Processing method and device of cerebral hemorrhage segmentation prediction model based on self-supervised learning

Abstract

The present application provides a processing method and device for a cerebral hemorrhage segmentation prediction model based on self-supervised learning, which is used to train the cerebral hemorrhage segmentation prediction model and provides a novel model pre-training architecture. The pre-trained model obtained based on the processing can help to carry out more efficient and high-precision model training in the future, and can promote the acquisition of a cerebral hemorrhage segmentation prediction model with higher processing accuracy, so that the model has stronger processing performance and better application value.

Description

Technical Field

[0001] The present application relates to the field of medical technology, and in particular to a processing method and device for a cerebral hemorrhage segmentation prediction model based on self-supervised learning. Background Art

[0002] For cerebral hemorrhage, computed tomography (CT) and magnetic resonance imaging (MRI) are the two main diagnostic tools in actual situations. Among them, computed tomography is often used for initial diagnosis in the emergency room due to its fast imaging speed and high sensitivity to acute hemorrhage, while magnetic resonance imaging can provide more detailed soft tissue contrast, which is helpful for subsequent evaluation and monitoring.

[0003] In traditional work, the interpretation of the above two types of images and the segmentation prediction of cerebral hemorrhage areas mainly rely on the experience and manual operation of radiologists, which is obviously time-consuming and prone to subjective errors. In this traditional context, the introduction of deep learning into image interpretation and segmentation prediction of cerebral hemorrhage areas can not only improve the accuracy and efficiency of diagnosis and assist in treatment decisions, but also support the development of clinical research and personalized medicine, reduce the workload of doctors, and has important clinical significance and application prospects.

[0004] However, the inventors of the present application have discovered that the prior art cerebral hemorrhage segmentation prediction method based on deep learning has certain defects in terms of model training efficiency and model segmentation prediction accuracy, which is caused by the model training scheme adopted. Summary of the invention

[0005] The present application provides a processing method and device for a cerebral hemorrhage segmentation prediction model based on self-supervised learning, which is used to train the cerebral hemorrhage segmentation prediction model and provides a novel model pre-training architecture. The pre-trained model obtained based on the processing can help to carry out more efficient and high-precision model training in the future, and can promote the acquisition of a cerebral hemorrhage segmentation prediction model with higher processing accuracy, so that the model has stronger processing performance and better application value.

[0006] In a first aspect, the present application provides a processing method for a cerebral hemorrhage segmentation prediction model based on self-supervised learning, the method comprising:

[0007] Acquire sample images for pre-training, wherein the sample images include at least one of computer tomography images and magnetic resonance imaging images, and the sample images are images without annotations;

[0008] In the first stage, m consecutive sample images are input into the pre-trained model, and the pre-trained model is trained to output m consecutive sample images so that the input and output are consistent;

[0009] In the second stage, the image obtained by randomly erasing the image content of m consecutive sample images is input into the pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent;

[0010] In the third stage, m consecutive sample images are divided into three equal parts: upper, middle and lower. The middle part is randomly discarded and only the middle part is randomly retained. The images are input into the pre-trained model. The pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent.

[0011] In the fourth stage, m consecutive sample images including the cerebral hemorrhage image area to a preset degree, the images obtained by randomly erasing the image content, dividing them into three equal parts of upper, middle and lower, and then randomly discarding the middle part, and dividing them into three equal parts of upper, middle and lower, and then randomly retaining only the middle part, are input into the pre-trained model, and the pre-trained model is trained to output m consecutive sample images so that the input and output are consistent, and the model pre-training is completed, wherein the pre-trained model that has completed the model pre-training is used to train the cerebral hemorrhage segmentation prediction model based on the labeled new sample images, and the cerebral hemorrhage segmentation prediction model is used to perform segmentation prediction processing of the cerebral hemorrhage area on the image input to the model.

[0012] In a second aspect, the present application provides a processing device for a cerebral hemorrhage segmentation prediction model based on self-supervised learning, the device comprising:

[0013] An acquisition unit, used to acquire sample images for pre-training, wherein the sample images include at least one of computer tomography images and magnetic resonance imaging images, and the sample images are images without annotations;

[0014] A first training unit is used to input m consecutive sample images into a pre-training model in a first stage, and train the pre-training model to output m consecutive sample images so that the input and output are consistent;

[0015] The second training unit is used for, in the second stage, inputting the images obtained by randomly erasing the image contents of m consecutive sample images into the pre-training model, and training the pre-training model to output the m consecutive sample images before processing, so that the input and output are consistent;

[0016] A third training unit is used for, in the third stage, dividing m continuous sample images into three equal parts, namely, upper, middle and lower parts, randomly discarding the middle part and randomly retaining only the middle part, inputting the obtained images into the pre-trained model, and training the pre-trained model to output the m continuous sample images before processing, so that the input and output are consistent;

[0017] The fourth training unit is used to input the m consecutive sample images, which include the cerebral hemorrhage image area to a preset degree, into a pre-trained model in the fourth stage, randomly erase the image content, divide the image into three equal parts of upper, middle and lower, and then randomly discard the middle part, and divide the image into three equal parts of upper, middle and lower, and then randomly retain only the middle part, train the pre-trained model to output m consecutive sample images, so that the input and output are consistent, and complete the model pre-training, wherein the pre-trained model that has completed the model pre-training is used to train the cerebral hemorrhage segmentation prediction model based on the labeled new sample images, and the cerebral hemorrhage segmentation prediction model is used to perform segmentation prediction processing of the cerebral hemorrhage area on the image input by the model.

[0018] In a third aspect, the present application provides a processing device, including a processor and a memory, wherein a computer program is stored in the memory, and when the processor calls the computer program in the memory, the method provided in the first aspect of the present application or any possible implementation method of the first aspect of the present application is executed.

[0019] In a fourth aspect, the present application provides a computer-readable storage medium, which stores multiple instructions, and the instructions are suitable for a processor to load to execute the method provided by the first aspect of the present application or any possible implementation of the first aspect of the present application.

[0020] From the above content, it can be concluded that the present application has the following beneficial effects:

[0021] Corresponding to the high-quality training goal of the cerebral hemorrhage segmentation prediction model, the present application provides a novel model pre-training architecture for the training of the cerebral hemorrhage segmentation prediction model. The pre-trained model obtained based on the processing can help to carry out more efficient and high-precision model training in the future, and can promote the acquisition of a cerebral hemorrhage segmentation prediction model with higher processing accuracy, so that the model has stronger processing performance and better application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 A flowchart of a processing method of a cerebral hemorrhage segmentation prediction model based on self-supervised learning in this application;

[0024] Figure 2 A schematic diagram of a processing architecture of a processing method of a cerebral hemorrhage segmentation prediction model based on self-supervised learning in this application;

[0025] Figure 3 A structural schematic diagram of a processing device for a cerebral hemorrhage segmentation prediction model based on self-supervised learning in this application;

[0026] Figure 4 A schematic diagram of the structure of the processing equipment of this application. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0028] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or modules is not necessarily limited to those steps or modules clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices. The naming or numbering of steps in this application does not mean that the steps in the method flow must be executed in the time / logical sequence indicated by the naming or numbering. The process steps that have been named or numbered can change the execution order according to the technical purpose to be achieved, as long as the same or similar technical effects can be achieved.

[0029] The division of modules in this application is a logical division. There may be other division methods when it is implemented in actual applications. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection between modules can be electrical or other similar forms, which are not limited in this application. In addition, the modules or submodules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed in multiple circuit modules, and some or all of the modules may be selected according to actual needs to achieve the purpose of the present application.

[0030] Before introducing the processing method of the cerebral hemorrhage segmentation prediction model based on self-supervised learning provided by the present application, the background content involved in the present application is first introduced.

[0031] The processing method, device and computer-readable storage medium of the cerebral hemorrhage segmentation prediction model based on self-supervised learning provided in the present application can be applied to processing equipment to train the cerebral hemorrhage segmentation prediction model, and provide a novel model pre-training architecture. The pre-trained model obtained based on the processing can help to carry out more efficient and high-precision model training in the future, and can promote the acquisition of a cerebral hemorrhage segmentation prediction model with higher processing accuracy, so that the model has stronger processing performance and better application value.

[0032] The processing method of the cerebral hemorrhage segmentation prediction model based on self-supervised learning mentioned in this application, its execution subject can be a processing device of the cerebral hemorrhage segmentation prediction model based on self-supervised learning, or a server, physical host or user equipment (UE) and other different types of processing devices integrated with the processing device of the cerebral hemorrhage segmentation prediction model based on self-supervised learning. Among them, the processing device of the cerebral hemorrhage segmentation prediction model based on self-supervised learning can be implemented in hardware or software, and the UE can be a terminal device such as a smart phone, tablet computer, laptop computer, desktop computer or personal digital assistant (PDA), and the processing device can be set in the form of a device cluster.

[0033] It can be understood that the processing device that executes the processing method of the cerebral hemorrhage segmentation prediction model of the present application, or the processing device equipped with the application service corresponding to the processing method of the cerebral hemorrhage segmentation prediction model of the present application, considers that the focus of the present application scheme is on the pre-training of the model, and therefore only needs to have the corresponding data processing capabilities.

[0034] If the processing device also needs to involve model training based on the labeled training samples, the same applies to the above.

[0035] If the processing equipment needs to continue to involve the actual application of the model, it needs to be combined with clinical applications. It can be deployed on equipment that is specifically responsible for the diagnosis of cerebral hemorrhage, or it can be deployed on equipment that is responsible for other tasks (such as other clinical diagnoses or even routine office work) in addition to the diagnosis of cerebral hemorrhage.

[0036] Of course, the processing equipment can also be divided into the first part responsible for pre-model training, the second part responsible for model training and the third part responsible for model application according to the pre-training, training and application stages of the model. For example, the background server performs pre-training and model training, and then sends the model to the front-end (doctor side) through remote services for specific application. This is also possible.

[0037] Next, we will introduce the processing method of the cerebral hemorrhage segmentation prediction model based on self-supervised learning provided by this application.

[0038] First, see Figure 1 , Figure 1 A flowchart of a processing method of a brain hemorrhage segmentation prediction model based on self-supervised learning is shown in the present application, and then refer to Figure 2 A processing architecture diagram of a processing method of a brain hemorrhage segmentation prediction model based on self-supervised learning in the present application is shown. The processing method of a brain hemorrhage segmentation prediction model based on self-supervised learning provided by the present application may specifically include the following steps S101 to S105:

[0039] Step S101, obtaining sample images for pre-training, wherein the sample images include at least one of computer tomography images and magnetic resonance imaging images, and the sample images are images without annotations;

[0040] It can be understood that model pre-training is similar to model training in that both are based on corresponding training samples. The difference is that the training samples involved in the model pre-training architecture configured in this application, that is, the sample images obtained here, do not have the annotations involved in conventional model training.

[0041] The labeling of training samples can be understood as predicting the true value of the model configured by manual or automatic labeling tools. In the process of model training, the model is guided to optimize the model by providing theoretical prediction results to achieve effective training goals.

[0042] The model pre-training architecture involved in this application does not require labeling of training samples. The purpose is to use self-supervised learning (a training method that mines effective information from the data itself) during the pre-training process to allow the model to mine useful information related to the cerebral hemorrhage image area from the training samples themselves. In this way, in the subsequent model training phase after the model pre-training phase, fine-tuning training based on the labeled training samples can more effectively improve the prediction accuracy of the model.

[0043] It can be understood that the acquisition and processing of the sample image here is usually the acquisition and processing of the ready-made image, which can be either manually input image, or image received from other devices, or local image read, all of which are possible.

[0044] Of course, it is not excluded that in actual situations, the sample image may be collected in real time when the present application solution is applied, and it can be adjusted according to the flexible and changeable application requirements in actual situations.

[0045] Furthermore, in order to further enhance the data quality and sample diversity of the sample images, corresponding data enhancement processing / data preprocessing may be performed thereon.

[0046] As an exemplary embodiment, after obtaining the sample images for pre-training, the method of the present application may further include:

[0047] The sample image is subjected to data enhancement processing, wherein the data enhancement processing includes rotation, translation, and noise addition.

[0048] It can be understood that the specific processing content of data enhancement processing is not the focus of the present application and the existing technology can be used. Therefore, the specific processing content including rotation, translation and noise addition will not be further explained here.

[0049] At the same time, the data enhancement processing involved in the present application can also be directly included in the process of acquiring the sample image, which is also possible in actual situations and can be configured according to actual needs.

[0050] In addition, the data enhancement processing involved in the present application can also be included in the subsequent different model pre-training stages, that is, the subsequent four stages may all involve data enhancement processing, and the specific data enhancement processing content adopted can also be fine-tuned, thereby achieving different data enhancement processing / different data enhancement processing contents in different stages.

[0051] Step S102, in the first stage, m consecutive sample images are input into the pre-trained model, and the pre-trained model is trained to output m consecutive sample images, so that the input and output are consistent;

[0052] It can be understood that after obtaining the sample image, the four-stage model pre-training processing involved in the model pre-training architecture of the present application can be carried out.

[0053] For the four stages of model pre-training processing, it is necessary to extract m (specified number) continuous sample images (images / slices) from the initial sample image set. Although each stage involves m images, these m images are not necessarily the same. The only similarity is in quantity. In this case, it also means that the number of sample images obtained in step S101 is greater than m, and usually much greater than m.

[0054] Of course, in some cases, the four stages can directly use the same m images, which is also possible.

[0055] For the first stage here, the purpose is to train the model to calculate the output of m consecutive images that are the same as the input, and because the input provides complete image information, the training task of this stage is the simplest.

[0056] As an exemplary embodiment, in the first stage, m consecutive sample images are input into the pre-trained model, and the pre-trained model is trained to output m consecutive sample images so that the input and output are consistent, which may specifically include:

[0057] In the first stage, m continuous sample images are input into the pre-trained model at the x-axis level, y-axis level and z-axis level respectively, and the pre-trained model is trained to output m continuous sample images so that the input and output are consistent, wherein the processing sequence of the x-axis level and the y-axis level are swapped as required.

[0058] It is easy to see that in specific operations, training at three levels, namely, x-axis, y-axis and z-axis, may be involved. The x-axis level training may be performed first, followed by the y-axis level training, and then the z-axis level training. Alternatively, the y-axis level training may be performed first, followed by the x-axis level training, and then the z-axis level training. The order of the x-axis level and the y-axis level training may be adjusted / swapped (the same applies to the implementation examples involved in the subsequent stages).

[0059] By carrying out more refined training from three specific dimensions / levels, more efficient and precise training results can be achieved.

[0060] As an example, the processing here may include:

[0061] (1) Extract and input m consecutive sample images on the x-axis plane, and let the model predict the corresponding m consecutive images so that the input and output contents are consistent;

[0062] (2) Extract and input m consecutive sample images on the y-axis level, and let the model predict the corresponding m consecutive images so that the input and output contents are consistent;

[0063] (3) Extract and input m consecutive sample images on the z-axis plane, and let the model predict the corresponding m consecutive images so that the input and output contents are consistent.

[0064] It can be noted that the m images involved in the specific training of each level can be extracted separately from the multiple sample images in step S101, rather than the m images involved in the three levels being consistent (the same applies to the instances involved in the subsequent stages).

[0065] In addition, for the extracted sample image, a fixed image size configuration may also be involved, for example, it may be configured as a fixed size of h×w×d (height×width×depth), where It is regarded as the channel dimension. Of course, the configuration of fixed image size here can also be directly included in the scope of the early data enhancement processing of sample images (the same applies to the examples involved in the subsequent stages).

[0066] During the training process, the mean square error loss function (MSE) can be used to update the model parameters to achieve the purpose of model optimization / training (the same applies to the examples involved in the subsequent stages).

[0067] During the training process, it can be understood that in the case of an original model, different training processes will produce different models. In this process, the model with the best processing effect can be selected as the output based on the evaluation of indicators such as the root mean square error loss function.

[0068] As an example, the pre-trained model involved in this application can specifically adopt the model architecture of the Unet model. Of course, other types of model architectures can also be adopted, or corresponding improvements can be made on the basis of the existing model architecture, or even a completely self-developed, novel model architecture can be adopted. All of these are possible.

[0069] Step S103, in the second stage, the image obtained by randomly erasing the image contents of m consecutive sample images is input into the pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent;

[0070] For the second stage here, the purpose is to train the model's ability to reconstruct missing images, that is, input the original m consecutive images, then randomly erase some content, and use the model to calculate and output the same m consecutive images, so that the original input and output are consistent. Due to the lack of input information, the prediction difficulty of the model will also increase, but the model can still complete the reconstruction through adjacent image information. The training task of this stage is the second most difficult task.

[0071] As an exemplary embodiment, in the second stage, an image obtained by randomly erasing the image contents of m consecutive sample images is input into a pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing so that the input and output are consistent, which may specifically include:

[0072] In the second stage, images obtained by randomly erasing the image contents of m sample images at the x-axis level, y-axis level, and z-axis level are input into the pre-trained model, and the pre-trained model is trained to output m consecutive sample images before processing, so that the input and output are consistent. The processing sequence of the x-axis level and the y-axis level is swapped as required.

[0073] As an example, the processing here may include:

[0074] (1) Extract m consecutive sample images from the x-axis plane, randomly erase the image content and input them, and let the model predict the corresponding m consecutive images in the same way, so that the input and output contents are consistent;

[0075] (2) Extract m consecutive sample images on the y-axis plane, randomly erase the image content and input them, and let the model predict the corresponding m consecutive images in the same way, so that the input and output contents are consistent;

[0076] (3) Extract m consecutive sample images on the z-axis plane, randomly erase the image content and input them, and let the model predict the corresponding m consecutive images in the same way, so that the input and output contents are consistent.

[0077] Among them, the random erasing of image content can be restricted to not erasing all of it, retaining the content of at least one image, and the number of erased images increases from small to large, that is, as the number of training times increases, the erased content also increases (the same applies to the instances involved in the subsequent stages).

[0078] Step S104, in the third stage, m consecutive sample images are divided into three equal parts, namely, upper, middle and lower parts, and then the middle part is randomly discarded and only the middle part is randomly retained, and the images are input into the pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent;

[0079] For the third stage here, the purpose is to train the model to predict contextual content, that is, input the original continuous m images, then divide the m images into three equal parts, record them as upper, middle and lower (upper part, middle part and lower part, can also be referred to as front part, middle part and back part), finally randomly erase the content of the middle part and keep the content of the upper and lower parts, or keep the content of the middle part and erase the content of the upper and lower parts, and calculate and output the same m consecutive images through the model, so that the input and output content are consistent, and reconstruct the erased content. Since in this task, part of the input continuous image content is missing (the continuous image content is randomly erased), it will further increase the difficulty of model prediction, so the training task of this stage is a medium-difficulty task.

[0080] As an exemplary embodiment, in the third stage, m continuous sample images are divided into three equal parts: upper, middle and lower, and then the middle part is randomly discarded and only the middle part is randomly retained. The images are input into the pre-trained model, and the pre-trained model is trained to output m continuous sample images so that the input and output are consistent. Specifically, the following may be included:

[0081] In the third stage, m continuous sample images are divided into three equal parts, namely upper, middle and lower parts, at the x-axis level, y-axis level and z-axis level respectively, and the middle part is randomly discarded or only the middle part is randomly retained, and the images are input into the pre-trained model. The pre-trained model is trained to output the m continuous sample images before processing, so that the input and output are consistent. Among them, the processing sequence of the x-axis level and the y-axis level is swapped as required.

[0082] As an example, the processing here may include:

[0083] (1) Extract m consecutive sample images from the x-axis plane, divide them into three equal parts: upper, middle and lower. Then randomly discard the middle part or randomly keep only the middle part. Then input the obtained images and let the model predict the corresponding m consecutive images in the same way, so that the input and output contents are consistent.

[0084] (2) Extract m consecutive sample images on the y-axis plane, divide them into three equal parts: upper, middle and lower. Then randomly discard the middle part or randomly keep only the middle part. Then input the obtained image and let the model predict the corresponding m consecutive images in the same way, so that the input and output contents are consistent.

[0085] (3) Extract m consecutive sample images from the z-axis plane, divide them into three equal parts: upper, middle and lower. Then randomly discard the middle part or randomly keep only the middle part to obtain the image. Then input the image and let the model predict the corresponding m consecutive images in the same way, so that the input and output contents are consistent.

[0086] Step S105, in the fourth stage, m consecutive sample images including the cerebral hemorrhage image area to a preset degree, the image contents are randomly erased, the image is divided into three equal parts of upper, middle and lower, and then the middle part is randomly discarded, and the image is divided into three equal parts of upper, middle and lower, and then only the middle part is randomly retained, and the image is input into the pre-trained model, the pre-trained model is trained to output m consecutive sample images, so that the input and output are consistent, and the model pre-training is completed, wherein the pre-trained model that has completed the model pre-training is used to train the cerebral hemorrhage segmentation prediction model based on the labeled new sample images, and the cerebral hemorrhage segmentation prediction model is used to perform segmentation prediction processing of the cerebral hemorrhage area on the image input to the model.

[0087] For the fourth stage here, the purpose is to focus on the brain hemorrhage image area and train the model to predict the context content, that is, input the original continuous m images containing as many brain hemorrhage areas as possible (the preset degree of the brain hemorrhage image area can be used as the boundary to divide, mainly for the area, and secondly for the number of areas and other elements), and then randomly erase some image content, or divide the image containing the brain hemorrhage image area into three equal parts of the same number, recorded as upper, middle and lower (upper part, middle part and lower part, can also be referred to as the front part, middle part and back part), erase the content of the middle part and retain the content of the upper and lower parts, or retain the content of the middle part and erase the content of the upper and lower parts, and finally calculate and output the same continuous m images through the model, so that the input and output content are consistent, and the erased content is reconstructed. The task of this stage is directly related to the segmentation of brain hemorrhage (randomly erase the continuous lesion image content), which can purposefully and targetedly allow the model to reconstruct a large number of brain hemorrhage area images, thereby increasing the pre-training effect.

[0088] As an exemplary embodiment, in the fourth stage, m consecutive sample images including the cerebral hemorrhage image area to a preset degree are randomly erased, divided into three equal parts, and then the middle part is randomly discarded, and divided into three equal parts, and then only the middle part is randomly retained, and the images are input into the pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent, which may specifically include:

[0089] In the fourth stage, m continuous sample images that include the cerebral hemorrhage image area to a preset degree are randomly erased, divided into three equal parts of upper, middle and lower, and then the middle part is randomly discarded, and divided into three equal parts of upper, middle and lower, and then only the middle part is randomly retained, respectively, at the x-axis level, y-axis level and z-axis level, and input into the pre-trained model, and the pre-trained model is trained to output the m continuous sample images before processing, so that the input and output are consistent, wherein the processing sequence of the x-axis level and the y-axis level is swapped as required.

[0090] As an example, the processing here may include:

[0091] (1) Extract m consecutive sample images on the x-axis plane, randomly erase the image content, divide it into three equal parts, randomly discard the middle part, and randomly keep only the middle part. Then input the obtained image and let the model predict the corresponding m consecutive images in the same way, so that the input and output contents are consistent;

[0092] (2) Extract m consecutive sample images on the y-axis plane, randomly erase the image content, divide it into three equal parts: upper, middle and lower, randomly discard the middle part, and randomly retain only the middle part. Then input the obtained image and let the model predict the corresponding m consecutive images in the same way, so that the input and output contents are consistent;

[0093] (3) Extract m consecutive sample images on the z-axis plane, randomly erase the image content, divide the image into three equal parts, randomly discard the middle part, and randomly retain only the middle part to obtain the image, and then input it, and let the model predict the corresponding m consecutive images in the same way, so that the input and output contents are consistent.

[0094] It can also be seen from the specific contents of the model training architecture in the above stages that this application adopts a self-supervised learning method of first reconstructing the image, then repairing the missing image, then repairing the missing continuous image, and finally repairing the lesion site image (brain hemorrhage area image), so that the model can learn the image features related to the lesion, thereby improving the final model segmentation accuracy.

[0095] At the same time, this application also follows the process from easy to difficult, gradually increasing the difficulty of the pre-training tasks, and gradually improving the robustness and efficiency of the model.

[0096] Finally, this application gradually increases the relevance of the tasks, from random image restoration to image restoration of areas related to the lesion, thereby gradually increasing the model's ability to perceive the lesion.

[0097] After completing the pre-training of the model, the subsequent model training based on the labeled training samples mentioned above can be continued.

[0098] In this regard, as an exemplary embodiment, the method of the present application may further include:

[0099] Acquire a sample image with a label of a cerebral hemorrhage region, wherein the sample image with a label of a cerebral hemorrhage region includes at least one of a computer tomography image and a magnetic resonance imaging image;

[0100] Based on the sample images with labeled cerebral hemorrhage areas, the pre-trained model of the model pre-training is trained to obtain a cerebral hemorrhage segmentation prediction model.

[0101] It can be understood that the model training process here is not the focus of the present application, and the model training architecture and loss function involved can refer to the existing technology, so the specific description will not be given here.

[0102] Among them, since the pre-trained model used for model training is a model enhanced by the present application scheme, the model training link here can be carried out more efficiently and accurately, which in turn helps to obtain a brain hemorrhage segmentation prediction model with high-quality processing effects that can be put into practical use.

[0103] It is worth noting that the sample image marked with the cerebral hemorrhage area label involved here is not necessarily the same image as the sample image obtained in the previous step S101. In other words, the difference is not just whether the cerebral hemorrhage area label is marked. The sample image before the cerebral hemorrhage area label is marked by manual or automatic labeling tools can be completely different or not completely the same as the sample image obtained in step S101. Of course, it can also be completely the same, which can be configured according to actual needs.

[0104] After completing the training of the model, it is obvious that the practical application of the model can continue. In this regard, as an exemplary embodiment, the method of the present application may also include:

[0105] Obtain the target image to be segmented and predicted;

[0106] Input the target image into the cerebral hemorrhage segmentation prediction model;

[0107] Extract the cerebral hemorrhage area segmentation prediction result corresponding to the target image output by the cerebral hemorrhage segmentation prediction model.

[0108] In addition, after obtaining the prediction results of the cerebral hemorrhage area segmentation, further data applications may be involved in accordance with clinical work needs, such as result display, result preservation, result forwarding, output completion processing prompts or corresponding data analysis and processing. The specific data application can be configured according to actual needs, and this application does not make specific limitations.

[0109] In general, for the above solution content, corresponding to the high-quality training goal of the cerebral hemorrhage segmentation prediction model, this application provides a novel model pre-training architecture for the training of the cerebral hemorrhage segmentation prediction model. The pre-trained model obtained based on this processing will help to carry out more efficient and high-precision model training in the future, and can promote the acquisition of a cerebral hemorrhage segmentation prediction model with higher processing accuracy, so that the model has stronger processing performance and better application value.

[0110] The above is an introduction to the processing method of the cerebral hemorrhage segmentation prediction model based on self-supervised learning provided by the present application. In order to facilitate better implementation of the processing method of the cerebral hemorrhage segmentation prediction model based on self-supervised learning provided by the present application, the present application also provides a processing device of the cerebral hemorrhage segmentation prediction model based on self-supervised learning from the perspective of functional modules.

[0111] See also Figure 3 , Figure 3 This is a structural diagram of a processing device for a cerebral hemorrhage segmentation prediction model based on self-supervised learning in this application. In this application, the processing device 300 based on a cerebral hemorrhage segmentation prediction model based on self-supervised learning may specifically include the following structure:

[0112] An acquisition unit 301 is used to acquire sample images for pre-training, wherein the sample images include at least one of computer tomography images and magnetic resonance imaging images, and the sample images are images without annotations;

[0113] The first training unit 302 is used to input m consecutive sample images into the pre-training model in the first stage, and train the pre-training model to output m consecutive sample images so that the input and output are consistent;

[0114] The second training unit 303 is used for, in the second stage, inputting the images obtained by randomly erasing the image contents of m consecutive sample images into the pre-training model, and training the pre-training model to output the m consecutive sample images before processing, so that the input and output are consistent;

[0115] The third training unit 304 is used to divide the m continuous sample images into three equal parts, namely, upper, middle and lower parts, and then randomly discard the middle part and randomly retain only the middle part, input the obtained images into the pre-training model, and train the pre-training model to output the m continuous sample images before processing, so that the input and output are consistent;

[0116] The fourth training unit 305 is used to input the m consecutive sample images, which include the cerebral hemorrhage image area to a preset degree, into a pre-trained model in the fourth stage, randomly erase the image content, divide the image into three equal parts of upper, middle and lower, and then randomly discard the middle part, and divide the image into three equal parts of upper, middle and lower, and then randomly retain only the middle part, train the pre-trained model to output m consecutive sample images, so that the input and output are consistent, and complete the model pre-training, wherein the pre-trained model that has completed the model pre-training is used to train the cerebral hemorrhage segmentation prediction model based on the labeled new sample images, and the cerebral hemorrhage segmentation prediction model is used to perform segmentation prediction processing of the cerebral hemorrhage area on the image input by the model.

[0117] In an exemplary embodiment, the first training unit 302 is specifically configured to:

[0118] In the first stage, m continuous sample images are input into the pre-trained model at the x-axis level, y-axis level and z-axis level respectively, and the pre-trained model is trained to output m continuous sample images so that the input and output are consistent, wherein the processing sequence of the x-axis level and the y-axis level are swapped as required.

[0119] In another exemplary embodiment, the second training unit 303 is specifically configured to:

[0120] In the second stage, images obtained by randomly erasing the image contents of m consecutive sample images at the x-axis level, y-axis level, and z-axis level are input into the pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent. The processing sequence of the x-axis level and the y-axis level is swapped as required.

[0121] In another exemplary embodiment, the third training unit 304 is specifically configured to:

[0122] In the third stage, m continuous sample images are divided into three equal parts, namely upper, middle and lower parts, at the x-axis level, y-axis level and z-axis level respectively, and the middle part is randomly discarded or only the middle part is randomly retained, and the images are input into the pre-trained model. The pre-trained model is trained to output the m continuous sample images before processing, so that the input and output are consistent. Among them, the processing sequence of the x-axis level and the y-axis level is swapped as required.

[0123] In another exemplary embodiment, the fourth training unit 305 is specifically configured to:

[0124] In the fourth stage, m continuous sample images that include the cerebral hemorrhage image area to a preset degree are randomly erased, divided into three equal parts of upper, middle and lower, and then the middle part is randomly discarded, and divided into three equal parts of upper, middle and lower, and then only the middle part is randomly retained, respectively, at the x-axis level, y-axis level and z-axis level, and input into the pre-trained model, and the pre-trained model is trained to output the m continuous sample images before processing, so that the input and output are consistent, wherein the processing sequence of the x-axis level and the y-axis level is swapped as required.

[0125] In yet another exemplary embodiment, the apparatus further comprises:

[0126] The data enhancement unit 306 is used to perform data enhancement processing on the sample image, wherein the data enhancement processing includes rotation, translation and noise addition.

[0127] In yet another exemplary embodiment, the apparatus further includes a fifth training unit 307, configured to:

[0128] Acquire a sample image with a label of a cerebral hemorrhage region, wherein the sample image with a label of a cerebral hemorrhage region includes at least one of a computer tomography image and a magnetic resonance imaging image;

[0129] Based on the sample images with labeled cerebral hemorrhage areas, the pre-trained model of the model pre-training is trained to obtain a cerebral hemorrhage segmentation prediction model.

[0130] This application also provides a processing device from the perspective of hardware structure, see Figure 4 , Figure 4 401, a computer program stored in the memory 402, and an input / output device 403. The processor 401 is used to execute the computer program stored in the memory 402. Figure 1 The steps of the processing method of the cerebral hemorrhage segmentation prediction model based on self-supervised learning in the corresponding embodiment; or, the processor 401 is used to execute the computer program stored in the memory 402 to implement the following Figure 3 Corresponding to the functions of each unit in the embodiment, the memory 402 is used to store the processor 401 executing the above Figure 1 The computer program required for the processing method of the cerebral hemorrhage segmentation prediction model based on self-supervised learning in the corresponding embodiment.

[0131] Exemplarily, the computer program may be divided into one or more modules / units, one or more modules / units are stored in the memory 402, and executed by the processor 401 to complete the present application. One or more modules / units may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program in the computer device.

[0132] The processing device may include, but is not limited to, a processor 401, a memory 402, and an input / output device 403. Those skilled in the art will appreciate that the illustration is merely an example of a processing device and does not constitute a limitation on the processing device, and may include more or fewer components than shown in the illustration, or a combination of certain components, or different components. For example, the processing device may also include a network access device, a bus, etc., and the processor 401, the memory 402, the input / output device 403, etc. are connected via a bus.

[0133] The processor 401 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the processing device, and uses various interfaces and lines to connect various parts of the entire device.

[0134] The memory 402 can be used to store computer programs and / or modules. The processor 401 implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory 402 and calling the data stored in the memory 402. The memory 402 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created according to the use of the processing device, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0135] When the processor 401 is used to execute the computer program stored in the memory 402, the following functions can be implemented:

[0136] Acquire sample images for pre-training, wherein the sample images include at least one of computer tomography images and magnetic resonance imaging images, and the sample images are images without annotations;

[0137] In the first stage, m consecutive sample images are input into the pre-trained model, and the pre-trained model is trained to output m sample images so that the input and output are consistent;

[0138] In the second stage, m first enhanced images after randomly erasing the image contents of m consecutive sample images are input into the pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent;

[0139] In the third stage, m consecutive sample images are divided into three equal parts: upper, middle and lower. The middle part is randomly discarded and only the middle part is randomly retained to obtain m second enhanced images, which are input into the pre-trained model. The pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent.

[0140] In the fourth stage, m consecutive sample images including the cerebral hemorrhage image area to a preset degree, m third enhanced images obtained by randomly erasing the image content, dividing the images into three equal parts of upper, middle and lower, and then randomly discarding the middle part, and dividing the images into three equal parts of upper, middle and lower, and then randomly retaining only the middle part, are input into the pre-trained model, and the pre-trained model is trained to output m consecutive sample images so that the input and output are consistent, and the model pre-training is completed, wherein the pre-trained model that has completed the model pre-training is used to train the cerebral hemorrhage segmentation prediction model based on the labeled new sample images, and the cerebral hemorrhage segmentation prediction model is used to perform segmentation prediction processing of the cerebral hemorrhage area on the image input to the model.

[0141] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the processing device, processing equipment and corresponding units of the above-described cerebral hemorrhage segmentation prediction model based on self-supervised learning can refer to the following. Figure 1 The description of the processing method of the cerebral hemorrhage segmentation prediction model based on self-supervised learning in the corresponding embodiment will not be repeated here.

[0142] A person of ordinary skill in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be completed by instructions, or by controlling related hardware through instructions. The instructions may be stored in a computer-readable storage medium and loaded and executed by a processor.

[0143] To this end, the present application provides a computer-readable storage medium, in which a plurality of instructions are stored, and the instructions can be loaded by a processor to execute the present application as follows: Figure 1 The steps of the processing method of the cerebral hemorrhage segmentation prediction model based on self-supervised learning in the corresponding embodiment, the specific operation can be referred to as Figure 1 The description of the processing method of the cerebral hemorrhage segmentation prediction model based on self-supervised learning in the corresponding embodiment will not be repeated here.

[0144] The computer-readable storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0145] Due to the instructions stored in the computer-readable storage medium, the present application can be executed. Figure 1The steps of the processing method of the cerebral hemorrhage segmentation prediction model based on self-supervised learning in the corresponding embodiment, therefore, the present application can be implemented as follows Figure 1 The beneficial effects that can be achieved by the processing method of the cerebral hemorrhage segmentation prediction model based on self-supervised learning in the corresponding embodiment are detailed in the previous description and will not be repeated here.

[0146] The above is a detailed introduction to the processing method, device, processing equipment and computer-readable storage medium of the self-supervised learning based cerebral hemorrhage segmentation prediction model provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A processing method for a cerebral hemorrhage segmentation prediction model based on self-supervised learning, characterized in that: The method comprises: Acquire sample images for pre-training, wherein the sample images include at least one of computer tomography images and magnetic resonance imaging images, and the sample images are images without annotations; In the first stage, m consecutive sample images are input into a pre-trained model, and the pre-trained model is trained to output m consecutive sample images so that the input and output are consistent; In the second stage, an image obtained by randomly erasing the image contents of m consecutive sample images is input into the pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing so that the input and output are consistent; In the third stage, m consecutive sample images are divided into three equal parts: upper, middle and lower, and then the middle part is randomly discarded and only the middle part is randomly retained to obtain an image, which is input into the pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent; In the fourth stage, m consecutive sample images including the cerebral hemorrhage image area to a preset degree, the image contents are randomly erased, the middle part is randomly discarded after being divided into three equal parts, and only the middle part is randomly retained after being divided into three equal parts, the images are input into the pre-trained model, the pre-trained model is trained to output m consecutive sample images, so that the input and output are consistent, and the model pre-training is completed, wherein the pre-trained model that has completed the model pre-training is used to train the cerebral hemorrhage segmentation prediction model based on the labeled new sample images, and the cerebral hemorrhage segmentation prediction model is used to carry out segmentation prediction processing of the cerebral hemorrhage area on the image input to the model; In the first stage, m consecutive sample images are input into a pre-trained model, and the pre-trained model is trained to output m consecutive sample images so that the input and output are consistent, including: In the first stage, m consecutive sample images are input into the pre-trained model at the x-axis level, the y-axis level and the z-axis level respectively, and the pre-trained model is trained to output m consecutive sample images so that the input and output are consistent, wherein the processing timing of the x-axis level and the y-axis level is swapped as required.

2. The method according to claim 1, characterized in that In the second stage, the image obtained by randomly erasing the image contents of m consecutive sample images is input into the pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing so that the input and output are consistent, including: In the second stage, images obtained by randomly erasing image contents of m consecutive sample images are input into the pre-trained model at the x-axis level, y-axis level and z-axis level respectively, and the pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent, wherein the processing timing of the x-axis level and the y-axis level are swapped as required.

3. The method according to claim 1, characterized in that: In the third stage, m consecutive sample images are divided into three equal parts, namely, upper, middle and lower parts, and then the middle part is randomly discarded and only the middle part is randomly retained, and the images are input into the pre-trained model, and the pre-trained model is trained to output m consecutive sample images so that the input and output are consistent, including: In the third stage, m consecutive sample images are divided into three equal parts, upper, middle and lower, at the x-axis level, y-axis level and z-axis level respectively, and the middle part is randomly discarded and only the middle part is randomly retained to obtain the image, which is input into the pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent, wherein the processing timing of the x-axis level and the y-axis level is swapped as required.

4. The method according to claim 1, characterized in that: In the fourth stage, m consecutive sample images including the cerebral hemorrhage image area to a preset degree are randomly erased, divided into three equal parts, the middle part is randomly discarded, and divided into three equal parts, the middle part is randomly retained, and the images are input into the pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent, including: In the fourth stage, at the x-axis level, y-axis level and z-axis level, respectively, m consecutive sample images including the cerebral hemorrhage image area to a preset extent are randomly erased of image content, divided into three equal parts of upper, middle and lower, and then the middle part is randomly discarded, and divided into three equal parts of upper, middle and lower, and then only the middle part is randomly retained, and input into the pre-trained model, and the pre-trained model is trained to output the m consecutive sample images before processing, so that the input and output are consistent, wherein the processing timing of the x-axis level and the y-axis level is swapped as required.

5. The method according to claim 1, characterized in that After obtaining the sample image for pre-training, the method further includes: Performing data enhancement processing on the sample image, wherein the data enhancement processing includes rotation, translation and noise addition.

6. The method according to claim 1, characterized in that The method further comprises: Acquire a sample image with a label of a cerebral hemorrhage region, wherein the sample image with a label of a cerebral hemorrhage region includes at least one of the computer tomography image and the magnetic resonance imaging image; Based on the sample image marked with the cerebral hemorrhage area label, the pre-trained model of the model pre-training is trained to obtain the cerebral hemorrhage segmentation prediction model.

7. A processing device for a cerebral hemorrhage segmentation prediction model based on self-supervised learning, characterized in that: The device comprises: An acquisition unit, used to acquire sample images for pre-training, wherein the sample images include at least one of computer tomography images and magnetic resonance imaging images, and the sample images are images without annotations; A first training unit is used for, in a first stage, inputting m consecutive sample images into a pre-training model, and training the pre-training model to output m consecutive sample images so that the input and output are consistent; A second training unit is used for, in the second stage, inputting the images obtained by randomly erasing the image contents of m consecutive sample images into the pre-training model, and training the pre-training model to output the m consecutive sample images before processing so that the input and output are consistent; A third training unit is used for, in the third stage, dividing the m consecutive sample images into three equal parts, namely, upper, middle and lower parts, randomly discarding the middle part and randomly retaining only the middle part, inputting the obtained images into the pre-trained model, and training the pre-trained model to output the m consecutive sample images before processing, so that the input and output are consistent; The fourth training unit is used to input the m consecutive sample images, which include the cerebral hemorrhage image area to a preset degree, into the pre-trained model in the fourth stage, randomly erase the image content, divide them into three equal parts of upper, middle and lower, and then randomly discard the middle part, and divide them into three equal parts of upper, middle and lower, and then randomly retain only the middle part, to train the pre-trained model to output the m consecutive sample images, so that the input and output are consistent, and complete the model pre-training, wherein the pre-trained model that has completed the model pre-training is used to train the cerebral hemorrhage segmentation prediction model based on the labeled new sample images, and the cerebral hemorrhage segmentation prediction model is used to carry out segmentation prediction processing of the cerebral hemorrhage area on the image input by the model; The first training unit is specifically used for: In the first stage, m consecutive sample images are input into the pre-trained model at the x-axis level, the y-axis level and the z-axis level respectively, and the pre-trained model is trained to output m consecutive sample images so that the input and output are consistent, wherein the processing timing of the x-axis level and the y-axis level is swapped as required.

8. A processing device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program, and when the processor calls the computer program in the memory, the method according to any one of claims 1 to 6 is executed.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the method according to any one of claims 1 to 6.

Images