Recognition model, training method and prediction method for osteoporosis and osteoporosis centrum compression fracture

Abstract

The invention provides an osteoporosis and osteoporosis centrum compression fracture identification model, a training method and a prediction method, and relates to a technical scheme for predicting and identifying osteoporosis and osteoporosis centrum compression fracture by repeatedly utilizing chest CT (Computed Tomography) scanning images of a patient for pulmonary nodule screening based on a deep learning technology. The first input data of the model comprises 2D image information data cut from a chest CT image slice and 2D segmented mask information data cut from a vertebral mask slice; the second input data comprises clinical information of age and gender of the patient; the multi-layer feature extraction layer of the model is used for processing first input data, and a final feature extraction graph is obtained after the first input data passes through the multi-layer feature extraction layer; and the prediction layer is used for splicing the processed final feature extraction image and second input data, and finally performing osteoporosis category prediction and vertebral compression fracture category prediction through the two parallel full-connection layers.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image recognition, and in particular to a technology related to an osteoporosis and osteoporotic vertebral compression fracture recognition model. Background Art

[0002] With the aging of the population, osteoporosis has become a major public health challenge facing our country. Osteoporosis patients usually have no obvious symptoms before a fracture occurs, so it is of great significance to identify these patients and take appropriate treatment measures in time to prevent further bone loss. Currently, bone mineral density (BMD) screening tests are the only way to identify osteoporosis patients, but the coverage of this screening is limited.

[0003] In recent years, a large number of chest CT scans have been performed each year for lung nodule screening. Using these chest CT images for opportunistic osteoporosis screening provides a great opportunity for early detection of osteopenia and osteoporosis without additional radiation. Common methods for bone density screening tests include dual-energy X-ray absorptiometry (DXA) and quantitative CT (QCT), both of which involve radiation.

[0004] Vertebral compression fracture (OVCF) is the most common type of osteoporotic fracture in osteoporosis. It mainly relies on radiologists to manually judge each vertebral segment by vertebral segment, which is not only time-consuming, but also the accuracy is related to the experience of radiologists. Therefore, it is clinically significant to automatically judge each segment of vertebral compression fracture (OVCF) based on vertebral CT images. Using the Res2Net module as the basis for feature extraction enhances the multi-scale representation capability, thereby better capturing detailed features at different levels. Summary of the invention

[0005] The present invention application provides an osteoporosis and osteoporotic vertebral compression fracture identification model, training and prediction method. Based on deep learning technology, the chest CT scan images of patients used for lung nodule screening are reused to perform a technical solution for the prediction and identification of osteoporosis and osteoporotic vertebral compression fractures.

[0006] To achieve the above object, the present invention adopts the following technical solutions: The present application provides an osteoporosis and osteoporotic vertebral compression fracture identification model, wherein the model is used to predict osteoporosis category and vertebral compression fracture category according to first input data and second input data; The first input data includes 2D image information data cut from chest CT image slices and 2D segmentation mask information data cut from vertebral mask slices; The second input data includes clinical information of the patient's age and gender; The model includes multiple layers of feature extraction layers and prediction layers; The multi-layer feature extraction layer is used to process the first input data, and obtain a final feature extraction graph after passing through the multi-layer feature extraction layer; The prediction layer includes an adaptive mean pooling layer and two parallel fully connected layers. The adaptive mean pooling layer is used to splice the processed final feature extraction map with the second input data. Finally, the osteoporosis category prediction and the vertebral compression fracture category prediction are respectively performed through the two parallel fully connected layers.

[0007] The present application also provides a method for training an osteoporosis and osteoporotic vertebral compression fracture recognition model, the method comprising the following steps: Constructing a training sample set, wherein each sample in the training sample set includes its chest CT image and clinical information of the patient's age and gender; Preprocessing first input data, wherein the first input data includes 2D image information data cut out from chest CT image slices and 2D segmentation mask information data cut out from vertebral mask slices; Preprocessing second input data, wherein the second input data includes patient age, gender, and clinical information of the sample; In the training stage, the manually labeled osteoporotic vertebral compression fracture category information of each vertebra and the osteoporosis information extracted from the patient's dual-energy X-ray bone density test report are used as the supervision information for model training; the first input data and the second input data are input into the osteoporosis and osteoporotic vertebral compression fracture recognition model for training; During the training phase of the model, the prediction confidences of all CT image slices in each vertebral segment are counted, and the preliminary prediction results of osteoporosis and osteoporotic vertebral compression fractures of each vertebral segment are obtained by accumulating and averaging the confidences; by comparing the accuracy of osteoporosis prediction of all vertebrae during the model training process, the n vertebrae with the highest accuracy are selected for confidence fusion as the final osteoporosis prediction result; In the inference stage of the model, the osteoporosis confidence of all slices of the n vertebrae selected in the training stage is integrated as the final osteoporosis prediction result; and the osteoporotic vertebral compression fracture confidence of all slices of each vertebra is integrated as the osteoporotic vertebral compression fracture prediction result of the corresponding vertebra.

[0008] At the same time, the present application also provides a method for predicting osteoporosis and osteoporotic vertebral compression fractures based on the above model, the method comprising the following steps: Obtain the target's chest CT image, patient age, and gender clinical information; Preprocessing first input data, wherein the first input data includes 2D image information data cut out from chest CT image slices and 2D segmentation mask information data cut out from vertebral mask slices; Preprocessing second input data, wherein the second input data includes patient age, gender, and clinical information of the sample; Inputting the first input data and the second input data into the trained osteoporosis and osteoporotic vertebral compression fracture recognition model to obtain a preliminary prediction result of osteoporosis and osteoporotic vertebral compression fracture of each vertebra; The osteoporosis confidence of all slices of the n vertebrae with the highest osteoporosis accuracy selected by the fusion model during the training phase is used as the final osteoporosis prediction result; The osteoporotic vertebral compression fracture confidence of all slices of each vertebra is fused to obtain the osteoporotic vertebral compression fracture prediction result of the corresponding vertebra.

[0009] This solution uses deep learning technology to achieve intelligent solutions for CT images to predict osteoporosis and osteoporotic vertebral compression fractures. A major advantage of this solution is that it does not require the use of chest CT scan images used by patients for lung nodule screening to predict and identify osteoporosis and osteoporotic vertebral compression fractures. The application of this solution can effectively replace the commonly used methods of bone density screening tests, including dual-energy X-ray absorptiometry (DXA) and quantitative CT (QCT), to avoid excessive radiation affecting the body, while greatly reducing costs. It can also make up for the problem in existing technologies that vertebral compression fractures (OVCF) mainly rely on radiologists to manually view films to judge each vertebra, further improving efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a structural diagram of an osteoporosis and osteoporotic vertebral compression fracture identification model of the present invention.

[0011] Figure 2 An example diagram of the application of an osteoporosis and osteoporotic vertebral compression fracture identification model of the present invention.

[0012] Figure 3 It is an example diagram of the sagittal center plane of the vertebral body of the chest CT image in the application of the present invention. DETAILED DESCRIPTION

[0013] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0014] The existing methods for diagnosing osteoporosis mainly include dual-energy X-ray absorptiometry (DXA) and quantitative CT (QCT).

[0015] Dual-energy X-ray absorptiometry (DXA): DXA is a low-dose X-ray imaging technique that uses two different energy peaks to differentiate between soft tissue and bone. By measuring the bone mineral content (primarily calcium), bone density can be calculated.

[0016] This method is often used for bone density measurement in the hip, spine, and other areas because it provides precise results for these areas.

[0017] DXA scans are quick and safe, involve very low radiation exposure, and are considered the gold standard for bone density measurement.

[0018] However, when measuring spinal bone density, degenerative changes in the vertebrae, arteriosclerosis, or vertebral fractures may interfere with the accuracy of the results.

[0019] Quantitative CT (QCT): QCT is a special type of computed tomography (CT) that provides three-dimensional images and separately assesses bone density of cortical bone (the hard outer part of the bone) and cancellous bone (the softer, spongy inner part of the bone).

[0020] Compared with DXA, QCT can more accurately detect changes in bone density in complex structures such as the spine and is not affected by spinal degenerative diseases or arteriosclerosis.

[0021] The radiation dose with QCT is significantly higher, which is an important consideration for patients who require frequent monitoring.

[0022] Cost issues: The cost of QCT is generally much higher than that of DXA, which may limit its widespread use.

[0023] Equipment and skills required: Not all medical facilities are equipped to perform QCT scans, and operating this technology also requires more specialized skills.

[0024] Vertebral compression fracture (OVCF) refers to compression fracture of the spinal vertebrae caused by osteoporosis. DXA and QCT can measure bone density and assist in assessing the degree of osteoporosis, thereby indirectly indicating the risk of compression fracture. However, it still needs to be combined with screening tools such as CT and magnetic resonance imaging (MRI). Relying on radiologists to manually look at the film and judge the problem of each vertebra is inefficient and costly.

[0025] After the epidemic, chest CT scans have been increasingly valued for lung nodule screening. Therefore, if chest CT images can be reused to identify osteoporotic vertebral compression fractures and assist in diagnosis, it will greatly improve efficiency, reduce diagnosis and treatment costs, and reduce radiation risks for patients.

[0026] The present application provides an osteoporosis and osteoporotic vertebral compression fracture identification model, wherein the model is used to predict osteoporosis category and vertebral compression fracture category according to first input data and second input data; The first input data includes 2D image information data cut from chest CT image slices and 2D segmentation mask information data cut from vertebral mask slices; The second input data includes patient age, gender and clinical information.

[0027] The method for preprocessing the first input data comprises the following steps: The three-dimensional chest CT image is input into a mature vertebral segmentation model to obtain vertebral segmentation mask information corresponding to the CT image, and the layer thickness of the CT image is interpolated to the same preset layer thickness value. In this embodiment, the preset layer thickness value is 1 mm.

[0028] Based on the segmentation mask information of each vertebral segment, the center coordinates of the sagittal center layer of each vertebral segment and the mask corresponding to the sagittal center layer are obtained. The sagittal center layer is usually the layer with the most complete and clear vertebral structure.

[0029] In the embodiment of the present application, the central coordinates of the sagittal center plane of each vertebral segment and the mask corresponding to the sagittal center plane are obtained. The specific processing method of this step is: like Figure 2 As shown, set the coordinate system of the CT image and mask information, x is the left-right direction (sagittal separation left-right), y is the front-back direction (coronal separation front-back), z is the up-down direction (transverse separation up-down), and the sagittal center plane is X. Mask is a three-dimensional array of (x, y, z).

[0030] Calculate the sum of the elements in mask on the (y, z)th axis to get the one-dimensional array sums.

[0031] Normalized_sums is obtained by dividing sums by the sum of sums. This step converts sums into a probability distribution form normalized_sums so that the sum of all elements of normalized_sums is equal to 1.

[0032] Calculate the cumulative sum pos of the values ​​between 0 and mask.shape[x] but not including mask.shape[x], where each value is multiplied by its corresponding normalized_sums and pos is rounded to the sagittal center plane position X.

[0033] The calculation formula is:

[0034] in: N is mask.shape[0], which is the number of layers along the left and right directions; i represents the index of the i-th level (counting starts from 0); normalized_sums[i] is the normalized weight of the i-th level.

[0035] All CT image slices within a preset distance from the sagittal center plane and 2D mask slices corresponding to the CT image slices are obtained; in this embodiment, the preset distance is 5 mm.

[0036] The preset distance is an integer multiple of the preset layer thickness value, and the CT image slice blocks of fixed size are divided according to the center coordinates of the mask; each slice block is combined with the slice blocks of the adjacent upper and lower layers to form a fixed-size three-channel 2D image information data; the 2D image information data is an array of size [64,64,3].

[0037] All 2D mask slices within a preset distance from the sagittal center plane are obtained, and a fixed-size channel of 2D segmentation mask information data is cut out from the vertebral segmentation mask according to the center coordinates of the obtained mask; the 2D segmentation mask information data is an array of size [64, 64, 1].

[0038] The 2D image information data and 2D segmentation mask information data of the slice corresponding to each vertebra are spliced ​​into four channels; and the data block of [4, 64, 64] is obtained by channel adjustment, which is the first input data.

[0039] like Figure 3 The figure below shows an example of a chest CT image, which is a sagittal center plane. The center coordinates [y, z] are marked. First, multiple sagittal planes within the range are selected, such as [x-2, x-1, x, x+1, x+2], and the above planes are cut into blocks of size [64, 64] according to the center point [y, z]. Then, the upper and lower adjacent planes are combined. For example, the x-2 plane, that is, the three planes [x-3, x-2, x-1] form an image information block of [64x64x3]. The mask of size [64x64x1] corresponding to the x-2 plane block is used as the segmentation information.

[0040] Each slice is combined with its upper and lower layers to form a 3-channel image, which not only retains the information of a single slice, but also increases the spatial correlation between the upper and lower adjacent slices, providing richer local context information to the deep learning model. The center coordinate [x, y] cuts out the image information data of fixed size [64, 64, 3] and the vertebral mask information data of [64, 64, 1] from the original image and mask, respectively, to ensure the consistency and standardization of the input data, so that the model can stably receive data in the same format, which is conducive to the convergence and generalization performance during the training process.

[0041] The normalized sums array is used to calculate the cumulative sum pos, and the position x of the sagittal center layer is rounded, and then the center coordinates [y, z] are obtained from the mask of this layer to ensure the accurate positioning of each vertebral segment in space. Based on this center position, the region of interest (ROI) related to the vertebra can be effectively located and extracted, which improves the accuracy of the model in identifying the target structure.

[0042] like Figure 1 As shown, an osteoporosis and osteoporotic vertebral compression fracture recognition model of the present application includes multiple layers of feature extraction layers and prediction layers.

[0043] The multi-layer feature extraction layer is used to process the first input data, and a final feature extraction graph is obtained after passing through the multi-layer feature extraction layer.

[0044] The prediction layer includes an adaptive mean pooling layer and two parallel fully connected layers. The adaptive mean pooling layer is used to splice the processed final feature extraction map with the second input data. Finally, the osteoporosis category prediction and the vertebral compression fracture category prediction are respectively performed through the two parallel fully connected layers.

[0045] The present application discloses an osteoporosis and osteoporotic vertebral compression fracture recognition model, wherein the feature extraction layer comprises a combination of three Res2Net module groups and pooling modules, and a fourth Res2Net module group. The feature blocks corresponding to the central feature areas of the first input data after being processed by the three combinations of the Res2Net module groups and the pooling modules are concatenated and fused with the feature maps after the fourth Res2Net module group to obtain a final feature extraction map.

[0046] The feature extraction layer mainly uses the Res2Net module, which adopts a simple and effective multi-scale processing method, and improves the multi-scale representation capability at a finer level through multiple available receptive fields. The feature map after the third Res2Net module group is used to obtain the central 6×6×6 feature area, and a feature block of size [128, 6, 6, 6] is obtained, which is then spliced ​​and fused with the feature map after the fourth Res2Net module group to obtain the final feature extraction map.

[0047] The prediction layer is a multi-task prediction branch, and the two parallel fully connected layer prediction branches respectively perform osteoporosis category prediction and vertebral compression fracture category prediction.

[0048] The design of the feature extraction layer of the model takes into account the importance of the central feature area. The final feature extraction map is obtained by splicing and fusing feature maps at different stages, which helps to improve the expressiveness of the model.

[0049] The data post-processing methods of the model include: During the training phase of the model, the prediction confidences of all CT image slices in each vertebral segment are counted, and the preliminary prediction results of osteoporosis and osteoporotic vertebral compression fractures of each vertebral segment are obtained by accumulating and averaging the confidences; by comparing the accuracy of osteoporosis prediction of all vertebrae during the model training process, the n vertebrae with the highest accuracy are selected for confidence fusion as the final osteoporosis prediction result; In the inference stage of the model, the osteoporosis confidence of all slices of the n vertebrae selected in the training stage is integrated as the final osteoporosis prediction result; and the osteoporotic vertebral compression fracture confidence of all slices of each vertebra is integrated as the osteoporotic vertebral compression fracture prediction result of the corresponding vertebra.

[0050] In the training phase, the prediction confidence of each vertebral segment slice is counted and fused, and the vertebrae with higher accuracy are selected for result fusion to ensure the reliability of the final prediction result. The same strategy is used in the inference phase to further ensure the consistency and stability of the output results.

[0051] like Figure 2 As shown, a training method for osteoporosis and osteoporotic vertebral compression fracture recognition model comprises the following steps: Constructing a training sample set, wherein each sample in the training sample set includes its chest CT image and clinical information of the patient's age and gender; Preprocess the first input data, the first input data including 2D image information data cut from chest CT image slices and 2D segmentation mask information data cut from vertebral mask slices. The preprocessing method of the first input data has been described in detail in the previous text of this embodiment and will not be repeated here.

[0052] Preprocessing second input data, wherein the second input data includes patient age, gender, and clinical information of the sample; In the training stage, the manually labeled osteoporotic vertebral compression fracture category information of each vertebra and the osteoporosis information extracted from the patient's dual-energy X-ray bone density test report are used as the supervision information for model training; the first input data and the second input data are input into the osteoporosis and osteoporotic vertebral compression fracture recognition model for training; During the training phase of the model, the prediction confidences of all CT image slices in each vertebral segment are counted, and the preliminary prediction results of osteoporosis and osteoporotic vertebral compression fractures of each vertebral segment are obtained by accumulating and averaging the confidences; by comparing the accuracy of osteoporosis prediction of all vertebrae during the model training process, the n vertebrae with the highest accuracy are selected for confidence fusion as the final osteoporosis prediction result; In the inference stage of the model, the osteoporosis confidence of all slices of the n vertebrae selected in the training stage is integrated as the final osteoporosis prediction result; and the osteoporotic vertebral compression fracture confidence of all slices of each vertebra is integrated as the osteoporotic vertebral compression fracture prediction result of the corresponding vertebra.

[0053] like Figure 2 As shown, the present application implements a method for predicting osteoporosis and osteoporotic vertebral compression fractures. The method is based on the osteoporosis and osteoporotic vertebral compression fracture model described above, and realizes the inference prediction of osteoporosis and osteoporotic vertebral compression fractures using chest CT images.

[0054] Obtain the target's chest CT image, patient age, and gender clinical information; Preprocessing first input data, wherein the first input data includes 2D image information data cut out from chest CT image slices and 2D segmentation mask information data cut out from vertebral mask slices; Preprocessing second input data, wherein the second input data includes patient age, gender, and clinical information of the sample; Inputting the first input data and the second input data into the osteoporosis and osteoporotic vertebral compression fracture recognition model trained by the training method described above, to obtain a preliminary prediction result of osteoporosis and osteoporotic vertebral compression fracture of each vertebra; The osteoporosis confidence of all slices of the n vertebrae with the highest osteoporosis accuracy selected by the fusion model during the training phase is used as the final osteoporosis prediction result; The osteoporotic vertebral compression fracture confidence of all slices of each vertebra is fused to obtain the osteoporotic vertebral compression fracture prediction result of the corresponding vertebra.

[0055] One of the major advantages of this solution is that it does not require the use of chest CT scan images used by patients for lung nodule screening to predict and identify osteoporosis and osteoporotic vertebral compression fractures. The application of this solution can effectively replace the commonly used methods of bone density screening tests, including dual-energy X-ray absorptiometry (DXA) and quantitative CT (QCT), to avoid excessive radiation affecting the body, while greatly reducing costs. It can also make up for the problem that vertebral compression fractures (OVCF) in the existing technology mainly rely on radiologists to manually view the film to judge each vertebra, further improving efficiency.

[0056] The clinical data involved in this application have been obtained with full consent and authorization, and the collection, use and processing of relevant information must comply with the relevant laws, regulations and standards of the relevant countries and regions.

Claims

1. A model for identifying osteoporosis and osteoporotic vertebral compression fractures, characterized in that: The model is used to predict the osteoporosis category and the vertebral compression fracture category according to the first input data and the second input data; The first input data includes 2D image information data cut from chest CT image slices and 2D segmentation mask information data cut from vertebral mask slices; The second input data includes clinical information of the patient's age and gender; The model includes multiple layers of feature extraction layers and prediction layers; The multi-layer feature extraction layer is used to process the first input data, and obtain a final feature extraction graph after passing through the multi-layer feature extraction layer; The prediction layer includes an adaptive mean pooling layer and two parallel fully connected layers. The adaptive mean pooling layer is used to splice the processed final feature extraction map with the second input data. Finally, the osteoporosis category prediction and the vertebral compression fracture category prediction are respectively performed through the two parallel fully connected layers.

2. The osteoporosis and osteoporotic vertebral compression fracture identification model according to claim 1, characterized in that: The method for preprocessing the first input data comprises the following steps: Input the three-dimensional chest CT image into a mature vertebral segmentation model to obtain the vertebral segmentation mask information corresponding to the CT image, and interpolate the layer thickness of the CT image to the same preset layer thickness value; Based on the segmentation mask information of each vertebral segment, the center coordinates of the sagittal center layer of each vertebral segment and the mask corresponding to the sagittal center layer are obtained; Acquire all CT image slices within a preset distance from the sagittal center plane, and 2D mask slices corresponding to the CT image slices; The preset distance is an integer multiple of the preset layer thickness value, and the CT image slice blocks of fixed size are segmented according to the center coordinates of the mask; each slice block is combined with the slice blocks of the adjacent upper and lower layers to form a fixed-size three-channel 2D image information data; Acquire all 2D mask slices within a preset distance from the sagittal center plane, and cut out a channel of 2D segmentation mask information data of a fixed size from the vertebral segmentation mask according to the center coordinates of the acquired mask; The 2D image information data and 2D segmentation mask information data of the slice corresponding to each vertebra are spliced ​​into the first input data of four channels.

3. The osteoporosis and osteoporotic vertebral compression fracture identification model according to claim 2, characterized in that: The preset layer thickness value is 1mm, and the preset distance is 5mm; The 2D image information data is an array of size [64, 64, 3]; The 2D segmentation mask information data is an array of size [64, 64, 1]; The first input data is a data block of size [64, 64, 4] obtained by splicing 2D image information data and 2D segmentation mask information data in the third dimension, and a data block of size [4, 64, 64] is obtained by channel adjustment.

4. The osteoporosis and osteoporotic vertebral compression fracture identification model according to claim 1, characterized in that: The feature extraction layer includes three combinations of Res2Net module groups and pooling modules, and a fourth Res2Net module group. The feature blocks corresponding to the central feature areas of the first input data after being processed by the three combinations of Res2Net module groups and pooling modules are spliced ​​and fused with the feature maps after the fourth Res2Net module group to obtain a final feature extraction map.

5. The osteoporosis and osteoporotic vertebral compression fracture identification model according to claim 1, characterized in that: The prediction layer is a multi-task prediction branch, and the two parallel fully connected layer prediction branches respectively perform osteoporosis category prediction and vertebral compression fracture category prediction.

6. The osteoporosis and osteoporotic vertebral compression fracture identification model according to claim 1, characterized in that: The data post-processing methods of the model include: During the training phase of the model, the prediction confidences of all CT image slices in each vertebral segment are counted, and the preliminary prediction results of osteoporosis and osteoporotic vertebral compression fractures of each vertebral segment are obtained by accumulating and averaging the confidences; by comparing the accuracy of osteoporosis prediction of all vertebrae during the model training process, the n vertebrae with the highest accuracy are selected for confidence fusion as the final osteoporosis prediction result; In the inference stage of the model, the osteoporosis confidence of all slices of the n vertebrae selected in the training stage is integrated as the final osteoporosis prediction result; and the osteoporotic vertebral compression fracture confidence of all slices of each vertebra is integrated as the osteoporotic vertebral compression fracture prediction result of the corresponding vertebra.

7. A training method for an osteoporosis and osteoporotic vertebral compression fracture recognition model, characterized in that: The method comprises the following steps: Constructing a training sample set, wherein each sample in the training sample set includes its chest CT image and clinical information of the patient's age and gender; Preprocessing first input data, wherein the first input data includes 2D image information data cut out from chest CT image slices and 2D segmentation mask information data cut out from vertebral mask slices; Preprocessing second input data, wherein the second input data includes patient age, gender, and clinical information of the sample; In the training stage, the manually labeled osteoporotic vertebral compression fracture category information of each vertebra and the osteoporosis information extracted from the patient's dual-energy X-ray bone density test report are used as the supervision information for model training; the first input data and the second input data are input into the osteoporosis and osteoporotic vertebral compression fracture recognition model described in any one of claims 1 to 5 for training; During the training phase of the model, the prediction confidences of all CT image slices in each vertebral segment are counted, and the preliminary prediction results of osteoporosis and osteoporotic vertebral compression fractures of each vertebral segment are obtained by accumulating and averaging the confidences; by comparing the accuracy of osteoporosis prediction of all vertebrae during the model training process, the n vertebrae with the highest accuracy are selected for confidence fusion as the final osteoporosis prediction result; In the inference stage of the model, the osteoporosis confidence of all slices of the n vertebrae selected in the training stage is integrated as the final osteoporosis prediction result; and the osteoporotic vertebral compression fracture confidence of all slices of each vertebra is integrated as the osteoporotic vertebral compression fracture prediction result of the corresponding vertebra.

8. The method for training an osteoporosis and osteoporotic vertebral compression fracture recognition model according to claim 7, characterized in that: The method for preprocessing the first input data comprises the following steps: Input the three-dimensional chest CT image into a mature vertebral segmentation model to obtain the vertebral segmentation mask information corresponding to the CT image, and interpolate the layer thickness of the CT image to the same preset layer thickness value; Based on the segmentation mask information of each vertebral segment, the center coordinates of the sagittal center layer of each vertebral segment and the mask corresponding to the sagittal center layer are obtained; Acquire all CT image slices within a preset distance from the sagittal center plane, and 2D mask slices corresponding to the CT image slices; The preset distance is an integer multiple of the preset layer thickness value, and the CT image slice blocks of fixed size are segmented according to the center coordinates of the mask; each slice block is combined with the slice blocks of the adjacent upper and lower layers to form a fixed-size three-channel 2D image information data; Acquire all 2D mask slices within a preset distance from the sagittal center plane, and cut out a channel of 2D segmentation mask information data of a fixed size from the vertebral segmentation mask according to the center coordinates of the acquired mask, the said; The 2D image information data and 2D segmentation mask information data of the slice corresponding to each vertebra are spliced ​​into the first input data of four channels.

9. A method for predicting osteoporosis and osteoporotic vertebral compression fractures, characterized in that: The method comprises the following steps: Obtain the target's chest CT image, patient age, and gender clinical information; Preprocessing first input data, wherein the first input data includes 2D image information data cut out from chest CT image slices and 2D segmentation mask information data cut out from vertebral mask slices; Preprocessing second input data, wherein the second input data includes patient age, gender, and clinical information of the sample; Inputting the first input data and the second input data into the osteoporosis and osteoporotic vertebral compression fracture recognition model trained by the training method of any one of claims 6 or 7 to obtain a preliminary prediction result of osteoporosis and osteoporotic vertebral compression fracture of each vertebra; The osteoporosis confidence of all slices of the n vertebrae with the highest osteoporosis accuracy selected by the fusion model during the training phase is used as the final osteoporosis prediction result; The osteoporotic vertebral compression fracture confidence of all slices of each vertebra is fused to obtain the osteoporotic vertebral compression fracture prediction result of the corresponding vertebra.

10. The method for predicting osteoporosis and osteoporotic vertebral compression fractures according to claim 9, characterized in that: The method for preprocessing the first input data comprises the following steps: Input the three-dimensional chest CT image into a mature vertebral segmentation model to obtain the vertebral segmentation mask information corresponding to the CT image, and interpolate the layer thickness of the CT image to the same preset layer thickness value; Based on the segmentation mask information of each vertebral segment, the center coordinates of the sagittal center layer of each vertebral segment and the mask corresponding to the sagittal center layer are obtained; Acquire all CT image slices within a preset distance from the sagittal center plane, and 2D mask slices corresponding to the CT image slices; The preset distance is an integer multiple of the preset layer thickness value, and the CT image slice blocks of fixed size are segmented according to the center coordinates of the mask; each slice block is combined with the slice blocks of the adjacent upper and lower layers to form a fixed-size three-channel 2D image information data; Acquire all 2D mask slices within a preset distance from the sagittal center plane, and cut out a channel of 2D segmentation mask information data of a fixed size from the vertebral segmentation mask according to the center coordinates of the acquired mask, the said; The 2D image information data and 2D segmentation mask information data of the slice corresponding to each vertebra are spliced ​​into the first input data of four channels.

Images