Sacroiliac joint X-ray image intelligent grading method and system

Abstract

The invention provides a sacroiliac joint X-ray image intelligent grading method comprising the following steps: carrying out CT value remapping on a DR image through an adaptive filtering core to obtain an artifact-removed image; segmenting the sacrum and the ilium through a neural network model to obtain sacroiliac joints; performing affine transformation on the artifact-removed image to obtain a standardized image; performing three-dimensional reconstruction on the standardized image to obtain a reconstructed image, and extracting an articular cavity depth coefficient; performing multi-scale feature fusion on the reconstructed image to obtain multi-scale features; based on the reconstructed image, performing time domain evolution through a space-time double-flow network to obtain a bone erosion area change rate; and performing logistic regression fusion on the articular cavity depth coefficient, the bone erosion area change rate and the multi-scale features to obtain a grading result. According to the method, through adaptive artifact filtering, multi-scale spatial-temporal feature fusion and pathological region weight optimization, the problems of structural offset and dynamic feature deficiency caused by noise interference are solved, and the recognition precision and efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image recognition, in particular to a sacroiliac joint X-ray image intelligent grading method and system. BACKGROUND

[0002] Sacroiliac arthritis is the core diagnostic indicator of ankylosing spondylitis and other diseases, and its grading directly affects the development of treatment plans. Currently, doctors rely on visual assessment of features such as joint space bone erosion in X-ray images. However, manual reading is highly subjective and inefficient, and it is difficult to quantify dynamic changes. In traditional computer-aided processes, device electromagnetic noise in DR images can cause bone texture details to be lost during the filtering process, and the anatomical structure of the sacroiliac joint is complex, so traditional threshold segmentation is easily disturbed by osteoporosis, resulting in inaccurate image segmentation. Moreover, existing methods only analyze static images, ignoring the time-domain evolution of bone erosion. Therefore, it is necessary to design a sacroiliac joint X-ray image intelligent grading method and system. SUMMARY

[0003] The purpose of the present application is to provide a sacroiliac joint X-ray image intelligent grading method and system to solve the problems of structural deviation and dynamic feature loss caused by noise interference through adaptive artifact filtering, multi-scale spatiotemporal feature fusion, and pathological region weight optimization, and to improve recognition accuracy and efficiency.

[0004] To achieve the above purpose, the present application provides the following scheme:

[0005] A sacroiliac joint X-ray image intelligent grading method, comprising the following steps:

[0006] Generate an adaptive filter kernel based on the electromagnetic noise spectrum, and perform CT value remapping on the DR image through the adaptive filter kernel to obtain a de-artifact image;

[0007] Segment the sacrum and ilium in the de-artifact image through a neural network model to obtain a sacroiliac joint;

[0008] Perform a standardization operation on the de-artifact image in the IS coordinate system through an affine transformation to obtain a standardized image;

[0009] Perform three-dimensional reconstruction of the coronal and sagittal planes of the standardized image to obtain a reconstructed image, and extract the joint cavity depth coefficient of the sacroiliac joint;

[0010] Perform multi-scale feature fusion on the reconstructed image through a multi-scale image pyramid to obtain multi-scale features;

[0011] Based on the reconstructed image, perform time-domain evolution through a pre-constructed spatiotemporal dual-stream network to obtain a bone erosion area change rate;

[0012] The joint cavity depth coefficient, the bone erosion area change rate and the multi-scale features are fused by a logistic regression to obtain a grading result.

[0013] Optionally, an adaptive filtering kernel is generated based on the electromagnetic noise spectrum, and CT value remapping is performed on the DR image through the adaptive filtering kernel to obtain a deartifact image, including:

[0014] The DR image is subjected to wavelet packet decomposition, and high-frequency sub-band components are extracted, and the device electromagnetic noise and anatomical structure signals in the DR image are separated through a blind source separation algorithm to generate a spectral feature map;

[0015] A mapping relationship between the device electromagnetic noise and the spatial domain is established according to the spectral feature map, the mapping relationship is associated with the spatial position features of the device electromagnetic noise in the DR image, and a space-frequency association relationship is obtained;

[0016] An adaptive filtering kernel is constructed according to the spectral feature map and the space-frequency association relationship to perform noise filtering on the DR image to obtain a filtered image; the adaptive filtering kernel is a rotatable elliptical kernel, and the long axis direction of the adaptive filtering kernel is orthogonal to the artifact texture direction;

[0017] A segmented remapping function is constructed, and the local variance entropy of the filtered image is taken as an optimization objective to perform iterative optimization to obtain a deartifact image.

[0018] Optionally, the sacrum and the ilium in the deartifact image are segmented through a neural network model to obtain a sacroiliac joint, including:

[0019] Whole-image bone texture features in the deartifact image are extracted through a depth separable convolution;

[0020] Based on the whole-image bone texture features, segmentation masks of the sacrum and the ilium are respectively generated through an improved U-Net network;

[0021] The receptive field shape of the improved U-Net network is adjusted according to the HU value in the deartifact image to perform attention extraction on the segmentation masks;

[0022] The physical edges of the HU value mutation area in the deartifact image are detected through a Canny operator;

[0023] Based on the whole-image bone texture features, the texture edges are obtained by capturing the trabecular bone orientation through a Gabor filter;

[0024] Based on the whole-image bone texture features, the topological edges are obtained by repairing the broken edges of the bone through a graph convolution network;

[0025] The Hausdorff distance between the segmentation masks is calculated;

[0026] The sacroiliac joint is identified by the discriminator based on the value of the Hausdorff distance.

[0027] Optionally, the artifact-removed image is standardized in the IS coordinate system through an affine transformation to obtain a standardized image, including:

[0028] The key anatomical structures are located through a key point detection network; the key anatomical structures include: a center point of the upper edge of the sacrum, a center point of the left iliac articular surface, and a center point of the right iliac articular surface.

[0029] An affine transformation matrix of the key anatomical structures is calculated through a least square method.

[0030] Based on the affine transformation matrix, the artifact-removed image is resampled through bilinear interpolation to align the artifact-removed image to the IS coordinate system.

[0031] The positional deviation of the aligned key anatomical structures is calculated, and if the deviation is greater than a preset deviation threshold, the artifact-removed image is aligned again through an ICP algorithm to obtain the standardized image.

[0032] Optionally, the standardized image is three-dimensionally reconstructed in the coronal plane and the sagittal plane to obtain a reconstructed image, and a joint cavity depth coefficient of the sacroiliac joint is extracted, including:

[0033] The slices are stacked layer by layer along the Y axis of the IS coordinate system to generate an interarticular space panoramic view.

[0034] A geometric constraint from the coronal plane to the sagittal plane is constructed according to the tangent direction of the iliac articular surface of the coronal plane.

[0035] Based on the geometric constraint, a joint cavity profile sequence is generated along the X axis of the IS coordinate system.

[0036] The joint cavity profile sequence is anisotropically interpolated through a Bézier curve driven voxel interpolation algorithm to obtain the reconstructed image.

[0037] Regions with a HU value less than -350 and a gradient amplitude greater than 100 in the reconstructed image are marked as candidate erosion regions.

[0038] An erosion region depth value of the candidate erosion region is taken as an erosion influence factor to construct a joint cavity depth coefficient; the joint cavity depth coefficient includes: an average depth, a minimum depth, a depth variation, and the erosion influence factor.

[0039] Optionally, multi-scale features are obtained by fusing multi-scale image pyramids of the reconstructed image, including:

[0040] constructing a multi-scale image pyramid; the multi-scale image pyramid is a 5-level pyramid, and the feature sizes of each level of the pyramid are 256*256, 128*128, 64*64, 32*32 and 16*16 respectively;

[0041] calculating a local gray co-occurrence matrix of the reconstructed image in each level of the multi-scale image pyramid;

[0042] generating a distance transform map in each level of the multi-scale image pyramid based on the segmentation mask;

[0043] performing channel subtraction on the local gray co-occurrence matrix and the distance transform map that have been normalized respectively to obtain multi-scale features.

[0044] Optionally, based on the reconstructed image, temporal evolution is performed through a pre-constructed spatio-temporal dual-flow network to obtain a bone erosion area change rate, including:

[0045] extracting static pathological features of the reconstructed image through a ResNet-34 network;

[0046] extracting a bone erosion heat map of the static pathological features through an LSTM network;

[0047] establishing a reference coordinate system with the sacral S1 pedicle as a dynamic origin, the sacral midline direction as the Z-axis and the plane parallel to the sacrum as the X-plane;

[0048] calculating a sagittal plane projection difference of the bone erosion heat map in the reference coordinate system;

[0049] area-dividing the sagittal plane projection difference to obtain the bone erosion area change rate.

[0050] Optionally, the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale features are fused through logistic regression to obtain a grading result, including:

[0051] constructing a five-dimensional learnable matrix according to the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale features;

[0052] generating a feature modulation vector according to a clinical test index;

[0053] marking a region with a HU value less than -300 and a Hausdorff distance less than 5mm in the reconstructed image as a pathological region, and marking the remaining regions as non-pathological regions;

[0054] performing 3 times weight enhancement on the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale features in the pathological region, and performing 0.3 times weight attenuation on the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale features in the non-pathological region to obtain an optimized matrix;

[0055] The optimized matrix is fused with the characteristic modulation vector by logistic regression to obtain a pathological index.

[0056] The pathological index is graded according to its size to obtain a grading result.

[0057] A sacroiliac joint X-ray image intelligent grading system comprises:

[0058] An image mapping module is configured to generate an adaptive filter kernel based on an electromagnetic noise spectrum, and perform CT value remapping on a DR image through the adaptive filter kernel to obtain a deartifact image.

[0059] A joint recognition module is configured to segment a sacrum and an ilium in the deartifact image through a neural network model to obtain a sacroiliac joint.

[0060] A coordinate conversion module is configured to perform a standardization operation on the deartifact image in an IS coordinate system through an affine transformation to obtain a standardized image.

[0061] An image reconstruction module is configured to perform three-dimensional reconstruction of a coronal plane and a sagittal plane on the standardized image to obtain a reconstructed image, and extract a joint cavity depth coefficient of the sacroiliac joint.

[0062] A feature extraction module is configured to perform multi-scale feature fusion on the reconstructed image through a multi-scale image pyramid to obtain multi-scale features.

[0063] An erosion prediction module is configured to perform time-domain evolution on the reconstructed image through a pre-constructed spatiotemporal dual-flow network to obtain a bone erosion area change rate.

[0064] An intelligent grading module is configured to fuse the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale features by logistic regression to obtain a grading result.

[0065] According to the specific embodiments of the present application, the following technical effects are disclosed: the sacroiliac joint X-ray image intelligent grading method provided by the present application, the method comprises: generating an adaptive filter kernel based on an electromagnetic noise spectrum, and performing CT value remapping on the DR image through the adaptive filter kernel to obtain a deartifact image; segmenting the sacrum and ilium in the deartifact image through a neural network model to obtain a sacroiliac joint; performing a standardization operation on the deartifact image in the IS coordinate system through affine transformation to obtain a standardized image; performing three-dimensional reconstruction of the coronal and sagittal planes of the standardized image to obtain a reconstructed image, and extracting a joint cavity depth coefficient of the sacroiliac joint; performing multi-scale feature fusion on the reconstructed image through a multi-scale image pyramid to obtain a multi-scale feature; based on the reconstructed image, performing time domain evolution through a pre-constructed spatiotemporal double-flow network to obtain a bone erosion area change rate; and performing logic regression fusion on the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale feature to obtain a grading result. Through adaptive artifact filtering, multi-scale spatiotemporal feature fusion and pathological region weight optimization, the method solves the problems of structure deviation and dynamic feature loss caused by noise interference, and improves the recognition accuracy and efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0067] Figure 1 The sacroiliac joint X-ray image intelligent grading method of the present application is shown in the flowchart.

[0068] Figure 2 The DR image remapping flowchart of the present application is shown in the flowchart. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0070] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0071] As shown in the drawings, Figure 1 The present application provides a sacroiliac joint X-ray image intelligent grading method, comprising the following steps:

[0072] Step 100: generating an adaptive filter kernel based on the electromagnetic noise spectrum, and performing CT value remapping on the DR image through the adaptive filter kernel to obtain a deartifact image; the specific steps are as shown in Figure 2

[0073] Step 101: performing wavelet packet decomposition on the DR image, extracting a high-frequency subband component, and separating the device electromagnetic noise and the anatomical structure signal in the DR image through a blind source separation algorithm to generate a spectrum feature map;

[0074] Specifically, first, a wavelet packet decomposition operation is performed on the input DR image, a Daubechies-4 wavelet basis function is used for multi-scale decomposition, and the image is divided into a high-frequency subband component and a low-frequency subband component; wherein the high-frequency subband component focuses on the high-frequency noise component in the image, and a subband signal in the 1.5-3.0 MHz frequency band is extracted through three decomposition levels, so as to retain the mixed features of the bone edge and the noise, while eliminating the low-frequency anatomical structure interference. Then, based on the high-frequency subband component, a fast independent component analysis algorithm is used for blind source separation: first, two source signals of the device electromagnetic noise signal and the anatomical structure signal are initialized, and the separation matrix is iteratively optimized through the negative entropy maximization criterion until the signal converges to non-Gaussian; after separation, the noise signal is subjected to short-time Fourier transform to generate a spectrum feature map, and the spectrum feature map has feature dimensions including noise power spectral density and main frequency distribution, so as to quantify the amplitude-phase characteristics of the electromagnetic noise in the frequency domain.

[0075] Step 102: establishing a mapping relationship between the device electromagnetic noise and the spatial domain according to the spectrum feature map, associating the mapping relationship with the spatial position characteristics of the device electromagnetic noise in the DR image, and obtaining a space-frequency association relationship;

[0076] Specifically, the noise PSD peak coordinates in the spectrum feature map are extracted as key anchor points, the frequency domain coordinates are associated with the image spatial coordinates, and a two-dimensional mapping function is constructed, the expression of which is:

[0077] ;

[0078] wherein is the noise anchor point, is the PSD weight, is the bandwidth parameter, and finally the probability density distribution map of the noise in the spatial domain is output. Then, the direction field of the artifact texture in the DR image is extracted through the Hessian matrix, and the covariance matrix of the calculated noise spatial position and the direction field is taken as the space-frequency association relationship matrix.

[0079] ​Step 103: constructing an adaptive filter kernel according to the spectral feature map and the space-frequency correlation relationship to perform noise filtering on the DR image to obtain a filtered image; the adaptive filter kernel is a rotatable elliptical kernel, and a long axis direction of the adaptive filter kernel is orthogonal to a texture direction of the artifact;

[0080] Specifically, the dominant direction of the artifact is determined according to a texture direction angle channel in the space-frequency correlation relationship matrix , and an elliptical kernel parameter equation is generated, and the expression is:

[0081] ;

[0082] wherein, is a coordinate of an arbitrary point in the filter kernel relative to the center of the kernel, is an elliptical long semi-axis length, is an elliptical short semi-axis length, and the long axis direction of the kernel is set to + 90° to be orthogonal to the texture of the artifact. Then, the kernel size is dynamically adjusted through bilinear interpolation, so that the center weight of the kernel is the highest, and the edge weight is zero, to obtain an adaptive rotating elliptical kernel.

[0083] Step 104: constructing a segmented remapping function, taking local variance entropy of the filtered image as an optimization objective to perform iterative optimization to obtain a de-artifact image.

[0084] Specifically, the CT value range is divided into three intervals of a bone tissue region of [-1000, 200], a soft tissue region of [200, 1000] and a residual artifact region of [1000, 3000], and a linear-convex combination function is defined in each interval, and the combination function of the bone tissue region in this embodiment is f(v) = a log(v+β), wherein a and β are slope compensation coefficients. The variance entropy of the three intervals is calculated in an 8x8 local window, and the combination function is iteratively updated by gradient ascent method until the entropy value change rate is less than a threshold e -4 , so as to obtain the de-artifact image.

[0085] It should be noted that the adaptive filter kernel constructed by the space-frequency joint modeling realizes accurate directional suppression of electromagnetic artifacts, and the local variance entropy driven remapping optimization significantly retains the bone microstructure while removing noise.

[0086] Step 200: segmenting the sacrum and ilium in the de-artifact image through a neural network model to obtain a sacroiliac joint;

[0087] Specifically, depthwise separable convolutional layers are used to process the artifact-removed image. Each input channel is independently scanned using a 3×3 spatial convolutional kernel to extract local bone texture. A 1×1 convolutional kernel is then used to fuse multi-channel features, generating a complete bone texture feature map. The artifact-removed image is divided into three regions based on the HU value: soft tissue, cancellous bone, and cortical bone, corresponding to dilation rates of 1, 2, and 4, respectively. In the U-Net decoder stage, the convolutional kernel size is dynamically adjusted based on the dilation rate indexed by the HU value of the current pixel, focusing the network attention on anatomical structure boundaries. Simultaneously, a channel attention module is embedded in the skip connections of the U-Net, weighting the feature channels of HU value abrupt changes (gradient > 150) by a factor of 3 to enhance edge response, thereby identifying and segmenting the sacrum and iliac bone, resulting in two segmentation masks. Then, the Canny operator is used to detect the physical edges of HU value abrupt change regions in the artifact-free image; the trabecular bone texture amplitudes in the four directions of 0°, 45°, 90°, and 135° are extracted from the bone texture feature map using an 8-pixel Gabor filter, and summed to obtain the texture edge map; the bone edge pixels are modeled as a graph structure, with nodes representing edge point coordinates, and edge weights determined by Euclidean distance and gradient direction difference; then, three-layer neighborhood aggregation is used to repair bone fracture edges caused by artifacts or low contrast, resulting in topologically continuous bone edges. Next, the Hausdorff distance d between the sacrum and ilium is calculated. H Through the discriminator, when d H When the distance is less than 5mm, the area of ​​contact between the two bones is determined to be the sacroiliac joint.

[0088] It should be noted that the multimodal edge fusion of Canny physical edge + Gabor texture edge + GCN topological edge and the dynamic receptive field mechanism driven by HU value significantly improves the segmentation accuracy of complex anatomical structures. At the same time, the discriminator guided by Hausdorff distance also reduces joint localization error.

[0089] Step 300: Standardize the artifact-removed image in the IS coordinate system using affine transformation to obtain a standardized image;

[0090] Specifically, a key point detection network is used to locate the center point of the superior border of the sacrum. The center point of the left iliac joint surface and the center point of the right iliac joint surface Three key anatomical structures. Let the original image coordinate system be... The IS coordinate system is The reference points are as follows: The least squares method minimizes the sum of squared distances between corresponding points, generating a 3×3 affine transformation matrix. to satisfy conditions, It consists of three parameters: translation, rotation, and scaling, and its expression is:

[0091] ;

[0092] wherein is the rotation and scaling component in the joint surface direction, is the translation component of the center point of the superior edge of the sacrum. The position of each pixel in the IS coordinate system is traversed The coordinates of each pixel in the original image are calculated by the affine transformation matrix If the original image coordinates are not integer coordinates, take the HU values of the adjacent 4 and pixel points, and perform bilinear interpolation according to the distance weight, and the expression is:

[0093] ;

[0094] so that the artifact-removed image is aligned to the IS coordinate system. The positional deviation of the key anatomical structure after alignment is calculated by the Euclidean distance, and if the deviation is greater than > 2mm, the artifact-removed image is aligned again by the ICP algorithm to obtain the standardized image.

[0095] Step 400: three-dimensional reconstruction of the coronal and sagittal planes of the standardized image to obtain a reconstructed image, and extraction of the joint cavity depth coefficient of the sacroiliac joint;

[0096] Specifically, along the Y axis of the IS coordinate system, i.e. the sagittal direction, the coronal plane slices of the standardized image are extracted layer by layer at an interval of 0.5mm, the adjacent slices are fused by the maximum intensity projection (MIP) algorithm to generate an interarticular space panoramic view, which retains the continuity feature of the sacroiliac joint surface and enhances the cartilage boundary contrast through histogram equalization, ensuring the visualization integrity of the joint cavity structure. Then in the coronal plane view, i.e. the interarticular space panoramic view, a cubic B-spline curve is fitted based on the iliac articular surface edge points, and the tangent direction vector of the curve at each point is calculated, and the angle between the tangent and the normal vector of the sagittal plane is taken as the geometric constraint function, so that the sagittal plane profile is perpendicular to the local curvature of the articular surface. At the same time, along the X axis of the IS coordinate system, i.e. the coronal direction, profile lines are generated at an interval of 0.5mm, and the direction of each profile line is dynamically adjusted by the geometric constraint function, and at the position , the profile direction is N(x, y) = [sin θ(y), cos θ(y)], to ensure that the profile is always orthogonal to the articular surface tangent, the profile pixel value is extracted by bilinear interpolation to form a sequence containing 512 profiles, and each profile has a length of 128 pixels.

[0097] The end points of each profile line are taken as the Bezier curve control points, and the adjacent control points are connected by a third-order Bezier function to generate a smooth spatial curve path. The curve path is resampled at a resolution of 0.1 mm, and adaptive interpolation is performed in the normal plane of the profile direction N(x, y), that is, interpolation is performed in the direction of the normal vector of the joint cavity with a step size of 0.1 mm, and interpolation is performed in the tangential direction with a step size of 0.5 mm, to obtain a reconstructed image of 512x128x128 voxels. The region with a HU value less than -350 and a gradient amplitude greater than 100 in the reconstructed image is marked as a candidate erosion region; the average depth , minimum depth , depth variation and erosion region depth value of the candidate erosion region are calculated, and the calculation formulas are as follows:

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] wherein, is the depth value, which is obtained by measuring the Euclidean distance from the surface of the erosion region to the nearest articular surface, and the erosion region depth value of the candidate erosion region is taken as an erosion impact factor to construct a joint cavity depth coefficient, which is a four-tuple .

[0103] It should be noted that the image reconstruction error is reduced by dynamic geometric constraints and Bezier anisotropic interpolation; the pathological relevance of the joint cavity depth coefficient is improved by quantifying the candidate erosion region and the depth variation of the HU gradient double threshold marking.

[0104] Step 500: multi-scale feature fusion is performed on the reconstructed image by a multi-scale image pyramid to obtain a multi-scale feature;

[0105] Specifically, a 5-level multi-scale image pyramid is constructed, and the feature size of each level of the pyramid is 256x256, 128x128, 64x64, 32x32 and 16x16 respectively. The reconstructed image is down-sampled level by level, and the local gray level co-occurrence matrix is calculated on each level of the pyramid image using the sliding window method. The window size is fixed at 16x16 pixels, and the step size is 8 pixels. The spatial distance of each pixel in the segmentation mask to the nearest joint boundary is calculated to generate a distance transform map, where the distance value in the joint cavity is positive, the bone region is negative, and the joint boundary is zero. The local gray level co-occurrence matrix is normalized by Z-score, the distance value is linearly mapped to the interval [0, 1] by min-max normalization, and the multi-scale features are obtained by channel-by-channel subtraction operation, which highlights the pathological features in the joint cavity.

[0106] Step 600: Based on the reconstructed image, time-domain evolution is performed by a pre-constructed spatio-temporal dual-flow network to obtain the bone erosion area change rate;

[0107] Specifically, the reconstructed image is input into the ResNet-34 network, the global average pooling layer and the fully connected layer of the original ResNet-34 network model are removed, and only the convolutional backbone is retained. The four groups of residual blocks capture the bone edge microstructure and identify the macroscopic pathology to obtain the static pathological feature map. The static pathological feature map is expanded along the time dimension to a sequence, which is input into the bidirectional LSTM network. The number of hidden layer units of the network is 256, and the hidden state is output through the time step. The transposed convolution layer is decoded to generate the time sequence heat map of bone erosion. The lowest point of the sacrum S1 pedicle is taken as the dynamic origin, the sacrum midline direction is taken as the Z axis, and the plane parallel to the sacrum is taken as the X plane to establish a reference coordinate system. The time sequence heat map is projected onto the sagittal plane of the reference coordinate system, and the projection difference between adjacent time sequences is calculated. The calculation formula is:

[0108] ;

[0109] wherein is the heat map difference value at the projection coordinate , is the value of the heat map at time point t at coordinate , is the HU value of point, is an indicator function, which takes a value of 1 when the condition in the parentheses is satisfied, and 0 otherwise. The projection difference is then area-integrated to obtain the bone erosion area change rate, and the calculation formula is:

[0110] ;

[0111] ;

[0112] wherein Reference area.

[0113] Step 700: Joint cavity depth coefficient, bone erosion area change rate and multi-scale feature are fused by logistic regression to obtain grading results.

[0114] Specifically, the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale feature are compressed to a five-dimensional feature vector through a full connection layer, and the five-dimensional feature vectors compressed multiple times are spliced into a five-dimensional learnable matrix, and the compression times of the embodiment are 5 times. The clinical test indexes of the patient, C-reactive protein (CRP), erythrocyte sedimentation rate (ESR) and disease duration, are standardized to generate a three-dimensional vector, which is expanded to a five-dimensional feature modulation vector through a linear transformation layer, and is used to adjust the pathological sensitivity of the image feature. The region with a HU value less than -300 and a Hausdorff distance less than 5mm in the reconstructed image is marked as a pathological region, and the remaining region is marked as a non-pathological region. The joint cavity depth coefficient, the bone erosion area change rate and the multi-scale feature in the pathological region are all enhanced by 3 times weight, and the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale feature in the non-pathological region are all attenuated by 0.3 times weight, and the weighted features are globally averaged to generate an optimized matrix, which significantly strengthens the contribution weight of the pathological region. The optimized matrix and the modulation vector are added element by element, and the pathological index P is calculated through a sigmoid function, and the calculation formula is:

[0115] ;

[0116] Wherein, is the i-th dimension feature of the optimized matrix, is the i-th dimension feature of the feature modulation vector, is a logistic regression coefficient, which is not limited in the embodiment.

[0117] If P<0.3, it is I level, if 0.3≤P<0.6, it is II level, if 0.6≤P<0.8, it is III level, and if P≥0.8, it is IV level.

[0118] The application also provides a sacroiliac joint X-ray image intelligent grading system, comprising:

[0119] An image mapping module is configured to generate an adaptive filtering kernel based on an electromagnetic noise spectrum, and to perform CT value remapping on a DR image through the adaptive filtering kernel to obtain a deartifact image.

[0120] A joint recognition module is configured to segment a sacrum and an ilium in the deartifact image through a neural network model to obtain a sacroiliac joint.

[0121] A coordinate conversion module is configured to perform a standardization operation on the deartifact image in an IS coordinate system through an affine transformation to obtain a standardized image.

[0122] an image reconstruction module, configured to perform three-dimensional reconstruction of the coronal plane and the sagittal plane on the standardized image to obtain a reconstructed image, and extract a joint cavity depth coefficient of the sacroiliac joint;

[0123] a feature extraction module, configured to perform multi-scale feature fusion on the reconstructed image through a multi-scale image pyramid to obtain multi-scale features;

[0124] an erosion prediction module, configured to perform time domain evolution on the reconstructed image through a pre-constructed spatio-temporal double-flow network to obtain a bone erosion area change rate;

[0125] an intelligent grading module, configured to perform logic regression fusion on the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale features to obtain a grading result.

[0126] The beneficial effects of the present application are as follows:

[0127] 1) The electromagnetic noise and the anatomical structure signal are accurately separated through wavelet packet decomposition, and the bone texture loss problem is solved by combining a direction-sensitive adaptive elliptical filter kernel, so that the bone edge sharpness is improved;

[0128] 2) The accurate segmentation of complex anatomical structures is realized through a HU value dynamically adjusted U-Net network, a Canny detection algorithm and a graph convolution topology repair technology, and the segmentation accuracy is greatly improved;

[0129] 3) The anatomical landmarks are located based on a key point detection network, and the influence of body position difference is eliminated through affine transformation and ICP secondary calibration, so as to provide a spatial consistent IS coordinate system reference for three-dimensional reconstruction;

[0130] 4) The candidate erosion area is marked through a HU gradient double threshold value, and the pathological relevance of feature extraction is improved by combining the joint cavity depth coefficient and the multi-scale features.

[0131] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between the embodiments can be mutually referred to.

[0132] The principles and implementation manners of the present application are described by using specific examples in the present application. The above description of the embodiments is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range will be changed. In conclusion, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for intelligent grading of sacroiliac joint X-ray images, characterized in that, The method comprises the following steps: An adaptive filter kernel is generated based on an electromagnetic noise spectrum, and CT value remapping of a DR image is performed through the adaptive filter kernel to obtain a deartifact image; A sacroiliac joint is obtained by segmenting the sacrum and ilium in the deartifact image through a neural network model; The deartifact image is standardized in an IS coordinate system through affine transformation to obtain a standardized image; Three-dimensional reconstruction of the standardized image in coronal and sagittal planes is performed to obtain a reconstructed image, and a joint cavity depth coefficient of the sacroiliac joint is extracted; Multi-scale feature fusion of the reconstructed image is performed through a multi-scale image pyramid to obtain multi-scale features; Based on the reconstructed image, time-domain evolution is performed through a pre-constructed spatio-temporal double-flow network to obtain a bone erosion area change rate; Logical regression fusion of the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale features is performed to obtain a grading result.

2. The method of claim 1, wherein the method further comprises: An adaptive filter kernel is generated based on an electromagnetic noise spectrum, and CT value remapping of a DR image is performed through the adaptive filter kernel to obtain a deartifact image, comprising: Wavelet packet decomposition is performed on the DR image to extract high-frequency sub-band components, and blind source separation algorithm is used to separate device electromagnetic noise and anatomical structure signals in the DR image to generate a spectrum feature map; A mapping relationship between the device electromagnetic noise and a spatial domain is established according to the spectrum feature map, the mapping relationship is associated with spatial position characteristics of the device electromagnetic noise in the DR image to obtain a space-frequency association relationship; The adaptive filter kernel is constructed according to the spectrum feature map and the space-frequency association relationship to perform noise filtering on the DR image to obtain a filtered image; the adaptive filter kernel is a rotatable elliptical kernel, and a long axis direction of the adaptive filter kernel is orthogonal to a artifact texture direction; A segmented remapping function is constructed, and local variance entropy of the filtered image is taken as an optimization target to perform iterative optimization to obtain the deartifact image.

3. The method of claim 1, wherein the method further comprises: A sacroiliac joint is obtained by segmenting the sacrum and ilium in the deartifact image through a neural network model, comprising: Whole-image bone texture features in the deartifact image are extracted through a depth separable convolution; Based on the whole-image bone texture features, segmentation masks of the sacrum and ilium are respectively generated through an improved U-Net network; The receptive field shape of the improved U-Net network is adjusted according to HU values in the deartifact image to perform attention extraction on the segmentation masks; Physical edges of HU value mutation zones in the deartifact image are detected through a Canny operator; Based on the whole-image bone texture features, texture edges are obtained by capturing bone trabecula directions through a Gabor filter; Based on the whole-image bone texture features, topological edges are obtained by repairing broken bone edges through a graph convolution network; Hausdorff distances between the segmentation masks are calculated; Based on the value of the Hausdorff distance, the sacroiliac joint is identified through a discriminator.

4. The method of claim 1, wherein the method further comprises: The deartifact image is standardized in an IS coordinate system through affine transformation to obtain a standardized image, comprising: Positioning key anatomical structures through a key point detection network; the key anatomical structures include: a sacral upper edge center point, a left iliac articular surface center point, and a right iliac articular surface center point; An affine transformation matrix of the key anatomical structures is calculated through a least square method; Based on the affine transformation matrix, an image resampling is performed on the artifact-removed image through a bilinear interpolation, so that the artifact-removed image is aligned to the IS coordinate system; A position deviation of the aligned key anatomical structures is calculated, and if the deviation is greater than a preset deviation threshold, a secondary alignment is performed on the artifact-removed image through an ICP algorithm, to obtain the standardized image.

5. The method of claim 1, wherein the method further comprises: A three-dimensional reconstruction of the standardized image in a coronal plane and a sagittal plane is performed, to obtain a reconstructed image, and a joint cavity depth coefficient of the sacroiliac joint is extracted, including: Slices are stacked along a Y axis of the IS coordinate system, to generate an interarticular space panoramic view; A geometric constraint of the coronal plane to the sagittal plane is constructed according to a tangent direction of an iliac articular surface of the coronal plane; Based on the geometric constraint, an interarticular space profile sequence is generated along an X axis of the IS coordinate system; Anisotropic interpolation is performed on the interarticular space profile sequence through a Bézier curve driven voxel interpolation algorithm, to obtain the reconstructed image; An area with a HU value less than -350 and a gradient amplitude greater than 100 in the reconstructed image is marked as a candidate erosion area; An erosion area depth value of the candidate erosion area is taken as an erosion influence factor to construct the joint cavity depth coefficient; the joint cavity depth coefficient includes: an average depth, a minimum depth, a depth variation degree, and the erosion influence factor.

6. The method of claim 3, wherein the method further comprises: Multi-scale features are obtained through multi-scale image pyramid fusion on the reconstructed image, including: A multi-scale image pyramid is constructed; the multi-scale image pyramid is a 5-level pyramid, and the feature sizes of each level of the pyramid are 256×256, 128×128, 64×64, 32×32, and 16×16, respectively; A local gray level co-occurrence matrix of the reconstructed image is calculated in each level of the multi-scale image pyramid; Based on the segmentation mask, a distance transform map is generated in each level of the multi-scale image pyramid; The local gray level co-occurrence matrix and the distance transform map, which are respectively normalized, are subjected to channel subtraction, to obtain the multi-scale features.

7. The method of claim 1, wherein the method further comprises: determining a first angle between the first line and the second line; and determining a second angle between the first line and the third line. Based on the reconstructed image, a bone erosion area change rate is obtained through a pre-constructed spatio-temporal dual-flow network, including: Static pathological features of the reconstructed image are extracted through a ResNet-34 network; A bone erosion heat map of the static pathological features is extracted through an LSTM network; A reference coordinate system is established with a sacral S1 pedicle as a dynamic origin, a sacral midline direction as a Z axis, and a plane parallel to the sacrum as an X plane; A sagittal plane projection difference value of the bone erosion heat map in the reference coordinate system is calculated; An area integral is performed on the sagittal plane projection difference value, to obtain the bone erosion area change rate.

8. The method of claim 1, wherein the method further comprises: determining a first angle between the first line and the second line; determining a second angle between the first line and the third line; and determining a third angle between the second line and the third line. Logical regression fusion is performed on the joint cavity depth coefficient, the bone erosion area change rate, and the multi-scale features, to obtain a grading result, including: According to the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale feature, a five-dimensional learnable matrix is constructed; According to the clinical test index, a feature modulation vector is generated; Regions in the reconstructed image with HU values less than -300 and Hausdorff distances less than 5mm are marked as pathological regions, and the remaining regions are marked as non-pathological regions; The joint cavity depth coefficient, the bone erosion area change rate and the multi-scale feature in the pathological region are all enhanced by 3 times weight, and the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale feature in the non-pathological region are all attenuated by 0.3 times weight, to obtain an optimized matrix; The optimized matrix and the feature modulation vector are fused by logistic regression to obtain a pathological index; According to the size of the pathological index, a grading result is obtained.

9. A sacroiliac joint X-ray image intelligent grading system, characterized in that, Comprise: An image mapping module is configured to generate an adaptive filter kernel based on an electromagnetic noise spectrum, and to perform CT value remapping on a DR image through the adaptive filter kernel to obtain a deartifact image; A joint recognition module is configured to segment the sacrum and ilium in the deartifact image through a neural network model to obtain a sacroiliac joint; A coordinate conversion module is configured to perform standardization operation on the deartifact image in an IS coordinate system through affine transformation to obtain a standardized image; An image reconstruction module is configured to perform three-dimensional reconstruction of the standardized image in coronal and sagittal planes to obtain a reconstructed image, and to extract a joint cavity depth coefficient of the sacroiliac joint; A feature extraction module is configured to perform multi-scale feature fusion on the reconstructed image through a multi-scale image pyramid to obtain a multi-scale feature; An erosion prediction module is configured to perform time domain evolution on the reconstructed image through a pre-constructed spatio-temporal double-flow network to obtain a bone erosion area change rate; An intelligent grading module is configured to fuse the joint cavity depth coefficient, the bone erosion area change rate and the multi-scale feature by logistic regression to obtain a grading result.

Images