Method, device and storage medium for training a recognition model for recognizing ischemic stroke

Abstract

The application discloses a method, device and storage medium for training an identification model for identifying ischemic stroke. The method comprises: acquiring diffusion weighted imaging; preprocessing the diffusion weighted imaging to obtain preprocessed diffusion weighted imaging; inputting the preprocessed diffusion weighted imaging into a first model and a second model respectively, extracting corresponding image features via an encoder in the first model and the second model; performing a segmentation task and a scoring task via a decoder in the first model and the second model based on the corresponding image features, and calculating a supervised loss with labels and an unsupervised loss without labels; and determining a total loss according to the supervised loss and the unsupervised loss to train the identification model for identifying ischemic stroke. The application can improve the accuracy of the identification model in identifying ischemic stroke, reduce the difficulty of labeling, and improve the performance of the identification model in identifying ischemic stroke under the condition of scarce samples.

Description

TECHNICAL FIELD

[0001] The present application generally relates to the technical field of ischemic stroke recognition. More particularly, the present application relates to a method, device and computer readable storage medium for training a recognition model for recognizing ischemic stroke. BACKGROUND

[0002] Stroke is a common cerebrovascular disease with high morbidity, mortality and disability rate. Acute ischemic stroke (AIS) accounts for 60-80% of stroke, and its cause is due to the obstruction of blood supply to brain tissue. In the assessment of the severity and prognosis of ischemic stroke, neuroimaging examination plays an important role in guiding clinical treatment. Within a few hours of acute ischemic stroke, the water content changes caused by local brain tissue ischemia will show high signal lesions on diffusion weighted imaging (DWI), and early ischemic changes usually reflect cytotoxic edema or irreversible damage.

[0003] In order to accurately evaluate early ischemic changes, the current common way is to assist in recognition and measurement by lesion segmentation and Alberta Stroke Program Early CT Score (ASPECTS). For example, by extracting the target region from the image, a cerebral ischemia classification model is trained based on machine learning (or deep learning) to predict the ASPECTS score; in addition, there is also a method of segmenting the lesion by machine learning (or deep learning), and obtaining the ASPECTS score according to the overlap of the lesion region and the ASPECTS region. However, the former needs to realize the extraction of the target brain area, the preprocessing is complex and cannot fully utilize the end-to-end characteristics of the deep learning model, and only the ASPECTS score can be obtained, which cannot provide accurate positioning of the stroke lesion, affecting the recognition accuracy. The effectiveness of the latter may be affected by the errors introduced in the early ischemic segmentation and ASPECTS region registration process, resulting in inaccurate recognition results. In addition, the performance of the existing deep learning model depends on a large number of segmentation annotations and ASPECTS score annotations, and there are many pseudo lesions in the DWI image that mimic the intensity and shape of stroke lesions, making it difficult to segment and score the lesions, which severely limits the performance of the automated ASPECTS scoring method.

[0004] Therefore, there is an urgent need to provide a scheme for training a recognition model for recognizing ischemic stroke, so as to improve the accuracy of the recognition model in recognizing ischemic stroke, reduce the difficulty of annotation, and improve the performance of the recognition model in recognizing ischemic stroke under the condition of scarce samples. SUMMARY

[0005] To at least solve one or more technical problems as mentioned above, the present application proposes, in multiple aspects, a scheme for training a recognition model for recognizing ischemic stroke.

[0006] In a first aspect, the present application provides a method for training a recognition model for recognizing ischemic stroke, wherein the recognition model comprises a first model and a second model formed based on an exponential moving average of weights in the first model, and the first model and the second model both comprise an encoder and a decoder, and the method comprises: obtaining diffusion weighted imaging, wherein the diffusion weighted imaging contains labeled and unlabeled; pre-processing the diffusion weighted imaging to obtain pre-processed diffusion weighted imaging; inputting the pre-processed diffusion weighted imaging into the first model and the second model respectively, extracting corresponding image features via the encoders in the first model and the second model respectively; performing a segmentation task and a scoring task via the decoders in the first model and the second model respectively based on the corresponding image features, and calculating a supervised loss under labeling and an unsupervised loss under unlabeled; and determining a total loss according to the supervised loss and the unsupervised loss to train the recognition model for recognizing ischemic stroke.

[0007] In some embodiments, wherein the labeled at least includes lesion annotation and ASPECT score annotation.

[0008] In some embodiments, wherein the pre-processing at least includes one or more of skull stripping, image registration or normalization operation.

[0009] In some embodiments, wherein the encoders in the first model and the second model both comprise a plurality of first convolutional layers of different scales, and each of the convolutional layers is connected with a batch normalization layer and a first ReLU activation layer, wherein the pre-processed diffusion weighted imaging is down-sampled via the first convolutional layers, the batch normalization layers and the first ReLU activation layers of different scales of the encoders in the first model and the second model respectively to extract the corresponding image features.

[0010] In some embodiments, wherein in performing the segmentation task, the decoders in the first model and the second model at least comprise a plurality of deconvolutional layers, a second convolutional layer and a second ReLU activation layer, wherein the corresponding image features are up-sampled via the deconvolutional layers of the decoders in the first model and the second model respectively to obtain up-sampled features; the up-sampled features are spliced with image features of the corresponding first convolutional layers of the encoders in the first model and the second model, and the splicing result is subjected to the segmentation task via the second convolutional layer and the second ReLU activation layer to obtain a segmentation result of a lesion.

[0011] In some embodiments, wherein in performing the scoring task, the decoder in the first model and the second model comprises a global average pooling layer, a fully connected layer and a Sigmoid activation layer, and wherein the corresponding image features are respectively classified via the global average pooling layer, the fully connected layer and the Sigmoid activation layer of the decoder in the first model and the second model to obtain a classification result of the target scoring labeled region.

[0012] In some embodiments, wherein the supervised loss is calculated by: calculating a segmentation loss and a classification loss corresponding to the segmentation task and the scoring task respectively under the labeled condition; and calculating the supervised loss according to a weighted sum of the segmentation loss and the classification loss.

[0013] In some embodiments, wherein the unsupervised loss is calculated by: calculating a segmentation consistency loss and a classification consistency loss for the segmentation task and the scoring task respectively under the unlabeled condition; and calculating the unsupervised loss according to a weighted sum of the segmentation consistency loss and the classification consistency loss.

[0014] In a second aspect, the present application provides a method for identifying ischemic stroke, comprising: obtaining diffusion weighted imaging; and inputting the diffusion weighted imaging into the identification model trained according to one or more embodiments of the first aspect, to obtain an identification result of ischemic stroke via the first model in the identification model.

[0015] In a third aspect, the present application provides a device, comprising: a processor; and a memory, wherein the memory stores program instructions for training an identification model for identifying ischemic stroke, which, when executed by the processor, causes the device to implement one or more embodiments of the first aspect; or the memory stores program instructions for identifying ischemic stroke, which, when executed by the processor, causes the device to implement an embodiment of the second aspect.

[0016] In a fourth aspect, the present application provides a computer readable storage medium, which stores computer readable instructions for training an identification model for identifying ischemic stroke, which, when executed by one or more processors, implements one or more embodiments of the first aspect; or the computer readable storage medium stores computer readable instructions for identifying ischemic stroke, which, when executed by one or more processors, implements an embodiment of the second aspect.

[0017] By the scheme for training the recognition model for recognizing ischemic stroke as provided above, the embodiments of the present application train the recognition model by constructing the recognition model with the semi-supervised average teacher theoretical framework containing the first model and the second model, and simultaneously performing the segmentation task and the scoring task in the training, and calculating the total loss corresponding to the supervised loss and the unsupervised loss to train the recognition model. Based on this, the embodiments of the present application can reduce the need to determine the optimal threshold, obtain no additional ASPECTS score, and utilize shared information through the correlation task, so that the segmentation task and the scoring task can assist and enhance each other, improve the accuracy of the recognition model, and thus improve the accuracy of the recognition model in recognizing ischemic stroke. Further, the embodiments of the present application can be trained under a small amount of labeled data and a large amount of unlabeled data, reduce the difficulty of labeling, and improve the performance of the recognition model in recognizing ischemic stroke under the condition of scarce samples. BRIEF DESCRIPTION OF DRAWINGS

[0018] The above and other objects, features and advantages of the present exemplary embodiments will become more apparent from the following detailed description read in conjunction with the accompanying drawings, in which like reference numerals refer to like elements throughout. The drawings illustrate several embodiments of the present application and, as such, a further understanding of the present application can be obtained by reference to a

[0019] Figure 1 A target region in a middle cerebral artery supply area is shown;

[0020] Figure 2 is an exemplary flow block diagram illustrating a method for training a recognition model for recognizing ischemic stroke according to an embodiment of the present application;

[0021] Figure 3 is an exemplary schematic diagram illustrating the whole of a method for training a recognition model for recognizing ischemic stroke according to an embodiment of the present application;

[0022] Figure 4 is an exemplary flow block diagram illustrating a method for recognizing ischemic stroke according to an embodiment of the present application;

[0023] Figure 5 is an exemplary schematic diagram illustrating a stroke segmentation result according to an embodiment of the present application;

[0024] Figure 6 is an exemplary structural block diagram illustrating an apparatus according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] With reference to the drawings and the embodiments described herein, it will be understood that the embodiments described are intended in all respects to be illustrative rather than restrictive. Numerous modifications and adaptations thereof will be apparent to those skilled in the art without departing from the scope of the present application as set forth in the following claims.

[0026] It will be understood that the terms "comprises" and "comprising", when used in this specification and claims, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0027] It will be understood that the terms "comprises" and "comprising", when used in this specification and claims, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0028] As used in this specification and claims, the terms "if" and "when" can be interpreted to mean "upon" or "in response to determining," or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [a described condition or event] is detected" can be interpreted to mean "upon determining" or "in response to determining," or "upon detecting [a described condition or event]" or "in response to detecting [a described condition or event]," depending on the context.

[0029] As known from the background art description above, target identification is a common way to identify and measure ischemic stroke, which is assisted by lesion segmentation and ASPECTS scoring. Among them, lesion segmentation can help doctors accurately locate the damaged area of stroke in clinic, provide more detailed lesion information (such as the size, shape and location of the lesion, etc.), and thus develop more personalized treatment plans. Lesion segmentation is a method for evaluating early ischemic changes in the middle cerebral artery supply area of ischemic stroke patients, which can be performed by selecting 10 regions in the middle cerebral artery supply area on the image (such as Figure 1For example, the 10 regions are divided into 7 regions at the level of the nuclei (i.e., the thalamic and striatal planes), including M1, M2, M3, insular I, lenticular L, caudate C, and internal capsule posterior limb IC, and 3 regions above the level of the nuclei (2 cm above the level of the nuclei), including M4, M5, and M6. The aforementioned 10 regions have the same weight, and a full score is 10 points. The score is deducted according to whether there is an identifiable ischemic region, so as to evaluate early ischemic changes.

[0030] Based on the target brain regions (for example, the aforementioned 10 regions) extracted on the image data, the ischemic stroke image data marked by judging whether ischemia exists can be used to train a brain ischemia classification model based on a machine learning model or a deep learning model. However, this method needs to realize the extraction of the target brain region, and the preprocessing is complex and cannot fully utilize the end-to-end characteristics of the deep learning model, which limits the clinical practice, and this method can only obtain the ASPECTS score and cannot provide accurate positioning of the stroke lesion, affecting the recognition accuracy.

[0031] In addition, the commonly used method can also perform lesion segmentation on the stroke image by using a segmentation network first, and then generate a regional ASPECTS prediction according to the overlap of the ASPECTS region and the lesion region by applying a specific involvement threshold (for example, there is any lesion or the lesion covers more than 5% of the ASPECTS region). This method can provide scores for all ASPECT regions and provide more detailed location information, but its effectiveness can be affected by errors introduced in the early ischemic segmentation and ASPECTS region registration processes, resulting in inaccurate recognition results.

[0032] In addition, the performance of the existing deep learning model depends on a large number of segmentation annotations and ASPECTS score labels, and there are many artifacts in the DWI image that mimic the intensity and shape of the stroke lesion. It can be understood that there are many subtypes of stroke, resulting in a large difference in lesion size and location. The stroke lesion volume can range from hundreds to tens of thousands of cubic millimeters and can occur in any brain region such as the cerebral lobe, cerebellum, and brainstem. Therefore, even for experienced radiologists, accurately delineating detailed lesion regions is still a challenging task. In addition to lesion annotation, the complexity of some acute stroke lesions requires neuro-radiologists to spend a lot of time and effort in performing ASPECTS region scoring by visual inspection, and at the same time, there may be a large subjective bias. In summary, due to the doubled difficulty of obtaining these two labels, this seriously limits the performance of the automated ASPECTS scoring method, and there may be problems in reliability in more diversified actual clinical scenarios.

[0033] Based on this, the embodiment of the present application provides a scheme for training a recognition model for recognizing ischemic stroke, which can realize joint training of the lesion segmentation task and the scoring task while using a small amount of labeled data and a large amount of unlabeled data, thereby obtaining richer feature representation of ischemic stroke lesions, improving the accuracy of the recognition model in recognizing ischemic stroke, and improving the performance of the segmentation task and the scoring task in recognizing ischemic stroke, and improving the accuracy of the recognition model.

[0034] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0035] Figure 2 is an exemplary flow block diagram illustrating a method 200 for training a recognition model for recognizing ischemic stroke according to an embodiment of the present application. In some embodiments, the recognition model includes a first model and a second model formed based on an exponential moving average of weights in the first model. The aforementioned first model can correspond to a student model, and the second model can correspond to a teacher model. That is, in the training process, the exponential moving average of the weights of the student model is assigned to the teacher model, so that the recognition model has a semi-supervised average teacher theoretical framework. In some embodiments, the first model and the second model both include an encoder and a decoder, and the recognition of the second model can be regarded as additional supervision of the first model. In some implementation scenarios, the aforementioned encoder can be shared between different tasks, while the decoder is specific to the segmentation task or the scoring task.

[0036] As shown in Figure 2 At step S201, a diffusion weighted imaging is acquired, wherein the diffusion weighted imaging contains labeled and unlabeled. It can be understood that the acquisition of medical image data covers cerebral infarction lesions in multiple parts of the anterior circulation region, which can present a variety of different types and characteristics. DWI imaging is very sensitive to the detection of acute cerebral infarction, and can show abnormal signals in a very short time after ischemia occurs, which is earlier than traditional CT scanning. Therefore, acquiring DWI images of patients with ischemic stroke helps to accurately identify and analyze the characteristics and ASPECTS scores of cerebral infarction lesions, thereby providing an important basis for clinical diagnosis and treatment.

[0037] In some embodiments, the labeled at least includes lesion annotation and ASPECT score annotation. For lesion annotation, in order to ensure the accuracy and consistency of diagnosis, image processing software (such as ITK-SNAP) can be used to view DWI images layer by layer, and the patient's brain image is analyzed in detail, so as to complete the accurate delineation and annotation of the lesion of cerebral infarction, so as to obtain the lesion annotation.

[0038] For ASPECT score annotation, first, the brain images of patients can be interpreted by two experienced doctors independently according to clear scoring criteria. Specifically, according to the conditions of 10 target regions including the caudate head, the putamen, the posterior limb of the internal capsule, the insular ribbon and the M1-M6 regions, 1 or 0 is used for preliminary scoring. According to the preliminary score, a third senior doctor will review these scores and make necessary integration and adjustment. If there are differences in the scoring results or disputes in the assessment of some regions during this process, the final score will be determined by a three-person vote to complete the ASPECT score annotation. This three-verification process ensures the accuracy and reliability of the score and reduces the errors that may be caused by subjective judgment.

[0039] Based on the acquired diffusion weighted imaging, at step S202, the diffusion weighted imaging is preprocessed to obtain preprocessed diffusion weighted imaging. In some embodiments, the foregoing preprocessing can include, but is not limited to, one or more of skull stripping, image registration or normalization operation. For example, it can also include image cropping, image denoising and other preprocessing.

[0040] In some implementation scenarios, the pixel points corresponding to the outer skull of the brain can be determined from the DWI image according to a preset determination manner about the pixel points corresponding to the outer skull of the brain. Then, the grayscale of the pixel points corresponding to the outer skull of the brain is set to 0 to realize skull stripping. In other implementation scenarios, the original DWI image can be registered to the internationally recognized MNI152 standard space template by, for example, an affine registration algorithm to realize image registration. The foregoing affine registration algorithm can also process translation, rotation, scaling and other transformations of the image to ensure the spatial consistency of the DWI image. In yet other embodiments, the DWI image can be normalized by, for example, a z-score method. In addition, the DWI image can also be cropped to the same size (for example, 128*128*128) by image cropping.

[0041] As an example, assume that the dataset of preprocessed diffusion weighted imaging contains N labeled datasets and M unlabeled datasets, and the labeled dataset is defined as The unlabeled dataset is defined as N << M. Wherein, represents the size of the input image, H, W and D represent the height, width and depth of the input image, respectively; i ∈{0,1} H×WxD represents the corresponding true lesion annotation (or label). Y i ∈{0,1} 10The ASPECTS score label is represented, and the number 10 represents that the ASPECTS scoring mechanism divides the middle cerebral artery blood supply area into 10 regions, in which 0 and 1 represent normal regions and ischemic infarction respectively.

[0042] Then, at step S203, the pre-processed diffusion weighted imaging is input into the first model and the second model respectively, and corresponding image features are extracted via the encoders in the first model and the second model. For example, 50 and 150 examples of data can be selected from the data set of the pre-processed diffusion weighted imaging as labeled samples and unlabeled data samples respectively, and different proportions of the data are input into the first model and the second model, and the corresponding image features are extracted via the encoders in the first model and the second model.

[0043] In some embodiments, the encoders in the first model and the second model can each include a plurality of first convolutional layers of different scales, and a batch normalization layer and a first ReLU activation layer are connected after each convolutional layer. The pre-processed diffusion weighted imaging is down-sampled via the plurality of first convolutional layers of different scales, the batch normalization layers and the first ReLU activation layers of the encoders in the first model and the second model respectively, to extract corresponding image features. That is, the main task of the encoders in the first model and the second model is to extract the features of the image, and gradually reduce the resolution of the feature map, so as to capture semantic information at different levels.

[0044] As an example, the encoders in the first model and the second model can include four first convolutional layers, each layer can have convolutional blocks of different feature scales, and each layer can include two convolutional layers with a convolution kernel size of 3x3x3. Further, a batch normalization layer and a ReLU activation layer can be connected after each convolutional layer. In an implementation scenario, the input image is sequentially down-sampled via the first convolutional layers, the batch normalization layers and the ReLU activation layers, and the image features are output.

[0045] Further, at step S204, based on the corresponding image features, the decoders in the first model and the second model perform segmentation tasks and scoring tasks respectively, and calculate supervised losses with labels and unsupervised losses without labels. In some embodiments, in performing the segmentation task, the decoders in the first model and the second model at least include a plurality of deconvolutional layers, a second convolutional layer and a second ReLU activation layer, wherein the corresponding image features are up-sampled via the plurality of deconvolutional layers in the decoders of the first model and the second model to obtain up-sampled features; the up-sampled features are spliced with the image features of the corresponding first convolutional layers of the encoders in the first model and the second model, and the spliced results are subjected to the segmentation task via the second convolutional layer and the second ReLU activation layer to obtain the segmentation results of the lesions.

[0046] In other embodiments, during the scoring task, the decoders in the first and second models include a global average pooling layer, a fully connected layer, and a sigmoid activation layer, wherein the corresponding image features are classified through the global average pooling layer, fully connected layer, and sigmoid activation layer of the decoders in the first and second models, respectively, to obtain the classification result of the target scoring label region.

[0047] In other words, in the segmentation and scoring tasks, the decoder corresponds to the lesion segmentation module and the ASPECT scoring module in the recognition model, respectively. The main task of the lesion segmentation module is to restore the spatial resolution of the image (i.e., upsampling), gradually reconstructing detailed information, and ultimately generating a segmentation map with the same size as the input image. In some implementation scenarios, upsampling can be performed first through deconvolutional layers. For example, an upsampling operation with a 3×3×3 kernel and a stride of 2 can double the size of the image features. Then, the upsampled feature map is concatenated with the feature map of the corresponding layer in the encoder to obtain the concatenated result. This allows for the fusion of features at different resolutions, which helps to better restore spatial information. Furthermore, the aforementioned concatenated result can be used for segmentation tasks, for example, through two second convolutional layers and one second ReLU activation layer, until the input feature map size is restored, obtaining the classification result of the target scoring labeled region.

[0048] For the ASPECT scoring module, in the implementation scenario, 10 pairs of ASPECTS regions can be extracted based on the ASPECTS annotations. It can be understood that the ASPECTS region labels in the image can be obtained when registering the DWI image to an Atlas image. Based on these ASPECTS region labels, 10 pairs of regions from the affected and unaffected hemispheres of the 10 ASPECTS regions are extracted from the image features obtained by the encoder. For each region, the dimensionality of the features can be reduced via a global average pooling layer, and the features in each region block can be combined by applying concatenation to combine them into a fully connected layer. Finally, the response of each ASPECTS region is activated via a Sigmoid activation layer to obtain the classification result (e.g., the classification probability of each region) of the target scoring labeled region. In some implementation scenarios, the ASPECTS region labels are... k-th The classification probability (P) of the ASPECTS region k This can be represented as follows:

[0049]

[0050] Where k = 1, 2, ..., 10, σ represents the sigmoid function, and R k Indicates the first extracted from. k-th In the ASPECTS region, FAP represents the global average pooling operation, and W represents the fully connected layer parameters of the k-th ASPECTS region.

[0051] According to the foregoing, the identification model of the embodiments of the present application is constructed based on the semi-supervised average teacher theoretical framework, and is trained by using a small amount of labeled DWI images and a large amount of unlabeled DWI images. The corresponding loss function can include a supervised loss and an unsupervised consistency loss.

[0052] In some embodiments, the supervised loss can be calculated by calculating the segmentation loss and the classification loss corresponding to the segmentation task and the scoring task respectively under the labeled condition, and calculating the supervised loss according to the weighted sum of the segmentation loss and the classification loss. In some embodiments, the unsupervised loss can be calculated by calculating the segmentation consistency loss and the classification consistency loss of the first model and the second model for the segmentation task and the scoring task respectively under the unlabeled condition, and calculating the unsupervised loss according to the weighted sum of the segmentation consistency loss and the classification consistency loss.

[0053] Specifically, in one implementation scenario, the above-mentioned supervised loss can be represented by the following formula:

[0054]

[0055] wherein, the above-mentioned classification loss is represented by, the above-mentioned segmentation loss is represented by, and β represents a weighting coefficient. In some implementation scenarios, the above-mentioned classification loss is a binary cross-entropy loss, which is used for binary classification of each ASPECTS region, and is specifically represented by N represents the number of training samples, and respectively represent the label and the affected probability of the kth ASPECTS region of the ith sample. In other implementation scenarios, the segmentation loss can be a Dice loss, which is used to measure the overlap between the model prediction result and the true label, and is specifically represented by P represents the segmentation region predicted by the model, and G represents the true segmentation region.

[0056] It can be understood that, since the second model is formed based on the exponentially moving average of the weights in the first model, the outputs of the first model and the second model for segmentation and classification remain consistent for the same image input, and the corresponding unsupervised loss has consistency. Specifically, in one implementation scenario, the above-mentioned unsupervised loss can be represented by the following formula:

[0057]

[0058] wherein, the classification consistency loss is represented by, Let represent the segmentation consistency loss, and γ be the weighting coefficient. Specifically, let γ represent the classification consistency loss. Represented as This represents the prediction result of the first model for the score. This represents the second model's prediction of the score. (Segment consistency loss) Represented as This represents the prediction result of the first model for segmentation. This represents the prediction result of the second model for segmentation.

[0059] After obtaining the supervised loss and unsupervised loss, in step S205, a total loss is determined based on the supervised loss and unsupervised loss to train the identification model for ischemic stroke. In some embodiments, the aforementioned total loss... Supervised loss and unsupervised losses The weighted sum, as shown in the following example:

[0060]

[0061] In this scenario, the total loss is calculated. To update the parameters of the recognition model, the learning rate is adjusted using a linear decay strategy during training until the maximum number of training iterations or the total loss is reached. The process continues until convergence, thus obtaining a well-trained recognition model.

[0062] As described above, this embodiment constructs a recognition model with a semi-supervised average teacher theoretical framework, incorporating a first and a second model. During training, it simultaneously performs segmentation and scoring tasks, calculating the total loss corresponding to supervised and unsupervised losses to train the recognition model. Based on this, this embodiment reduces the need to determine the optimal threshold, obtains unattached ASPECTS scores, and utilizes shared information through association tasks, allowing the segmentation and scoring tasks to mutually reinforce each other, improving the accuracy of the recognition model and thus enhancing its ability to identify ischemic stroke. Furthermore, this embodiment can be trained on both small amounts of labeled data and large amounts of unlabeled data, reducing annotation difficulty and improving the performance of the recognition model in identifying ischemic stroke under conditions of scarce samples.

[0063] Figure 3 This is an exemplary schematic diagram illustrating the overall structure of a recognition model for identifying ischemic stroke according to an embodiment of this application. It should be understood that... Figure 3 The above Figure 2 A specific embodiment of method 200, therefore the above regarding Figure 2 The description also applies toFigure 3 .

[0064] like Figure 3 As shown, the recognition model in this embodiment may include a first model (corresponding to a student model) 301 and a second model (corresponding to a teacher model) 302, wherein the second model is formed based on the exponential moving average of the weights in the first model. That is, the exponential moving average (“EMA”) of the weights of the student model is assigned to the teacher model. In some embodiments, the first model 301 and the second model 302 include an encoder and a decoder, wherein the decoder corresponds to the lesion segmentation module and the ASPECT scoring module in the recognition model in the segmentation task and the scoring task, respectively. As an example, the first model 301 includes an encoder 301-1 and a decoder, and the decoder corresponds to the lesion segmentation module 301-21 and the ASPECT scoring module 301-22, respectively. The second model 302 includes an encoder 302-1 and a decoder, and the decoder corresponds to the lesion segmentation module 302-21 and the ASPECT scoring module 302-22, respectively.

[0065] In the implementation scenario, encoders 301-1 and 302-1 can each include multiple first convolutional layers, batch normalization layers, and first ReLU activation layers of different scales. Lesion segmentation modules 301-21 and 302-21 can each include multiple deconvolutional layers, second convolutional layers, and second ReLU activation layers; ASPECT scoring modules 301-22 and ASPECT scoring modules 302-22 can each include a global average pooling layer, a fully connected layer, and a Sigmoid activation layer. In this scenario, the acquired diffusion-weighted images (including labeled 303 and unlabeled 304) are first processed by encoders 301-1 of the first model 301 and encoders 302-1 of the second model 302 to extract their respective image features 305-1 and 305-2. Specifically, the input image is downsampled sequentially through the first convolutional layers, batch normalization layers, and ReLU activation layers in the encoders before the output image features are generated.

[0066] Based on the image features 305-1 and 305-2, the classification task is performed via the respective ASPECT score module 301-22 and ASPECT score module 302-22 of the first model 301 and the second model 302 respectively to obtain the respective corresponding classification results 306-1 and 306-2 of the target score labeling region. Further, the segmentation task is performed via the respective lesion segmentation module 301-21 and lesion segmentation module 302-21 of the first model 301 and the second model 302 respectively to obtain the respective corresponding segmentation results 307-1 and 307-2 of the lesion. Specifically, in the classification task, the classification result is obtained via the global average pooling layer, the fully connected layer and the Sigmoid activation layer in the ASPECT score module. In the segmentation task, the segmentation result is obtained via upsampling, concatenation and convolution in the lesion segmentation module.

[0067] Further, the respective segmentation loss 308 and classification loss 309 corresponding to the labeled segmentation task and the scoring task are calculated respectively, and the weighted sum between the segmentation loss and the classification loss is calculated based on the above formula (2) to obtain the supervised loss. In addition, the segmentation consistency loss 310 and the classification consistency loss 311 are also calculated respectively, and then the weighted sum of the segmentation consistency loss and the classification consistency loss is calculated based on the above formula (3) to obtain the unsupervised loss. Finally, the total loss determined by the supervised loss and the unsupervised loss is used to train the recognition model for identifying ischemic stroke. More details about the calculation of the aforementioned losses can be referred to the description of the above formula (2) and formula (3), which will not be repeated here. Figure 2

[0068] In an exemplary scenario, the recognition model can be trained using, for example, Adam optimizer, wherein the above weighting coefficients can be set as a = 0.9, β = 0.8, γ = 0.9. Under the condition that the total training period is 300 epochs and the initial learning rate is 0.0001, the total loss is calculated to update the model parameters, and the learning rate is adjusted using, for example, a linear decay strategy during the training process. That is, the learning rate is decayed once every 30 cycles, and the decay coefficient is 0.5, until the maximum number of training cycles or the total loss function converges. In addition, the trained recognition model can also be obtained by setting the proportion of labeled DWI images to be 10%, 20% and 50% for training and testing. Based on the recognition model, the stroke DWI medical image data can be processed, and accurate stroke lesion segmentation and ASPECTS scoring can be obtained for clinical auxiliary decision-making.

[0069] Figure 4 is an exemplary flow chart showing a method 400 for identifying ischemic stroke according to an embodiment of the present application. As Figure 4 ​As shown in FIG. 4, at step S401, the diffusion weighted imaging is acquired. Then, at step S402, the diffusion weighted imaging is input into the trained recognition model, and the recognition result of ischemic stroke is obtained via the first model in the recognition model. More details about the training of the recognition model can be referred to the description of the above Figures 2-3 , which will not be repeated here. It can be understood that after the training of the recognition model is completed, the inference process can only need to input the diffusion weighted imaging into the first model for recognition to obtain the recognition result of ischemic stroke, which can include the lesion segmentation result and the score result at the same time.

[0070] In some embodiments, in order to ensure the performance of the recognition model, the results of the stroke lesion segmentation and the ASPECTS classification can be evaluated by the following evaluation indexes. For example, for the stroke segmentation, the Dice coefficient is used to measure the similarity of two samples:

[0071]

[0072] where P and G represent the predicted and true segmentation regions, respectively.

[0073] The following evaluation indexes are used for the ASPECTS score. For example, the accuracy is used to calculate the proportion of the number of samples correctly predicted by the model to the total number of samples.

[0074]

[0075] where TP is the true positive, TN is the true negative, FP is the false positive, and FN is the false negative rate.

[0076] For another example, the F1 score is the harmonic mean of the precision and the recall.

[0077]

[0078]

[0079] where TP represents the true positive, FP represents the false positive, and FN represents the false negative rate.

[0080] In addition, the Kappa coefficient is a statistical index for measuring the classification accuracy, which considers the probability of accidental agreement between classifiers. Specifically, the calculation formula of the Kappa coefficient is as follows:

[0081]

[0082] where P0 represents the observed consistency ratio, i.e., the proportion of actual agreement between the raters. P eThe expected proportion of chance agreement, i.e. the proportion of agreement between raters that would be expected in the absence of any actual agreement.

[0083] Figure 5 is an exemplary schematic diagram showing a stroke segmentation result according to an embodiment of the present application. As shown in Figure 5 Fig. (a) of shows a large area of lesion region segmented by the recognition model based on the embodiments of the present application, which can accurately capture its boundary. This shows that the recognition model of the embodiments of the present application has excellent recognition ability when dealing with large volume lesions. Figure 5 Figs. (b) and (d) show that multiple small lesions are segmented by the recognition model based on the embodiments of the present application. This shows that the recognition model of the embodiments of the present application still maintains high accuracy when dealing with scattered small lesions, and can effectively distinguish lesions from normal tissues. Figure 5 Fig. (c) shows the ability of the recognition model of the embodiments of the present application in segmenting irregularly shaped lesions, which shows that the recognition model of the embodiments of the present application can adapt to lesions of different shapes, and ensure the integrity and accuracy of segmentation. As can be seen from the figure, the recognition model of the embodiments of the present application shows superior segmentation performance when dealing with lesions of different characteristics, and can provide reliable support for clinical diagnosis.

[0084] Meanwhile, the recognition model of the embodiments of the present application can also get ASPETCS score results, for example, as shown in Table 1 below.

[0085] Table 1 Performance comparison of recognition model and other methods

[0086]

[0087] Among them, other methods include the method of extracting brain regions and then classifying by Wu et al.; the method of segmenting and then getting score results based on threshold by Naganuma et al.; the method of combining segmentation and ASPECTS score classification for joint training by Baseline; the existing semi-supervised training method Self-training and the training method of the recognition model of the embodiments of the present application.

[0088] As can be seen from Table 1, compared with the first three methods, the embodiment of the application trains by adopting segmentation and ASPECTS scoring in combination, different tasks can share the underlying feature representation, which can help the model to learn more abundant and fine features of ischemic stroke lesions, and thus improve the performance of the model. Compared with the existing semi-supervised learning method, the innovative method proposed in the application can simultaneously utilize limited labeled data and a large amount of unlabeled data for training. The embodiment of the application not only effectively solves the problem of obtaining labeled data in lesion segmentation and ASPECTS scoring, but also significantly improves the accuracy of cerebral stroke lesion segmentation and optimizes the accuracy of ASPECT scoring by deeply mining the potential information in the unlabeled data, thereby providing more reliable technical support for clinical diagnosis and treatment.

[0089] Figure 6 is an exemplary structural block diagram illustrating the device 600 according to an embodiment of the application. As shown in Figure 6 the device 600 of the application can include a processor 601 and a memory 602, wherein the processor 601 and the memory 602 communicate with each other through a bus. The memory 602 stores program instructions for training a recognition model for recognizing ischemic stroke, and when the program instructions are executed by the processor 601, the method steps described in the foregoing in conjunction with the accompanying Figure 2 description are implemented: obtaining diffusion weighted imaging, wherein the diffusion weighted imaging contains labeled and unlabeled; pre-processing the diffusion weighted imaging to obtain pre-processed diffusion weighted imaging; inputting the pre-processed diffusion weighted imaging into the first model and the second model respectively, extracting corresponding image features via the encoders in the first model and the second model respectively; performing segmentation task and scoring task via the decoders in the first model and the second model respectively based on the corresponding image features, and calculating supervised loss under labeling and unsupervised loss under unlabeled; and determining total loss according to the supervised loss and the unsupervised loss to train the recognition model for recognizing ischemic stroke.

[0090] In some embodiments, the memory 602 stores program instructions for recognizing ischemic stroke, and when the program instructions are executed by the processor 601, the method steps described in the foregoing in conjunction with the accompanying Figure 4 description are implemented: obtaining diffusion weighted imaging; and inputting the diffusion weighted imaging into the trained recognition model, and obtaining a recognition result of ischemic stroke via the first model in the recognition model.

[0091] According to the above description in combination with the drawings, those skilled in the art can also understand that the embodiments of the present application can also be implemented by a software program. Therefore, the present application also provides a computer readable storage medium. The computer readable storage medium has stored thereon computer readable instructions for training a recognition model for recognizing ischemic stroke or for recognizing ischemic stroke, and the computer readable instructions are executed by one or more processors to implement the embodiments of the present application described in combination with the drawings. Figure 2 Or Figure 4 The method for training a recognition model for recognizing ischemic stroke or the method for recognizing ischemic stroke is described.

[0092] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be implemented by means of software and the necessary general hardware platform, and of course can also be implemented by hardware. Based on such understanding, the above technical solutions essentially or in other words the part of the prior art that contributes to the present application can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, server, or network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0093] It should be noted that although the operations of the method of the present application are described in a specific order in the drawings, this does not require or imply that the operations must be performed in that specific order, or that all of the shown operations must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can change the order of execution. Additionally or alternatively, some steps can be omitted, a plurality of steps can be combined into one step, and / or one step can be divided into a plurality of steps.

[0094] It should be understood that when the terms "first", "second", "third" and "fourth" are used in the claims, the description and the drawings of the present application, they are only used to distinguish different objects, and are not used to describe a specific order. The terms "include" and "contain" used in the specification and claims of the present application indicate the presence of the described features, whole, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, whole, steps, operations, elements, components and / or sets thereof.

[0095] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. As used in this specification and the claims, the singular forms "a," "an" and "the" include plural referents unless the context clearly dictates otherwise. It is further to be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items, as well as the lack of combinations when interpreted in the alternative.

[0096] Although the embodiments of the present application are as described above, the content is only for the purpose of facilitating the understanding of the application, and is not intended to limit the scope and application of the application. Any person skilled in the art of the technology described in the present application can make any modification and change in the form of implementation and details without departing from the spirit and scope of the present application, but the patent protection scope of the present application shall be subject to the scope defined by the appended claims.

Claims

1. A method for training a recognition model to identify ischemic stroke, wherein the recognition model includes a first model and a second model formed based on an exponential moving average of the weights in the first model, and both the first model and the second model include an encoder and a decoder, and the method includes: Acquire diffusion-weighted images, wherein the diffusion-weighted images include both labeled and unlabeled images; The diffusion-weighted imaging is preprocessed to obtain preprocessed diffusion-weighted imaging; The preprocessed diffusion-weighted imaging is input into the first model and the second model respectively, and the corresponding image features are extracted by the encoders in the first model and the second model respectively. Based on the corresponding image features, segmentation and scoring tasks are performed by the decoders in the first and second models, respectively, and supervised loss with annotation and unsupervised loss without annotation are calculated. as well as The total loss is determined based on the supervised loss and the unsupervised loss to train the identification model for ischemic stroke. The encoders in both the first and second models include multiple first convolutional layers of different scales, and each convolutional layer is followed by a batch normalization layer and a first ReLU activation layer. The preprocessed diffusion-weighted imaging is downsampled by the encoders of the first and second models at multiple different scales of the first convolutional layers, batch normalization layers and first ReLU activation layers to extract the corresponding image features. In the segmentation task, the decoders in the first model and the second model include at least multiple deconvolutional layers, a second convolutional layer, and a second ReLU activation layer. The corresponding image features are upsampled through the multiple deconvolutional layers of the decoders in the first model and the second model to obtain upsampled features. The upsampled features are concatenated with the image features of the corresponding first convolutional layer of the encoder in the first model and the second model, and the concatenated result is used for segmentation through the second convolutional layer and the second ReLU activation layer to obtain the lesion segmentation result.

2. The method according to claim 1, wherein the labeled elements include at least lesion labels and ASPECT score labels.

3. The method according to claim 1, wherein the preprocessing includes at least one or more of skull dissection, image registration, or normalization operations.

4. The method according to claim 1, wherein in the scoring task, the decoders in the first model and the second model include a global average pooling layer, a fully connected layer and a sigmoid activation layer, wherein the corresponding image features are classified through the global average pooling layer, the fully connected layer and the sigmoid activation layer of the decoders in the first model and the second model to obtain the classification result of the target scoring annotation region.

5. The method of claim 1, wherein the supervised loss is calculated by: Calculate the segmentation loss and classification loss for the segmentation task and the scoring task respectively, under the labeled conditions; and The supervised loss is calculated based on the weighted sum of the segmentation loss and the classification loss.

6. The method of claim 1, wherein the unsupervised loss is calculated by: Calculate the segmentation consistency loss and classification consistency loss of the first model and the second model respectively for the segmentation task and the scoring task under unlabeled conditions; and The unsupervised loss is calculated based on the weighted sum of the segmentation consistency loss and the classification consistency loss.

7. A method for identifying ischemic stroke, comprising: Obtain diffusion-weighted imaging; as well as The diffusion-weighted imaging is input into the recognition model trained by the method according to any one of claims 1-6, and the recognition result of ischemic stroke is obtained by the first model in the recognition model.

8. An apparatus comprising: processor; as well as A memory storing program instructions for training a recognition model for ischemic stroke, wherein when the program instructions are executed by the processor, the device performs the method according to any one of claims 1-6; or storing program instructions for recognizing ischemic stroke, wherein when the program instructions are executed by the processor, the device performs the method according to claim 7.

9. A computer-readable storage medium storing computer-readable instructions for training a recognition model for ischemic stroke, wherein the computer-readable instructions, when executed by one or more processors, implement the method as described in any one of claims 1-6; or storing computer-readable instructions for recognizing ischemic stroke, wherein the computer-readable instructions, when executed by one or more processors, implement the method as described in claim 7.

Images