A method, apparatus and device for CT bone microstructure and bone quality analysis

Abstract

The application discloses a CT bone microstructure and bone quality analysis method, device and equipment, and relates to the technical field of medical image analysis. The method comprises the following steps: acquiring ordinary CT images and high-resolution CT images of a same target bone region of a same subject, performing pretreatment, spatial registration and target bone region segmentation, and constructing paired three-dimensional image blocks; training is performed by using a super-resolution reconstruction network fused with bone microstructure parameter constraints, and super-resolution reconstruction is performed on to-be-analyzed ordinary CT images to obtain predicted bone microstructure three-dimensional images; further, bone density, bone microstructure parameters are calculated based on the predicted bone microstructure three-dimensional images, bone mechanical performance analysis is performed, and bone quality analysis results are output.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, and more specifically to a method, apparatus, and device for analyzing bone microstructure and bone quality using CT scans. Background Technology

[0002] Computed tomography (CT) has been widely used in the diagnosis of bone diseases and the assessment of bone quality. Clinically common spiral CT scans are characterized by their fast scanning speed, wide coverage, and abundant existing image data, making them an important source of imaging data for bone quality analysis. However, due to factors such as slice thickness, detector size, reconstruction algorithms, and radiation dose, the spatial resolution of conventional spiral CT is usually limited. The resulting three-dimensional images cannot fully represent the subtle morphology of bone microstructures such as trabeculae, thus limiting the ability to perform detailed quantitative analysis of bone microstructures based on conventional CT images.

[0003] To obtain clearer images of bone microstructure, existing technologies include high-resolution CT, photon-counting CT, phased-array CT, MicroCT, or HR-pQCT. However, these devices have limitations in terms of cost, scanning range, radiation dose, and clinical accessibility, making it difficult to replace conventional spiral CT as a common source of images for clinical bone quality analysis. Therefore, how to utilize existing conventional spiral CT images to obtain high-resolution imaging information suitable for quantitative analysis of bone microstructure without altering existing examination procedures has become a problem that needs to be addressed when conducting bone quality analysis using conventional CT.

[0004] In recent years, artificial intelligence and deep learning technologies have been increasingly applied to CT image segmentation, bone mass identification, and image super-resolution reconstruction. For example, models based on convolutional neural networks, 3D convolutional networks, attention mechanisms, generative networks, and Transformers can extract bone structure features from low-resolution images and generate high-resolution images; furthermore, image features and bone microstructure parameters can be combined to analyze bone quality.

[0005] However, existing technologies still have the following problems: On the one hand, some bone analysis methods directly calculate bone density or bone microstructure parameters based on ordinary CT images. The analysis results are limited by the original CT spatial resolution, making it difficult to obtain structural information corresponding to high-resolution bone microstructure. On the other hand, existing CT segmentation, super-resolution reconstruction, and bone density calculation functions are usually independent of each other, lacking a continuous processing relationship from ordinary CT image input, target bone region extraction, super-resolution bone microstructure reconstruction to quantitative analysis of bone microstructure and bone quality.

[0006] Furthermore, when using generative artificial intelligence models for medical image super-resolution reconstruction, if optimization is primarily based on pixel or visual features, the model may generate virtual high-frequency textures that are inconsistent with the actual anatomical structures. For applications requiring further calculation of microstructural parameters such as trabecular volume fraction, trabecular thickness, trabecular spacing, and trabecular number, simply improving the visual clarity of the image does not guarantee that the bone microstructure in the reconstructed image maintains a correspondence with the actual high-resolution image. This application further points out that existing related solutions have not fully utilized paired data of ordinary CT and high-resolution CT scans of the same subject and the same bone region to establish a supervised training relationship, nor have they directly used trabecular microstructural parameters as training constraints for the super-resolution model.

[0007] Therefore, there is an urgent need for a method for analyzing bone microstructure and bone quality based on ordinary CT images. By establishing a pairing relationship and spatial correspondence between ordinary CT images and high-resolution CT images, bone microstructure information can be introduced into the super-resolution reconstruction process. Furthermore, the reconstructed bone microstructure images can be used to analyze bone density, bone microstructure, and bone mechanics-related parameters, thus forming a complete processing flow from ordinary CT image input to bone quality analysis result output. Summary of the Invention

[0008] The purpose of this invention is to provide a method, apparatus, and device for CT bone microstructure and bone quality analysis. By establishing a pairing and spatial correspondence between ordinary CT images and high-resolution CT images of the same target bone region from the same subject, and by incorporating bone microstructure parameters into the training constraints of the super-resolution reconstruction model, ordinary CT images can generate high-resolution three-dimensional images with corresponding bone microstructure, reducing the influence of virtual textures generated by reconstruction relying solely on visual features. Furthermore, bone density and bone microstructure parameters are simultaneously acquired from the reconstructed images, generating bone quality analysis results. This establishes a continuous data processing chain for ordinary CT image processing, bone microstructure reconstruction, and quantitative bone quality analysis, providing a structured analytical basis for bone quality evaluation without the need for additional high-resolution CT images.

[0009] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for CT bone microstructure and bone quality analysis, the method comprising: We acquired conventional CT images and high-resolution CT images of the same target bone region from the same subject, and obtained bone density data and bone microstructure parameters corresponding to the target bone region to construct paired training data between conventional CT images and high-resolution CT images. Gray-scale normalization and spatial registration are performed on ordinary CT images and high-resolution CT images. The corresponding target bone regions and regions of interest in cancellous bone are extracted from the registered images to form ordinary CT three-dimensional image blocks and high-resolution CT three-dimensional image blocks with spatial correspondence. Ordinary CT 3D image blocks are input into the super-resolution reconstruction network to generate predicted bone microstructure 3D images. Based on the pixel differences, structural differences, high-frequency structural differences, and bone microstructure parameter differences between the predicted bone microstructure 3D images and the corresponding high-resolution CT 3D images, a composite loss is constructed to supervise the training of the super-resolution reconstruction network. This enables the super-resolution reconstruction network to simultaneously establish the mapping relationship between ordinary CT images and high-resolution bone microstructure images, as well as the correspondence between ordinary CT images and bone microstructure parameters. The trained super-resolution reconstruction network is used to perform super-resolution reconstruction on the ordinary CT images to be analyzed, and obtain three-dimensional images of predicted bone microstructure. Based on the masks corresponding to the target bone region and the region of interest in cancellous bone, bone density parameters and bone microstructure parameters are extracted from the three-dimensional images of predicted bone microstructure. The bone quality analysis results are determined based on bone mineral density parameters and bone microstructure parameters, and the predicted three-dimensional images of bone microstructure, bone mineral density parameters, bone microstructure parameters, and bone quality analysis results are output.

[0010] Secondly, the present invention also provides a CT bone microstructure and bone quality analysis device, the device comprising: The image acquisition module is used to acquire ordinary CT images and high-resolution CT images of the same target bone region of the same subject, and to acquire bone density data and bone microstructure parameters corresponding to the target bone region, and to construct paired training data between ordinary CT images and high-resolution CT images. The image registration module is used to perform grayscale normalization and spatial registration on ordinary CT images and high-resolution CT images, and extract the corresponding target bone region and cancellous bone region of interest from the registered images to form ordinary CT three-dimensional image blocks and high-resolution CT three-dimensional image blocks with spatial correspondence. The image generation module is used to input ordinary CT 3D image blocks into the super-resolution reconstruction network to generate predicted bone microstructure 3D images. Based on the pixel differences, structural differences, high-frequency structural differences, and bone microstructure parameter differences between the predicted bone microstructure 3D images and the corresponding high-resolution CT 3D images, a composite loss is constructed to supervise the training of the super-resolution reconstruction network. This enables the super-resolution reconstruction network to simultaneously establish the mapping relationship between ordinary CT images and high-resolution bone microstructure images, as well as the correspondence between ordinary CT images and bone microstructure parameters. The parameter acquisition module is used to perform super-resolution reconstruction on the ordinary CT images to be analyzed using the trained super-resolution reconstruction network, obtain the predicted three-dimensional images of bone microstructure, and extract bone density parameters and bone microstructure parameters from the predicted three-dimensional images of bone microstructure based on the masks corresponding to the target bone region and the region of interest in cancellous bone. The results determination module is used to determine the bone quality analysis results based on bone mineral density parameters and bone microstructure parameters, and outputs the predicted three-dimensional image of bone microstructure, bone mineral density parameters, bone microstructure parameters, and bone quality analysis results.

[0011] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the CT bone microstructure and bone quality analysis method provided in the first aspect.

[0012] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the CT bone microstructure and bone quality analysis method provided in the first aspect.

[0013] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the CT bone microstructure and bone quality analysis method provided in the first aspect.

[0014] The beneficial effects of this invention are as follows: The method of this invention establishes a pairing and spatial correspondence between ordinary CT images and high-resolution CT images of the same target bone region of the same subject, and introduces bone microstructure parameters into the training constraints of the super-resolution reconstruction model. This enables ordinary CT images to generate high-resolution three-dimensional images with bone microstructure correspondence, reducing the influence of virtual textures generated by reconstruction relying solely on visual features. Furthermore, it utilizes the reconstructed images to simultaneously acquire bone density and bone microstructure parameters, forming bone quality analysis results. This establishes a continuous data processing chain for ordinary CT image processing, bone microstructure reconstruction, and quantitative bone quality analysis, providing a structured analytical basis for bone quality evaluation without the need for additional high-resolution CT images.

[0015] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart illustrating a CT bone microstructure and bone quality analysis method according to an embodiment of the present invention; Figure 2This is a schematic diagram of a CT bone microstructure and bone quality analysis device according to an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0017] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] It should be noted that references to "an embodiment," "embodiment," "example embodiment," etc., in this specification refer to the described embodiment including specific features, structures, or characteristics; however, not every embodiment must include these specific features, structures, or characteristics. Furthermore, such expressions do not refer to the same embodiment. Moreover, when describing specific features, structures, or characteristics in conjunction with embodiments, whether or not explicitly described, it is indicated that incorporating such features, structures, or characteristics into other embodiments is within the knowledge of those skilled in the art.

[0019] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0020] Example 1 like Figure 1 As shown, this embodiment provides a method for CT bone microstructure and bone quality analysis. This embodiment uses the lumbar spine as the target bone region, conventional spiral CT as the conventional CT image, and high-resolution phased array CT as the high-resolution CT image. Paired image data are established for the same vertebral body region of the same subject, and a super-resolution reconstruction model is obtained by training with the paired images. The trained super-resolution reconstruction model is then used to process the conventional CT images to be analyzed, obtaining a three-dimensional image of the predicted bone microstructure, and further performing bone mineral density, bone microstructure, and bone mechanical property analysis.

[0021] In this embodiment, ordinary CT images refer to three-dimensional CT images obtained from routine clinical CT examinations, which have a lower spatial resolution compared to high-resolution CT. High-resolution CT images refer to CT images that can provide reference information on bone microstructure. In different implementations, high-resolution CT images can be obtained using photon-counting CT, phased-array CT, MicroCT, HR-pQCT, or other CT methods that can provide reference information on bone microstructure, and are not limited to the phased-array CT used in this embodiment. Bone microstructure refers to the three-dimensional structural information of trabeculae formed within cancellous bone, including information such as the volume distribution, thickness, spacing, number, connectivity, and spatial orientation of trabeculae. Bone quality is a data set that characterizes the state of the target bone based on information such as bone density, bone microstructure, and further obtained bone mechanical properties. The data connection relationships between the above processing modules are as follows: Figure 1 As shown, the model building-related data is processed sequentially by modules such as paired training data management, spatial registration and segmentation, bone microstructure constraints and loss calculation, and model training and parameter management to form a model parameter set; the bone quality analysis-related data is processed sequentially by image data interface, image preprocessing and standardization, target bone and sub-region segmentation, super-resolution reconstruction, bone density and bone microstructure quantification, and mechanical performance and risk index processing, and is finally summarized by the quality control and result output module.

[0022] S101. Obtain ordinary CT images and high-resolution CT images of the same target bone region of the same subject, and obtain bone density data and bone microstructure parameters corresponding to the target bone region to construct paired training data between ordinary CT images and high-resolution CT images.

[0023] Conventional CT images refer to CT images obtained using routine clinical CT scanning methods for routine clinical examinations; in this embodiment, conventional spiral CT images are used. High-resolution CT images refer to CT images that can provide relatively complete reference information on bone microstructures; in this embodiment, high-resolution phased array CT images are used. In other embodiments, high-resolution CT images can also be obtained using photon counting CT, MicroCT, HR-pQCT, or other CT methods that can provide reference information on bone microstructures, and conventional CT images are not limited to a specific spiral CT device.

[0024] The term "same subject, same target skeletal region" refers to bone tissue in both conventional and high-resolution CT images originating from the same subject, with both types of images corresponding to the same anatomical region. For example, in this embodiment, conventional spiral CT images and high-resolution phased-array CT images are simultaneously acquired for the same lumbar vertebrae of the same subject, ensuring that the target vertebrae in the conventional spiral CT and the target vertebrae in the high-resolution phased-array CT have corresponding anatomical objects. Establishing paired data in this way allows subsequent model training to use corresponding high-resolution CT 3D image blocks as references for conventional CT 3D image blocks, rather than training with images that do not have a strict anatomical correspondence between different subjects.

[0025] In this embodiment, 3D images of the lumbar spine of 120 clinical subjects were collected, of which 80 were used as training data, 20 as validation data, and 20 as independent test data. Low-resolution conventional spiral CT and high-resolution phased array CT images of the same vertebral body segment were acquired from the same subject. Conventional spiral CT was performed using a standard clinical 64-slice spiral CT scanner with a tube voltage of 120kV, automatic tube current, a reconstruction slice thickness of 0.625mm, and an in-plane voxel resolution of 0.65mm×0.65mm. High-resolution phased array CT was used to perform high-resolution imaging of the same target vertebral segment, with a reconstruction slice thickness of 0.165mm.

[0026] In another data construction approach, phased array CT and spiral CT data from 100 subjects can be collected, with 80 cases used for training and 20 cases for validation. The acquisition time interval between the two types of images should be controlled within 3 months to reduce the impact of bone structure changes at different time points on the image pairing relationship. The above different data scales are used to illustrate the construction methods of paired training data. In practical applications, training data, validation data, and test data can be divided according to the amount of available clinical imaging data.

[0027] In addition to conventional CT images and high-resolution CT images, bone mineral density (BMD) data and bone microstructure parameters corresponding to the target skeletal region are acquired simultaneously. BMD data describes the bone mass status of the target skeletal region and can be obtained using BMD data acquired via DXA or BMD-related data based on the HU value of CT images. Bone microstructure parameters describe the three-dimensional tissue structure of trabeculae within cancellous bone, including bone volume fraction (BV / TV), trabecular thickness (Tb.Th), trabecular separation (Tb.Sp), and trabecular number (Tb.N). When further description of trabecular network connectivity and spatial orientation is required, connectivity density (Conn.D), spatial anisotropy (DA), and other parameters characterizing bone microstructure can also be included.

[0028] Based on the aforementioned data, paired training datasets are created, ensuring that each training dataset includes at least two CT images of the same target bone region: a standard CT image, a high-resolution CT image, and corresponding bone density data and bone microstructure parameters. These paired training data are then processed through image preprocessing, spatial registration, and image patch construction. This ensures a clear data correspondence between the subsequently generated standard CT 3D image patches, high-resolution CT 3D image patches, and bone microstructure parameters. Establishing paired data for the same target bone region from the same subject provides high-resolution reference data with anatomical correspondence for subsequent supervised training and allows bone microstructure parameters to serve as quantitative constraints for subsequent model training.

[0029] S102 performs grayscale normalization and spatial registration on ordinary CT images and high-resolution CT images, and extracts the corresponding target bone region and cancellous bone region of interest from the registered images to form ordinary CT three-dimensional image blocks and high-resolution CT three-dimensional image blocks with spatial correspondence.

[0030] Gray-scale normalization refers to converting CT images obtained from different scanning devices and under different scanning conditions into a unified CT value expression space, enabling the comparison of similar tissues in ordinary CT images and high-resolution CT images using a unified gray-scale measurement. This embodiment first converts the original CT images into standard HU values ​​and truncates abnormal gray levels, while suppressing noise and artifacts in the images. Then, a three-dimensional interpolation method is used to resample the ordinary CT images and high-resolution CT images to a unified isotropic spatial resolution, and the coordinate systems of the two images are aligned. After completing the above processing, a three-dimensional region of interest (ROI) image block of the target vertebra is cropped according to the target vertebral body location.

[0031] After grayscale normalization, spatial registration is performed on the standard CT images and the high-resolution CT images. Spatial registration refers to establishing a correspondence between tissues belonging to the same target bone region in two images in spatial coordinates through spatial transformation. In this embodiment, a three-dimensional rigid body transformation based on mutual information is first used to coarsely register the standard CT images to the coordinate system of the high-resolution CT images to correct for overall positional and directional differences between the two images caused by subject posture, scanning position, etc. For cases requiring further correction of scale differences, affine registration can also be used.

[0032] After completing the overall registration, local structural registration is further performed within the vertebral cancellous bone region. Local structural registration refers to performing local spatial correspondence on the fine structures such as trabeculae within the cancellous bone, rather than performing large-scale deformation on the entire vertebral body. Specifically, a sliding window is used to extract ordinary CT image blocks and high-resolution CT image blocks within the cancellous bone region. The cross-correlation coefficients of the two image blocks at the gradient edges of the trabeculae are calculated, and a 3D mesh sampler is used to perform inverse rigid body transformation displacement compensation on the ordinary CT image blocks, thereby correcting the microscopic anatomical misalignment between the ordinary and high-resolution CT image blocks.

[0033] This embodiment allows for a certain degree of sub-millimeter spatial residuals between conventional CT images and high-resolution CT images, rather than forcibly eliminating all residuals through excessive non-rigid deformation. Sub-millimeter spatial residuals refer to the minute spatial differences that still exist after global and local registration. During subsequent model training, these residuals are tolerated and eliminated through adversarial constraints and bone microstructure-related losses to avoid distortion of the true trabecular structure caused by excessive non-rigid deformation.

[0034] To enable the spatially registered images to be further used for bone microstructure analysis, it is also necessary to extract the target bone region and the region of interest (ROI) of cancellous bone from the registered images. The target bone region refers to the complete bone structure targeted in this analysis; in this embodiment, it is specifically the target vertebral body. The ROI of cancellous bone refers to the cancellous bone region located inside the vertebral body, used for bone density and bone microstructure analysis. The target bone region and the ROI of cancellous bone are obtained using a three-dimensional deep learning segmentation network.

[0035] This embodiment constructs a target skeleton segmentation network based on 3D U-Net to automatically segment ordinary CT 3D images, obtaining the overall mask of the target vertebral body and candidate trabecular bone region masks. An attention module is introduced during the network downsampling process to enhance the network's ability to model the importance of different spatial locations and feature responses in the downsampled feature map. Specifically, for the 3D downsampled feature tensor output by the encoder, the spatial feature elements are statistically analyzed according to feature channels, the attention score of each feature element is calculated, and attention weights are generated based on the scores. The original feature elements are then weighted and fused, subsequently restored to their original feature size, and further upsampling processing is performed to obtain the final target skeleton segmentation result.

[0036] The attention score for each element in this vector is calculated using the following formula: ; In the formula, e i Let x represent the attention score for the i-th feature element. i σ represents the numerical value of the current feature element. 2μ and μ represent the variance and mean of the current feature element, respectively. To prevent the denominator from being too small, a stability constant is used, where k represents a proportionality coefficient. In this embodiment, k = 5. Currently, it can be confirmed that this module needs to calculate the element importance based on the variance, mean, and preset constant of the element values, then multiply the importance by the original feature element, and restore it to the original feature size.

[0037] After obtaining the overall mask of the target vertebral body, the vertebral body region is further segmented into sub-regions and abnormal regions are excluded. In this embodiment, a 3×3×3 cross-shaped structural element is used to perform morphological opening operations on the vertebral body mask, and connected component analysis is combined to identify degenerative osteophytes at the edge of the vertebral body. The identified osteophyte regions are removed from the effective bone tissue region; further, masks of the superior endplate, inferior endplate, cortical lateral plate, and cancellous bone region of interest are obtained.

[0038] Simultaneously, anomaly masks are created to exclude interference. These anomaly regions include one or more of the following: osteophytes, fracture lines, sclerotic foci, tumor destruction areas, metal artifacts, and non-bone areas. The resulting overall target vertebral body mask, cancellous bone candidate region mask, and anomaly region mask are not discarded after the segmentation step but are used as spatial constraint data for subsequent super-resolution reconstruction and bone quality quantification.

[0039] After extracting the target bone and cancellous bone regions, spatially corresponding 3D image blocks are extracted from ordinary CT images and high-resolution CT images based on the spatial registration results and corresponding masks. In this embodiment, a 64×64×64 voxel ordinary CT 3D image block is cut out within the cancellous bone ROI mask using a sliding window, and a 384×384×384 voxel high-resolution CT 3D image block corresponding to the geometric center is extracted. After local displacement compensation is completed through trabecular gradient edge cross-correlation measurement and a 3D mesh sampler, a set of ordinary CT 3D image blocks and high-resolution CT 3D image blocks with spatial correspondence is formed.

[0040] Therefore, the second stage ultimately generates ordinary CT 3D image blocks, high-resolution CT 3D image blocks, a target vertebral body mask, a cancellous bone region of interest mask, and anomaly region masks. Ordinary CT 3D image blocks serve as input to the subsequent super-resolution reconstruction network, high-resolution CT 3D image blocks serve as training labels, the cancellous bone region of interest mask is used to define the training area related to bone microstructure, and the anomaly region mask is used to avoid interference from abnormal bone tissue and artifact areas in subsequent quantitative analysis. By performing overall registration first and then local misalignment correction in the cancellous bone region, the spatial correspondence between ordinary and high-resolution CT 3D image blocks can be improved, while excessive deformation that could artificially alter the trabecular bone structure can be avoided.

[0041] S103, ordinary CT 3D image blocks are input into the super-resolution reconstruction network to generate predicted bone microstructure 3D images. Based on the pixel differences, structural differences, high-frequency structural differences, and bone microstructure parameter differences between the predicted bone microstructure 3D images and the corresponding high-resolution CT 3D images, a composite loss is constructed to supervise the training of the super-resolution reconstruction network. This enables the super-resolution reconstruction network to simultaneously establish the mapping relationship between ordinary CT images and high-resolution bone microstructure images, as well as the correspondence between ordinary CT images and bone microstructure parameters.

[0042] This embodiment employs a super-resolution reconstruction network consisting of a generator and a discriminator. The generator receives ordinary CT 3D image blocks and generates a predicted 3D image of bone microstructure. The discriminator determines the local structural differences between the generated predicted 3D image of bone microstructure and the real high-resolution CT 3D image. In a specific implementation, a 3D generative adversarial network based on residual dense blocks (RRDB) and spatial attention mechanism can be used, with a 3D sub-pixel convolutional layer (3DPixel Shuffle) added at the end of the network to achieve 6x spatial upsampling reconstruction.

[0043] As other implementations, the generator of the super-resolution reconstruction network can also employ 3D U-Net, 3DResUNet, 3D SRGAN, 3D ESRGAN, 3D CycleGAN, a Transformer-based 3D super-resolution network, a conditional diffusion model, or a hybrid network composed of 3D convolution, residual connections, skip connections, and attention mechanisms. Although the structures of these networks differ, they all receive the aforementioned ordinary CT 3D image blocks and output predicted 3D images of bone microstructure that can be subsequently analyzed for bone density and bone microstructure parameters.

[0044] During training, the ordinary CT 3D image blocks generated in the second stage are used as input to the generator, while high-resolution CT 3D image blocks with spatial correspondence are used as supervision labels. After processing by the generator, the ordinary CT 3D image blocks yield predicted bone microstructure 3D image blocks. These predicted bone microstructure 3D image blocks are then used together with their corresponding high-resolution CT 3D image blocks in the composite loss calculation process. Therefore, the predicted bone microstructure 3D images generated by the generator are not merely for display purposes, but serve as a direct data source for subsequent loss calculations and model parameter updates. Furthermore, after model training is complete, they will also serve as input data for the actual bone quality analysis stage.

[0045] To simultaneously constrain the image quality and quantitative consistency of predicted 3D bone microstructure images, a composite loss function is established, including pixel-level reconstruction loss, structural similarity loss, 3D gradient loss, adversarial loss, bone microstructure consistency loss, and bone density consistency loss. The pixel-level reconstruction loss evaluates the grayscale difference between the predicted 3D bone microstructure image and the real high-resolution CT 3D image at corresponding voxel locations, and can be calculated using mean absolute error (MAE) or mean squared error (MSE). The structural similarity loss evaluates the consistency between the predicted image and the real high-resolution CT image in terms of local structure, contrast, and texture, and can be calculated based on SSIM. The 3D gradient loss constrains high-frequency structures such as bone cortical boundaries and trabecular bone edges. The adversarial loss, formed jointly by the generator and discriminator, constrains the local structural realism of the predicted image.

[0046] Furthermore, bone microstructure parameters are directly used as quantization constraints during the training process of the super-resolution network. The bone microstructure consistency loss is determined based on the difference between the bone microstructure parameters calculated from the predicted 3D bone microstructure image and the corresponding high-resolution CT 3D image. Bone microstructure parameters include one or more of BV / TV, Tb.Th, Tb.Sp, Tb.N, Conn.D, and DA. BV / TV characterizes the proportion of bone tissue volume in the analysis region, Tb.Th characterizes trabecular thickness, Tb.Sp characterizes the degree of trabecular separation, Tb.N characterizes the number of trabeculae, Conn.D characterizes the connectivity of the trabecular network, and DA characterizes the anisotropy of the trabecular spatial structure.

[0047] Bone mineral density consistency loss is determined based on the differences in average HU value, HU distribution, or volumetric bone mineral density (vBMD) between the predicted 3D bone microstructure image and the corresponding high-resolution CT 3D image. Therefore, during super-resolution network training, not only are the voxel values ​​of the predicted and ground truth images compared, but also their structures, high-frequency edges, bone microstructure parameters, and bone mineral density-related parameters, thus establishing a direct correspondence between the model output and subsequent quantitative analysis.

[0048] The composite loss function is defined as: ; Where α, β, γ, δ, ε, and η represent the weighting coefficients of each loss term. Represents pixel-level reconstruction loss, used to evaluate the difference in corresponding voxel gray values ​​between AI-generated high-resolution CT images and real high-resolution CT images. It is calculated using mean absolute error (MAE) or mean square error (MSE). The structural similarity loss is used to evaluate the consistency between the generated image and the real high-resolution CT image in terms of local structure, contrast, and texture. It is calculated based on the structural similarity index SSIM. The three-dimensional gradient loss is used to constrain the generated image to preserve high-frequency structural information such as the bone cortex boundary and trabecular bone edge. It is calculated by comparing the spatial gradient difference between the real image and the generated image to reduce edge blurring after super-resolution. The adversarial loss is derived from the generative adversarial network structure, in which the generator generates high-resolution CT images, and the discriminator determines whether the input images come from real high-resolution CT data. Through adversarial optimization between the generator and the discriminator, the generated images are made closer to the distribution of real images. The loss of bone microstructure consistency is indicated by comparing the differences in bone microstructure parameters (including but not limited to BV / TV, Tb.Th, Tb.Sp, Tb.N, Conn.D, DA) calculated by comparing the actual high-resolution CT with the generated high-resolution CT. The loss of bone mineral density consistency can be calculated based on the average HU value, HU distribution, or volumetric bone mineral density (vBMD). Pixel-level reconstruction loss, structural similarity loss, 3D gradient loss, adversarial loss, bone microstructure consistency loss, and bone mineral density consistency loss each bear their respective constraint relationships.

[0049] In one specific implementation, the weight coefficients of each loss term are set as follows: α=1.0, β=0.3, γ=0.1, δ=0.05, ε=1.0, η=0.5. These weight coefficients are used to adjust the contribution of different loss terms during model training. and Using the same weights ensures both the grayscale accuracy of the generated high-resolution CT images and the consistency of bone microstructure parameters. The weights are approximately 50% of the pixel-level reconstruction loss, used to further constrain the HU value distribution and volumetric bone density information of the generated image. and As a constraint for maintaining the image structure, it is used to enhance the ability of bone boundaries to recover. Setting the minimum weighting avoids the generation of unrealistic bone structures due to excessively strong adversarial constraints. In other embodiments, the weighting coefficients can be adjusted based on the sample size, CT scan parameters, model structure, and validation set evaluation results.

[0050] During training, data augmentation, such as rotation, scaling, and noise simulation, is performed synchronously on paired ordinary CT 3D image blocks and high-resolution CT 3D image blocks. Synchronous execution means that when an ordinary CT 3D image block undergoes a spatial transformation, its corresponding high-resolution CT 3D image block undergoes a corresponding spatial transformation to avoid losing the voxel correspondence between the two image blocks after data augmentation. After data augmentation, the ordinary CT 3D image blocks are input into the super-resolution generation model to obtain predicted bone microstructure 3D image blocks. Then, the predicted bone microstructure 3D image blocks and the high-resolution CT 3D image block labels are compared to calculate losses at the pixel, structure, gradient, bone microstructure, and bone density levels.

[0051] The calculated composite loss is used to perform backpropagation of the error, obtaining the gradient flow of each layer of the super-resolution reconstruction network. The gradient optimizer is then used to iteratively update the network parameters of the generator and discriminator. In this embodiment, AdamW is used as the gradient optimizer with an initial learning rate of 2×10⁻⁶. -4 The batch size was set to 4, and the training was conducted for a total of 200 epochs. The first 50 epochs mainly focused on enhancing the reconstruction of the macroscopic morphology, while the last 150 epochs focused on enhancing the reconstruction of the topological details of the trabecular bone.

[0052] During training, the image reconstruction quality and consistency of bone microstructure indices of the predicted 3D bone microstructure images are evaluated on validation data. Iteration stops when the composite loss gradually stabilizes and meets the preset convergence criteria, and the generator's model structure, weight parameters, calibration parameters, and version information are solidified to form a model parameter set. This model parameter set is then written to the model storage module as the super-resolution reconstruction model to be used in subsequent processing of ordinary CT images.

[0053] Through the above training process, the ordinary CT 3D image blocks and high-resolution CT 3D image blocks formed in the second stage become the supervised training data for the third stage. The model parameter set generated in the third stage becomes the model input for the next stage of super-resolution reconstruction of the ordinary CT images to be analyzed, thus completing the data transfer from paired data to model parameter set.

[0054] S104 utilizes the trained super-resolution reconstruction network to perform super-resolution reconstruction on the ordinary CT images to be analyzed, obtaining a three-dimensional image of the predicted bone microstructure. Based on the mask corresponding to the target bone region and the region of interest in cancellous bone, bone density parameters and bone microstructure parameters are extracted from the three-dimensional image of the predicted bone microstructure.

[0055] The plain CT images to be analyzed can be acquired from the CT scanner or PACS via an image data interface. The image data interface can receive DICOM image sequences and transmit the received plain CT images to the image preprocessing module. The plain CT images to be analyzed undergo HU value normalization, abnormal grayscale processing, noise and artifact suppression, and spatial resampling, consistent with the model training phase, ensuring that the plain CT images to be analyzed are in the same input feature space as the plain CT 3D images used during model training.

[0056] Subsequently, the previously trained and fixed model parameter set is invoked, and the preprocessed ordinary CT images to be analyzed are input into the super-resolution reconstruction network to obtain three-dimensional images of predicted bone microstructures. In this embodiment, the super-resolution reconstruction result can achieve an isotropic resolution of 0.165 mm, thereby obtaining predicted images that can be used for subsequent three-dimensional bone microstructure analysis.

[0057] After obtaining the 3D image of the predicted bone microstructure, the target vertebral body mask, the cancellous bone region of interest mask, and the abnormal region mask generated in the second stage are invoked to determine the effective analysis area in the 3D image of the predicted bone microstructure. In other words, the mask generated in the second stage is not only used for training data construction, but also continues to serve as spatial constraint input in the actual inference process, and is entered into the bone density and bone microstructure quantification module together with the 3D image of the predicted bone microstructure.

[0058] First, an abnormal region mask is used to exclude one or more interfering regions, including cortical bone, venous plexus, osteophytes, sclerotic lesions, fracture areas, tumor destruction areas, and metal artifact areas. Within the excluded region of interest in cancellous bone, the CT values ​​of all effective voxels are extracted, and the average HU value is calculated. The average HU value is the statistical average of the HU values ​​of effective cancellous bone voxels after excluding abnormal regions. Then, based on a pre-established HU value-bone mineral density mapping relationship, the average HU value is mapped to volumetric bone mineral density. The HU value-bone mineral density mapping relationship can be obtained through phantom calibration, QCT comparison, DXA comparison, or clinical data fitting.

[0059] When relative bone mineral density (cBMD) is needed, pre-stored standard population baseline values ​​are retrieved, and dimensionless calculations are performed based on the average CT values ​​within the region of interest in the cancellous bone to obtain the cBMD. ; in, mean To predict the average CT value of all voxels within a ROI in cancellous bone in high-resolution CT images, The average CT value of normal bone tissue in a pre-defined standard population.

[0060] After obtaining bone mineral density parameters, bone microstructure parameters are further extracted using the predicted three-dimensional images of bone microstructure. Specifically, within the previously identified region of interest in cancellous bone, three-dimensional morphological analysis is performed on the predicted three-dimensional images of bone microstructure. One or more of the following parameters are extracted based on the spatial distribution relationship between bone and non-bone tissues: bone volume fraction (BV / TV), trabecular thickness (Tb.Th), trabecular separation (Tb.Sp), trabecular number (Tb.N), connectivity density (Conn.D), and spatial anisotropy (DA).

[0061] In this stage, the predicted 3D bone microstructure image is directly generated by the super-resolution reconstruction network. Bone density and bone microstructure parameters are calculated based on this predicted 3D bone microstructure image and the aforementioned target bone and cancellous bone masks. Therefore, a direct data connection is formed between the model output, the region mask, and the quantitative parameters. By using bone microstructure parameters as constraints in the model training stage, and simultaneously using the predicted 3D bone microstructure image for actual bone microstructure parameter calculation, the constraint objectives in the training stage and the quantification objectives in the inference stage can be kept consistent.

[0062] S105 determines the bone quality analysis results based on bone mineral density parameters and bone microstructure parameters, and outputs the predicted three-dimensional image of bone microstructure, bone mineral density parameters, bone microstructure parameters, and bone quality analysis results.

[0063] In this embodiment, the bone quality analysis results include not only bone mineral density and bone microstructure parameters, but also bone mechanical parameters obtained from predicted three-dimensional images of bone microstructure, bone mineral density parameters, and bone microstructure parameters, as well as finite element analysis results. Firstly, based on the average HU value of the cancellous bone region, and through a pre-established mapping relationship between the HU value and bone mineral density and tissue mechanical properties, the cancellous bone mineral density, elastic modulus, and mechanical strength are calculated. Among these, the cancellous bone mineral density... The mapping relationship between elastic modulus E, mechanical strength criterion, and average HU value of cancellous bone region is as follows: =1.122×Hu+47; E=1.92× -170 (E≥5MPa); =0.2%; Where E represents the elastic modulus of cancellous bone. The criterion for tensile strength of cancellous bone is represented by the maximum tensile failure strain. HU represents the HU value of the cancellous bone region. The bone mineral density, elastic modulus, and mechanical strength parameters are obtained by fitting experimental data, clinical data, or stratified population data.

[0064] Furthermore, a three-dimensional finite element model can be established using the predicted three-dimensional bone microstructure images. The finite element model is used to convert the predicted three-dimensional bone microstructure images into discrete structures that can be used for mechanical calculations. In this embodiment, the reconstructed super-resolution three-dimensional bone microstructure images are combined with the upper endplate mask, lower endplate mask, and cortical lateral plate mask obtained in the second stage, and converted into a hexahedral linear solid finite element mesh, so that the finite element mesh corresponds to the actual bone structure in the predicted three-dimensional bone microstructure images.

[0065] After completing the finite element mesh construction, based on the local HU values ​​in the predicted bone microstructure 3D image and the previously extracted bone volume fraction BV / TV, a mapping relationship between image parameters and bone tissue mechanical properties is established, and each finite element is assigned a corresponding elastic modulus Ei. For the i-th finite element, the BV / TV of its corresponding region is denoted as BV / TVi, the CT value at the corresponding location is denoted as HUi, the average HU value of the target analysis region is denoted as HUmean, α represents the nonlinear relationship exponent between bone volume fraction and elastic modulus, and k represents the proportionality coefficient between image parameters and material mechanical parameters.

[0066] ; Among them, BV / TV i HU represents the bone volume fraction of the region corresponding to the finite element element; α represents the nonlinear relationship exponent between bone volume fraction and elastic modulus; i HU represents the CT value at the location corresponding to the i-th finite element element; mean α represents the average HU value within the target analysis area; k represents the proportionality coefficient between the imaging parameters and the material mechanical parameters. α and k can be obtained by fitting data from bone mechanics experiments, finite element calibration experiments, or clinical sample data. In this embodiment, α = 1.4, which is obtained by fitting mechanical experimental data. This means that as the bone volume fraction increases, the elastic modulus exhibits a superlinear increase. The rationality of this parameter setting can be supported by the reference (Wald MJ, Magland JF, Rajapakse CS, Bhagat YA, Wehrli FW. Predicting trabecular bone elastic properties from measures of bone volume fraction and fabric on the basis of micromagnetic resonanceimages. Magn Reson Med. 2012 Aug;68(2):463-73.). In this embodiment, k = 2110, which is calibrated based on existing bone mechanics experimental data.

[0067] After obtaining the material properties corresponding to the finite element elements, boundary conditions and load conditions are set for the finite element model. In this embodiment, the lower end plate node is set as a fully fixed constraint boundary, and axial compression, bending, and shear load displacements are applied to the upper end plate node respectively, so that the finite element model can simulate the mechanical response of the target cone structure under different stress conditions.

[0068] After the load is applied, the finite element model is solved mechanically to obtain the maximum equivalent stress on the trabecular micronetwork. The safety factor SF of the vertebral structure is then calculated based on the system's preset material failure threshold. The safety factor represents the safety margin of the target vertebral body from material failure under the current load conditions, and its specific calculation formula is as follows: ; in, For the maximum equivalent stress on the trabecular micronetwork, This is the material failure threshold of the system.

[0069] The calculated safety factor is compared with a preset risk threshold. When the calculated safety factor is lower than the corresponding risk threshold, the target subject can be classified as being at high risk of fracture. The risk threshold can be determined based on material experiments, clinical validation, or equipment calibration, and is not limited to a single fixed value.

[0070] Therefore, the three-dimensional images of predicted bone microstructure, bone mineral density parameters, and bone microstructure parameters obtained in the previous stage continue to serve as data inputs for the bone mechanics analysis in this stage, ultimately forming the bone quality analysis results. The bone quality analysis results may include one or more of the following: bone mineral density, bone microstructure parameters, elastic modulus, compressive strength, finite element stress distribution, structural safety factor, and fracture risk level.

[0071] After completing the bone quality analysis, quality control is performed on the input data, target bone segmentation results, predicted 3D images of bone microstructure, and quantitative analysis results. Quality control confirms whether the current analysis task is within the model's applicable range and checks the completeness of the segmentation results, whether the predicted images meet preset quality conditions, and whether there are any anomalies in the quantitative analysis results. For the trained super-resolution model, comprehensive validation can also be performed through image reconstruction quality, consistency of bone microstructure parameters, accuracy of bone density prediction, accuracy of bone quality prediction, accuracy of osteoporosis grading, and fracture risk prediction capabilities.

[0072] In practical deployment, the quality control module can verify the input scanning parameters, segmentation integrity, reconstruction confidence, and model applicability. When the input data does not meet the model's applicability conditions, the target bone segmentation is incomplete, or the prediction results do not meet the preset quality requirements, the corresponding task is marked as requiring manual review or inapplicable, instead of directly outputting abnormal results as the final bone quality analysis results. In this way, the quality control results, together with the predicted 3D bone microstructure images, bone density parameters, bone microstructure parameters, and mechanical parameters, constitute the final result data.

[0073] Finally, the predicted 3D bone microstructure images, bone mineral density parameters, bone microstructure parameters, and bone quality analysis results are output. The output results can be displayed on a monitor or transmitted back via a PACS interface, DICOM structured report, PDF report, or database interface. The display interface allows for synchronized display of the original plain CT images and the predicted 3D bone microstructure images, simultaneously showing the target vertebral body mask, cancellous bone ROI mask, and the corresponding bone mineral density, bone microstructure, and mechanical analysis parameters. When data archiving is required, the original image index, predicted bone microstructure images, masks, bone mineral density parameters, bone microstructure parameters, mechanical parameters, model version, and quality control conclusions are combined into a structured result.

[0074] Through the above processing, this embodiment ultimately forms a complete data processing chain from acquiring paired CT images to outputting bone quality analysis results. Specifically, the first stage generates ordinary CT images, high-resolution CT images, bone density data, and bone microstructure parameters to construct training data; the second stage standardizes, spatially registers, and extracts target bone and cancellous bone regions from the training data, forming ordinary CT 3D image blocks and high-resolution CT 3D image blocks; the third stage uses these ordinary CT 3D image blocks to train a super-resolution reconstruction network and forms a model parameter set; the fourth stage calls the model parameter set to process the ordinary CT images to be analyzed, forming predicted bone microstructure 3D images, and extracts bone density parameters and bone microstructure parameters based on the aforementioned target bone and cancellous bone masks; the fifth stage further utilizes the predicted bone microstructure 3D images, bone density parameters, and bone microstructure parameters for mechanical analysis and risk analysis, and outputs the predicted bone microstructure 3D images, bone density parameters, bone microstructure parameters, and bone quality analysis results uniformly. Therefore, the data generated in each stage is used in subsequent stages, and there is no data bypass that is only used for illustration and not involved in subsequent processing.

[0075] Example 2 This embodiment is used to validate the above-mentioned CT bone microstructure and bone quality analysis method. The test data, evaluation indicators, and comparison results in this embodiment are used to verify the model reconstruction capability, the consistency of bone microstructure parameters, and the bone quality analysis capability, and are not intended as a necessary processing step for each bone quality analysis.

[0076] This embodiment employs a deep learning computing platform for super-resolution model training, using two NVIDIA RTX 4090 graphics processors, each with 24GB of video memory, and the Python 3.10 deep learning framework. A dataset containing 3D images of the lumbar spine of 120 clinical subjects was established, with 80 cases used as the training set, 20 cases as the validation set, and 20 cases as the independent test set. For each subject, conventional spiral CT and high-resolution phased array CT images of the same vertebral body were acquired. The conventional spiral CT used a 64-slice spiral CT scanner with a tube voltage of 120kV, automatic tube current, a reconstruction slice thickness of 0.625mm, and an in-plane voxel resolution of 0.65mm×0.65mm. The high-resolution phased array CT images had a reconstruction slice thickness of 0.165mm.

[0077] First, spatial registration was performed on conventional CT images and high-resolution CT images as described in Example 1, and local paired training image blocks were established within the region of interest in cancellous bone. For each paired region, a 64×64×64 voxel conventional CT image block and a corresponding 384×384×384 voxel high-resolution CT image block were extracted. Local displacement was determined by cross-correlation measurement of trabecular gradient edges, and rigid body inverse transformation displacement compensation was performed using a 3D mesh sampler. This resulted in a set of local training image blocks with spatial correspondence.

[0078] A 3D generative adversarial network based on RRDB and spatial attention mechanisms was used to train the super-resolution model. At the network endpoint, 3D pixel shuffle was used to achieve a 6x spatial upsampling. Before training, the HU values ​​in the paired images were normalized, and data augmentation such as rotation, scaling, and noise simulation were performed simultaneously to maintain the spatial correspondence between the augmented ordinary CT image blocks and the high-resolution CT image blocks. During training, AdamW or other gradient optimizers were used, with an initial learning rate of 2×10⁻⁶. -4 The batch size is 4, and the training lasts for 200 epochs. The first 50 epochs are used to enhance the reconstruction of macroscopic morphology, and the last 150 epochs are used to enhance the reconstruction of trabecular topology details.

[0079] During model training, ordinary CT image patches are input into the generator to obtain predicted 3D image patches of bone microstructure. Pixel-level reconstruction loss, structural similarity loss, 3D gradient loss, adversarial loss, and bone microstructure and bone density consistency loss are calculated separately. The gradients of each network layer are obtained through error backpropagation, and the generator and discriminator parameters are updated alternately using a gradient optimizer. The image reconstruction quality and microstructure consistency are evaluated on the validation set. The optimal model is solidified when the composite loss meets the preset convergence condition.

[0080] After model training, ordinary CT images from an independent test set were input into the trained super-resolution model to obtain predicted 3D images of bone microstructures. These predicted 3D images were then compared with corresponding high-resolution CT images at the image level. Peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), mean absolute error (MAE), and root mean square error (RMSE) were used to evaluate the reconstruction quality of the predicted images. Test results showed that the average PSNR of the predicted images was 32.8 dB, the SSIM was 0.90, and the MAE and RMSE were reduced by more than 35% compared to the basic interpolation method and the standard SRGAN model.

[0081] Further, bone microstructure parameters such as BV / TV, Tb.Th, and Tb.Sp were extracted from the predicted 3D bone microstructure images and their corresponding high-resolution CT images, respectively. The two sets of parameters were then compared in pairs to verify whether the predicted 3D bone microstructure images could maintain the quantitative information of bone microstructure in the high-resolution CT images. The test results showed no statistically significant difference between the relevant bone microstructure parameters extracted from the predicted 3D bone microstructure images and the corresponding physical measurements from the high-resolution CT images. Bland-Altman consistency analysis showed that 96% of the data were within the corresponding standard deviation range.

[0082] Further analysis was conducted by combining bone mineral density and bone microstructure parameters extracted from 3D images of predicted bone microstructure with structural safety factors obtained from finite element analysis to verify the correlation between bone quality parameters at the three levels of imaging, structure, and mechanics. A multimodal fracture risk classification prediction model was constructed based on the test set results. Test results showed that the model achieved an area under the receiver operating characteristic (AUC) of 0.94, a sensitivity of 91.2%, and a specificity of 92.5%; in contrast, the traditional diagnostic baseline AUC relying solely on 2D DXA bone mineral density was 0.72.

[0083] The finite element analysis results were further validated. The predicted three-dimensional images of bone microstructure, along with masks of the upper and lower endplates and cortical lateral plates, were converted into a hexahedral linear solid finite element mesh. Material mechanical properties were assigned to the finite element elements based on the local HU value and BV / TV. Fully fixed constraints were applied to the lower endplate nodes, while axial compression, bending, and shear load displacements were applied to the upper endplate nodes. The maximum equivalent stress in the trabecular micronetwork was obtained through finite element analysis, and the structural safety factor was calculated based on a preset material failure threshold.

[0084] During model validation, quality control can also be performed on the applicability of the super-resolution model. This includes checking whether the scanning parameters of the input ordinary CT images are within the model's applicable range, whether the target bone segmentation is complete, whether the ROI of cancellous bone meets the requirements for quantitative analysis, and whether there are obvious abnormalities in the predicted bone microstructure 3D images. Test data that do not meet the preset conditions are marked as unsuitable or require manual review, and are not directly used for model performance evaluation, thereby avoiding the impact of abnormal input on the model validation results.

[0085] After model performance validation, the validated model parameter set is permanently saved, along with information such as model identifier, model version, and applicable scanning protocol. Subsequent inference devices can load the validated model parameter set and perform super-resolution reconstruction, bone density and bone microstructure parameter extraction, and further mechanical analysis using only the plain CT images to be analyzed. The model building server and inference server can be deployed separately. When the model is updated, the validated new model parameter set is transmitted to the inference device. After model integrity verification, the model storage module is updated, while existing model versions are retained for future reference to trace historical analysis results based on model version.

[0086] It should be noted that the PSNR, SSIM, MAE, RMSE, Bland-Altman consistency analysis, AUC, sensitivity, and specificity in this embodiment are all validation metrics used to evaluate the effectiveness of this method, rather than technical steps necessarily performed when conducting ordinary CT bone quality analysis. Therefore, in practical applications, appropriate metrics can be selected for validation based on model training, product validation, or clinical evaluation needs, without affecting the aforementioned processing flow consisting of ordinary CT image input, generation of predicted three-dimensional bone microstructure images, extraction of bone density and bone microstructure parameters, and output of bone quality analysis results.

[0087] Based on the same inventive concept, this application also provides a CT bone microstructure and bone quality analysis device for implementing the CT bone microstructure and bone quality analysis method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more CT bone microstructure and bone quality analysis device embodiments provided below can be found in the limitations of the CT bone microstructure and bone quality analysis method described above, and will not be repeated here.

[0088] In one embodiment, such as Figure 2 As shown, a CT bone microstructure and bone quality analysis device is provided, the device comprising: The image acquisition module 30 is used to acquire ordinary CT images and high-resolution CT images of the same target bone region of the same subject, and to acquire bone density data and bone microstructure parameters corresponding to the target bone region, and to construct paired training data between ordinary CT images and high-resolution CT images. The image registration module 31 is used to perform grayscale standardization and spatial registration on ordinary CT images and high-resolution CT images, and extract the corresponding target bone region and cancellous bone region of interest from the registered images to form ordinary CT three-dimensional image blocks and high-resolution CT three-dimensional image blocks with spatial correspondence. Image generation module 32 is used to input ordinary CT three-dimensional image blocks into the super-resolution reconstruction network to generate predicted bone microstructure three-dimensional images. Based on the pixel differences, structural differences, high-frequency structural differences, and bone microstructure parameter differences between the predicted bone microstructure three-dimensional images and the corresponding high-resolution CT three-dimensional images, a composite loss is constructed to supervise the training of the super-resolution reconstruction network, so that the super-resolution reconstruction network can simultaneously establish the mapping relationship between ordinary CT images and high-resolution bone microstructure images, as well as the correspondence between ordinary CT images and bone microstructure parameters. The parameter acquisition module 33 is used to perform super-resolution reconstruction on the ordinary CT image to be analyzed using the trained super-resolution reconstruction network, obtain a three-dimensional image of the predicted bone microstructure, and extract bone density parameters and bone microstructure parameters from the three-dimensional image of the predicted bone microstructure based on the mask corresponding to the target bone region and the region of interest of cancellous bone. The result determination module 34 is used to determine the bone quality analysis results based on bone mineral density parameters and bone microstructure parameters, and output the predicted three-dimensional image of bone microstructure, bone mineral density parameters, bone microstructure parameters, and bone quality analysis results.

[0089] This application also provides an electronic device, in some embodiments, referring to... Figure 3As shown, the electronic device 700 includes an input unit 710, a memory 720, a processor 730, and an output unit 740. The memory 720 stores program instructions that can be executed on the processor 730. The processor 730 can execute the CT bone microstructure and bone quality analysis methods and / or technical solutions based on the foregoing embodiments by calling the program instructions. The electronic device 700 can be a mobile terminal device such as a mobile phone or a computer.

[0090] Furthermore, embodiments of this application also provide a computer-readable storage medium for storing a computer program that performs CT bone microstructure and bone quality analysis methods. For example, computer program instructions, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. The program instructions that invoke the methods of this application may be stored in a fixed or removable storage medium, and / or transmitted via data streams in broadcast or other signal carrying media, and / or stored in a storage medium that operates according to the program instructions.

[0091] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0092] The technical features of the above embodiments can be arbitrarily integrated. For the sake of brevity, not all possible integrations of the technical features in the above embodiments are described. However, as long as the integration of these technical features does not contradict each other, they should be considered to be within the scope of this specification.

[0093] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for analyzing bone microstructure and bone quality using CT scans, characterized in that, The method includes: Obtain ordinary CT images and high-resolution CT images of the same target bone region of the same subject, and obtain bone density data and bone microstructure parameters corresponding to the target bone region to construct paired training data between ordinary CT images and high-resolution CT images; The ordinary CT images and high-resolution CT images are grayscale normalized and spatially registered, and the corresponding target bone regions and cancellous bone regions of interest are extracted from the registered images to form ordinary CT three-dimensional image blocks and high-resolution CT three-dimensional image blocks with spatial correspondence. The ordinary CT 3D image blocks are input into the super-resolution reconstruction network to generate a predicted bone microstructure 3D image. Based on the pixel differences, structural differences, high-frequency structural differences, and bone microstructure parameter differences between the predicted bone microstructure 3D image and the corresponding high-resolution CT 3D image, a composite loss is constructed to supervise the training of the super-resolution reconstruction network. This enables the super-resolution reconstruction network to simultaneously establish the mapping relationship between ordinary CT images and high-resolution bone microstructure images, as well as the correspondence between ordinary CT images and bone microstructure parameters. The trained super-resolution reconstruction network is used to perform super-resolution reconstruction on the ordinary CT images to be analyzed, and a three-dimensional image of the predicted bone microstructure is obtained. Based on the mask corresponding to the target bone region and the region of interest in cancellous bone, bone density parameters and bone microstructure parameters are extracted from the three-dimensional image of the predicted bone microstructure. The bone quality analysis results are determined based on the bone mineral density parameters and bone microstructure parameters, and the predicted three-dimensional image of bone microstructure, bone mineral density parameters, bone microstructure parameters, and bone quality analysis results are output.

2. The CT bone microstructure and bone quality analysis method as described in claim 1, characterized in that, The spatial registration includes first performing rigid registration or affine registration on ordinary CT images and high-resolution CT images to correct the overall position and orientation differences between the two images, and then performing non-rigid registration or local displacement compensation on the region of interest in the cancellous bone to correct local spatial misalignment at the trabecular scale. In the local displacement compensation process, the local displacement is determined based on the correlation between the ordinary CT image block and the high-resolution CT image block at the edge or gradient region of the trabecular bone, and the ordinary CT image block is inversely transformed by three-dimensional mesh sampling to obtain a paired three-dimensional image block with the corresponding local anatomical structure.

3. The CT bone microstructure and bone quality analysis method as described in claim 2, characterized in that, The extraction of the target skeletal region includes: Three-dimensional target bone segmentation is performed on ordinary CT images to obtain the overall mask of the target bone, and further, the candidate region mask of cancellous bone and the interference region mask used to exclude abnormal regions are obtained. Based on the interference region mask, one or more of osteophytes, fracture areas, sclerotic foci, tumor destruction areas, metal artifact areas, and non-bone areas are excluded from the overall target bone mask to form the target bone region of interest for super-resolution reconstruction and bone quality quantification. The three-dimensional target bone segmentation adopts a three-dimensional segmentation network based on an encoder-decoder structure, and introduces an attention mechanism for weighting the importance of different feature elements in the downsampling feature processing to obtain the overall mask of the target bone and the candidate region mask of cancellous bone.

4. The CT bone microstructure and bone quality analysis method as described in claim 1, characterized in that, The composite loss includes at least a variety of the following: pixel-level reconstruction loss, structural similarity loss, three-dimensional gradient loss, adversarial loss, and bone microstructure consistency loss. The bone microstructure consistency loss is determined based on the difference between one or more of the following: bone volume fraction, trabecular thickness, trabecular separation, trabecular number, connectivity density, and spatial anisotropy, calculated from the predicted three-dimensional bone microstructure image and the high-resolution CT image, respectively. The composite loss also includes bone mineral density consistency loss, which is determined based on the difference between the average HU value, HU distribution, or volumetric bone mineral density of the corresponding region in the predicted three-dimensional bone microstructure image and the high-resolution CT image, and the composite loss is obtained by weighting each loss term according to the weight of the corresponding loss term.

5. The CT bone microstructure and bone quality analysis method as described in claim 1, characterized in that, The extraction of bone mineral density parameters and bone microstructure parameters includes: Based on the mask corresponding to the region of interest in the cancellous bone, exclude one or more of the following: cortical bone, venous plexus, osteophytes, sclerotic foci, fracture areas, tumor destruction areas, and metal artifact areas, and calculate the average HU value of the cancellous bone region after excluding interfering areas. The volumetric bone mineral density or relative bone mineral density parameter is determined based on the average HU value and the pre-established HU value-bone mineral density mapping relationship, which is established through phantom calibration, quantitative CT comparison, dual-energy X-ray absorptiometry comparison, or clinical data fitting. Three-dimensional morphological analysis is performed on the predicted bone microstructure three-dimensional image to obtain one or more of the following: bone volume fraction, trabecular thickness, trabecular separation, trabecular number, connectivity density, and spatial anisotropy.

6. The CT bone microstructure and bone quality analysis method as described in claim 5, characterized in that, The bone biomechanical analysis results are determined based on the predicted three-dimensional bone microstructure image, the bone density parameters, and the bone microstructure parameters, including: The bone density, elastic modulus and mechanical strength of the cancellous bone are determined based on the HU value of the cancellous bone region, or the predicted three-dimensional image of bone microstructure is converted into a three-dimensional finite element solid mesh, and the corresponding material mechanical properties are assigned to the finite element elements based on the local HU value and bone volume fraction. One or more of axial compressive load, bending load, and shear load are applied to the finite element solid mesh, and the stress distribution and structural safety factor of the target skeleton structure are calculated. The fracture risk level or bone quality auxiliary assessment result is determined based on one or more of the structural safety factor, bone mineral density parameter, bone microstructure parameter and bone biomechanical parameter. The system performs quality control on the input images, target bone segmentation results, super-resolution reconstruction results, bone density parameters, bone microstructure parameters, and bone biomechanics analysis results, and outputs the quality control conclusions along with the corresponding analysis results.

7. A CT bone microstructure and bone quality analysis device, characterized in that, The device includes: The image acquisition module is used to acquire ordinary CT images and high-resolution CT images of the same target bone region of the same subject, and to acquire bone density data and bone microstructure parameters corresponding to the target bone region, and to construct paired training data between ordinary CT images and high-resolution CT images. The image registration module is used to perform grayscale normalization and spatial registration on the ordinary CT images and high-resolution CT images, and extract the corresponding target bone region and cancellous bone region of interest from the registered images to form ordinary CT three-dimensional image blocks and high-resolution CT three-dimensional image blocks with spatial correspondence. The image generation module is used to input the ordinary CT three-dimensional image block into the super-resolution reconstruction network to generate a predicted bone microstructure three-dimensional image. Based on the pixel difference, structural difference, high-frequency structural difference, and bone microstructure parameter difference between the predicted bone microstructure three-dimensional image and the corresponding high-resolution CT three-dimensional image, a composite loss is constructed to supervise the training of the super-resolution reconstruction network, so that the super-resolution reconstruction network can simultaneously establish the mapping relationship between ordinary CT images and high-resolution bone microstructure images, as well as the correspondence between ordinary CT images and bone microstructure parameters. The parameter acquisition module is used to perform super-resolution reconstruction on the ordinary CT image to be analyzed using the trained super-resolution reconstruction network, obtain a three-dimensional image of predicted bone microstructure, and extract bone density parameters and bone microstructure parameters from the three-dimensional image of predicted bone microstructure based on the mask corresponding to the target bone region and the region of interest of cancellous bone. The result determination module is used to determine the bone quality analysis results based on the bone density parameters and bone microstructure parameters, and output the predicted three-dimensional image of bone microstructure, bone density parameters, bone microstructure parameters, and bone quality analysis results.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the CT bone microstructure and bone quality analysis method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the CT bone microstructure and bone quality analysis method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the CT bone microstructure and bone quality analysis method as described in any one of claims 1 to 6.

Images