Methods for training a predictive model for bleeding risk and related products

Abstract

The application discloses a method for training a prediction model for a bleeding risk and related products. The method comprises: collecting medical image data, genotypes and bleeding risk labels of a subject, wherein the medical image data is labeled with a bleeding mask and a classification type of cerebral cavernous malformations; based on the medical image data, feature extraction is performed using the feature extraction unit to obtain image features; the genotypes are embedded into the image features using the embedding unit to obtain embedded features; the embedded features are input into the prediction unit for bleeding risk prediction and classification prediction, and a loss function is calculated based on the bleeding risk labels and the classification type; and the prediction model is trained based on the loss function, so that the prediction unit outputs bleeding risk prediction results and classification prediction results. The scheme of the application can improve accurate bleeding risk and classification prediction, and provide comprehensive decision support for clinics.

Description

Technical Field

[0001] This application generally relates to the field of bleeding risk prediction technology. More specifically, this application relates to a method for training a prediction model for bleeding risk, an apparatus for bleeding risk prediction, an electronic device, and a computer-readable storage medium. Background Technology

[0002] Cavernous malformations of the brain and spinal cord are a specific type of low-flow-velocity vascular malformation. Its pathological features include abnormal vascular structure: the vessels consist of dilated lumens with walls containing only a single layer of endothelial cells, lacking the smooth muscle and elastic layers required in normal blood vessel walls, thus failing to maintain the structural stability and function of normal blood vessels. This disease can occur in critical areas of the central nervous system, such as the brain parenchyma or spinal cord, and has significant clinical consequences. Some patients experience repeated small hemorrhages over a long period, which gradually damage nerve tissue, ultimately leading to progressive neurological dysfunction such as hemiplegia, sensory disturbances, and motor coordination disorders, severely impacting the patient's quality of life.

[0003] In the field of predicting the risk of bleeding in cavernous malformations of the brain and spinal cord, the core idea of ​​existing technologies is feature extraction and machine learning training: First, relevant information is extracted from the patient's clinical information (such as medical history, symptoms, etc.) and imaging data (such as MRI images). Then, this information is transformed into features using traditional image processing methods (such as pixel statistics, edge detection, etc.) or deep learning algorithms (such as basic CNN networks). Finally, the extracted features are input into a conventional machine learning model to complete the training. However, existing technologies have three major drawbacks: First, the modeling method is mismatched with the characteristics of disease development. Bleeding is not treated as a dynamic event in a continuous time dimension. Only discrete labels (such as bleeding / no bleeding) are used for modeling, and no attenuation mechanism is included. This fails to reflect the clinical pattern that patients with recent bleeding or large bleeding volumes have a higher probability of rebleeding. Second, the efficiency of image feature learning is low. The correlation between bleeding risk prediction and classification judgment is not established. The two tasks are isolated from each other, which prevents the model from mining deep features in the images that simultaneously support both tasks. Third, the system stability is disconnected from clinical priors. The significant impact of genotype (especially PIK3CA mutant and non-mutant types) on bleeding risk is not considered. The risk differences output by the model may deviate from clinical priors, resulting in insufficient reliability.

[0004] In view of this, there is an urgent need to provide a scheme for training a predictive model for bleeding risk in order to effectively solve the problems of accuracy and stability of existing technologies, while outputting classification results to provide comprehensive decision support for clinical practice. Summary of the Invention

[0005] In order to at least address one or more of the technical problems mentioned above, this application proposes a scheme for training a predictive model for bleeding risk in several aspects.

[0006] In a first aspect, this application provides a method for training a prediction model for bleeding risk, wherein the prediction model includes a feature extraction unit, an embedding unit, and a prediction unit, and the method includes: acquiring medical imaging data, genotype, and bleeding risk labels reflecting continuous-time decay characteristics related to cavernous malformations in the brain of a subject, wherein the medical imaging data is labeled with bleeding masks and classification types of cavernous malformations in the brain; based on the medical imaging data, performing feature extraction using the feature extraction unit to obtain image features; using the embedding unit to embed the genotype into the image features to obtain embedded features; inputting the embedded features into the prediction unit to perform bleeding risk prediction and classification prediction, and calculating a loss function based on the bleeding risk label and the classification type; training the prediction model based on the loss function, such that the prediction unit outputs bleeding risk prediction results and classification prediction results.

[0007] In some embodiments, the medical imaging data includes T1-weighted imaging data, T2-weighted imaging data, and magnetic susceptibility-weighted imaging data; the genotype includes PIK3CA mutant and non-PIK3CA mutant.

[0008] In some embodiments, the classification type includes the Zabramski classification, which includes type I, type II, or type III.

[0009] In some embodiments, the bleeding risk label reflecting the continuous time decay characteristics is acquired by the following operations: acquiring the subject's follow-up bleeding record; in response to the subject's follow-up bleeding record not bleeding within a preset follow-up period, setting the bleeding risk label to 0; in response to the subject's follow-up bleeding record experiencing rebleeding within the preset follow-up period, setting the bleeding risk label to (V1 / (V1 + V0) + (ΔT_max -(t1-t0)) / ΔT_max) / 2, where V1 represents the rebleeding amount, V0 represents the baseline bleeding amount, ΔT_max represents the preset follow-up period, and (t1-t0) represents the bleeding time interval.

[0010] In some embodiments, the amount of bleeding is calculated by: calculating the voxel size of the medical image data and the number of pixels in the bleeding region of the bleeding mask; and calculating the amount of bleeding based on the product of the voxel size and the number of pixels.

[0011] In some embodiments, the loss function includes a total loss function calculated from a first loss function, a second loss function, and a third loss function, wherein the first loss function is the regression loss between the hemorrhage risk prediction result and the hemorrhage risk label; the second loss function is the classification loss between the genotyping prediction result and the genotyping label; and the third loss function is the counterfactual regularization loss based on the influence of different genotypes on the risk of hemorrhage from cavernous malformations in the brain.

[0012] In some embodiments, the method further includes: extracting texture features from the medical image data; and embedding the texture features into the embedding features using the embedding unit.

[0013] In a second aspect, this application provides an apparatus for predicting bleeding risk, comprising: a processor; and a memory storing computer instructions for predicting bleeding risk, wherein when the computer instructions are executed by the processor, the following operations are performed: acquiring medical image data and genotype of a subject, wherein the medical image data is annotated with a bleeding mask; inputting the medical image data and the genotype into a prediction model trained in the first aspect above to perform bleeding risk prediction and subtyping prediction, thereby obtaining bleeding risk prediction results and subtyping prediction results.

[0014] In a third aspect, this application provides an electronic device comprising: a processor; and a memory having computer instructions stored thereon for training a bleeding risk prediction model, wherein when the computer instructions are executed by the processor, the method described in the first aspect is performed; or having computer instructions stored thereon for bleeding risk prediction, wherein when the computer instructions are executed by the processor, the operation performed by the apparatus described in the second aspect is performed.

[0015] In a fourth aspect, this application provides a computer-readable storage medium having stored thereon computer program instructions for training a bleeding risk prediction model, which, when executed by one or more processors, cause the method described in the first aspect to be implemented; or having stored thereon computer program instructions for bleeding risk prediction, which, when executed by one or more processors, cause the operation performed by the apparatus described in the second aspect to be implemented.

[0016] By employing the training scheme for the prediction model of bleeding risk provided above, this embodiment of the application effectively solves the problems of mismatch between existing modeling and disease development characteristics, low efficiency of image feature learning, and insufficient system stability by collecting bleeding risk labels with continuous time decay characteristics, combining genotype embedding and dual-task prediction of bleeding risk and subtyping, and calculating loss function to train the model. This not only improves the accuracy of bleeding risk prediction but also outputs subtyping results simultaneously, providing comprehensive and reliable technical support for clinical assessment, treatment plan formulation, and follow-up monitoring of cavernous malformations of the brain and spinal cord. Attached Figure Description

[0017] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein:

[0018] Figure 1 This is an exemplary flowchart illustrating a method 100 for training a prediction model for bleeding risk according to an embodiment of this application;

[0019] Figure 2 This is an exemplary flowchart illustrating the overall training of a predictive model for bleeding risk according to an embodiment of this application;

[0020] Figure 3 This is an exemplary structural block diagram illustrating an apparatus 300 for predicting bleeding risk according to an embodiment of this application;

[0021] Figure 4 This is an exemplary flowchart illustrating a method for predicting bleeding risk according to an embodiment of this application;

[0022] Figure 5 An exemplary structural block diagram of an electronic device 500 according to an embodiment of this application is shown. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0024] It should be understood that the terms "comprising" and "including" as used in the specification and claims of this application indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0025] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0026] As used in this specification and claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0027] The specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0028] Figure 1 This is an exemplary flowchart illustrating a method 100 for training a prediction model for bleeding risk according to an embodiment of this application. In some embodiments, the prediction model may include a feature extraction unit, an embedding unit, and a prediction unit. In some implementations, the prediction model may be, for example, a convolutional neural network or a hybrid network of convolutional neural networks and Transformers, including a feature extractor, an embedding layer, and a prediction head.

[0029] like Figure 1As shown, the method 100 may include: Step S101: Collecting medical imaging data, genotype, and bleeding risk labels reflecting continuous time decay characteristics related to cavernous malformations in the brain of the subject, wherein the medical imaging data are marked with bleeding masks and classification types of cavernous malformations in the brain; Step S102: Based on the medical imaging data, performing feature extraction using a feature extraction unit to obtain image features; Step S103: Using an embedding unit to embed the genotype into the image features to obtain embedded features; Step S104: Inputting the embedded features into a prediction unit to perform bleeding risk prediction and classification prediction, and calculating a loss function based on the bleeding risk label and classification type; Step S105: Training a prediction model based on the loss function, so that the prediction unit outputs bleeding risk prediction results and classification prediction results.

[0030] First, in step S101, medical imaging data, genotype, and hemorrhage risk labels reflecting continuous-time decay characteristics related to the cavernous malformation (CVM) of the subject are collected. The medical imaging data includes annotations of the hemorrhage mask and the classification of the CVM. It can be understood that CVM refers to vascular malformations of the central nervous system, composed of abnormally dilated blood vessels lacking a smooth muscle layer in the vessel walls, making them prone to repeated bleeding and potentially leading to neurological dysfunction.

[0031] In some embodiments, the medical imaging data may include T1-weighted imaging data, T2-weighted imaging data, and susceptibility-weighted imaging (SWI) data; the genotype may include PIK3CA mutant and non-PIK3CA mutant.

[0032] In some implementation scenarios, MRI equipment can be used to acquire the aforementioned medical imaging data. By pixel-level labeling of the bleeding areas in the medical images, for example, a bleeding area is 1 and a non-bleeding area is 0, a binary image is formed as a bleeding mask, which is used to accurately locate the bleeding foci. In some embodiments, the classification type of cavernous malformations in the brain can include the Zabramski classification, which includes type I, type II, or type III. It can be understood that the Zabramski classification is a classification system of cavernous malformations based on MRI features, reflecting the bleeding stage and pathological evolution of the lesion. Type I corresponds to subacute hemorrhage; type II corresponds to "popcorn" hemorrhage; and type III corresponds to chronic hemorrhage. The Zabramski classification is directly related to the bleeding activity of the lesion (type I > type II > type III). The model uses classification prediction as an auxiliary task, which can mine classification features related to bleeding risk in the images (such as the thickness of the hemosiderin ring), improving the interpretability of risk prediction.

[0033] In other implementation scenarios, DNA is extracted from peripheral venous blood samples (e.g., 5 mL), and next-generation sequencing (NGS) is used to detect hotspot mutations in the PIK3CA gene (such as E542K, E545K, H1047R) to determine whether the mutation is PIK3CA mutant or non-PIK3CA mutant. PIK3CA mutants refer to gain-of-function mutations in the PIK3CA gene (such as E545K), leading to abnormal activation of the PI3K / AKT signaling pathway, promoting vascular endothelial cell proliferation and vascular malformation, and significantly increasing the risk of bleeding. Non-PIK3CA mutants refer to cases where no PIK3CA mutation is detected, which may involve carrying other gene mutations (such as MAP3K3) or are sporadic cases, with a relatively low risk of bleeding.

[0034] In some embodiments, a bleeding risk label reflecting continuous time decay characteristics can be collected through the following operations: collecting the subject's follow-up bleeding records; in response to the subject's follow-up bleeding records not bleeding within a preset follow-up period, setting the bleeding risk label to 0; in response to the subject's follow-up bleeding records experiencing rebleeding within the preset follow-up period, setting the bleeding risk label to (V1 / (V1+V0) + (ΔT_max -(t1-t0)) / ΔT_max) / 2, where V1 represents the rebleeding amount, V0 represents the baseline bleeding amount, ΔT_max represents the preset follow-up period, and (t1-t0) represents the bleeding time interval.

[0035] The aforementioned follow-up bleeding records can be obtained from the hospital's electronic medical record system by extracting the patient's medical records within a preset follow-up period after baseline (e.g., ΔT_max = 2 years). Combined with MRI re-examination results, rebleeding events can be confirmed. For example, if a new SWI low-signal lesion appears or the volume of the original bleeding lesion increases by ≥20%, it is determined as rebleeding, and the bleeding time t1 is recorded; otherwise, it is considered non-bleeding. Furthermore, the bleeding risk label for non-bleeding is set to 0; the bleeding risk label for rebleeding is set to (V1 / (V1+V0) +(ΔT_max -(t1-t0)) / ΔT_max) / 2. This label breaks through the limitations of the traditional 0 / 1 discrete label, quantifying risk through both time and bleeding volume dimensions, which better reflects the dynamic bleeding patterns of cavernous malformations.

[0036] In some embodiments, the amount of bleeding can be calculated by: calculating the voxel size of the medical image data and the number of pixels in the bleeding area of ​​the bleeding mask, and then calculating the amount of bleeding based on the product of the voxel size and the number of pixels. As an example, the amount of bleeding V = N × unit, where N represents the number of pixels, which is the total number of pixels with a value of 1 in the bleeding mask, reflecting the spatial extent of the hematoma; unit represents the voxel size, which is the actual size (in mm³) of a single pixel in the medical image in three-dimensional space, determined by the image resolution.

[0037] Next, in step S102, feature extraction is performed using a feature extraction unit based on the medical image data to obtain image features. That is, the feature extraction unit extracts discriminative deep features (such as lesion morphology, edges, signal intensity distribution, etc.) from the medical image, which are quantitative parameters that can characterize the attributes of the lesion.

[0038] In some implementation scenarios, the feature extraction unit may include at least one convolutional block or a hybrid of convolutional blocks and Transformers. The convolutional block can consist of, for example, convolutional layers, BatchNorm layers, ReLU activation functions, and max-pooling layers. The Transformer can flatten the feature maps output by the convolutional block into a sequence, learn long-distance dependencies between features through a self-attention mechanism, and output global image features. In some implementation scenarios, before inputting medical image data into the prediction model, preprocessing operations such as rigid registration and alignment to the same coordinate system can be performed on T1, T2, and SWI medical image data to ensure that the lesion location is consistent across the three sequences. The fused data is then used as input to the feature extraction unit, improving the comprehensiveness of the features.

[0039] Further, in step S103, the genotype is embedded into the image features using an embedding unit to obtain embedded features. In some implementations, the embedding unit may include, for example, an embedding layer and a fully connected layer. The embedding layer can map the genotype into, for example, a 64-dimensional embedding vector, and the fully connected layer concatenates the image features and the genotype embedding vector to form the embedded features. By fusing the joint feature vector of image features and genotype information, a more comprehensive basis for risk prediction is included.

[0040] In some embodiments, the model may further include extracting texture features from medical image data and embedding these texture features into the aforementioned embedding features using an embedding unit. Texture features supplement the microstructural information of the image (e.g., the texture entropy of type I lesions is significantly higher than that of type III lesions), making the model more sensitive to early hemorrhage and further improving accuracy. In some embodiments, texture features may include, for example, mean, variance, skewness, contrast, correlation, or surface area, sphericity, etc. In implementation scenarios, new embedding features are obtained by Z-score normalization of the texture features, compression through a fully connected layer, and concatenation with the embedding vector and image features.

[0041] Based on the obtained embedding features, in step S104, the embedding features are input into the prediction unit to perform bleeding risk prediction and classification prediction, and a loss function is calculated based on the bleeding risk label and classification type. That is, the prediction unit includes a bleeding risk prediction head and a classification prediction head, which can respectively output bleeding risk prediction results and classification prediction results through a fully connected layer and a sigmoid activation function, and a fully connected layer and a softmax activation function.

[0042] In some embodiments, the loss function may include a total loss function calculated from a first loss function, a second loss function, and a third loss function. The first loss function is the regression loss between the hemorrhage risk prediction result (risk_pred) and the hemorrhage risk label (risk_label); the second loss function is the classification loss between the genotyping prediction result (Z_pred) and the genotyping label (Z_label); and the third loss function is the counterfactual regularization loss based on the impact of different genotypes on the risk of hemorrhage from cavernous malformations in the brain.

[0043] As an example, the first loss function is MSE(risk_pred, risk_label), the second loss function is CrossEntropy(Z_pred, Z_label), and the third loss function is Model(Conv(X, M), 2) - Model(Conv(X, M), 1), where Model(*) is the predicted risk value of the prediction model when the input genotype is 1 or 2, X and M are medical image data and bleeding mask, respectively, and 1 and 2 correspond to PIK3CA mutant and non-PIK3CA mutant, respectively. The total loss function Loss = MSE(risk_pred, risk_label) + CrossEntropy(Z_pred, Z_label) + Model(Conv(X, M), 2) - Model(Conv(X, M), 1).

[0044] Finally, in step S105, a prediction model is trained based on the loss function, enabling the prediction unit to output bleeding risk prediction results and classification prediction results. Specifically, based on the aforementioned total loss function Loss, the parameters in the model are updated through backpropagation to train the prediction model, enabling the prediction unit to output bleeding risk prediction results and classification prediction results. The three-loss synergistic optimization enables the model to simultaneously meet the three requirements of accurate prediction, reliable classification, and compliance with clinical priors. Furthermore, by introducing counterfactual loss, the stability of the model's risk prediction for both PIK3CA mutant and non-mutant types is significantly improved.

[0045] As described above, the embodiments of this application construct a prediction model that includes a feature extraction unit, an embedding unit, and a prediction unit. This model collects medical image data (including bleeding masks and classification types), genotypes, and bleeding risk labels that reflect continuous time decay characteristics. Image features are extracted and embedded with genotypes to obtain embedded features. Through dual-task prediction and training based on a multi-dimensional loss function, the collaborative prediction of bleeding risk and classification is achieved. This solves the problems of isolated risk prediction and classification in the prior art, and the failure to combine dynamic bleeding characteristics and the influence of genotypes, thus improving the comprehensiveness and accuracy of the prediction.

[0046] Furthermore, the embodiments of this application employ multiple loss functions: regression loss ensures risk prediction accuracy, classification loss improves genotyping reliability, and counterfactual regularization loss ensures compliance with the clinical prior knowledge of genotype's impact on risk, thereby enhancing model stability. Furthermore, by embedding texture features, microstructural information of lesions is supplemented, further improving prediction accuracy.

[0047] Figure 2 This is an exemplary flowchart illustrating the overall training of a predictive model for bleeding risk according to an embodiment of this application. Figure 2 As shown, in steps S201 and S202, medical imaging data and genotypes related to cavernous malformations in the brain are collected from the subject, respectively. The medical imaging data includes annotations of the hemorrhage mask and the Zabramski classification of the cavernous malformation. In some embodiments, the medical imaging data may be, for example, T1, T2, and SWI data. The genotype is either PIK3CA mutant or non-PIK3CA mutant.

[0048] Based on the acquired medical image data and the labeled bleeding mask, in step S203, bleeding risk labels reflecting continuous time decay characteristics are acquired. Specifically, based on the subject's follow-up bleeding records, it is determined whether bleeding occurred within a preset follow-up period, and the amount of bleeding is calculated based on the voxel size of the medical image data and the number of pixels in the bleeding area of ​​the bleeding mask, and corresponding bleeding risk labels are set. In some embodiments, in step S204, texture features can also be extracted based on the medical image data.

[0049] Next, in step S205, the feature extraction unit is used to extract features from the medical image data to obtain image features. In step S206, the embedding unit is used to embed genotype and texture features into the image features to obtain embedded features. Further, in step S207, the embedded features are input into the prediction unit to perform bleeding risk prediction and classification prediction, and in step S208, the first loss function, the second loss function, and the third loss function are calculated by combining the bleeding risk label and the classification type to obtain the total loss function.

[0050] Finally, a prediction model is trained based on the loss function, enabling the prediction unit to output bleeding risk prediction results and classification prediction results, thus obtaining the trained prediction model. For more details, please refer to the above. Figure 1 The description of the subject matter will not be repeated here.

[0051] In some embodiments, this application provides an apparatus for predicting bleeding risk, for example... Figure 3 As shown, Figure 3 This is an exemplary structural block diagram illustrating an apparatus 300 for predicting bleeding risk according to an embodiment of this application.

[0052] like Figure 3 As shown, the device 300 may include a processor 301 and a memory 302. The processor 301 may include, for example, a general-purpose processor (“CPU”) or a dedicated graphics processor (“GPU”), and the memory 302 stores program instructions executable on the processor. In some embodiments, the memory 302 may include, but is not limited to, resistive random access memory (RRAM), dynamic random access memory (DRAM), static random access memory (SRAM), and enhanced dynamic random access memory (EDRAM).

[0053] Furthermore, the aforementioned memory 302 may store program instructions for bleeding risk prediction. When the program instructions are executed by the processor, the device 300 performs the following operations: acquiring the medical image data and genotype of the subject, wherein the medical image data is marked with a bleeding mask; inputting the medical image data and genotype into a trained prediction model to perform bleeding risk prediction and subtyping prediction, and obtaining bleeding risk prediction results and subtyping prediction results.

[0054] Figure 4 This is an exemplary flowchart illustrating a method for predicting bleeding risk according to an embodiment of this application. Figure 4 As shown, in steps S401 and S402, the subject's medical imaging data and genotype are collected. The medical imaging data includes a bleeding mask.

[0055] Next, in step S403, the prediction model trained on the medical image data is processed by the feature extraction unit to obtain image features. In step S404, the genotype is embedded into the image features by the embedding unit to obtain embedded features. Further, in step S405, the embedded features are input into the prediction unit to perform bleeding risk prediction and typing prediction, so that the bleeding risk prediction results and typing prediction results are output in steps S406 and S407 respectively.

[0056] Figure 5 An exemplary structural block diagram of an electronic device 500 according to an embodiment of this application is shown. It will be understood that the device implementing the solution of this application may be a single device (e.g., a computing device) or a multifunctional device including various peripheral devices.

[0057] like Figure 5As shown, the electronic device of this application may include a central processing unit (“CPU”) 511, which may be a general-purpose CPU, a dedicated CPU, or other information processing and program execution unit. Further, the electronic device 500 may also include a mass storage device 512 and a read-only memory (“ROM”) 513. The mass storage device 512 may be configured to store various types of data, including various medical imaging data, genotypes and bleeding risk labels reflecting continuous-time decay characteristics, loss functions, algorithm data, intermediate results, and various programs required to run the electronic device 500. The ROM 513 may be configured to store power-on self-tests for the electronic device 500, initialization of various functional modules in the system, drivers for the system's basic input / output, and data and instructions required to boot the operating system.

[0058] Optionally, the electronic device 500 may also include other hardware platforms or components, such as the tensor processing unit (“TPU”) 514, graphics processing unit (“GPU”) 515, field-programmable gate array (“FPGA”) 516, and machine learning unit (“MLU”) 517 shown. It is understood that although various hardware platforms or components are shown in the electronic device 500, they are merely exemplary and not limiting, and those skilled in the art can add or remove appropriate hardware as needed. For example, the electronic device 500 may include only a CPU, associated storage devices, and interface devices to implement the operations performed by the method for training a predictive model for bleeding risk or the apparatus for predicting bleeding risk as described in this application.

[0059] In some embodiments, to facilitate data transmission and interaction with external networks, the electronic device 500 of this application further includes a communication interface 518, through which it can connect to a local area network / wireless local area network (“LAN / WLAN”) 505, and further through the LAN / WLAN to connect to a local server 506 or to the Internet (“Internet”) 507. Alternatively or additionally, the electronic device 500 of this application can also directly connect to the Internet or cellular network via the communication interface 518 based on wireless communication technology, such as wireless communication technology based on 3G (“3G”), 4G (“4G”), or 5G (“5G”). In some application scenarios, the electronic device 500 of this application can also access the server 508 and database 509 of an external network as needed to obtain various known algorithms, data, and modules, and can remotely store various data, such as various data or instructions for presenting, for example, medical image data, genotypes, and bleeding risk labels reflecting continuous time decay characteristics, loss functions, etc.

[0060] Peripherals of the electronic device 500 may include a display device 502, an input device 503, and a data transmission interface 504. In one embodiment, the display device 502 may include, for example, one or more speakers and / or one or more visual displays, configured to provide voice prompts and / or display images and videos for operations performed by the device for training a bleeding risk prediction model or for bleeding risk prediction. The input device 503 may include, for example, a keyboard, mouse, microphone, posture capture camera, or other input buttons or controls, configured to receive audio data input and / or user commands. The data transmission interface 504 may include, for example, a serial interface, parallel interface, or Universal Serial Bus interface (“USB”), Small Computer System Interface (“SCSI”), Serial ATA, FireWire (“FireWire”), PCI Express, and High Definition Multimedia Interface (“HDMI”), configured for data transmission and interaction with other devices or systems. According to the scheme of this application, the data transmission interface 504 can receive acquired medical image data from an MRI device and transmit medical image data or various other types of data or results to the electronic device 500.

[0061] The aforementioned CPU 511, mass storage 512, ROM 513, TPU 514, GPU 515, FPGA 516, MLU 517, and communication interface 518 of the electronic device 500 of this application can be interconnected via bus 519, and can interact with peripheral devices through this bus. In one embodiment, the CPU 511 can control other hardware components and peripheral devices in the electronic device 500 through this bus 519.

[0062] The above combination Figure 5 An electronic device that can be used to execute this application is described. It should be understood that the device structure or architecture described herein is merely exemplary, and the implementation method and implementation entity of this application are not limited thereto, but can be changed without departing from the spirit of this application.

[0063] Based on the foregoing description in conjunction with the accompanying drawings, those skilled in the art will understand that the embodiments of this application can also be implemented by software programs. Therefore, this application also provides a computer-readable storage medium storing computer-readable instructions for training a predictive model for bleeding risk or for predicting bleeding risk. When executed by one or more processors, these computer-readable instructions can be used to implement the embodiments of this application in conjunction with the accompanying drawings. Figure 1 The described method for training a predictive model for bleeding risk or Figure 3 The operations performed by the device used for bleeding risk prediction.

[0064] It should be noted that although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0065] It should be understood that when the terms "first," "second," "third," and "fourth," etc., are used in the claims, specification, and drawings of this application, they are used only to distinguish different objects and not to describe a specific order. The terms "comprising" and "including" as used in the specification and claims of this application indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof.

[0066] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0067] While numerous embodiments of this application have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will arise for those skilled in the art without departing from the spirit and intent of this application. It should be understood that various alternatives to the embodiments of this application described herein may be employed in the practice of this application. The appended claims are intended to define the scope of protection of this application and therefore cover equivalents or alternatives within the scope of these claims.

Claims

1. A method for training a prediction model for bleeding risk, wherein the prediction model includes a feature extraction unit, an embedding unit, and a prediction unit, and the method includes: Medical imaging data, genotype, and hemorrhage risk labels reflecting continuous time decay characteristics related to cavernous malformations in the brain were collected from the subjects. The medical imaging data included hemorrhage masks and classification types of cavernous malformations in the brain. Based on the medical image data, feature extraction is performed using the feature extraction unit to obtain image features; The genotype is embedded into the image feature using the embedding unit to obtain the embedded feature; The embedded features are input into the prediction unit to perform bleeding risk prediction and classification prediction, and a loss function is calculated based on the bleeding risk label and the classification type. The prediction model is trained based on the loss function, so that the prediction unit outputs the bleeding risk prediction result and the classification prediction result; The bleeding risk label, which reflects the continuous time decay characteristics, is collected through the following operations: Collect follow-up bleeding records of the subjects; If the follow-up bleeding record of the examinee shows no bleeding within the preset follow-up period, the bleeding risk label is set to 0. In response to the rebleeding recorded in the follow-up bleeding record of the subject occurring within the preset follow-up period, the bleeding risk label is set as (V1 / (V1 + V0) + (ΔT_max - (t1-t0)) / ΔT_max) / 2, where V1 represents the amount of rebleeding, V0 represents the baseline bleeding amount, ΔT_max represents the preset follow-up period, and (t1-t0) represents the bleeding time interval.

2. The method according to claim 1, wherein the medical imaging data includes T1-weighted imaging data, T2-weighted imaging data, and magnetic susceptibility-weighted imaging data; and the genotype includes PIK3CA mutant and non-PIK3CA mutant.

3. The method according to claim 1, wherein the classification type includes the Zabramski classification, and the Zabramski classification includes type I, type II or type III.

4. The method of claim 1, wherein the blood loss is calculated by the following operations: Calculate the voxel size of the medical image data and the number of pixels in the bleeding region of the bleeding mask; The bleed is calculated based on the product of the voxel size and the number of pixels.

5. The method according to claim 1, wherein the loss function includes a total loss function calculated by a first loss function, a second loss function, and a third loss function, wherein the first loss function is the regression loss between the hemorrhage risk prediction result and the hemorrhage risk label; the second loss function is the classification loss between the genotyping prediction result and the genotyping label; and the third loss function is the counterfactual regularization loss based on the influence of different genotypes on the risk of hemorrhage from cavernous malformations in the brain.

6. The method of claim 1, further comprising: Extract the texture features from the medical image data; The embedding unit is used to embed the texture feature into the embedding feature.

7. A device for predicting bleeding risk, comprising: processor; as well as A memory storing computer instructions for predicting bleeding risk, which, when executed by a processor, cause the following operations to be performed: Medical imaging data and genotype of the subjects were collected, wherein the medical imaging data was marked with a bleeding mask; The medical imaging data and the genotype are input into the prediction model trained by the method according to any one of claims 1-6 to perform bleeding risk prediction and genotyping prediction, and the bleeding risk prediction results and genotyping prediction results are obtained.

8. An electronic device, comprising: processor; as well as A memory having computer instructions stored thereon for training a predictive model for bleeding risk, which, when executed by a processor, cause the method according to any one of claims 1-6 to be performed; or having computer instructions stored thereon for predicting bleeding risk, which, when executed by a processor, cause the operation performed by the apparatus according to claim 7 to be performed.

9. A computer-readable storage medium storing computer program instructions for training a predictive model for bleeding risk, the computer program instructions, when executed by one or more processors, causing the method according to any one of claims 1-6 to be implemented; or storing computer program instructions for predicting bleeding risk, the computer program instructions, when executed by one or more processors, causing the operation performed by the apparatus according to claim 7 to be implemented.

Images