Machine learning model construction method for predicting lumbar vertebra bone mineral density based on CT image

Abstract

The invention discloses a machine learning model construction method for predicting lumbar vertebra bone mineral density based on a CT image. The method comprises the following steps: step 1, image acquisition: collecting images and data including lumbar vertebra CT of a patient; step 2, image segmentation: on the cross section of each CT image, using ITK-SNAP software to carry out ROI segmentation on the cancellous bone part of the vertebral body so as to ensure that the construction of the model can reflect a cancellous bone part result which can best represent the real bone mineral density; step 3, performing deep learning by using DenseNet: realizing training and testing of a Densenet model on the basis of a Pytorch framework; and 4, performing verification by using the trained model: after the training is completed, evaluating the performance of the trained model by using a verification data set. According to the method, the patients in the training set and the test set are strictly screened, influence factors possibly causing inaccurate bone mineral density results are eliminated, results more conforming to actual bone mineral density values of vertebral bodies can be given, and the method has more clinical significance on lumbar surgery and bone mineral density treatment.

Description

Technical Field

[0001] The present invention relates to a method for constructing a machine learning model for predicting lumbar bone mineral density based on CT images, belonging to the technical field of biomedical engineering. Background Art

[0002] Osteoporosis is a systemic bone disease characterized by low bone mass, damaged bone tissue microstructure, resulting in increased bone fragility and prone to fractures. The national epidemiological survey of osteoporosis shows that the prevalence of osteoporosis in people over 50 years old is 19.2%, of which 32.1% are women and 6.9% are men; the prevalence of osteoporosis in people over 65 years old is 32.0%, of which 51.6% are women and 10.7% are men. With the aggravation of population aging in China, the prevalence of osteoporosis has risen rapidly, becoming an important public health problem.

[0003] Osteoporotic fractures (or fragility fractures) refer to fractures that occur after minor trauma (equivalent to falling from standing height or lower), which are serious consequences of osteoporosis. Patients with osteoporotic fractures may experience great pain and suffering, disability or even death, bringing huge losses to society due to this disease. For example, within 1 year after a hip fracture, 20% of patients may die from various complications; about 50% of patients become disabled and their quality of life decreases significantly. Moreover, the medical treatment and care of osteoporosis and fractures will also impose a heavy burden on the families and society of patients.

[0004] Currently, BMD measurement is a reliable and ideal method for early diagnosis of osteoporosis. The assessment methods of osteoporosis include commonly used methods such as dual-energy X-ray absorptiometry (DXA), quantitative computed tomography (QCT), quantitative ultrasound (QUS), etc. and emerging imaging technologies such as dual-layer osteoporosis. DXA is a commonly used tool for measuring BMD of the spine and hip. Considering the measurement principle of DXA, which projects the three-dimensional structure of bones onto a two-dimensional plane to calculate its areal bone density, the BMD measured by DXA includes the sum of cortical bone and cancellous bone. However, DXA cannot eliminate the influence of cortical, osteophyte and sclerosis on BMD measurement, which may cause the bone mineral density (BMD) and T value measured by DEXA to increase. Although QCT is a recognized 3D bone density assessment method and many studies have shown that the detection rate of QCT for osteoporosis is significantly higher than that of DXA, it requires calibration and standardization software, which means that there is complex processing after bone density examination. And compared with DXA, the radiation dose of QCT is much higher. Therefore, the application of QCT as a screening technology has been limited so far.

[0005] Computed tomography (CT) examinations are being applied more and more widely, especially in abdominal or lumbar spine CT. Currently, there are models for opportunistic screening of osteoporosis and osteopenia by computed tomography (CT) to determine osteoporosis and osteopenia. However, due to its simple binary classification model, limited numerical prediction accuracy, and predicting the bone condition of one site from the measurement values of another site (as the bone loss rate varies at different sites with age), its guiding significance for orthopedic clinical practice is relatively limited. Therefore, we plan to develop a model using deep learning to predict bone mineral density values based on CT scans including the lumbar spine. Moreover, CT scans including the lumbar spine are widely used in clinical practice, especially before lumbar spine surgery. Therefore, developing a bone mineral density prediction model based on CT scans including the lumbar spine will not increase the additional radiation exposure and medical costs for patients, and can greatly improve the screening rate of patients' BMD, thereby reducing osteoporosis-related complications.

[0006] The application of deep learning in radiological diagnosis has developed very rapidly. In the field of osteoporosis, there have been many works applying deep learning techniques. Therefore, this study plans to use deep learning methods to develop a bone mineral density prediction model based on CT scans including the lumbar spine, with the DEXA standard of patients after inclusion and exclusion criteria as the standard. Dual-energy X-ray absorptiometry (DXA) for bone mineral density measurement usually measures through anteroposterior or lateral X-rays. For patients with vertebral calcification or ligament calcification, the bone mineral density often shows serious deviation, resulting in the patients being unable to obtain accurate treatment. Such patients with abnormalities around the vertebrae were excluded when establishing this model, and the region of interest (ROI) only segmented and collected the cancellous bone part of the vertebrae, enabling an accurate mapping between the cancellous bone CT image information of normal patients and bone mineral density, which can be extrapolated and applied to patients with abnormal vertebral calcification or ligament ossification.

[0007] A vertebral bone mineral density classification method based on the fusion of radiomics and deep learning features with the patent application number 2022103786966 classifies vertebrae under CT images into osteoporosis, low bone mass, and normal groups, including the following steps:

[0008] S1: Establish a vertebral segmentation network based on conditional random fields (CRF) and attention guidance to obtain the cancellous bone segmentation masks of L1 and L2 vertebrae;

[0009] S2: Perform deep learning feature extraction on the feature maps after fusing L1 and L2 vertebrae through GCAM-Net, and perform radiomics feature extraction using the CT images and masks of L1 and L2 vertebrae;

[0010] S3: Use the differential evolution algorithm to extract the optimal feature set from the deep learning features, screen the radiomics features through the SVM-RFE method, finally perform feature fusion on the radiomics features and deep learning features using the maximum correlation fusion algorithm, and classify using a neural network.

[0011] In the above solution, the L1 and L2 vertebral bodies are used to extrapolate the overall bone density of the patient. However, there are still differences among different vertebral bodies, and it is difficult to truly reflect the bone density of each lumbar vertebra of the patient. The present invention includes the vertebral bodies of L1-L4 of the patient and predicts for each vertebral body, and can accurately predict the values of each vertebral body, showing its differences. At the same time, the above solution uses a three-classification method to divide the patients into a normal group, a osteopenia group, and an osteoporosis group. However, in clinical applications, even for patients in the same group, their treatment plans, including medications and surgical strategies, are still different. The prediction result of the present invention is the specific value of the lumbar bone density, which has more clinical guiding significance. In addition, the ROI is drawn for the entire vertebral body, including the cortical bone and cancellous bone, and it is impossible to accurately reflect the bone density of the cancellous bone part of the vertebral body, which is the most clinically significant. For patients with abnormal vertebral body calcification or ossification of the ligament, etc., the results cannot reflect the true bone density value. The ROI of the present invention only segments and collects the cancellous bone part of the vertebral body, so that an accurate mapping is established between the CT image information of the cancellous bone of normal patients and the bone density. Summary of the Invention

[0012] The purpose of the present invention is to provide a method for constructing a machine learning model for predicting lumbar bone density based on CT images, which is used to predict lumbar bone density from CT scans; this method uses the CT image data of patients with lumbar vertebrae after strict inclusion and exclusion, and takes the DEXA result as the output target to establish a deep learning model to predict lumbar bone density.

[0013] To achieve the above purpose, the technical solution adopted by the present invention is: a method for constructing a machine learning model for predicting lumbar bone density based on CT images, which includes the following steps:

[0014] Step 1, Image acquisition: Collect the images and data of the patients including lumbar CT.

[0015] Step 2, Image segmentation: On the cross-section of each CT image, use ITK-SNAP software to segment the ROI of the cancellous bone part of the vertebral body to ensure that the construction of the model can reflect the result of the cancellous bone part that best represents the true bone density.

[0016] Step 3, Use DenseNet for deep learning: Based on the Pytorch framework, implement the training and testing of the Densenet model.

[0017] Step 4, Use the trained model for verification: After completing the training, use the verification data set to evaluate the performance of the trained model.

[0018] Further, in the step 1, the lumbar spine CT images are obtained by scanning with a spiral 64-channel CT scanner. The CT scanning parameters are: tube voltage 120 kVp, automatic tube current, slice thickness 1.25 mm or 5 mm, and interval 0.625 mm.

[0019] Further, in the step 2, after segmentation, the images are unified. After processing, the slice thickness is 5 mm and the interval is 0.625 mm.

[0020] Further, in the step 2, first, several patients are randomly selected, and manual segmentation and modification are performed. Then, according to the outlined ROI results, a semi-automatic segmentation model for lumbar spine CT is trained. After that, the semi-automatic segmentation model is used to automatically segment the remaining images, and the final ROI results are modified manually. Finally, review and annotation modification are carried out to ensure the accuracy of the final ROI results.

[0021] Further, in the step 3, first, the previously outlined ROI part is preprocessed. Then, the model is trained using the Adam optimizer for 200 iterations, and the base learning rate is set to 1×10-4. The cross-entropy and MSE are used as loss functions for the classification and regression models, respectively.

[0022] Even further, the data preprocessing includes: first, each vertebra is cropped into an image with a voxel size of 128*128*64 according to the mask, and the pixel values of the image are normalized to the range [0, 1] by the maximum and minimum values method. At the same time, to increase the diversity of the data and prevent overfitting, horizontal flipping and rotation are used to augment the images.

[0023] Further, in the DenseNet network, the DenseBlock+Transition structure is used. The DenseBlock contains multiple layers of modules, the feature map sizes of each layer are the same, and dense connections are used between layers. The Transition module is used to connect two adjacent DenseBlocks and reduce the feature map size through Pooling.

[0024] Even further, the DenseNet network structure contains a total of 4 DenseBlocks, and the DenseBlocks are connected by Transition Layers.

[0025] Further, the training work is carried out on a workstation with the following configuration: Core i9-7900X (Intel) central processing unit and Nvidia RTX 3090.

[0026] Further, in the step 4, any data in the validation dataset is not included in the training dataset.

[0027] The beneficial effects of the present invention are as follows:

[0028] (1) For the delineation of ROI in the present invention, a method of only delineating cancellous bone is adopted. And in order for the DEXA results to accurately reflect the specific situation of cancellous bone, we have strictly screened the patients in the training set and the test set, excluding the influencing factors that may lead to inaccurate bone density results.

[0029] (2) The present invention includes the vertebral bodies of L1-L4 of the patients and makes predictions for each vertebral body, which can accurately predict the values of each vertebral body and show their differences.

[0030] (3) The regression model of the present invention can give results that are more in line with the actual bone density values of the vertebral bodies, and has more clinical significance for clinical guidance, including lumbar spine surgery and bone density treatment.

[0031] (4) Any CT containing the lumbar spine can be used to predict bone density through this model, improving the opportunistic screening for osteoporosis and osteopenia. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is the schematic diagram of the implementation of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0033] The present invention will be described in detail below with reference to the drawings and specific embodiments.

[0034] As Figure 1 , a method for constructing a machine learning model for predicting lumbar spine bone density based on CT images, the construction process is as follows:

[0035] Image acquisition: Images and data containing lumbar spine CT of patients who visited our hospital from January 2021 to September 2023 were collected. All CT images and DEXA results were collected from the Picture Archiving and Communication System (PACS) of our department. Each patient received a lumbar spine CT scan using a spiral 64-channel CT scanner. The CT scan parameters were: tube voltage 120 kVp, automatic tube current, slice thickness 1.25 mm or 5 mm, interval 0.625 mm. Due to the different thicknesses and intervals of different images, the images will be unified after segmentation, and the processed slice thickness is 5 mm and the interval is 0.625 mm. Patients who had a lumbar spine DEXA scan within two weeks were selected from the above patients, and the following operations were performed on their CT images.

[0036] Image segmentation: On the cross-section of the CT image containing the lumbar vertebrae of each patient, the ITK-SNAP software was used to segment the ROI of the cancellous bone of the vertebral body to ensure that the model construction method could reflect the results of the cancellous bone part that best represented the true bone density. To reduce the time consumption of manual annotation and ensure the accuracy of vertebral body delineation, 30 patients were randomly selected first, and three orthopedic surgeons performed segmentation and modification. According to the outlined ROI results, a semi-automatic segmentation model of lumbar CT was trained, and then the remaining images were automatically segmented using the semi-automatic segmentation model, and the final ROI results were modified by two experienced orthopedic surgeons. Finally, an orthopedic surgeon with 11 years of work experience reviewed and modified the annotations to ensure the accuracy of the final ROI results.

[0037] Deep learning using DenseNet: This study implemented the training and testing of the Densenet model based on the Pytorch framework. First, the previously outlined ROI part was preprocessed. We first cropped each vertebra into an image with a voxel size of 128*128*64 according to the mask, and normalized the pixel values of the image to the range [0,1] by the maximum and minimum values. To increase the diversity of the data and prevent overfitting, horizontal flipping and rotation were used to augment the images. The model was trained using the Adam optimizer for 200 iterations, and the initial learning rate was set to 1×10-4. The cross-entropy and MSE were used as the loss functions for the classification and regression models respectively. All experiments were carried out on a workstation with the following configuration: Core i9-7900X (Intel) central processing unit and Nvidia RTX 3090. The structure of DenseBlock+Transition was used in the DenseNet network, where DenseBlock is a module containing many layers, the feature map sizes of each layer are the same, and dense connections are used between layers. The Transition module is used to connect two adjacent DenseBlocks and reduce the feature map size through Pooling. The DenseNet-121 network structure contains a total of 4 DenseBlocks, and each DenseBlock is connected by a TransitionLayer. In this study, the softmax function and the linear activation function were used for the last layer of the network for classification and regression problems respectively.

[0038] Verification using the trained model: After completing the supervised training, the performance of the trained model was evaluated using the validation dataset. None of the data in these datasets was included in the training dataset.

[0039] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those of ordinary skill in the art should understand that the above embodiments do not limit the protection scope of the present invention in any form. Any technical solutions obtained by means of equivalent replacement and the like all fall within the protection scope of the present invention.

[0040] Parts not involved in the present invention are the same as or can be implemented by using the prior art.

Claims

1. A method for constructing a machine learning model for predicting lumbar bone density based on CT images, characterized in that The steps include: Step 1: Image acquisition: Collect the patient's lumbar spine CT images and data; Step 2: Image segmentation: On each cross-section of the CT image, use ITK-SNAP software to perform ROI segmentation of the cancellous bone of the vertebrae to ensure that the model construction can reflect the cancellous bone results that best represent the actual bone density; Step 3: Use DenseNet for deep learning: Implement the training and testing of the Densenet model based on the Pytorch framework; Step 4: Use the trained model for validation: After training, use the validation dataset to evaluate the performance of the trained model.

2. The method for constructing a machine learning model for predicting lumbar vertebrae bone density based on CT images according to claim 1, characterized in that: In step 1, the image containing the lumbar CT is obtained by scanning with a spiral 64-channel CT scanner, and the CT scanning parameters are: tube voltage 120 kVp, automatic tube current, slice thickness 1.25 mm or 5 mm, and interval 0.625 mm.

3. The method for constructing a machine learning model for predicting lumbar vertebrae bone density based on CT images according to claim 1, characterized in that: In step 2, the image is processed uniformly after segmentation, and the slice thickness after processing is 5 mm and the interval is 0.625 mm.

4. The method for constructing a machine learning model for predicting lumbar vertebrae bone density based on CT images according to claim 1, characterized in that: In step 2, several patients are first randomly selected, and the images are segmented and modified manually. Then, the semi-automatic segmentation model of lumbar CT is trained based on the outlined ROI results. After that, the remaining images are automatically segmented using the semi-automatic segmentation model, and the final ROI results are modified manually. Finally, the annotations are reviewed and modified to ensure the accuracy of the final ROI results.

5. The method for constructing a machine learning model for predicting lumbar vertebrae bone density based on CT images according to claim 1, characterized in that: In step 3, the previously drawn ROI part is firstly subjected to data preprocessing, and then the model training is performed using the Adam optimizer for 200 iterations, with the basic learning rate set to 1×10-4; the classification and regression models use cross entropy and MSE as loss functions, respectively.

6. The method for constructing a machine learning model for predicting lumbar vertebrae bone density based on CT images according to claim 5, characterized in that: Data preprocessing includes: first, cropping each vertebra into an image of 128*128*64 voxels according to the mask, and normalizing the pixel values ​​of the image to the interval [0,1] by the maximum and minimum value method; at the same time, in order to increase the diversity of the data and prevent overfitting, horizontal flipping and rotation are used to enhance the image.

7. The method for constructing a machine learning model for predicting lumbar vertebrae bone density based on CT images according to claim 1, characterized in that: The DenseNet network uses the DenseBlock+Transition structure, where DenseBlock contains multiple layers of modules, the feature map of each layer has the same size, and dense connections are used between layers; the Transition module is used to connect two adjacent DenseBlocks, and the feature map size is reduced through Pooling.

8. The method for constructing a machine learning model for predicting lumbar vertebrae bone density based on CT images according to claim 7, characterized in that: The DenseNet network structure contains a total of 4 DenseBlocks, and each DenseBlock is connected by a TransitionLayer.

9. The method for constructing a machine learning model for predicting lumbar vertebrae bone density based on CT images according to claim 1, characterized in that: The training was performed on a workstation with the following configuration: Core i9-7900X CPU and Nvidia RTX 3090.

10. The method for constructing a machine learning model for predicting lumbar vertebrae bone density based on CT images according to claim 1, characterized in that: In step 4, any data in the data set used for verification is not included in the training data set.

Images