An automatic positioning and classification method for ossification of the posterior longitudinal ligament of the cervical spine

Abstract

The present invention belongs to the field of medical image processing technology, and specifically relates to a method for automatic positioning and classification of ossification of the posterior longitudinal ligament of the cervical spine. The method is based on the simultaneous analysis of the spatial distribution characteristics of ossification foci in CT images and the spinal cord morphological parameters in MRI images by CT and MRI images. The spatial relationship between the vertebral body, ossification and spinal cord is explicitly modeled through anatomical structure topological constraints. The method overcomes diffuse calcification artifacts, heterogeneous signal interference and multimodal data alignment deviation, achieves accurate positioning and classification diagnosis of ossification, realizes multi-dimensional feature fusion analysis, breaks through the limitations of traditional single imaging indicator diagnosis, integrates spatial continuity and clinical parameters through a hybrid classifier, improves the accuracy and robustness of classification judgment, and automatically generates a structured diagnostic report throughout the entire process, significantly improving the efficiency and standardization of clinical decision-making. The method has a scientific and reliable principle, and automatically calculates the ossification continuity index, spinal canal occupancy rate and spinal cord compression grading parameters according to the classification standard.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing technology, and specifically relates to an automatic positioning and classification method for ossification of the posterior longitudinal ligament of the cervical spine based on CT (computed tomography) and MRI (magnetic resonance imaging) images, which can accurately locate and classify ossifications. Background Art

[0002] With the aging population and the increasing severity of spinal degenerative diseases, the incidence of ossification of the posterior longitudinal ligament (OPLL) has been increasing year by year. The resulting spinal stenosis and nerve compression symptoms seriously threaten patients' quality of life. In clinical diagnosis and treatment, the combined application of CT and MRI imaging technologies provides key support for the accurate analysis of OPLL, while artificial intelligence-based medical image processing technology is gradually breaking through the efficiency bottleneck of traditional diagnostic models.

[0003] In the imaging assessment of OPLL, CT scans, with their high spatial resolution and superior bone tissue visualization, are the preferred imaging method for morphological analysis of ossified structures. CT is a medical imaging technique that uses a precise X-ray beam and highly sensitive detectors to scan the human body layer by layer. Computer processing of the scanned data generates high-resolution images of the body's interior in cross-sectional, coronal, or sagittal planes. CT images use varying grayscale to reflect the degree of X-ray absorption by organs and tissues, offering high density resolution and clear visualization of soft tissue and bone structures. Multi-planar CT observations allow clinicians to clearly identify the longitudinal extent and cross-sectional occupancy of ossified structures, thereby quantifying the degree of spinal stenosis. MRI, with its excellent soft tissue contrast, can precisely demonstrate spinal cord compression and deformation, providing a crucial basis for assessing neurological impairment. MRI utilizes the principle of nuclear magnetic resonance. Based on the differential attenuation of energy released in different structural environments within a substance, the emitted electromagnetic waves are detected by an applied gradient magnetic field, revealing the location and type of atomic nuclei that constitute the object and creating a structural image of the object. The complementary imaging characteristics of CT and MRI at the bone-soft tissue interface together constitute the imaging basis for the diagnosis and treatment of OPLL.

[0004] Current AI (artificial intelligence)-assisted diagnosis technology focuses on three core aspects: 1) First, in the field of medical image segmentation, deep learning methods construct a hierarchical segmentation model of vertebral body-ossification material-spinal cord to achieve precise separation of calcified ligaments and the posterior edge of the vertebral body in CT images, as well as spatial decoupling of the spinal cord and vertebral body in MRI images; 2) Secondly, in terms of ossification detection, based on a three-dimensional convolutional neural network, parameters such as the continuity, thickness and spinal canal occupancy rate of ossification foci in CT images are automatically extracted. At the same time, combined with the T2-weighted image signal changes of MRI, the inflammatory response stage in the ligament ossification process is evaluated; 3) Finally, at the classification diagnosis level, by fusing the ossification morphological characteristics of CT and the spinal cord compression parameters of MRI, the AI system performs intelligent classification based on the classification (focal, segmental, continuous, mixed) and spinal canal occupancy rate grading standards to assist in formulating surgical approach plans.

[0005] Existing automatic classification technologies have developed primarily along two technical paths: First, deep regression networks directly predict ossification type. However, these networks lack ossification segmentation capabilities and offer limited clinical guidance. Second, end-to-end segmentation models based on deep learning architectures simultaneously extract the spatial distribution characteristics of ossifications from CT scans and spinal cord morphological parameters from MRI, combining the spatial topological relationship between vertebrae, ossifications, and spinal cord to make classification decisions. While these models can achieve three-dimensional visualization of ossifications, they can still present issues such as blurred ossification boundaries and discrepancies in spinal cord contour recognition when encountering diffuse calcification, osteophyte interference at the vertebral margins, or metal implant artifacts. Furthermore, the heterogeneous signal characteristics of ligament ossification pose challenges to the quantitative analysis of MRI images, making them a key area of current AI algorithm optimization.

[0006] Therefore, it is necessary to break through the limitations of traditional imaging diagnosis and develop a method for automatic positioning and classification of ossification of the posterior longitudinal ligament of the cervical spine. Based on a multi-task deep learning network, the spatial distribution characteristics of ossification foci in CT images and the spinal cord morphological parameters in MRI images are synchronously analyzed. The spatial relationship between vertebral body, ossification and spinal cord is explicitly modeled through anatomical structure topological constraints, effectively overcoming diffuse calcification artifacts, heterogeneous signal interference and multimodal data registration deviation, and realizing accurate positioning and classification diagnosis of ossification. Summary of the Invention

[0007] The purpose of the present invention is to overcome the shortcomings of the existing technology and develop a method for automatically locating and classifying ossification of the posterior longitudinal ligament of the cervical spine. By constructing a deep learning architecture, the ossification map of the CT image and the spinal cord compression deformation field of the MRI image are extracted respectively, and the ossification continuity index, spinal canal occupancy rate and spinal cord compression grading parameters are automatically calculated according to the classification criteria, and an intelligent classification report is output.

[0008] In order to achieve the above-mentioned object, the present invention relates to an automatic positioning and classification method for ossification of the posterior longitudinal ligament of the cervical spine, comprising the following steps:

[0009] S1. Multimodal dataset construction

[0010] Collect the OPLL special dataset containing cervical spine CT images and MRI images, convert the CT image and MRI image data to obtain a multimodal dataset;

[0011] For CT and MRI sagittal images, 5-10 clear images adjacent to the mid-sagittal plane were selected to effectively avoid the influence of image overlap and blur on the analysis results;

[0012] For CT axial images, the focus is on extracting images of the cervical spine area. By randomly eliminating redundant images, the risk of mutual interference between images is reduced while ensuring the image sample size.

[0013] S2. Data Processing

[0014] Based on the collected multimodal dataset, a three-level anatomical topological map of the vertebral body, ossification, and spinal cord was constructed. The contours of the vertebral body and ossification at the posterior edge of the vertebral body were annotated on CT images, and the compressed and deformed areas of the vertebral body and spinal cord were simultaneously annotated on MRI images to generate annotation masks.

[0015] For sagittal CT images, vertebral annotation is completed within the cervical C2-C7 segments and thoracic T1-T2 segments. The left and right vertebral morphologies are divided into independent categories based on anatomical characteristics, and ossification is also classified separately. Ultimately, 17 orderly arranged annotation categories are formed, providing a structured data foundation for subsequent ossification location analysis.

[0016] For axial CT images, the four anatomical structural categories of spinal canal, vertebral body, lamina, and ossification are marked to effectively support the subsequent quantitative analysis of spinal stenosis rate;

[0017] For sagittal MRI images, a unified labeling strategy was adopted for the cervical C2-C7 segments and thoracic T1-T2 segments, labeling only the left vertebra as a single category, while independently labeling the spinal cord area to accurately assess the spinal cord compression status and its spatial distribution characteristics.

[0018] S3. Deep learning network architecture design

[0019] Construct a multi-task segmentation network. At the same time, perform feature restoration through a two-layer fusion module and add a multi-layer supervision loss function to enhance the spatial continuity of vertebrae, ossification, and spinal cord.

[0020] Among them, the process of building a multi-task segmentation network is:

[0021] CT and MRI images are fed into a deep neural network (Cer-Net) separately. After a convolution operation to extract low-level image features, they are fed into a U-shaped architecture encoder and decoder. The encoder uses a multi-scale residual downsampling module to extract features from medical images and adds a feature fusion module at the skip connection to enhance the representation of cervical vertebrae semantic information. The decoder uses a residual upsampling fusion module.

[0022] First, four different feature mapping heads are set based on the multi-layer feature map of the encoder. The four feature mapping heads include convolution, batch normalization, and ReLU activation functions. The convolution kernel results are output in the original size and 、 、 、 The feature map of the size is calculated and the loss is calculated with the real result;

[0023] Then, multiple layers of supervision are carried out;

[0024] Finally, through a fully connected layer, a feature map with the same number of layers as the segmentation result category is obtained, and the multi-layer results are merged into one map to obtain the segmentation result map.

[0025] The process of feature restoration in the double-layer fusion module is as follows:

[0026] In the U-shaped architecture encoder, the upper and lower layer feature fusion modules integrate multi-scale information through cross-layer connections. The output features of the ) layer are , the next layer (the ) layer) obtains features by downsampling ;

[0027] The mathematical process of fusion and skip connection is as follows:

[0028] First, high-level features are upsampled

[0029] right Perform upsampling (such as transposed convolution or bilinear interpolation) to restore its spatial resolution: ,in Represents an upsampling operation;

[0030] Then, the feature fusion mechanism

[0031] Will After upsampling By splicing and fusion:

[0032] ;

[0033] Followed by Convolution adjusts the number of channels:

[0034] ,in is the convolution kernel, is bias;

[0035] Finally, the skip connection is concatenated with the encoder

[0036] Decoder Layers get features by upsampling , and merge it with the Splicing along the channel dimension: ;

[0037] The concatenated features are processed by the convolution module of the decoder:

[0038] ,in Contains convolution, normalization and activation functions for feature refinement;

[0039] It improves the decoder's performance in object localization and boundary recovery by fusing the encoder's shallow (high-resolution details) and deep (high-level semantics) features;

[0040] The process of multi-layer supervision is:

[0041] The U-shaped architecture decoder introduces supervision signals on feature maps of different scales through a multi-layer supervision mechanism to enhance feature expression and optimize gradient propagation. Each layer output of the decoder is connected to an independent feature mapping head, and supervision is performed through the joint loss calculation of multi-scale predictions and true labels.

[0042] Feature map header structure

[0043] Each feature map head is composed of convolution layer + batch normalization + ReLU activation function. Finally, the number of channels and resolution are adjusted through convolution to output prediction results of different scales. The output feature map of the layer is , the corresponding feature map head operations include:

[0044] Convolution operation: using Convolution extracts features while maintaining spatial resolution:

[0045] ;

[0046] Batch Normalization and Activation: = ;

[0047] Resolution Adjustment Convolution: Through Convolution generates the target number of channels, and introduces transposed convolution to adjust the resolution: the output size is the original image ((s=0,1,2,3) corresponds to the original size, 1 / 2, 1 / 4, 1 / 8 respectively), with a step size of Convolution or upsampling implementation;

[0048] Multi-scale prediction and loss calculation

[0049] The feature map resolutions output by the four feature mapping heads are 、 、 、 , multi-scale supervision with true labels:

[0050] The real label is downsampled, Perform average pooling to generate multi-scale labels: = ;

[0051] Multi-scale loss function: Using CELoss and DiceLoss, jointly optimize all scales: ,in , is the loss weight of each scale (usually set to 0.5);

[0052] In the process of training deep neural networks, random flipping, random scaling, and contrast changes are used for data enhancement. The Adam optimizer is used for training, the learning rate is set to 0.001, the batch size is 8, and a total of 400 rounds of training are performed.

[0053] S4, accurate segmentation of vertebral bodies and ossified areas

[0054] A deep neural network is trained using a multimodal dataset, and a U-shaped encoder is used to extract the vertebral body and ossification morphological features from CT images and the vertebral body morphological features from MRI images.

[0055] Input any original CT image and MRI image into the trained deep neural network to obtain the predicted CT sagittal vertebral body and ossification segmentation, CT axial vertebral body and ossification segmentation, and MRI sagittal vertebral body and spinal cord segmentation results;

[0056] The specific process is:

[0057] First, multimodal data preprocessing is performed on the original CT and MRI images, including standardized spatial resolution adjustment, heterogeneous image intensity normalization, and cross-modality registration based on anatomical landmarks to ensure that the anatomical structures of different imaging sequences are aligned in space.

[0058] Then, feature extraction is performed. CT images focus on analyzing the density gradient characteristics of the bony structure. Multi-scale convolution operations are used to capture the calcification boundaries of the vertebral contour and the morphological characteristics of the ossified material. At the same time, the annular structural characteristics of the axial tomographic images are combined to extract the cross-sectional spinal canal morphology. MRI images focus on analyzing the sagittal soft tissue contrast characteristics, using the signal intensity differences between sequences to enhance the distinction between the spinal cord and vertebral boundaries.

[0059] Secondly, the input image undergoes feature encoding and enters the multi-task reasoning module. By establishing a cross-modal feature correlation map, it simultaneously generates binary masks of the vertebral bodies and ossifications in the CT sagittal plane, reconstructed contours of the vertebral bodies and ossifications in the CT axial plane, and the anatomical boundary between the vertebral bodies and the spinal cord in the MRI sagittal plane.

[0060] Thirdly, we optimized the results by adopting a cascade post-processing strategy. We eliminated small noise based on morphological operations, corrected the segmentation boundaries through edge continuity detection, and performed structural verification based on the real segmentation images of the vertebrae to ensure that each segmentation result conformed to the physiological structural characteristics of the cervical spine.

[0061] Finally, quantitative measurement parameters of each structure are generated, including clinical key indicators such as the sagittal diameter of the vertebral canal, ossification volume, and spinal cord compression index, providing a reliable anatomical basis for subsequent visualization reconstruction and surgical path planning.

[0062] S5. CT ossification location analysis

[0063] Based on the segmented CT sagittal data, the spatial distribution of ossification is located, the morphological characteristics of the ossification area are extracted, and its extension along the longitudinal axis of the spinal canal is calculated. The ossification boundary is dynamically evaluated in combination with the vertebral reference plane. For uneven calcification or artifact interference, an adaptive threshold is used to optimize boundary recognition and output the ossification core positioning parameters.

[0064] The specific process is:

[0065] Perform morphological analysis on the ossified area R, calculate the extension length, and project along the longitudinal axis of the vertebral canal (set as the y-axis): ;

[0066] Adaptive threshold boundary optimization, designing local adaptive thresholds for uneven calcification or artifact interference: ,in, is the local window grayscale mean, is the standard deviation, and k is the empirical coefficient (usually 0.5-1.5).

[0067] S6, MRI spinal cord compression analysis

[0068] In the MRI modality branch, computer vision methods are used to capture the morphological variation characteristics of the spinal cord, and the spinal cord compression deformation field is constructed in combination with the changes in cerebrospinal fluid signal intensity to calculate the spinal cord angle;

[0069] The specific process is as follows: the anterior and posterior edges of the spinal cord are automatically detected on the sagittal image, the anterior and posterior edge curves are fitted using the least squares method, and the spinal cord angle is defined as the angle between the tangent vectors of the upper and lower marking points:

[0070] ,in , The tangent vectors are taken 5 mm above and below the compression point respectively.

[0071] S7. Vertebral canal parameter measurement

[0072] CT axial images were analyzed and the ratio of the thickness of the ossification to the anteroposterior diameter of the spinal canal was calculated to obtain the spinal canal stenosis rate.

[0073] The specific process is as follows: locate the pedicle level on the axial CT image and select the narrowest cross section of the spinal canal as the measurement plane;

[0074] Maximum thickness of ossification (T): the vertical distance from the anterior edge to the posterior edge of the ossification along the sagittal direction of the spinal canal;

[0075] Anteroposterior diameter of the spinal canal (D): the shortest distance from the posterior edge of the vertebral body to the anterior edge of the lamina in the same plane;

[0076] The formula for calculating spinal canal stenosis rate is: S=(T / D)×100%;

[0077] When there are multi-segment lesions, the maximum S value of each segment is taken as the final evaluation index.

[0078] S8, Intelligent Classification Decision

[0079] Based on the classification criteria, the spatial continuity characteristics of ossification foci are analyzed through deep learning classification methods. Combined with the spinal canal stenosis rate threshold and spinal cord compression, it automatically distinguishes focal, segmental, continuous, and mixed types, and generates a comprehensive diagnostic report.

[0080] The specific process is:

[0081] First, an analysis model was established based on the distribution characteristics of ossified lesions. Sagittal and axial images of the cervical spine were reconstructed using a deep learning network. The spatial relationship and continuity of the ossified lesions within the vertebral sequence were extracted. A convolutional neural network was used to identify the longitudinal extension of the ossified lesions and accurately divide the number of vertebral segments involved. Furthermore, the distance between adjacent ossified lesions was measured, laying the foundation for spatial analysis for subsequent classification.

[0082] Then, a dynamic assessment module for spinal canal morphological parameters was constructed. Based on the stenosis threshold standards set by clinical guidelines, a dynamic assessment model for the effective volume of the spinal canal was established. This model focused on monitoring changes in the degree of stenosis in key segments, established a correlation mapping with spinal cord morphological changes, and introduced an evaluation mechanism for the state of spinal cord compression, forming a compression assessment system from macromorphology to microstructure.

[0083] Finally, an intelligent decision-making system with multi-dimensional feature fusion is established. By constructing a hierarchical classification model, the spatial continuity characteristics, the degree of spinal canal stenosis and the spinal cord compression parameters are fused to accurately distinguish focal, segmental, continuous and mixed types. The diagnostic report generation module adopts a structured output framework to integrate key imaging feature parameters, classification basis and clinical recommendations to form standardized diagnostic documents that comply with diagnosis and treatment standards.

[0084] Compared with the existing technology, the present invention synchronously analyzes the spatial distribution characteristics of ossification foci in CT images and the spinal cord morphological parameters in MRI images based on a multi-task deep learning network, explicitly models the spatial relationship between vertebrae, ossifications and spinal cord through anatomical structure topological constraints, overcomes diffuse calcification artifacts, heterogeneous signal interference and multimodal data alignment deviation, realizes precise positioning and typing diagnosis of ossifications, realizes multi-dimensional feature fusion analysis, breaks through the limitations of traditional single imaging indicator diagnosis, integrates spatial continuity and clinical parameters through a hybrid classifier, improves the accuracy and robustness of typing judgment, and automatically generates structured diagnostic reports throughout the entire process, significantly improving clinical decision-making efficiency and standardization level; its principle is scientific and reliable. By constructing a deep learning architecture, the ossification map of the CT image and the spinal cord compression deformation field of the MRI image are extracted respectively, and the ossification continuity index, spinal canal occupancy rate and spinal cord compression grading parameters are automatically calculated according to the typing standard, and an intelligent typing report that meets clinical guidelines is output. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 It is a process flow chart of the present invention.

[0086] Figure 2 This is a schematic diagram of segmentation information of the vertebral body, ossification and spinal cord involved in the present invention.

[0087] Figure 3 This is an example diagram of the CT ossification position analysis involved in the present invention, wherein a is the ossification of the cervical vertebra C2-6 segment, b is the ossification of the cervical vertebra C2-3 and C4-6 segments, and c is the ossification of the cervical vertebra C5-6 intervertebral disc.

[0088] Figure 4 This is an example diagram of the MRI spinal cord compression analysis involved in the present invention, wherein a is a spinal cord angle of 102°, b is a spinal cord angle of 97°, and c is a spinal cord angle of 113°.

[0089] Figure 5This is an example diagram of the spinal canal stenosis rate measurement involved in the present invention, wherein a represents a spinal canal stenosis rate of 36.6%, b represents a spinal canal stenosis rate of 41.0%, and c represents a spinal canal stenosis rate of 45.1%.

[0090] Figure 6 This is a schematic diagram of the diagnostic results involved in the present invention. DETAILED DESCRIPTION

[0091] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0092] Example 1:

[0093] This embodiment relates to a method for automatically locating and classifying ossification of the posterior longitudinal ligament of the cervical spine, comprising the following steps:

[0094] S1. Multimodal dataset construction

[0095] Collect cervical spine CT and MRI images, analyze and screen the original DICOM format;

[0096] Among the sagittal images, select 5-10 images near the mid-sagittal position and export them to nii format;

[0097] In the axial images, 30-35 images from the C2-C7 interval were randomly selected and exported to nii format for subsequent annotation using ITK-SNAP software;

[0098] S2. Data Processing

[0099] 700 cervical spine CT sagittal images, 2600 cervical spine CT axial images, and 600 cervical spine MRI sagittal images provided by the hospital's spine surgery department were annotated. The vertebral position, ossification, and clarity of each cervical spine image were different, such as Figure 2 As shown, each cervical spine CT sagittal image is marked with 17 annotations, each cervical spine CT axial image is marked with 4 annotations, and each cervical spine MRI sagittal image is marked with 2 annotations;

[0100] S3. Deep learning network architecture design

[0101] In the U-shaped architecture encoder, the upper and lower layers of feature fusion integrate multi-scale information through cross-layer connections. The output features of the ) layer are , the next layer (the ) layer) obtains features by downsampling ;

[0102] S31, high-level feature upsampling

[0103] right Upsample to restore its spatial resolution: ;

[0104] S32. Feature fusion mechanism

[0105] Will After upsampling By splicing and fusion: , followed by Convolution adjusts the number of channels: ;

[0106] S33, skip connection and encoder splicing

[0107] Decoder Layers get features by upsampling , and merge it with the Splicing along the channel dimension: , the concatenated features are processed by the decoder convolution module: ;

[0108] S34, feature map header structure

[0109] Adjust the number of channels and resolution through convolution, output prediction results of different scales, and set the decoder The output feature map of the layer is , the corresponding feature map head operations include:

[0110] Convolution operation: using Convolution extracts features while maintaining spatial resolution:

[0111] ;

[0112] Batch Normalization and Activation: = ;

[0113] Resolution Adjustment Convolution: Through Convolution generates the target number of channels, and introduces transposed convolution (TransposedConv) to adjust the resolution: the output size is the original image ((s=0,1,2,3) corresponds to the original size, 1 / 2, 1 / 4, 1 / 8 respectively), with a step size of Convolution or upsampling implementation;

[0114] S35. Multi-scale prediction and loss calculation

[0115] The feature map resolutions output by the four feature mapping heads are 、 、 、 , multi-scale supervision with true labels:

[0116] The real label is downsampled, Perform average pooling to generate multi-scale labels: = ;

[0117] Multi-scale loss function: Using CELoss and DiceLoss, jointly optimize all scales: ;

[0118] In the process of training deep neural networks, random flipping, random scaling, and contrast changes are used for data enhancement. The Adam optimizer is used for training, the learning rate is set to 0.001, the batch size is 8, and a total of 400 rounds of training are performed.

[0119] S4, accurate segmentation of vertebral bodies and ossified areas

[0120] First, the original CT and MRI images were standardized in spatial resolution, normalized in intensity, and cross-modality registration based on anatomical landmarks.

[0121] Then, the density gradient features of the bony structure of the CT image are extracted, and the morphological features of the calcification boundary and ossification of the vertebral body contour are captured through multi-scale convolution operation. At the same time, the cross-sectional spinal canal morphology is extracted by combining the annular structure characteristics of the axial tomographic image.

[0122] Extract the soft tissue contrast features of sagittal MRI images and use the signal intensity difference between sequences to enhance the distinction between the spinal cord and vertebral boundaries;

[0123] Secondly, the input image undergoes feature encoding and enters the multi-task reasoning module. By establishing a cross-modal feature correlation map, it simultaneously generates binary masks of the vertebral bodies and ossifications in the CT sagittal plane, reconstructed contours of the vertebral bodies and ossifications in the CT axial plane, and the anatomical boundary between the vertebral bodies and the spinal cord in the MRI sagittal plane.

[0124] Thirdly, to optimize the results, a cascade post-processing strategy was adopted. Based on morphological operations, small noises were eliminated, segmentation boundaries were corrected through edge continuity detection, and structural verification was performed by combining the real segmentation images of the vertebrae.

[0125] Finally, key clinical indicators including sagittal diameter of the spinal canal, ossification volume and spinal cord compression index were generated.

[0126] S5. CT ossification location analysis

[0127] The morphological analysis of the ossification region R was performed, such as Figure 3 As shown, the extension length is calculated and projected along the longitudinal axis of the vertebral canal (set as the y-axis): ;

[0128] Adaptive threshold boundary optimization, designing local adaptive thresholds for uneven calcification or artifact interference: .

[0129] S6, MRI spinal cord compression analysis

[0130] The anterior and posterior edges of the spinal cord are automatically detected on the sagittal image, and the anterior and posterior edge curves are fitted using the least squares method, as shown in Figure 2. Figure 4 As shown in the figure, the spinal cord angle is defined as the angle between the tangent vectors of the upper and lower markers:

[0131] .

[0132] S7. Vertebral canal parameter measurement

[0133] Locate the pedicle level on the axial CT image and select the narrowest cross section of the spinal canal as the measurement plane, such as Figure 5 shown.

[0134] S8, Intelligent Classification Decision

[0135] First, an analysis model was established based on the distribution characteristics of ossified lesions. Sagittal and axial images of the cervical spine were reconstructed using a deep learning network. The spatial positional relationship and continuity of the ossified lesions within the vertebral sequence were extracted. A convolutional neural network was then used to identify the longitudinal extension of the ossified lesions, accurately dividing the number of vertebral segments involved in the lesions. Furthermore, the distance between adjacent ossified lesions was detected.

[0136] Then, a dynamic assessment module for spinal canal morphological parameters was constructed. Based on the stenosis threshold standards set by clinical guidelines, a dynamic assessment model for the effective volume of the spinal canal was established. This model focused on monitoring changes in the degree of stenosis in key segments, established a correlation mapping with spinal cord morphological changes, and introduced an evaluation mechanism for the state of spinal cord compression, forming a compression assessment system from macromorphology to microstructure.

[0137] Finally, an intelligent decision-making system with multi-dimensional feature fusion is established. By constructing a hierarchical classification model, spatial continuity features, spinal canal stenosis degree and spinal cord compression parameters are integrated to accurately distinguish focal, segmental, continuous and mixed types. The diagnostic report generation module adopts a structured output framework to integrate key imaging feature parameters, classification basis and clinical recommendations to form standardized diagnostic documents, such as Figure 6 shown.

Claims

1. A method for automatically locating and classifying ossification of the posterior longitudinal ligament of the cervical spine, characterized in that: The following steps are involved: S1. Multimodal dataset construction Collect cervical spine CT images and MRI images and convert them to obtain a multimodal dataset; S2. Data Processing Construct a topological map of the three-level anatomical structure of vertebral body, ossification, and spinal cord, annotate the contours of the vertebral body and ossification at the posterior edge of the vertebral body on CT images, and simultaneously annotate the compressed and deformed areas of the vertebral body and spinal cord on MRI images to generate an annotation mask. S3. Deep learning network architecture design Construct a multi-task segmentation network. At the same time, perform feature restoration through a two-layer fusion module and add a multi-layer supervision loss function. S4, accurate segmentation of vertebral bodies and ossified areas A deep neural network is trained using a multimodal dataset, and a U-shaped encoder is used to extract the vertebral body and ossification morphological features from CT images and the vertebral body morphological features from MRI images. Input any original CT image and MRI image into the trained deep neural network to obtain the predicted CT sagittal vertebral body and ossification segmentation, CT axial vertebral body and ossification segmentation, and MRI sagittal vertebral body and spinal cord segmentation results; S5. CT ossification location analysis Based on the segmented CT sagittal data, the spatial distribution of ossification is located, the morphological characteristics of the ossification area are extracted, and its extension along the longitudinal axis of the spinal canal is calculated. The ossification boundary is dynamically evaluated in combination with the vertebral reference plane. For uneven calcification or artifact interference, an adaptive threshold is used to optimize boundary recognition and output the ossification core positioning parameters. S6, MRI spinal cord compression analysis In the MRI modality branch, computer vision methods are used to capture the morphological variation characteristics of the spinal cord, and the spinal cord compression deformation field is constructed in combination with the changes in cerebrospinal fluid signal intensity to calculate the spinal cord angle; S7. Vertebral canal parameter measurement CT axial images were analyzed and the ratio of the thickness of the ossification to the anteroposterior diameter of the spinal canal was calculated to obtain the spinal canal stenosis rate. S8, Intelligent Classification Decision Based on the classification criteria, the spatial continuity characteristics of ossification foci are analyzed through deep learning classification methods. Combined with the spinal canal stenosis rate threshold and spinal cord compression, automatic identification is performed and a comprehensive diagnostic report is generated. Among them, the process of building a multi-task segmentation network is: CT images and MRI images are fed into a deep neural network respectively. After a convolution operation to extract low-level image features, they are fed into a U-shaped architecture encoder and decoder. The encoder uses a multi-scale residual downsampling module to extract features from medical images, and adds a feature fusion module at the jump connection. First, four different feature mapping heads are set based on the multi-layer feature map of the encoder. The four feature mapping heads include convolution, batch normalization, and ReLU activation functions. The convolution kernel results are output in the original size and 、 、 、 The feature map of the size is calculated and the loss is calculated with the real result; Then, multiple layers of supervision are carried out; Finally, through a fully connected layer, a feature map with the same number of layers as the segmentation result category is obtained, and the multi-layer results are merged into one image to obtain the segmentation result image; The process of feature restoration in the double-layer fusion module is as follows: In the U-shaped architecture encoder, the upper and lower layer feature fusion modules integrate multi-scale information through cross-layer connections. The output features of the ) layer are , the next layer (the ) layer) obtains features by downsampling ; The mathematical process of fusion and skip connection is as follows: First, high-level features are upsampled right Upsample to restore spatial resolution: ,in Represents an upsampling operation; Then, the feature fusion mechanism Will After upsampling By splicing and fusion: ； Followed by Convolution adjusts the number of channels: ,in is the convolution kernel, is bias; Finally, the skip connection is concatenated with the encoder Decoder Layers get features by upsampling , and merge it with the Splicing along the channel dimension: ; The concatenated features are processed by the convolution module of the decoder: ,in Contains convolution, normalization and activation functions for feature refinement; The process of multi-layer supervision is: The U-shaped architecture decoder introduces supervision signals on feature maps of different scales through a multi-layer supervision mechanism to enhance feature expression and optimize gradient propagation. Each layer output of the decoder is connected to an independent feature mapping head, and supervision is performed through the joint loss calculation of multi-scale predictions and true labels. Feature map header structure Each feature map head is composed of convolution layer + batch normalization + ReLU activation function. Finally, the number of channels and resolution are adjusted through convolution to output prediction results of different scales. The output feature map of the layer is , the corresponding feature map head operations include: Convolution operation: using Convolution extracts features while maintaining spatial resolution: ； Batch Normalization and Activation: = ; Resolution Adjustment Convolution: Through Convolution generates the target number of channels and introduces transposed convolution to adjust the resolution: the output size is the original image ((s=0,1,2,3) corresponds to the original size, 1 / 2, 1 / 4, 1 / 8 respectively), with a step size of Convolution or upsampling implementation; Multi-scale prediction and loss calculation The feature map resolutions output by the four feature mapping heads are 、 、 、 , multi-scale supervision with true labels: The real label is downsampled, Perform average pooling to generate multi-scale labels: = ; Multi-scale loss function: Using CELoss and DiceLoss, jointly optimize all scales: ,in , is the loss weight of each scale.

2. The method for automatic positioning and classification of ossification of the posterior longitudinal ligament of the cervical spine according to claim 1, characterized in that: In S1, for CT and MRI sagittal images, 5–10 clear images adjacent to the midsagittal plane were selected, and for CT axial images, images of the cervical spine region were randomly extracted.

3. The method for automatic positioning and classification of ossification of the posterior longitudinal ligament of the cervical spine according to claim 1, characterized in that: In S2, for sagittal CT images, vertebral annotation is completed within the C2-C7 segment of the cervical spine and the T1-T2 segment of the thoracic spine, and the left and right vertebral morphologies are divided into independent categories, and the ossification is classified separately, forming 17 annotation categories; For CT axial images, the four anatomical structure categories of spinal canal, vertebral body, lamina, and ossification were marked; For sagittal MRI images, the left vertebral body of the cervical C2-C7 segments and the thoracic T1-T2 segments were annotated as a single category, and the spinal cord region was annotated independently.

4. The method for automatic location and classification of ossification of the posterior longitudinal ligament of the cervical spine according to claim 1, characterized in that: The specific process of S4 is: First, the original CT and MRI images were subjected to multimodal data preprocessing, including standardized spatial resolution adjustment, heterogeneous image intensity normalization, and cross-modality registration based on anatomical landmarks. Then, feature extraction is performed. CT images analyze the density gradient characteristics of the bony structure, and multi-scale convolution operations are used to capture the calcification boundaries of the vertebral contour and the morphological characteristics of the ossification. At the same time, the annular structure characteristics of the axial tomographic images are combined to extract the cross-sectional spinal canal morphology. MRI images analyze the sagittal soft tissue contrast characteristics, and the signal intensity difference between sequences is used to enhance the distinction between the spinal cord and vertebral boundaries. Secondly, the input image undergoes feature encoding and enters the multi-task reasoning module. By establishing a cross-modal feature correlation map, it simultaneously generates binary masks of the vertebral bodies and ossifications in the CT sagittal plane, reconstructed contours of the vertebral bodies and ossifications in the CT axial plane, and the anatomical boundary between the vertebral bodies and the spinal cord in the MRI sagittal plane. Thirdly, to optimize the results, a cascade post-processing strategy was adopted. Based on morphological operations, small noises were eliminated, segmentation boundaries were corrected through edge continuity detection, and structural verification was performed by combining the real segmentation images of the vertebrae. Finally, quantitative measurement parameters of each structure are generated.

5. The method for automatic location and classification of ossification of the posterior longitudinal ligament of the cervical spine according to claim 1, characterized in that: The specific process of S5 is: Perform morphological analysis on the ossified area R, calculate its extension length, and project along the longitudinal axis of the vertebral canal: ; Adaptive threshold boundary optimization, designing local adaptive thresholds for uneven calcification or artifact interference: ,in, is the local window grayscale mean, is the standard deviation, and k is the empirical coefficient.

6. The method for automatic positioning and classification of ossification of the posterior longitudinal ligament of the cervical spine according to claim 1, characterized in that: The specific process of S6 is: The anterior and posterior edges of the spinal cord were automatically detected on the sagittal image. The anterior and posterior edge curves were fitted using the least squares method. The spinal cord angle was defined as the angle between the tangent vectors of the upper and lower markers: ,in , The tangent vectors are taken 5 mm above and below the compression point respectively.

7. The method for automatic location and classification of ossification of the posterior longitudinal ligament of the cervical spine according to claim 1, characterized in that: In S7, the pedicle level was located on the axial CT image, and the narrowest cross section of the spinal canal was selected as the measurement plane; Maximum thickness of ossification (T): the vertical distance from the anterior edge to the posterior edge of the ossification along the sagittal direction of the spinal canal; Anteroposterior diameter of the spinal canal (D): the shortest distance from the posterior edge of the vertebral body to the anterior edge of the lamina in the same plane; The formula for calculating spinal canal stenosis rate is: S=(T / D)×100%; When there are multi-segment lesions, the maximum S value of each segment is taken as the final evaluation index.

8. The method for automatic location and classification of ossification of the posterior longitudinal ligament of the cervical spine according to claim 1, characterized in that: The specific process of S8 is as follows: First, an analysis model was established based on the distribution characteristics of ossified lesions. Sagittal and axial images of the cervical spine were reconstructed using a deep learning network. The spatial positional relationship and continuity of the ossified lesions within the vertebral sequence were extracted. A convolutional neural network was then used to identify the longitudinal extension of the ossified lesions, accurately dividing the number of vertebral segments involved in the lesions. Furthermore, the distance between adjacent ossified lesions was detected. Then, a dynamic assessment module for spinal canal morphological parameters was constructed. Based on the stenosis threshold standards set by clinical guidelines, a dynamic assessment model for the effective volume of the spinal canal was established. This model monitored changes in the degree of stenosis in key segments and correlated this with changes in spinal cord morphology. An evaluation mechanism for the state of spinal cord compression was introduced, forming a compression assessment system from macromorphology to microstructure. Finally, an intelligent decision-making system with multi-dimensional feature fusion was established. By constructing a hierarchical classification model, the spatial continuity features, the degree of spinal canal stenosis and the spinal cord compression parameters were fused to distinguish focal, segmental, continuous and mixed types. The diagnostic report generation module adopted a structured output framework to integrate key imaging feature parameters, classification basis and clinical recommendations to form a standardized diagnostic document.

9. The method for automatic positioning and classification of ossification of the posterior longitudinal ligament of the cervical spine according to any one of claims 1 to 8, characterized in that: In the process of training deep neural networks, random flipping, random scaling, and contrast changes are used for data enhancement. The Adam optimizer is used for training, the learning rate is set to 0.001, the batch size is 8, and a total of 400 rounds of training are performed.

Images