Method, device and equipment for identifying rotator cuff tearing, storage medium and product

Abstract

The invention discloses a rotator cuff tearing identification method and device, equipment, a storage medium and a product, and relates to the technical field of computer vision, and the method comprises the steps: carrying out the preprocessing of a magnetic resonance imaging image; performing feature extraction on the to-be-recognized torn image through target feature extraction equipment; based on a target recognition model, recognition is carried out according to the high-dimensional feature image, and the target recognition model is composed of a pre-training model trained according to a self-supervised training strategy of random occlusion, a deep residual network and a convolutional neural network; through the above mode, after the high-dimensional feature image is extracted, the target recognition model trained by the randomly-shielded self-supervised training strategy is used for recognition, the training strategy can force the model to learn a mode of deducing lost information from the remaining visible image part, and the trained target recognition model has good representation capability. Therefore, the accuracy and efficiency of identifying the tear category and the tear segmentation data can be effectively improved, and tear repair is assisted.

Description

Technical Field

[0001] The present application relates to the field of computer vision technology, and in particular to methods, devices, equipment, storage media and products for identifying rotator cuff tears. Background Art

[0002] Rotator cuff tear is a common cause of shoulder disability. Acute or chronic rotator cuff tendon tear causes shoulder pain and decreased shoulder function in users. Therefore, rotator cuff tear identification is particularly important. Currently, magnetic resonance imaging (MRI) is widely used in the identification of rotator cuff tears. The commonly used identification method is for radiologists to manually interpret MRI images for rotator cuff tear identification, which requires a lot of time and effort. In addition, the accuracy of identification is affected by the subjective changes of radiologists, resulting in low accuracy and efficiency in identifying tear types and tear segmentation.

[0003] The above contents are only used to assist in understanding the technical solution of the present application and do not constitute an admission that the above contents are prior art. Summary of the invention

[0004] The main purpose of this application is to provide a rotator cuff tear identification method, device, equipment, storage medium and product, aiming to solve the technical problems of low accuracy and efficiency in identifying tear types and tear segmentation in the prior art.

[0005] To achieve the above objectives, the present application proposes a method for identifying a rotator cuff tear, the method comprising:

[0006] Acquire a magnetic resonance imaging image of the target object, and preprocess the magnetic resonance imaging image to obtain a tear image to be identified;

[0007] Extracting features of the torn image to be identified by a target feature extraction device to obtain a high-dimensional feature image;

[0008] The rotator cuff tear category and rotator cuff tear segmentation data of the target object are identified according to the high-dimensional feature image based on a target recognition model, wherein the target recognition model is composed of a pre-trained model trained according to a self-supervised training strategy with random occlusion, a deep residual network, and a convolutional neural network.

[0009] In one embodiment, before the step of identifying the rotator cuff tear category and rotator cuff tear segmentation data of the target object based on the high-dimensional feature image based on the target recognition model, the step further includes:

[0010] Acquire various types of initial rotator cuff tear sample data, and perform pixel value restriction on the various types of initial rotator cuff tear sample data;

[0011] The various types of initial rotator cuff tear sample data after restriction are divided to obtain a rotator cuff tear sample training set and a rotator cuff tear sample test set;

[0012] Performing data enhancement on the rotator cuff tear sample training set to obtain a rotator cuff tear sample augmentation set;

[0013] The current recognition model is composed of a pre-trained model trained according to the rotator cuff tear sample augmentation set and the self-supervised training strategy of random occlusion, a deep residual network, and a convolutional neural network;

[0014] Testing the current recognition model according to the rotator cuff tear sample test set;

[0015] When the current test result meets the preset requirements, the current recognition model is determined to be the target recognition model.

[0016] In one embodiment, the step of performing data enhancement on the rotator cuff tear sample training set to obtain a rotator cuff tear sample augmentation set includes:

[0017] Rotating the images in the rotator cuff tear sample training set according to a preset rotation angle based on a target interpolation strategy;

[0018] Determine the target flip axis;

[0019] The images in the rotated rotator cuff tear sample training set are flipped according to the target flip axis to obtain a rotator cuff tear sample augmentation set.

[0020] In one embodiment, before the step of forming the current recognition model using the pre-trained model, deep residual network and convolutional neural network trained according to the rotator cuff tear sample augmentation set and the random occlusion self-supervised training strategy, the step further includes:

[0021] Performing preliminary segmentation on the images in the rotator cuff tear sample amplification set based on a target mainstream image segmentation model;

[0022] Determine an approximate location region of the tendon according to the preliminary image segmentation result, and mark a boundary box of the tendon according to the approximate location region;

[0023] enlarging the approximate location region according to the bounding box;

[0024] The enlarged approximate location region is cropped, and the cropped approximate location region is scaled;

[0025] Performing dimension conversion on the scaled rotator cuff image in the approximate position area to obtain a target rotator cuff tear sample training set;

[0026] The current recognition model is composed of a pre-trained model, a deep residual network, and a convolutional neural network trained according to the target rotator cuff tear sample training set and a self-supervised training strategy of random occlusion.

[0027] In one embodiment, the step of forming the current recognition model using the pre-trained model, the deep residual network, and the convolutional neural network trained according to the target rotator cuff tear sample training set and the random occlusion self-supervised training strategy includes:

[0028] Training a pre-trained model based on the target rotator cuff tear sample training set and a self-supervised training strategy with random occlusion;

[0029] Determine a current classification model according to the pre-trained model and the deep residual network;

[0030] Determining a current segmentation model according to the pre-trained model and the convolutional neural network;

[0031] A current recognition model is constructed according to the current classification model and the current segmentation model.

[0032] In one embodiment, the step of training the pre-trained model according to the target rotator cuff tear sample training set and the self-supervised training strategy of random occlusion includes:

[0033] Determine the mask ratio based on the self-supervised training strategy of random occlusion;

[0034] Randomly masking the image in the target rotator cuff tear sample training set according to the mask ratio to obtain a current mask image;

[0035] Performing feature extraction on the current mask image in the Encoder stage;

[0036] Encode the extracted image features to obtain highly abstract image features;

[0037] In the decoder stage, the highly abstract image features are decoded by a transposed convolutional network to restore the image;

[0038] The pre-trained model is trained based on the mean square error between the restored image and the image in the target rotator cuff tear sample training set as a loss function.

[0039] In addition, to achieve the above-mentioned purpose, the present application also proposes a rotator cuff tear identification device, the rotator cuff tear identification device comprising:

[0040] An acquisition module, used for acquiring a magnetic resonance imaging image of a target object, and preprocessing the magnetic resonance imaging image to obtain a tearing image to be identified;

[0041] A feature extraction module, used for extracting features of the torn image to be identified through a target feature extraction device to obtain a high-dimensional feature image;

[0042] An identification module is used to identify the rotator cuff tear category and rotator cuff tear segmentation data of the target object according to the high-dimensional feature image based on a target recognition model, wherein the target recognition model is composed of a pre-trained model trained according to a self-supervised training strategy with random occlusion, a deep residual network, and a convolutional neural network.

[0043] In addition, to achieve the above-mentioned purpose, the present application also proposes a rotator cuff tear identification device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, and the computer program is configured to implement the steps of the rotator cuff tear identification method described above.

[0044] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the rotator cuff tear identification method described above are implemented.

[0045] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the steps of the rotator cuff tear identification method as described above are implemented.

[0046] One or more technical solutions proposed in the present application have at least the following technical effects: by acquiring a magnetic resonance imaging image of a target object, preprocessing the magnetic resonance imaging image to obtain a tear image to be identified; extracting features of the tear image to be identified by a target feature extraction device to obtain a high-dimensional feature image; based on a target recognition model, identifying the rotator cuff tear category and rotator cuff tear segmentation data of the target object according to the high-dimensional feature image, wherein the target recognition model is composed of a pre-trained model trained according to a self-supervised training strategy with random occlusion, a deep residual network, and a convolutional neural network; through the above method, after acquiring the tear image to be identified, the high-dimensional feature image is extracted by a target feature extraction device, and then the tear image is identified by a target recognition model trained by a self-supervised training strategy with random occlusion, the above training strategy can force the model to learn a way to infer lost information from the remaining visible image part, and the trained target recognition model has good representation ability, thereby effectively improving the accuracy and efficiency of identifying tear categories and tear segmentation data, thereby assisting tear repair. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings herein are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present application, and together with the description, are used to explain the principles of the present application.

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0049] Figure 1 A schematic diagram of a flow chart provided for Embodiment 1 of the rotator cuff tear identification method of the present application;

[0050] Figure 2 A schematic diagram of a flow chart provided for Embodiment 2 of the rotator cuff tear identification method of the present application;

[0051] Figure 3 A schematic diagram of a brief process of a rotator cuff tear identification method provided in Embodiment 2 of the present application;

[0052] Figure 4 This is a schematic diagram of the module structure of the rotator cuff tear identification device according to an embodiment of the present application;

[0053] Figure 5 Schematic diagram of the device structure of the hardware operating environment involved in the rotator cuff tear identification method in the embodiment of the present application.

[0054] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0055] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, a rotator cuff tear identification system, etc. The rotator cuff tear identification system is taken as an example to illustrate this embodiment and the following embodiments.

[0056] Based on this, the present application embodiment provides a method for identifying a rotator cuff tear, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the rotator cuff tear identification method of the present application.

[0057] In this embodiment, the rotator cuff tear identification method includes steps S10 to S40:

[0058] Step S10, obtaining a magnetic resonance imaging image of the target object, and preprocessing the magnetic resonance imaging image to obtain a tearing image to be identified.

[0059] It should be noted that the target object refers to the object undergoing rotator cuff tear detection, and the target object may be a user undergoing a physical examination or a user who feels shoulder pain and decreased shoulder function. Magnetic resonance imaging refers to an imaging image that uses magnetic fields and radio waves to generate detailed images of the internal structure of the human body.

[0060] It should be understood that the torn image to be identified refers to the torn image for identification by the target recognition model, and the torn image to be identified is obtained by preprocessing the magnetic resonance imaging image, and the preprocessing operations include but are not limited to segmentation, labeling, enlargement, cropping and scaling.

[0061] Step S20: extracting features of the torn image to be identified by a target feature extraction device to obtain a high-dimensional feature image.

[0062] It can be understood that the target feature extraction device refers to a device for high-dimensional feature extraction. After the torn image to be identified is obtained, the target feature extraction device is used to extract a high-dimensional feature image from the torn image to be identified.

[0063] Step S30, identifying the rotator cuff tear category and rotator cuff tear segmentation data of the target object according to the high-dimensional feature image based on a target recognition model, wherein the target recognition model is composed of a pre-trained model trained according to a self-supervised training strategy with random occlusion, a deep residual network, and a convolutional neural network.

[0064] It should be understood that the target recognition model refers to a model used to identify rotator cuff tear categories and rotator cuff tear segmentation data. The rotator cuff tear categories include but are not limited to normal, mild and severe. In addition, the target recognition model is composed of a pre-trained model trained according to a self-supervised training strategy with random occlusion, a deep residual network and a convolutional neural network. The beneficial effect of model training using a self-supervised training strategy with random occlusion is that it forces the model to learn how to infer lost information from the remaining visible image parts. The trained target recognition model has good representation capabilities, which can effectively improve the accuracy and efficiency of identifying tear categories and tear segmentation data. The goal of this pre-training stage is to allow the model to learn to reconstruct images without the original image category annotations, so that it can better generalize in subsequent feature extraction tasks.

[0065] It can be understood that identifying the rotator cuff tear category and rotator cuff tear segmentation data of the target object is conducive to further assisting tear repair by using the rotator cuff tear category and rotator cuff tear segmentation data.

[0066] In this embodiment, a magnetic resonance imaging image of a target object is acquired, and the magnetic resonance imaging image is preprocessed to obtain a tear image to be identified; a target feature extraction device is used to extract features of the tear image to be identified to obtain a high-dimensional feature image; a target recognition model is used to identify the rotator cuff tear category and rotator cuff tear segmentation data of the target object according to the high-dimensional feature image, wherein the target recognition model is composed of a pre-trained model trained according to a self-supervised training strategy with random occlusion, a deep residual network, and a convolutional neural network; in the above manner, after acquiring the tear image to be identified, the high-dimensional feature image is extracted by a target feature extraction device, and then the tear image is identified by a target recognition model trained by a self-supervised training strategy with random occlusion, and the training strategy can force the model to learn a way to infer lost information from the remaining visible image part, and the trained target recognition model has good representation ability, thereby effectively improving the accuracy and efficiency of identifying tear categories and tear segmentation data, thereby assisting tear repair.

[0067] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can be referred to the above introduction, and will not be repeated in the following. Figure 2 , before step S30, it also includes steps S201 to S206:

[0068] Step S201, obtaining various types of initial rotator cuff tear sample data, and performing pixel value restriction on the various types of initial rotator cuff tear sample data.

[0069] It should be noted that the initial rotator cuff tear sample data includes three categories, namely normal (category 0), partial tear (category 1) and complete tear (category 2). In order to effectively improve the image quality, it is necessary to limit the pixel values ​​of the images in each type of initial rotator cuff tear sample data, specifically limiting the pixel values ​​between 0 and 5000, which helps to remove the overly bright parts of the images in the rotator cuff tear sample training set and improve the contrast and the overall visualization effect of the image.

[0070] Step S202, dividing the various types of restricted initial rotator cuff tear sample data to obtain a rotator cuff tear sample training set and a rotator cuff tear sample test set.

[0071] It can be understood that after the pixel value is restricted, the various types of initial rotator cuff tear sample data after restriction are divided into a rotator cuff tear sample training set and a rotator cuff tear sample test set. Taking the actual 1058 cases in a certain hospital as an example, the rotator cuff tear sample training set is divided into: 70 in category 0, 701 in category 1, and 76 in category 2, and the rotator cuff tear sample test set is 18 in category 0, 175 in category 1, and 18 in category 2.

[0072] Step S203, performing data enhancement on the rotator cuff tear sample training set to obtain a rotator cuff tear sample augmentation set.

[0073] It should be understood that in order to increase the diversity of data samples, alleviate the problem of data imbalance, and help improve the generalization ability of the model, data enhancement is also required, that is, using data enhancement to expand the rotator cuff tear sample training set into a rotator cuff tear sample expansion set.

[0074] Further, step S203 includes: rotating the images in the rotator cuff tear sample training set according to a preset rotation angle based on a target interpolation strategy; determining a target flip axis; flipping the rotated images in the rotator cuff tear sample training set according to the target flip axis to obtain a rotator cuff tear sample expansion set.

[0075] It can be understood that the target interpolation strategy can be a bilinear interpolation strategy, and the number of preset rotation angles can be 8, that is, the images in the rotator cuff tear sample training set are rotated according to the set 8 rotation angles, and a bilinear interpolation algorithm is used to ensure the smoothness of the rotator cuff tear sample training set during the rotation process, and then the images in the rotated rotator cuff tear sample training set are flipped according to the target flip axis. For example, the number of Class 0 and Class 2 samples in the rotator cuff tear sample training set is expanded to 24 times the original, that is, the number of Class 0 samples is increased to 1680, and the number of Class 2 samples is increased to 1824.

[0076] Step S204, a current recognition model is constructed based on the pre-trained model, deep residual network and convolutional neural network trained according to the rotator cuff tear sample augmentation set and the random occlusion self-supervised training strategy.

[0077] It should be noted that after obtaining the augmented set of rotator cuff tear samples, the pre-trained model was trained using the self-supervised training strategy with random occlusion, and then the current recognition model was formed by combining the deep residual network and the convolutional neural network.

[0078] Furthermore, before step S204, it also includes: performing preliminary segmentation on the images in the rotator cuff tear sample amplification set based on the target mainstream image segmentation model; determining the approximate position area of ​​the tendon according to the preliminary image segmentation result, and marking the bounding box of the tendon according to the approximate position area; enlarging the approximate position area according to the bounding box; cropping the enlarged approximate position area, and scaling the cropped approximate position area; performing dimension conversion on the rotator cuff image in the scaled approximate position area to obtain the target rotator cuff tear sample training set; the pre-trained model, deep residual network and convolutional neural network trained according to the target rotator cuff tear sample training set and the self-supervised training strategy of random occlusion constitute the current recognition model.

[0079] It should be understood that the target mainstream image segmentation model refers to the mainstream model used for image segmentation in the medical field. The target mainstream image segmentation model can be a nnUNet model. After obtaining the magnetic resonance imaging image of the target object, the target mainstream image segmentation model is used to perform preliminary segmentation to determine the approximate position area of ​​the tendon of the target object, and the bounding box of the tendon is marked according to the approximate position area using preset labels.

[0080] It is understandable that in order to facilitate fine classification and segmentation tasks, it is also necessary to use a bounding box to enlarge the approximate position area, and the enlargement ratio can be flexibly set according to actual needs. After enlargement, a series of operations such as cropping and scaling are required to increase the proportion of tendons in the image. The specific execution process is: first determine the boundary of the image mask, then appropriately expand the cropping range to include the necessary edge information, then crop, and then scale the cropped approximate position area. In addition, the dimension value of the dimension conversion can be [32,160,160].

[0081] Furthermore, the step of forming the current recognition model from the pre-trained model, deep residual network and convolutional neural network trained according to the target rotator cuff tear sample training set and the self-supervised training strategy of random occlusion includes: training the pre-trained model according to the target rotator cuff tear sample training set and the self-supervised training strategy of random occlusion; determining the current classification model according to the pre-trained model and the deep residual network; determining the current segmentation model according to the pre-trained model and the convolutional neural network; and forming the current recognition model according to the current classification model and the current segmentation model.

[0082] It can be understood that after obtaining the target rotator cuff tear sample training set, a pre-training process is carried out. The training strategy adopted in the pre-training is a self-supervised training strategy with random occlusion. Then, the current classification model is determined in combination with the deep residual network, and the current segmentation model is determined in combination with the convolutional neural network. At this time, the current recognition model is composed of the current classification model and the current segmentation model.

[0083] It should be understood that after the pre-trained model is trained, the feature extraction and current classification model training phase begins. The Encoder part of the pre-trained Encoder-Decoder architecture is now used as an independent target feature extraction device E c The main goal of this stage is to use these pre-trained deep features to improve the performance and accuracy of subsequent classification tasks, specifically:

[0084] e=E c (x).

[0085] Where e represents the high-dimensional embedding vector, E crepresents the target feature extraction device, and x represents the image in the target rotator cuff tear sample training set.

[0086] It should be noted that the above high-dimensional embedding vector captures the key features of the input target rotator cuff tear sample training set, providing valuable information for subsequent classification tasks. For high-dimensional embedded information, this embodiment can use a deep residual network for classification. The deep residual network can be ResNet-50, which is widely used in various image recognition tasks due to its powerful feature learning ability. In this application, the ResNet-50 network uses the target feature extraction device E c The obtained high-dimensional embedding vector e is used to identify the category of the image. For each high-dimensional embedding vector e:

[0087]

[0088] Among them, AdaptiveAvgPool3d(·) represents the maximum average pooling, W5 and b5 represent the weight and bias of the fifth layer of convolution, W6 and b6 represent the weight and bias of the sixth layer of convolution, and ResidualBlock3d(·) represents a layer of the ResNet network, which is specifically expressed as:

[0089] It should be noted that in order to solve the defect of imbalanced classification, this embodiment uses the Focal Loss function to reduce the attention paid to easy-to-classify samples to increase the model's attention to difficult-to-classify samples, specifically:

[0090]

[0091] Among them, m represents the total number of categories, p c is the probability that the model predicts that the sample belongs to category c, α c is the weight of category c, and γ represents the adjustment factor for adjusting the impact of difficult and easy classification samples.

[0092] It should be understood that in order to ensure that the Encoder continues to improve its feature extraction ability in the classification task, the reconstruction loss of the image, that is, the mean square error, is recorded. This not only helps maintain the feature extraction ability of the Encoder, but also promotes the further optimization of the Decoder. In order to optimize the classification performance and reconstruction quality at the same time, the loss function of the final model is the weighted sum of the pixel value difference and the Focal Loss function, specifically:

[0093] Loss = β·MES + FL.

[0094] Among them, Loss represents the loss function of the model, FL represents the Focal Loss function, and β represents the hyperparameter, which is used to control the contribution of reconstruction loss to the total loss. By adjusting β, the best balance between reconstruction accuracy and classification accuracy is found to achieve the best model performance.

[0095] It should be noted that during the training process, the gradient descent strategy is continuously used to adjust the network parameters to minimize the total loss. In particular, the parameters of the feature extractor and classifier are updated to ensure that the entire model can effectively learn from the training data. The model trained by combining feature extraction and classification training can more accurately identify the category of the image and maintain high performance even when facing a dataset with imbalanced categories.

[0096] On the other hand, after the pre-training model is trained, the training phase of 3D feature embedding and segmentation will begin. The Encoder-Decoder structure obtained through pre-training will be used as an independent 3D feature extractor. The main goal of this phase is to use deep features to improve the accuracy and efficiency of the segmentation task. In this phase, the images in the input target rotator cuff tear sample training set are first passed through the Encoder module and the Decoder module composed of a 3D convolutional network to obtain a high-dimensional 3D feature image (3D Embedding). This high-dimensional 3D feature image can preferentially encode important information in the image while ensuring that the size is the same as the original image, providing valuable information for subsequent segmentation tasks.

[0097] It should be understood that after obtaining the high-dimensional 3D feature image, the high-dimensional 3D feature image is input into the UNet network for detailed segmentation. The structure of the UNet network is particularly suitable for medical image segmentation, and performs the task through a symmetrical encoder-decoder structure and jump connections, which includes multiple convolutional blocks and jump connections. In this application, the high-dimensional 3D feature image first passes through the encoder part of UNet, then through the decoder part, and finally through an output layer to generate a segmented image.

[0098] In the encoding stage, the encoding device includes a plurality of convolution blocks, each of which includes two convolution layers and a maximum pooling layer for downsampling. The encoding device may be an encoder. For the i-th convolution block in the encoder, specifically:

[0099]

[0100] Among them, e0 represents the high-dimensional 3D feature image, W ei and b ei represents the weight and bias of the convolutional layer, BN(·) represents normalization, and ReLU(·) represents the activation function.

[0101] In the decoding stage, the decoding device includes multiple convolution blocks and upsampling operations. The decoding device can be a decoder, and each decoder block receives the skip connection output from the corresponding encoder block and the output of the previous decoder block. Output layer: The output of the last decoder passes through a convolution layer to predict the category of each pixel.

[0102] In addition, for the training of the segmentation task, a combination of the target coefficient loss function and the target cross entropy loss function is used to optimize the segmentation effect. The target coefficient loss function is used to optimize the class imbalance problem and focus on the global integrity of the structure. The target cross entropy loss function provides strict penalties for classification errors and effectively enhances the model's ability to capture details. This combined loss function shows excellent performance when processing medical images with complex structures and size changes. The target coefficient loss function can be a Dice Loss function, which is based on the Dice coefficient and is used to evaluate the similarity of two samples and deal with the defects of class imbalance. Specifically:

[0103]

[0104] Among them, Dice Loss represents the similarity value between samples, X represents the predicted segmentation area, and Y represents the actual segmentation area. When the predicted segmentation area is completely consistent with the actual segmentation area, the similarity value between samples is 0, thereby driving the model to optimize towards a more accurate segmentation direction.

[0105] In addition, the target cross entropy loss function can be a two-class cross entropy loss function, which is used to calculate the difference between the predicted probability and the actual label. Each pixel is regarded as an independent two-class problem. This form of loss function can effectively drive model optimization, so that the predicted probability is close to the real label. It is very suitable for pixel-level two-class image segmentation tasks. During the training process, the network parameters are continuously adjusted through the gradient descent strategy to minimize the total loss. In particular, the parameters of the 3D feature extractor and segmenter will be updated to ensure that the entire model can effectively learn from the training data. Through this combination of feature extraction and classification training, the target recognition model in this embodiment can more accurately identify and segment images, and maintain high performance even when facing class-imbalanced data sets.

[0106] Furthermore, the step of training the pre-trained model according to the target rotator cuff tear sample training set and the self-supervised training strategy of random occlusion includes: determining the mask ratio according to the self-supervised training strategy of random occlusion; randomly occluding the image in the target rotator cuff tear sample training set according to the mask ratio to obtain the current mask image; performing feature extraction on the current mask image in the Encoder stage; encoding the extracted image features to obtain highly abstract image features; decoding the highly abstract image features through a transposed convolutional network in the Decoder stage to restore the image; and training the pre-trained model based on the mean square error between the restored image and the image in the target rotator cuff tear sample training set as the loss function.

[0107] It should be understood that the mask ratio refers to the ratio of image blocks covered by random occlusion to the total image blocks. The mask ratio can be determined by a self-supervised training strategy of random occlusion. For example, the mask ratio is 70%, specifically:

[0108] x'=Mask(x).

[0109] Where x' represents the current mask image, Mask(·) represents the random masking process, and x represents the image in the target rotator cuff tear sample training set.

[0110] It can be understood that after obtaining the current mask image, the Encoder-Decoder architecture is used to process the current mask image. Specifically, the image features are extracted in the Encoder stage, and then the extracted image features are encoded into a highly abstract representation, that is, a highly abstract image feature, specifically:

[0111]

[0112] Among them, W1 and b1 represent the weight and bias of the first layer of convolution, W2 and b2 represent the weight and bias of the second layer of convolution, Conv3d(·) represents the 3D convolution process, ReLU(·) represents the activation function, MaxPool3d(·) represents the maximum pooling process, and z represents the highly abstract image feature.

[0113] It can be understood that after obtaining the highly abstract image features in the Encoder stage, the highly abstract image features are decoded by the transposed convolutional network in the Decoder stage to try to reconstruct the complete visual content of the images in the rotator cuff tear sample amplification set, specifically:

[0114]

[0115] Among them, W3 and b3 represent the weight and bias of the first layer of deconvolution, W4 and b4 represent the weight and bias of the second layer of deconvolution, ConvTranspose3d(·) represents the 3D deconvolution process, Sigmoid(·) represents the activation function, Represents the restored image.

[0116] It should be noted that after the reconstruction process is completed, in order to quantify the reconstruction quality, the pre-training model is trained based on the mean square error of the restored image and the image in the target rotator cuff tear sample training set as the loss function. Therefore, it is necessary to calculate the mean square error between the restored image and the image before occlusion (i.e., the image in the target rotator cuff tear sample training set), specifically:

[0117]

[0118] Among them, y i represents the pixel value of the image in the training set of target rotator cuff tear samples, represents the pixel value of the restored image, n represents the total number of pixels, and MSE represents the mean square error.

[0119] Step S205: testing the current recognition model according to the rotator cuff tear sample test set.

[0120] Step S206: When the current test result meets the preset requirement, determine that the current recognition model is a target recognition model.

[0121] It should be understood that after testing the current recognition model using the rotator cuff tear sample test set, it will be determined whether the current test results meet the preset requirements. If so, the current recognition model is determined to be the target recognition model. If not, it indicates that the current recognition model does not meet the preset requirements and training must continue.

[0122] In this embodiment, various types of initial rotator cuff tear sample data are obtained, and pixel values ​​of the various types of initial rotator cuff tear sample data are restricted; the various types of initial rotator cuff tear sample data after restriction are divided to obtain a rotator cuff tear sample training set and a rotator cuff tear sample test set; the rotator cuff tear sample training set is data enhanced to obtain a rotator cuff tear sample expansion set; a pre-trained model, a deep residual network and a convolutional neural network trained according to the rotator cuff tear sample expansion set and a random occlusion self-supervised training strategy constitute a current recognition model; the current recognition model is tested according to the rotator cuff tear sample test set; when the current test result meets the preset requirements, the current recognition model is determined to be a target recognition model; in the above manner, after a series of processing such as pixel value restriction, data division and data enhancement are performed on various types of initial rotator cuff tear sample data, the pre-trained model is trained using a random occlusion self-supervised training strategy, and the current recognition model is formed by combining a deep residual network and a convolutional neural network, and finally when the current test result meets the preset requirements, the target recognition model is determined, thereby effectively improving the accuracy of the target recognition model.

[0123] For example, to help understand the implementation process of the rotator cuff tear identification method obtained by combining this embodiment with the above-mentioned embodiment 1, please refer to Figure 3 , Figure 3 A brief flowchart of a rotator cuff tear identification method is provided, specifically:

[0124] After acquiring the magnetic resonance imaging image of the target object, the magnetic resonance imaging image is preprocessed, and then a high-dimensional feature image is extracted from the tear image to be identified by a target feature extraction device, and the rotator cuff tear category and rotator cuff tear segmentation data of the target object are identified according to the high-dimensional feature image using a pre-trained target recognition model. In addition, a target rotator cuff tear sample training set is required when training the target recognition model. The target rotator cuff tear sample training set is obtained by preprocessing and coarse segmentation of various types of initial rotator cuff tear sample data, wherein the preprocessing operations include but are not limited to pixel value restriction, data division, image rotation, etc., and the coarse segmentation includes but is not limited to preliminary segmentation, marking bounding boxes, magnification, cropping, and dimensional conversion, etc. It should also be noted that after identifying the rotator cuff tear category and rotator cuff tear segmentation data of the target object, the rotator cuff tear category and rotator cuff tear segmentation data of the target object can be used to assist in tear repair.

[0125] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the rotator cuff tear identification method of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0126] This application also provides a rotator cuff tear identification device, please refer to Figure 4, the rotator cuff tear identification device comprises:

[0127] The acquisition module 10 is used to acquire a magnetic resonance imaging image of the target object, and pre-process the magnetic resonance imaging image to obtain a tear image to be identified.

[0128] The feature extraction module 20 is used to extract features of the torn image to be identified by using a target feature extraction device to obtain a high-dimensional feature image.

[0129] The recognition module 30 is used to identify the rotator cuff tear category and rotator cuff tear segmentation data of the target object according to the high-dimensional feature image based on a target recognition model, wherein the target recognition model is composed of a pre-trained model trained according to a self-supervised training strategy with random occlusion, a deep residual network, and a convolutional neural network.

[0130] In this embodiment, a magnetic resonance imaging image of a target object is acquired, and the magnetic resonance imaging image is preprocessed to obtain a tear image to be identified; a target feature extraction device is used to extract features of the tear image to be identified to obtain a high-dimensional feature image; a target recognition model is used to identify the rotator cuff tear category and rotator cuff tear segmentation data of the target object according to the high-dimensional feature image, wherein the target recognition model is composed of a pre-trained model trained according to a self-supervised training strategy with random occlusion, a deep residual network, and a convolutional neural network; in the above manner, after acquiring the tear image to be identified, the high-dimensional feature image is extracted by a target feature extraction device, and then the tear image is identified by a target recognition model trained by a self-supervised training strategy with random occlusion, and the training strategy can force the model to learn a way to infer lost information from the remaining visible image part, and the trained target recognition model has good representation ability, thereby effectively improving the accuracy and efficiency of identifying tear categories and tear segmentation data, thereby assisting tear repair.

[0131] The rotator cuff tear identification device provided by the present application adopts the rotator cuff tear identification method in the above embodiment, which can solve the technical problems of low accuracy and efficiency in identifying tear types and tear segmentation in the prior art. Compared with the prior art, the beneficial effects of the rotator cuff tear identification device provided by the present application are the same as the beneficial effects of the rotator cuff tear identification method provided by the above embodiment, and the other technical features of the rotator cuff tear identification device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0132] In one embodiment, the recognition module 30 is also used to obtain various types of initial rotator cuff tear sample data, and perform pixel value restrictions on the various types of initial rotator cuff tear sample data; divide the various types of initial rotator cuff tear sample data after restriction to obtain a rotator cuff tear sample training set and a rotator cuff tear sample test set; perform data enhancement on the rotator cuff tear sample training set to obtain a rotator cuff tear sample expansion set; construct a current recognition model based on a pre-trained model, a deep residual network, and a convolutional neural network trained with the rotator cuff tear sample expansion set and a self-supervised training strategy with random occlusion; test the current recognition model based on the rotator cuff tear sample test set; and when the current test result meets the preset requirements, determine that the current recognition model is the target recognition model.

[0133] In one embodiment, the recognition module 30 is also used to rotate the images in the rotator cuff tear sample training set according to a preset rotation angle based on a target interpolation strategy; determine a target flip axis; and flip the rotated images in the rotator cuff tear sample training set according to the target flip axis to obtain a rotator cuff tear sample expansion set.

[0134] In one embodiment, the recognition module 30 is also used to perform preliminary segmentation of the images in the rotator cuff tear sample amplification set based on the target mainstream image segmentation model; determine the approximate position area of ​​the tendon according to the preliminary image segmentation result, and mark the bounding box of the tendon according to the approximate position area; enlarge the approximate position area according to the bounding box; crop the enlarged approximate position area, and scale the cropped approximate position area; perform dimension conversion on the rotator cuff image in the scaled approximate position area to obtain the target rotator cuff tear sample training set; the pre-trained model, deep residual network and convolutional neural network trained according to the target rotator cuff tear sample training set and the self-supervised training strategy with random occlusion constitute the current recognition model.

[0135] In one embodiment, the recognition module 30 is also used to train a pre-trained model based on the target rotator cuff tear sample training set and the self-supervised training strategy of random occlusion; determine a current classification model based on the pre-trained model and the deep residual network; determine a current segmentation model based on the pre-trained model and the convolutional neural network; and construct a current recognition model based on the current classification model and the current segmentation model.

[0136] In one embodiment, the recognition module 30 is also used to determine the mask ratio according to the self-supervised training strategy of random occlusion; randomly occlude the image in the target rotator cuff tear sample training set according to the mask ratio to obtain the current mask image; extract features of the current mask image in the Encoder stage; encode the extracted image features to obtain highly abstract image features; decode the highly abstract image features through a transposed convolutional network in the Decoder stage to restore the image; train the pre-trained model based on the mean square error between the restored image and the image in the target rotator cuff tear sample training set as the loss function.

[0137] The present application provides a rotator cuff tear identification device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the rotator cuff tear identification method in the above-mentioned embodiment one.

[0138] Reference below Figure 5 , which shows a schematic diagram of the structure of a rotator cuff tear identification device suitable for implementing the embodiment of the present application. The rotator cuff tear identification device in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 5 The rotator cuff tear identification device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0139] like Figure 5As shown, the rotator cuff tear identification device may include a processing device 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the rotator cuff tear identification device are also stored. The processing device 1001, ROM1002, and RAM1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the rotator cuff tear identification device to communicate wirelessly or wired with other devices to exchange data. Although the rotator cuff tear identification device with various systems is shown in the figure, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented or have alternatively.

[0140] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0141] The rotator cuff tear identification device provided by the present application adopts the rotator cuff tear identification method in the above embodiment, which can solve the technical problems of low accuracy and efficiency in identifying tear types and tear segmentation in the prior art. Compared with the prior art, the beneficial effects of the rotator cuff tear identification device provided by the present application are the same as the beneficial effects of the rotator cuff tear identification method provided by the above embodiment, and the other technical features in the rotator cuff tear identification device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0142] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0143] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0144] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer program) stored thereon, and the computer-readable program instructions are used to execute the rotator cuff tear identification method in the above-mentioned embodiment.

[0145] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM: Erasable Programmable Read Only Memory or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM: CD-Read Only Memory), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0146] The computer-readable storage medium may be included in the rotator cuff tear identification device; or may exist independently without being assembled into the rotator cuff tear identification device.

[0147] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0148] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0149] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0150] The readable storage medium provided in the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned rotator cuff tear identification method, and can solve the technical problems of low accuracy and efficiency in identifying tear types and tear segmentation in the prior art. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in the present application are the same as the beneficial effects of the rotator cuff tear identification method provided in the above-mentioned embodiment, and will not be repeated here.

[0151] The present application also provides a computer program product, including a computer program, which implements the steps of the above-mentioned rotator cuff tear identification method when executed by a processor.

[0152] The computer program product provided by the present application can solve the technical problems of low accuracy and efficiency in identifying tear types and tear segmentation in the prior art. Compared with the prior art, the beneficial effects of the computer program product provided by the present application are the same as the beneficial effects of the rotator cuff tear identification method provided by the above embodiment, which will not be described in detail here.

[0153] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A method for identifying a rotator cuff tear, characterized in that: The method comprises: Acquire a magnetic resonance imaging image of the target object, and preprocess the magnetic resonance imaging image to obtain a tear image to be identified; Extracting features of the torn image to be identified by a target feature extraction device to obtain a high-dimensional feature image; The rotator cuff tear category and rotator cuff tear segmentation data of the target object are identified according to the high-dimensional feature image based on a target recognition model, wherein the target recognition model is composed of a pre-trained model trained according to a self-supervised training strategy with random occlusion, a deep residual network, and a convolutional neural network.

2. The method according to claim 1, characterized in that Before the step of identifying the rotator cuff tear category and rotator cuff tear segmentation data of the target object according to the high-dimensional feature image based on the target recognition model, the method further includes: Acquire various types of initial rotator cuff tear sample data, and perform pixel value restriction on the various types of initial rotator cuff tear sample data; The restricted initial rotator cuff tear sample data are divided into a rotator cuff tear sample training set and a rotator cuff tear sample test set; Performing data enhancement on the rotator cuff tear sample training set to obtain a rotator cuff tear sample augmentation set; The current recognition model is composed of a pre-trained model trained according to the rotator cuff tear sample augmentation set and the self-supervised training strategy of random occlusion, a deep residual network, and a convolutional neural network; Testing the current recognition model according to the rotator cuff tear sample test set; When the current test result meets the preset requirements, the current recognition model is determined to be the target recognition model.

3. The method according to claim 2, characterized in that The step of performing data enhancement on the rotator cuff tear sample training set to obtain a rotator cuff tear sample augmentation set comprises: Rotating the images in the rotator cuff tear sample training set according to a preset rotation angle based on a target interpolation strategy; Determine the target flip axis; The images in the rotated rotator cuff tear sample training set are flipped according to the target flip axis to obtain a rotator cuff tear sample augmentation set.

4. The method according to claim 2, characterized in that Before the step of forming the current recognition model with the pre-trained model, deep residual network and convolutional neural network trained according to the rotator cuff tear sample augmentation set and the random occlusion self-supervised training strategy, the method further includes: Performing preliminary segmentation on the images in the rotator cuff tear sample amplification set based on a target mainstream image segmentation model; Determine an approximate location region of the tendon according to the preliminary image segmentation result, and mark a boundary box of the tendon according to the approximate location region; enlarging the approximate location region according to the bounding box; The enlarged approximate location region is cropped, and the cropped approximate location region is scaled; Performing dimension conversion on the scaled rotator cuff image in the approximate position area to obtain a target rotator cuff tear sample training set; The current recognition model is composed of a pre-trained model, a deep residual network, and a convolutional neural network trained according to the target rotator cuff tear sample training set and a self-supervised training strategy of random occlusion.

5. The method according to claim 4, characterized in that The step of forming the current recognition model using the pre-trained model, the deep residual network and the convolutional neural network trained according to the target rotator cuff tear sample training set and the self-supervised training strategy of random occlusion includes: Training a pre-trained model based on the target rotator cuff tear sample training set and a self-supervised training strategy with random occlusion; Determine a current classification model according to the pre-trained model and the deep residual network; Determining a current segmentation model according to the pre-trained model and the convolutional neural network; A current recognition model is constructed according to the current classification model and the current segmentation model.

6. The method according to claim 5, characterized in that The step of training the pre-trained model according to the target rotator cuff tear sample training set and the self-supervised training strategy of random occlusion comprises: Determine the mask ratio based on the self-supervised training strategy of random occlusion; Randomly masking the image in the target rotator cuff tear sample training set according to the mask ratio to obtain a current mask image; Performing feature extraction on the current mask image in the Encoder stage; Encode the extracted image features to obtain highly abstract image features; In the decoder stage, the highly abstract image features are decoded by a transposed convolutional network to restore the image; The pre-trained model is trained based on the mean square error between the restored image and the image in the target rotator cuff tear sample training set as a loss function.

7. A rotator cuff tear identification device, characterized in that: The device comprises: An acquisition module, used for acquiring a magnetic resonance imaging image of a target object, and preprocessing the magnetic resonance imaging image to obtain a tearing image to be identified; A feature extraction module, used for extracting features of the torn image to be identified through a target feature extraction device to obtain a high-dimensional feature image; An identification module is used to identify the rotator cuff tear category and rotator cuff tear segmentation data of the target object according to the high-dimensional feature image based on a target recognition model, wherein the target recognition model is composed of a pre-trained model trained according to a self-supervised training strategy with random occlusion, a deep residual network, and a convolutional neural network.

8. A rotator cuff tear identification device, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the rotator cuff tear identification method according to any one of claims 1 to 6.

9. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the rotator cuff tear identification method according to any one of claims 1 to 6 are implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the rotator cuff tear identification method according to any one of claims 1 to 6 are implemented.

Images