each group of training prediction images respectively;

[0087] The training module is used to train the multi-stage risk model and the three-image feature fusion risk prediction model respectively through each of the fusion features, so as to obtain a trained multi-stage risk model and a trained three-image feature fusion risk prediction model.

[0088] This application also provides a prediction method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images. The prediction method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images includes:

[0089] Acquire a trained multi-stage risk model and a trained three-image feature fusion risk prediction model obtained through the above-mentioned prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images;

[0090] Acquire an image to be recognized;

[0091] Input the image to be recognized into the trained multi-stage risk model to obtain a first classification result;

[0092] Input the image to be recognized into the trained three-image feature fusion risk prediction model, so as to obtain a second classification result.

[0093] It should be noted that the foregoing explanations of the method embodiments also apply to the devices in this embodiment, and will not be repeated here.

[0094] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the above-mentioned prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images.

[0095] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement the above-mentioned prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images.

[0096] Figure 2 It is an exemplary structural diagram of an electronic device capable of implementing the prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images provided by an embodiment of this application.

[0097] As Figure 2 shown, the electronic device includes an input device 501, an input interface 502, a central processing unit 503, a memory 504, an output interface 505, and an output device 506. Among them, the input interface 502, the central processing unit 503, the memory 504, and the output interface 505 are connected to each other through a bus 507. The input device 501 and the output device 506 are respectively connected to the bus 507 through the input interface 502 and the output interface 505, and then connected to other components of the electronic device. Specifically, the input device 504 receives input information from the outside and transmits the input information to the central processing unit 503 through the input interface 502; the central processing unit 503 processes the input information based on the computer-executable instructions stored in the memory 504 to generate output information, temporarily or permanently stores the output information in the memory 504, and then transmits the output information to the output device 506 through the output interface 505; the output device 506 outputs the output information to the outside of the electronic device for the user to use.

[0098] That is to say, Figure 2 the electronic device shown can also be implemented as including: a memory storing computer-executable instructions; and one or more processors, and when the one or more processors execute the computer-executable instructions, they can implement the combination Figure 1 described prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images.

[0099] In one embodiment, Figure 2 The electronic device shown can be implemented to include: a memory 504 configured to store executable program code; and one or more processors 503 configured to run the executable program code stored in the memory 504 to execute the prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images in the above embodiments.

[0100] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0101] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM), and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of computer-readable media.

[0102] Computer-readable media includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory, or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD), or other optical storage, magnetic cassette tapes, magnetic disk storage, or other magnetic storage devices, or any other non-transmission media that can be used to store information accessible by a computing device.

[0103] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, system, or computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented 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.

[0104] In addition, it is obvious that the term "including" does not exclude other units or steps. The multiple units, modules, or devices stated in the apparatus claims can also be implemented by one unit or a general apparatus through software or hardware.

[0105] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart or block diagram may represent a module, a segment of a program, or a part of code, which includes one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that marked in the accompanying drawings. For example, two consecutive blocks marked may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, as well as the combinations of blocks in the block diagram and / or overall flowchart, may be implemented by a dedicated hardware-based system for performing the specified functions or operations, or may be implemented by a combination of dedicated hardware and computer instructions.

[0106] Those skilled in the art should understand that the embodiments of the present application may be provided as a method, a system, or a computer program product. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0107] In addition, it is obvious that the term "including" does not exclude other units or steps. The multiple units, modules, or devices stated in the apparatus claims may also be implemented by one unit or an overall device through software or hardware.

[0108] Although the present application is disclosed above with preferred embodiments, it is not intended to limit the present application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the protection scope of the present application should be determined by the scope defined by the claims of the present application.

[0109] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for training a prediction model for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images, characterized in that The method for training a prediction model for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images includes: Obtaining a multi-stage risk model and a three-image feature fusion risk prediction model; Obtaining a group of training prediction images, where one group of training prediction images includes one GS image, one RMT image, and one CDFI image; Respectively extracting the fusion features of each group of training prediction images; Training the multi-stage risk model and the three-image feature fusion risk prediction model respectively through each of the fusion features, so as to obtain a trained multi-stage risk model and a trained three-image feature fusion risk prediction model.

2. The prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images according to claim 1, wherein, The three-image feature fusion risk prediction model includes an image feature extraction module, and the image feature extraction module includes a ResNet-50 module initialized with weights obtained from the natural scene ImageNet-1k data, and the first three-layer parameter modules of ResNet-50 are frozen.

3. The prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images according to claim 2, wherein The image feature extraction module is used to extract the image features of the GS image, RMT image, and CDFI image respectively, and the parameter sharing is performed among the feature extraction branches for extracting the image features of the GS image, RMT image, and CDFI image.

4. The prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images according to claim 3, characterized in that, The loss function of the three-image feature fusion risk prediction model includes: Weight i = n / (m i ×C); Where x represents the output of the model, y represents the true label, C represents the number of classes, n represents the number of samples, and m i is the total number of samples belonging to the class of the samples of class i.

5. The method for training a prediction model for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images according to claim 4, wherein The multi-stage risk model includes a first risk model and a second risk model. Among them, the first risk model and the second risk model have the same structure as the three-image feature fusion risk prediction model.

6. The prediction model training method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images according to claim 5, characterized in that Both the first risk model and the second risk model are binary classification models, where When the output of the first risk model is a first classification result, the fusion feature used to obtain the first classification result is input into the second classification model.

7. A prediction model training device for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images, characterized in that, The prediction model training device for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images includes: A multi-stage risk model acquisition module, which is used to acquire a multi-stage risk model; A three-image feature fusion risk prediction model acquisition module, which is used to acquire a three-image feature fusion risk prediction model; A training data acquisition module, which is used to acquire a group of training prediction images, where one group of training prediction images includes one GS image, one RMT image, and one CDFI image; A fusion feature acquisition module, which is used to extract the fusion features of each group of training prediction images respectively; A training module, which is used to train the multi-stage risk model and the three-image feature fusion risk prediction model respectively through each of the fusion features, so as to obtain a trained multi-stage risk model and a trained three-image feature fusion risk prediction model.

8. A prediction method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasonic images, characterized in that, The prediction method for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images includes: Obtaining a trained multi-stage risk model and a trained three-image feature fusion risk prediction model trained by the method for training a prediction model for predicting the prognosis risk of cesarean scar pregnancy based on ultrasound images according to any one of claims 1 to 6; Obtaining an image to be recognized; Input the image to be recognized into the trained multi-stage risk model to obtain a first classification result; Input the image to be recognized into the trained three-image feature fusion risk prediction model to obtain a second classification result.

Images