Axial spondyloarthritis sacroiliac joint MRI synthetic CT image processing system and method based on deep learning algorithm

Abstract

The application discloses a kind of middle axis type spondyloarthritis sacroiliac joint MRI synthetic CT image processing system and method based on deep learning algorithm, system includes: data acquisition unit, for obtaining the magnetic resonance sequence data obtained by magnetic resonance imaging to target user sacroiliac joint;Deep learning model unit, for the magnetic resonance sequence data is input to the deep learning algorithm model trained using federation learning, obtain the bone cortex feature and soft tissue feature of sacroiliac joint;Image synthesis unit, for synthesizing the synthetic CT image corresponding to the sacroiliac joint of the target user based on the bone cortex feature and the soft tissue feature, realize the high-precision CT image acquisition of no radiation, both can satisfy the diagnosis demand to middle axis type spondyloarthritis sacroiliac joint, and can reduce the influence of radiation to target user.

Description

Technical Field

[0001] This disclosure generally relates to the field of medical image processing technology, and specifically to a system and method for processing MRI-synthesized CT images of the sacroiliac joint in axial spondyloarthritis based on deep learning algorithms. Background Technology

[0002] Axial spondyloarthritis (axSpA) is a chronic inflammatory arthritis characterized primarily by arthritis of the axial spine. This disease commonly affects the sacroiliac joints and can lead to bamboo spine changes, severely impacting patients' quality of life. Accurate detection of sacroiliac joint structural damage is crucial for the radiographically positive stage of axial spondyloarthritis (i.e., ankylosing spondylitis).

[0003] Currently, clinical practice mainly relies on the combined use of computed tomography (CT) and magnetic resonance imaging (MRI) for image interpretation. Specifically, CT can clearly show structural damage such as cortical bone erosion, sclerosis, and joint ankylosis through its high X-ray attenuation coefficient, but its ionizing radiation risk limits repeated examinations, especially in pediatric patients. While MRI is radiation-free, it has insufficient resolution and low sensitivity for the edge of bone erosion.

[0004] Therefore, how to provide a CT image that can clearly display structural damage such as cortical erosion and sclerosis without radiation scanning has become an urgent problem to be solved. Summary of the Invention

[0005] In view of the above-mentioned defects or deficiencies in the existing technology, it is desirable to provide a deep learning algorithm-based MRI-to-CT image processing system and method for axial spondyloarthritis of the sacroiliac joint, which can achieve radiation-free high-precision CT image acquisition, meet the diagnostic needs of axial spondyloarthritis of the sacroiliac joint, and reduce the impact of radiation on the target user.

[0006] In a first aspect, embodiments of this application provide a deep learning algorithm-based MRI-to-CT image processing system for the sacroiliac joint in axial spondyloarthritis, comprising:

[0007] The data acquisition unit is used to acquire magnetic resonance sequence data obtained by magnetic resonance imaging of the sacroiliac joint of the target user.

[0008] A deep learning model unit is used to input the magnetic resonance sequence data into a deep learning algorithm model trained using federated learning to obtain the cortical bone features and soft tissue features of the sacroiliac joint.

[0009] An image synthesis unit is used to synthesize a synthetic CT image of the target user's sacroiliac joint based on the bone cortex features and the soft tissue features.

[0010] In some embodiments, the magnetic resonance sequence data includes a three-dimensional T1-weighted multigradient echo sequence, a T2-weighted sequence, and a water-lipid separation sequence;

[0011] The settings for the three-dimensional T1-weighted multigradient echo sequence include TR = 5.2 ms, TE = 2.5 ms, and a layer thickness of 0.8 mm; the settings for the T2-weighted sequence are TR = 3000 ms and TE = 80 ms.

[0012] In some embodiments, the data acquisition unit is further configured to:

[0013] Identify the age correction amount for the target user, and adjust the flip angle of the three-dimensional T1-weighted multigradient echo sequence based on the age correction amount.

[0014] In some embodiments, the deep learning algorithm model unit includes an anatomy-specific attention subunit.

[0015] The anatomically specific attention subunit includes a spatial-channel dual-path attention mechanism.

[0016] In some embodiments, deformable convolution is used in the spatial path to dynamically adjust the sampling points, so that the convolution operation adapts to the geometric deformation of the sacroiliac joint anatomy.

[0017] In some embodiments, the image synthesis unit includes an anatomical prior constraint subunit.

[0018] The anatomical prior constraint subunit is used to provide an anatomical loss function for the deep learning algorithm model during the training phase of the deep learning algorithm model. The anatomical prior constraints include at least the cortical bone continuity loss constraint, the interarticular distance loss constraint, and the bone erosion area resistance loss constraint.

[0019] In some embodiments, the image synthesis unit includes a feature alignment subunit.

[0020] The feature alignment subunit is used during the training phase of the deep learning algorithm model to make the distribution of bone cortex features and soft tissue features in the synthesized CT image approximate the distribution of bone cortex features and soft tissue features in the real CT image.

[0021] In some embodiments, the image synthesis unit is further configured to:

[0022] Based on the soft tissue features, HU values ​​are mapped in the synthetic CT images.

[0023] In some embodiments, quantitative analysis is also included.

[0024] The quantitative analysis unit is used to determine the bone erosion volume and maximum depth in the synthetic CT image based on the HU value mapped in the synthetic CT image.

[0025] Secondly, embodiments of this application provide a method for processing MRI-to-CT images of the sacroiliac joint in axial spondyloarthritis based on deep learning algorithms, including:

[0026] Acquire magnetic resonance sequence data obtained from magnetic resonance imaging of the sacroiliac joint of the target user;

[0027] The magnetic resonance sequence data is input into a deep learning algorithm model trained using federated learning to obtain the cortical bone and soft tissue features of the sacroiliac joint.

[0028] Based on the bone cortex features and the soft tissue features, a synthetic CT image corresponding to the sacroiliac joint of the target user is synthesized.

[0029] Thirdly, embodiments of this application provide 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 program to implement the method described in embodiments of this application.

[0030] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in embodiments of this application.

[0031] Fifthly, embodiments of this application provide a computer program product, including a computer program, characterized in that, when the computer program is executed by a processor, it implements the method described in embodiments of this application.

[0032] This application provides a deep learning algorithm-based MRI-to-CT image processing system and method for axial spondyloarthritis of the sacroiliac joint. It can obtain the cortical bone and soft tissue features of the sacroiliac joint using the magnetic resonance sequence data of the target user's sacroiliac joint, and then use the cortical bone and soft tissue features to synthesize the corresponding synthetic CT image of the target user's sacroiliac joint. This achieves radiation-free, high-precision CT image acquisition, which can meet the diagnostic needs of axial spondyloarthritis of the sacroiliac joint and reduce the impact of radiation on the target user.

[0033] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0034] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0035] Figure 1 An MRI image of the sacroiliac joint is shown;

[0036] Figure 2 A flowchart illustrating a method for processing MRI-to-CT images of the sacroiliac joint in axial spondyloarthritis based on a deep learning algorithm, according to an embodiment of this application, is shown.

[0037] Figure 3 This illustration shows a schematic diagram of the structure of a deep learning algorithm-based MRI-to-CT image processing system for axial spondyloarthritis of the sacroiliac joint, according to an embodiment of this application.

[0038] Figure 4 A schematic diagram of the structure of a deep learning algorithm-based MRI-to-CT image processing system for axial spondyloarthritis of the sacroiliac joint is shown in another embodiment of this application. Detailed Implementation

[0039] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0040] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0041] axSpA is a chronic inflammatory rheumatic disease that severely impacts patients' quality of life, and the sacroiliac joint is a key site for early diagnosis and treatment evaluation. MRI primarily uses magnetic fields and radio waves for imaging, emphasizing the contrast and detail of soft tissues. MRI is highly effective when detecting changes in soft tissues (such as inflammation, edema, etc.). Figure 1 As indicated by the arrows in (a) and (b). However, MRI has limited imaging capabilities for bone structures because the minerals in bone have a relatively small impact on the MRI signal, resulting in insufficient sensitivity in detecting bone destruction. In MRI images, bone tissue typically appears as a low signal intensity, making it difficult to separate from surrounding tissues, thus further complicating the interpretation of bone destruction, such as... Figure 1 As shown by * in (a) and (b).

[0042] CT's high spatial resolution and bone density-sensitive imaging characteristics make it more accurate in identifying minute bone damage, better revealing subtle changes in bone, including bone destruction and sclerosis. However, CT carries the risk of ionizing radiation and is not suitable for frequent repeat examinations, especially for young patients. Many patients may have already undergone MRI scans, and repeating CT scans not only increases radiation risk but also leads to wasted time and effort. Therefore, developing a radiation-free imaging method that can accurately reconstruct bone details has significant clinical value. Deep learning has shown great potential in medical image synthesis and reconstruction. By constructing accurate cross-modal image generation models, it is hoped that non-destructive conversion from MRI to CT can be achieved, providing innovative solutions for the early diagnosis and treatment monitoring of axSpA.

[0043] Based on this, this application proposes a deep learning algorithm-based MRI-to-CT image processing system and method for axial spondyloarthritis of the sacroiliac joint, to achieve radiation-free CT image acquisition and provide high-precision examination images for more accurate diagnosis and treatment decisions.

[0044] To further illustrate the technical solutions provided in the embodiments of this application, a detailed description is provided below in conjunction with the accompanying drawings and specific implementation methods. Although the embodiments of this application provide method operation instruction steps as shown in the following embodiments or drawings, the method may include more or fewer operation instruction steps based on conventional or non-creative effort. In steps where there is no logically necessary causal relationship, the execution order of these steps is not limited to the execution order provided in the embodiments of this application. In actual processing or when the device executes the method, it may be executed sequentially or in parallel according to the method shown in the embodiments or drawings.

[0045] It should be noted that the acquisition or use of data in the embodiments of this application requires the user's consent. The relevant data can only be obtained after the user's authorization and permission, and the acquisition or use of the data complies with the provisions of relevant laws and regulations.

[0046] Please refer to Figure 2 , Figure 2 This illustration shows a flowchart of a method for processing MRI-to-CT images of the sacroiliac joint in axial spondyloarthritis based on a deep learning algorithm, according to an embodiment of this application. Figure 2 As shown, the method includes:

[0047] Step 201: Obtain magnetic resonance sequence data obtained from magnetic resonance imaging of the target user's sacroiliac joint.

[0048] It should be noted that the target users are those who are about to undergo a sacroiliac joint axSpA diagnostic assessment.

[0049] Specifically, magnetic resonance imaging was performed on the sacroiliac joint of the target user. By setting the parameters TR=5.2ms, TE=2.5ms, and slice thickness to 0.8mm, a three-dimensional T1-weighted multi-gradient echo sequence was obtained. By setting the parameters TR=3000ms and TE=80ms, a T2-weighted sequence was obtained, as well as a water-fat separation sequence of the sacroiliac joint of the target user was acquired.

[0050] In magnetic resonance imaging (MRI), the three-dimensional T1-weighted multi-gradient echo sequence is abbreviated as 3D-T1-MGE. 3D refers to three-dimensional volumetric imaging, which achieves arbitrary plane reconstruction (e.g., sagittal, coronal, axial) by continuously acquiring isotropic voxels, avoiding the gaps between slices in traditional 2D layering. T1-weighting highlights the T1 relaxation differences of tissues and displays anatomical structures by adjusting repetition time (TR) and echo time (TE). Multi-gradient echo (MGE) sequence acquires signals with different TE values ​​using multiple gradient echoes after a single radiofrequency excitation, improving scanning efficiency and supporting quantitative analysis, providing a data basis for subsequent synthesis of synthetic CT with HU (Hounsfield Hnit) values.

[0051] In a preferred embodiment, before performing magnetic resonance imaging on the target user, the age correction amount of the target user is further identified, and the inversion angle of the three-dimensional T1-weighted multigradient echo sequence is adjusted based on the age correction amount.

[0052] In other words, the inversion angle of the 3D T1-weighted gradient echo sequence can be adjusted based on the user's age. Specifically, the target user's age is obtained. If the target user's age is greater than or equal to a preset age, it means the target user is an adult and no adjustment is needed; the corresponding age correction is 0. If the target user's age is less than the preset age, it means the target user is a child and adjustment is needed. The age correction is then obtained, and the inversion angle of the 3D T1-weighted gradient echo sequence is adjusted based on the age correction.

[0053] Optionally, the age correction amount can be a fixed value, such as 5°, or a value that is linearly related to age; this application does not impose any specific limitations.

[0054] For example, the age correction is a fixed value of 5°, and when the target user is identified as a child, the inversion angle of the three-dimensional T1-weighted gradient echo sequence is adjusted from 15° to 10°.

[0055] Therefore, this application reduces the specific absorption ratio (SAR) by adjusting the inversion angle of the three-dimensional T1-weighted gradient echo sequence to meet the safety requirements for pediatric scanning.

[0056] In some embodiments, after acquiring the magnetic resonance sequence data, a rigid registration algorithm based on mutual information is further employed to align the three-dimensional T1-weighted multi-gradient echo sequence, T2-weighted sequence, and water-lipid separation sequence with the reference CT, with a registration error of less than 0.5 mm, so as to facilitate high-precision feature extraction in subsequent magnetic resonance sequence data based on multiple sequences.

[0057] Preferably, for the aligned magnetic resonance sequence data, the Nyul histogram matching algorithm is further used to eliminate signal differences in multi-center magnetic resonance devices.

[0058] Step 202: Input the magnetic resonance sequence data into a deep learning algorithm model trained using federated learning to obtain the cortical bone features and soft tissue features of the sacroiliac joint.

[0059] It should be noted that the deep learning algorithm model is a feature analysis model trained using a federated learning framework. Specifically, the federated learning framework allows each medical institution to train its own model using local data, sharing only encrypted gradient parameters with each other. This ensures both the generalization ability of the model training and solves the privacy barrier problem among multiple medical centers.

[0060] In some embodiments, the deep learning algorithm model is a hybrid model of adversarial networks and U-Net. That is, the backbone network of the deep learning algorithm model adopts a U-Net architecture, fusing multi-scale features through nested dense skip connections to preserve local details and global anatomical structures of the sacroiliac joint. Preferably, residual blocks are introduced between the encoder and decoder to alleviate the gradient vanishing problem while enhancing the transmission of low-frequency features of the bone cortex signal, thereby improving the accuracy of the synthesized CT images.

[0061] In one feasible embodiment, the deep learning algorithm model includes an anatomy-specific attention subunit. This anatomy-specific attention subunit includes a spatial-channel dual-path attention mechanism.

[0062] It should be noted that the spatial-channel dual-path attention mechanism includes a spatial path and a channel path. The spatial path is used to introduce deformable convolution to capture irregular edges of the sacroiliac joint, dynamically weighting key anatomical regions through spatial attention maps. The channel path is used to employ channel attention to enhance the signal response features of the cortical bone and sclerotic regions. It should be understood that the dual-path outputs undergo feature fusion via an adaptive fusion module.

[0063] Specifically, the multi-scale features obtained by fusing multimodal features from 3D T1-weighted multi-gradient echo sequences, T2-weighted sequences, and water-lipid separation sequences can be input into the spatial path and channel path, respectively, to extract deformation features y through deformable convolution in the spatial path.def Generate spatial attention map A s To generate channel weights A through channel attention in the channel path. c Finally, the outputs from the two paths are weighted and fused to obtain the enhanced fused feature y. out .

[0064] On the one hand, due to the complex structure of the sacroiliac joint, it is normally composed of an uneven auricular surface covered with fibrocartilage and hyaline cartilage. During inflammation or degeneration, the cartilage peels off, exposing the cortical bone, and the natural irregular shape is magnified. Ankylosing spondylitis can cause inflammation of the sacroiliac joint, leading to subchondral bone erosion, followed by fibrosis and osteophyte formation. The joint surface gradually hardens, but the surface remains uneven.

[0065] Based on this, this application uses deformable convolution to dynamically adjust sampling points in the spatial path, fits the geometric deformation of the articular surface, and improves the expression of key anatomical features in MRI data, such as the articular surface of the sacroiliac joint, the edge of bone erosion, the sclerotic region, and ankylosis.

[0066] Specifically, this application employs deformable convolution to introduce a dynamic offset Δp:

[0067]

[0068] Among them, y def (p) is a deformable convolution at the current position p of the feature map to extract deformation features, p k For the predefined convolution kernel offset, Δp k (p) represents the dynamic offset of the kth sampling point.

[0069] It should be understood that the offset Δp is determined entirely by the content of the input feature map, rather than being predefined. Therefore, different offset patterns will be generated in different regions of the multi-scale feature map generated based on sacroiliac joint MRI data. For example, Δp approaches 0 in the hardened region, maintaining standard convolution behavior, indicating that the edge of the region is regular. Δp is significantly non-zero at the edge of the articular surface, aligning the sampling points to the joint anatomy, indicating that the edge of the region is irregular.

[0070] Furthermore, by using deformable convolution to introduce a dynamic offset Δp, the sampling points at the sacroiliac joint surface can be concentrated towards the protruding part of the joint surface edge, avoiding the problem of difficult alignment in traditional convolution. At the same time, the sampling points can be made to penetrate into the concave area in the bone erosion area, avoiding the problem that small concavities are easily ignored by smooth convolution.

[0071] Therefore, this application introduces a dynamic offset Δp through deformable convolution in the spatial path, enabling the convolution operation to adapt to the geometric deformation of the anatomical structure of the sacroiliac joint, thereby accurately capturing irregular structures such as the sacroiliac joint surface and bone erosion, improving edge alignment capability, providing more accurate image data support for subsequent CT image synthesis, and compensating for the inherent data insensitivity of MRI data.

[0072] Furthermore, a spatial attention map is generated based on the deformation features in the spatial path. The specific address can be obtained using the following formula:

[0073] A s =σ(Conv(ReLU(Conv(y) def ))))

[0074] Among them, A s This is the spatial weighted attention map corresponding to the deformation features. Conv is the convolutional layer, ReLU is the activation function, and σ is the normalized weight.

[0075] On the other hand, this application utilizes the SE-Net channel attention mechanism to dynamically enhance high-frequency feature channels related to the continuity of the bone cortex.

[0076] Specifically, the weights of feature channels relevant to the target region are enhanced by leveraging the dependencies between channels. For example, the following formula can be used:

[0077] A c =σ(W u ·δ(W d ·GAP(y channel )))

[0078] Among them, A c For channel weights, GAP(y) channel W is a channel descriptor obtained based on global average pooling and compressed spatial dimensions. d and W u For dimensionality reduction and dimensionality increase, δ is the activation function and σ is the normalized weight.

[0079] Furthermore, a gating mechanism is adopted to weight and fuse the dual-path outputs, so as to adaptively fuse the attention information of the spatial path and the channel path, and prioritize the reconstruction weight of key areas such as the sacroiliac joint surface and the sclerotic area.

[0080] For example, the following formula can be used:

[0081] y out =α·(A s ⊙y def )+(1-α)·(A c ⊙y channel )

[0082] Among them, y out For the fused features of the output, α is the contribution coefficient of the dynamic equilibrium space and channel attention, and y def To extract deformation features using deformable convolution, y channel As a channel feature, A s For spatial attention weights, A c This represents the channel attention weight.

[0083] Therefore, this application can amplify the details of key anatomical sites and enhance the channel response of specific pathological features through the anatomically specific attention subunit composed of a spatial-channel dual-path attention mechanism.

[0084] Preferably, after obtaining the fusion features, this application further classifies the fusion features to obtain the cortical bone features and soft tissue features of the sacroiliac joint, so as to synthesize CT images based on the cortical bone features and soft tissue features.

[0085] Step 203: Synthesize the target user's sacroiliac joint based on the bone cortex features and soft tissue features.

[0086] In other words, after obtaining more accurate cortical bone and soft tissue features through feature analysis using magnetic resonance sequence data, CT images are then synthesized using these features to meet the diagnostic needs of the sacroiliac joint axSpA.

[0087] In some embodiments, an anatomical loss function is provided to the deep learning algorithm model during the training phase to ensure that the CT images generated by the trained deep learning algorithm model strictly conform to the anatomical features of the sacroiliac joint. The anatomical prior constraints include at least cortical bone continuity loss constraints, joint space distance loss constraints, and bone erosion area resistance loss constraints.

[0088] It should be understood that anatomical loss constraints can ensure that the generated CT images strictly conform to the anatomical features of the sacroiliac joint, including cortical continuity, interarticular distance, and bone erosion morphology, thereby effectively improving the usability of synthetic CT images in clinical diagnosis.

[0089] For example, the cortical bone continuity loss constraint can be expressed by the following formula:

[0090]

[0091] in, M is the continuity loss constraint function for the bone cortex, Dice is the similarity coefficient used to calculate the overlap between real and synthetic CT scans in the bone cortex. real To use real CT scans in training, Mfake For the synthetic CT during training, λ g These are the weighting coefficients for the gradient difference loss. This is a gradient operator used to extract the edge gradient of the bone cortex.

[0092] Therefore, by using the continuity loss constraint of the bone cortex, it is possible to ensure that the bone cortex region of the synthetic CT is spatially aligned with the real CT, while using gradient difference to enhance the continuity of the bone cortex edge and avoid generating broken or blurred edges.

[0093] For example, the joint gap distance loss constraint can be expressed by the following formula:

[0094]

[0095] Hausdorff(A, B)=max(sup infd(a, b), sup infd(a, b))

[0096] in, Let d(a, b) be the Euclidean distance loss constraint function, d(a, b) be the Euclidean distance between points a and b, sup be the supremum, inf be the infremum, and S be the inequality. real For joint space segmentation in real CT scans, S fake Joint space segmentation for synthetic CT.

[0097] Therefore, by using the joint space distance loss constraint, the anatomical rationality of the joint space can be constrained, ensuring that the width and shape of the joint space in the synthesized CT image conform to the real anatomical structure.

[0098] For example, the resistance to loss constraints in the bone erosion region can be expressed by the following formula:

[0099]

[0100] in, D is the constraint function for the resistance loss in the bone erosion region. patch G(x) is a local discriminator used for true synthesis discrimination of regions of bone erosion (ROI). ROI To synthesize local image patches of bone erosion areas in CT images, y ROI The ROI is a local image patch of the bone erosion area in a real CT image, and it is determined manually during the training process.

[0101] Therefore, by utilizing the bone erosion region to counteract loss constraints, we can enhance the realism of local textures by targeting the peripheral features of the bone erosion region, such as irregular bone destruction and edge sharpness, and avoid ignoring local pathological details during the synthetic CT process when the overall image quality meets the standards.

[0102] In a preferred embodiment, during the deep learning algorithm model training phase, the distribution of cortical bone and soft tissue features in the synthesized CT image is made to approximate the distribution of cortical bone and soft tissue features in the real CT image.

[0103] Specifically, adversarial training is conducted using adversarial networks during the training process, which makes the deep feature distribution of synthetic CT and real CT tend to be consistent.

[0104] In a preferred embodiment, during the synthesis of the synthetic CT image, HU values ​​are also mapped in the synthetic CT image based on soft tissue features.

[0105] It should be noted that the HU value is a core physical property of CT images, directly reflecting tissue density differences. In the task of synthesizing CT images, ensuring that the synthesized CT image conforms to the HU calibration range of the real CT image is one of the indicators for judging whether the synthesized CT image is successful.

[0106] It should also be noted that traditional generative adversarial networks (GANs) typically use the Tanh activation function to limit the output to the range [-1, 1]. However, the HU values ​​of real CT images have a wide range; for example, the HU value for air is -1000 HU, for water it is 0 HU, for fat it is -70 to -90 HU, and for cortical bone it is +1000 HU. Therefore, to ensure the reliability of the HU values ​​in the synthesized CT images, this application employs a linear activation function, allowing the generator to directly output an unrestricted range of values. Then, data standardization and destandardization steps are used to adapt the output to the physical range of real HU values.

[0107] Specifically, during the training phase, the HU values ​​of real CT are first standardized, and then the last layer of the generator is trained to use linear activation to directly predict the standardized HU values. During the application phase, the output of the synthetic CT generator is destandardized to recover the HU values ​​corresponding to the synthetic CT, thereby ensuring that the HU values ​​corresponding to the synthetic CT are consistent with the HU values ​​of the real CT.

[0108] In a preferred embodiment, based on the HU value mapped when synthesizing the synthetic CT image, the volume and maximum depth of bone erosion in the synthetic CT image can be further determined according to the HU value mapped in the synthetic CT image.

[0109] Specifically, the region is segmented according to a preset HU value range based on the HU value in the synthetic CT image, and the bone erosion volume and maximum depth are calculated based on the image region that meets the bone erosion area.

[0110] Preferably, the HU value corresponding to the bone erosion region is greater than 400. That is, the region with a HU value greater than 400 in the synthesized CT image is defined as the synthesized bone erosion region, and the pixel range occupied by the bone erosion region is calculated to determine the volume of the bone erosion region. Thus, based on the ability to automatically calculate the bone erosion region, the calculation error is limited to ±0.15cm. 3 Inside.

[0111] Therefore, the present application provides a method for processing MRI-synthesized CT images of the sacroiliac joint based on deep learning algorithms for axial spondyloarthritis. This method can obtain the cortical bone and soft tissue features of the sacroiliac joint using the magnetic resonance sequence data of the target user's sacroiliac joint, and then synthesize the corresponding synthetic CT image of the target user's sacroiliac joint using the cortical bone and soft tissue features. This achieves high-precision CT image acquisition without radiation, which can not only meet the diagnostic needs of axial spondyloarthritis of the sacroiliac joint, but also reduce the impact of radiation on the target user.

[0112] It should be noted that although the operation of the method of the present invention is described in a specific order in the accompanying drawings, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed in order to achieve the desired result.

[0113] Figure 3 This paper illustrates a schematic diagram of the structure of a deep learning algorithm-based MRI-to-CT image processing system for the sacroiliac joint in axial spondyloarthritis, according to an embodiment of this application.

[0114] like Figure 3 As shown in the embodiment of this application, the MRI-to-CT image processing system 10 for axial spondyloarthritis of the sacroiliac joint based on deep learning algorithms includes:

[0115] Data acquisition unit 11 is used to acquire magnetic resonance sequence data obtained by magnetic resonance imaging of the sacroiliac joint of the target user;

[0116] The deep learning model unit 12 is used to input the magnetic resonance sequence data into a deep learning algorithm model trained by federated learning to obtain the cortical bone features and soft tissue features of the sacroiliac joint.

[0117] Image synthesis unit 13 is used to synthesize a synthetic CT image of the target user's sacroiliac joint based on the bone cortex features and the soft tissue features.

[0118] In some embodiments, the magnetic resonance sequence data includes a three-dimensional T1-weighted multigradient echo sequence, a T2-weighted sequence, and a water-lipid separation sequence;

[0119] The settings for the three-dimensional T1-weighted multigradient echo sequence include TR = 5.2 ms, TE = 2.5 ms, and a layer thickness of 0.8 mm; the settings for the T2-weighted sequence are TR = 3000 ms and TE = 80 ms.

[0120] In some embodiments, the data acquisition unit 11 is further configured to:

[0121] Identify the age correction amount for the target user, and adjust the flip angle of the three-dimensional T1-weighted multigradient echo sequence based on the age correction amount.

[0122] In some embodiments, such as Figure 4 As shown, the deep learning algorithm model unit 12 includes an anatomy-specific attention subunit 121.

[0123] The anatomical-specific attention subunit 121 includes a spatial-channel dual-path attention mechanism.

[0124] In some embodiments, deformable convolution is used in the spatial path to dynamically adjust the sampling points, so that the convolution operation adapts to the geometric deformation of the sacroiliac joint anatomy.

[0125] In some embodiments, such as Figure 4 As shown, the image synthesis unit 13 includes an anatomical prior constraint subunit 131.

[0126] The anatomical prior constraint subunit 131 is used to provide an anatomical loss function for the deep learning algorithm model during the training phase of the deep learning algorithm model. The anatomical prior constraints include at least the cortical bone continuity loss constraint, the interarticular distance loss constraint, and the bone erosion area resistance loss constraint.

[0127] In some embodiments, such as Figure 4 As shown, the image synthesis unit 13 includes a feature alignment subunit 132.

[0128] The feature alignment subunit 132 is used during the deep learning algorithm model training phase to make the distribution of bone cortex features and soft tissue features in the synthesized CT image approximate the distribution of bone cortex features and soft tissue features in the real CT image.

[0129] In some embodiments, such as Figure 4 As shown, the image synthesis unit 13 is further configured to:

[0130] Based on the soft tissue features, HU values ​​are mapped in the synthetic CT images.

[0131] In some embodiments, such as Figure 4 As shown, it also includes quantitative analysis 14.

[0132] The quantitative analysis unit 14 is used to determine the bone erosion volume and maximum depth in the synthetic CT image based on the HU value mapped in the synthetic CT image.

[0133] It should be understood that the modules or modules described in the deep learning algorithm-based deep learning algorithm-based MRI-to-CT image processing system for the sacroiliac joint in axial spondyloarthritis are similar to those in the reference... Figure 2 The steps in the described method correspond accordingly. Therefore, the operations and features described above for the method are also applicable to the deep learning algorithm-based MRI-to-CT image processing system 10 for axial spondyloarthritis and its included modules, and will not be repeated here. The deep learning algorithm-based MRI-to-CT image processing system 10 for axial spondyloarthritis and its included modules can be pre-implemented in the browser or other secure applications of an electronic device, or can be loaded into the browser or its secure applications of an electronic device through download or other means. The corresponding modules in the deep learning algorithm-based MRI-to-CT image processing system 10 for axial spondyloarthritis and its included modules can cooperate with modules in the electronic device to implement the solutions of the embodiments of this application.

[0134] The division of modules or units mentioned in the detailed description above is not mandatory. In fact, according to the embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0135] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operational instructions of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two connected blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified functions or operational instructions, or using a combination of dedicated hardware and computer instructions.

[0136] In another aspect, this application also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments, or may exist independently and not assembled into the electronic device. The aforementioned computer-readable storage medium stores one or more programs that, when used by one or more processors, execute the deep learning algorithm-based MRI-to-CT image processing method for axial spondyloarthritis of the sacroiliac joint described in this application.

[0137] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A deep learning algorithm-based axial-type sacroiliac joint MRI-synthetic CT image processing system for ankylosing spondylitis, characterized by, The method comprises the following steps: a data acquisition unit is configured to acquire magnetic resonance sequence data obtained by performing magnetic resonance imaging on a sacroiliac joint of a target user; a deep learning model unit is configured to input the magnetic resonance sequence data into a deep learning algorithm model trained using federated learning to obtain bone cortex features and soft tissue features of the sacroiliac joint; an image synthesis unit is configured to synthesize a synthetic CT image corresponding to the sacroiliac joint of the target user based on the bone cortex features and the soft tissue features; wherein the image synthesis unit comprises an anatomical prior constraint subunit, the anatomical prior constraint subunit is configured to provide an anatomical loss function for the deep learning algorithm model during a training phase of the deep learning algorithm model, and the anatomical prior constraint comprises at least a bone cortex continuity loss constraint, a joint space distance loss constraint, and a bone erosion region adversarial loss constraint; the bone cortex continuity loss constraint adopts the following formula: ， wherein, is a bone cortex continuity loss constraint function, Dice is a similarity coefficient used to calculate the overlap between the real CT and the synthetic CT in the bone cortex, is the real CT used in training, is the synthetic CT in training, is the weight coefficient of the gradient difference loss, is a gradient operator used to extract the edge gradient of the bone cortex; the joint space distance loss constraint adopts the following formula: ， ， wherein, is a joint space distance loss constraint function, is the Euclidean distance between a and b, sup is the supremum, and inf is the infimum, is the joint space segmentation of the real CT, is the joint space segmentation of the synthetic CT; the bone erosion region adversarial loss constraint adopts the following formula: ， wherein, is a bone erosion region against loss constraint function, is a local discriminator for real-synthetic discrimination for bone erosion region (ROI), is a local image patch of bone erosion region in synthetic CT image, is a local image patch of bone erosion region in real CT image, ROI is determined by manual labeling during training process.

2. The deep learning algorithm-based axial spondyloarthritis sacroiliac joint MRI to CT synthetic image processing system of claim 1, wherein, the magnetic resonance sequence data comprises a three-dimensional T1-weighted multi-echo sequence, a T2-weighted sequence, and a water-fat separation sequence; wherein the setting parameters of the three-dimensional T1-weighted multi-echo sequence comprise TR=5.2 ms, TE=2.5 ms, and a layer thickness of 0.8 mm; and the setting parameters of the T2-weighted sequence comprise TR=3000 ms and TE=80 ms.

3. The deep learning algorithm-based axial spondyloarthritis sacroiliac joint MRI to CT synthetic image processing system of claim 2, wherein, The data acquisition unit is further configured to: identify an age correction amount of the target user, and adjust a flip angle of the three-dimensional T1-weighted multi-echo sequence based on the age correction amount.

4. The deep learning algorithm-based axial spondyloarthritis sacroiliac joint MRI to CT synthetic image processing system of claim 1, wherein, The deep learning algorithm model unit comprises an anatomical specificity attention subunit, the anatomical specificity attention subunit comprises a space-channel dual-path attention mechanism.

5. The deep learning algorithm-based axial spondyloarthritis sacroiliac joint MRI to CT synthetic image processing system of claim 4, wherein, In the space path, a deformable convolution is used to dynamically adjust sampling points, so that the convolution operation is adapted to the geometric deformation of the sacroiliac joint anatomical structure.

6. The deep learning algorithm-based axial spondyloarthritis sacroiliac joint MRI to CT synthetic image processing system of claim 1, wherein, The image synthesis unit comprises a feature alignment subunit, the feature alignment subunit is configured to make the distribution of the bone cortex features and the soft tissue features in the synthetic CT image tend to the distribution of the bone cortex features and the soft tissue features in a real CT image during the training phase of the deep learning algorithm model.

7. The deep learning algorithm-based axial spondyloarthritis sacroiliac joint MRI to CT synthetic image processing system of claim 1, wherein, The image synthesis unit is further configured to: map HU values in the synthetic CT image based on the soft tissue features.

8. The deep learning algorithm-based axial spondyloarthritis sacroiliac joint MRI to CT synthetic image processing system of claim 7, wherein, Further comprising a quantitative analysis, the quantitative analysis unit is configured to determine a bone erosion volume and a maximum depth in the synthetic CT image according to the mapped HU values in the synthetic CT image.

9. A method for processing MRI synthetic CT images of sacroiliac joints of axial type spondyloarthritis based on deep learning algorithm, characterized in that, The method comprises the following steps: acquiring magnetic resonance sequence data obtained by performing magnetic resonance imaging on a sacroiliac joint of a target user; inputting the magnetic resonance sequence data into a deep learning algorithm model trained using federated learning to obtain bone cortex features and soft tissue features of the sacroiliac joint; synthesizing a synthetic CT image corresponding to the sacroiliac joint of the target user based on the bone cortex features and the soft tissue features; In the deep learning algorithm model training phase, an anatomical loss function is provided for the deep learning algorithm model, and the anatomical prior constraint at least includes a bone cortex continuity loss constraint, a joint space distance loss constraint, and a bone erosion region adversarial loss constraint. The bone cortex continuity loss constraint adopts the following formula: ， wherein, is a bone cortex continuity loss constraint function, Dice is a similarity coefficient used to calculate the overlap between the real CT and the synthetic CT in the bone cortex, is the real CT used in training, is the synthetic CT in training, is the weight coefficient of the gradient difference loss, is a gradient operator used to extract the edge gradient of the bone cortex; The joint space distance loss constraint adopts the following formula: ， ， wherein, is a joint space distance loss constraint function, is the Euclidean distance between a and b, sup is the supremum, and inf is the infimum, is the joint space segmentation of the real CT, is the joint space segmentation of the synthetic CT; The bone erosion region adversarial loss constraint adopts the following formula: ， wherein, is a bone erosion region against loss constraint function, is a local discriminator for real-synthetic discrimination for bone erosion region (ROI), is a local image patch of bone erosion region in synthetic CT image, is a local image patch of bone erosion region in real CT image, ROI is determined by manual labeling during training process.

Images