A method and system for bone imaging based on dual-view depth enhancement CT

Abstract

This invention discloses a method and system for bone imaging based on dual-view depth-enhanced CT. Utilizing deep learning and multi-strategy fusion techniques, it achieves high-quality image restoration and bone microstructure analysis. The system comprises the following modules: a Cascade-SwinUNETR backbone module for feature extraction and enhancement, enabling efficient multi-scale feature aggregation; a dual-view feature fusion module that extracts and integrates complementary information by combining CT image data from different perspectives, improving image detail and structural accuracy; and an unsupervised domain adaptation (UDA) module to adapt to different data distributions without additional annotation, enabling cross-domain learning and model generalization, ensuring stable performance in various medical scenarios. This invention improves the clarity and accuracy of CT imaging, significantly enhancing the quality and efficiency of physicians' decisions in bone health assessments while reducing reliance on invasive sampling.

Description

Technical Field

[0001] This invention relates to the field of medical imaging, and in particular to a method and system for bone imaging based on dual-view depth-enhanced CT. Background Technology

[0002] With the increasing aging of the global population, osteoporosis and other bone-related diseases are gradually becoming major health problems. Traditional CT imaging technology is widely used in clinical practice to obtain three-dimensional structural images of bones. However, conventional CT images have limited spatial resolution, making it difficult to fully display the details of microstructures such as trabecular bone, which are crucial for assessing bone biomechanical properties and disease risk. While microCT can provide higher resolution, its applicability is limited, typically only used for in vitro specimens or small animal experiments.

[0003] In recent years, with the rapid development of deep learning technology, image restoration and super-resolution techniques have been increasingly applied in medical imaging to improve the quality and detail of CT images. However, most existing methods focus on single-view image processing, failing to fully utilize the complementary information from different viewpoints. Furthermore, the difficulty in obtaining large-scale labeled data limits the widespread application of supervised learning models.

[0004] Therefore, a deep learning method combining dual-view data and unsupervised domain adaptation is proposed, which can make full use of CT data from different perspectives for feature fusion and enhancement, while achieving adaptive adjustment of the model under different data distributions. This is of great significance for improving the resolution of medical imaging and the accuracy of diagnosis. Summary of the Invention

[0005] This invention proposes a dual-view depth-enhanced CT bone imaging method and system. It addresses the limitations of existing CT imaging technologies in terms of detail display and resolution, especially in the analysis of complex bone structures such as trabecular bone. By combining deep learning technology and unsupervised domain adaptation strategies, it can significantly improve the quality and detail display of CT images.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] According to a first aspect of the present invention, a method for bone imaging based on dual-view depth enhancement CT is provided. This method combines feature extraction using a deep learning model and depth feature fusion of dual-view data to implement unsupervised domain adaptation to accommodate different input data distributions, thereby achieving high-resolution image restoration and detail enhancement. Specifically, the method includes the following steps:

[0008] (1) Feature extraction using Cascade-SwinUNETR backbone network: The Cascade-SwinUNETR backbone network is used for efficient feature extraction. Combining the win-Transformer module with the feature aggregation technology of U-Net++, through layer-by-layer feature fusion and deep supervision mechanism, the backbone network can capture complex image details and feature distribution, thereby achieving higher resolution and enhancement of fine structure. The network architecture utilizes multi-scale feature representation to ensure stability and robustness in various imaging noise environments.

[0009] (2) Using dual-view data for deep feature fusion and aggregation: By inputting dual-view data, including CT scan data from different perspectives, the overall clarity and accuracy of bone imaging can be improved. The dual-view feature fusion module uses a deep feature fusion strategy to jointly process the data from the two perspectives and extract complementary information from each perspective. This process uses a deep fusion method based on attention mechanism. By weighting global and local features, the integrity and detail clarity of the image are ensured. The advantage of dual-view fusion is that it compensates for the information loss or artifacts that may exist in a single view, thereby generating a more comprehensive bone structure image.

[0010] (3) Implement unsupervised domain adaptation (UDA) to adapt to different input distributions. The domain adaptation module adopts an unsupervised learning strategy to adapt to different input data distributions, enabling the model to be trained on unlabeled data, thus expanding its adaptability and application scope. By aligning the feature distributions of the target domain and the source domain, the unsupervised domain adaptation module can effectively reduce distribution differences and improve the model's generalization ability, thereby maintaining high-efficiency imaging quality in different environments. This module adopts adversarial training and self-supervised loss function to ensure the model's stable performance on diverse datasets.

[0011] As a preferred embodiment, in step (1), the deep learning architecture DVSR3D utilizes an attention mechanism and autoencoder-style reconstruction to achieve better noise reduction and feature clarity. The architecture adopts an autoencoder-style design and performs image reconstruction by introducing an attention mechanism and deep feature aggregation. The architecture includes an encoder, a feature fusion module, and a decoder module. The encoder is responsible for extracting multi-scale features from dual-view data and weighting them through an attention mechanism to obtain the most relevant feature representations. The fusion module combines local and global attention to ensure the effective integration of input features. The decoder module performs reconstruction, restoring the extracted deep features to a high-resolution image and reducing image artifacts and noise through specific noise reduction strategies, thereby improving image clarity and detail.

[0012] As a preferred embodiment, the dual-view data fusion in step (2) enhances the detail and accuracy of trabecular bone images by combining orthogonal CT views. By inputting CT scan data from orthogonal perspectives (such as horizontal and vertical directions), the microstructural details of trabecular bone are captured. The combination of orthogonal perspectives ensures that the data contains more dimensional information, reducing artifacts and information redundancy. Through the feature fusion module, the system can supplement and balance information from different perspectives, enhancing the structural consistency and detail display of the image. This fusion technology is particularly suitable for medical image analysis that requires high resolution and multi-level details, such as the diagnosis of osteoporosis and bone health assessment.

[0013] As a preferred embodiment, the unsupervised domain adaptation method in step (3) enables the model to be trained on limited-label data, thereby adapting to new medical datasets and enabling the model to be trained efficiently on limited-label or unlabeled data. The specific steps of this method include:

[0014] (31) Feature alignment: By calculating the statistical difference (such as distribution distance) of the feature distributions of the source domain and the target domain, and reducing this difference through the domain adaptation loss function, the robustness of the model under the new data distribution is ensured.

[0015] (32) Adversarial training: By using Generative Adversarial Networks (GANs) or their variants, the model can adaptively adjust to the image features of new environments, reducing the problem of model overfitting to a single data source;

[0016] (33) Self-supervised learning strategy: In the case of no label, the model is trained by using self-supervised mechanisms such as pseudo-label generation or image transformation adversarial, which enhances the generalization ability of new data. In practical applications, it is especially suitable for the field of medical image analysis, such as the consistency problem of imaging data across hospitals or across devices, which improves the applicability and reliability of the model in different medical environments.

[0017] According to a second aspect of the present invention, a dual-view depth-enhanced CT bone imaging system is also provided for CT image restoration tasks. This system includes a deep learning model framework for CT image restoration to perform data acquisition, image registration, and data preprocessing. The framework comprises Cascade-SwinUNETR, a single-view restoration architecture, a deep feature fusion module, and a dual-view restoration architecture. Each module performs a specific function, and these modules can be combined and reused to improve training efficiency and extend generalization capabilities. Cascade-SwinUNETR serves as the foundation module, providing a robust architectural foundation to support the construction of subsequent methods. The dual-view restoration architecture integrates a deep feature fusion module to process dual-view input data.

[0018] In a preferred embodiment, the Cascade-SwinUNETR serves as the base model in the framework, acting as a deep feature recovery module that can be integrated with different input / output modules. This network is adapted to the 3D Swin-Transformer module, U-Net++ structure, and DropKey technology. The single-view recovery architecture, as an independent model for the single-view task, is pre-trained before dual-view learning to accelerate training and enhance generalization ability. After pre-training, the trained recovery model is extracted and integrated into the overall architecture. The deep feature fusion module uses deep feature fusion technology to enable the model to learn, recognize, and combine features from two orthogonal viewpoints in CT scans. This deep feature fusion technology combines features extracted from depth rather than directly merging low-level images to improve performance. The dual-view recovery architecture is a dual-view architecture DVSR3D, including an encoder, a restorer, and a decoder. The encoder part upsamples and fuses deep features, the restorer generates the output image, and the decoder prevents overfitting through regularization.

[0019] As a preferred embodiment, the data acquisition involves a dataset designed to support dual-view tasks, comprising clinical CT scans and micro-CT scans obtained from four different human bone specimens. The experimental setup includes acquiring CT scans from both horizontal and vertical directions, while the micro-CT scans are acquired from only a single direction. Horizontal and vertical directions refer to manually rotating the human bone fragments after the scanning procedure is completed. In practice, this process can be achieved by having the patient lie in different positions, such as supine or prone, or by using different CT scanning devices. It should be noted that all original CT scans were initially unregistered because they were collected under different real-world environmental conditions, thus containing intractable noise distributions and different spatial information.

[0020] As a preferred embodiment, the image registration is performed as follows: Since the data was initially acquired under different settings, accurate image correspondence is required when creating the dataset. To achieve this goal, image registration was performed using the open-source software ITK-SNAP and the SimpleITK Python library. ITK-SNAP was used for both manual and automatic registration, with 16 levels of coarse registration and 8 levels of fine registration. Subsequently, the registration algorithm in SimpleITK was used to optimize the obtained transformation parameters, generating registered data for use in dual-view CT scans.

[0021] In a preferred embodiment, the data preprocessing involves the following steps: Since the original registration data is large in volume, a preprocessing step is required to make it suitable for model training, validation, and testing. Preprocessing includes randomly cropping from the registration data, with each cropping unit containing three types of samples, corresponding to different views. Figure 1 CT scan, visual Figure 2For this task, a super-resolution scale of four times was used for both CT and Micro-CT scans, so the pixel dimension was aligned with this scale. The random cropping process was limited to specific boundaries, with the cropping range restricted to within 100 pixels on each side of the registered data width, height, and depth. The first 60 samples corresponded to the first three bone samples, forming the training dataset. The next 20 samples were associated with the last bone sample and used as the validation dataset. In addition, 20 samples were extracted from the same bone sample to create an independent test dataset. Ideally, if more bone samples were available, it would be best to choose different bone samples as the test dataset. However, due to the challenges of obtaining the original data, this method is the relatively optimal solution in this study. Furthermore, the dynamic range of all data was adjusted to the range of 0 to 255 and stored in grayscale image format.

[0022] According to a third aspect of the present invention, a computer-readable storage medium is also provided, having stored thereon computer-executable instructions that, when executed by a processor, implement the dual-view depth-enhanced CT bone imaging method as described above.

[0023] The technical principle of this invention is as follows: A dual-view depth-enhanced CT bone imaging method and system, combined with deep learning technology, is used to achieve high-precision image restoration and detail enhancement of complex bone structures such as trabecular bone, aiming to improve the imaging quality and reliability in medical diagnosis and bone health assessment. It includes a Cascade-SwinUNETR backbone network for feature extraction, a dual-view feature fusion module to integrate data from different perspectives, a DVSR3D architecture for image reconstruction and noise reduction, and an unsupervised domain adaptation module for adaptive adjustment of cross-domain features. The entire system works collaboratively to achieve high-resolution restoration of CT images.

[0024] Beneficial Effects: This invention significantly improves the detail and structural fidelity of CT images through deep fusion of dual-view data and unsupervised domain adaptation strategies, effectively solving the problems of insufficient resolution and artifacts in traditional CT images. Furthermore, the technical solution of this application does not rely on large-scale labeled data, enabling consistent and robust high-quality image output across different devices and environments, making it suitable for osteoporosis diagnosis and other bone health assessments. This technical solution also reduces reliance on costly and invasive sampling, shortens diagnostic time, and improves diagnostic efficiency, allowing hospitals and clinics to quickly acquire high-precision bone images in their daily work. This helps doctors make diagnostic decisions more quickly, optimize patient treatment plans, and thus improve the overall quality and efficiency of medical services. Attached Figure Description

[0025] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, wherein:

[0026] Figure 1 This is a flowchart illustrating the single-view recovery architecture of the present invention.

[0027] Figure 2 This is a schematic diagram of the dual-view super-resolution 3D (DVSR3D) model architecture of the present invention.

[0028] Figure 3 This is a schematic diagram of the overall process of the unsupervised domain adaptation (UDA) method of the present invention.

[0029] Figure 4 This is a schematic diagram of the overall process of the dual-view depth-enhanced CT bone imaging system of the present invention. Detailed Implementation

[0030] To make the objectives, working principles, and technical solutions of this invention clearer, the invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0031] The following detailed description of the present invention, based on dual-view depth-enhanced CT bone imaging, is provided in conjunction with the accompanying drawings and specific embodiments.

[0032] like Figure 1 As shown, the single-view reconstruction architecture is one of the core components of this embodiment. Its function is to recover high-resolution bone structure images from single-view CT data using a deep learning model. The architecture consists of three main parts: an input layer, a feature extraction backbone network, and an output layer.

[0033] The input layer receives preprocessed single-view CT scan data. The CT data undergoes standardization, denoising, and cropping to ensure data consistency and input quality. Standardization ensures the grayscale values ​​remain within a certain range, preventing noise in the original data from affecting the training performance of subsequent models. The cropping step removes irrelevant regions from the original image, allowing the network to focus on the region of interest and improving computational efficiency.

[0034] In this embodiment, the backbone network for feature extraction adopts the Cascade-SwinUNETR structure, which combines the advantages of Swin-Transformer and U-Net++. Swin-Transformer excels at handling multi-scale features and can capture details and global information in images; while U-Net++, through its hierarchical aggregation capabilities, enhances the network's ability to fuse information at different scales and levels, thereby better recovering the subtle features of bone structures. The backbone network, through multi-level convolutional operations, gradually refines important information in the image while extracting features. Especially for the subtle parts of bone structures (such as the microscopic features of trabecular bone), multi-scale convolutional operations enable relatively fine image restoration. Through inter-layer skip connections and deep supervision mechanisms, the model can more effectively recover the details and structure of images during training.

[0035] The output layer is responsible for generating the restored high-resolution image. During training, the model optimizes multiple loss functions, including mean squared error (MSE) loss and structural similarity index (SSIM) loss, to ensure that the output image is as close as possible to the true label in terms of quality and detail. MSE loss helps minimize pixel-level errors, while SSIM loss focuses more on the fidelity of image structure, better preserving the texture and detail of the image.

[0036] like Figure 2 A depth feature fusion module for dual-view data fusion is demonstrated. This module further enhances the accuracy of image reconstruction by effectively integrating CT scan data from different perspectives (such as horizontal and vertical directions).

[0037] The deep feature fusion module first stitches together feature maps from different perspectives. This process fuses data from different viewpoints, providing richer information for subsequent processing. The stitched feature map contains local details and global structural information from different perspectives, which is particularly important for complex bone structure images (especially trabecular bone with microstructures).

[0038] The stitched feature maps are then processed using convolution operations and non-linear activation functions. This process not only helps extract local features from the image but also enhances the expressive power of global features. Through convolution operations, the network can extract more refined features, while non-linear activation functions enhance the model's non-linear modeling capabilities, enabling the network to better adapt to complex medical image restoration tasks.

[0039] To further optimize the allocation of feature importance, this embodiment introduces an attention mechanism from deep learning. In the fusion module, local and global attention mechanisms are used to weight and fuse features from different perspectives. This mechanism automatically adjusts the weights of different features in the final output based on their contribution, thereby improving the accuracy and detail of image restoration.

[0040] By employing this method, this embodiment can improve the accuracy and detail of image restoration by leveraging complementary information from different perspectives while preserving the details of the input data. Particularly in the field of medical image processing, minute features in bone structure imaging are often crucial for diagnosing diseases (such as osteoporosis), and dual-view fusion technology can significantly improve the resolution of these features.

[0041] like Figure 3 As shown, the unsupervised domain adaptation method in this embodiment aims to improve the model's generalization ability under different data distributions. The UDA module generates a reconstructed image by inputting source and target domain data into a pre-trained single-view reconstruction architecture, and uses a domain adaptation strategy to compare the feature distributions between different domains. Figure 3 The UDA process involves parallel input of source and target domain data. The reconstructed images are then passed to two discriminators to evaluate whether they belong to the same domain. The first discriminator is dedicated to domain discrimination of the reconstructed image, while the second discriminator performs domain discrimination on the reconstructed input image. These two discriminators employ an adversarial training method to reduce the distributional differences between the source and target domains.

[0042] During training, the intensity of the output image of the target domain data may vary significantly due to different acquisition environments. To address this, a simple correction method is employed: mean correction is applied to the recovered target domain image to ensure that the output image intensity is consistent with the input data. The discriminator's output guides the model's updates to align the feature distribution. Through this unsupervised domain adaptation, the model in this embodiment can learn the feature distribution of data in different environments without additional labels, thereby improving the model's adaptability and stability in cross-device and cross-environment applications.

[0043] like Figure 4As shown, in practical implementation, CT images are first standardized and cropped through preprocessing steps to ensure the consistency of input data. Then, single-view CT images are input into the Cascade-SwinUNETR backbone network for pre-training to extract multi-scale features. During training, the model uses MSE and SSIM as loss functions for optimization to ensure that the generated images are highly consistent with the real labels in terms of structure and quality. Next, the dual-view data are integrated through a deep feature fusion module to extract complementary information from different perspectives. The fused feature maps are then input into the deep reconstruction architecture to generate high-resolution output.

[0044] To further enhance the model's generalization ability, an unsupervised domain adaptation module was introduced. Source and target domain data are simultaneously input into the model. After reconstruction, the generated image is passed to the discriminator for domain discrimination. Through adversarial training, the model gradually learns and aligns with features from different data distributions, thereby maintaining consistent image output quality across various environments.

[0045] The technical solution of this invention has significant application value in the field of medical imaging, particularly in the diagnosis of osteoporosis and bone structure analysis. By enhancing the resolution and detail of CT images, doctors can acquire high-quality bone images without invasive sampling, thereby improving diagnostic efficiency and accuracy.

[0046] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A bone imaging method based on dual-view depth-enhanced CT, characterized in that, By combining deep learning models for feature extraction and deep feature fusion of dual-view data, unsupervised domain adaptation is implemented to adapt to different input data distributions, achieving high-resolution image restoration and detail enhancement. Specifically, the steps include: (1) Feature extraction using Cascade-SwinUNETR backbone network: The Cascade-SwinUNETR backbone network is used for efficient feature extraction. Combined with the win-Transformer module and U-Net++ feature aggregation technology, through layer-by-layer feature fusion and deep supervision mechanism, the backbone network can capture complex image details and feature distribution, thereby achieving higher resolution and enhancement of fine structure. The network architecture utilizes multi-scale feature representation to ensure stability and robustness in various imaging noise environments. The deep learning model utilizes attention mechanism and autoencoder style reconstruction to achieve better noise reduction and feature clarity. The architecture adopts autoencoder style design and performs image reconstruction by introducing attention mechanism and deep feature aggregation. The architecture includes encoder, feature fusion module and decoder module. The encoder part is responsible for extracting multi-scale features from dual-view data and weighting them through attention mechanism to obtain the most relevant feature representation. The fusion module combines local and global attention to ensure effective integration of input features. The decoder module performs reconstruction, restores the extracted deep features to high-resolution image, and reduces image artifacts and noise through specific noise reduction strategies to improve image clarity and detail performance. (2) Using dual-view data for deep feature fusion and aggregation: By inputting dual-view data, including CT scan data from different perspectives, the overall clarity and accuracy of bone imaging can be improved. The dual-view feature fusion module uses a deep feature fusion strategy to jointly process the data from the two perspectives and extract complementary information from each perspective. This process uses a deep fusion method based on attention mechanism. By weighting global and local features, the integrity and detail clarity of the image are ensured. The advantage of dual-view fusion is that it compensates for the information loss or artifacts that may exist in a single view, thereby generating a more comprehensive bone structure image. (3) Implement unsupervised domain adaptation to adapt to different input distributions. The domain adaptation module adopts an unsupervised learning strategy to adapt to different input data distributions, enabling the model to be trained on unlabeled data, thus expanding its adaptability and application scope. By aligning the feature distributions of the target domain and the source domain, the unsupervised domain adaptation module can effectively reduce distribution differences and improve the model's generalization ability, thereby maintaining high-efficiency imaging quality in different environments. This module adopts adversarial training and self-supervised loss function to ensure the model's stable performance on diverse datasets.

2. The bone imaging method based on dual-view depth-enhanced CT according to claim 1, characterized in that, The dual-view data fusion in step (2) enhances the detail and accuracy of trabecular bone images by combining orthogonal CT views. By inputting CT scan data from orthogonal perspectives, the microstructural details of the trabecular bone are captured. The combination of orthogonal perspectives ensures that the data contains more dimensional information, reducing artifacts and information redundancy. Through the feature fusion module, the system can supplement and balance information from different perspectives, enhancing the structural consistency and detail display of the image.

3. The bone imaging method based on dual-view depth-enhanced CT according to claim 1, characterized in that, The unsupervised domain adaptation method in step (3) enables the model to be trained on limited labeled data, thereby adapting to new medical datasets and enabling the model to be trained efficiently on limited or unlabeled data. The specific steps of this method include: (31) Feature alignment: By calculating the statistical difference between the feature distributions of the source domain and the target domain, and reducing this difference through the domain adaptation loss function, the robustness of the model under the new data distribution is ensured; (32) Adversarial training: By using generative adversarial networks or their variants, the model can adaptively adjust to the image features of new environments, reducing the problem of model overfitting to a single data source; (33) Self-supervised learning strategy: In the case of no label, the model is trained by using self-supervised mechanisms such as pseudo-label generation or image transformation adversarial, which enhances the generalization ability of new data. In practical applications, it is applicable to the field of medical image analysis, such as the consistency problem of imaging data across hospitals or across devices, and improves the applicability and reliability of the model in different medical environments.

4. A dual-view depth-enhanced CT bone imaging system for performing the method according to any one of claims 1 to 3, characterized in that, This system is designed for CT image restoration tasks and includes a deep learning model framework for CT image restoration, enabling data acquisition, image registration, and data preprocessing. It comprises Cascade-SwinUNETR, a single-view restoration architecture, a deep feature fusion module, and a dual-view restoration architecture. Each module performs a specific function and can be combined and reused to improve training efficiency and extend generalization capabilities. Cascade-SwinUNETR serves as the foundational module, providing a robust architectural foundation to support the construction of subsequent methods. The dual-view restoration architecture integrates a deep feature fusion module to handle dual-view input data.

5. A bone imaging system based on dual-view depth-enhanced CT according to claim 4, characterized in that, The Cascade-SwinUNETR serves as the base model in the framework, acting as a deep feature recovery module. It can be integrated with different input / output modules. The network is adapted to the 3D Swin-Transformer module, U-Net++ structure, and DropKey technology. The single-view recovery architecture, as an independent model for the single-view task, is pre-trained before dual-view learning to accelerate training and enhance generalization ability. After pre-training, the trained recovery model is extracted and integrated into the overall architecture. The deep feature fusion module uses deep feature fusion technology to enable the model to learn, recognize, and combine features from two orthogonal viewpoints in CT scans. This deep feature fusion technology combines features extracted from depth rather than directly merging low-level images to improve performance. The dual-view recovery architecture is a dual-view architecture DVSR3D, including an encoder, a restorer, and a decoder. The encoder part upsamples and fuses deep features, the restorer generates the output image, and the decoder prevents overfitting through regularization.

6. A bone imaging system based on dual-view depth-enhanced CT according to claim 4, characterized in that, The data acquisition: The compiled dataset was designed to support dual-view tasks and included clinical CT scans and micro-CT scans obtained from four different human bone specimens. The experimental setup included acquiring CT scans from both horizontal and vertical directions, while the micro-CT scans were acquired from only one direction. Horizontal and vertical directions refer to manually rotating the human bone fragments after the scanning procedure is completed. In practical applications, this process can be achieved by having the patient lie in different directions, such as supine or prone, or by using different CT scanning devices. It should be noted that all raw CT scans were initially unregistered because they were collected under different real-world environmental conditions and therefore contained intractable noise distributions and different spatial information.

7. A bone imaging system based on dual-view depth-enhanced CT according to claim 4, characterized in that, Image registration: Since the data was initially acquired under different settings, accurate image correspondence was required when creating the dataset. To achieve this, image registration was performed using the open-source software ITK-SNAP and the Simple ITK Python library. ITK-SNAP was used for both manual and automatic registration, with 16 levels of coarse registration and 8 levels of fine registration. Subsequently, the registration algorithm in Simple ITK was used to optimize the obtained transformation parameters, generating registered data for use in dual-view CT scans.

8. A bone imaging system based on dual-view depth-enhanced CT according to claim 4, characterized in that, Data preprocessing: Due to the large volume of the original registration data, preprocessing steps are required to make it suitable for model training, validation, and testing. Preprocessing includes random cropping from the registration data. Each cropping unit contains three types of samples, corresponding to View 1 CT scan, View 2 CT scan, and Micro-CT scan, respectively. The super-resolution ratio used in this task is four times, so the pixel dimension is aligned with this ratio. The random cropping process is restricted to specific boundaries, with the cropping range limited to within 100 pixels on each side of the width, height, and depth of the registered data. The first 60 samples correspond to the first three bone samples, constituting the training dataset. The next 20 samples are associated with the last bone sample and used as the validation dataset. In addition, 20 more samples are extracted from the same bone sample to create an independent test dataset. Ideally, if more bone samples are available, different bone samples would be selected as the test dataset. However, due to the challenges of obtaining the original data, this method is the relatively optimal solution in this study. Furthermore, the dynamic range of all data is adjusted to the range of 0 to 255 and stored in grayscale image format.

9. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that, When the executable instructions are executed by the processor, they implement the dual-view depth-enhanced CT bone imaging method as described in any one of claims 1 to 3.

Images