Osteoporosis evaluation method and system based on spine CT image and application method of system

Abstract

The present application relates to a kind of based on vertebral CT image osteoporosis evaluation method and system, solve the problem of high requirement of equipment, site in prior art osteoporosis evaluation.Method includes: obtaining vertebral CT image, pre-processing, obtain multiple vertebral slice images;Each vertebral slice image is extracted, and basic vertebral feature map is obtained;According to basic vertebral feature map, spatial dimension attention calculation, feature re-labeling on channel dimension are carried out, respectively, global perception feature map, channel enhancement feature map are obtained;Feature fusion is carried out to basic vertebral feature map, global perception feature map and channel enhancement feature map, and optimization vertebral feature map is obtained, using global classifier, determine osteoporosis basic evaluation result, and then determine osteoporosis final evaluation result.The present application realizes the automatic segmentation of vertebral CT image, obtains high-quality feature map using attention framework, and then evaluates osteoporosis probability, realizes osteoporosis self-evaluation.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, and in particular to a method and system for assessing osteoporosis based on spinal CT images. Background Technology

[0002] Osteoporosis has a high incidence and mortality rate worldwide, making it a major public health problem. However, the current screening rate for osteoporosis is low, leading to a significant number of missed diagnoses. These missed diagnoses cause enormous physical and psychological harm to patients and impose a huge economic burden on society. Technically, the main reason is that existing screening and diagnostic technologies and equipment are only available in specialized medical examination and treatment institutions, lacking methods that are more readily available to the general population, such as blood pressure monitors.

[0003] Existing technologies mainly include traditional methods and innovative methods that combine AI technology. Traditional methods include the gold standard dual-energy X-ray absorptiometry (DXA), which assesses bone mineral density by reflecting the characteristics of bone tissue absorption of X-rays; and quantitative CT technology, which is similar in principle to DXA. Innovative methods primarily utilize automated identification and processing of X-ray and CT images to diagnose osteoporosis. For example, Harbour Light Solutions LCC's smartphone osteoporosis screening application (BoneGauge) requires devices capable of clearly capturing posteroanterior (PA) X-ray images of the hand. The algorithm calculates the second metacarpal cortical thickness ratio (2MCP) based on the Schreiber formula and then classifies patients by threshold, making it suitable for patients aged 50 and above with low-energy distal radius fractures. A radiomics prediction model for osteoporosis based on lumbar spine CT images uses the L1-L4 vertebral bodies as regions of interest (ROIs). The SenseCare platform automatically segments the vertebral bodies, and professional physicians manually correct the boundaries. PyRadiomics is then used to extract 1197 radiomics features from the ROIs, suitable for the preliminary diagnosis of osteoporosis in patients with low back pain. A deep learning-based lumbar spine bone mineral density prediction system requires access to non-enhanced abdominal CT images containing the lumbar region and uses a convolutional neural network (CNN) to estimate the bone mineral density (BMD) of the L1-L4 vertebral bodies. Value (bone mineral density); The AI-based chest CT opportunistic osteoporosis screening system requires access to chest CT images, sets cylindrical regions of interest (ROIs) for different vertebrae, and extracts the average CT value (HU) within the ROI as the basis for bone mineral density assessment.

[0004] Traditional methods, limited by equipment costs, are only available in large medical institutions such as top-tier hospitals. Health checkup centers typically use limb bone ultrasound for screening, which is only effective for those undergoing regular checkups. Innovative methods incorporating AI technology still require imaging resources within medical institutions and necessitate physician operation or assistance, making timely outpatient assessment at home difficult. A small portion of these methods utilize X-rays for evaluation, particularly when assessing bone density in the limbs rather than axial skeletons like the spine, resulting in lower accuracy in predicting osteoporosis.

[0005] Therefore, there is an urgent need for an osteoporosis assessment method and system based on spinal CT images to achieve low-threshold, low-cost, and more convenient osteoporosis assessment, thereby improving the osteoporosis screening rate under the current conditions. Summary of the Invention

[0006] In view of the above analysis, the present invention aims to provide a method and system for osteoporosis assessment based on spinal CT images, addressing the problem that existing osteoporosis assessments have high requirements for equipment and facilities.

[0007] On one hand, embodiments of the present invention provide a method for osteoporosis assessment based on spinal CT images, including:

[0008] Acquire spinal CT images, preprocess the spinal CT images to obtain multiple spinal slice images; Feature extraction is performed on each of the spinal slice images to obtain a basic spinal feature map; Based on the aforementioned basic spinal feature map, spatial dimension attention calculation is performed to obtain a global perception feature map; Based on the basic spinal feature map, feature recalibration is performed in the channel dimension to obtain the channel-enhanced feature map; The basic spinal feature map, the global perception feature map, and the channel enhancement feature map are fused to obtain an optimized spinal feature map. Based on the optimized spinal feature map, a global classifier is used to determine the basic assessment results of osteoporosis. Based on the baseline osteoporosis assessment results corresponding to each vertebral slice image, the final osteoporosis assessment result is determined.

[0009] Furthermore, based on the aforementioned basic spinal feature map, spatial attention calculation is performed to obtain a global perception feature map, including: The basic spine feature map is input into two first convolutional layers to obtain the query matrix and the key matrix, respectively. The query matrix and the key matrix are flattened into query-space point sequences and key-space point sequences, respectively. Matrix multiplication is then performed on the query-space point sequences and the key-space point sequences to obtain a spatial attention map. The basic spine feature map is convolved to obtain a value matrix; Based on the spatial attention map, spatial attention enhancement is performed on the value matrix to obtain a spatial fusion feature map; The spatial fusion feature map and the basic spine feature map are fused to obtain the global perception feature map.

[0010] Furthermore, based on the aforementioned basic spine feature map, feature recalibration is performed along the channel dimension to obtain a channel-enhanced feature map, including: The basic spine feature map is flattened into a channel-space point matrix, and the channel-space point matrix is ​​input into two second convolutional layers to obtain the query vector and the health vector, respectively. Perform matrix multiplication on the query vector and the key vector to obtain the channel relevance matrix; The channel correlation matrix is ​​normalized to obtain the channel attention weights; Based on the channel attention weights, attention enhancement is performed on the basic spine feature map to obtain an attention-enhanced spine feature map; The attentional spine feature map and the basic spine feature map are fused to obtain the channel enhancement feature map.

[0011] Further, feature fusion is performed on the basic spinal feature map, the globally perceived feature map, and the channel-enhanced feature map to obtain the optimized spinal feature map, including: The global perception feature map and the channel enhancement feature map are concatenated to obtain a concatenated spine feature map. Based on the spliced ​​spine feature map, a gated map is obtained using a two-level convolutional network; Based on the gating map, the global perception feature map and the channel enhancement feature map are weighted and fused to obtain a spatial-channel fusion feature map; The optimized spine feature map is obtained by fusing the spatial-channel fusion feature map, the global perception feature map, and the channel enhancement feature map.

[0012] Furthermore, based on the baseline osteoporosis assessment results corresponding to each vertebral slice image, the final osteoporosis assessment results are determined, including: Identify the target orientation category of each of the spinal slice images, and divide all the basic osteoporosis assessment results into multiple assessment result sets based on the target orientation category of each spinal slice image; Based on the osteoporosis baseline assessment results of each set of assessment results, calculate the corresponding osteoporosis unidirectional assessment result; Based on all the osteoporosis unidirectional assessment results, a voting method was used to determine the final osteoporosis assessment result.

[0013] Further, the target orientation category of each of the said spinal slice images is identified, including: The CLIP model is used to obtain the image feature vector corresponding to the spinal slice image; Obtain the label text vector for each directional category; Calculate the similarity between the image feature vector and each of the label text vectors to determine the target orientation category corresponding to the spine slice image.

[0014] Further, the spinal CT images are preprocessed to obtain multiple spinal slice images, including: Calculate the gradient magnitude and gradient direction of each pixel in the spinal CT image; Based on the gradient magnitude and gradient direction of each pixel, the thinning edge of the spinal CT image is determined using non-maximum suppression. Based on the refined edges, a dual-threshold hysteresis process is used to determine the final edges of the spinal CT image; Based on the final edge, a contour finding algorithm is used to obtain the outer contours of multiple slice regions of the spinal CT image; The spinal CT image is cropped based on the outer contour of all the slice regions to obtain multiple spinal slice images.

[0015] Further, for each of the said spinal slice images, feature extraction is performed to obtain a basic spinal feature map, including: The vertebral slice image is input into a basic feature extractor to obtain the basic vertebral feature map; wherein, the basic feature extractor is a ResNet-50 model.

[0016] On the other hand, embodiments of the present invention provide an osteoporosis assessment system based on spinal CT images, comprising: The image acquisition module is used to acquire spinal CT images and preprocess the spinal CT images to obtain spinal slice images. The feature extraction module is used to extract features from the spinal slice image to obtain a basic spinal feature map; The attention module is used to perform spatial dimension attention calculation based on the basic spinal feature map to obtain a global perception feature map; to perform feature recalibration in the channel dimension based on the basic spinal feature map to obtain a channel enhancement feature map; and to perform feature fusion on the basic spinal feature map, the global perception feature map, and the channel enhancement feature map to obtain an optimized spinal feature map. The evaluation module is used to determine the basic assessment results of osteoporosis based on the optimized spinal feature map using a global classifier.

[0017] On the other hand, embodiments of the present invention provide an application method for an osteoporosis assessment system, characterized in that it includes: Obtain an osteoporosis assessment system; wherein the osteoporosis assessment system is determined based on the former aspect; The osteoporosis assessment system is weighted and quantized, and the quantized osteoporosis assessment system is converted into an osteoporosis computation graph; wherein, the osteoporosis computation graph includes computation nodes and data flow edges; The osteoporosis calculation map is fused into nodes to obtain a lightweight osteoporosis calculation map; A lightweight osteoporosis assessment system is obtained based on the aforementioned lightweight osteoporosis calculation diagram. The lightweight osteoporosis assessment system is packaged into an application.

[0018] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. By preprocessing spinal CT images to obtain spinal slice images, and then extracting features from these slice images to obtain baseline feature maps, optimized spinal feature maps are obtained through spatial dimension attention calculation and channel dimension feature restandardization. Using a global classifier, the final osteoporosis assessment result is determined. Existing spinal images can be used to automatically process and analyze non-standardized spinal CT images or screenshots, reducing the equipment and site requirements for osteoporosis detection, improving the convenience of osteoporosis assessment, and eliminating the need for professional medical personnel in the assessment process. This improves the accessibility of osteoporosis screening, allowing the general public to obtain osteoporosis assessment results independently. The operation is simple.

[0019] 2. Spatial attention calculation is performed on the basic spinal feature map to form a globally perceptive feature map that retains local details. Through channel-dimensional feature recalibration, a channel-enhanced feature map is obtained, which enables the feature fusion process to focus on key anatomical structures in the spine that are closely related to bone density, forming an optimized spinal feature map. This captures long-distance spatial dependencies and optimizes channel feature representation. The optimized spinal feature map has both global perception capability and local detail fidelity, improving the perception accuracy of bone loss areas in complex medical imaging backgrounds and enhancing the accuracy of osteoporosis assessment.

[0020] 3. Recognize the orientation category of spinal slice images to obtain corresponding basic osteoporosis assessment results, realize automatic osteoporosis classification, enhance the interpretability of assessment results, and provide convenience for remote diagnosis and cloud collaboration.

[0021] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0022] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart illustrating an osteoporosis assessment method based on spinal CT images in an embodiment of the present invention. Figure 2 This is a flowchart illustrating another osteoporosis assessment method based on spinal CT images, as described in this embodiment of the invention. Figure 3 This is a schematic diagram of the main modules of an osteoporosis assessment system based on spinal CT images in an embodiment of the present invention. Detailed Implementation

[0023] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0024] One specific embodiment of the present invention discloses a method for osteoporosis assessment based on spinal CT images, such as... Figure 1 As shown, it includes: Step S1: Acquire spinal CT images, preprocess the spinal CT images to obtain multiple spinal slice images.

[0025] The original CT images generated by medical institutions usually include multiple slice images. In this embodiment, after obtaining the spinal CT image, the spinal CT image is preprocessed, the Canny edge detection algorithm is used to locate the slice area, and the contour search algorithm is used to determine the outer contour of the slice area to obtain multiple spinal slice images. Specifically, it includes steps S11-S16.

[0026] Step S11: Obtain spinal CT images.

[0027] In this embodiment, spinal CT images are acquired by capturing images with an imaging device, such as a mobile phone. In other embodiments, screenshots of CT cloud images can also be obtained, or raw spinal CT images generated by medical institutions can be directly accessed.

[0028] Furthermore, considering that image format conversion of spinal CT images obtained by shooting or screenshotting leads to a loss of image accuracy, image enhancement processing is performed on the acquired spinal CT images, including grayscale conversion and Gaussian filtering, to improve image accuracy, eliminate noise, and facilitate subsequent slice area localization.

[0029] It is understandable that when capturing or screenshotting spinal CT images, multiple spinal CT images can be obtained from different directions to ensure information integrity.

[0030] Step S12: Calculate the gradient magnitude and gradient direction of each pixel in the spinal CT image.

[0031] Calculate each pixel in a spinal CT image lateral gradient magnitude in the direction and longitudinal gradient magnitude in the direction Determine the gradient magnitude of each pixel. and gradient direction In this embodiment, each pixel is selected. The gradient magnitude and direction are calculated within the neighborhood of the current pixel (centered on the current pixel and surrounding 8 neighboring pixels) to ensure computational efficiency and local feature capture. The gradient magnitude represents the strength of the brightness change in the spinal CT image. The larger the gradient magnitude, the greater the brightness difference between the current pixel and the neighboring pixels, and the more likely it is to be the true edge of the slice area. The gradient direction represents the direction of the brightness change, that is, "the direction in which the brightness changes from dark to bright or from bright to dark", which prepares for subsequent non-maximum suppression.

[0032] The slice regions in spinal CT images are mostly regular geometric edges. Gradient calculation can accurately capture edge information and locate edge candidate points.

[0033] Step S13: Based on the gradient magnitude and gradient direction of each pixel, non-maximum suppression is used to determine the thinning edge of the spinal CT image.

[0034] For each pixel in the spinal CT image, along the gradient direction, determine whether the current pixel is a local maximum. If it is, retain it as an edge candidate point; otherwise, suppress it and mark it as a non-edge point. All edge candidate points corresponding to the spinal CT image constitute the refined edge.

[0035] By using nonmaximum suppression, the refined edges of spinal CT images are determined, avoiding the identification of slice area edges as multi-pixel-wide blurred regions, thus ensuring the continuity and accuracy of edges during subsequent contour extraction.

[0036] Step S14: Based on the refined edges, use dual threshold hysteresis processing to determine the final edges of the spinal CT image.

[0037] By setting dual thresholds, noise filtering is applied to the refined edges to determine the final edges of the spinal CT image.

[0038] First, based on a dual threshold, strong and weak edge points in the thinned edge are determined. Specifically, the dual threshold includes a high threshold and a low threshold. For each pixel in the thinned edge, if the gradient magnitude is greater than the high threshold, it is determined to be a real edge, and the pixel is a strong edge point and is retained; if the gradient magnitude is less than the low threshold, it is determined to be noise, and the pixel is a non-edge point and is discarded; if the gradient magnitude is between the low and high thresholds, it is determined to be a suspected edge, and the pixel is a weak edge point. For example, the high threshold... Low threshold .

[0039] Secondly, based on strong edge points, connectivity is assessed for weak edge points to determine the final edge. Specifically, for each weak edge point, it is determined whether it is connected to any strong edge point. If a directly connected strong edge point exists, the pixel is re-identified as a strong edge point; otherwise, it is re-identified as a non-edge point and discarded. The final edge is constructed based on all strong edge points (including the re-identified strong edge points), ensuring edge continuity and the integrity of the sliced ​​region.

[0040] Step S15: Based on the final edge, use a contour lookup algorithm to obtain the outer contours of multiple slice regions of the spinal CT image.

[0041] Based on the final edges, a contour-finding algorithm is used to obtain the external contours of multiple slice regions in the spinal CT image. Based on the final edges (set of strong edge points) determined in step S14, the strong edge points in the final edges are connected to form a closed contour, and the closed contours belonging to the slice regions are selected, which are the external contours of the slice regions.

[0042] In this embodiment, closed contours belonging to the slice area are selected based on shape. According to the standardization of medical institution examinations, the slice area in the spinal CT image is a regular shape, such as a rectangle. The estimated rectangularity of each closed contour is calculated. If the estimated rectangularity reaches the rectangularity threshold (e.g., 0.95), it is determined to be the outer contour of the slice area.

[0043] In other embodiments, area-based filtering can also be used. An area benchmark value can be set based on the estimated average area of ​​all closed contours, or an area benchmark value can be set based on the area of ​​the slice region specified by the medical institution's examination equipment. Slice regions whose estimated area differs from the area benchmark value by less than an area difference threshold (e.g., 0.1) can be retained as the outer contour of the slice region, while excessively small noise contours and excessively large background contours can be excluded.

[0044] It is understandable that when a spinal CT image includes multiple slice regions, the outer contours of multiple slice regions may be determined based on a single spinal CT image; if a spinal CT image only includes one slice region, then based on the multiple spinal CT images obtained in step S11, at least one outer contour of a slice region is determined for each spinal CT image, ultimately resulting in the outer contours of multiple slice regions.

[0045] Step S16: Based on the outer contours of all the slice regions, the spinal CT image is cropped to obtain multiple spinal slice images.

[0046] Based on the outer contours of all slice regions, the spinal CT images are cropped, and each slice region forms a spinal slice image, resulting in multiple spinal slice images.

[0047] By determining the outer contour of the slice region through edge detection and contour finding, multiple spinal slice images are obtained. The CT image containing multiple slices is converted into a single slice image, which facilitates subsequent feature extraction, enhancement and osteoporosis assessment of the single slice image.

[0048] Step S2: Extract features from each of the vertebral slice images to obtain a basic vertebral feature map.

[0049] Feature extraction is performed on each vertebral slice image to obtain a basic vertebral feature map. In this embodiment, the vertebral slice image is input into a basic feature extractor to obtain the basic vertebral feature map; wherein, the basic feature extractor uses the ResNet-50 model.

[0050] First, each spine slice image is standardized, including size normalization and pixel normalization. For example, for each spine slice image, the image size is uniformly scaled to a fixed resolution (e.g., 224x224 pixels), and the original pixel value of each pixel is normalized to... The interval is used to obtain the normalized pixel value corresponding to each pixel, for example... In the formula, , They represent the first Normalized pixel value and original pixel value of each pixel. , These represent the maximum and minimum values ​​of the original pixel values ​​for each pixel, respectively.

[0051] Secondly, a basic feature extractor is used to extract features from each standardized vertebral slice image. In this embodiment, the ResNet-50 model is used as the basic feature extractor. Thanks to its residual connection structure, the ResNet-50 model can effectively avoid the vanishing gradient problem while maintaining network depth, thus robustly learning image features from low to high levels. Furthermore, the ResNet-50 model is pre-trained using the ImageNet dataset to ensure that it can extract generalized image features from vertebral slice images.

[0052] Step S3: Based on the basic spinal feature map, perform spatial dimension attention calculation to obtain a global perception feature map.

[0053] Based on the basic spinal feature map, attention calculation is performed in the spatial dimension to obtain the global perception feature map, specifically including steps S31-S35.

[0054] Step S31: Input the basic spine feature map into two first convolutional layers to obtain the query matrix and key matrix respectively.

[0055] The basic spine feature map is fed into two first convolutional layers to obtain the query matrix and key matrix, respectively. For example, the size of the basic spine feature map is... ,in, , , The basic spine feature map is represented by its number of channels, height, and width, respectively. The basic spine feature map is simultaneously input into two first convolutional layers and processed independently. Convolution generates a query matrix and a key matrix, compressing the channel dimension and reducing the number of channels. .

[0056] Furthermore, multiple basic spinal features can be used. Figure 1 And input, improve data processing efficiency.

[0057] Step S32: Flatten the query matrix and the key matrix into query-space point sequences and key-space point sequences, respectively, and perform matrix multiplication on the query-space point sequences and the key-space point sequences to obtain a spatial attention map.

[0058] First, the query matrix and key matrix are flattened into query-space point sequences and key-space point sequences, respectively. Then, matrix multiplication of the query-space point sequences and key-space point sequences yields the similarity matrix. By multiplying the expanded matrices, the influence weight of each pixel in the basic spine feature map on all other pixels is determined.

[0059] In this embodiment, the query matrix and key matrix are expanded while channel compression is performed to obtain query-space point sequences and key-space point sequences. The query-space point sequences are then transposed and multiplied with the key-space point sequences to obtain a similarity matrix. In other embodiments, the key-space point sequences can also be transposed, and the query-space point sequences can be multiplied with the transposed key-space point sequences to obtain a similarity matrix.

[0060] For example, the query matrix (query) is of size . The key matrix (key) has a size of After expansion and channel compression, query-spatial point sequences and key-spatial point sequences are obtained, both with dimensions of [missing information]. ,in Multiplying the query-space point sequence and the key-space point sequence by a matrix yields a similarity matrix of size [size missing]. ,Right now .

[0061] Secondly, the elements in the similarity matrix are normalized using the softmax function to obtain the spatial attention map.

[0062] Step S33: Perform convolution processing on the basic spine feature map to obtain the value matrix.

[0063] The basic spine feature map is convolved to obtain a value matrix. The basic spine feature map is then processed... Convolution generates a value matrix. It's important to note that channel compression is not performed during convolution.

[0064] Step S34: Based on the spatial attention map, perform spatial attention enhancement on the value matrix to obtain a spatial fusion feature map.

[0065] The spatial attention map is multiplied with the value matrix to enhance spatial attention, resulting in a spatial fusion feature map that aggregates global contextual information spatially. Specifically, based on the dimension of the spatial attention map, the single-channel matrix of each channel of the value matrix is ​​transformed to form a single-channel vector, ensuring that the dimension of the single-channel vector matches that of the spatial attention map. The spatial attention map is then broadcast and multiplied with the single-channel vector of each channel of the value matrix to enhance the spatial attention of each channel's single-channel vector. The enhanced single-channel vector of each channel is then transformed again to restore its original dimension, resulting in a single-channel enhancement matrix. The single-channel enhancement matrices of all channels constitute the spatial fusion feature map. For example, the spatial attention map and the similarity matrix have the same size. ,Right now The size of the value matrix is , value matrix Single-channel matrix ( ) The size is ), respectively converted into single-channel vectors ( The size is The spatial attention map is multiplied by each single-channel vector to obtain the corresponding enhanced single-channel vector. The size is ), and then perform another dimensional transformation to restore the original dimensions, resulting in a single-channel enhancement matrix ( The size is ), The single-channel enhancement matrices of each channel constitute a spatial fusion feature map (size: ).

[0066] Step S35: Perform feature fusion on the spatial fusion feature map and the basic spine feature map to obtain the global perception feature map.

[0067] Feature fusion is performed on the spatial fusion feature map and the basic spine feature map to obtain a globally perceptual feature map. In this embodiment, learnable parameters are utilized. The spatial fusion feature map is weighted, and then added to the base spine feature map to form a globally perceptual feature map that preserves local details. For example, learnable parameters... It is a vector, and each element can take values ​​ranging from 1 to 2. It can be optimized through training and learning. The weighted spatial fusion feature map is added element-wise to the basic spine feature map to obtain a globally perceived feature map.

[0068] Step S4: Based on the basic spine feature map, perform feature recalibration in the channel dimension to obtain the channel-enhanced feature map.

[0069] Based on the basic spinal feature map, feature recalibration is performed in the channel dimension to obtain the channel-enhanced feature map, specifically including steps S41-S45.

[0070] Step S41: Flatten the basic spine feature map into a channel-space point matrix, and input the channel-space point matrix into two second convolutional layers to obtain the query vector and the health vector, respectively.

[0071] First, for each dimension of the basic spinal feature map, it is flattened according to the height and width of the basic spinal feature map to form a channel-space point matrix. For example, the size of the basic spinal feature map is... ,in, , , The number of channels, height, and width of the basic spine feature map are represented respectively. Each dimension is flattened to form a channel-space point matrix with dimensions of [missing information]. ,in , which serves as the original representation of the query vector and the key vector.

[0072] Secondly, the channel-spatial point matrix is ​​simultaneously input into two second convolutional layers, where convolution is performed independently to generate query vectors and key vectors. In this embodiment, the query vectors and key vectors focus on channel features, compressing spatial information. For example, the dimensions of the query vectors and key vectors are... .

[0073] Step S42: Perform matrix multiplication on the query vector and the key vector to obtain the channel relevance matrix.

[0074] The query vector and key vector are multiplied by a matrix to obtain the channel correlation matrix, which quantifies the mutual information strength of different feature channels. In this embodiment, the query vector is transposed and then multiplied by the key vector to obtain the channel correlation matrix. In other embodiments, the key vector can also be transposed, and the query vector can be multiplied by the transposed key vector to obtain the channel correlation matrix. For example, the dimensions of the query vector and key vector are... The channel correlation matrix size is .

[0075] Step S43: Normalize the channel correlation matrix to obtain the channel attention weights.

[0076] First, a maximum value normalization strategy is employed to perform a first normalization process on the channel correlation matrix, yielding the channel attention weights. Specifically, for the channel correlation matrix, the maximum value is extracted from the vector corresponding to each channel. Based on the maximum values ​​for each channel, a difference matrix is ​​constructed to enhance the discriminative power of the dominant channel. For example, the channel correlation matrix has the following dimensions: The difference matrix has a size of or .

[0077] Secondly, the softmax function is used to normalize the elements in the difference matrix to obtain the channel attention weights.

[0078] Step S44: Based on the channel attention weights, perform attention enhancement on the basic spine feature map to obtain an attention spine feature map.

[0079] Channel attention weights are multiplied by the base spine feature map using a matrix multiplication. This matrix multiplication enhances the channel attention of the base spine feature map, resulting in an attention-focused spine feature map. This achieves dynamic reweighting of the dominant channels in the base spine feature map. For example, the channel attention weights are broadcast, and the single-channel spine feature matrix of each channel in the base spine feature map is multiplied by the corresponding weight for that channel. For instance, if the channel attention weights are... ,in Indicates the first The single-channel weights of each channel, for the basic spinal feature map ( ) No. Single-channel spine characteristic matrix Channel attention enhancement was performed to obtain the enhanced single-channel spine characteristic matrix. ,in .

[0080] Step S45: Perform feature fusion on the attention spine feature map and the basic spine feature map to obtain the channel enhancement feature map.

[0081] The attentional spine feature map and the baseline spine feature map are fused to obtain the channel enhancement feature map. In this embodiment, the attentional spine feature map and the baseline spine feature map are added element-wise to obtain the channel enhancement feature map.

[0082] Step S5: Perform feature fusion on the basic spinal feature map, the global perception feature map, and the channel enhancement feature map to obtain an optimized spinal feature map.

[0083] The basic spinal feature map, the global perception feature map, and the channel enhancement feature map are fused to obtain the optimized spinal feature map, specifically including steps S51-S54.

[0084] Step S51: The global perception feature map and the channel enhancement feature map are spliced ​​together to obtain a spliced ​​spine feature map.

[0085] The global-aware feature map and the channel-enhanced feature map are concatenated along the channel dimension to obtain a concatenated spine feature map. For example, the dimensions of both the global-aware feature map and the channel-enhanced feature map are [size missing]. The size of the spliced ​​spine feature map is then... ,in .

[0086] Step S52: Based on the spliced ​​spine feature map, a gated map is obtained using a two-level convolutional network.

[0087] First, the concatenated spine feature map is input into a two-stage convolutional network to learn a spatially adaptive fusion weight map. In this embodiment, the two-stage convolutional network includes batch normalization and activation functions, and learns a spatially adaptive fusion weight map. The first-stage convolution uses a 3×3 convolution to compress the channel dimension to [missing information]. The second-stage convolution uses a 3×3 convolution to restore the channel dimension to the number of channels in the globally perceived feature map (i.e., ).

[0088] Secondly, based on the fused weight graph, a 0-1 gated graph is output using the Sigmoid function. For example, In the formula, Indicates the first in the gating diagram The element value of each element. In the fusion weight graph, the first... The element value of each element.

[0089] Step S53: Based on the gating map, the global perception feature map and the channel enhancement feature map are weighted and fused to obtain a spatial-channel fused feature map.

[0090] A gated weighted fusion strategy is adopted, using the global perception feature map as the base feature. Based on the gated map, spatial attention weighting is applied to the channel enhancement feature map. The global perception feature map and the weighted channel enhancement feature map are then added element-wise to obtain the spatial-channel fused feature map. For example, the gated map and the channel enhancement feature map have the same size; the gated map and the channel enhancement feature map are multiplied element-wise to obtain the weighted channel enhancement feature map, with the size remaining unchanged.

[0091] Step S54: Perform feature fusion on the spatial-channel fusion feature map, the global perception feature map, and the channel enhancement feature map to obtain the optimized spine feature map.

[0092] Feature fusion is performed on the spatial-channel fusion feature map, the globally perceived feature map, and the channel-enhanced feature map to obtain an optimized spine feature map. In this embodiment, learnable residual parameters are utilized. and The global perception feature map and the channel enhancement feature map are weighted separately. The weighted global perception feature map and the channel enhancement feature map are then added element-wise to the spatial-channel fusion feature map to obtain the optimized spine feature map. For example, In the formula, This indicates an optimized spine feature map. Represents the global perception feature map. Represents the channel enhancement feature map, Representation of spatial-channel fusion feature maps. Learnable residual parameters. and It allows the network to adaptively adjust the contribution ratio of the global perception feature map and the channel enhancement feature map, achieving more flexible gradient propagation, which can be optimized through training.

[0093] Through steps S3-S5, after obtaining the basic spinal feature map, attention calculation in the spatial dimension and feature recalibration in the channel dimension are performed in parallel on the basic spinal feature map. The focus is on the key anatomical structures in the basic spinal feature map that are closely related to bone density. Through feature fusion, an optimized spinal feature map is formed, which captures long-distance spatial dependencies and optimizes channel feature representation. The final output optimized spinal feature map has both global perception capability and local detail fidelity.

[0094] Step S6: Based on the optimized spinal feature map, use a global classifier to determine the basic assessment results of osteoporosis.

[0095] Based on the optimized spinal feature map, a global classifier is used to determine the basic assessment result of osteoporosis. The global classifier includes a Softmax function, which generates a scalar value between 0 and 1, representing the positive probability of osteoporosis, and serves as the basic assessment result. Steps S3-S5 perform multi-mechanism collaborative attention calculations, improving the global perception capability and local detail fidelity of the optimized spinal feature map. Combined with the global classifier, this significantly enhances the perception accuracy of bone loss areas in complex medical imaging contexts.

[0096] Step S7: Determine the final osteoporosis assessment result based on the basic osteoporosis assessment results corresponding to each vertebral slice image.

[0097] Based on the baseline osteoporosis assessment results corresponding to each vertebral slice image, the final osteoporosis assessment result is determined. This specifically includes steps S71-S73.

[0098] Step S71: Identify the target orientation category of each of the spinal slice images, and classify all the osteoporosis baseline assessment results according to the target orientation category of each spinal slice image.

[0099] Specifically, this includes steps S711-S714.

[0100] Step S711: Use the CLIP model to obtain the image feature vector corresponding to the spinal slice image.

[0101] The standardized spine slice images are input into the CLIP model to calculate the corresponding image feature vectors. In this embodiment, the CLIP model (contrastive language-image pre-training) is obtained through contrastive language-image pre-training.

[0102] Step S712: Obtain the label text vector for each of the stated directional categories.

[0103] In this embodiment, the orientation categories include axial, sagittal, and coronal. The text descriptions of each orientation category are input into the CLIP model to obtain the corresponding label text vectors. For example, the text descriptions corresponding to axial, sagittal, and coronal orientations are "a sagittal CT image of the spine," "a transverse CT image of the spine," and "a coronal CT image of the spine," respectively. These are input into the CLIP model sequentially to obtain the label text vector for each orientation category.

[0104] Step S713: Calculate the similarity between the image feature vector and each of the label text vectors to determine the target orientation category corresponding to the spine slice image.

[0105] The similarity between the image feature vector and each label text vector is calculated to determine the target orientation category corresponding to the spine slice image. In this embodiment, cosine similarity is used to calculate the similarity between the image feature vector and each label text vector. For example, In the formula, Indicates similarity. Indicates the first Image feature vectors, Indicates the first Each label text vector. From multiple similarity values ​​corresponding to the image feature vectors, the direction category corresponding to the highest similarity is selected as the target direction category for the spine slice image.

[0106] Step S714: Based on the target orientation category of each spinal slice image, divide all the basic osteoporosis assessment results into multiple assessment result sets.

[0107] For all basic osteoporosis assessment results, based on the target orientation category of each spinal slice image, spinal slice images with the same target orientation category are grouped into the same assessment result set, resulting in multiple assessment result sets. For example, orientation categories include axial, sagittal, and coronal. For spinal slice images with a target orientation category of axial, the corresponding basic osteoporosis assessment results are grouped into the axial assessment result set; for spinal slice images with a target orientation category of sagittal, the corresponding basic osteoporosis assessment results are grouped into the sagittal assessment result set; and for spinal slice images with a target orientation category of coronal, the corresponding basic osteoporosis assessment results are grouped into the coronal assessment result set.

[0108] When obtaining multiple spinal slice images according to step S1, it is difficult to directly determine the orientation category of each spinal slice image when acquiring the spinal CT image because the photographer may not understand the spinal orientation discrimination method, or a single spinal CT image may contain spinal slice images from different orientations. The CLIP model can automatically, efficiently, and accurately identify the target orientation category of the spinal slice images so as to classify the basic assessment results of osteoporosis.

[0109] Step S72: Calculate the corresponding osteoporosis unidirectional assessment result based on the osteoporosis baseline assessment result of each assessment result set.

[0110] Based on the baseline osteoporosis assessment results for each assessment result set, the corresponding unidirectional osteoporosis assessment result is calculated. For example, for the axial assessment result set, the mean of all corresponding baseline osteoporosis assessment results is calculated, that is, the mean of the positive probability of all osteoporosis is calculated, which is taken as the corresponding unidirectional osteoporosis assessment result, representing the assessment result of all axial spinal slice images.

[0111] Step S73: Based on all the osteoporosis one-way assessment results, determine the final osteoporosis assessment result by voting.

[0112] Based on all osteoporosis one-way assessment results, a voting method is used to determine the final osteoporosis assessment result. In this embodiment, based on each osteoporosis one-way assessment result and a positive threshold, a corresponding classification result is determined. Among multiple classification results, the one with the most votes is selected as the final osteoporosis assessment result. For example, if the positive threshold is 0.75, and the classification results include osteoporosis present and osteoporosis absent, and the osteoporosis one-way assessment results corresponding to the axial, sagittal, and coronal assessment result sets are 0.8, 0.9, and 0.6 respectively, then the corresponding classification results are osteoporosis present, osteoporosis present, and osteoporosis absent. That is, the number of votes for osteoporosis present is 2, and the number of votes for osteoporosis absent is 1. Therefore, osteoporosis present is selected as the final osteoporosis assessment result.

[0113] In other embodiments, the contribution of different directional categories of spinal slice images to osteoporosis assessment can be statistically quantified using historical data based on their different clinical values. Weights can then be assigned to the votes of different directional categories, and the final osteoporosis assessment result can be determined by weighted summation.

[0114] Furthermore, when outputting the final osteoporosis assessment results, the unidirectional assessment results for each osteoporosis category can also be output simultaneously, improving clinical interpretability. The unidirectional assessment results for each osteoporosis category can also be mapped to confidence levels. For example, when the positive probability of osteoporosis is greater than 0.8 or less than 0.2, it is mapped to high confidence, meaning the judgment of the presence or absence of osteoporosis is credible. When the positive probability is between 0.8 and 0.2, it is mapped to low confidence, and reminders can be set to prompt patients to seek follow-up examinations at professional medical institutions, achieving simultaneous output of diagnostic results and confidence assessments.

[0115] like Figure 2 As shown, this embodiment of the invention provides yet another method for osteoporosis assessment based on spinal CT images, including: 1) Obtain a vertebral slice image, and use the ResNet-50 model to extract features from the vertebral slice image to obtain a basic vertebral feature map; 2) A dual-path parallel attention block is embedded after the ResNet-50 model. The basic spine feature map is input into the spatial attention block and the channel spatial attention block respectively for decoupled learning to obtain the global perception feature map and the channel enhanced feature map. 3) Integrate global-aware feature maps and channel-enhanced features Figure 1 The input is a gated multi-attention block (GAM), and feature adaptive fusion is performed based on the gated attention mechanism to obtain an optimized spine feature map; 4) Input the optimized spinal feature map into the global classifier to determine the basic assessment result of osteoporosis. For example, based on the optimized spinal feature map, the global classifier outputs an osteoporosis probability of 0.906 and a non-osteoporosis probability of 0.094, thus determining the basic assessment result of osteoporosis as osteoporosis with a probability value of 0.906.

[0116] This architecture, which combines spatial-channel dual-branch enhancement and gated adaptive fusion, captures long-range spatial dependencies and optimizes channel feature representations. The resulting optimized spine feature map possesses both global perception and local detail fidelity.

[0117] Furthermore, the ResNet-50 model, dual-path parallel attention blocks, gated multi-attention blocks, and global classifier are obtained through global training. Specifically, the ResNet-50 model is pre-trained using the ImageNet dataset, and then the parameters of the pre-trained ResNet-50 model are frozen. Spine slice images and corresponding bone labels are obtained using historical case data. The dual-path parallel attention blocks, gated multi-attention blocks, and global classifier are then trained. Based on the osteoporosis baseline assessment results output by the global classifier, the bone labels are aggregated, the cross-entropy function is calculated, and the parameters of the dual-path parallel attention blocks, gated multi-attention blocks, and global classifier are globally trained using a backpropagation mechanism. By freezing the parameters of the pre-trained ResNet-50 model, generalizable image features can be directly extracted during global training, improving the convergence speed and training efficiency of global training.

[0118] This invention provides an osteoporosis assessment system based on spinal CT images, such as... Figure 3 As shown, it includes: The image acquisition module is used to acquire spinal CT images and preprocess the spinal CT images to obtain spinal slice images. The feature extraction module is used to extract features from the spinal slice image to obtain a basic spinal feature map; The attention module is used to perform spatial dimension attention calculation based on the basic spinal feature map to obtain a global perception feature map; to perform feature recalibration in the channel dimension based on the basic spinal feature map to obtain a channel enhancement feature map; and to perform feature fusion on the basic spinal feature map, the global perception feature map, and the channel enhancement feature map to obtain an optimized spinal feature map. The evaluation module is used to determine the basic assessment results of osteoporosis based on the optimized spinal feature map using a global classifier.

[0119] The above-described method and system embodiments are based on the same principles, and their related aspects can be referenced from each other to achieve the same technical effects. For specific implementation processes, please refer to the foregoing embodiments, which will not be repeated here.

[0120] This invention provides an application method for an osteoporosis assessment system, comprising: Step A: Obtain an osteoporosis assessment system; wherein the osteoporosis assessment system is determined based on the previous embodiment.

[0121] Obtain an osteoporosis assessment system, wherein the osteoporosis assessment system is the spinal CT image-based osteoporosis assessment system determined in the previous embodiment, which can be used to perform steps S1-S7.

[0122] Step B: Perform weight quantization on the osteoporosis assessment system, and convert the quantized osteoporosis assessment system into an osteoporosis computation graph; wherein the osteoporosis computation graph includes computation nodes and data flow edges.

[0123] First, the osteoporosis assessment system is weighted. In this embodiment, GPTD is used to weight the osteoporosis assessment system, quantizing FP32 precision to INT8 precision, thereby significantly reducing the system size and improving the system inference speed with minimal precision loss. In other embodiments, AWQ can also be used for weight quantization.

[0124] Secondly, the quantified osteoporosis assessment system is converted into an osteoporosis computation graph, which includes computation nodes and data flow edges. In this embodiment, the osteoporosis assessment system is in PyTorch format. The PyTorch Mobile inference framework is used to perform conversion, node fusion (step C), and inverse conversion (step D) on the osteoporosis assessment system to reduce computational overhead. Specifically, all computational operations in the osteoporosis assessment system are identified, and each computational operation is used as a computation node in the osteoporosis computation graph. For example, in step S11, grayscale conversion is one computational operation, Gaussian filtering is one computational operation, and the two first convolutional layers in step S31 are two independent computational operations, each with its corresponding computation node. Based on the data flow between each computational operation, data flow edges are determined. For example, if the grayscale image after grayscale conversion in step S11 is the input to Gaussian filtering, then a directed edge (data flow edge) is established pointing from the grayscale computation node to the Gaussian filtering computation node.

[0125] Step C: Perform node fusion on the osteoporosis calculation map to obtain a lightweight osteoporosis calculation map.

[0126] Node fusion is performed on the osteoporosis computational graph to obtain a lightweight osteoporosis computational graph. In this embodiment, node fusion is performed through node rewriting and node merging. For example, the operation nodes corresponding to steps S13 and S14 are rewritten. Step S13, non-maximum suppression, is reduced to one operation node, and step S14, double threshold lag processing, is reduced to one operation node. Currently, non-maximum suppression requires traversing each pixel, determining the corresponding neighborhood, and comparing it with the gradient magnitude of the neighborhood pixels. By rewriting the operation nodes, the double threshold lag processing operation node is executed first to filter strong and weak edge points. Then, non-maximum suppression is performed on the strong and weak edge points. Since the double threshold lag processing only performs judgment with fixed values, the number of strong and weak edge points will not exceed the number of pixels in the spinal CT image. Then, non-maximum suppression is performed, reducing the computational load. The operation nodes corresponding to step S73 are merged. Step S73 includes two operation nodes: classification result and voting to determine the final osteoporosis assessment result. The two operation nodes are merged into one operation node, and the classification result and the final osteoporosis assessment result are determined by voting within this operation node.

[0127] Step D: Obtain a lightweight osteoporosis assessment system based on the aforementioned lightweight osteoporosis calculation chart.

[0128] Based on the lightweight osteoporosis computation graph, and using the fused computation nodes and data flow edges, a lightweight osteoporosis assessment system is obtained through reverse transformation.

[0129] Step E: Package the lightweight osteoporosis assessment system into the application.

[0130] By encapsulating a lightweight osteoporosis assessment system into an application, the application gains osteoporosis assessment capabilities, enhancing the system's ease of use and versatility. For example, a user takes a raw CT image using a mobile phone with the application installed. The application automatically runs the lightweight osteoporosis assessment system, preprocessing the raw CT image, identifying multiple spinal slices, extracting and fusing features, and providing the user with the final osteoporosis assessment result (e.g., osteoporosis, confidence level: 92%). This provides instant feedback to the user (doctor or screening participant), achieving a "get results instantly with a photo" user experience.

[0131] This invention provides another method for applying an osteoporosis assessment system, including: Step F: Obtain the osteoporosis assessment system and the large-scale model of bone risk.

[0132] In this embodiment, the osteoporosis assessment system is the spinal CT image-based osteoporosis assessment system determined in the previous embodiment, which can be used to execute steps S1-S7. The large bone risk model is a large language model finely tuned by medical knowledge, based on a built-in medical knowledge graph and clinical guidelines (such as the "Guidelines for the Diagnosis and Treatment of Primary Osteoporosis"), to comprehensively reason about multi-source information.

[0133] Step G: Obtain the patient's spinal CT images and clinical text data, and use the osteoporosis assessment system and large bone risk model to determine the patient's osteoporosis risk assessment.

[0134] Specifically, based on the patient's spinal CT images, an osteoporosis assessment system is used to determine the corresponding final osteoporosis assessment result. This final osteoporosis assessment result, along with the patient's clinical text data, is input into a large-scale osteoporosis risk model to determine the patient's osteoporosis risk assessment. The patient's clinical text data includes age, gender, history of fractures, family history, and current medications. For example, if the final osteoporosis assessment result corresponding to the patient's spinal CT images shows "Osteoporosis, confidence level 70% (i.e., probability of osteoporosis 70%)", even though the probability is not high, the large-scale osteoporosis risk model will give a higher risk level to an elderly woman with a history of fractures. Ultimately, the large-scale bone risk model automatically generates a personalized diagnostic report with a standardized format and professional language. This report includes "Imaging Findings," "Final Osteoporosis Assessment Results," "Comprehensive Clinical Assessment," "Treatment Recommendations (such as calcium supplementation, vitamin D, and selection of anti-osteoporosis drugs)," and "Lifestyle Intervention and Follow-up Plan." It can be printed directly or distributed electronically, providing powerful support for doctors' decision-making and offering patients a one-stop service experience from screening to management recommendations, completely streamlining the entire process from "photographing to diagnosis to recommendations."

[0135] The application method provided in this embodiment can be applied to medical institutions, using the final osteoporosis assessment result as input to a large-scale bone risk model, integrating isolated imaging diagnostic results into a complete clinical workflow to form a service loop.

[0136] In summary, the osteoporosis assessment method, system, and system optimization method based on spinal CT images of this invention have at least one of the following beneficial effects: 1. By preprocessing spinal CT images to obtain spinal slice images, and then extracting features from these slice images to obtain baseline feature maps, optimized spinal feature maps are obtained through spatial dimension attention calculation and channel dimension feature restandardization. Using a global classifier, the final osteoporosis assessment result is determined. Existing spinal images can be used to automatically process and analyze non-standardized spinal CT images or screenshots, reducing the equipment and site requirements for osteoporosis detection, improving the convenience of osteoporosis assessment, and eliminating the need for professional medical personnel in the assessment process. This improves the accessibility of osteoporosis screening, allowing the general public to obtain osteoporosis assessment results independently. The operation is simple.

[0137] 2. Spatial attention calculation is performed on the basic spinal feature map to form a globally perceptive feature map that retains local details. Through channel-dimensional feature recalibration, a channel-enhanced feature map is obtained, which enables the feature fusion process to focus on key anatomical structures in the spine that are closely related to bone density, forming an optimized spinal feature map. This captures long-distance spatial dependencies and optimizes channel feature representation. The optimized spinal feature map has both global perception capability and local detail fidelity, improving the perception accuracy of bone loss areas in complex medical imaging backgrounds and enhancing the accuracy of osteoporosis assessment.

[0138] 3. Recognize the orientation category of spinal slice images to obtain corresponding basic osteoporosis assessment results, realize automatic osteoporosis classification, enhance the interpretability of assessment results, and provide convenience for remote diagnosis and cloud collaboration.

[0139] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0140] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for assessing osteoporosis based on spinal CT images, characterized in that, include: A spinal CT image is acquired, and the spinal CT image is preprocessed to obtain multiple spinal slice images; wherein, the spinal CT image includes one or more slice regions; the slice regions are of regular shape; Feature extraction is performed on each of the spinal slice images to obtain a basic spinal feature map; Based on the aforementioned basic spinal feature map, spatial dimension attention calculation is performed to obtain a global perception feature map; Based on the basic spinal feature map, feature recalibration is performed in the channel dimension to obtain the channel-enhanced feature map; The basic spinal feature map, the global perception feature map, and the channel enhancement feature map are fused to obtain an optimized spinal feature map. Based on the optimized spinal feature map, a global classifier is used to determine the basic assessment results of osteoporosis. Based on the baseline osteoporosis assessment results corresponding to each vertebral slice image, the final osteoporosis assessment result is determined. The basic spinal feature map, the globally perceived feature map, and the channel-enhanced feature map are fused to obtain an optimized spinal feature map, including: The global perception feature map and the channel enhancement feature map are concatenated to obtain a concatenated spine feature map. Based on the spliced ​​spine feature map, a gated map is obtained using a two-level convolutional network; Based on the gating map, the global perception feature map and the channel enhancement feature map are weighted and fused to obtain a spatial-channel fusion feature map; The spatial-channel fusion feature map, the global perception feature map, and the channel enhancement feature map are fused to obtain the optimized spine feature map. Based on the baseline osteoporosis assessment results corresponding to each vertebral slice image, the final osteoporosis assessment results are determined, including: Identify the target orientation category of each of the spinal slice images, and divide all the basic osteoporosis assessment results into multiple assessment result sets based on the target orientation category of each spinal slice image; Based on the osteoporosis baseline assessment results of each set of assessment results, calculate the corresponding osteoporosis unidirectional assessment result; The final osteoporosis assessment result is determined based on all the osteoporosis unidirectional assessment results.

2. The method according to claim 1, characterized in that, Based on the aforementioned basic spinal feature map, spatial dimension attention calculation is performed to obtain a global perception feature map, including: The basic spine feature map is input into two first convolutional layers to obtain the query matrix and the key matrix, respectively. The query matrix and the key matrix are flattened into query-space point sequences and key-space point sequences, respectively. Matrix multiplication is then performed on the query-space point sequences and the key-space point sequences to obtain a spatial attention map. The basic spine feature map is convolved to obtain a value matrix; Based on the spatial attention map, spatial attention enhancement is performed on the value matrix to obtain a spatial fusion feature map; The spatial fusion feature map and the basic spine feature map are fused to obtain the global perception feature map.

3. The method according to claim 1, characterized in that, Based on the aforementioned basic spinal feature map, feature recalibration is performed along the channel dimension to obtain a channel-enhanced feature map, including: The basic spine feature map is flattened into a channel-space point matrix, and the channel-space point matrix is ​​input into two second convolutional layers to obtain the query vector and the health vector, respectively. Perform matrix multiplication on the query vector and the key vector to obtain the channel relevance matrix; The channel correlation matrix is ​​normalized to obtain the channel attention weights; Based on the channel attention weights, attention enhancement is performed on the basic spine feature map to obtain an attention-enhanced spine feature map; The attentional spine feature map and the basic spine feature map are fused to obtain the channel enhancement feature map.

4. The method according to claim 1, characterized in that, Based on the baseline osteoporosis assessment results corresponding to each vertebral slice image, the final osteoporosis assessment result is determined, which also includes: Based on all the osteoporosis unidirectional assessment results, a voting method was used to determine the final osteoporosis assessment result.

5. The method according to claim 1, characterized in that, Identify the target orientation category for each of the said spinal slice images, including: The CLIP model is used to obtain the image feature vector corresponding to the spinal slice image; Obtain the label text vector for each directional category; Calculate the similarity between the image feature vector and each of the label text vectors to determine the target orientation category corresponding to the spine slice image.

6. The method according to claim 1, characterized in that, The spinal CT images are preprocessed to obtain multiple spinal slice images, including: Calculate the gradient magnitude and gradient direction of each pixel in the spinal CT image; Based on the gradient magnitude and gradient direction of each pixel, the thinning edge of the spinal CT image is determined using non-maximum suppression. Based on the refined edges, a dual-threshold hysteresis process is used to determine the final edges of the spinal CT image; Based on the final edge, a contour finding algorithm is used to obtain the outer contours of multiple slice regions of the spinal CT image; The spinal CT image is cropped based on the outer contour of all the slice regions to obtain multiple spinal slice images.

7. The method according to claim 1, characterized in that, For each of the said spinal slice images, feature extraction is performed to obtain a basic spinal feature map, including: The vertebral slice image is input into a basic feature extractor to obtain the basic vertebral feature map; wherein, the basic feature extractor is a ResNet-50 model.

8. An osteoporosis assessment system based on spinal CT images, characterized in that, include: The image acquisition module is used to acquire spinal CT images, preprocess the spinal CT images to obtain multiple spinal slice images; wherein, the spinal CT images include one or more slice regions; the slice region is a regular shape; The feature extraction module is used to extract features from each of the spinal slice images to obtain a basic spinal feature map; The attention module is used to perform spatial dimension attention calculation based on the basic spinal feature map to obtain a global perception feature map; to perform feature recalibration in the channel dimension based on the basic spinal feature map to obtain a channel enhancement feature map; and to perform feature fusion on the basic spinal feature map, the global perception feature map, and the channel enhancement feature map to obtain an optimized spinal feature map. The evaluation module is used to determine the basic osteoporosis assessment result based on the optimized spinal feature map using a global classifier; and to determine the final osteoporosis assessment result based on the basic osteoporosis assessment result corresponding to each spinal slice image. The basic spinal feature map, the globally perceived feature map, and the channel-enhanced feature map are fused to obtain an optimized spinal feature map, including: The global perception feature map and the channel enhancement feature map are concatenated to obtain a concatenated spine feature map. Based on the spliced ​​spine feature map, a gated map is obtained using a two-level convolutional network; Based on the gating map, the global perception feature map and the channel enhancement feature map are weighted and fused to obtain a spatial-channel fusion feature map; The spatial-channel fusion feature map, the global perception feature map, and the channel enhancement feature map are fused to obtain the optimized spine feature map. Based on the baseline osteoporosis assessment results corresponding to each vertebral slice image, the final osteoporosis assessment results are determined, including: Identify the target orientation category of each of the spinal slice images, and divide all the basic osteoporosis assessment results into multiple assessment result sets based on the target orientation category of each spinal slice image; Based on the osteoporosis baseline assessment results of each set of assessment results, calculate the corresponding osteoporosis unidirectional assessment result; The final osteoporosis assessment result is determined based on all the osteoporosis unidirectional assessment results.

9. A method for applying an osteoporosis assessment system, characterized in that, include: Obtain an osteoporosis assessment system; wherein the osteoporosis assessment system is determined based on claim 8; The osteoporosis assessment system is weighted and quantized, and the quantized osteoporosis assessment system is converted into an osteoporosis computation graph; wherein, the osteoporosis computation graph includes computation nodes and data flow edges; The osteoporosis calculation map is fused into nodes to obtain a lightweight osteoporosis calculation map; A lightweight osteoporosis assessment system is obtained based on the aforementioned lightweight osteoporosis calculation diagram. The lightweight osteoporosis assessment system is packaged into an application.

Images