TWI743969B - System and method for detection of thoracic and lumbar vertebral fractures - Google Patents

Abstract

Provided herein is a method for detection of thoracic and lumbar vertebral fractures on plain lateral radiographs (PLRs), comprising the steps of data preprocessing, developing model ensemble, and marking vertebral fractures according to position detected by data pre-processing and prediction result of model ensemble.

Description

System and method for detecting chest and lumbar vertebral fractures

The invention relates to a system and method for detecting chest and lumbar vertebral fractures on plain lateral radiographs (PLRs).

Vertebral fractures (VFs) are the most common fractures in osteoporosis-related fragility fractures. After vertebral fractures, they can cause back pain, kyphosis, and poor quality of life, which in turn can cause fragility fractures in other parts of the body. Increased mortality is associated (1,2). Although the proportion of clinical diagnosis has increased with the increase in the severity of vertebral deformation, due to clinically asymptomatic and no back pain (4-6), some radiographic vertebral fractures (VFs) are clinically Misdiagnosed (3), so-called "clinically silent" vertebral fractures (VFs) (7). However, despite the importance of detecting vertebral fractures (VFs), some of these fractures have not been fully diagnosed. In a multicenter, multinational study, the global incidence of false negative diagnoses of vertebral fractures (VFs) was 34% (8). In summary, it is important for physicians to provide early diagnosis and treatment of vertebral fractures (VFs) and avoid delays in treatment.

The latest advances in artificial intelligence (AI) deep learning show significant progress in computer-aided diagnosis of non-medical images, and AI is considered the next generation of technological revolution (9). Some authors reported that the artificial intelligence module system automatically detects vertebral fractures with a sensitivity of 95.7% on Computed Tomography (CT) scans (10).

However, compared to computed tomography (CT) scans, plain X-ray radiography is still the first-line screening and initial inspection method, although the detection accuracy of vertebral fractures (VFs) in computed tomography (CT) scans is higher and the visualization effect is better. Good (13), plain X-ray radiography still has the advantages of high availability, less radiation, and wide medical coverage (11,12). In addition, in the emergency room or peak time of clinical visits in smaller hospitals, radiologists or orthopedists cannot always consult and read the radiographs immediately (9). In this case, AI assists in the diagnosis of vertebral fractures ( VFs) identification may be one of the alternative methods. So far, few literatures have studied the application of AI deep learning module in the detection of vertebral fractures (VFs) through plain X-ray radiographic images.

There is a need to develop a more effective and time-saving system and method for automatic detection of chest and lumbar vertebral fractures (VFs) based on plain radiography (PLRs).

The purpose of this disclosure is to provide a method for detecting chest and lumbar vertebral fractures, including: obtaining a plain X-ray lateral radiographic image; preprocessing the data, including: providing the plain X-ray lateral radiographic image; X-ray profile image, using object detection algorithm to locate the spine; Preprocess the plain X-ray profile image by Gaussian blur, median filter, and Contrast Adaptive Histogram Equalization; Redefine the width and height of the plain X-ray profile image by adjusting the method; and optimize the positioning of the plain X-ray profile image; develop model integration, including: defrosting and hiding according to transfer learning methods and ImageNet training weights Layer; and the characteristics of the plain X-ray side radiographic image preprocessed by the classification model; and the position detected based on the data preprocessing and model integration prediction results to indicate the chest and lumbar vertebral fractures.

The method according to claim 1, wherein the method for optimizing the positioning of the plain X-ray side-radiation image is selected from (1) image size adjustment and no image preprocessing; (2) image size adjustment and self-comparison The adaptive histogram equalization; (3) maintain the original image ratio and no image preprocessing; and (4) maintain the original image ratio and implement the group of adaptive histogram equalization.

The method according to claim 2, wherein the method of optimizing the positioning of the plain X-ray lateral radiation image is to adjust the image size and without image preprocessing.

The method according to claim 1, wherein the model integration is pre-trained in a fast artificial intelligence (fast.AI) library.

The method according to claim 4, wherein the model integration is pre-training of a combination of ResNet34, DenseNet121, and DenseNet201 models.

The method according to claim 1, wherein the step of preprocessing the data further includes: maintaining the image ratio and filling the image with color blocks to correspond to the fracture classification model in a 224*224 pixel output format.

When read in conjunction with the accompanying drawings, one will better understand the foregoing content of the invention and the following detailed description of the invention. To illustrate the present invention, the presently preferred embodiments are shown in the drawings.

In the diagram: Figure 1 provides an illustration of the original model of AI deep learning according to the present invention. Design YOLOv3 and its classification model in the fast artificial intelligence (fast.AI) library to detect vertebral fractures at the level of each spine. The classification ensemble model is composed of ResNet34, DenseNet121, and DenseNet201, and these models are designed for crack recognition. These models are combined and defined as artificial intelligence deep learning ensemble model (AIDLEM).

Figure 2 shows a collective module composed of ResNet34, DenseNet121, and DenseNet201. Compared with ResNet34, DenseNet121, and DenseNet models, the collective module shows higher accuracy and stability.

Figure 3 shows the receiver operating characteristic (ROC) curve and the area under the curve (AUC) of each fracture severity level. The area under the curve (AUC) for fractures of grade 1, grade 2, and grade 3 were 0.917, 0.989, and 0.990, respectively. AIDLEM's difference between grade 2 and grade 3 fractures is better than the difference between grade 1 fractures.

Figure 4 provides an illustration of vertebral fracture detection performed by AIDLEM. The output is an accurate result, and the red box is the position of the fractured vertebral body with the predicted probability.

Figures 5a, 5b, 5c, and 5d provide illustrations of the detection of vertebral fractures through AIDLEM, including Figure 5a: false positive result: severe osteoporosis (DEXA T score=-3.3); Figure 5b: false positive result: lung lines Or diaphragm; Figure 5c: false negative result: mild (grade I) vertebral fracture; and Figure 5d: false negative result: more severe fracture (grade III) vertebral fracture (fractured vertebral body height reduced by more than 80 %) (vertebral plana flat spine).

In conjunction with the drawings, the embodiments of this specification can clearly reveal the above-mentioned and other technical contents, features and effects of the present invention. Through the description of the specific embodiments, those skilled in the art can further understand the technical means and effects used to achieve the above-mentioned objectives. In addition, since those skilled in the art can clearly understand and implement the content disclosed in the present invention, all equivalent changes or modifications that do not deviate from the concept of the present invention are covered by the claims.

The present invention provides a method for detecting chest and lumbar vertebral fractures (VFs) on plain X-ray profile radiography (PLRs) through an integrated artificial intelligence deep learning model (AIDLEM).

AIDLEM can be used clinically to identify vertebral fractures (VFs) through plain X-ray lateral radiography (PLRs) of the spine, with high accuracy, sensitivity, and specificity. The physician must upload the original DICOM version of plain X-ray profile radiography (PLRs) in advance without any image preprocessing, which may be more effective and time-saving for the physician.

It has been previously reported that an automatic computerized computer system has been developed for the detection of vertebral fractures (VFs) on computerized tomography (CT) scans of the thoracic and lumbar spine, with an accuracy rate of 95% and a sensitivity of 95.7% (10). However, radiation exposure is still a problem. The radiation dose of a computed tomography (CT) scan of the spine may depend on the computed tomography (CT) scanner and its standard methods (22,23). Generally, the calculated effective doses (CED) in the computed tomography (CT) scans of the thoracic and lumbar spine are approximately 10 and 5.6 mSv, respectively (22, 24). The calculated effective dose (CED) of the waist AP and side plain X-ray radiographic images were 2.2 and 1.5 mSv, respectively (22). Compared with radiographic images of the lumbar spine, the radiation dose of a computed tomography (CT) scan of the spine is much higher. for For radiation exposure and medical treatment, physicians usually use plain X-ray radiography as the first method to evaluate various spinal diseases including vertebral fractures (VFs).

We combine YOLO with an integrated classification model to detect and distinguish cracks separately. In addition, since we only have 941 radiographic images, we use transfer learning and data augmentation techniques to overcome the problem of less training materials and train new models with limited clinical data. Taking into account the difference between the number of training data and the number of model parameters, we use the weights trained on the ImageNet dataset as the initial value of the weights for this study, and then use the data from the Taipei President for transfer learning. The training process is divided into two stages: freezing (freeze) and gradual thawing (unfreeze) to adjust the model parameters. We first assume that the weights obtained by ImageNet training can extract sufficiently good image features. At this time, the weights of all convolutional network layers are frozen, and the weights that are randomly initialized due to different categories are adjusted. However, the images of the ImageNet dataset are still different from radiological images. Therefore, we gradually unfreeze and train the remaining layers of the Convolutional Neural Network (CNN) so that the model can learn the features of radiological images while avoiding overfitting, and obtain better results. All frozen layers (accuracy 86%) have better results (accuracy 92.20%).

Factors related to lung markings, thoracic spine diaphragm, degree of osteoporosis (DEXA T score≦-2.5), and bean pot effect (26) may affect the recognition of AIDLEM. When acquiring plain X-ray side radiography (PLRs), technicians must ensure that each spine is parallel to the projector to avoid deviation caused by the tilt of the projector. The biconcave normal appearance of the endplate associated with aging is called the "bean pot effect" (26). In addition, aging-related spinal changes have also been reported (27,28), and their appearance may be similar to those reported in osteoporotic vertebral fractures (VFs). Degenerative scoliosis may have multiple endplate orientations on each spine in the same radiographic image, and the projector can be tilted on multiple spine levels even in the proper patient position.

Genant fracture classification may bias these results. Guglielmi et al. (29) reported that a special 6-point morphological study has a failure rate of greater than 20% in the identification of T5-T10 fractures. Lung markers may be one of the factors explaining these results. Burns et al. reported that in the automatic detection of computerized tomography (CT) images, Genant grade 3 fractures are more sensitive than Genant grade 1 and grade 2 fractures (10). Therefore, a more deformed spine with a higher Genant grade can be easily detected on a computerized tomography (CT) image.

Human radiologists have a poor diagnosis of Genant grade 1 vertebral fractures (VFs) (30,31). For Genant grade 3 vertebral fractures (VFs), the vertebral body is severely deformed locally, causing the upper and lower endplates to close. Due to the bounding box of severely deformed vertebral fractures (VFs), YOLOv3 may define the lower endplate of the head vertebrae and the upper endplate of the caudate vertebrae. AIDLEM's misidentification may identify severe vertebral fractures (VFs) as normal spine. Moreover, severe deformities not only cause kyphosis in the sagittal plane, but also degenerative scoliosis in the coronal plane (32).

Only plain X-ray profile radiographs (PLRs) are used to identify vertebral fractures (VFs). Radiographic images are projected before and after plain X-rays. Some burst fractures are characterized by widening of the intervertebral space (33), and plain X-ray lateral radiography (PLRs) at the thoracolumbar junction may not be clear Fractures.

According to the present invention, AIDLEM has better performance in the detection of lumbar spine fractures (VFs), and its accuracy, sensitivity, and specificity are 92.41% (260/282; 95% CI, 92.15-92.67%) and 91.23% ( 129/141; 95% CI, 90.83-91.62%), and 93.61% (132/141; 95% CI, 93.27-93.92%). Regarding the statistical data of misclassification, the most common causes of false positives and negatives were osteoporosis (42.9%) and Genant grade 1 fractures (68.8%). Regarding the severity of osteoporosis, AIDLEM is difficult to distinguish severe osteoporosis (T score≦-2.5), and its specificity is 91.13% (260/285; 95% CI, 90.91-91.35), which is lower than non-severe osteoporosis The specificity is 94.85% (222/234; 95% CI, 94.76-94.94). Regarding the grades of Genant fractures, the classification AUCs of grade 1, grade 2, and grade 3 vertebral fractures (VFs) are 0.917, 0.989, and 0.990, respectively.

The present invention is further illustrated through the following examples, which are provided for illustration rather than limitation.

Example

Materials and Methods

The data set of 941 patients (mean age 76±12.4; 1,111 cases of thoracic and lumbar spine fractures (VFs)) were retrospectively included, and AIDLEM training (n=565) and verification ( n=188), and test (n=188). Perform target detection, data preprocessing of plain X-ray profile radiography (PLRs), transfer learning, and integrated learning to improve accuracy and develop stable integrated models. Accuracy, sensitivity, specificity, 95% confidence interval, ROC (receiver operating characteristic) curve, and area under the curve (AUC) are used to analyze AIDLEM.

1. Research object

The retrospective study does not require informed consent, and in this study we analyzed previously obtained images. From January 2016 to December 2018, all lumbar and thoracolumbar vertebral X-ray radiographic images are included.

The keywords of "compression" or "burst" fractures investigated in the radiographic report included databases used for training, verification, and testing, and the patients were over 60 years old. The classification of fractures is based on the Genant classification (14), which is considered to be one of the classifications of spine fractures assessed by plain radiographs (PLRs).

2. X-ray radiography technology

The X-ray method is described as follows: AP and profile radiographic images of the lumbar and thoracolumbar spine are performed with an X-ray high-pressure generator (SHIMADZU, UD150B-40). The voltage is 94kVp, the average current is 56mAs, and the duration is 360 milliseconds, depending on the patient's physical habits.

3. Data preprocessing

In order to identify thoracic and lumbar vertebral fractures (VFs), we need to first locate the vertebral body in plain X-ray lateral radiography (PLRs). YOLO (You Only Look Once) is one of the object detection algorithms proposed in 2015. The YOLOv1 (15), YOLOv2 (16), and YOLOv3 (17) generations were developed respectively. The main advantage of the YOLO method is that it can quickly predict the location, category, and bound box of an object in a round of CNN (Convolution Neural Network) and perform regression. In addition, the breakthrough of the YOLOv3 method is more suitable for the detection of severe vertebral fractures (VFs) or deformed small objects (17). Therefore, we use the YOLOv3 model to detect vertebral bodies (Figure 1).

Because the thoracic vertebrae (T1-T12) are located near the lungs and diaphragm, these anatomical positions may affect the clarity and contrast of plain radiographs (PLRs). The spine detection processed by YOLOv3 may result in textures, different brightness, and different image sizes. To overcome these shortcomings, we use Gaussian blur, median filter, and Contrast Adaptive Histogram Equalization (CLAHE) image preprocessing to reduce noise and adjust brightness and contrast. Secondly, we use the Resize method to redefine the width and height of the image, or keep the original image ratio, and fill it with color blocks to conform to the 224*224 pixel crack classification model input format.

We tried four different data preprocessing methods to optimize plain X-ray profile radiography (PLRs) positioned by YOLOv3: (1) Re-adjust the image size without image pre-processing (2) Re-adjust the image size and compare Adaptive Histogram Equalization (CLAHE) (18) maintains the original image aspect ratio and does not Perform image preprocessing (4) to maintain the original image aspect ratio and perform contrast adaptive histogram equalization (CLAHE). Combine the above data preprocessing method with ResNet and DenseNet's classification model framework to determine whether the data preprocessing method has the best performance. According to the preliminary preprocessing results, it can be found that data preprocessing is performed while maintaining the original image aspect ratio, and no image preprocessing has the best performance than other methods. Therefore, we choose this method as the main program for data preprocessing (see Figure 1).

4. Model integration and AI deep learning model development

After data preprocessing, we developed a classification model to detect vertebral fractures (VFs) (Figure 1). In the deep learning library, we chose the fast.AI library based on the PyTorch framework. The advantage of this model library is that it contains common deep learning models and provides software packages for optimizing and adjusting parameters. Combine this model and define it as an Artificial Intelligence Deep Learning Integrated Model (AIDLEM) (see Figure 1).

Due to the limitation of a small number of medical images, this strategy uses a pre-trained model in the fast artificial intelligence (fast.AI) library. According to the transfer learning method (19) and ImageNet training weights, we unfreeze the hidden layer and make the classification model learn the features of medical images. When choosing a pre-training model, we chose a simpler model, ResNet34. In addition, we refer to the CheXNet model, which also performs well on chest radiographs (20), and try its basic frameworks DenseNet121 and DenseNet201. Faced with different data sets, the performance of each model is not stable. Therefore, we apply ensemble learning to combine the prediction results of the three models to clarify more accurate and stable results (see Figure 2). Finally, according to the position detected by the YOLOv3 method and the prediction result of the integrated model, the vertebral fractures (VFs) are marked through plain X-ray profile radiography (PLRs).

5. Statistical analysis

We use analysis methods such as accuracy, sensitivity, specificity, 95% confidence interval, and ROC curve to evaluate the performance of AIDLEM in anatomical locations, osteoporosis, and Genant fracture classification.

In order to obtain point estimates and interval estimates, we used the bootstrap method to perform 1000 replacements and resampling of the test data, and then SPSS software was used to calculate the average accuracy, average sensitivity, average specificity, and 95% confidence of the T test. Interval. In addition, we set different classification probability thresholds for AIDLEM, draw ROC (receiver operating characteristic) curve, calculate area under the curve (AUC) to study model performance.

接收者操作特徵曲線(ROC曲線)：設置不同的分類閾值，記錄對應的靈敏度與1-特異性，然後繪製ROC曲線

formula:

Receiver operating characteristic curve (ROC curve): set different classification thresholds, record the corresponding sensitivity and 1-specificity, and then draw the ROC curve

：樣本平均值

: Sample average

S : sample standard deviation

n : sample size

t _0.025 : degree of freedom = n -1 95% confidence interval (95% CI): use the bootstrap method to replace the test data to perform 1000 re-sampling, and then use SPSS to calculate the 95% confidence interval through T test.

6. Results

6.1 Demographic data of the data set

AIDLEM included 941 patients with plain radiographs (PLRs) of 1,111 spine fractures (VFs) in the chest (T1-T12) or lumbar (L1-L5). Table 1 illustrates demographic data and fracture location.

6.2 Data set

A retrospective study of plain radiographs (PLRs) of the thoracolumbar and lumbar spine in 941 patients. However, since the thoracic spine is covered by lung markings and diaphragm, we investigate the edges of the thoracic spine The fact that the margin is more unclear than the lumbar spine. Therefore, we divided all plain X-ray lateral radiographs (PLRs) into thoracic vertebrae (T1-T12), lumbar vertebrae (L1-L5), and the entire spine (T1-L5), and assessed whether the anatomical location is a vertebral fracture ( VFs) have a role in detection.

After collecting these plain X-ray profile radiography (PLRs), the number of records related to training data, verification data, and test data were 565, 188, and 188 (Table 2). The number of normal vertebrae is 5 to 6 times that of fractured vertebrae, so we use sub-sampling method to solve the problem of data imbalance, and then establish a fracture classification model.

6.3 Regular partition performance

Table 3 lists the test results of the thoracic spine (T1-T12), lumbar spine (L1-L5), and the entire spine (T1-L5). It is found that AIDLEM has the best performance in the detection of lumbar vertebral fractures (VFs). The average accuracy, sensitivity, and specificity of 1,000 bootstraps for detection of lumbar spine fractures (VFs) were 92.41% (260/282; 95% CI, 92.15-92.67%) and 91.23% (129/141; 95% CI, 90.83-91.62%), and 93.61% (132/141; 95% CI, 93.27-93.92%). The results were similar for the thoracic spine (T1-T12) and the entire spine (T1-L5) (Table 3).

In addition, the confusion matrix of the thoracic spine (T1-T12) and lumbar spine (L1-L5) (Table 4) shows that the AIDLEM of the lumbar spine is fairer than the AIDLEM of the thoracic spine, and this model tends to predict fractures.

In addition, we divided the lumbar spine data set into two subgroups according to the severity of osteoporosis (DEXA T score>-2.5 for non-severe osteoporosis group or DEXA T score ≦-2.5 for severe osteoporosis group).

The sensitivity of detection of non-severe and severe osteoporotic lumbar spine fractures (VFs) was 83.1% (39/47; 95% CI, 82.76-83.45%) and 96.69% (59/62; 95% CI, 96.55-96.83%), respectively ). The specificities of non-severe and severe osteoporotic lumbar spine fractures (VFs) were 94.85% (222/234; 95% CI, 94.76-94.94%) and 91.13% (260/285; 95% CI, 90.91-91.35, respectively) %) (Table 3), which indicates that the severity of osteoporosis plays an important role in the detection of vertebral fractures (VFs) of the AIDLEM lumbar spine. Vertebral fractures (VFs) with severe osteoporosis are easier to detect than vertebral fractures (VFs) with non-severe osteoporosis, which indicates a higher sensitivity. On the other hand, AIDLEM sometimes makes mistakes in detecting normal spine with severe osteoporosis as vertebral fractures (VFs), and therefore exhibits low specificity.

According to the Genant classification (14), vertebral fractures (VFs) are divided into: Grade 1 (20%-25% height reduction), Grade 2 (25%-40% height reduction), and Grade 3 (>40% height reduction). The results of fracture classification are shown in Table 3. We investigated the performance of this model at level 2 and level 3 better than level 1. The accuracy and sensitivity of grade 2 fracture detection were 94.94% (99/104; 95% CI, 94.74-95.14%) and 99.21% (53/54; 95% CI, 98.41-100.0%), respectively. In addition, in Figure 3, the ROC curves of all fracture types are shown. The AUC areas of Grade 2 and Grade 3 are close and higher, which means that the higher the severity of vertebral fractures, the easier the model can predict.

6.4 Correct and wrong classification

Figure 4 shows an example of accurate vertebral fractures (VFs) detection through AIDLEM. The output image shows the location and likelihood of the fracture. The doctor only needs to input the original radiological image, the model will preprocess the data, automatically determine the fracture, and then output the result to help the orthopedist make a judgment or diagnosis. When used in the clinic, the program that directly uploads without manual preprocessing is more efficient.

However, the model also has some misclassifications. For misclassified vertebral fractures (VFs), we used the Grad-CAM (21) method to draw a heat map to visually attract model attention. Inspiration and the location of the hotspot area are indicated in red. The common causes of false positives are severe osteoporosis (DEXA T score≦-2.5) (42.9%) and lung markers (23.8%). The Grad-CAM heat map shows that the AI model focuses on the concave appearance of the spine (Figure 5(a)) and the lung markings or the local area of the diaphragm (Figure 5(b)), and then predicts the normal spine as a fracture. In addition, false negatives often occur in grade 1 fractures (68.8%) and lung markers (18.7%), and then due to upper endplate (Figure 5(c)) or affected by lung markers, diaphragm (Figure 5(d) )) and the detection of vertebral fractures (VFs) is judged to be normal, which is displayed from the Grad-CAM heat map.

The conclusion is that AIDLEM can detect vertebral fractures (VFs) on plain X-ray profile radiographs (PLRs) with high accuracy, sensitivity, and specificity, especially for non-Genant grades 2 and 3 Severe osteoporotic lumbar spine fractures (VFs). This model can help physicians effectively diagnose vertebral fractures (VFs).

From the contents disclosed in the embodiments of this specification, those with ordinary knowledge in the field to which the present invention belongs can clearly understand that the foregoing embodiments are only examples and can be implemented in combination; those with ordinary knowledge in the technical field to which the present invention belongs can be implemented by many transformations and substitutions. , Does not depart from the present invention. According to the embodiments of the specification, the present invention can have many variations without hindering its implementation. The claims provided in this specification define the scope of the present invention, which covers the aforementioned methods and structures and their equivalent creations.

references:

1. O'Neill TW, Cockerill W, Matthis C, Raspe HH, Lunt M, Cooper C, Banzer D, Cannata JB, Naves M, Felsch B, Felsenberg D, Janott J, Johnell O, Kanis JA, Kragl G, Lopes Vaz A, Lyritis G, Masaryk P, Poor G, Reid DM, Reisinger W, Scheidt-Nave C, Stepan JJ, Todd CJ, Woolf AD, Reeve J, Silman AJ. Back pain, disability, and radiographic vertebral fracture in European women : a prospective study. Osteoporosis international: a journal established as result of cooperation between the European Foundation for Osteoporosis and the National Osteoporosis Foundation of the USA 2004;15(9):760-765. doi: 10.1007/s00198-004-1615-4

2. Cockerill W, Lunt M, Silman AJ, Cooper C, Lips P, Bhalla AK, Cannata JB, Eastell R, Felsenberg D, Gennari C, Johnell O, Kanis JA, Kiss C, Masaryk P, Naves M, Poor G, Raspe H, Reid DM, Reeve J, Stepan J, Todd C, Woolf AD, O'Neill TW. Health-related quality of life and radiographic vertebral fracture. Osteoporosis international: a journal established as result of cooperation between the European Foundation for Osteoporosis and the National Osteoporosis Foundation of the USA 2004;15(2):113-119. doi: 10.1007/s00198-003-1547-4

3. Fink HA, Milavetz DL, Palermo L, Nevitt MC, Cauley JA, Genant HK, Black DM, Ensrud KE. What proportion of incident radiographic vertebral deformities is clinically diagnosed and vice versa? Journal of bone and mineral research: the official journal of the American Society for Bone and Mineral Research 2005;20(7):1216-1222. doi: 10.1359/jbmr.050314

4. Lentle BC, Brown JP, Khan A, Leslie WD, Levesque J, Lyons DJ, Siminoski K, Tarulli G, Josse RG, Hodsman A. Recognizing and reporting vertebral fractures: reducing the risk of future osteoporotic fractures. Can Assoc Radiol J 2007;58(1):27-36.

5. Francis RM, Aspray TJ, Hide G, Sutcliffe AM, Wilkinson P. Back pain in osteoporotic vertebral fractures. Osteoporosis international: a journal established as result of cooperation between the European Foundation for Osteoporosis and the National Osteoporosis Foundation of the USA 2008; 19(7):895-903. doi: 10.1007/s00198-007-0530-x

6. Wong CC, McGirt MJ. Vertebral compression fractures: a review of current management and multimodal therapy. J Multidiscip Healthc 2013;6:205-214. doi: 10.2147/jmdh.S31659

7. Lindsay R, Silverman SL, Cooper C, Hanley DA, Barton I, Broy SB, Licata A, Benhamou L, Geusens P, Flowers K, Stracke H, Seeman E. Risk of new vertebral fracture in the year following a fracture. Jama 2001;285(3):320-323. doi: 10.1001/jama.285.3.320

8. Delmas PD, van de Langerijt L, Watts NB, Eastell R, Genant H, Grauer A, Cahall DL. Underdiagnosis of vertebral fractures is a worldwide problem: the IMPACT study. Journal of bone and mineral research: the official journal of the American Society for Bone and Mineral Research 2005;20(4):557-563. doi: 10.1359/jbmr.041214

9. Olczak J, Fahlberg N, Maki A, Razavian AS, Jilert A, Stark A, Skoldenberg O, Gordon M. Artificial intelligence for analyzing orthopedic trauma radiographs. Acta Orthop 2017;88(6):581-586. doi: 10.1080 /17453674.2017.1344459

10. Burns JE, Yao J, Summers RM. Vertebral Body Compression Fractures and Bone Density: Automated Detection and Classification on CT Images. Radiology 2017;284(3):788-797. doi: 10.1148/radiol.2017162100

11. Lateef H, Patel D. What is the role of imaging in acute low back pain? Curr Rev Musculoskelet Med 2009;2(2):69-73. doi: 10.1007/s12178-008-9037-0

12. Berry GE, Adams S, Harris MB, Boles CA, McKernan MG, Collinson F, Hoth JJ, Meredith JW, Chang MC, Miller PR. Are plain radiographs of the spine necessary during evaluation after blunt trauma? Accuracy of screening torso computed tomography in thoracic/lumbar spine fracture diagnosis. J Trauma 2005;59(6):1410-1413; discussion 1413. doi: 10.1097/01.ta.0000197279.97113.0e

13. VandenBerg J, Cullison K, Fowler SA, Parsons MS, McAndrew CM, Carpenter CR. Blunt Thoracolumbar-Spine Trauma Evaluation in the Emergency Department: A Meta-Analysis of Diagnostic Accuracy for History, Physical Examination, and Imaging. J Emerg Med 2019;56(2):153-165. doi: 10.1016/j.jemermed.2018.10.032

14. Genant HK, Wu CY, van Kuijk C, Nevitt MC. Vertebral fracture assessment using a semiquantitative technique. Journal of bone and mineral research: the official journal of the American Society for Bone and Mineral Research 1993;8(9):1137 -1148. doi: 10.1002/jbmr.5650080915

15. Redmon J, Divvala S, Girshick R, Farhadi A. You Only Look Once: Unified, Real-Time Object Detection. Proc Cvpr Ieee 2016:779-788. doi: 10.1109/Cvpr.2016.91

16. Redmon J, Farhadi A. YOLO9000: Better, Faster, Stronger. 30th Ieee Conference on Computer Vision and Pattern Recognition (Cvpr 2017) 2017:6517-6525. doi: 10.1109/Cvpr.2017.690

17. Redmon J, Farhadi AJapa. Yolov3: An incremental improvement. 2018.

18. Melton LJ, 3rd. Epidemiology of spinal osteoporosis. Spine 1997;22(24 Suppl):2s-11s. doi: 10.1097/00007632-199712151-00002

19. Pan SJ, Yang QA. A Survey on Transfer Learning. Ieee T Knowl Data En 2010;22(10):1345-1359. doi: 10.1109/Tkde.2009.191

20. Rajpurkar P, Irvin J, Zhu K, Yang B, Mehta H, Duan T, Ding D, Bagul A, Langlotz C, Shpanskaya KJapa. Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning. 2017.

21. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-cam: Visual explanations from deep networks via gradient-based localization. Proceedings of the IEEE International Conference on Computer Vision2017; p. 618-626 .

22. Tozakidou M, Reisinger C, Harder D, Lieb J, Szucs-Farkas Z, Muller-Gerbl M, Studler U, Schindera S, Hirschmann A. Systematic Radiation Dose Reduction in Cervical Spine CT of Human Cadaveric Specimens: How Low Can We Go? AJNR American journal of neuroradiology 2018;39(2):385-391. doi: 10.3174/ajnr.A5490

23. Simpson AK, Whang PG, Jonisch A, Haims A, Grauer JN. The radiation exposure associated with cervical and lumbar spine radiographs. Journal of spinal disorders & techniques 2008;21(6):409-412. doi: 10.1097/BSD .0b013e3181568656

24. Richards PJ, George J, Metelko M, Brown M. Spine computed tomography doses and cancer induction. Spine 2010;35(4):430-433. doi: 10.1097/BRS.0b013e3181cdde47

25. Cheng CT, Ho TY, Lee TY, Chang CC, Chou CC, Chen CC, Chung IF, Liao C-HJEr. Application of a deep learning algorithm for detection and visualization of hip fractures on plain pelvic radiographs. 2019:1- 9.

26. Griffith JF, Guglielmi G. Vertebral fracture. Radiologic clinics of North America 2010;48(3):519-529. doi: 10.1016/j.rcl.2010.02.012

27. Diacinti D, Acca M, D'Erasmo E, Tomei E, Mazzuoli GF. Aging changes in vertebral morphometry. Calcified tissue international 1995;57(6):426-429. doi: 10.1007/bf00301945

28. Frobin W, Brinckmann P, Kramer M, Hartwig E. Height of lumbar discs measured from radiographs compared with degeneration and height classified from MR images. European radiology 2001;11(2):263-269. doi: 10.1007/s003300000556

29. Guglielmi G, Palmieri F, Placentino MG, D'Errico F, Stoppino LP. Assessment of osteoporotic vertebral fractures using specialized workflow software for 6-point morphometry. European journal of radiology 2009;70(1):142-148. doi : 10.1016/j.ejrad.2007.12.001

30. Binkley N, Krueger D, Gangnon R, Genant HK, Drezner MK. Lateral vertebral assessment: a valuable technique to detect clinically significant vertebral fractures. Osteoporosis international: a journal established as result of cooperation between the European Foundation for Osteoporosis and the National Osteoporosis Foundation of the USA 2005;16(12):1513-1518. doi: 10.1007/s00198-005-1891-7

31. Vokes TJ, Dixon LB, Favus MJ. Clinical utility of dual-energy vertebral assessment (DVA). Osteoporosis international: a journal established as result of cooperation between the European Foundation for Osteoporosis and the National Osteoporosis Foundation of the USA 2003;14 (11):871-878. doi: 10.1007/s00198-003-1461-9

32. Rea JA, Li J, Blake GM, Steiger P, Genant HK, Fogelman I. Visual assessment of vertebral deformity by X-ray absorptiometry: a highly predictive method to exclude vertebral deformity. Osteoporosis international: a journal established as result of cooperation between the European Foundation for Osteoporosis and the National Osteoporosis Foundation of the USA 2000;11(8):660-668. doi: 10.1007/s001980070063

33. Ruiz Santiago F, Tomas Munoz P, Moya Sanchez E, Revelles Paniza M, Martinez Martinez A, Perez Abela AL. Classifying thoracolumbar fractures: role of quantitative imaging. Quantitative imaging in medicine and surgery 2016;6(6):772- 784. doi: 10.21037/qims.2016.12.04

Claims (6)

A method for detecting chest and lumbar vertebral fractures, including: obtaining a plain X-ray lateral radiographic image; preprocessing data, including: providing the plain X-ray lateral radiographic image; using the plain X-ray lateral radiographic image Object detection algorithm locates the spine; Preprocess the plain X-ray profile image with Gaussian blur, median filter, and Contrast Adaptive Histogram Equalization; redefine the image by adjusting the method The width and height of the plain X-ray lateral radiation image; and optimize the positioning of the plain X-ray lateral radiation image; develop model integration, including: unfreeze the hidden layer according to the transfer learning method and ImageNet training weights; and learn the classification model Preprocess the features of the plain X-ray side radiography; and mark the chest and lumbar vertebral fractures based on the data preprocessing and the prediction results of the model integration. The object detection algorithm is the YOLO method. The method according to claim 1, wherein the method for optimizing the positioning of the plain X-ray side-radiation image is selected from (1) image size adjustment and no image preprocessing; (2) image size adjustment and self-comparison The adaptive histogram equalization; (3) maintain the original image ratio and no image preprocessing; and (4) maintain the original image ratio and implement the group of adaptive histogram equalization. The method according to claim 2, wherein the method of optimizing the positioning of the plain X-ray lateral radiation image is to adjust the image size and without image preprocessing. The method according to claim 1, wherein the model integration is pre-trained in a fast artificial intelligence (fast.AI) library. The method according to claim 4, wherein the model integration is pre-training of a combination of ResNet34, DenseNet121, and DenseNet201 models. The method according to claim 1, wherein the step of preprocessing the data further includes: maintaining the image ratio and filling the image with color blocks to correspond to the 224*224 pixel output format of the fracture classification model.

Images