A method and system for multi-planar feature quantification analysis of spinal facet joint images

Abstract

This invention relates to the field of medical image processing technology, specifically to a method and system for multi-planar feature quantification analysis of spinal facet joint images. The method includes the following steps: acquiring CT scan data of the target lumbar spine segment and extracting the three-dimensional regions of interest (ROIs) of both facet joints; reconstructing and extracting two-dimensional image sequences from axial, coronal, and sagittal perspectives; extracting deep spatial features from each perspective; dynamically weighting and fusion the deep spatial features to generate a global three-dimensional fusion feature vector; inputting the global three-dimensional fusion feature vector into a multi-task prediction head, simultaneously outputting joint space width scores, osteophyte formation scores, and bone quality change scores; accumulating the scores and outputting the degeneration grading result based on the total accumulated score. This invention overcomes the limitations of single-planar assessment by fusing axial, coronal, and sagittal three-dimensional image features, achieving automated and standardized quantitative assessment of spinal facet joint degeneration, and improving the objectivity and consistency of the assessment.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method and system for multiplanar feature quantification analysis of spinal facet joint images. Background Technology

[0002] In the field of spinal imaging analysis, objective and quantitative morphological assessment of the lumbar facet joints is fundamental to understanding the biomechanical properties and kinematic function of the spine. Morphological changes in the lumbar facet joints involve the comprehensive evolution of various structural features, including joint space width, osteophyte formation, and bone changes. Precise quantification of their three-dimensional structure is of great significance for spinal functional analysis.

[0003] However, most current clinical and imaging methods for small joint assessment are based on single-plane analysis, which has the following technical limitations:

[0004] First, relying on a single-plane assessment ignores the morphological features of facet joints in the coronal and sagittal planes. While axial views can better present mid-to-late-stage structural changes, coronal and sagittal views are more advantageous for identifying early joint space narrowing and bone changes. The single-plane assessment system cannot comprehensively represent three-dimensional spatial information, resulting in incomplete extraction of joint morphological features.

[0005] Second, in traditional radiographic interpretation, the determination of structural features such as joint spaces, osteophyte formation, and bone changes is highly dependent on the reader's experience. This process is not only time-consuming and laborious, but also difficult to maintain a high degree of consistency among readers with different experience levels, making it impossible to achieve large-scale, standardized extraction of quantitative parameters.

[0006] Third, there is currently no mature automated system that can extract and integrate image features from the axial, coronal, and sagittal planes to output a set of quantitative parameters that conform to the three-dimensional anatomical structure of the facet joints.

[0007] Therefore, there is an urgent need for a spinal facet joint image analysis method that can integrate multi-planar image features and achieve automated quantitative parameter extraction. Summary of the Invention

[0008] To address the technical problems in existing technologies, such as the inability of single-plane assessment to fully characterize the three-dimensional morphological features of facet joints, the strong subjectivity and poor consistency of manual image interpretation, and the lack of automated multi-plane fusion quantitative parameter extraction methods, this invention provides a multi-plane feature quantitative analysis method and system for spinal facet joint images.

[0009] In a first aspect, the present invention provides a method for multiplanar feature quantification analysis of spinal facet joint images, specifically including the following steps: Obtain CT scan data of the target lumbar vertebral segment and extract the three-dimensional regions of interest of the bilateral facet joints; Based on the three-dimensional region of interest, two-dimensional image sequences from three independent perspectives—axial, coronal, and sagittal—are reconstructed and extracted. The two-dimensional image sequence is synchronously input into a multi-branch convolutional neural network, and deep spatial features from each viewpoint are extracted by three parallel feature extractors. Dynamic weight allocation and weighted fusion are performed on the deep spatial features to generate a global three-dimensional fusion feature vector; The global three-dimensional fusion feature vector is input into the multi-task prediction head, and the joint space width score, osteophyte formation score, and bone quality change score are output simultaneously. The joint space width score, osteophyte formation score, and bone quality change score are summed, and the degeneration grading result is output based on the total sum.

[0010] Furthermore, the steps for generating the global 3D fusion feature vector include: Weights are assigned to each viewpoint using an attention mechanism, as shown in the formula:

[0011] The deep spatial features from various perspectives are weighted and fused using the following formula:

[0012] in, Indicates attention weights. This represents the transpose of the learnable weight matrix. Indicates characteristics of deep space. and Indicates the bias term. This represents the summation index variable. Indicates the axial perspective. Indicates the coronal view. Indicates the sagittal viewpoint. This represents the global 3D fusion feature vector.

[0013] Furthermore, the multi-task prediction head includes: The first subnetwork is used to output the joint space width score; The second subnetwork is used to output the osteophyte formation score; The third subnetwork is used to output a bone change score; Among them, the first subnetwork, the second subnetwork, and the third subnetwork are parallel fully connected network branches.

[0014] Furthermore, the joint space width score ranges from 0 to 3 points as a discrete level value, where: A score of 0 indicates that the joint space width is normal, while the joint space width is ≥2mm. A score of 1 indicates a slight narrowing of the joint space width, where 1 mm ≤ joint space width < 2 mm; A score of 2 indicates moderate narrowing of the joint space, with a joint space width <1mm; A score of 3 indicates severe narrowing of the joint space, disappearance of the joint space, or the presence of a vacuum.

[0015] Furthermore, the osteophyte formation score ranges from 0 to 3 points, which are discrete gradations. A score of 0 indicates that the osteophytes are normal, there are no osteophytes, and there is no facet hypertrophy. A score of 1 indicates mild osteophytes, with small, blunt, and short osteophytes and mild articular process hypertrophy. A score of 2 indicates moderate osteophytes, which are sharp and long, without forming inclusive osteophytes or bone bridges, and moderate articular process hypertrophy. A score of 3 indicates severe osteophytes, which can form bone bridges and severe articular process hypertrophy.

[0016] Furthermore, the bone change score ranges from 0 to 3 points as a discrete rank value, where: A score of 0 indicates normal bone quality, smooth joint surfaces, and no bone sclerosis or cystic changes. A score of 1 indicates mild bone abnormalities, subchondral bone sclerosis, and mild subchondral erosion or irregular areas on the joint surface. A score of 2 indicates moderate bone abnormalities, with fine cystic changes or punctate and septate sclerosis within the bone; A score of 3 indicates severe bone abnormalities, with large areas of heterogeneous sclerosis or cystic changes within the bone.

[0017] Furthermore, based on the accumulated total score, the degradation level is output as follows: a total score of 0 points is degradation level 0, 1 to 3 points is degradation level I, 4 to 6 points is degradation level II, and 7 to 9 points is degradation level III.

[0018] Furthermore, the multi-branch convolutional neural network employs a multi-task joint loss function for backpropagation optimization during the training phase:

[0019] in, This represents the joint loss function for multiple tasks. This represents the loss term for predicting the joint space width score. This represents the loss term in the osteophyte formation score prediction. This represents the loss term in the prediction of bone change scores. This represents the loss term predicted by the degradation grading results. , , and This represents the weighting coefficient of each loss term.

[0020] Secondly, the present invention provides a multi-planar feature quantification analysis system for spinal facet joint images, comprising: The data acquisition and multi-plane reconstruction module is used to acquire CT scan data of the target lumbar vertebral segment, extract the three-dimensional region of interest of the bilateral facet joints, and reconstruct two-dimensional image sequences with three independent perspectives: axial, coronal, and sagittal, based on the three-dimensional region of interest. The multi-view feature extraction and fusion module is used to synchronously input the two-dimensional image sequence into a multi-branch convolutional neural network, extract the deep spatial features of each view through three parallel feature extractors, and perform dynamic weight allocation and weighted fusion on the deep spatial features to generate a global three-dimensional fusion feature vector. The multi-task parameter generation module is used to input the global three-dimensional fusion feature vector into the multi-task prediction head, synchronously output the joint space width score, osteophyte formation score and bone quality change score, and accumulate the joint space width score, osteophyte formation score and bone quality change score, and output the degeneration grading result based on the accumulated total score.

[0021] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention extracts two-dimensional image sequences from three independent perspectives: axial, coronal, and sagittal, and uses an attention mechanism to dynamically assign weights and weighted fuse the deep spatial features of each perspective, thus overcoming the limitations of single-plane evaluation and achieving a comprehensive characterization of the three-dimensional anatomical features of small joint degeneration.

[0022] (2) This invention synchronously outputs joint space width score, osteophyte formation score and bone quality change score through multi-task prediction head, and automatically maps and outputs degeneration grading results based on the total score of the three scores, thus constructing an end-to-end automated quantitative assessment system, eliminating the subjective differences of manual image reading, and improving the consistency and reproducibility of assessment.

[0023] (3) The present invention uses a multi-task joint loss function to optimize the multi-branch convolutional neural network. By simultaneously optimizing the three vital sign prediction tasks and the hierarchical prediction task, the generalization ability and robustness of the model in the data sparse scenario are improved. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1This is a flowchart of a multiplanar feature quantification analysis method for spinal facet joint images according to the present invention.

[0026] Figure 2 This is a schematic diagram of the three signs grading standards for lumbar facet joint CT bone window axial images of the present invention. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0029] This invention provides the following technical solutions: like Figure 1 As shown, a multiplanar feature quantification analysis method for spinal facet joint images mainly includes the following steps: S1. Obtain CT scan data of the target lumbar vertebral segment and extract the three-dimensional region of interest of the bilateral facet joints.

[0030] S2. Based on the three-dimensional region of interest, reconstruct and extract two-dimensional image sequences from three independent perspectives: axial, coronal, and sagittal.

[0031] S3. The two-dimensional image sequence is synchronously input into a multi-branch convolutional neural network, and the deep spatial features of each viewpoint are extracted by three parallel feature extractors.

[0032] S4. Perform dynamic weight allocation and weighted fusion on the deep spatial features to generate a global three-dimensional fusion feature vector.

[0033] S41. Weights are assigned to each viewpoint using an attention mechanism, with the following formula:

[0034] S42. Weighted fusion of deep spatial features from various perspectives is performed using the following formula:

[0035] in, Indicates attention weights. This represents the transpose of the learnable weight matrix. Indicates characteristics of deep space. and Indicates the bias term. This represents the summation index variable. Indicates the axial perspective. Indicates the coronal view. Indicates the sagittal viewpoint. This represents the global 3D fusion feature vector.

[0036] S5. Input the global three-dimensional fusion feature vector into the multi-task prediction head and simultaneously output the joint space width score, osteophyte formation score, and bone quality change score.

[0037] In a preferred embodiment of this application, the multi-task prediction head includes: a first sub-network for outputting joint space width scores; a second sub-network for outputting osteophyte formation scores; and a third sub-network for outputting bone quality change scores; wherein the first sub-network, the second sub-network, and the third sub-network are parallel fully connected network branches.

[0038] In a preferred embodiment of this application, the joint space width score is a discrete level value ranging from 0 to 3 points, where: 0 points indicates that the joint space width is normal, and the joint space width is ≥2mm; 1 point indicates that the joint space width is slightly narrowed, and 1mm≤joint space width<2mm; 2 points indicates that the joint space width is moderately narrowed, and the joint space width is <1mm; 3 points indicates that the joint space width is severely narrowed, and the joint space disappears or a vacuum phenomenon exists.

[0039] In a preferred embodiment of this application, the osteophyte formation score is a discrete grade value ranging from 0 to 3 points, where: 0 points indicates normal osteophytes, no osteophytes, and no articular process hypertrophy; 1 point indicates mild osteophytes, with small, blunt, and short osteophytes and mild articular process hypertrophy; 2 points indicate moderate osteophytes, with moderate osteophytes that are sharp and long, do not form inclusive osteophytes or bone bridges, and moderate articular process hypertrophy; and 3 points indicate severe osteophytes, which may form bone bridges and severe articular process hypertrophy.

[0040] In a preferred embodiment of this application, the bone change score ranges from 0 to 3 points, which are discrete gradations: 0 points indicate normal bone quality, smooth joint surface, and no sclerosis or cystic changes; 1 point indicates mild bone abnormality, subchondral bone sclerosis, and mild subchondral erosion or irregular areas on the joint surface; 2 points indicate moderate bone abnormality, with fine cystic changes or patchy or septate sclerosis within the bone; and 3 points indicate severe bone abnormality, with large areas of heterogeneous sclerosis or cystic changes within the bone.

[0041] S6. The joint space width score, osteophyte formation score and bone change score are summed up, and the degeneration grading result is output based on the total sum.

[0042] In a preferred embodiment of this application, the degradation classification result is output based on the accumulated total score, including: a total score of 0 points is degradation level 0, 1 to 3 points is degradation level I, 4 to 6 points is degradation level II, and 7 to 9 points is degradation level III.

[0043] As a preferred embodiment of this application, the multi-branch convolutional neural network employs a multi-task joint loss function for backpropagation optimization during the training phase:

[0044] in, This represents the joint loss function for multiple tasks. This represents the loss term for predicting the joint space width score. This represents the loss term in the osteophyte formation score prediction. This represents the loss term in the prediction of bone change scores. This represents the loss term predicted by the degradation grading results. , , and This represents the weighting coefficient of each loss term.

[0045] This invention discloses a multi-plane feature quantification analysis system for spinal facet joint images, including a data acquisition and multi-plane reconstruction module, a multi-view feature extraction and fusion module, and a multi-task parameter generation module.

[0046] The data acquisition and multi-plane reconstruction module is used to acquire CT scan data of the target lumbar vertebral segment, extract the three-dimensional region of interest of the bilateral facet joints, and reconstruct a two-dimensional image sequence with three independent perspectives: axial, coronal, and sagittal, based on the three-dimensional region of interest.

[0047] The multi-view feature extraction and fusion module is used to synchronously input the two-dimensional image sequence into a multi-branch convolutional neural network, extract deep spatial features from each viewpoint through three parallel feature extractors, and perform dynamic weight allocation and weighted fusion on the deep spatial features to generate a global three-dimensional fusion feature vector.

[0048] The multi-task parameter generation module is used to input the global three-dimensional fusion feature vector into the multi-task prediction head, synchronously output the joint space width score, osteophyte formation score and bone quality change score, and accumulate the joint space width score, osteophyte formation score and bone quality change score, and output the degeneration grading result based on the accumulated total score.

[0049] Example This embodiment verifies the multiplanar feature quantification analysis method for spinal facet joint images proposed in this invention by following the steps below.

[0050] S1. Acquire data and extract regions of interest. CT imaging data of 226 patients who underwent lumbar spine surgery were acquired, involving a total of 1356 facet joint units, covering three segments: L3 / L4, L4 / L5, and L5 / S1. The sample age ranged from 16 to 81 years, with a mean age of 51.1 ± 16.0 years. Three-dimensional regions of interest of bilateral facet joints were extracted from the raw CT data.

[0051] S2, Reconstructing the Multiplanar Image Based on the extracted 3D region of interest, 2D image sequences from three independent perspectives—axial, coronal, and sagittal—are reconstructed and extracted.

[0052] S3. Extract deep spatial features Three independent 2D image sequences are synchronously input into a multi-branch convolutional neural network, and three parallel feature extractors extract the deep spatial features of each viewpoint.

[0053] S4. Dynamic Weight Allocation and Weighted Fusion By using an attention mechanism to assign weights to each perspective, the deep spatial features of each perspective are dynamically weighted and fused to generate a global 3D fusion feature vector.

[0054] S5, Output Multi-task Prediction The global 3D fusion feature vector is input into the multi-task prediction head, which simultaneously outputs the joint space width score, osteophyte formation score, and bone quality change score.

[0055] S6. Output degradation grading results The joint space width score, osteophyte formation score, and bone quality change score are summed up, and the degeneration grade is output based on the total sum.

[0056] like Figure 2 The diagram shows the grading criteria for three signs in the axial view of the lumbar facet joint CT bone window of this invention. In the figure, a) normal joint space (0 points), b) mild narrowing of joint space (1 point), c) moderate narrowing of joint space (2 points), d) severe narrowing of joint space (3 points), e) no osteophytes (0 points), f) mild osteophytes (1 point), g) moderate osteophytes (2 points), h) severe osteophytes (3 points), i) normal bone quality (0 points), j) mild bone changes (1 point), k) moderate bone changes (2 points), l) severe bone changes (3 points).

[0057] Experimental results show that the axial image features extracted by the method of this invention on the entire L3 to S1 segment have the most significant response to the feature response of mid-to-late stage structural changes (P<0.001). The coronal plane has the highest detection rate in feature identification of stage I degeneration (84.4%~86.8%), indicating that the coronal and sagittal planes have higher sensitivity in identifying early joint space narrowing and subtle bone changes.

[0058] Spearman rank correlation analysis was used to evaluate the correlation between the facet joint image features extracted by the method of this invention and the Pfirrmann grading and Modic changes of the intervertebral disc. Table 1 below shows the correlation results between facet joint degeneration and Pfirrmann grading, and Table 2 shows the correlation results between facet joint degeneration and Modic changes.

[0059] Table 1. Correlation results between facet joint degeneration and Pfirrmann's theory.

[0060] Table 2. Correlation results between facet joint degeneration and Modic changes

[0061] Experimental results show that the facet joint image features extracted from the axial, coronal, and sagittal planes are significantly positively correlated with the intervertebral disc Pfirrmann grade and Modic changes (P<0.001), verifying the effectiveness of the parameter extraction method of the present invention.

[0062] Further refine the segment analysis: L3 / L4 segment: Facet joint imaging features showed a moderate positive correlation with intervertebral disc degeneration indicators, with the strongest correlation between axial and sagittal views (P<0.05), demonstrating the high synchronicity between facet joint and intervertebral disc degeneration in this segment.

[0063] L4 / L5 segment: The facet joint imaging features in all three orientations were significantly positively correlated with intervertebral disc degeneration indices (P<0.001).

[0064] L5 / S1 segment: showed obvious orientation selectivity. The facet joint imaging features in axial and coronal views were not statistically correlated with the Pfirrmann grade, while only the facet joint imaging features in sagittal views were positively correlated with the Pfirrmann grade (P<0.05).

[0065] The above results verify the unique advantage of the multi-plane fusion method adopted in this invention: that is, in a specific anatomical segment, only through specific perspective features can the relationship between facet joints and spinal function be accurately revealed, and single-plane assessment cannot fully characterize the three-dimensional morphological features of facet joints.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for multiplanar feature quantification analysis of spinal facet joint images, comprising: Includes the following steps: Obtain CT scan data of the target lumbar vertebral segment and extract the three-dimensional regions of interest of the bilateral facet joints; Based on the three-dimensional region of interest, two-dimensional image sequences from three independent perspectives—axial, coronal, and sagittal—are reconstructed and extracted. The two-dimensional image sequence is synchronously input into a multi-branch convolutional neural network, and deep spatial features from each viewpoint are extracted by three parallel feature extractors. Dynamic weight allocation and weighted fusion are performed on the deep spatial features to generate a global three-dimensional fusion feature vector; The global three-dimensional fusion feature vector is input into the multi-task prediction head, and the joint space width score, osteophyte formation score, and bone quality change score are output simultaneously. The joint space width score, osteophyte formation score, and bone quality change score are summed, and the degeneration grading result is output based on the total sum.

2. A method of multiplanar feature quantification analysis of spinal facet joint images according to claim 1, wherein, The steps for generating a global 3D fusion feature vector include: Weights are assigned to each viewpoint using an attention mechanism, as shown in the formula: The deep spatial features from various perspectives are weighted and fused using the following formula: in, Indicates attention weights. This represents the transpose of the learnable weight matrix. Indicates characteristics of deep space. and Indicates the bias term. This represents the summation index variable. Indicates the axial perspective. Indicates the coronal view. Indicates the sagittal viewpoint. This represents the global 3D fusion feature vector.

3. A method for multiplanar feature quantification analysis of spinal facet joint images as claimed in claim 1, wherein, The multi-task prediction head includes: The first subnetwork is used to output the joint space width score; The second subnetwork is used to output the osteophyte formation score; The third subnetwork is used to output a bone change score; Among them, the first subnetwork, the second subnetwork, and the third subnetwork are parallel fully connected network branches.

4. A method for multiplanar feature quantification analysis of spinal facet joint images as claimed in claim 1, wherein, The joint space width score ranges from 0 to 3 points, which are discrete level values. A score of 0 indicates that the joint space width is normal, while the joint space width is ≥2mm. A score of 1 indicates a slight narrowing of the joint space width, where 1 mm ≤ joint space width < 2 mm; A score of 2 indicates moderate narrowing of the joint space, with a joint space width <1mm; A score of 3 indicates severe narrowing of the joint space, disappearance of the joint space, or the presence of a vacuum.

5. The method of claim 1 wherein the method is a method of multiplanar feature quantification analysis of the facet joints of the spine. The osteophyte formation score ranges from 0 to 3 points, which are discrete gradations. A score of 0 indicates that the osteophytes are normal, there are no osteophytes, and there is no facet hypertrophy. A score of 1 indicates mild osteophytes, with small, blunt, and short osteophytes and mild articular process hypertrophy. A score of 2 indicates moderate osteophytes, which are sharp and long, without forming inclusive osteophytes or bone bridges, and moderate articular process hypertrophy. A score of 3 indicates severe osteophytes, which can form bone bridges and severe articular process hypertrophy.

6. The method for multiplanar feature quantification analysis of spinal facet joint images according to claim 1, characterized in that, The bone change score ranges from 0 to 3 points, which are discrete gradations. A score of 0 indicates normal bone quality, smooth joint surfaces, and no bone sclerosis or cystic changes. A score of 1 indicates mild bone abnormalities, subchondral bone sclerosis, and mild subchondral erosion or irregular areas on the joint surface. A score of 2 indicates moderate bone abnormalities, with fine cystic changes or punctate and septate sclerosis within the bone; A score of 3 indicates severe bone abnormalities, with large areas of heterogeneous sclerosis or cystic changes within the bone.

7. The method for multiplanar feature quantification analysis of spinal facet joint images according to claim 1, characterized in that, The degradation level is determined based on the total accumulated score, including: 0 points for degradation level 0, 1 to 3 points for degradation level I, 4 to 6 points for degradation level II, and 7 to 9 points for degradation level III.

8. The method for multiplanar feature quantification analysis of spinal facet joint images according to claim 1, characterized in that, The multi-branch convolutional neural network employs a multi-task joint loss function for backpropagation optimization during the training phase. in, This represents the joint loss function for multiple tasks. This represents the loss term for predicting the joint space width score. This represents the loss term in the osteophyte formation score prediction. This represents the loss term in the prediction of bone change scores. This represents the loss term predicted by the degradation grading results. , , and This represents the weighting coefficient of each loss term.

9. A system for multi-planar quantification of features of an image of a facet joint of a spine, implemented based on the method of any one of claims 1 to 8, characterized in that, include: The data acquisition and multi-plane reconstruction module is used to acquire CT scan data of the target lumbar vertebral segment, extract the three-dimensional region of interest of the bilateral facet joints, and reconstruct two-dimensional image sequences with three independent perspectives: axial, coronal, and sagittal, based on the three-dimensional region of interest. The multi-view feature extraction and fusion module is used to synchronously input the two-dimensional image sequence into a multi-branch convolutional neural network, extract the deep spatial features of each view through three parallel feature extractors, and perform dynamic weight allocation and weighted fusion on the deep spatial features to generate a global three-dimensional fusion feature vector. The multi-task parameter generation module is used to input the global three-dimensional fusion feature vector into the multi-task prediction head, synchronously output the joint space width score, osteophyte formation score and bone quality change score, and accumulate the joint space width score, osteophyte formation score and bone quality change score, and output the degeneration grading result based on the accumulated total score.

Images