Neural network semi-supervised learning-based spinal fracture segmentation and training method and system, and related equipment

Abstract

The invention relates to a spinal fracture segmentation and training method and system based on neural network semi-supervised learning and related equipment. The method comprises the following steps: acquiring a spine CT image of a target patient, and performing vertebra segmentation on the spine CT image based on a 3D U-Net model to obtain a vertebra CT image; classifying the vertebra CT image by using a 2.5 D vertebra fracture classification model to obtain a vertebra fracture image and a fracture-free vertebra image; constructing an artificial vertebra fracture area segmentation mask based on the fracture CT image of the specific part and the fracture-free vertebra image, and generating a vertebra fracture area segmentation mask according to the vertebra fracture image; and performing semi-supervised learning training on the 3D U-Net model by taking the artificial vertebral fracture region segmentation mask and the vertebral fracture region segmentation mask as training data to obtain a vertebral fracture segmentation model. According to the method, training of the spinal fracture segmentation model is realized under the condition of limited real patient spinal fracture image data, the accuracy and efficiency of fracture classification are improved, and the spinal fracture segmentation precision is improved.

Description

Technical Field

[0001] The present application relates to the technical field of spinal image processing, and in particular to a spinal fracture segmentation and training method, system and related equipment based on neural network semi-supervised learning. Background Art

[0002] Traditional spinal fracture diagnosis relies primarily on doctors' subjective judgment of CT images. This process is not only time-consuming but also prone to diagnostic bias due to subjective factors. Therefore, the application of deep learning technology has become a research hotspot in the field of medical image analysis, especially in the diagnosis and classification of spinal fractures.

[0003] In order to improve diagnostic efficiency and accuracy, deep learning technology based on 3D (3D) convolutional neural networks (CNNs) is often used in related technologies. However, this type of method faces the following challenges: (1) Spinal fractures are pathologically specific, and there is currently a lack of accurate labeled data for spinal fracture areas; (2) Traditional 3D convolutional methods have problems such as low learning efficiency and high model complexity, making it difficult to meet the demand for fast computing in resource-limited medical scenarios; (3) Related technologies focus on binary discrimination of fracture presence, making it difficult to achieve accurate segmentation of fracture areas, which affects the formulation of treatment plans.

[0004] Currently, no effective solution has been proposed to the above-mentioned problems in related technologies. Summary of the Invention

[0005] The embodiments of the present invention provide a spinal fracture segmentation and training method, system and related equipment based on neural network semi-supervised learning, which at least solves the problems of large deviation in diagnostic results in traditional diagnostic methods in related technologies, and the large amount of data required for training in the solution based on 3D convolutional neural network, resulting in difficulty in data acquisition, high labeling cost and low accuracy in fracture area segmentation.

[0006] To solve the above problems, one aspect of an embodiment of the present invention provides a vertebral fracture segmentation and training method based on neural network semi-supervised learning, comprising:

[0007] Obtain a spinal CT image of the target patient and perform initial vertebra segmentation on the spinal CT image based on the 3D U-Net model to obtain vertebral CT images corresponding to each vertebra;

[0008] The vertebral CT images were classified using a 2.5D vertebral fracture classification model to obtain images of vertebral fractures and images of vertebrae without fractures.

[0009] A CT image of a fracture at a specific site is obtained, an artificial vertebral fracture region segmentation mask is constructed based on the CT image of the fracture at the specific site and the image of a vertebra without a fracture, and a vertebral fracture region segmentation mask is generated based on the vertebral fracture image; training data is constructed based on the artificial vertebral fracture region segmentation mask and the vertebral fracture region segmentation mask, and a 3D U-Net model is semi-supervisedly trained using an uncertainty-aware cross-entropy loss function to obtain a vertebral fracture segmentation model; wherein the amount of training data is greater than the amount of data corresponding to the spinal CT image; the specific site includes limbs and / or ribs.

[0010] In some embodiments, the step of generating a vertebral fracture region segmentation mask based on the vertebral fracture image includes:

[0011] determining the vertebral fracture area of ​​the vertebral fracture image based on resolution visualization technology;

[0012] The vertebral fracture area was scanned along the z-axis, and the scan results were stacked to generate a thermal map image of the fracture area in a three-dimensional spatial structure. The thermal map image of the fracture area was manually verified and annotated to generate a segmentation mask of the vertebral fracture area.

[0013] In some embodiments, the step of constructing an artificial vertebral fracture region segmentation mask based on a CT image of a fracture at a specific location and an image of a vertebra without a fracture comprises:

[0014] Determine a minimum bounding box corresponding to the fracture area in the CT image of the fracture at the specific site, and crop the CT image of the fracture at the specific site based on the minimum bounding box to obtain a segmentation mask of the fracture area at the specific site;

[0015] A non-fractured vertebra segmentation mask is generated according to the non-fractured vertebra image, and an artificial vertebra fracture region segmentation mask is constructed based on the non-fractured vertebra segmentation mask and the fracture region segmentation mask of the specific position.

[0016] In some embodiments, a 3D U-Net model is initially trained based on historical patient spinal CT sample data to predict and generate segmentation mask labels corresponding to each vertebra to implement initial vertebral segmentation. The steps of performing initial vertebral segmentation on the image preprocessed spinal CT image based on the 3D U-Net model to obtain vertebral CT images corresponding to each vertebra include:

[0017] Based on the 3D U-Net model, the initial vertebrae segmentation processing is performed on the spinal CT images after image preprocessing, and the segmentation mask label of each vertebra is extracted;

[0018] According to the segmentation mask label and the spinal CT image, the vertebral CT image corresponding to each vertebra is extracted.

[0019] In some embodiments, the method further includes the steps of initial training and optimization of the 3D U-Net model:

[0020] Acquire spinal CT sample data from multiple historical patients, perform image preprocessing on the spinal CT sample images in the spinal CT sample data, and perform data augmentation processing on the preprocessed spinal CT sample images; wherein the spinal CT sample data includes the spinal CT sample images, labels indicating whether multiple vertebrae have fractures, and true segmentation mask labels corresponding to the multiple vertebrae; the data augmentation processing includes elastic deformation processing, noise addition processing, and geometric change processing;

[0021] A 3D U-Net model was constructed, and the data-augmented spinal CT sample data was used as input. The 3D U-Net model output the predicted segmentation mask labels corresponding to each vertebra. The cross-entropy loss function between the predicted segmentation mask labels and the true segmentation mask labels was calculated. The model parameters were adjusted to minimize the cross-entropy loss function, completing the initial training and optimization of the 3D U-Net model.

[0022] In some embodiments, a 2.5D vertebral fracture classification model is constructed based on a 2.5D convolutional neural network and an LSTM network, and is trained based on vertebral CT sample images of historical patients, segmentation mask labels corresponding to each vertebra, and a label indicating whether a fracture exists. The method further includes the steps of training and optimizing the 2.5D vertebral fracture classification model:

[0023] Acquire M vertebral CT sample images, extract multiple slices of preset resolution along the z-axis from each vertebral CT sample image, and combine each slice with its adjacent slices to form an N-channel vertebral CT image; generate a vertebral mask image based on the segmentation mask label corresponding to each vertebra, and merge the N-channel vertebral CT images with the vertebral mask image on a channel-by-channel basis to form target input data; wherein the vertebral CT sample images include the segmentation mask label corresponding to the vertebra, and at least m sample images include a label indicating whether a fracture is present, where m is a positive integer less than or equal to M;

[0024] A 2.5D vertebral fracture classification model was constructed. The 2.5D vertebral fracture classification model includes an encoder, a reshaping layer, an LSTM network layer, and a head layer. The target input data is fed into the 2.5D vertebral fracture classification model. The encoder extracts feature information from the target input data. The reshaping layer reshapes and adjusts the feature tensor corresponding to the feature information. The LSTM network layer captures contextual information in the z-axis direction to obtain feature representation information corresponding to each sample image. The head layer compresses and classifies the feature representation information to output a label indicating whether a vertebral fracture exists.

[0025] The binary cross entropy loss function was used to optimize the 2.5D vertebral fracture classification model to determine the classification model parameters corresponding to the optimized 2.5D vertebral fracture classification model.

[0026] In some embodiments, after the step of acquiring a spine CT image of the target patient, the method further includes the step of performing image preprocessing on the spine CT image:

[0027] Perform unified processing of the medical image axis direction and the physical scale of the vertebrae on spinal CT images;

[0028] Perform image denoising on spine CT images;

[0029] The pixel values ​​in the spine CT image are normalized based on the maximum and minimum normalization method.

[0030] To solve the above problems, one aspect of an embodiment of the present invention provides a spinal fracture segmentation and training system based on neural network semi-supervised learning, comprising:

[0031] A vertebral segmentation processing module is used to obtain a spinal CT image of a target patient and perform initial vertebral segmentation processing on the spinal CT image based on a 3D U-Net model to obtain a vertebral CT image corresponding to each vertebra;

[0032] A vertebra classification module is used to classify vertebral CT images using a 2.5D vertebral fracture classification model to obtain vertebral fracture images and non-fractured vertebrae images;

[0033] The vertebral fracture segmentation and training module is used to obtain CT images of fractures in specific locations, construct artificial vertebral fracture region segmentation masks based on the CT images of fractures in specific locations and images of vertebrae without fractures, and generate vertebral fracture region segmentation masks based on the vertebral fracture images; construct training data based on the artificial vertebral fracture region segmentation masks and the vertebral fracture region segmentation masks, and use the uncertainty-aware cross-entropy loss function to perform semi-supervised learning training on the 3D U-Net model to obtain a vertebral fracture segmentation model; wherein the amount of training data is greater than the amount of data corresponding to the vertebral CT images; the specific locations include limbs and / or ribs.

[0034] In order to solve the above problems, one aspect of an embodiment of the present invention provides an electronic device, including: a processor, and a memory for storing a program, the program including instructions, which, when executed by the processor, enable the processor to execute any of the above-mentioned spinal fracture segmentation and training methods based on neural network semi-supervised learning.

[0035] In order to solve the above problems, one aspect of an embodiment of the present invention provides a non-transitory machine-readable medium storing computer instructions, where the computer instructions are used to enable a computer to execute any of the above-mentioned spinal fracture segmentation and training methods based on neural network semi-supervised learning.

[0036] The beneficial effects of the embodiments of the present invention are as follows: by obtaining a spinal CT image of a target patient, performing initial vertebral segmentation processing on the spinal CT image based on a 3D U-Net model to obtain vertebral CT images corresponding to each vertebra; using a 2.5D vertebral fracture classification model to classify the vertebral CT image to obtain a vertebral fracture image and a vertebral image without fracture; obtaining a fracture CT image of a specific part, constructing an artificial vertebral fracture region segmentation mask based on the fracture CT image of the specific part and the vertebral image without fracture, and generating a vertebral fracture region segmentation mask based on the vertebral fracture image; constructing training data based on the artificial vertebral fracture region segmentation mask and the vertebral fracture region segmentation mask, using an uncertainty-aware cross-entropy loss function to perform semi-supervised learning training on the 3DU-Net model to obtain a spinal fracture segmentation model; wherein the amount of training data is greater than the amount of data corresponding to the spinal CT image; the technical means of specific parts including limbs and / or ribs overcomes the large deviation in diagnostic results of traditional diagnostic methods in related technologies, and the large amount of data required for training the scheme based on the 3D convolutional neural network, the problems of difficult data acquisition, high labeling cost, and low accuracy of fracture region segmentation, thereby realizing the method of segmenting the vertebral fracture through 3D convolutional neural network. The U-Net model obtains the vertebral CT images corresponding to each vertebra, realizes effective and fast initial segmentation of the vertebrae, reduces the interference of other vertebrae on subsequent classification, and enables the subsequent vertebral classification model to focus on the classification of the target vertebrae, thereby improving the accuracy and precision of classification detection; and then uses the 2.5D vertebral fracture classification model constructed based on the 2.5D convolutional neural network and LSTM network to classify the vertebral CT images. This vertebral classification model efficiently integrates two-dimensional spatial information and depth-direction contextual information, avoids the high computational cost of 3D convolution, and can also use the pre-trained weights of two-dimensional convolution to reduce training time while The accuracy of fracture classification is further improved; finally, artificial vertebral fracture data is constructed using fracture images of specific parts, which overcomes the influence of less training data and difficulty in data acquisition in related technologies. At the same time, semi-supervised learning training is used to enhance the vertebral fracture segmentation model's ability to recognize vertebral fracture features. Relying on some annotated data of vertebral fracture images of real patients combined with annotated data of fracture images of some specific parts (such as limbs and ribs), the training data required for model training is constructed, which can realize the training of the vertebral fracture segmentation model, reduce the difficulty of data acquisition, and achieve the technical effect of significantly improving the accuracy of vertebral fracture segmentation under the condition of a limited number of samples.

[0037] The details of one or more embodiments of the invention are set forth in the accompanying drawings and the description below so that other features, objects, and advantages of the invention are more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be derived from these drawings without inventive effort.

[0039] Figure 1 This is a schematic diagram of the main process of a spinal fracture segmentation and training method based on neural network semi-supervised learning according to an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of the main flow of a spinal fracture segmentation and training method based on neural network semi-supervised learning according to another embodiment of the present invention.

[0041] Figure 3 3D U-Net model according to an embodiment of the present invention.

[0042] Figure 4 Schematic diagram of the architecture of a 2.5D vertebral fracture classification model provided by one embodiment of the present invention.

[0043] Figure 5 It is a schematic diagram of the main framework of a spinal fracture segmentation and training system based on neural network semi-supervised learning according to an embodiment of the present invention.

[0044] Figure 6 It is a structural schematic diagram of the electronic device of the present invention. DETAILED DESCRIPTION

[0045] Embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0046] After a patient with a spinal fracture presents to the doctor, quickly and accurately determining the location of the fractured vertebra is a prerequisite for subsequent treatment. Traditional spinal fracture diagnosis relies primarily on the physician's subjective judgment of CT images. This process requires the physician to review each CT image slice, requiring both expertise and experience and taking considerable time. Furthermore, the subjective judgment can lead to significant deviations in the diagnostic results.

[0047] In view of the problems existing in traditional subjective diagnostic schemes, the application of deep learning technology has gradually become a research hotspot in the field of medical image analysis, especially in the diagnosis and classification of spinal fractures. In order to improve the efficiency and accuracy of diagnosis in related technologies, deep learning technology based on 3D (3 Dimensions, three-dimensional) convolutional neural network (CNN) is usually adopted. However, this type of method itself has the following challenges: (1) The reasoning time required for 3D CNN is long and the training data required is large, resulting in low detection efficiency and difficulty in data acquisition; (2) 3D CNN cannot effectively utilize the pre-trained weights of two-dimensional convolution, resulting in high computational complexity and slow model training speed. (3) Algorithm reasoning can only provide a conclusion on whether there is a fracture or not, and cannot provide more detailed visual information to assist experts or labelers in better locating the fracture type and area. In addition, due to the specificity of spinal fractures in pathological characteristics and the lack of accurate labeled data of spinal fracture areas, the accuracy of existing fracture segmentation models is poor. At the same time, related technologies focus on binary discrimination of the existence of fractures, which makes it difficult to achieve accurate segmentation of the fracture area and also affects the formulation of subsequent treatment plans.

[0048] In order to solve the above problems, the embodiment of the present invention provides a vertebral fracture segmentation and training method based on neural network semi-supervised learning, such as Figure 1 As shown in FIG, the vertebral fracture segmentation and training method based on neural network semi-supervised learning mainly includes:

[0049] Step S101 : Acquire a spinal CT image of a target patient, and perform initial vertebra segmentation processing on the spinal CT image based on a 3D U-Net model to obtain a vertebral CT image corresponding to each vertebra.

[0050] Among them, the above-mentioned CT (Computed Tomography) uses precisely collimated X-ray beams, gamma rays, ultrasound waves, etc., together with extremely sensitive detectors to perform cross-sectional scans around a certain part of the human body one by one. It has the characteristics of fast scanning time and clear images, and can be used to detect various diseases.

[0051] In some embodiments, after the step of acquiring the spinal CT image of the target patient, the above method further includes the step of performing image preprocessing on the spinal CT image: performing uniform processing on the medical image axis direction and the physical scale of the spinal vertebrae on the spinal CT image; performing image denoising on the spinal CT image; and normalizing the pixel values ​​in the spinal CT image based on the maximum and minimum normalization method.

[0052] Based on the above settings, before using the target patient's spinal CT images for vertebral fracture detection, image preprocessing is performed on the spinal CT images, which helps to enhance image features and unify the image format, thereby improving the generalization ability and classification accuracy of the subsequent vertebral segmentation model and fracture classification model, reducing the misdiagnosis rate, and improving the efficiency of fracture segmentation.

[0053] It should be noted that the target patient referred to in this embodiment of the present invention is a historical patient identified in the spinal CT sample data used for initial training of the 3D U-Net model. The target patient's spinal CT image is part of the raw data used to subsequently construct the spinal fracture segmentation model.

[0054] According to embodiments of the present invention, the aforementioned axial alignment of spinal CT images is performed to correct the orientation of spinal CT images, ensuring that all CT images are aligned in the same direction. Specifically, image processing techniques can be used to adjust the orientation of CT images to conform to a unified axial standard. This image preprocessing step ensures spatial consistency between the target patient's CT images and the sample CT images of the patient used in model training, eliminating deviations caused by inconsistent image orientation.

[0055] According to an embodiment of the present invention, the above-mentioned unified processing of the physical scale of the spinal vertebrae of the spinal CT images is to ensure that the CT images from different sources (including CT images of the patient to be tested and CT sample images of historical patients used for model training) are consistent in physical scale. According to a specific implementation of an embodiment of the present invention, all CT images can be scaled according to the physical voxel size of [1, 1, 1]. This means that the size of each voxel is unified to 1mm×1mm×1mm in three dimensions. This image preprocessing step helps to eliminate size differences caused by different devices or scanning parameters, ensures that the model can accurately process data from different sources, and helps subsequent models to more accurately identify and locate vertebral structures.

[0056] According to embodiments of the present invention, the aforementioned image denoising process for spinal CT images can effectively remove noise, such as Gaussian noise and salt-and-pepper noise, from the images, thereby improving image clarity and detail. By removing noise, the model can more accurately analyze image content, thereby reducing the misdiagnosis rate caused by noise. The denoised images can also more clearly display the structure of the vertebrae and potential fracture areas, helping the model to more accurately extract features. According to a specific implementation of an embodiment of the present invention, in addition to image denoising methods used in related arts, noise removal can also be achieved by limiting the pixel value range. For example, since the pixel values ​​of CT images generally represent the degree of tissue attenuation of X-rays, and different tissues (such as air, soft tissue, and bone) have different attenuation values, the CT data pixel range can be limited to [-200, 1000] based on the tissue attenuation of X-rays. By limiting the pixel value range, irrelevant background information (such as low-value pixels in air areas) can be removed while retaining diagnostically relevant tissue information (such as high-value pixels in bone). This step helps improve the robustness of the model and detection accuracy.

[0057] According to an embodiment of the present invention, the above-mentioned normalization processing of the pixel values ​​in the spinal CT image based on the maximum and minimum normalization method can unify the pixel values ​​of the image into a fixed range (such as [0, 1]), which helps the subsequent model to learn and process image data more stably; at the same time, the normalization processing can avoid the gradient disappearance or gradient explosion problems caused by the pixel value range being too large or too small, which helps to improve the training efficiency of the model; in addition, the normalized image can enable the model to learn the features in the image more effectively, thereby improving the overall performance of the model. According to a specific implementation method of an embodiment of the present invention, the maximum and minimum normalization method is used to normalize the pixel values ​​of the CT image to the range of [0, 1]. The specific expression is:

[0058] Normalized_pixel=(pixel-min_pixel) / (max_pixel-min_pixel)

[0059] Where pixel is the original pixel value, min_pixel and max_pixel are the minimum and maximum pixel values ​​in the CT image, respectively. Normalization ensures that all data features are on the same scale.

[0060] Based on the above image preprocessing steps, the quality and consistency of images can be significantly improved. For the model training step, image preprocessing improves the model's generalization and feature capabilities and reduces the misdiagnosis rate. For the detection of spinal CT images of target patients, image preprocessing helps eliminate noise and differences in the data and enhances the subsequent model's ability to recognize fracture features, thereby improving the accuracy and efficiency of classification.

[0061] The 3D U-Net model provided in the embodiment of the present invention is initially trained based on spinal CT sample data of historical patients, and is used to predict and generate segmentation mask labels corresponding to each vertebra to achieve initial vertebrae segmentation processing.

[0062] Specifically, the segmentation mask label provided by an embodiment of the present invention is a pixel-level annotation data used to indicate whether each pixel or voxel in a medical image (such as CT or MRI) belongs to a specific anatomical structure or lesion area. In 3D medical image segmentation tasks, the segmentation mask label is usually a binary (or multi-valued) mask with the same spatial dimension as the original image, where: the background area is usually represented by 0; the target area (such as vertebrae, tumors, etc.) is represented by 1 or other integers. For example, in the task of spinal fracture detection and diagnosis, the segmentation mask label will clearly mark the boundaries and positions of each vertebra to help the model learn the morphology and spatial distribution of the vertebrae.

[0063] Furthermore, unlike the traditional 2D U-Net model, the 3D U-Net model can directly process three-dimensional data, preserving the spatial information of the image. The 3D U-Net model is also more efficient when processing block-like images, resolving the efficiency issues of traditional 2D models when processing 3D data. This enables the model to complete vertebrae segmentation tasks more quickly, improving overall fracture detection efficiency. Through this setup, the 3D U-Net model can improve the accuracy of subsequent vertebrae classification by quickly and accurately segmenting vertebrae regions in CT images.

[0064] In some embodiments, the step of performing initial vertebrae segmentation processing on the image-preprocessed spinal CT image based on the 3D U-Net model to obtain vertebrae CT images corresponding to each vertebra includes: performing initial vertebrae segmentation processing on the image-preprocessed spinal CT image based on the 3D U-Net model to extract a segmentation mask label for each vertebra; and extracting a vertebrae CT image corresponding to each vertebra based on the segmentation mask label and the spinal CT image.

[0065] The 3D U-Net model provided by the embodiments of the present invention, through its unique encoder-decoder structure and skip connections, can effectively extract features from spinal CT images and generate high-precision vertebral segmentation masks. Based on this high-precision segmentation, individual vertebrae can be clearly separated, providing accurate regional information for subsequent vertebral fracture classification. Specifically, the encoder in the 3DU-Net model uses convolution and downsampling layers to gradually extract multi-scale features and compress spatial information. The decoder uses upsampling and convolution layers to gradually reconstruct spatial resolution. At the same time, skip connections are used to fuse features from each encoder stage, ultimately achieving accurate segmentation prediction.

[0066] In some embodiments, the above method also includes the initial training and optimization steps of the 3D U-Net model: obtaining spinal CT sample data of multiple historical patients, performing image preprocessing on the spinal CT sample images in the spinal CT sample data, and performing data enhancement processing on the spinal CT sample images that have undergone image preprocessing; wherein the spinal CT sample data includes spinal CT sample images, labels of whether multiple vertebrae have fractures, and true segmentation mask labels corresponding to multiple vertebrae; data augmentation processing includes elastic deformation processing, noise addition processing, and geometric change processing; constructing a 3D U-Net model, and using the spinal CT sample data after data enhancement processing as input, outputting predicted segmentation mask labels corresponding to each vertebra based on the 3D U-Net model, calculating the cross-entropy loss function between the predicted segmentation mask labels and the true segmentation mask labels, adjusting the model parameters with the purpose of minimizing the cross-entropy loss function, and completing the initial training and optimization of the 3D U-Net model.

[0067] The historical patient spinal CT sample data provided by the present invention includes the following: the patient's spinal CT image, which is the primary data source for analyzing vertebral fractures; a vertebral fracture label, which indicates whether a vertebra is fractured; and a corresponding segmentation mask label, which identifies the specific region of the vertebra. The fracture label and segmentation mask label do not need to be included for every vertebra in the patient's spinal CT image; only a portion of them can be used to complete the initial training and optimization of the 3D U-Net model.

[0068] Given the limited availability of real-world annotated data, data augmentation can be performed on pre-processed spinal CT sample images to increase data diversity and richness, thereby improving the range of the 3D U-Net model. Specifically, elastic deformation processing involves applying random elastic deformations to CT images to simulate different vertebral variations; noise addition processing involves adding random noise to CT images to enhance the model's robustness to noise; and geometric transformation processing, including operations such as rotation, scaling, and translation, can simulate vertebral images from different perspectives and positions.

[0069] Based on the above settings, data augmentation techniques such as elastic deformation, noise addition, and geometric changes enable the model to learn more diverse image features, thereby improving its adaptability to different patients and imaging conditions. Data augmentation also increases the diversity of training data and reduces overfitting of the model to the training data. Furthermore, by simulating different image changes, the model can more accurately identify vertebral boundaries, thereby improving segmentation accuracy.

[0070] During 3D U-Net model optimization, the number of background and target (vertebrae) pixels is often unbalanced during vertebrae segmentation. By calculating the cross-entropy loss function between the predicted segmentation mask label and the true segmentation mask label and optimizing it with the goal of minimizing this cross-entropy loss function, and updating the model weights through backpropagation, the model parameters are tuned. This enables the 3D U-Net model to more accurately learn the segmentation features of the vertebrae, thereby helping the model achieve better segmentation results in the case of unbalanced data.

[0071] In some embodiments, after performing vertebra segmentation processing on the image pre-processed spinal CT image based on the 3D U-Net model to obtain vertebral CT images corresponding to each vertebra, the above method further includes: performing morphological dilation processing on the vertebral CT image to fill holes and gaps in the segmentation result; and performing noise point removal processing on the vertebral CT image to eliminate noise points around the vertebral area.

[0072] Morphological dilation effectively fills holes and gaps in the segmentation results, enhancing the continuity of the vertebral region and making it more complete, reducing subsequent misclassifications due to incomplete segmentation. Noise point removal effectively removes noise points around the vertebral region, improving the purity of the segmentation results and reducing misclassifications due to noise. These operations help further improve the accuracy of vertebral fracture classification.

[0073] Step S102 : using the 2.5D vertebral fracture classification model to classify the vertebral CT image to obtain vertebral fracture images and non-fracture vertebral images.

[0074] Among them, the 2.5D vertebral fracture classification model provided by the embodiment of the present invention is constructed based on a 2.5D convolutional neural network and an LSTM network, and is trained based on vertebral CT sample images of historical patients, the segmentation mask labels corresponding to each vertebra, and the label of whether a fracture exists.

[0075] Furthermore, the 2.5D vertebral fracture classification model provided by an embodiment of the present invention combines a 2.5D convolutional neural network (CNN) and a long short-term memory network (LSTM). This vertebral classification model efficiently integrates two-dimensional spatial information with depth-direction contextual information, avoiding the high computational cost of 3D convolution. It can also utilize the pre-trained weights of two-dimensional convolution to further improve the accuracy of fracture classification detection while reducing training time. Based on the 2.5D vertebral fracture classification model, it can quickly process vertebral CT images and output a prediction result of whether there is a fracture (including vertebral fracture images and non-fracture vertebral images). This automated classification process significantly improves the accuracy of vertebral classification while reducing the efficiency of vertebral classification.

[0076] According to an embodiment of the present invention, the vertebral CT sample images of historical patients used in training the aforementioned 2.5D vertebral fracture classification model can be obtained by segmenting the vertebral CT sample images of historical patients using the aforementioned 3D U-Net model. Furthermore, for vertebral CT sample images with a small number of annotations (i.e., the number of segmentation mask labels corresponding to vertebrae and the number of labels indicating the presence of fractures in the vertebral CT images), data augmentation techniques can be used to process the vertebral CT sample images to increase the diversity of the training data and reduce overfitting of the model to the training data.

[0077] In some embodiments, the method further includes the steps of training and optimizing a 2.5D vertebral fracture classification model: obtaining M vertebral CT sample images, extracting multiple slices of preset resolution along the z-axis for each vertebral CT sample image, and combining each slice with its adjacent slices to form an N-channel vertebral CT image; generating a vertebral mask image based on the segmentation mask label corresponding to each vertebra, and merging the N-channel vertebral CT image and the vertebral mask image by channel to form target input data; wherein the vertebral CT sample images include the segmentation mask label corresponding to the vertebra, and at least m sample images include a label indicating whether a fracture exists, where m is a positive integer less than or equal to M; constructing a 2.5D vertebral fracture classification model, 2 The .5D vertebral fracture classification model includes an encoder, a reshaping layer, an LSTM network layer, and a Head layer. The target input data is input into the 2.5D vertebral fracture classification model. The encoder extracts the feature information in the target input data, and the reshaping layer reshapes and adjusts the feature tensor corresponding to the feature information. The LSTM network layer is used to capture the context information in the z-axis direction to obtain the feature representation information corresponding to each sample image. The feature representation information is compressed and classified by the Head layer to output a label indicating whether the vertebra has a fracture. The binary cross entropy loss function is used to optimize the 2.5D vertebral fracture classification model to determine the classification model parameters corresponding to the optimized 2.5D vertebral fracture classification model.

[0078] The z-axis provided in the embodiments of the present invention refers to the z-axis in medical images, i.e., the direction along the long axis of the human body (from head to toe). Specifically, the direction of the z-axis coincides with the vertical axis of the human body and is typically used to indicate the depth direction of an image slice. In CT or MRI scans, the z-axis typically coincides with the direction of motion of the scanning bed, i.e., from toe to head.

[0079] The 2.5D vertebral fracture classification model provided by this embodiment includes an encoder, a reshaping layer, an LSTM network layer, and a head layer. Specifically, the encoder extracts feature information, the reshaping layer adjusts the shape of the feature tensor, the LSTM network layer captures contextual information in the z-axis direction, and the head layer performs feature compression and classification.

[0080] By training with historical patient vertebral CT sample images, segmentation mask labels, and fracture presence / absence labels, the 2.5D vertebral fracture classification model is able to learn characteristics from different patients and imaging conditions. This diverse training data helps improve the model's generalization ability, enabling it to maintain high performance even when faced with new, unseen data. During the model optimization process, the 2.5D vertebral fracture classification model is optimized using a binary cross-entropy loss function. By minimizing the binary cross-entropy loss function, the classification model parameters are adjusted to ensure that the model can accurately output the label of whether a vertebral fracture exists.

[0081] Step S103: Acquire a CT image of a fracture at a specific site, construct an artificial vertebral fracture region segmentation mask based on the CT image of the fracture at the specific site and the image of a vertebra without a fracture, and generate a vertebral fracture region segmentation mask based on the vertebral fracture image; construct training data based on the artificial vertebral fracture region segmentation mask and the vertebral fracture region segmentation mask, and use an uncertainty-aware cross-entropy loss function to perform semi-supervised learning training on a 3D U-Net model to obtain a vertebral fracture segmentation model; wherein the amount of training data is greater than the amount of data corresponding to the spinal CT image; the specific site includes limbs and / or ribs.

[0082] The vertebral fracture segmentation model trained in the embodiment of the present invention is used to perform vertebral fracture segmentation operations on vertebral CT images of real patients to assist in the diagnosis and classification of vertebral fractures.

[0083] Through the above settings, vertebral fracture region segmentation masks are generated based on vertebral fracture images to achieve effective classification and labeling of vertebral fracture regions. Combined with artificial vertebral fracture region segmentation masks constructed based on fracture images of specific locations (such as common fracture locations such as limbs and ribs, where there is a large amount of existing labeled data for these common fractures), artificial vertebral fracture data can be constructed using fracture images of specific locations, overcoming the impact of the limited training data and data acquisition difficulties in related technologies. Semi-supervised learning training is then used to enhance the vertebral fracture segmentation model's ability to recognize vertebral fracture features. The vertebral fracture segmentation model can be trained with only a small amount of patient spinal CT image data, achieving the technical effect of significantly improving vertebral fracture segmentation accuracy under the condition of a limited number of samples. Furthermore, by adopting a semi-supervised learning strategy, not only is the model's ability to recognize fracture features enhanced, but the generalization of the constructed vertebral fracture segmentation model is also significantly improved, as well as the segmentation accuracy of the vertebral fracture segmentation model for actual vertebral fracture cases.

[0084] In some embodiments, the step of generating a vertebral fracture region segmentation mask based on a vertebral fracture image includes: determining the vertebral fracture region of the vertebral fracture image based on resolution visualization technology; scanning the vertebral fracture region along the z-axis, and stacking the scanning results to generate a thermal map image of the fracture region under a three-dimensional spatial structure, manually verifying the thermal map image of the fracture region, and annotating and generating a vertebral fracture region segmentation mask.

[0085] The resolution visualization technology provided by the embodiments of the present invention can be HiResCAM (High-Resolution Class Activation Mapping), a visualization technology for deep convolutional neural networks that aims to provide high-resolution heat maps that are more faithful to the model's decision-making process. By directly multiplying the feature map by the gradient element-by-element, it avoids the positioning bias caused by gradient averaging in traditional methods, significantly improving the model's interpretation reliability and positioning accuracy for fracture area identification.

[0086] Through the above settings, high-resolution visualization technology (such as HiResCAM technology) can provide high-resolution feature maps by retaining more spatial details, which can more accurately locate important areas in the image, thereby more clearly identifying the specific location and range of vertebral fractures, and improving the accuracy of fracture area identification. Specifically, by scanning the vertebral fracture area along the z-axis and stacking the scan results, a thermal image of the fracture area in a three-dimensional spatial structure is generated. This three-dimensional thermal image can intuitively display the spatial distribution of the fracture area in the vertebrae, and also help doctors more comprehensively understand the severity and location of the fracture.

[0087] In some of the embodiments, the step of constructing an artificial vertebral fracture area segmentation mask based on the fracture CT image of the specific part and the vertebral image without fracture includes: determining the minimum bounding box corresponding to the fracture area in the fracture CT image of the specific part, and cutting out the fracture area segmentation mask of the specific part from the fracture CT image of the specific part based on the minimum bounding box; generating a non-fracture vertebral segmentation mask based on the non-fracture vertebral image, and constructing an artificial vertebral fracture area segmentation mask based on the non-fracture vertebral segmentation mask and the fracture area segmentation mask of the specific part.

[0088] The above steps provide a specific implementation method for constructing an artificial vertebral fracture region segmentation mask based on CT images of fractures at specific locations. Specifically, taking rib fracture CT images as an example, rib fracture CT images can be obtained from an existing publicly available rib fracture dataset. For each fracture region in each rib fracture CT image, a minimum bounding box is found, and the rib fracture region is cropped according to the size of the bounding box. The non-fractured vertebrae images obtained in the above steps are used to generate a non-fractured vertebrae segmentation mask. This non-fractured vertebrae segmentation mask is then overlapped with the cropped specific fracture region segmentation mask through random region matching. The overlapping portion constitutes the artificial vertebral fracture region segmentation mask.

[0089] An embodiment of the present invention also provides an embodiment for constructing an artificial vertebral CT image based on an artificial vertebral fracture region segmentation mask, comprising: filling a normal vertebral CT image according to the artificial vertebral fracture region segmentation mask to construct an artificial vertebral fracture CT image, wherein the filling method can be randomly selected from "Gaussian noise," "tissue pixel value," "overlapping region pixel value mean," or "overlapping region pixel value median." Based on this embodiment, the vertebral fracture CT images that can be effectively applied to vertebral fracture segmentation model training can be expanded.

[0090] In some embodiments, in order to improve the model's ability to learn detailed features in vertebral fracture area segmentation, the weights of the 3D U-Net model applied in step S101 and initially trained are used, its encoder structure is fixed, and only the decoder part is trained for parameter optimization.

[0091] In order to prevent the 3D U-Net model from learning only simple artificial vertebral fracture features during the training process (the training data are the artificial vertebral fracture region segmentation mask R and the real vertebral fracture region segmentation mask S (i.e., the vertebral fracture region segmentation mask generated according to the vertebral fracture image)), and to improve its segmentation accuracy in scenarios with uneven data quality, the embodiment of the present invention uses uncertainty-aware cross-entropy L u-ce Perform semi-supervised learning. Among them, the uncertainty-aware cross entropy L u-ce The expression is as follows:

[0092]

[0093] w n =(1+σ n ) α

[0094] Among them, L u-ce is the uncertainty-aware cross entropy loss for a single image, N is the number of pixels in the image, C is the number of categories, and yn,c is the corresponding segmentation mask, p n,c is the corresponding predicted softmax probability. n is the pixel-level uncertainty weight, where σ n represents the prediction uncertainty, and α controls the influence of uncertainty in an exponential manner.

[0095] Prediction uncertainty α represents the standard deviation of the softmax probability of the predicted category in the segmented sample. n When is 1, the uncertainty-aware cross entropy loss degenerates into a conventional cross entropy loss. Therefore, for the artificial vertebral fracture region segmentation mask R and the real vertebral fracture region segmentation mask S, in specific applications, the expression of the cross entropy loss function is as follows:

[0096]

[0097] The above-mentioned vertebral fracture segmentation and training method based on neural network semi-supervised learning provided by the embodiment of the present invention adopts the method of obtaining a vertebral CT image of a target patient, performing initial vertebral segmentation processing on the vertebral CT image based on a 3D U-Net model to obtain vertebral CT images corresponding to each vertebra; using a 2.5D vertebral fracture classification model to classify the vertebral CT image to obtain vertebral fracture images and vertebral images without fractures; obtaining a CT image of a fracture at a specific location, constructing an artificial vertebral fracture region segmentation mask based on the CT image of the fracture at the specific location and the vertebral image without fractures, and generating a vertebral fracture region segmentation mask based on the vertebral fracture image; constructing training data based on the artificial vertebral fracture region segmentation mask and the vertebral fracture region segmentation mask, and using an uncertainty-aware cross-entropy loss function to classify the 3D The U-Net model is trained for semi-supervised learning to obtain a vertebral fracture segmentation model; wherein, the amount of training data is greater than the amount of data corresponding to the spinal CT image; the technical means of specific parts including limbs and / or ribs overcomes the large deviation in diagnostic results of traditional diagnostic methods in related technologies, while the training of the scheme based on 3D convolutional neural network requires a large amount of labeled data for vertebral fracture images of real patients, which has the problems of difficulty in data acquisition and low accuracy of fracture area segmentation. The vertebral CT images corresponding to each vertebra are obtained through the 3D U-Net model, and effective and fast initial segmentation processing of the vertebra is achieved, which reduces the interference of other vertebrae on subsequent classification, so that the subsequent vertebral classification model can focus on the classification of the target vertebra, thereby improving the accuracy and precision of classification detection; and then the vertebral CT images are classified using a 2.5D vertebral fracture classification model constructed based on a 2.5D convolutional neural network and an LSTM network. The vertebral classification model efficiently integrates two-dimensional spatial information and depth-direction contextual information, avoids the high computational cost of 3D convolution, and can also use the pre-trained weights of two-dimensional convolution to reduce training time while achieving Further improve the accuracy of fracture classification; finally, use fracture images of specific parts to construct artificial vertebral fracture data, overcoming the influence of less training data and difficulty in data acquisition in related technologies, and at the same time use semi-supervised learning training to enhance the vertebral fracture segmentation model's ability to recognize vertebral fracture features. Relying on the annotated data of vertebral fracture images of some real patients combined with the annotated data of fracture images of specific parts (such as limbs and ribs) to construct the training data required for model training, the training of the vertebral fracture segmentation model can be realized, which reduces the difficulty of data acquisition and achieves the technical effect of significantly improving the accuracy of vertebral fracture segmentation under the condition of a limited number of real patient vertebral fracture samples.

[0098] The embodiment of the present invention also provides a vertebral fracture segmentation and training method based on neural network semi-supervised learning, which involves optimizing the training of a vertebral fracture segmentation model based on a small amount of historical patient vertebral CT sample data, specifically including initial vertebral segmentation processing based on vertebral CT images, vertebral fracture detection and fracture area visualization for target patients, construction of an artificial vertebral fracture dataset, and training of a vertebral fracture segmentation model. Figure 2 As shown, the spinal fracture segmentation and training method based on neural network semi-supervised learning provided by the embodiment of the present invention mainly includes:

[0099] Step S210 , performing image preprocessing on the spinal CT sample data of historical patients and the spinal CT images of target patients; wherein the spinal CT sample data includes the patient's spinal CT images, 14,811 labels indicating whether each vertebra has a fracture, and segmentation mask labels corresponding to each vertebra of 87 patients.

[0100] The spinal CT sample data of historical patients are image preprocessed for model training and optimization, and the spinal CT images of target patients are image preprocessed for performing current spinal fracture detection.

[0101] Specifically, the image preprocessing steps in the above steps include:

[0102] S211: Directionally correct the CT images to ensure that all images are aligned in the same direction, that is, the axial directions of the medical images are unified.

[0103] S212: Scale all images according to the [1, 1, 1] physical voxel size to ensure consistency in the physical scale of vertebrae from different image sources.

[0104] S213: According to the attenuation degree of the tissue to the X-ray, the pixel range of the CT data is limited to [-200, 1000].

[0105] S214: Use the maximum and minimum normalization technology to normalize the pixel values ​​in the CT image to ensure the consistency and comparability of the data feature distribution and avoid the gradient disappearance phenomenon in the subsequent model learning process.

[0106] In step S220, a 3D U-Net model is constructed and initially trained based on the spinal CT images after image preprocessing and the true segmentation mask labels corresponding to each vertebra. The 3D U-Net model that has completed the initial training is then used to perform inference prediction on other unlabeled spinal CT images to obtain the 3D segmentation mask labels of all vertebrae.

[0107] Since there is little real label data for the segmentation mask corresponding to each vertebra of the patient, the present invention adopts a 3D U-Net model (a deep learning model based on the U-Net architecture, specifically for three-dimensional medical image segmentation tasks. It replaces the two-dimensional operations in U-Net with three-dimensional operations, can process volumetric data (such as CT or MRI images), and generate dense three-dimensional segmentation masks) as the basic network.

[0108] S221: For the data processed in step S210, the richness of the data is increased through data enhancement methods such as elastic deformation, noise addition, and geometric change.

[0109] S222: Build a 3D U-Net model, which consists of an encoder and a decoder. The encoder gradually extracts high-level image features and reduces spatial dimensions through convolution and downsampling operations. The decoder gradually restores the image's spatial resolution through upsampling and convolution operations, and combines the encoder's features for segmentation prediction.

[0110] S223: Use the data generated by the enhancement method as the model data input to perform initial training on the model, calculate the cross entropy loss function (Cross Entropy Loss) between the model's predicted mask labels and the true mask labels, update the model parameters through backpropagation, and obtain the optimal training model through tuning.

[0111] S224: After fixing the model parameters, the unlabeled vertebral CT image is used for inference prediction, and the predicted segmentation mask label is output. The vertebral CT image corresponding to each vertebra is extracted based on the predicted segmentation mask label. Then, a morphological dilation operation is performed on the vertebral CT image to fill the holes and gaps, and eliminate the noise points around the vertebral area, thereby improving the accuracy and robustness of the segmentation result. According to a specific implementation of an embodiment of the present invention, the architecture diagram of the above-mentioned 3D UNet model is as follows: Figure 3 As shown, the 3D UNet model consists of an encoder and a decoder. Specifically,

[0112] The encoder extracts features and downsamples the input 3D image through a series of 3D convolutional layers and maximum pooling layers, gradually reducing the spatial dimension while extracting more abstract features. In each encoder layer, two 3D convolution operations (with a kernel size of 3×3×3) are performed to extract local features. Then, a 2×2×2 maximum pooling operation is performed to halve the spatial size of the feature map, thereby extracting features at a coarser scale.

[0113] The decoder uses upsampling (using 3×3×3 transposed convolutions) and skip connections to gradually restore the spatial resolution of the image and fuse the high-resolution features from the encoder stage with those from the decoder stage to fully utilize contextual information and details. In each decoder layer, upsampling is first performed to expand the spatial size of the feature map, which is then concatenated with the feature map from the corresponding encoder layer (i.e., skip connections, which bridge the semantic gap between the encoder and decoder). The fused features are then processed through two 3D convolution operations. Finally, the segmentation result is output through a 3D convolution layer in the last layer. The number of output channels is usually consistent with the number of categories of the segmentation target.

[0114] Step S230: Construct a 2.5D vertebral fracture classification model including a long short-term memory network. Use the CT image, the 3D segmentation mask label obtained by the above inference prediction, and the label of whether each vertebra has a fracture to train and optimize the model. Based on the trained and optimized 2.5D vertebral fracture classification model, detect whether the corresponding vertebra in the target patient's spinal CT image has a fracture.

[0115] S231: Use the 3D U-Net model of step S220 to perform inference prediction on the spinal CT images of historical patients to obtain segmentation mask labels of all vertebrae.

[0116] S232: Locate the coordinate position of each vertebra in the CT sample image by segmenting the mask label, cut and separate the single vertebra from the entire spine CT sample image, and obtain the corresponding vertebra CT sample image I_CT.

[0117] S233: For each vertebral CT sample image I_CT, 15 224×224 slices are uniformly extracted along the z-axis. Each slice, along with its two adjacent slices, forms a 5-channel vertebral CT image to provide spatial information, denoted as I_(CT_slice). The vertebral mask image corresponding to the predicted vertebral segmentation mask label is also used as the sixth channel, I_mask. Using the vertebral mask image effectively reduces the influence of multiple vertebrae in a single sample, allowing classification to focus more on the target vertebra. Finally, the 5-channel vertebral CT image and the vertebral mask image are merged channel by channel, so that each vertebra constitutes target input data with a shape of 15×6×224×224, and the presence or absence of a fracture is the label of that vertebra.

[0118] S234: Construct a 2.5D-based vertebral fracture classification model, which includes an encoder, a reshape layer, an LSTM layer, and a head layer. The encoder can use ResNet18 (which consists of an 18-layer convolutional neural network and includes multiple residual modules. Each module passes the input directly to the subsequent layer through a skip connection, allowing the model to learn the residual mapping between input and output, thereby improving the model's training efficiency and performance. The reshape layer reshapes the input tensor to a specified shape, meeting the input requirements of subsequent model layers without changing its data content. The LSTM layer is mainly composed of three gating mechanisms: a forget gate, an input gate, and an output gate. It uses its modeling ability for sequential data to capture contextual information in the z-axis direction, thereby capturing the continuity and correlation of the image in the depth direction, further improving the model's accuracy in judging vertebral fractures. The head layer uses a multilayer perceptron (MLP) and a dropout layer to compress feature information for classification. The head layer is stacked in the form of an input layer, a dropout layer, an intermediate layer, a dropout layer, and an output layer. The dropout rate of the dropout layer is 0.2 during training.

[0119] S235: A binary cross entropy loss function is used to train and optimize a 2.5D vertebral fracture classification model. The trained and optimized 2.5D vertebral fracture classification model is then used to classify and detect vertebral CT images of target patients to detect vertebral CT images with fractures.

[0120] According to a specific implementation of an embodiment of the present invention, the schematic diagram of the architecture of the above-mentioned 2.5D vertebral fracture classification model is as follows: Figure 4 As shown in the figure, the input image shape is BS×15×6×224×224, where BS represents the batch size. This 2.5D vertebral fracture classification model uses ResNet18 as the encoder structure to extract sample features. After a reshape operation, the feature information is processed to form a feature vector with a shape of [BS×15,1280]. To capture the spatial information of the vertebra along the z-axis, the model stacks a long short-term memory network (LSTM) to extract contextual information, thereby better assisting the model in determining the fracture location. Finally, the classification task is completed by reshaping the features and using a multi-layer perceptron (MLP) as the head layer.

[0121] In step S240 , the model parameters of the optimized 2.5D vertebral fracture classification model are fixed, and the potential fracture area is visualized using HiResCAM high-resolution visualization technology and displayed in the form of a heat map to effectively assist doctors and data labelers in determining the potential fracture area of ​​the vertebra.

[0122] S241: Using the 2.5D vertebral fracture classification model trained in step S230, and fixing its classification model parameters.

[0123] S242: HiResCAM technology is used to generate high-resolution class activation maps that are faithful to the model's decision process, indicating possible vertebral fracture areas in each planar CT image. Compared to traditional techniques, HiResCAM directly multiplies the feature map by the gradient element-by-element rather than averaging the gradients in spatial dimensions. This avoids the misleading interpretation caused by gradient averaging in Grad-CAM, ensuring the model's reliability and accuracy in interpreting fracture areas.

[0124] S243: Vertebral CT images are repeatedly scanned along the z-axis. HiResCAM technology is used to generate potential fracture areas for each slice. These results are stacked into a complete three-dimensional thermal map of the potential fracture areas. The thermal map of the fracture areas is manually verified and annotated to generate a segmentation mask for the vertebral fracture area.

[0125] It should be pointed out that the numerical values ​​involved in the above specific embodiments are not intended to limit the present application, and the selection of specific numerical values ​​can be adjusted according to actual conditions.

[0126] Step S250, generating a vertebral fracture region segmentation mask based on the vertebral fracture image; obtaining a CT image of the fracture at a specific location, and constructing an artificial vertebral fracture region segmentation mask based on the CT image of the fracture at the specific location and the image of the vertebra without a fracture; constructing training data based on the vertebral fracture region segmentation mask and the artificial vertebral fracture region segmentation mask, and using an uncertainty-aware cross-entropy loss function to perform semi-supervised learning training on a 3D U-Net model to obtain a vertebral fracture segmentation model; wherein the amount of training data is greater than the amount of data corresponding to the spinal CT image, and the specific location includes limbs and / or ribs.

[0127] S251: Based on the potential fracture area heat map image generated in step S240, fine annotation is performed through manual verification to generate a true vertebral fracture area segmentation mask.

[0128] S252: Obtain rib fracture CT images from an existing publicly available rib fracture dataset. For each fracture region in each rib fracture CT image, find a minimum bounding box and crop the rib fracture region according to the size of the bounding box. Generate a non-fracture vertebrae segmentation mask using the non-fracture vertebrae image obtained in the previous step. Perform random region matching on the non-fracture vertebrae segmentation mask and the cropped specific fracture region segmentation mask to overlap them. The overlapping portion constitutes the artificial vertebrae fracture region segmentation mask.

[0129] S253: Using the artificial vertebral fracture region segmentation mask and the real vertebral fracture region segmentation mask as training data (thereby achieving amplification of the training data, so that the vertebral fracture segmentation model can be trained based on a small number of spinal CT images of target patients), semi-supervised training is performed on the 3D U-Net model trained in step S220 to obtain a vertebral fracture segmentation model.

[0130] The embodiment of the present invention provides a training scheme for a vertebral fracture segmentation model based on semi-supervised learning of a neural network. Through the three links of vertebral fracture classification → vertebral fracture area labeling → vertebral fracture area segmentation, the classification labels are effectively used in combination with artificial vertebral fracture area segmentation masks to improve the accuracy of the model in vertebral fracture area segmentation. Among them, the 3D U-Net model is used to cut the vertebrae, which not only reduces the interference of other vertebrae on the classification, but also enables the model to focus on the classification of the target vertebrae, thereby improving the accuracy and precision of the classification. The 2.5D vertebral fracture classification model combines two-dimensional convolution with LSTM structure, which can not only efficiently integrate two-dimensional spatial information and contextual information in the depth direction, avoiding the high computational cost of pure three-dimensional convolution, but also utilize pre-trained models such as ImageNet to improve the accuracy of fracture classification while reducing training time, thereby achieving fast and high-precision judgment of whether the vertebra is fractured. Based on HiResCAM vertebral fracture visualization technology, an optimized gradient calculation algorithm is adopted to achieve vertebrae visualization at a high-resolution three-dimensional spatial scale. This method significantly improves the annotation accuracy of vertebral fracture area segmentation mask data, providing intuitive and accurate imaging support for clinical diagnosis. Finally, the artificial vertebral fracture region segmentation mask data was constructed using fracture data of specific parts (such as rib fracture data), and the real vertebral fracture region segmentation mask data was integrated for semi-supervised learning. The training data required for model training was constructed by relying on the spinal CT image data of some real patients and combining easily obtained labeled data (such as labeled data of limbs and fracture images), thereby completing the training of the vertebral fracture segmentation model, significantly reducing the difficulty of data acquisition, and achieving the effect of significantly improving the accuracy of vertebral fracture segmentation under the condition of a limited number of samples.

[0131] Based on the above-mentioned vertebral fracture segmentation and training method based on neural network semi-supervised learning provided by the embodiment of the present invention, the embodiment of the present invention also provides a vertebral fracture segmentation and training system based on neural network semi-supervised learning, which is used to construct and train a vertebral fracture segmentation model, and the vertebral fracture segmentation model can be used to perform vertebral fracture segmentation on actual patient spinal CT images, such as Figure 5 As shown, the vertebral fracture segmentation and training system 500 based on neural network semi-supervised learning includes:

[0132] The vertebra segmentation processing module 501 is used to obtain a spinal CT image of a target patient and perform initial vertebra segmentation processing on the spinal CT image based on a 3D U-Net model to obtain a vertebral CT image corresponding to each vertebra.

[0133] In some of the embodiments, after the step of acquiring the spinal CT image of the target patient, the above-mentioned vertebrae segmentation processing module 501 is also used to perform image preprocessing on the spinal CT image, specifically including: performing unified processing of the medical image axis direction and the physical scale of the spinal vertebrae on the spinal CT image; performing image denoising processing on the spinal CT image; and normalizing the pixel values ​​in the spinal CT image based on the maximum and minimum normalization method.

[0134] Based on the above settings, before using the target patient's spinal CT images for vertebral fracture detection, image preprocessing is performed on the spinal CT images, which helps to enhance image features and unify the image format, thereby improving the generalization ability and classification accuracy of the subsequent vertebral segmentation model and fracture classification model, reducing the misdiagnosis rate, and improving the efficiency of fracture segmentation.

[0135] Based on the specific processing steps of the above-mentioned image preprocessing, the quality and consistency of the image can be significantly improved. For the model training step, image preprocessing improves the generalization and feature capabilities of the model and reduces the misdiagnosis rate. For the spinal CT image detection of the target patient, image preprocessing helps to eliminate noise and differences in the data, enhances the subsequent model's ability to recognize fracture characteristics, and thus improves the accuracy and efficiency of classification.

[0136] The 3D U-Net model provided in the embodiment of the present invention is initially trained based on spinal CT sample data of historical patients, and is used to predict and generate segmentation mask labels corresponding to each vertebra to achieve initial vertebrae segmentation processing.

[0137] Unlike traditional 2D U-Net models, 3D U-Net models can directly process three-dimensional data, preserving the spatial information of the image. Furthermore, 3D U-Net models are more efficient when processing block-like images, addressing the efficiency issues of traditional 2D models when processing 3D data. This enables the model to complete vertebrae segmentation tasks more quickly, improving overall fracture detection efficiency. By quickly and accurately segmenting vertebrae regions in CT images, the 3D U-Net model can improve the accuracy of subsequent vertebrae classification.

[0138] In some embodiments, the vertebral segmentation processing module 501 is further used to: perform initial vertebral segmentation processing on the spinal CT image that has undergone image preprocessing based on the 3D U-Net model to extract the segmentation mask label of each vertebra; and extract the vertebral CT image corresponding to each vertebra based on the segmentation mask label and the spinal CT image.

[0139] The 3D U-Net model provided in the embodiment of the present invention, through its unique encoder-decoder structure and skip connections, can effectively extract features from spinal CT images and generate high-precision vertebral segmentation masks. Based on this high-precision segmentation, individual vertebrae can be clearly separated, providing accurate regional information for subsequent vertebral fracture classification.

[0140] In some embodiments, the above-mentioned spinal fracture segmentation and training system 500 based on neural network semi-supervised learning also includes an initial training and optimization module of a 3D U-Net model, which is used to: obtain spinal CT sample data of multiple historical patients, perform image preprocessing on the spinal CT sample images in the spinal CT sample data, and perform data enhancement processing on the spinal CT sample images after image preprocessing; wherein the spinal CT sample data includes spinal CT sample images, labels of whether multiple vertebrae have fractures, and true segmentation mask labels corresponding to multiple vertebrae; data augmentation processing includes elastic deformation processing, noise addition processing, and geometric change processing; construct a 3D U-Net model, and use the spinal CT sample data after data enhancement processing as input, output the predicted segmentation mask label corresponding to each vertebra based on the 3D U-Net model, calculate the cross-entropy loss function between the predicted segmentation mask label and the true segmentation mask label, adjust the model parameters with the purpose of minimizing the cross-entropy loss function, and complete the initial training and optimization of the 3D U-Net model.

[0141] Based on the above settings, data augmentation techniques such as elastic deformation, noise addition, and geometric changes enable the model to learn more diverse image features, thereby improving its adaptability to different patients and imaging conditions. Data augmentation also increases the diversity of training data and reduces overfitting of the model to the training data. Furthermore, by simulating different image changes, the model can more accurately identify vertebral boundaries, thereby improving segmentation accuracy.

[0142] During 3D U-Net model optimization, the number of background and target (vertebrae) pixels is often unbalanced during vertebrae segmentation. By calculating the cross-entropy loss function between the predicted segmentation mask label and the true segmentation mask label and optimizing it with the goal of minimizing this cross-entropy loss function, the 3D U-Net model can more accurately learn the segmentation features of the vertebrae, thereby helping the model achieve better segmentation results in the case of unbalanced data.

[0143] In some embodiments, the above-mentioned vertebral fracture segmentation and training system 500 based on neural network semi-supervised learning also includes a vertebral CT image processing module. After the step of performing vertebral segmentation processing on the vertebral CT image that has undergone image preprocessing based on the 3D U-Net model to obtain vertebral CT images corresponding to each vertebra, the above-mentioned vertebral CT image processing module is used to: perform morphological expansion processing on the vertebral CT image to fill the holes and gaps in the segmentation results; and perform noise point removal processing on the vertebral CT image to eliminate noise points around the vertebral area.

[0144] Morphological dilation effectively fills holes and gaps in the segmentation results, enhancing the continuity of the vertebral region and making it more complete, reducing subsequent misclassifications due to incomplete segmentation. Noise point removal effectively removes noise points around the vertebral region, improving the purity of the segmentation results and reducing misclassifications due to noise. These operations help further improve the accuracy of vertebral fracture classification.

[0145] The vertebra classification module 502 is used to classify the vertebral CT images using the 2.5D vertebral fracture classification model to obtain vertebral fracture images and non-fracture vertebral images.

[0146] Among them, the 2.5D vertebral fracture classification model provided by the embodiment of the present invention is constructed based on a 2.5D convolutional neural network and an LSTM network, and is trained based on vertebral CT sample images of historical patients, the segmentation mask labels corresponding to each vertebra, and the label of whether a fracture exists.

[0147] Furthermore, the 2.5D vertebral fracture classification model provided by an embodiment of the present invention combines a 2.5D convolutional neural network (CNN) and a long short-term memory network (LSTM). This vertebral classification model efficiently integrates two-dimensional spatial information with depth-direction contextual information, avoiding the high computational cost of 3D convolution. It can also utilize the pre-trained weights of two-dimensional convolution to further improve the accuracy of fracture classification detection while reducing training time. Based on the 2.5D vertebral fracture classification model, it can quickly process vertebral CT images and output a prediction result of whether there is a fracture (including vertebral fracture images and non-fracture vertebral images). This automated classification process significantly improves the accuracy of vertebral classification while reducing the efficiency of vertebral classification.

[0148] In some embodiments, the above-mentioned vertebral fracture segmentation and training system 500 based on neural network semi-supervised learning also includes a training and optimization module of a 2.5D vertebral fracture classification model, which is used to: obtain M vertebral CT sample images, each vertebral CT sample image extracts multiple slices of preset resolution along the z-axis, and combines each slice with its adjacent slices to form an N-channel vertebral CT image; generate a vertebral mask image based on the segmentation mask label corresponding to each vertebra, and merge the N-channel vertebral CT images and the vertebral mask image by channel to form target input data; wherein the vertebral CT sample images include the segmentation mask label corresponding to the vertebra, and at least m sample images include a label of whether there is a fracture, where m is a positive integer less than or equal to M; construct 2 .5D vertebral fracture classification model, the 2.5D vertebral fracture classification model includes an encoder, a reshaping layer, an LSTM network layer, and a Head layer. The target input data is input into the 2.5D vertebral fracture classification model, and the feature information in the target input data is extracted based on the encoder. The feature tensor corresponding to the feature information is reshaped and adjusted based on the reshaping layer. The LSTM network layer is used to capture the context information in the z-axis direction to obtain the feature representation information corresponding to each sample image. The feature representation information is compressed and classified through the Head layer to output a label indicating whether the vertebra has a fracture; the binary cross entropy loss function is used to optimize the 2.5D vertebral fracture classification model to determine the classification model parameters corresponding to the optimized 2.5D vertebral fracture classification model.

[0149] By training with historical patient vertebral CT sample images, segmentation mask labels, and fracture presence / absence labels, the 2.5D vertebral fracture classification model is able to learn characteristics from different patients and imaging conditions. This diverse training data helps improve the model's generalization ability, enabling it to maintain high performance even when faced with new, unseen data. During the model optimization process, the 2.5D vertebral fracture classification model is optimized using a binary cross-entropy loss function. By minimizing the binary cross-entropy loss function, the classification model parameters are adjusted to ensure that the model can accurately output the label of whether a vertebral fracture exists.

[0150] The vertebral fracture segmentation and training module 503 is used to obtain a CT image of a fracture at a specific site, construct an artificial vertebral fracture region segmentation mask based on the CT image of the fracture at the specific site and the image of a vertebra without a fracture, and generate a vertebral fracture region segmentation mask based on the vertebral fracture image; construct training data based on the artificial vertebral fracture region segmentation mask and the vertebral fracture region segmentation mask, and use the uncertainty-aware cross-entropy loss function to perform semi-supervised learning training on the 3D U-Net model to obtain a vertebral fracture segmentation model; wherein the amount of training data is greater than the amount of data corresponding to the vertebral CT image; the specific site includes limbs and / or ribs.

[0151] The vertebral fracture segmentation model trained in the embodiment of the present invention is used to perform vertebral fracture segmentation operations on vertebral CT images of real patients to assist in the diagnosis and classification of vertebral fractures.

[0152] Through the above settings, vertebral fracture region segmentation masks are generated according to vertebral fracture images to achieve effective classification and labeling of vertebral fracture regions. Combined with the artificial vertebral fracture region segmentation masks constructed based on fracture images of specific parts (such as common fracture parts such as limbs and ribs, which have more existing labeled data for these common fractures), a semi-supervised learning strategy is adopted to enhance the model's ability to recognize fracture features, thereby significantly improving the generalization of the constructed vertebral fracture segmentation model and the segmentation accuracy of the vertebral fracture segmentation model for actual vertebral fracture cases.

[0153] In some embodiments, the above-mentioned vertebral fracture segmentation and training module 503 is also used to: determine the vertebral fracture area of ​​the vertebral fracture image based on resolution visualization technology; scan the vertebral fracture area along the z-axis, and stack the scanning results to generate a thermal map image of the fracture area under a three-dimensional spatial structure, manually verify the thermal map image of the fracture area, and mark and generate a vertebral fracture area segmentation mask.

[0154] Through the above settings, high-resolution visualization technology (such as HiResCAM technology) can provide high-resolution feature maps by retaining more spatial details, which can more accurately locate important areas in the image, thereby more clearly identifying the specific location and range of vertebral fractures, and improving the accuracy of fracture area identification. Specifically, by scanning the vertebral fracture area along the z-axis and stacking the scan results, a thermal image of the fracture area in a three-dimensional spatial structure is generated. This three-dimensional thermal image can intuitively display the spatial distribution of the fracture area in the vertebrae, and also help doctors more comprehensively understand the severity and location of the fracture.

[0155] In some of the embodiments, the above-mentioned spinal fracture segmentation and training module 503 is also used to: determine the minimum bounding box corresponding to the fracture area in the fracture CT image of the specific part, and obtain a fracture area segmentation mask of the specific part from the fracture CT image of the specific part based on the minimum bounding box; generate a non-fracture vertebra segmentation mask based on the non-fracture vertebra image, and construct an artificial vertebral fracture area segmentation mask based on the non-fracture vertebra segmentation mask and the fracture area segmentation mask of the specific part.

[0156] Specifically, taking rib fracture CT images as an example, rib fracture CT images can be obtained from existing publicly available rib fracture datasets. For each fracture region in each rib fracture CT image, a minimum bounding box is found and the rib fracture region is cropped according to the size of the bounding box. A non-fractured vertebrae segmentation mask is generated using the non-fractured vertebrae image obtained in the previous step. This non-fractured vertebrae segmentation mask is then overlapped with the cropped fracture region segmentation mask at the specific location through random region matching. The overlapping portion is the artificial vertebrae fracture region segmentation mask.

[0157] The above-mentioned vertebral fracture segmentation and training system based on neural network semi-supervised learning provided by the embodiment of the present invention adopts a vertebral segmentation processing module for obtaining the vertebral CT image of the target patient, and performs initial vertebral segmentation processing on the vertebral CT image based on the 3D U-Net model to obtain the vertebral CT image corresponding to each vertebra; the vertebral classification module is used to classify the vertebral CT image using the 2.5D vertebral fracture classification model to obtain vertebral fracture images and non-fracture vertebral images; the vertebral fracture segmentation and training module is used to obtain the fracture CT image of a specific part, construct an artificial vertebral fracture area segmentation mask based on the fracture CT image of the specific part and the non-fracture vertebral image, and generate a vertebral fracture area segmentation mask according to the vertebral fracture image; construct training data based on the artificial vertebral fracture area segmentation mask and the vertebral fracture area segmentation mask, and use the uncertainty-aware cross-entropy loss function to perform semi-supervised learning training on the 3D U-Net model to obtain a vertebral fracture segmentation model; wherein the amount of training data is greater than the amount of data corresponding to the vertebral CT image; the specific parts include limbs and / or ribs. It is realized through 3D The U-Net model obtains the vertebral CT images corresponding to each vertebra, realizes effective and fast initial segmentation of the vertebrae, reduces the interference of other vertebrae on subsequent classification, and enables the subsequent vertebral classification model to focus on the classification of the target vertebrae, thereby improving the accuracy and precision of classification detection; and then uses the 2.5D vertebral fracture classification model constructed based on the 2.5D convolutional neural network and LSTM network to classify the vertebral CT images. The vertebral classification model efficiently integrates the two-dimensional spatial information and the contextual information in the depth direction, avoids the high computational cost of 3D convolution, and can also use the pre-trained weights of the two-dimensional convolution to reduce the training time while reducing the training time. When the image of fracture at a specific site is used, the accuracy of fracture classification is further improved; finally, artificial vertebral fracture data is constructed by using fracture images at a specific site, which overcomes the influence of less training data and difficulty in data acquisition in related technologies. At the same time, semi-supervised learning training is used to enhance the recognition ability of vertebral fracture segmentation model for vertebral fracture features. The training data required for model training is constructed by relying on the annotated data of vertebral fracture images of some real patients combined with the annotated data of fracture images at specific sites (such as limbs and ribs), which reduces the difficulty of data acquisition and realizes the training of vertebral fracture segmentation model. It also achieves the technical effect of significantly improving the accuracy of vertebral fracture segmentation under the condition of limited sample number.

[0158] An embodiment of the present invention further provides a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a processor of a computer, is used to cause the computer to perform the method of the embodiment of the present invention.

[0159] An embodiment of the present invention further provides a computer program product, including a computer program, wherein the computer program, when executed by a processor of a computer, is configured to cause the computer to perform the method of an embodiment of the present invention. A computer program product should be understood as a software product that primarily implements the above-described method of the present application through a computer program.

[0160] An embodiment of the present invention further provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, wherein the computer program, when executed by the at least one processor, causes the electronic device to perform the method of an embodiment of the present invention.

[0161] refer to Figure 6 , a block diagram of an electronic device that can serve as a server or client of an embodiment of the present invention will now be described, which is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or required herein.

[0162] like Figure 6 As shown, the electronic device includes a computing unit 601, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 602 or a computer program loaded from a storage unit 608 into a random access memory (RAM) 603. Various programs and data required for the operation of the electronic device can also be stored in the RAM 603. The computing unit 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0163] Multiple components within the electronic device are connected to the I / O interface 605, including an input unit 606, an output unit 607, a storage unit 608, and a communication unit 609. The input unit 606 can be any type of device capable of inputting information into the electronic device. The input unit 606 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. The output unit 607 can be any type of device capable of presenting information and may include, but is not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 608 may include, but is not limited to, a magnetic disk or an optical disk. The communication unit 609 allows the electronic device to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks and may include, but is not limited to, a modem, a network card, an infrared communication device, and / or a wireless communication transceiver, such as a Bluetooth device, a WiFi device, a WiMax device, a cellular communication device, and / or the like.

[0164] The computing unit 601 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a CPU, a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing units, various computing units that run machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 601 performs the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention may be implemented as a computer program that is tangibly contained in a machine-readable medium, such as a storage unit 608. In some embodiments, part or all of the computer program may be stored in a ROM.

[0165] 602 and / or the communication unit 609 are loaded and / or installed on the electronic device. In some embodiments, the computing unit 601 can be configured to execute the above method in any other appropriate manner (for example, by means of firmware).

[0166] The computer programs for implementing the methods of the embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the computer programs are executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The computer programs may be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0167] In the context of an embodiment of the present invention, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, or infrared system, device, or apparatus, or any suitable combination of the foregoing. More specific examples of machine-readable storage media can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0168] It should be noted that the term "including" and its variations used in the embodiments of the present invention are open-ended, i.e., "including but not limited to." The term "based on" means "based at least in part on." The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; and the term "some embodiments" means "at least some embodiments." The modifications of "one" and "a plurality of" mentioned in the embodiments of the present invention are illustrative and non-restrictive. Those skilled in the art should understand that, unless the context clearly indicates otherwise, they should be understood as "one or more."

[0169] The information / data involved in the embodiments of the present invention (including but not limited to user device information / data, user personal information / data, information / data used for analysis, stored information / data, displayed information / data, etc.) are all information / data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant information / data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0170] The various steps described in the method implementation scheme provided in the embodiments of the present invention may be performed in different orders and / or in parallel. In addition, the method implementation scheme may include additional steps and / or omit the steps shown. The scope of protection of the present invention is not limited in this respect.

[0171] The term "embodiment" in this specification refers to specific features, structures, or characteristics described in conjunction with the embodiment that can be included in at least one embodiment of the present invention. The appearance of this phrase in various places in the specification does not necessarily mean the same embodiment, nor does it mean that it is mutually exclusive with other embodiments and is independent or optional. The various embodiments in this specification are described in a related manner, and the same or similar parts between the various embodiments are referenced to each other. In particular, for the embodiments of the device, equipment, and system, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts are referred to the partial description of the method embodiment.

[0172] The above-described embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of protection. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A vertebral fracture segmentation and training method based on neural network semi-supervised learning, characterized in that: include: Obtaining a spinal CT image of a target patient, and performing initial vertebra segmentation processing on the spinal CT image based on a 3D U-Net model to obtain a vertebral CT image corresponding to each vertebra; Classifying the vertebral CT images using a 2.5D vertebral fracture classification model to obtain vertebral fracture images and non-fractured vertebral images; Acquire a CT image of a fracture at a specific site, construct an artificial vertebral fracture region segmentation mask based on the CT image of the fracture at the specific site and the image of the vertebra without a fracture, and generate a vertebral fracture region segmentation mask based on the vertebral fracture image; construct training data based on the artificial vertebral fracture region segmentation mask and the vertebral fracture region segmentation mask, and perform semi-supervised learning training on the 3D U-Net model using an uncertainty-aware cross-entropy loss function to obtain a vertebral fracture segmentation model; wherein the amount of the training data is greater than the amount of data corresponding to the vertebral CT image; and the specific site includes limbs and / or ribs.

2. The method according to claim 1, characterized in that The step of generating a vertebral fracture region segmentation mask according to the vertebral fracture image comprises: determining a vertebral fracture region of the vertebral fracture image based on a resolution visualization technique; The vertebral fracture area is scanned along the z-axis, and the scanning results are stacked to generate a thermal map image of the fracture area in a three-dimensional spatial structure. The thermal map image of the fracture area is manually verified and annotated to generate a segmentation mask of the vertebral fracture area.

3. The method according to claim 1, characterized in that The step of constructing an artificial vertebral fracture region segmentation mask based on the fracture CT image of the specific part and the vertebral image without fracture comprises: Determining a minimum bounding box corresponding to the fracture area in the CT image of the fracture at the specific site, and cropping the CT image of the fracture at the specific site based on the minimum bounding box to obtain a segmentation mask of the fracture area at the specific site; A non-fractured vertebra segmentation mask is generated according to the non-fractured vertebra image, and an artificial vertebra fracture region segmentation mask is constructed based on the non-fractured vertebra segmentation mask and the specific fracture region segmentation mask.

4. The method according to claim 1, wherein The 3D U-Net model is initially trained based on historical patient spinal CT sample data and is used to predict and generate segmentation mask labels corresponding to each vertebra to achieve initial vertebrae segmentation processing. The steps of performing initial vertebrae segmentation processing on the spinal CT image that has undergone image preprocessing based on the 3D U-Net model to obtain vertebrae CT images corresponding to each vertebra include: performing initial vertebra segmentation processing on the spinal CT image that has undergone image preprocessing based on the 3D U-Net model, and extracting a segmentation mask label for each vertebra; A vertebral CT image corresponding to each vertebra is extracted based on the segmentation mask label and the spinal CT image.

5. The method according to claim 4, characterized in that The method also includes the initial training and optimization steps of the 3D U-Net model: Acquire spinal CT sample data of multiple historical patients, perform image preprocessing on spinal CT sample images in the spinal CT sample data, and perform data enhancement processing on the spinal CT sample images that have undergone image preprocessing; wherein the spinal CT sample data includes spinal CT sample images, labels indicating whether multiple vertebrae have fractures, and true segmentation mask labels corresponding to the multiple vertebrae; the data augmentation processing includes elastic deformation processing, noise addition processing, and geometric change processing; A 3D U-Net model was constructed, and the spinal CT sample data after data augmentation was used as input. The predicted segmentation mask labels corresponding to each vertebra were output based on the 3DU-Net model. The cross-entropy loss function between the predicted segmentation mask labels and the true segmentation mask labels was calculated. The model parameters were adjusted to minimize the cross-entropy loss function, completing the initial training and optimization of the 3D U-Net model.

6. The method according to claim 1, characterized in that The 2.5D vertebral fracture classification model is constructed based on a 2.5D convolutional neural network and an LSTM network, and is trained based on historical patient vertebral CT sample images, segmentation mask labels corresponding to each vertebra, and labels indicating whether a fracture is present. The method further includes the following steps: training and optimizing the 2.5D vertebral fracture classification model: Acquire M vertebral CT sample images, extract multiple slices of preset resolution along the z-axis from each vertebral CT sample image, and combine each slice with its adjacent slices to form an N-channel vertebral CT image; generate a vertebral mask image based on the segmentation mask label corresponding to each vertebra, and merge the N-channel vertebral CT images with the vertebral mask image on a channel-by-channel basis to form target input data; wherein the vertebral CT sample images include the segmentation mask label corresponding to the vertebra, and at least m sample images include a label indicating whether a fracture is present, where m is a positive integer less than or equal to M; Constructing a 2.5D vertebral fracture classification model, the 2.5D vertebral fracture classification model comprising an encoder, a reshaping layer, an LSTM network layer, and a head layer, inputting target input data into the 2.5D vertebral fracture classification model, extracting feature information from the target input data based on the encoder, reshaping and adjusting a feature tensor corresponding to the feature information based on the reshaping layer, and capturing contextual information in the z-axis direction using the LSTM network layer to obtain feature representation information corresponding to each sample image, compressing and classifying the feature representation information through the head layer to output a label indicating whether a vertebral fracture exists; The 2.5D vertebral fracture classification model is optimized using a binary cross entropy loss function to determine classification model parameters corresponding to the optimized 2.5D vertebral fracture classification model.

7. The method according to claim 1, characterized in that After the step of acquiring a spine CT image of the target patient, the method further includes the step of performing image preprocessing on the spine CT image: Performing unified processing of the medical image axis direction and the physical scale of the vertebrae on the spinal CT image; performing image denoising processing on the spine CT image; The pixel values ​​in the spinal CT image are normalized based on the maximum-minimum normalization method.

8. A vertebral fracture segmentation and training system based on neural network semi-supervised learning, characterized by: include: a vertebral segmentation processing module, configured to obtain a spinal CT image of a target patient and perform initial vertebral segmentation processing on the spinal CT image based on a 3D U-Net model to obtain a vertebral CT image corresponding to each vertebra; a vertebra classification module, configured to classify the vertebral CT images using a 2.5D vertebral fracture classification model to obtain vertebral fracture images and non-fractured vertebral images; A vertebral fracture segmentation and training module is used to obtain a CT image of a fracture at a specific site, construct an artificial vertebral fracture region segmentation mask based on the CT image of the fracture at the specific site and the image of the vertebra without a fracture, and generate a vertebral fracture region segmentation mask based on the vertebral fracture image; construct training data based on the artificial vertebral fracture region segmentation mask and the vertebral fracture region segmentation mask, and use an uncertainty-aware cross-entropy loss function to perform semi-supervised learning training on the 3D U-Net model to obtain a vertebral fracture segmentation model; wherein the amount of the training data is greater than the amount of data corresponding to the vertebral CT image; the specific site includes limbs and / or ribs.

9. An electronic device comprising: A processor and a memory storing a program, wherein the program comprises instructions, which, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 7.

10. A non-transitory machine-readable medium storing computer instructions, characterized in that The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 7.

Images