An automatic segmentation system and method for femoral head images

Abstract

This invention discloses an automatic segmentation system and method for femoral head images, belonging to the field of computer vision and image processing analysis technology. The system includes: a data acquisition module for constructing multimodal data based on femoral head images; a preprocessing module for acquiring preprocessed images; a data augmentation module for performing online data augmentation based on the preprocessed images to obtain augmented images; an automatic segmentation module for constructing an initial image segmentation model and obtaining necrotic region prediction results based on the initial image segmentation model and the augmented images; and a clustering analysis module for extracting features based on the necrotic region prediction results and a preset learning algorithm to obtain radiomics features, feature clusters, and clustering staging suggestions. The automatic segmentation system and method for femoral head images provided by this invention can solve the problems of insufficient segmentation efficiency and accuracy in existing technologies, achieving a comprehensive and accurate reflection of femoral head image features and effectively improving image segmentation accuracy.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing analysis technology, and in particular to an automatic segmentation system and method for femoral head images. Background Technology

[0002] Existing staging techniques for avascular necrosis of the femoral head (ONFH) mainly rely on the experience of clinicians combined with imaging examinations (such as MRI, X-ray, etc.) to perform necrosis image analysis by manually delineating regions of interest (ROI). Traditional segmentation methods usually employ the following techniques: (1) Manual segmentation and annotation: Radiologists or orthopedic surgeons manually annotate necrotic lesions and cartilage regions using software (such as ITK-SNAP), which relies on subjective experience, is time-consuming, and has poor consistency; (2) Single-modal image analysis: Most studies are based on a single MRI sequence (such as T1WI or T2 fat suppression) for qualitative assessment, lacking comprehensive quantitative analysis of multimodal data; (3) Traditional machine learning methods: Staging is performed using superficial features (such as shape and texture) combined with classifiers (such as SVM), but feature extraction relies on manual design and has limited generalization ability; (4) Rule-based staging systems: such as ARCO staging or JIC classification, which only rely on the macroscopic manifestations of images (such as lesion extent and cartilage collapse), and do not make full use of the microscopic features of radiomics.

[0003] However, existing technologies suffer from insufficient segmentation efficiency and accuracy. Traditional segmentation techniques typically rely on manual or semi-automatic segmentation, depending on physician experience, which is time-consuming and prone to human error, especially in multi-center datasets where equipment differences can lead to low annotation consistency. Furthermore, traditional segmentation algorithms (such as region growing) are sensitive to noise and struggle to handle complex anatomical structures (such as the junction of cartilage and necrotic lesions). Secondly, manually designed radiomics features (such as GLCM texture) can only capture limited patterns and cannot fully characterize the heterogeneity of ONFH (such as early changes in small necrotic lesions). The correlation between features and clinical staging depends on expert experience and lacks data-driven automated correlation analysis, indicating a limitation in feature extraction. Further, traditional machine learning methods suffer from poor model generalization ability, requiring feature engineering readjustment for different devices or scanning parameters, resulting in significant performance degradation when applied across centers. Supervised learning relies on large amounts of labeled data, while accurate annotation of ONFH data is costly and scarce. Finally, existing staging systems are often clinically impractical; for example, ARCO only provides static classification and lacks automated tools, making large-scale screening difficult. Therefore, there is an urgent need for an analysis system that addresses the shortcomings of existing technologies in femoral head image analysis, and solves the technical problems of existing technologies in ensuring the consistency of multi-center data, inability to fully capture the complex heterogeneity of ONFH, unstable performance of traditional machine learning methods under different scanning devices or parameters, and the inability of existing staging systems to reflect the dynamic changes of microscopic image features. Summary of the Invention

[0004] This invention provides an automatic segmentation system and method for femoral head images, which can solve the problems of insufficient segmentation efficiency and accuracy in the prior art, achieve a comprehensive and accurate reflection of the femoral head image features, and effectively improve the image segmentation accuracy.

[0005] This invention provides an automatic segmentation system for femoral head images, comprising:

[0006] The data acquisition module is used to construct multimodal data based on femoral head images;

[0007] The preprocessing module is used to perform orientation normalization, voxel resampling, and normalization processing on the multimodal data to obtain a preprocessed image;

[0008] The data augmentation module is used to perform online data augmentation based on the preprocessed image to obtain an enhanced image;

[0009] An automatic segmentation module is used to construct an initial image segmentation model and build a historical multimodal dataset based on historical femoral head images and related historical image data. It then obtains corresponding real label data based on the historical multimodal dataset and a preset label dataset. Furthermore, it performs orientation standardization, voxel resampling, and normalization on the historical multimodal data and real label data in the historical multimodal dataset to obtain historical preprocessed images and preprocessed label data. Online data augmentation is then performed on the historical preprocessed images and preprocessed label data to obtain historical enhanced images and enhanced label data. Finally, based on the initial image segmentation model and the historical enhanced images, it obtains historical necrosis region prediction results. Based on the historical necrosis region prediction results, enhanced label data, and a preset hybrid loss function, it obtains hybrid loss data. Based on the hybrid loss data, it optimizes and adjusts the weights of the initial image segmentation model to obtain an optimized image segmentation model. Finally, based on the optimized image segmentation model and the enhanced images, it obtains necrosis region prediction results.

[0010] The clustering analysis module is used to extract features based on the necrotic region prediction results and a preset learning algorithm to obtain radiomics features, perform deep clustering analysis based on the radiomics features to obtain feature clusters, and obtain clustering analysis suggestions based on the feature clusters.

[0011] The results output module is used to output multidimensional necrosis image analysis results based on the necrosis region prediction results, radiomics features, and clustering staging suggestions.

[0012] This invention provides an automatic segmentation system for femoral head images. First, a data acquisition module constructs multimodal data, integrating femoral head images and related data to provide a richer information foundation for subsequent analysis, avoiding the limitations of a single data dimension and comprehensively reflecting the imaging characteristics of femoral head necrosis. Next, a preprocessing module performs orientation standardization, voxel resampling, and normalization to unify the format and characteristics of data from different sources, reducing interference caused by differences in data acquisition and providing standardized data input for subsequent model processing and analysis, improving the system's adaptability to multi-source data. Then, a data augmentation module performs online enhancement on the preprocessed images, increasing data diversity so that the subsequent automatic segmentation model can adapt to image changes under different imaging conditions, improving the generalization ability of the initial image segmentation model to complex scenes and reducing the risk of overfitting. An initial image segmentation model was then constructed using an automatic segmentation module, and a multimodal training set was built based on historical data. After preprocessing and data augmentation of the modal training set, the model was used to obtain initial historical necrotic region prediction results. Then, a hybrid loss function was used to optimize the model prediction results, thereby completing the optimized training of the image segmentation model, improving the model's sensitivity to the segmentation of small targets, and further enhancing the robustness of the system. Next, based on the optimized image segmentation model and the augmented image, the necrotic region prediction results were obtained, realizing the automated identification of necrotic regions, avoiding the subjectivity and time-consuming problems of manual segmentation, and improving segmentation efficiency and consistency of results. Finally, the result output module integrated the necrotic region prediction results, radiomics features, and clustering analysis suggestions to output multidimensional analysis results, providing comprehensive and structured information support for femoral head image analysis.

[0013] Furthermore, the preprocessing module includes: a normalization submodule, used to construct a three-dimensional coordinate system and perform orientation normalization processing based on the three-dimensional coordinate system and multimodal data to obtain a normalized image; a resampling submodule, used to perform voxel resampling processing based on the normalized image and a preset bilinear interpolation algorithm to obtain a resampled image; and a normalization submodule, used to obtain a normalized image based on the resampled image and a preset intensity normalization algorithm, and use the normalized image as a historical preprocessed image.

[0014] The above scheme eliminates the impact of orientation differences on segmentation by constructing a three-dimensional coordinate system with different scanning positions; and it solves the resolution difference problem by standardizing voxel resampling of the standardized image, providing consistent input for subsequent model training and improving segmentation accuracy.

[0015] Furthermore, the preprocessing module also includes a background removal submodule, which is used to: obtain an image foreground mask based on the normalized image and a preset foreground mask generation algorithm, obtain a background removal image based on the image foreground mask and the normalized image, and use the background removal image as the preprocessed image.

[0016] In the above scheme, the background removal submodule removes the background by generating a foreground mask and cropping non-interest areas, reducing computation and improving processing efficiency; it focuses on cartilage and necrotic lesion areas to avoid background noise interference, ensuring that subsequent feature extraction is only for the target area and improving analysis accuracy.

[0017] Furthermore, the data enhancement module includes:

[0018] The data augmentation submodule is used to obtain random subvolume samples of the image based on the preprocessed image;

[0019] The online data augmentation submodule is used to obtain an enhanced image by performing spatial adjustment, random cropping, and image intensity adjustment based on random sub-volume samples of the image.

[0020] In the above scheme, random subvolume sampling and online data augmentation (such as spatial adjustment, cropping, intensity adjustment, etc.) are performed through data augmentation submodule and online data augmentation submodule to simulate different imaging conditions (such as body position and equipment differences), enhance the model's adaptability to multi-center data, and improve generalization performance; balance positive and negative samples, reduce overfitting, and enable the system model to maintain stable segmentation effect on the test set.

[0021] Furthermore, the automatic segmentation module includes:

[0022] The data acquisition submodule is used to construct a historical multimodal dataset based on historical femoral head images and related historical image data, and to acquire corresponding real label data based on the historical multimodal dataset and the preset label dataset.

[0023] The preprocessing submodule is used to perform orientation standardization, voxel resampling, and normalization processing on historical multimodal data and real label data in the historical multimodal dataset to obtain historical preprocessed images and preprocessed label data.

[0024] The data augmentation submodule is used to perform online data augmentation based on the historical preprocessed images and preprocessed label data, and to obtain historical augmented images and augmented label data;

[0025] The model building submodule is used to build an initial image segmentation model and obtain the prediction results of historical necrotic regions based on the initial image segmentation model and historical enhanced images;

[0026] The model optimization submodule is used to obtain mixed loss data based on the historical necrotic region prediction results, enhanced label data and preset mixed loss function, and to optimize and adjust the weights of the initial image segmentation model based on the mixed loss data to obtain an optimized image segmentation model.

[0027] An automatic segmentation submodule is used to obtain prediction results of necrotic regions based on the optimized image segmentation model and the enhanced image.

[0028] The above scheme constructs a multimodal training set from historical data through a preprocessing submodule, and performs preprocessing and data augmentation on the modality training set in conjunction with a data augmentation submodule, providing a good data foundation for subsequent model processing. Then, the model is constructed and optimized through a model building submodule and a model optimization submodule, which are used to achieve automatic segmentation. An automatic segmentation module is designed that uses a small amount of labeled data combined with semi-supervised learning to reduce the dependence on a large number of accurate labels and reduce labeling costs. A hybrid loss function is used to optimize the model, improve the segmentation sensitivity of small targets (such as early necrotic foci), and enhance the robustness of the model.

[0029] Furthermore, the preprocessing submodule includes:

[0030] A standardized unit is used to construct a three-dimensional coordinate system and perform orientation standardization processing based on the three-dimensional coordinate system, historical multimodal data and real label data to obtain historical standardized images and standardized label data.

[0031] The resampling unit is used to perform voxel resampling processing based on the historical standardized image and a preset bilinear interpolation algorithm to obtain the historical resampled image; it is also used to obtain resampled label data based on the standardized label data and a preset nearest neighbor interpolation algorithm.

[0032] The normalization unit is used to obtain a historical normalized image based on the historical resampled image and a preset intensity normalization algorithm, and to use the historical normalized image as a historical preprocessed image.

[0033] Furthermore, the preprocessing submodule also includes a background removal unit, which is used to: obtain a historical image foreground mask based on the historical normalized image and a preset foreground mask generation algorithm, obtain a background removal historical image based on the historical image foreground mask and the historical normalized image, and use the background removal historical image as the historical preprocessed image.

[0034] Furthermore, the data augmentation submodule includes:

[0035] The data augmentation unit is used to obtain random sub-volume samples of historical images based on the historical preprocessed images, and to obtain random sub-volume samples of labels based on the preprocessed label data;

[0036] The online data augmentation unit is used to obtain historical enhanced images by performing spatial adjustment, random cropping, and image intensity adjustment based on the random sub-volume samples of the historical images; it is also used to obtain enhanced label data by performing spatial adjustment based on the random sub-volume samples of the labels.

[0037] In the above scheme, the preprocessing submodule uses bilinear interpolation and label nearest neighbor interpolation to standardize historical data and resample voxels, ensuring spatial consistency between historical training data and corresponding label data, and guaranteeing the accuracy of input data for subsequent image segmentation models. A background removal unit removes background from relevant historical images, reducing invalid information in the historical image data, allowing training to focus on the target region, effectively improving the model's learning performance on regions of interest and computational efficiency. Furthermore, a data augmentation submodule performs synchronous data augmentation of historical images and labels, ensuring the spatial correspondence between images and labels, avoiding misalignment during augmentation, and ensuring the effectiveness of the augmented data.

[0038] Furthermore, the clustering analysis module includes:

[0039] The feature extraction submodule is used to obtain radiomics features based on the prediction results of the necrotic region and a preset image analysis algorithm;

[0040] The clustering analysis submodule is used to initialize several centroids and perform deep clustering analysis based on a preset clustering algorithm: an initial cluster is obtained based on the radiomics features, the centroids, and a preset allocation method; the current centroid is obtained based on the initial cluster; and the deep clustering analysis is repeatedly performed based on the current centroid to obtain the target cluster and the target centroid. When the target centroid is determined to meet the preset iteration requirements, a feature cluster is obtained based on the target cluster.

[0041] The region synthesis submodule is used to merge clustering results and analyze cluster features based on the feature clusters to obtain clustering staging suggestions.

[0042] In the above scheme, feature extraction is combined with clustering and region synthesis through a cluster analysis module to comprehensively capture image information. Based on the characteristics of radiomics, clustering staging suggestions are obtained, which can reflect the dynamic changes of microscopic image features.

[0043] This invention also provides an automatic segmentation method for femoral head images, implemented using the aforementioned automatic segmentation system for femoral head images, comprising:

[0044] Constructing multimodal data based on femoral head images;

[0045] Based on the multimodal data, orientation normalization, voxel resampling, and normalization are performed to obtain a preprocessed image;

[0046] Online data augmentation is performed on the preprocessed image to obtain an enhanced image;

[0047] An initial image segmentation model is constructed, and a historical multimodal dataset is built based on historical femoral head images and related historical image data. Real label data is obtained based on the historical multimodal dataset and a preset label dataset. Orientation standardization, voxel resampling, and normalization are performed on the historical multimodal data and real label data in the historical multimodal dataset to obtain historical preprocessed images and preprocessed label data. Online data augmentation is performed on the historical preprocessed images and preprocessed label data to obtain historical enhanced images and enhanced label data. Then, based on the initial image segmentation model and the historical enhanced images, the prediction result of the historical necrosis region is obtained. Based on the historical necrosis region prediction result, the enhanced label data, and a preset hybrid loss function, hybrid loss data is obtained. The initial image segmentation model is then optimized and adjusted based on the hybrid loss data to obtain an optimized image segmentation model. Finally, based on the optimized image segmentation model and the enhanced images, the prediction result of the necrosis region is obtained.

[0048] Based on the predicted necrotic region and the preset learning algorithm, feature extraction is performed to obtain radiomics features. Based on the radiomics features, deep clustering analysis is performed to obtain feature clusters. Based on the feature clusters, clustering analysis suggestions are obtained.

[0049] Based on the predicted necrotic area, radiomics features, and clustering staging suggestions, multidimensional necrotic image analysis results are output.

[0050] This invention provides an automatic segmentation method for femoral head images. By constructing multimodal data and integrating femoral head images and related data, it provides a richer information foundation for subsequent analysis, avoiding the limitations of a single data dimension and comprehensively reflecting the imaging characteristics of femoral head necrosis. Through orientation standardization, voxel resampling, and normalization, it unifies the format and characteristics of data from different sources, reduces interference caused by differences in data acquisition, provides standardized data input for subsequent model processing and analysis, and improves the system's adaptability to multi-source data. Next, online enhancement is performed on the preprocessed images to increase data diversity, enabling the subsequent automatic segmentation model to adapt to image changes under different imaging conditions, improving the generalization ability of the initial image segmentation model to complex scenes, and reducing the risk of overfitting. A hybrid loss function is constructed and a hybrid loss function is used to optimize and train the image segmentation model. Based on the optimized image segmentation model, the predicted result of the necrotic region is obtained, realizing the automated identification of the necrotic region, avoiding the subjectivity and time-consuming problems of manual segmentation, and improving segmentation efficiency and result consistency. Finally, the predicted result of the necrotic region, radiomics features, and cluster analysis suggestions are integrated to output multidimensional analysis results, providing comprehensive and structured information support for femoral head image analysis. Attached Figure Description

[0051] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of an automatic segmentation method for femoral head images provided in this embodiment;

[0053] Figure 2 This is a schematic diagram illustrating the changing trends of the loss function and the Dice coefficients on the validation set during the model training process provided in this embodiment;

[0054] Figure 3 This is a schematic diagram of the t-SNE visualization of K-means tetraclustering results provided in this embodiment. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0056] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0057] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0058] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0059] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0060] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0061] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0062] This embodiment provides an automatic segmentation system for femoral head images, including:

[0063] The data acquisition module is used to construct multimodal data based on femoral head images;

[0064] The preprocessing module is used to perform orientation normalization, voxel resampling, and normalization processing on the multimodal data to obtain a preprocessed image;

[0065] The data augmentation module is used to perform online data augmentation based on the preprocessed image to obtain an enhanced image;

[0066] An automatic segmentation module is used to construct an initial image segmentation model and build a historical multimodal dataset based on historical femoral head images and related historical image data. It then obtains corresponding real label data based on the historical multimodal dataset and a preset label dataset. Furthermore, it performs orientation standardization, voxel resampling, and normalization on the historical multimodal data and real label data in the historical multimodal dataset to obtain historical preprocessed images and preprocessed label data. Online data augmentation is then performed on the historical preprocessed images and preprocessed label data to obtain historical enhanced images and enhanced label data. Finally, based on the initial image segmentation model and the historical enhanced images, it obtains historical necrosis region prediction results. Based on the historical necrosis region prediction results, enhanced label data, and a preset hybrid loss function, it obtains hybrid loss data. Based on the hybrid loss data, it optimizes and adjusts the weights of the initial image segmentation model to obtain an optimized image segmentation model. Finally, based on the optimized image segmentation model and the enhanced images, it obtains necrosis region prediction results.

[0067] The clustering analysis module is used to extract features based on the necrotic region prediction results and a preset learning algorithm to obtain radiomics features, perform deep clustering analysis based on the radiomics features to obtain feature clusters, and obtain clustering analysis suggestions based on the feature clusters.

[0068] The results output module is used to output multidimensional necrosis image analysis results based on the necrosis region prediction results, radiomics features, and clustering staging suggestions.

[0069] This embodiment provides an automatic segmentation system for femoral head images. First, a data acquisition module constructs multimodal data, integrating femoral head images and related data to provide a richer information foundation for subsequent analysis, avoiding the limitations of a single data dimension and comprehensively reflecting the imaging characteristics of femoral head necrosis. Next, a preprocessing module performs orientation standardization, voxel resampling, and normalization to unify the format and characteristics of data from different sources, reducing interference caused by differences in data acquisition and providing standardized data input for subsequent model processing and analysis, thus improving the system's adaptability to multi-source data. Then, a data augmentation module performs online enhancement on the preprocessed images, increasing data diversity so that the subsequent automatic segmentation model can adapt to image changes under different imaging conditions, improving the generalization ability of the initial image segmentation model to complex scenes and reducing the risk of overfitting. Then, an initial image segmentation model is constructed through an automatic segmentation module, and a multimodal training set is built based on historical data. After preprocessing and data augmentation of the modal training set, the model is used to obtain the initial historical necrotic region prediction results. Next, a hybrid loss function is used to optimize the model prediction results, thereby completing the optimized training of the image segmentation model, improving the model's sensitivity to the segmentation of small targets, and further enhancing the robustness of the system. Then, based on the optimized image segmentation model and augmented images after optimization training, the necrotic region prediction results are obtained, realizing the automated identification of necrotic regions, avoiding the subjectivity and time-consuming problems of manual segmentation, and improving segmentation efficiency and consistency of results. Finally, the result output module integrates the necrotic region prediction results, radiomics features, and clustering analysis suggestions to output multidimensional analysis results, providing comprehensive and structured information support for femoral head image analysis.

[0070] Optionally, the preprocessing module includes: a normalization submodule, used to construct a three-dimensional coordinate system and perform orientation normalization processing based on the three-dimensional coordinate system and multimodal data to obtain a normalized image; a resampling submodule, used to perform voxel resampling processing based on the normalized image and a preset bilinear interpolation algorithm to obtain a resampled image; and a normalization submodule, used to obtain a normalized image based on the resampled image and a preset intensity normalization algorithm, and use the normalized image as a historical preprocessed image.

[0071] In the specific implementation process, this embodiment eliminates the difference in scanning position by constructing a three-dimensional coordinate system-spatial orientation; it uses bilinear interpolation (image) and nearest neighbor interpolation (label) to standardize images of different resolutions to a voxel spacing of 1×1×1 mm³; and it linearly scales the image intensity to the range of [300, 600] Hounsfield units (HU) to achieve intensity normalization and solve the problem of differences between multi-center equipment.

[0072] Optionally, the preprocessing module further includes a background removal submodule, which is used to: obtain an image foreground mask based on the normalized image and a preset foreground mask generation algorithm, obtain a background removal image based on the image foreground mask and the normalized image, and use the background removal image as the preprocessed image.

[0073] In the specific implementation process, the background removal submodule removes the background by generating a foreground mask, clips non-interested regions, reduces computation, and improves processing efficiency; it focuses on cartilage and necrotic lesion regions to avoid background noise interference, ensuring that subsequent feature extraction is only for the target region, thus improving the accuracy of analysis.

[0074] Optionally, the data enhancement module includes:

[0075] The data augmentation submodule is used to obtain random subvolume samples of the image based on the preprocessed image;

[0076] The online data augmentation submodule is used to obtain an enhanced image by performing spatial adjustment, random cropping, and image intensity adjustment based on random sub-volume samples of the image.

[0077] In the specific implementation process, random subvolume sampling and online data augmentation (such as spatial adjustment, cropping, intensity adjustment, etc.) are performed through the data augmentation submodule and the online data augmentation submodule to simulate different imaging conditions (such as body position and equipment differences), enhance the model's adaptability to multi-center data, and improve generalization performance; balance positive and negative samples, reduce overfitting, and enable the system model to maintain a stable segmentation effect on the test set.

[0078] Optionally, the automatic segmentation module includes:

[0079] The data acquisition submodule is used to construct a historical multimodal dataset based on historical femoral head images and related historical image data, and to acquire corresponding real label data based on the historical multimodal dataset and the preset label dataset.

[0080] The preprocessing submodule is used to perform orientation standardization, voxel resampling, and normalization processing on historical multimodal data and real label data in the historical multimodal dataset to obtain historical preprocessed images and preprocessed label data.

[0081] The data augmentation submodule is used to perform online data augmentation based on the historical preprocessed images and preprocessed label data, and to obtain historical augmented images and augmented label data;

[0082] The model building submodule is used to build an initial image segmentation model and obtain the prediction results of historical necrotic regions based on the initial image segmentation model and historical enhanced images;

[0083] The model optimization submodule is used to obtain mixed loss data based on the historical necrotic region prediction results, enhanced label data and preset mixed loss function, and to optimize and adjust the weights of the initial image segmentation model based on the mixed loss data to obtain an optimized image segmentation model.

[0084] An automatic segmentation submodule is used to obtain prediction results of necrotic regions based on the optimized image segmentation model and the enhanced image.

[0085] In the specific implementation process, this embodiment integrates models within the PyTorch framework when constructing the initial image segmentation model. The V-Net model is used for model integration and training optimization to obtain the final image segmentation model. The V-Net model is a 3D convolutional neural network with encoder and decoder structures, capable of processing data across the entire 3D volume. It is specifically designed for 3D medical image segmentation and can automatically segment the ROI of all input images. Subsequently, when constructing and training the image segmentation model through the model construction and optimization submodules, the acquisition of historical enhanced images involves collecting and downloading historical ONFH-related imaging data (including DR, MRI, etc.). All patient MR imaging data includes coronal T1WI sequences and T2 fat-suppressed images, all scanned using a GE 1.5T or higher MR machine. All imaging data undergoes diagnostic review, staging, and classification registration and confirmation by dedicated researchers at each center to ensure the accuracy of the image training set data. MR images of patients clinically and radiologically diagnosed with ONFH are included as the training queue for machine learning to construct the image training set. To ensure training results, this embodiment adopts the following inclusion and exclusion criteria when including images in the training queue: images that are clinically and MR imaging diagnosed as ONFH, whose ARCO stage and JIC classification can be clearly defined, and which have undergone multiple modal hip joint MR scans are included; images that do not show necrotic lesions in MR images or whose MR image quality is too poor to be measured are excluded. Finally, a multimodal ONFH MR image set is constructed as the image training set.

[0086] After acquiring historical enhanced images, the Region of Interest (ROI) is first automatically segmented (the ROI includes two regions: cartilage and necrotic lesions. The cartilage is defined by its anatomical structure, while the necrotic lesions are defined by a predefined boundary) to obtain historical image samples, thus partially delineating the ROI. Simultaneously, a limited number of samples from the historical image samples are labeled as a training set (i.e., historical multimodal data). In this embodiment, the delineation scale is calculated based on 10% of the hip data in the existing historical image samples.

[0087] Optionally, the preprocessing submodule includes:

[0088] A standardized unit is used to construct a three-dimensional coordinate system and perform orientation standardization processing based on the three-dimensional coordinate system, historical multimodal data and real label data to obtain historical standardized images and standardized label data.

[0089] The resampling unit is used to perform voxel resampling processing based on the historical standardized image and a preset bilinear interpolation algorithm to obtain the historical resampled image; it is also used to obtain resampled label data based on the standardized label data and a preset nearest neighbor interpolation algorithm.

[0090] The normalization unit is used to obtain a historical normalized image based on the historical resampled image and a preset intensity normalization algorithm, and to use the historical normalized image as a historical preprocessed image.

[0091] In this implementation, a unified RAS (Right-Anterior-Superior) coordinate system is used to construct the three-dimensional coordinate system, standardizing the anatomical orientation. This coordinate system can adjust the input medical images and corresponding label data to a unified spatial orientation. In this embodiment, the coordinate system specifies that the right side of the image is the positive X-axis, the front side is the positive Y-axis, and the top is the positive Z-axis. This standardization ensures that all images and labels are in the same spatial orientation, thereby avoiding subsequent processing problems caused by inconsistent orientations.

[0092] In the specific implementation process, when using a resampling unit to resample to a uniform voxel spacing, the voxel sizes of medical images from different sources in the historical multimodal data may differ, affecting model training. To address this uniformity, this embodiment first determines the target voxel size to be 1×1×1 mm; then, it uses bilinear interpolation to resample the standardized image data (historical standardized images) to smoothly redistribute pixel values ​​and reduce image artifacts that may occur during interpolation. The preset bilinear interpolation algorithm formula in this embodiment is as follows: ;in: It is the intensity value of the historical normalized image at voxel position x. It is a resampled, standardized historical image at a new location The intensity value, The interpolation weights, typically the weight function of bilinear interpolation, are used to calculate the intensity value at the new voxel location to obtain the historical resampled image. For standardized label data, since the labels are discrete values ​​(usually integers), nearest neighbor interpolation is used to obtain the nearest original label value, thus obtaining the resampled label data. This ensures the clarity of label boundaries and prevents incorrect label values ​​from being introduced due to interpolation. The preset nearest neighbor interpolation algorithm used in this embodiment has the following calculation formula: ,in: Indicates separation The most recent original label value.

[0093] In practical implementation, since different medical images may come from different MRI devices and have different intensity ranges, to enhance the adaptability of the image segmentation model in this embodiment and to process images from different sources, it is necessary to standardize the image intensity. The specific approach adopted in this embodiment is to linearly scale the intensity range of the historical resampled image to between [300, 600]. This range is typically the Hounsfield unit value associated with a specific tissue or lesion. To avoid the influence of extreme values ​​(e.g., very high or very low intensity values ​​may be noise), cropping is optionally performed to ensure that all pixel values ​​are within this range, resulting in a historically normalized image. The preset intensity normalization algorithm used in this embodiment is calculated as follows: ,in: It is the normalized image intensity value. and These are the maximum and minimum intensity values ​​before normalization (300 and 600, respectively). A and B are the target ranges after normalization (usually [0, 1] or [300, 600]).

[0094] Optionally, the preprocessing submodule further includes a background removal unit, which is used to: obtain a historical image foreground mask based on the historical normalized image and a preset foreground mask generation algorithm, obtain a background removal historical image based on the historical image foreground mask and the historical normalized image, and use the background removal historical image as the historical preprocessed image.

[0095] In practical implementation, background regions increase the computational burden. Therefore, in this embodiment, a foreground mask is generated by a background removal unit to remove background regions from the historical normalized image. The foreground mask is generated through threshold segmentation or simple binarization based on image intensity. The foreground mask is then used to crop the historical normalized image and corresponding label data to include only the foreground region, reducing the amount of data and focusing on the region of interest, thereby improving the efficiency of the image segmentation model.

[0096] Optionally, the data augmentation submodule includes:

[0097] The data augmentation unit is used to obtain random sub-volume samples of historical images based on the historical preprocessed images, and to obtain random sub-volume samples of labels based on the preprocessed label data;

[0098] The online data augmentation unit is used to obtain historical enhanced images by performing spatial adjustment, random cropping, and image intensity adjustment based on the random sub-volume samples of the historical images; it is also used to obtain enhanced label data by performing spatial adjustment based on the random sub-volume samples of the labels.

[0099] In the specific implementation process, in the segmentation task, data augmentation helps to improve the robustness of the model and reduce overfitting. In this embodiment, data augmentation is performed through a data augmentation unit. The specific operations include randomly extracting fixed-size sub-volumes from historical preprocessed images and preprocessed label data. The sub-volumes can contain complete ROI regions or a part of the background region, resulting in random sub-volume samples of historical images and random sub-volume samples of labels. At the same time, in order to ensure the learning effect of the model, the balance of positive and negative samples needs to be considered when randomly extracting sub-volumes to ensure that the model can learn enough target regions (positive samples) and background regions (negative samples). Then, the online data augmentation unit is used in real time during the model training process, specifically including: (1) Spatial adjustment: randomly translating, rotating or scaling the sub-volumes to simulate the differences in body shape or changes in imaging angle of different patients; (2) Random cropping: randomly cropping the image to simulate ROI regions of different sizes and enhance the model's adaptability to different ROI sizes; (3) Image intensity adjustment: slightly perturbing the image intensity, such as increasing noise or adjusting contrast, to simulate image differences under different imaging conditions. By obtaining historical augmented images and corresponding augmented label data during real-time training, the training samples can be enriched and the realism improved.

[0100] In the specific implementation process, this embodiment uses DiceCELoss (a pre-defined hybrid loss function) in the model optimization submodule for model optimization. This hybrid loss function combines Dice Loss and cross-entropy loss to improve the model's accurate segmentation of ROI regions. Dice Loss maximizes the overlap between the predicted segmented region and the true label, making it particularly suitable for imbalanced datasets, i.e., when the target region accounts for a small proportion. Cross-entropy loss calculates the difference between the model's predicted probability distribution and the actual label, suitable for multi-class problems, and helps the model distinguish the boundaries between different categories. DiceCELoss is a weighted sum of Dice Loss and cross-entropy loss. By combining these two loss functions, it maintains sensitivity to small ROI regions while ensuring the overall segmentation performance of the model. The DiceCELoss calculation formula designed in this embodiment is as follows:

[0101]

[0102] Where: N represents the batch size. and Let represent the Dice loss and cross-entropy loss for the i-th sample, respectively. α is used to adjust the balance between the two loss functions. Unlabeled pixels are assigned zero weights, allowing the model to focus on labeled regions and improving generalization across the entire dataset.

[0103] In the specific implementation process, the image segmentation model trained and optimized in this embodiment also incorporates post-processing techniques when processing input image data to obtain classification prediction results in practical applications. These include: a sliding window technique, using a 96x96x96 voxel-sized sliding window to divide the entire input image into blocks, predicting each block separately, and then stitching these blocks together to obtain a complete classification prediction result. The sliding window avoids GPU memory shortages caused by excessively large input image sizes; a morphological processing operation to remove scattered points (i.e., misclassified voxels in image space) from the obtained complete classification prediction result, for example, by removing connected regions smaller than a certain threshold through voxel connectivity analysis; and smoothing processing, using Gaussian smoothing (σ=1.0) or other filters to smooth the segmentation result, reducing edge noise and jagged edges. After the sliding window, scattered point removal, and smoothing operations, the desired dead region prediction result is obtained.

[0104] Optionally, the clustering analysis module includes:

[0105] The feature extraction submodule is used to obtain radiomics features based on the prediction results of the necrotic region and a preset image analysis algorithm;

[0106] The clustering analysis submodule is used to initialize several centroids and perform deep clustering analysis based on a preset clustering algorithm: an initial cluster is obtained based on the radiomics features, the centroids, and a preset allocation method; the current centroid is obtained based on the initial cluster; and the deep clustering analysis is repeatedly performed based on the current centroid to obtain the target cluster and the target centroid. When the target centroid is determined to meet the preset iteration requirements, a feature cluster is obtained based on the target cluster.

[0107] The region synthesis submodule is used to merge clustering results and analyze cluster features based on the feature clusters to obtain clustering staging suggestions.

[0108] In this implementation, based on the necrotic region prediction results obtained from all training samples, and through a custom-defined preset image analysis algorithm and expert knowledge base, 1834 unique handcrafted radiomic features were extracted for each sample using mathematical formulas or image processing techniques. These radiomic features are quantitative descriptions of medical images, used to capture specific patterns, textures, shapes, and intensity variations within a region of interest (ROI) in medical imaging, providing interpretable metrics that can be correlated with specific anatomical or pathological features. The radiomic features are systematically categorized into three main types: geometry (shape), first-order features (intensity), and texture. Specifically, in this embodiment, the dataset includes 14 shape features, 360 first-order features, and a comprehensive set of texture features. The three categories aim to analyze different aspects of ROI features: geometric features define the shape of the ROI, intensity features assess the brightness level of voxels, and texture features explore spatial patterns through techniques such as GLCM (Gray Co-occurrence Matrix), GLRLM (Gray Run Length Matrix), GLSZM (Gray Block Size Matrix), and NGTDM (Neighborhood Gray Difference Matrix).

[0109] In the specific implementation process, the preset clustering algorithm used in this embodiment is deep clustering analysis (K-means clustering algorithm). In the multidimensional feature space, radiomics features usually exhibit complex distribution patterns. Therefore, this embodiment uses the K-means clustering algorithm to perform clustering analysis on radiomics features. As an unsupervised learning algorithm, K-means assigns data points to K different clusters and adjusts the centroids according to a preset assignment method, i.e., the distance minimization principle. The specific process includes: Step 1, initializing K centroids (usually randomly selected); Step 2, assigning each data point to the nearest centroid to form an initial cluster; Step 3, recalculating the centroid of each initial cluster to obtain the current centroid; repeating steps 2 and 3 until the centroids no longer change significantly or the preset iteration number is reached, thus meeting the preset iteration requirements, and thus obtaining feature clusters. The optimization of the K-means algorithm is driven by the following objective function:

[0110]

[0111] In the formula: J represents the objective function, N represents the total number of data points, and K represents the number of clusters. Represents a binary indicator (if data point i belongs to the k-th cluster). =1, otherwise (0) This represents the i-th data point. Denotes the centroid of the k-th cluster. This represents the square of the Euclidean distance between data point i and centroid k. The K-means clustering algorithm used in this embodiment divides the features into four clusters to correspond to the four-stage classification standards of JIC typing and ARCO staging in ONFH. To ensure cluster compactness and minimize the impact of outliers, this embodiment first sorts the feature clusters according to their distance from their respective centroids, retaining the closest 50%; then, outliers that do not conform to the expected features are identified and removed; finally, t-SNE (t-distributed Stochastic Neighbor Embedding, a nonlinear dimensionality reduction algorithm) is used for dimensionality reduction, reducing the high-dimensional features to 2D / 3D to visualize the radiomics features, thereby obtaining the desired feature cluster classification of the necrotic region in the femoral head image.

[0112] In the specific implementation process, the different clusters generated by cluster analysis represent samples with similar radiomics features in the multidimensional feature space. In this embodiment, the features are synthesized by the region synthesis module in the following steps to generate cluster staging suggestions: (1) Cluster result merging: Samples with the same cluster ID in the feature clusters are merged to form larger and more representative groups to capture specific radiomics patterns, such as low-density areas, high-texture complexity areas, etc.; (2) Cluster feature analysis: The features of each subgroup are statistically analyzed during the synthesis process to determine whether these groups have biological or clinical interpretability. Then, a four-class classification system of "necrosis ecology" is constructed based on the cluster staging suggestions. At the same time, this embodiment also eliminates the influence of dimensions through Z-score normalization to ensure the reliability of cluster analysis.

[0113] This embodiment provides an automatic segmentation method for femoral head images, implemented using the aforementioned automatic segmentation system for femoral head images, such as... Figure 1 As shown, it includes:

[0114] S1. Constructing multimodal data based on femoral head images;

[0115] S2. Based on the multimodal data, perform orientation standardization, voxel resampling, and normalization to obtain a preprocessed image;

[0116] S3. Perform online data enhancement based on the preprocessed image to obtain an enhanced image;

[0117] S4. Construct an initial image segmentation model and build a historical multimodal dataset based on historical femoral head images and related historical image data. Obtain corresponding real label data based on the historical multimodal dataset and a preset label dataset. Perform orientation standardization, voxel resampling, and normalization processing on the historical multimodal data and real label data in the historical multimodal dataset to obtain historical preprocessed images and preprocessed label data. Perform online data augmentation on the historical preprocessed images and preprocessed label data to obtain historical augmented images and augmented label data. Then, obtain historical necrosis region prediction results based on the initial image segmentation model and historical augmented images. Obtain mixed loss data based on the historical necrosis region prediction results, augmented label data, and a preset mixed loss function. Optimize and adjust the weights of the initial image segmentation model based on the mixed loss data to obtain an optimized image segmentation model. Then, obtain necrosis region prediction results based on the optimized image segmentation model and augmented images.

[0118] S5. Based on the prediction results of the necrotic area and the preset learning algorithm, feature extraction is performed to obtain radiomics features. Based on the radiomics features, deep clustering analysis is performed to obtain feature clusters. Based on the feature clusters, clustering analysis suggestions are obtained.

[0119] S6. Based on the predicted necrotic area, radiomics features, and clustering staging suggestions, output multidimensional necrotic image analysis results.

[0120] This embodiment provides an automatic segmentation method for femoral head images. By constructing multimodal data and integrating femoral head images and related data, it provides a richer information foundation for subsequent analysis, avoiding the limitations of a single data dimension and comprehensively reflecting the imaging characteristics of femoral head necrosis. Through orientation standardization, voxel resampling, and normalization, the format and characteristics of data from different sources are unified, reducing interference caused by differences in data acquisition and providing standardized data input for subsequent model processing and analysis, thus improving the system's adaptability to multi-source data. Next, the preprocessed images are enhanced online to increase data diversity, enabling the subsequent automatic segmentation model to adapt to image changes under different imaging conditions, improving the generalization ability of the initial image segmentation model to complex scenes, and reducing the risk of overfitting. An initial image segmentation model is constructed to obtain the prediction results of necrotic areas, realizing the automatic identification of necrotic areas, avoiding the subjectivity and time-consuming problems of manual segmentation, and improving segmentation efficiency and result consistency. Finally, the prediction results of necrotic areas, radiomics features, and cluster analysis suggestions are integrated to output multidimensional analysis results, providing comprehensive and structured information support for femoral head image analysis.

[0121] To validate the system's performance, this embodiment employs a multi-center retrospective study, collecting femoral head images from patients of different origins for training and validation. Training focuses on identifying Regions of Interest (ROIs) of cartilage and necrosis. A V-Net convolutional neural network model combined with a 96x96x96 voxel sliding window is used to process the input data. During segmentation, mislabeled voxels in the image space, i.e., the "scatter plot" problem, are addressed to improve the model's accuracy and stability. This embodiment uses Epoch Average Loss and Val Mean Dice (validation set average Dice coefficients) to evaluate the training results. The trends of the loss function and validation set Dice coefficients during model training are shown below. Figure 2 As shown, where, Figure 2 (a) shows the Epoch Average Loss result. As the curve shows, the loss function decreases rapidly as training progresses, especially in the first 100 training epochs, when the model converges quickly. After about 100 training epochs, the loss function tends to stabilize and remains at a low level in subsequent training, indicating that the model has a good training effect and has reached the convergence state. Figure 2 (b) represents the change in the Val Mean Dice coefficient on the validation set, which measures the overlap between the predicted and actual segments; a higher value indicates better segmentation performance. According to Figure 2 (b) It was found that the Dice coefficient increased rapidly in the first 200 training epochs, indicating that the model learned useful features in the early stages. After more than 200 training epochs, the Dice coefficient gradually stabilized, and the model finally achieved a maximum Dice coefficient of 0.783 on the test set, indicating that the model could segment the target ROI well. These results show that, through V-Net model training, sliding window technology, and error point handling, the model constructed in this embodiment can handle large medical images and improve the reliability of predictions. The automatic segmentation system provided in this embodiment provides reliable ROI segmentation results for subsequent analysis or clinical applications, and has the ability to accurately identify the presence of cartilage and necrotic lesions in complex anatomical structures and disease environments.

[0122] This embodiment uses t-SNE to reduce the dimensionality of radiomics features and visualizes the K-means clustering results. The visualized four-cluster results are as follows: Figure 3 As shown, Figure 3The four different colored dots represent four different clusters (cluster 0, cluster 1, cluster 2, and cluster 3). The visualization results show that the four clusters are clearly separated in low-dimensional space, indicating significant differences in their characteristics. This also verifies the accuracy of the automatic segmentation and multi-dimensional necrosis image analysis results of the system in this embodiment. This embodiment names this four-cluster system the "necrosis ecology" system. Within this system, different clusters exhibit significant differences in shape, texture, and intensity characteristics, visually displaying the different states or types of femoral head necrosis images and their diversity in image representation.

[0123] This embodiment provides an automatic segmentation system for femoral head images, achieving fully automatic and efficient segmentation, significantly improving analysis efficiency and accuracy. Compared to existing techniques that rely on manual or semi-automatic segmentation by doctors, this system uses a 3D V-Net deep learning model to achieve fully automatic segmentation. Combined with sliding window technology and morphological post-processing, it reduces the analysis time for a single case from over 30 minutes to seconds. Standardized preprocessing (RAS coordinate system alignment, voxel resampling) ensures consistency of multi-center data, achieving a segmentation Dice coefficient of 0.783, significantly better than traditional methods (usually <0.7). This not only solves the problems of low efficiency and strong subjectivity of manual annotation but also accurately handles complex anatomical structures such as the junction of cartilage and necrotic lesions. This embodiment achieves comprehensive capture of disease heterogeneity through multidimensional radiomics features: Existing technologies rely on limited hand-designed features (such as GLCM texture), while this embodiment innovatively extracts 1834-dimensional features (shape, first-order statistics, higher-order textures, etc.) to comprehensively characterize the microscopic heterogeneity of ONFH through a data-driven approach. In particular, the introduction of higher-order texture features such as GLRLM and GLSZM can detect subtle image changes in early necrosis (ARCO stage I), overcoming the limitation of traditional methods that can only analyze macroscopic manifestations. This multidimensional feature system also provides a new perspective for disease mechanism research and accurate staging. A data-driven "necrosis ecology" staging system is constructed, breaking through traditional standards: Unlike static staging standards such as ARCO / JIC that rely on macroscopic image manifestations, this embodiment constructs a four-class classification system of "necrosis ecology" based on K-means clustering, and reveals the classification and subtyping corresponding to radiomics features through t-SNE dimensionality reduction visualization; new image patterns (such as high-texture complexity clusters) are discovered, and the dynamic evolution of radiomics features is tracked. It exhibits superior cross-center applicability and practicality: Addressing the limitations of existing technologies in multi-scenario applications, this embodiment significantly improves the model's adaptability to different MR devices (such as GE / Siemens) through standardized preprocessing (intensity normalization to [300, 600] HU) and online data augmentation (random noise, rotation, etc.), giving it excellent generalization ability. The system's final structured report (including segmentation maps, feature values, and clustering suggestions) makes the feature analysis of femoral head necrosis images more comprehensive, complete, visualized, and interpretable. In summary, this embodiment, through the technical route of "deep learning segmentation + radiomics features + unsupervised clustering," comprehensively surpasses existing technologies in terms of accuracy (Dice coefficient 0.783), efficiency (10x improvement in analysis speed), and interpretability (t-SNE visualization). Its semi-supervised learning strategy (DiceCELoss optimization) effectively reduces dependence on labeled data, while the dynamically updated model architecture can be continuously optimized with new data. This technology not only sets a new standard for the automatic segmentation and analysis of femoral head necrosis images, but its methodology can also be extended to the field of intelligent analysis of other bone and joint images, which has important promotional value and commercial transformation potential.

[0124] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. An automatic segmentation system for femoral head images, characterized in that, include: The data acquisition module is used to construct multimodal data based on femoral head images; The preprocessing module is used to perform orientation normalization, voxel resampling, and normalization processing on the multimodal data to obtain a preprocessed image; The data augmentation module is used to perform online data augmentation based on the preprocessed image to obtain an enhanced image; An automatic segmentation module is used to construct an initial image segmentation model and build a historical multimodal dataset based on historical femoral head images and related historical image data. It then obtains corresponding real label data based on the historical multimodal dataset and a preset label dataset. Furthermore, it performs orientation standardization, voxel resampling, and normalization processing on the historical multimodal data and real label data in the historical multimodal dataset to obtain historical preprocessed images and preprocessed label data. Finally, it performs online data augmentation based on the historical preprocessed images and preprocessed label data to obtain historical enhanced images and enhanced label data. Finally, it obtains the prediction results of historical necrosis areas based on the initial image segmentation model and the historical enhanced images. Based on the historical necrotic region prediction results, enhanced label data, and a preset mixed loss function, mixed loss data is obtained. The initial image segmentation model is then optimized and adjusted based on the mixed loss data to obtain an optimized image segmentation model. Finally, based on the optimized image segmentation model and the enhanced image, necrotic region prediction results are obtained. The clustering analysis module is used to extract features based on the necrotic region prediction results and a preset learning algorithm to obtain radiomics features, perform deep clustering analysis based on the radiomics features to obtain feature clusters, and obtain clustering analysis suggestions based on the feature clusters. The results output module is used to output multidimensional necrosis image analysis results based on the necrosis region prediction results, radiomics features, and clustering staging suggestions.

2. The automatic segmentation system for femoral head images as described in claim 1, characterized in that, The preprocessing module includes: The standardization submodule is used to construct a three-dimensional coordinate system and perform orientation standardization processing based on the three-dimensional coordinate system and multimodal data to obtain a standardized image; The resampling submodule is used to perform voxel resampling processing based on the standardized image and a preset bilinear interpolation algorithm to obtain a resampled image. The normalization submodule is used to obtain a normalized image based on the resampled image and a preset intensity normalization algorithm, and to use the normalized image as a historical preprocessed image.

3. The automatic segmentation system for femoral head images as described in claim 2, characterized in that, The preprocessing module further includes a background removal submodule, which is used to: obtain an image foreground mask based on the normalized image and a preset foreground mask generation algorithm, obtain a background removal image based on the image foreground mask and the normalized image, and use the background removal image as the preprocessed image.

4. The automatic segmentation system for femoral head images as described in claim 1, characterized in that, The data enhancement module includes: The data augmentation submodule is used to obtain random subvolume samples of the image based on the preprocessed image; The online data augmentation submodule is used to obtain an enhanced image by performing spatial adjustment, random cropping, and image intensity adjustment based on random sub-volume samples of the image.

5. The automatic segmentation system for femoral head images as described in claim 1, characterized in that, The automatic segmentation module includes: The data acquisition submodule is used to construct a historical multimodal dataset based on historical femoral head images and related historical image data, and to acquire corresponding real label data based on the historical multimodal dataset and the preset label dataset. The preprocessing submodule is used to perform orientation standardization, voxel resampling, and normalization processing on historical multimodal data and real label data in the historical multimodal dataset to obtain historical preprocessed images and preprocessed label data. The data augmentation submodule is used to perform online data augmentation based on the historical preprocessed images and preprocessed label data, and to obtain historical augmented images and augmented label data; The model building submodule is used to build an initial image segmentation model and obtain the prediction results of historical necrotic regions based on the initial image segmentation model and historical enhanced images; The model optimization submodule is used to obtain mixed loss data based on the historical necrotic region prediction results, enhanced label data and preset mixed loss function, and to optimize and adjust the weights of the initial image segmentation model based on the mixed loss data to obtain an optimized image segmentation model. An automatic segmentation submodule is used to obtain prediction results of necrotic regions based on the optimized image segmentation model and the enhanced image.

6. The automatic segmentation system for femoral head images as described in claim 5, characterized in that, The preprocessing submodule includes: A standardized unit is used to construct a three-dimensional coordinate system and perform orientation standardization processing based on the three-dimensional coordinate system, historical multimodal data and real label data to obtain historical standardized images and standardized label data. The resampling unit is used to perform voxel resampling processing based on the historical standardized image and a preset bilinear interpolation algorithm to obtain the historical resampled image; it is also used to obtain resampled label data based on the standardized label data and a preset nearest neighbor interpolation algorithm. The normalization unit is used to obtain a historical normalized image based on the historical resampled image and a preset intensity normalization algorithm, and to use the historical normalized image as a historical preprocessed image.

7. The automatic segmentation system for femoral head images as described in claim 6, characterized in that, The preprocessing submodule further includes a background removal unit, which is used to: obtain a historical image foreground mask based on the historical normalized image and a preset foreground mask generation algorithm, obtain a background removal historical image based on the historical image foreground mask and the historical normalized image, and use the background removal historical image as the historical preprocessed image.

8. The automatic segmentation system for femoral head images as described in claim 5, characterized in that, The data augmentation submodule includes: The data augmentation unit is used to obtain random sub-volume samples of historical images based on the historical preprocessed images, and to obtain random sub-volume samples of labels based on the preprocessed label data; The online data augmentation unit is used to obtain historical enhanced images by performing spatial adjustment, random cropping, and image intensity adjustment based on the random sub-volume samples of the historical images; it is also used to obtain enhanced label data by performing spatial adjustment based on the random sub-volume samples of the labels.

9. The automatic segmentation system for femoral head images as described in claim 1, characterized in that, The cluster analysis module includes: The feature extraction submodule is used to obtain radiomics features based on the prediction results of the necrotic region and a preset image analysis algorithm; The clustering analysis submodule is used to initialize several centroids and perform deep clustering analysis based on a preset clustering algorithm: an initial cluster is obtained based on the radiomics features, the centroids, and a preset allocation method; the current centroid is obtained based on the initial cluster; and the deep clustering analysis is repeatedly performed based on the current centroid to obtain the target cluster and the target centroid. When the target centroid is determined to meet the preset iteration requirements, a feature cluster is obtained based on the target cluster. The region synthesis submodule is used to merge clustering results and analyze cluster features based on the feature clusters to obtain clustering staging suggestions.

10. An automatic segmentation method for femoral head images, characterized in that, This is achieved using an automatic segmentation system for femoral head images as described in any one of claims 1 to 9, comprising: Constructing multimodal data based on femoral head images; Based on the multimodal data, orientation normalization, voxel resampling, and normalization are performed to obtain a preprocessed image; Online data augmentation is performed on the preprocessed image to obtain an enhanced image; An initial image segmentation model is constructed, and a historical multimodal dataset is built based on historical femoral head images and related historical image data. Real label data is obtained based on the historical multimodal dataset and a preset label dataset. Orientation standardization, voxel resampling, and normalization are performed on the historical multimodal data and real label data in the historical multimodal dataset to obtain historical preprocessed images and preprocessed label data. Online data augmentation is performed on the historical preprocessed images and preprocessed label data to obtain historical enhanced images and enhanced label data. Then, based on the initial image segmentation model and the historical enhanced images, the prediction result of the historical necrosis region is obtained. Based on the historical necrosis region prediction result, the enhanced label data, and a preset hybrid loss function, hybrid loss data is obtained. The initial image segmentation model is then optimized and adjusted based on the hybrid loss data to obtain an optimized image segmentation model. Finally, based on the optimized image segmentation model and the enhanced images, the prediction result of the necrosis region is obtained. Based on the predicted necrotic region and the preset learning algorithm, feature extraction is performed to obtain radiomics features. Based on the radiomics features, deep clustering analysis is performed to obtain feature clusters. Based on the feature clusters, clustering analysis suggestions are obtained. Based on the predicted necrotic area, radiomics features, and clustering staging suggestions, multidimensional necrotic image analysis results are output.

Images