An image segmentation method and related apparatus
By training an image segmentation model using human model images with different attribute information and multi-angle projection images, the problems of accuracy and efficiency in skeletal image segmentation were solved, achieving high accuracy and high efficiency in skeletal segmentation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SPARTICLE HEALTHCARE CO LTD
- Filing Date
- 2025-10-16
- Publication Date
- 2026-07-03
AI Technical Summary
In existing technologies, the accuracy of skeletal image segmentation is poor, mainly relying on manual delineation of skeletal contours, which suffers from high subjectivity, low efficiency, and inconsistent annotation.
By generating human model images with different attribute information, multi-angle projection images are generated based on attenuation data and activity settings. The image segmentation model is trained using label information to improve segmentation accuracy.
It achieves high accuracy and efficiency in skeletal image segmentation, generates a large number of projection images with different individual features, provides effective samples for training models, and improves the accuracy of segmentation models and the diversity and consistency of labeled data.
Smart Images
Figure CN121304708B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of image processing technology, and in particular to an image segmentation method and related apparatus. Background Technology
[0002] In medical image analysis, accurate segmentation of organs throughout the body, such as bones, is crucial for a variety of clinical applications, including diagnosis of bone diseases (including osteoporosis, bone tumors, fractures, etc.), surgical planning (orthopedic surgery, maxillofacial surgery, etc.), and rehabilitation assessment.
[0003] When performing bone image segmentation, doctors mainly rely on manually outlining the bone contours on medical images (such as SPECT (Single-Photon Emission Computed Tomography)) to segment bones. However, this method is highly subjective, resulting in poor accuracy in bone image segmentation. Summary of the Invention
[0004] In view of the above problems, this application provides an image segmentation method and related apparatus to improve the accuracy of image segmentation. The specific solution is as follows:
[0005] The first aspect of this application provides an image segmentation method, comprising:
[0006] Obtain the target image for image segmentation;
[0007] The target image is input into the image segmentation model to obtain the image segmentation result of the target image;
[0008] The generation process of the image segmentation model includes:
[0009] Generate multiple human body model images with different attribute information;
[0010] Based on the attenuation data of each region in the human body model image, determine the attenuation image corresponding to the human body model image;
[0011] The activity level of each region in the human body model image is set to obtain the activity image corresponding to the human body model image.
[0012] Using the attenuation image and the activity image, a multi-angle projection image is generated;
[0013] The multi-angle projected image is adjusted to obtain the target projected image;
[0014] An image segmentation model is trained based on the target projection image and the label information of each region in the target projection image.
[0015] In one possible implementation, multiple human model images with different attribute information are generated, including:
[0016] The content of the attribute information in the parameter file of the simulation software was adjusted multiple times to obtain multiple attribute information; the attribute information includes at least one of gender, height, weight and organ size;
[0017] Generate a human model image corresponding to each of the aforementioned attribute information; each organ in the human model image is configured with corresponding index information.
[0018] In one possible implementation, determining the attenuation image corresponding to the human body model image based on attenuation data of various regions in the human body model image includes:
[0019] By setting the photon energy and voxel size of each region in the human body model image, the attenuation data of X-rays for each region is obtained.
[0020] Based on the attenuation data of X-rays in each of the aforementioned regions, attenuation images of each region in the human body model image are determined.
[0021] In one possible implementation, an activity setting operation is performed on each region of the human body model image to obtain an activity image corresponding to the human body model image, including:
[0022] The activity of each region in the human body model image is set to the same activity value to obtain the activity image corresponding to the human body model image;
[0023] Alternatively, the activity of each region in the human body model image can be set to the actual activity value of the region during the scanning operation to obtain the activity image corresponding to the human body model image; wherein the activity of different regions may be the same or different.
[0024] In one possible implementation, a multi-angle projection image is generated using the attenuation image and the activity image, including:
[0025] Based on the target device type, set the detector parameters, collimator parameters, radioactive source activity, acquisition distance, and acquisition angle of the projection program;
[0026] The attenuation image and the activity image are input into the projection program so that the projection program uses the attenuation image and the activity image to simulate the signal detection and data conversion process when the target device performs image acquisition, and obtains a multi-angle projection image in binary format.
[0027] In one possible implementation, image adjustment operations are performed on the multi-angle projected image to obtain the target projected image, including:
[0028] Gaussian smoothing is performed on the multi-angle projection image to obtain the intermediate projection image;
[0029] The intermediate projection image is subjected to noise addition to obtain the target projection image.
[0030] In one possible implementation, an image segmentation model is trained based on the target projection image and the label information of each region in the target projection image, including:
[0031] Based on the index information of each organ configuration in the human model image, the label information of the corresponding region in the target projection image is determined, and the label information is the region classification result of the region;
[0032] The image segmentation model is trained using the target projection image and the label information of each region in the target projection image until the training stops when the stopping condition is met. During the training process, the parameters of the image segmentation model are continuously adjusted through the backpropagation algorithm, so that the image segmentation model learns the features and segmentation patterns of different regions.
[0033] A second aspect of this application provides an image segmentation apparatus, comprising:
[0034] The image acquisition module is used to acquire the target image to be segmented.
[0035] Image generation module, used to generate image segmentation models;
[0036] The image segmentation module is used to input the target image into the image segmentation model to obtain the image segmentation result of the target image;
[0037] The image generation module includes:
[0038] The image generation submodule is used to generate multiple human model images with different attribute information;
[0039] The attenuation image determination submodule is used to determine the attenuation image corresponding to the human body model image based on the attenuation data of each region in the human body model image;
[0040] The activity image determination submodule is used to perform activity setting operations on each region in the human body model image to obtain the activity image corresponding to the human body model image;
[0041] The projection image determination submodule is used to generate a multi-angle projection image using the attenuation image and the activity image;
[0042] The image adjustment submodule is used to perform image adjustment operations on the multi-angle projection image to obtain the target projection image;
[0043] The model training submodule is used to train an image segmentation model based on the target projection image and the label information of each region in the target projection image.
[0044] A third aspect of this application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein:
[0045] The memory is used to store computer programs;
[0046] The processor is used to execute the computer program so that the electronic device can implement the image segmentation method described above.
[0047] A fourth aspect of this application provides a computer storage medium carrying one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the image segmentation method described above.
[0048] By employing the above technical solutions, this application provides an image segmentation method and related apparatus. In this application, a target image to be segmented is acquired, and the target image is input into an image segmentation model to obtain the image segmentation result. Since the image segmentation model is trained based on the target projection image and the label information of each region in the target projection image, the image segmentation model can achieve high segmentation accuracy, thereby improving the accuracy of target image segmentation. Furthermore, this application generates multiple human model images with different attribute information. Based on the attenuation data of each region in the human model image, an attenuation image corresponding to the human model image is determined. Activity setting operations are performed on each region in the human model image to obtain the activity image corresponding to the human model image. Using the attenuation image and the activity image, a multi-angle projection image is generated. Image adjustment operations are performed on the multi-angle projection image to obtain the target projection image. Through the above steps, a large number of projection images with different individual characteristics can be quickly generated, and the obtained target projection image is close to the projection image obtained from a real scanning operation, providing a large number of effective training samples for training the image segmentation model. Furthermore, using these training samples for model training can improve the accuracy of the image segmentation model when performing image segmentation operations. Attached Figure Description
[0049] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.
[0050] Figure 1 A flowchart of an image segmentation method provided in this application;
[0051] Figure 2 A schematic diagram of a human body model image provided in this application;
[0052] Figure 3 A schematic diagram of an attenuation image provided in this application;
[0053] Figure 4 A schematic diagram of an activity image provided in this application;
[0054] Figure 5 A schematic diagram of a multi-angle projection image provided in this application;
[0055] Figure 6 A schematic diagram of a target projection image provided in this application;
[0056] Figure 7 A flowchart of an image segmentation method provided in this application;
[0057] Figure 8 A schematic diagram of image segmentation provided in this application;
[0058] Figure 9 A schematic diagram of the structure of an image segmentation device provided in this application;
[0059] Figure 10 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation
[0060] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.
[0061] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.
[0062] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.
[0063] In medical image analysis, accurate segmentation of organs throughout the body, such as bones, is crucial for a variety of clinical applications, including diagnosis of bone diseases (including osteoporosis, bone tumors, fractures, etc.), surgical planning (orthopedic surgery, maxillofacial surgery, etc.), and rehabilitation assessment.
[0064] Taking bone image segmentation as an example, when performing bone image segmentation, doctors mainly rely on manually outlining the bone contour on medical images (such as SPECT (Single-Photon Emission Computed Tomography)) to segment bones. However, this method is highly subjective, resulting in poor accuracy of bone image segmentation.
[0065] With the development of artificial intelligence technology, deep learning-based AI (Artificial Intelligence) models have shown great potential in the field of medical image segmentation. By training AI models with a large amount of labeled medical image data, the models can automatically learn the features of bones, thereby achieving automatic segmentation of bones in new medical images. However, constructing high-quality labeled datasets has become a key bottleneck in training accurate and effective whole-body bone segmentation AI models.
[0066] Currently, the main methods for generating labeled data are manual annotation and semi-automatic annotation. Manual annotation requires a significant investment of time and effort from professional medical imaging experts, and is prone to inaccuracies and inconsistencies due to fatigue and subjective judgment. For example, a complete set of whole-body SPECT images contains hundreds or even thousands of tomographic images. It could take medical experts hours or even days to manually delineate the skeletal outlines in each image. This not only severely impacts the progress of annotation work but also limits the rapid construction of large-scale labeled datasets, thus hindering the rapid training and optimization of AI models. Furthermore, due to the complexity of the human skeletal structure and the significant differences in morphology, density, and other characteristics of bones in different parts of the body, experts may make annotation errors or omissions due to fatigue or lack of concentration during manual annotation. Moreover, different experts may have different understandings and judgments of bone boundaries, leading to a lack of consistency in annotation results. Such inaccurate and inconsistent labeled data will seriously affect the training effect of AI models, reducing the accuracy and reliability of model segmentation.
[0067] Although semi-automatic annotation utilizes some image segmentation algorithms, it still requires a significant amount of manual correction work, resulting in limited efficiency improvements.
[0068] Based on the above, this application provides an image segmentation method and related apparatus. In this application, a target image to be segmented is acquired, and the target image is input into an image segmentation model to obtain the image segmentation result. Since the image segmentation model is trained based on the target projection image and the label information of each region in the target projection image, the image segmentation model can achieve high segmentation accuracy, thereby improving the accuracy of target image segmentation. Furthermore, this application generates multiple human model images with different attribute information. Based on the attenuation data of each region in the human model image, the attenuation image corresponding to the human model image is determined. Activity setting operations are performed on each region in the human model image to obtain the activity image corresponding to the human model image. Using the attenuation image and the activity image, a multi-angle projection image is generated. Image adjustment operations are performed on the multi-angle projection image to obtain the target projection image. Through the above steps, a large number of projection images with different individual characteristics can be quickly generated, and the obtained target projection image is close to the projection image obtained from a real scanning operation, providing a large number of effective training samples for training the image segmentation model. Using these training samples for model training can improve the accuracy of the image segmentation model when performing image segmentation operations.
[0069] One embodiment of this application provides an image segmentation method, the execution entity of which can be a processor, server, or other device. To achieve image segmentation, this embodiment requires generating a large number of projected images with different individual features to train an image segmentation model. In one implementation, this application first describes the generation process of the image segmentation model, and then describes how to use the image segmentation model to perform image segmentation.
[0070] Reference Figure 1 The generation process of the image segmentation model includes:
[0071] S11. Generate multiple human body model images with different attribute information.
[0072] In this embodiment, attribute information refers to the attribute information of the human body. Different individuals may have different attribute information, which includes at least one of gender, height, weight, and organ size. In one implementation, to reflect the differences between individuals, the attribute information may simultaneously include gender, height, weight, and organ size. The organs may include the spine, skull, etc.
[0073] In one implementation, simulation software can be used to generate images of human models. These images can be SPECT, CT (Computed Tomography), PET (Positron Emission Tomography), MRI (Magnetic Resonance Imaging), etc. Subsequent embodiments will use SPECT as an example; the processing of other image types is similar.
[0074] Simulation software such as XCAT can be used. When using simulation software to generate images, the attribute information in the simulation software's parameter file can be adjusted multiple times to obtain multiple attribute information. For example, taking XCAT as an example, XCAT has a corresponding parameter file. Lines 7-11 of the parameter file are related to gender, allowing selection of gender, chest, organ files, and heart files. In addition, height, weight, and organ size can also be configured through XCAT. In one example, the attribute information can be:
[0075] 1) Gender: Female, height 160cm, weight 50KG, spine size X1, skull size Y1...
[0076] 2) Gender: Male, height 175cm, weight 70kg, spine size X2, spine size Y2...
[0077] 3) Gender: Male, height 180cm, weight 80kg, spine size X3, spine size Y3...
[0078] The number of specific attribute information configurations can be determined based on the actual scenario requirements. In one implementation, as many different attribute information configurations as possible can be set to comprehensively and accurately simulate the complex morphology and individual differences of the entire human skeleton, thus meeting the needs for diversity and accuracy of labeled data.
[0079] After the attribute information is determined, a human model image corresponding to each attribute can be generated. Specifically, after setting the attribute information for the simulation software, the simulation software can be used to generate the human model image corresponding to that attribute information. For example, simulation software (such as XCAT) can be used to generate several human model images containing different genders, heights, weights, and organ sizes (the number is generally greater than 100 for model training purposes). These human model images are binary raw data, and each organ in the human model image is configured with corresponding index information, such as: spine as 2, skull as 3, etc. In one example, the generated human model images can be referenced... Figure 2 As shown.
[0080] S12. Based on the attenuation data of each region in the human body model image, determine the attenuation image corresponding to the human body model image.
[0081] In this embodiment, each region in the human body model image can correspond to one of the aforementioned organs, i.e., one organ corresponds to one region. The attenuation data of a region refers to the attenuation effect of that region on X-rays.
[0082] In real-world scenarios, simulation software may have an option to generate attenuation images corresponding to human model images. This option can be selected according to requirements. In this embodiment, if the option to generate attenuation images corresponding to human model images is selected, the simulation software can be used to generate and output the attenuation images corresponding to human model images.
[0083] In one implementation, when generating the attenuation image corresponding to the human body model image, the photon energy and voxel size of each region in the human body model image can be set to obtain the attenuation data of each region to X-rays. Based on the attenuation data of each region to X-rays, the attenuation image of each region in the human body model image can be determined.
[0084] Specifically, by setting photon energy and voxel size, the attenuation effect of various human organs on X-rays is displayed, simulating the tomographic imaging during a CT scan, thus obtaining attenuated images. For details, please refer to... Figure 3 .
[0085] Setting the photon energy refers to simulating different CT scan conditions using the `photon_energy` parameter (50-150keV), such as low energy suitable for soft tissue imaging and high energy suitable for bone imaging. The attenuation coefficient is inversely proportional to the cube of the energy (μ∝1 / E³). Setting the voxel size refers to adjusting the image resolution using the `voxel_size` parameter (0.1-1cm). The smaller the voxel, the richer the detail (e.g., a 0.2cm voxel can display minute calcifications).
[0086] When generating the attenuation image, the following steps need to be performed:
[0087] Load the 3D organ segmentation data (.mat format) of the XCAT phantom. Each voxel is digitally encoded to correspond to a specific organ (such as bone, lung, liver, etc.) to ensure the realism of the anatomical structure. Based on the set photon energy (e.g., 100keV), voxel size, and tissue physical properties (atomic number Z, density ρ), calculate the X-ray linear attenuation coefficient (μ) of each organ using a physical formula. This reflects the differences in X-ray attenuation among different organs, thus obtaining attenuation images of various regions in the human model image. These attenuation images specifically include the attenuation data corresponding to each organ in each human model image. The attenuation data corresponding to each organ in different human model images may be the same or different.
[0088] S13. Perform activity setting operations on each region in the human body model image to obtain the activity image corresponding to the human body model image.
[0089] In this embodiment, based on the image segmentation training requirements, the human model image described above undergoes activity assignment operations for each organ to obtain an activity image that approximates the actual SPECT acquisition. An example of the activity image can be found in [reference needed]. Figure 4 As shown.
[0090] S14. Using the attenuation image and the activity image, generate a multi-angle projection image.
[0091] In this embodiment, after obtaining the attenuation image and activity image, Monte Carlo simulation or other analytical projection programs can be used to simulate the multi-angle projection images acquired by the SPECT device. An example of a multi-angle projection image can be found in [reference needed]. Figure 5 As shown.
[0092] S15. Perform image adjustment operations on the multi-angle projection image to obtain the target projection image.
[0093] In this embodiment, since nuclear medicine images generally have low resolution, image adjustment operations are required for the generated multi-angle projection images to increase realism and generate target projection images that more closely approximate the actual acquired images, thus completing the dataset creation. An example of the target projection image can be found in [reference needed]. Figure 6 As shown.
[0094] In one implementation, the image adjustment operation can be to adjust the noise in the image. Specifically, Gaussian smoothing is first performed on the multi-angle projection image to obtain an intermediate projection image, and then noise is added to the intermediate projection image to obtain the target projection image. Additionally, the image contrast can also be adjusted, depending on the actual configuration.
[0095] The implementation process of Gaussian smoothing includes:
[0096] The three main steps are: Gaussian kernel generation → image convolution → boundary processing, as follows:
[0097] 1. Generate Gaussian kernel
[0098] The weight matrix is calculated using a two-dimensional Gaussian function based on the input kernel size.
[0099] 2. Convolution operation of images with Gaussian kernels
[0100] Convolution is the core operation for smoothing: for each pixel in the image, the weights of the Gaussian kernel are used to sum the values of the neighboring pixels, and the original pixel value is replaced.
[0101] step:
[0102] (1) Traverse each pixel (i,j) of the image;
[0103] (2) Using (i,j) as the center, extract a neighborhood region of the same size as the Gaussian kernel;
[0104] (3) Multiply the neighboring pixel values by the weights of the corresponding positions of the Gaussian kernel and sum them up. The result is used as the new value of (i,j).
[0105] 3. Boundary processing (solving the problem of incomplete neighborhood of edge pixels)
[0106] If the neighborhood of pixels at the image edge (such as the first row or the last column) exceeds the image boundaries, special handling is required. Specifically, any of the following padding methods can be used:
[0107] Zero padding: Fills the excess area with 0s (simple but may cause the edges to darken);
[0108] Replicate Padding: Fills the excess area with the nearest edge pixel value (preserving edge brightness);
[0109] Mirror Padding: The overflowing portion is filled with pixel values at symmetrical positions (for a more natural effect).
[0110] When adding noise to an intermediate projected image, Poisson noise can be added. The specific process for adding Poisson noise is as follows:
[0111] Since the pixel values of the actual image may be floating-point numbers (such as the normalized range of [0,1]) or integers (such as a grayscale image of 0-255), numerical conversion is required before generating Poisson noise. The specific steps are as follows:
[0112] 1. Image preprocessing: Signal value range adjustment
[0113] The parameter λ of the Poisson distribution must be a non-negative value, and it is usually necessary to convert the image pixel values into a "non-negative signal proportional to the photon count":
[0114] If the image is a normalized floating-point number (e.g., [0,1]): it needs to be scaled to a suitable integer range first (e.g., multiplied by a coefficient s to convert it to a non-negative integer [0, s]. The larger s is, the higher the signal strength and the weaker the noise).
[0115] If the image is an integer grayscale image (e.g., 0-255): directly use the pixel value as an approximation of λ (ensure that the pixel value is non-negative; if there is a negative value, such as the HU value of CT, it needs to be shifted to a non-negative range first).
[0116] 2. Generate Poisson noise and add it to the image.
[0117] Based on the preprocessed parameter λ, a random value following a Poisson distribution is generated for each pixel, and then the original pixel value is replaced (or noise is superimposed).
[0118] S16. Based on the target projection image and the label information of each region in the target projection image, an image segmentation model is trained.
[0119] In this embodiment, after obtaining the target projection image, it is necessary to determine the label information of each region in the target projection image before model training can be performed.
[0120] In one implementation, the label information of the corresponding region in the target projection image can be determined based on the index information configured for each organ in the human body model image. The label information is the region classification result of the region, which can be skull, femoral neck, spine, non-skeleton, etc.
[0121] Based on the above discussion, each organ in the human model image is configured with index information. This index information can be directly set as the label information for the corresponding region in the target projection image, or it can be based on the index information to set corresponding label information. For example, if the index information is 3, representing the skull, then the label is configured as skull, or the identifier corresponding to the skull. In one example, "0 = non-skeleton, 1 = femoral head, 2 = femoral neck, 3 = spine, etc." The specific label information can be configured according to the actual needs of the scenario.
[0122] In one example, the complete coordinates (x, y, z) of each skeletal region and their matching labels (e.g., rib region coordinates are labeled 1, spine is labeled 2, skull is labeled 3, and so on) can be recorded for subsequent segmentation network training.
[0123] Then, the image segmentation model is trained using the target projection image and the label information of each region in the target projection image until the training stops when the stopping condition is met.
[0124] The image segmentation model can be a deep learning framework (such as TensorFlow, PyTorch, etc.) or a neural network architecture (such as UNet, nnUnet, SegNet), specifically designed for image segmentation. The model is trained using prepared target projection images, with the training labels being unique labels for each known skeletal region.
[0125] In this embodiment, UNet is used as an example to illustrate the image segmentation model. The core of UNet segmentation network training is the training samples of "paired images and annotations". Taking the segmentation of organs in medical images as an example, the training samples are a series of original medical images (i.e., the target projection images mentioned above) and "annotation images" that are exactly the same size as these images (each pixel in the annotation image is labeled, and the label can be the background or an organ, for example, the label is "0 = non-skeleton, 1 = femoral head, 2 = femoral neck, 3 = spine, etc.").
[0126] The training samples are then divided into three parts: one part is used as the training set for model training, one part is used as the validation set to adjust model parameters, and one part is used as the test set to test the model's learning performance.
[0127] In the process of training the image segmentation model, the parameters of the image segmentation model are continuously adjusted through the backpropagation algorithm, so that the image segmentation model learns the features and segmentation patterns of different regions, such as bones.
[0128] In this embodiment, a mature nuclear medicine image segmentation model can be obtained by creating an image segmentation model dataset and training the image segmentation model.
[0129] Once the image segmentation model has been trained, it can be used to perform image segmentation operations. Specifically, refer to... Figure 7 An image segmentation method may include:
[0130] S21. Obtain the target image for image segmentation operation.
[0131] In this embodiment, the target image can be SPECT, CT, PET, MRI, etc. This embodiment uses SPECT as an example for illustration.
[0132] S22. Input the target image into the image segmentation model to obtain the image segmentation result of the target image.
[0133] In this embodiment, before inputting the target image into the image segmentation model, preprocessing such as brightness adjustment and noise removal can be performed. Then, the processed image is input into the image segmentation model to obtain the image segmentation result of the target image. Alternatively, no preprocessing can be performed; the target image can be directly input into the image segmentation model to obtain the image segmentation result. A schematic diagram of the training result after segmentation using the image segmentation model can be found in one example. Figure 8 As shown.
[0134] In this embodiment, a target image to be segmented is acquired and input into an image segmentation model to obtain the image segmentation result. Since the image segmentation model is trained based on the target projection image and the label information of each region within the target projection image, it achieves high segmentation accuracy, thereby improving the accuracy of target image segmentation. Furthermore, this application generates multiple human model images with different attribute information. Based on the attenuation data of each region in the human model images, an attenuation image corresponding to the human model image is determined. Activity settings are performed on each region in the human model images to obtain the corresponding activity images. Using the attenuation images and activity images, multi-angle projection images are generated. Image adjustment operations are performed on the multi-angle projection images to obtain the target projection image. Through these steps, a large number of projection images with different individual characteristics can be quickly generated, and the obtained target projection images are close to the projection images obtained from actual scanning operations. This provides a large number of effective training samples for training the image segmentation model. Using these training samples for model training improves the accuracy of the image segmentation model during image segmentation operations.
[0135] Furthermore, this embodiment significantly improves annotation efficiency: leveraging XCAT's powerful simulation capabilities, it can quickly generate a large number of virtual full-body skeleton models with different individual characteristics. By combining automated annotation algorithms with XCAT, the skeleton annotation of these virtual models can be completed in a short time. Compared to traditional manual annotation methods, the annotation efficiency can be improved by tens or even hundreds of times. For example, what would have taken several days to manually annotate a set of full-body skeleton images could be completed in just a few hours using the method of this application, greatly accelerating the construction process of the annotated dataset.
[0136] Furthermore, this embodiment improves annotation accuracy and consistency: the skeletal model generated by XCAT in this invention has high accuracy and stability, and its skeletal morphology and features are constructed based on a large amount of real human data and medical knowledge. Automated annotation algorithms annotate these accurate models, avoiding the subjective errors and inconsistencies that may occur with manual annotation. Verification has shown that the annotation data generated using the method of this application has an accuracy improvement of more than 20% compared to traditional manual annotation, and annotation consistency is greatly guaranteed, providing a solid data foundation for training high-precision whole-body bone segmentation AI models.
[0137] Furthermore, this embodiment enhances the diversity of labeled data: XCAT can flexibly adjust various parameters to simulate full-body skeletal models of humans with different heights, weights, ages, genders, and various skeletal diseases. This allows for the generation of rich and diverse labeled data, covering a wide range of possible human skeletal morphologies and pathological conditions. By training the AI model using this diverse data, the model can learn a broader range of skeletal features, thereby significantly improving the segmentation ability of bones from different individuals. In practical clinical applications, the accuracy of bone segmentation for complex cases and special individuals is improved by more than 30% compared to existing technologies.
[0138] Based on any of the above embodiments, when performing activity setting operations on each region in the human body model image to obtain the activity image corresponding to the human body model image, the activity of each region in the human body model image can be set to the same activity value to obtain the activity image corresponding to the human body model image.
[0139] In practical implementation, when automatically segmenting the skeletal region, it is necessary to set the activity values of bones in different parts of the human body model. One approach is to set the entire skeleton to the same activity value, that is, to set the activity values of all regions in the human body model image to the same value, in order to achieve the activity value setting more quickly. A diagram illustrating the uniform activity assignment of the entire skeleton can be found here. Figure 4 As shown.
[0140] In another implementation, when performing activity setting operations on each region in the human body model image to obtain the activity image corresponding to the human body model image, the activity of each region in the human body model image can be set to the actual activity value when the region is scanned, thus obtaining the activity image corresponding to the human body model image; wherein, the activity of different regions may be the same or different.
[0141] Specifically, in real-world scenarios, the activity values of different regions, such as different bones, may vary. To more closely resemble actual image acquisition scenarios, a SPECT bone scan of the human body can be performed beforehand to obtain the corresponding acquired images. Then, the actual activity values of each region in these images, such as the bones, can be analyzed. The activity values of each region in the human model image can be set to the actual activity values of the regions during the scanning operation, so that the set activity values closely approximate the true activity of the regions. In one example, differentiated values can be assigned based on different bone regions such as ribs, spine, skull, and leg bones, with different regions having the same or different activity values.
[0142] In this embodiment, an appropriate activity setting method can be selected according to actual needs to realize the setting operation of bone activity.
[0143] Based on any of the above embodiments, generating a multi-angle projection image using attenuation images and activity images may include:
[0144] Based on the target device type, set the detector parameters, collimator parameters, radioactive source activity, acquisition distance, and acquisition angle of the projection program. Input the attenuation image and activity image into the projection program so that the projection program can use the attenuation image and activity image to simulate the signal detection and data conversion process when the target device acquires images, and obtain a multi-angle projection image in binary format.
[0145] Specifically, Monte Carlo simulation algorithms or professional medical image projection simulation programs (such as GATE, SIMIND, etc.) are used to simulate the acquisition process of the target medical imaging equipment (SPECT, PET, CT, MRI, etc.) and generate projection images. The specific steps are as follows:
[0146] Configure relevant equipment parameters: Based on the type of target medical imaging equipment, set the detector parameters (such as the number of detectors, detection efficiency, resolution, etc.), collimator parameters (such as collimator type, aperture size, etc.), radiation source activity (simulating the actual clinical injection dose), acquisition distance (distance between the detector and the human model), and acquisition angle (e.g., 360-degree circular acquisition, acquiring once at intervals of 1-5 degrees, generating 360-72 projected images). In one example, taking SPECT as the target medical imaging equipment, the detector parameters, collimator parameters, radiation source activity, and acquisition distance are set according to the parameters of the actual SPECT equipment during operation.
[0147] Run the simulation program: Input the generated attenuation image and activity image into the projection program (such as Monte Carlo simulation algorithm or professional medical image projection simulation program) to simulate the signal detection and data conversion process when the target medical imaging equipment acquires data, and generate a multi-angle projection image in binary format. The imaging effect of this multi-angle projection image is binary data and is consistent with the projection image actually acquired by the clinical equipment.
[0148] In this embodiment, by setting different detector parameters, collimator parameters, radioactive source activity, and acquisition distance, a clearer multi-angle projection image of the real SPECT acquisition is generated.
[0149] In summary, this application automatically generates SPECT images of simulated bone scans to form a training set. At the same time, it uses these training data to build a targeted AI segmentation network to achieve automatic segmentation of skeletal regions in nuclear medicine bone scan images, avoiding the workload of doctors in annotation, improving efficiency, and enhancing the convenience of clinical diagnosis.
[0150] Based on the above-described embodiments of the image segmentation method, another embodiment of this application provides an image segmentation apparatus, referring to... Figure 9 It can include:
[0151] Image acquisition module 11 is used to acquire the target image to be segmented.
[0152] Image generation module 12 is used to generate image segmentation models;
[0153] Image segmentation module 13 is used to input the target image into the image segmentation model to obtain the image segmentation result of the target image;
[0154] The image generation module 12 includes:
[0155] The image generation submodule is used to generate multiple human model images with different attribute information;
[0156] The attenuation image determination submodule is used to determine the attenuation image corresponding to the human model image based on the attenuation data of each region in the human model image.
[0157] The activity image determination submodule is used to perform activity setting operations on each region in the human body model image to obtain the activity image corresponding to the human body model image;
[0158] The projection image determination submodule is used to generate multi-angle projection images using attenuation images and activity images;
[0159] The image adjustment submodule is used to perform image adjustment operations on multi-angle projected images to obtain the target projected image;
[0160] The model training submodule is used to train an image segmentation model based on the target projection image and the label information of each region in the target projection image.
[0161] In one implementation, the image generation submodule is specifically used for:
[0162] The content of the attribute information in the parameter file of the simulation software was adjusted multiple times to obtain multiple attribute information; the attribute information includes at least one of gender, height, weight and organ size; a human body model image corresponding to each attribute information was generated; each organ in the human body model image is configured with corresponding index information.
[0163] In one implementation, the attenuation image determination submodule is specifically used for:
[0164] By setting the photon energy and voxel size of each region in the human body model image, the attenuation data of each region to X-rays is obtained; based on the attenuation data of each region to X-rays, the attenuation image of each region in the human body model image is determined.
[0165] In one implementation, the activity image determination submodule is specifically used for:
[0166] The activity of each region in the human model image is set to the same activity value to obtain the activity image corresponding to the human model image; or, the activity of each region in the human model image is set to the actual activity value when the region is scanned to obtain the activity image corresponding to the human model image; wherein, the activity of different regions may be the same or different.
[0167] In one implementation, the projected image determination submodule is specifically used for:
[0168] Based on the target device type, set the detector parameters, collimator parameters, radioactive source activity, acquisition distance, and acquisition angle of the projection program; input the attenuation image and activity image into the projection program so that the projection program can use the attenuation image and activity image to simulate the signal detection and data conversion process when the target device performs image acquisition, and obtain a multi-angle projection image in binary format.
[0169] In one implementation, the image adjustment submodule is specifically used for:
[0170] Gaussian smoothing is applied to the multi-angle projection image to obtain the intermediate projection image; noise is added to the intermediate projection image to obtain the target projection image.
[0171] In one implementation, the model training submodule is specifically used for:
[0172] Based on the index information of each organ configuration in the human model image, the label information of the corresponding region in the target projection image is determined. The label information is the region classification result of the region. Using the target projection image and the label information of each region in the target projection image, the image segmentation model is trained until the training stops when the stopping condition is met. During the training operation of the image segmentation model, the parameters of the image segmentation model are continuously adjusted through the backpropagation algorithm, so that the image segmentation model learns the features and segmentation patterns of different regions.
[0173] In this embodiment, a target image to be segmented is acquired and input into an image segmentation model to obtain the image segmentation result. Since the image segmentation model is trained based on the target projection image and the label information of each region within the target projection image, it achieves high segmentation accuracy, thereby improving the accuracy of target image segmentation. Furthermore, this application generates multiple human model images with different attribute information. Based on the attenuation data of each region in the human model images, an attenuation image corresponding to the human model image is determined. Activity settings are performed on each region in the human model images to obtain the corresponding activity images. Using the attenuation images and activity images, multi-angle projection images are generated. Image adjustment operations are performed on the multi-angle projection images to obtain the target projection image. Through these steps, a large number of projection images with different individual characteristics can be quickly generated, and the obtained target projection images are close to the projection images obtained from actual scanning operations. This provides a large number of effective training samples for training the image segmentation model. Using these training samples for model training improves the accuracy of the image segmentation model during image segmentation operations.
[0174] It should be noted that the working process of each module and sub-module in this embodiment is described in the corresponding descriptions in the above embodiments.
[0175] This application also provides an electronic device, including at least one processor and a memory connected to the processor, wherein:
[0176] Memory is used to store computer programs;
[0177] The processor is used to execute computer programs so that the electronic device can implement the image segmentation method described above.
[0178] refer to Figure 10 The diagram illustrates a structural schematic suitable for implementing the electronic device in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 10 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.
[0179] like Figure 10 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. When the electronic device is powered on, the RAM 603 also stores various programs and data required for the operation of the electronic device. The processing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.
[0180] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, memory cards, hard drives, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 10 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.
[0181] This application also provides a computer program product including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the image segmentation methods provided in this application.
[0182] This application also provides a computer-readable storage medium that carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the image segmentation methods provided in this application.
[0183] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.
[0184] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.
[0185] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.
[0186] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).
Claims
1. An image segmentation method characterized by, include: Obtain the target image for image segmentation; The target image is input into the image segmentation model to obtain the image segmentation result of the target image; The generation process of the image segmentation model includes: Generate multiple human body model images with different attribute information; Based on the attenuation data of each region in the human body model image, determine the attenuation image corresponding to the human body model image; The activity level of each region in the human body model image is set to obtain the activity image corresponding to the human body model image. Using the attenuation image and the activity image, a multi-angle projection image is generated; Gaussian smoothing is performed on the multi-angle projection image to obtain the intermediate projection image; Poisson noise is added to the intermediate projection image to obtain the target projection image. The process of adding Poisson noise includes: adjusting the signal value range of the intermediate projection image, converting the image pixel values into a non-negative signal proportional to the photon count, and obtaining the parameters of the Poisson distribution; wherein, if the intermediate projection image is a normalized floating-point number, it is scaled to an integer range; if the intermediate projection image is an integer grayscale image with negative values, it is shifted to a non-negative range; according to the parameters, a random value following a Poisson distribution is generated for each pixel, and then the original pixel value is replaced or superimposed on the original pixel value. An image segmentation model is trained based on the target projection image and the label information of each region in the target projection image.
2. The image segmentation method of claim 1, wherein, Generate multiple human body model images with different attribute information, including: The content of the attribute information in the parameter file of the simulation software was adjusted multiple times to obtain multiple attribute information; the attribute information includes at least one of gender, height, weight and organ size; Generate a human model image corresponding to each of the aforementioned attribute information; each organ in the human model image is configured with corresponding index information.
3. The image segmentation method of claim 1, wherein, Based on the attenuation data of each region in the human body model image, determine the attenuation image corresponding to the human body model image, including: By setting the photon energy and voxel size of each region in the human body model image, the attenuation data of X-rays for each region is obtained. Based on the attenuation data of X-rays in each of the aforementioned regions, attenuation images of each region in the human body model image are determined.
4. The image segmentation method according to claim 1, characterized in that, The activity level of each region in the human body model image is set to obtain the activity image corresponding to the human body model image, including: The activity of each region in the human body model image is set to the same activity value to obtain the activity image corresponding to the human body model image; Alternatively, the activity of each region in the human body model image can be set to the actual activity value of the region during the scanning operation to obtain the activity image corresponding to the human body model image; wherein the activity of different regions may be the same or different.
5. The image segmentation method according to claim 1, characterized in that, Using the attenuation image and the activity image, a multi-angle projection image is generated, including: Based on the target device type, set the detector parameters, collimator parameters, radioactive source activity, acquisition distance, and acquisition angle of the projection program; The attenuation image and the activity image are input into the projection program so that the projection program uses the attenuation image and the activity image to simulate the signal detection and data conversion process when the target device performs image acquisition, and obtains a multi-angle projection image in binary format.
6. The image segmentation method according to claim 2, characterized in that, Based on the target projection image and the label information of each region in the target projection image, an image segmentation model is trained, including: Based on the index information of each organ configuration in the human model image, the label information of the corresponding region in the target projection image is determined, and the label information is the region classification result of the region; The image segmentation model is trained using the target projection image and the label information of each region in the target projection image until the training stops when the stopping condition is met. During the training process, the parameters of the image segmentation model are continuously adjusted through the backpropagation algorithm, so that the image segmentation model learns the features and segmentation patterns of different regions.
7. An image segmentation apparatus, characterized in that, include: The image acquisition module is used to acquire the target image to be segmented. Image generation module, used to generate image segmentation models; The image segmentation module is used to input the target image into the image segmentation model to obtain the image segmentation result of the target image; The image generation module includes: The image generation submodule is used to generate multiple human model images with different attribute information; The attenuation image determination submodule is used to determine the attenuation image corresponding to the human body model image based on the attenuation data of each region in the human body model image; The activity image determination submodule is used to perform activity setting operations on each region in the human body model image to obtain the activity image corresponding to the human body model image; The projection image determination submodule is used to generate a multi-angle projection image using the attenuation image and the activity image; The image adjustment submodule is used to perform Gaussian smoothing on the multi-angle projected image to obtain an intermediate projected image; and to add Poisson noise to the intermediate projected image to obtain a target projected image. The process of adding Poisson noise includes: adjusting the signal value range of the intermediate projected image, converting the image pixel values into a non-negative signal proportional to the photon count, and obtaining parameters of the Poisson distribution; wherein, if the intermediate projected image is a normalized floating-point number, it is scaled to an integer range; if the intermediate projected image is an integer grayscale image with negative values, it is shifted to a non-negative range; according to the parameters, a random value following a Poisson distribution is generated for each pixel, and then the original pixel value is replaced or superimposed on the original pixel value. The model training submodule is used to train an image segmentation model based on the target projection image and the label information of each region in the target projection image.
8. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the image segmentation method as described in any one of claims 1 to 6.
9. A computer storage medium, characterized in that, The storage medium carries one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the image segmentation method as described in any one of claims 1 to 6.