WO2023010248A1 - 利用胸腹部正位x线片检测脊柱骨质疏松性压缩性骨折的装置 - Google Patents

利用胸腹部正位x线片检测脊柱骨质疏松性压缩性骨折的装置 Download PDF

Info

Publication number
WO2023010248A1
WO2023010248A1 PCT/CN2021/110109 CN2021110109W WO2023010248A1 WO 2023010248 A1 WO2023010248 A1 WO 2023010248A1 CN 2021110109 W CN2021110109 W CN 2021110109W WO 2023010248 A1 WO2023010248 A1 WO 2023010248A1
Authority
WO
WIPO (PCT)
Prior art keywords
chest
ray
abdomen
osteoporotic
anteroposterior
Prior art date
Application number
PCT/CN2021/110109
Other languages
English (en)
French (fr)
Inventor
王毅翔
肖本亨
闫泰毓
Original Assignee
香港中文大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 香港中文大学 filed Critical 香港中文大学
Priority to PCT/CN2021/110109 priority Critical patent/WO2023010248A1/zh
Publication of WO2023010248A1 publication Critical patent/WO2023010248A1/zh

Links

Images

Classifications

    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B6/00Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment

Definitions

  • the invention relates to the field of medical imaging, in particular to a device for detecting spinal osteoporotic compression fractures by using chest and abdomen anteroposterior X-ray films.
  • Osteoporotic vertebral fracture (OVF) of the spine is a common disease in the elderly, especially in elderly women. However, about 3/4 of OVF will not be accompanied by obvious pain, which often leads to the neglect of patients. Although physical examination is recommended for patients with osteoporosis, patients often ignore physical examination, so that the condition worsens, leading to more serious hip fractures. Elderly people often have the opportunity to take frontal X-ray images of the chest and abdomen or participate in chest X-ray examinations due to pneumonia, bronchitis, abdominal pain and other diseases, so OVF can be observed using Fontal view Radiograph (FR) of the chest and abdomen. However, OVF on anteroposterior radiographs is often missed.
  • FR Fontal view Radiograph
  • Anteroposterior X-ray films of the chest and abdomen are often used for the diagnosis of bronchitis, urinary calculi and other diseases, and clinicians do not pay too much attention to whether the spine is fractured.
  • doctors in other outpatient clinics are not professional enough in the diagnosis of compression fractures, so it is easy to cause missed diagnosis.
  • CT computed tomography
  • MRI magnetic resonance imaging
  • the contrast and clarity of the vertebrae are poor, and it is difficult to observe the vertebrae with compression fractures, especially mild fractures.
  • the purpose of the present invention is to help doctors find the possible locations of spinal osteoporotic compression fractures based on the very common frontal X-ray images of the chest and abdomen in clinical diagnosis, and to remind doctors and patients in time.
  • the present invention adopts the following technical solutions:
  • a device for detecting osteoporotic compression fractures of the spine including a processor, when the processor runs a computer program stored in a computer-readable storage medium, the following steps are performed:
  • step S1 the 4 pictures are randomly scaled, randomly cropped, and randomly arranged to be stitched together as the input of the convolutional neural network.
  • step S2 the network model structure of the convolutional neural network includes:
  • the downsampled residual convolution layer is used to extract strong semantic features.
  • the downsampled residual convolution layer transfers and fuses the high-level feature information through downsampling to obtain a feature map for prediction, which is conveyed from top to bottom Strong semantic features;
  • the feature pyramid structure behind the downsampling residual convolution layer conveys strong positioning features from bottom to top, and performs parameter aggregation on different detection layers from different backbone layers to extract strong positioning features.
  • the residual network structure includes two sets of convolutional layers, batch processing layers, activation functions, and feature fusion layers.
  • the prediction output part includes a classification branch and a regression branch
  • the classification branch is configured to determine the probability of fracture of the detected vertebrae, and in the process of training, after passing through the fully connected layer, the binary classification problem is obtained Probability for each category
  • the regression branch is configured to determine the location of vertebrae that may be fractured, and frame the vertebrae that are likely to be fractured with a detection box.
  • step S3 the classification branch performs backpropagation through a cross-entropy loss function to update network parameters.
  • step S3 the regression branch uses the GIoU loss function to perform backpropagation and update network parameters.
  • step S3 when there is an osteoporotic compression fracture in the output X-ray image, the vertebral body with the osteoporotic compression fracture is prompted, and the position of the osteoporotic compression fracture is identified in the X-ray image Make an annotation.
  • a computer-readable storage medium stores a computer program, and when the computer program is run by a processor, the following steps are performed:
  • Image preprocessing is performed on the original spinal X-ray image
  • the present invention has the following beneficial effects:
  • the present invention proposes a device for detecting osteoporotic compression fractures of the spine, which can use the very commonly used chest and abdomen anteroposterior X-ray images in clinic (these chest and abdomen anteroposterior X-ray images can be used for other diseases or routine examinations) Obtained), through the artificial intelligence algorithm, it can prompt the location of possible spinal compression fractures, automatically help doctors to detect the possible locations of spinal osteoporotic compression fractures early, and can promptly remind doctors and patients for early examination and treatment. This can not only greatly reduce the workload of doctors, but also avoid the situation that doctors in other departments cannot make a diagnosis due to insufficient professionalism.
  • an anteroposterior X-ray film is used as an input, and a convolutional neural network for target detection is used, and if there is an OVF in the output X-ray image, it is marked in the X-ray image.
  • the correct detection rate according to the number of OVF vertebral bodies is 0.944, and the false positive rate according to the number of patients with or without OVF is 0.09.
  • the device of the present invention can greatly reduce the workload of clinicians by using AI algorithms, and can be applied to other outpatient departments, thus avoiding clinical problems caused by image quality problems, different areas of attention of doctors, and other diseases.
  • the missed diagnosis caused by the lack of professionalism of doctors can achieve the purpose of early detection and early treatment, which greatly reduces the aggravation of patients.
  • the invention proposes for the first time the idea of diagnosing vertebral compression fractures with anteroposterior X-ray images of the chest and abdomen, and proposes an artificial intelligence algorithm to assist doctors in diagnosis, which has important clinical significance and market potential.
  • Using the device of the present invention can assist doctors to automatically detect spinal osteoporotic compression fractures using chest and abdomen X-ray images when diagnosing bronchitis, pneumonia and other diseases, and remind the doctor in time whether the patient has vertebral fractures or not. Compression fractures for further diagnosis and treatment of patients.
  • FIG. 1 is a schematic diagram of a processing flow of an embodiment of the present invention.
  • Fig. 2 is a schematic diagram of the overall structure of an embodiment of the present invention.
  • Fig. 3 is a schematic diagram of an image preprocessing effect according to an embodiment of the present invention.
  • FIG. 4 is a schematic diagram of a residual network structure according to an embodiment of the present invention.
  • Fig. 5 is a schematic diagram of a specific residual structure according to an embodiment of the present invention.
  • FIG. 6 is a schematic diagram of a network model structure according to an embodiment of the present invention.
  • the present invention provides a device for detecting osteoporotic compression fractures of the spine, including a processor, the processor is run by a computer-readable memory When storing a computer program on a medium, perform the following steps:
  • the present invention provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and the computer program is run by a processor , perform the following steps:
  • Image preprocessing is performed on the original spinal X-ray image
  • the device for detecting osteoporotic compression fractures of the spine proposed by the present invention uses very commonly used chest and abdomen anteroposterior X-ray films in clinical practice (these chest and abdomen anteroposterior X-ray films can be obtained during other diseases or routine examinations) , through the artificial intelligence algorithm, it automatically helps doctors to detect the possible locations of spinal compression fractures early, and reminds patients for early examination and treatment. This can not only greatly reduce the workload of doctors, but also avoid the situation that doctors in other departments cannot make a diagnosis due to insufficient professionalism.
  • the realization process of the present invention comprises the following parts: image preprocessing part, CNNs (Convolutional Neural Network) and prediction output part for extracting the OVF position information in the image.
  • image preprocessing part CNNs (Convolutional Neural Network)
  • prediction output part for extracting the OVF position information in the image.
  • Image preprocessing part Since X-ray images have the characteristics of low resolution, low contrast, and small vertebral targets, the preferred embodiment of the present invention proposes a data enhancement method, which randomly zooms and randomly crops 4 pictures , in a random arrangement, as shown in Figure 3, which can reduce the use of the GPU.
  • the data enhancement method not only enriches the detection data set, especially the random scaling adds many small targets, which makes the network more robust; it also reduces the GPU computing power requirements.
  • This data enhancement method for training The data of 4 pictures can be directly calculated, the Mini-batch size is reduced, and a single GPU can make the network loss function converge, and obtain a better model training effect.
  • This part is mainly a convolutional neural network.
  • a network module similar to a pyramid structure is also constructed to better extract semantic features.
  • the CNN model part mainly consists of a convolution layer composed of convolution kernels of different sizes, a pooling layer, an upsampling layer, and a feature merging layer.
  • Convolutional layers, pooling layers, batch normalization, and activation functions constitute the residual module.
  • Using a residual module combined with a convolutional layer and a pooling layer to form a network can better extract image features.
  • Image batch normalization makes the model have good generalization robustness, and the residual module can solve the model degradation problem caused by network deepening. Its structure is shown in Figure 4.
  • the residual module structure of the preferred embodiment of the present invention is shown in FIG. 5 .
  • the embodiment of the present invention can use a network module similar to a pyramid structure to better extract semantic features and ensure that the OVF detection of the model has good performance.
  • the overall network structure is shown in FIG. 6 .
  • Prediction output part This part mainly includes two parts: classification branch and regression branch.
  • the classification branch provides the probability of fracture of the detected vertebrae, and in the process of training, after passing through the fully connected layer, the probability of each category will be obtained for the binary classification problem; the preferred embodiment reverses through the cross entropy loss function Propagate, update and optimize network parameters.
  • the regression branch provides the position of the vertebrae that may be fractured, and uses the detection frame to frame the vertebrae that have OVF fractures.
  • the preferred embodiment uses the GIoU loss function to carry out backpropagation, and updates the training parameters of the network to obtain the best detection result (Bounding Box detection result).
  • Figure 1 shows the OVF detection process, in which the original frontal X-ray image of the chest and abdomen is first input, and after preprocessing and convolutional neural network model processing, the OVF detection result is finally output.
  • Fig. 2 shows that in the preferred embodiment of the present invention, the X-ray image input after preprocessing includes a downsampling residual convolution layer, a pyramid structure network model for extracting OVF position features, and finally the image features are classified and regressed, for existing
  • the vertebral body of the OVF can be indicated by a red bounding box.
  • Conv1x1 represents a 1X1 convolution kernel
  • Conv3x3 represents a 3x3 convolution kernel.
  • Fig. 3 shows the image preprocessing of the preferred embodiment of the present invention, in which four images are randomly scaled, randomly cropped, and randomly arranged for splicing as the input of the convolutional neural network.
  • Figure 4 shows the principle of the residual network structure of the preferred embodiment of the present invention.
  • Fig. 5 shows the specific structure of the residual module of the preferred embodiment of the present invention, including two sets of convolutional layers, batch processing layers, activation functions, and feature fusion layers.
  • Fig. 6 shows the network model structure of the preferred embodiment of the present invention
  • the left part is the downsampled residual convolution layer, which is used to extract strong semantic features
  • the structure in the dotted box is the feature pyramid module.
  • the downsampled residual convolutional layer transfers and fuses the high-level feature information by downsampling to obtain a feature map for prediction, and conveys strong semantic features from top to bottom; after the downsampled residual convolutional layer, an automatic
  • the bottom-up feature pyramid structure conveys strong positioning features from bottom to top, and aggregates parameters from different backbone layers to different detection layers to achieve the purpose of extracting strong positioning features and ensure that the OVF detection of the model has good performance.
  • an anteroposterior X-ray film is used as an input, and a convolutional neural network for object detection is used to output an OVF in the X-ray image, which is marked in the X-ray image.
  • the test results using the specific embodiment of the present invention are: the correct detection accuracy rate according to the number of OVF vertebral bodies is 0.944, and the false positive rate according to the number of patients with or without OVF is 0.09.
  • the artificial intelligence algorithm is used for the detection of compression fractures in X-ray images of the chest and abdomen for the first time, and the feasibility of the algorithm is verified by using a large amount of clinical data.
  • the inventor verified the feasibility of the algorithm using clinical data from different regions and different hospitals, and achieved good detection results. Count, Based on Patients) reached 0.944 and 0.09 respectively.
  • the algorithm can be written into a visual operation interface, which is convenient for clinicians to operate.
  • the use of artificial intelligence algorithms can greatly reduce the workload of clinicians, thereby avoiding missed diagnoses caused by image quality problems, different areas of attention of doctors, and insufficient professionalism of clinicians for other diseases, so as to achieve The purpose of early detection and early treatment.
  • the Background of the Invention section may contain background information about the problem or circumstances of the invention without necessarily describing prior art. Accordingly, inclusion in the Background section is not an admission by the applicant of prior art.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Public Health (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Pathology (AREA)
  • Primary Health Care (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Data Mining & Analysis (AREA)
  • Epidemiology (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biophysics (AREA)
  • High Energy & Nuclear Physics (AREA)
  • Databases & Information Systems (AREA)
  • Optics & Photonics (AREA)
  • Radiology & Medical Imaging (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Molecular Biology (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • Veterinary Medicine (AREA)
  • Image Analysis (AREA)
  • Apparatus For Radiation Diagnosis (AREA)

Abstract

一种利用胸腹部正位X线片检测脊柱骨质疏松性压缩性骨折的装置,包括处理器,处理器运行由计算机可读存储介质存储的计算机程序时,执行如下步骤:S1、对原始的胸腹部正位X射线图像进行图像预处理;S2、将预处理后的X射线图像输入卷积神经网络进行处理,其中,卷积神经网络经配置使用残差模块提取图像特征,并使用类金字塔结构的网络模块提取语义特征,以便实现X射线图像中的骨质疏松性压缩性骨折位置信息的提取;S3、将提取的特征经过预测输出部分进行分类和回归,输出骨质疏松性压缩性骨折的检测结果。基于在临床诊断中非常常见的胸腹部正位X射线图像,帮助医生发现有可能发生脊柱骨质疏松性压缩骨折的位置,并及时提醒医生及患者。

Description

利用胸腹部正位X线片检测脊柱骨质疏松性压缩性骨折的装置 技术领域
本发明涉及医学影像领域,特别是一种利用胸腹部正位X线片检测脊柱骨质疏松性压缩性骨折的装置。
背景技术
脊柱的骨质疏松性压缩性骨折(osteoporotic vertebral fracture,OVF)在老年人中,尤其是老年女性中,是常见病。但是约3/4的OVF不会伴有明显的疼痛,往往会造成患者的忽视。虽然建议骨质疏松患者体检,但患者常常忽视体检,以至于病情加重,导致更加严重的髋关节骨折的发生。老年人常常有机会因肺炎、支气管炎、腹痛等疾病拍摄胸腹部的正位X射线图像或者参加胸片体检,故可以使用胸腹部的正位X线片(Fontal view Radiograph,FR)观察OVF。但是正位X线片上的OVF常常漏诊。胸腹部的正位X线片往往用于支气管炎、尿路结石等其他疾病诊断,临床医生不会过多关注脊椎是否发生骨折。另外,其他门诊的医生对于压缩性骨折的诊断专业性也不够,因此很容易造成漏诊。相比较于CT、MRI等成像方式,对于胸腹部的正位的X射线图像,其椎骨的对比度和清晰度较差,不易观察到发生压缩性骨折的椎骨,尤其是轻度骨折。
需要说明的是,在上述背景技术部分公开的信息仅用于对本申请的背景的理解,因此可以包括不构成对本领域普通技术人员已知的现有技术的信息。
发明内容
本发明的目的是基于在临床诊断中非常常见的胸腹部正位X射线图像,帮助医生发现有可能发生脊柱骨质疏松性压缩骨折的位置,并及时提醒医生及患者。
为实现上述目的,本发明采用以下技术方案:
一种检测脊柱骨质疏松性压缩性骨折的装置,包括处理器,所述处理器运行由计算机可读存储介质存储的计算机程序时,执行如下步骤:
S1、对原始的胸腹部正位X射线图像(即正位X线片)进行图像预处理;
S2、将预处理后的X射线图像输入卷积神经网络CNNs进行处理,其中,所述卷积神经网络经配置使用残差模块提取图像特征,并使用类金字塔结构的网络模块提取语义特征,以便实现X射线图像中的骨质疏松性压缩性骨折位置信息的提取;
S3、将提取的特征经过预测输出部分进行分类和回归,输出骨质疏松性压缩性骨折的检测结果。
进一步地:
步骤S1中,将4张图片进行随机缩放、随机裁剪、随机排布的方式进行拼接,以作为卷积神经网络的输入。
步骤S2中,所述卷积神经网络的网络模型结构包括:
下采样残差卷积层,用于提取强语义特征,所述下采样残差卷积层将高层的特征信息通过下采样的方式进行传递融合,得到进行预测的特征图,自顶向下传达强语义特征;
在所述下采样残差卷积层后面的特征金字塔结构,所述特征金字塔结构自底向上传达强定位特征,从不同的主干层对不同的检测层进行参数聚合,以提取强定位特征。
步骤S2中,所述残差模块经配置利用中间的非线性层得到F(x)=H(x)–x的映射关系,以便使得网络收敛和参数优化。
步骤S2中,残差网络结构包括两组卷积层、批处理层、激活函数,以及特征融合层。
步骤S3中,预测输出部分包含分类分支和回归分支,所述分类分支经配置以确定检测到的椎骨发生骨折的概率,并且在训练的过程中,在通过全连接层之后,对于二分类问题得到每一个类别的概率;所述回归分支经配置以确定有可能发生骨折的椎骨的位置,并用检测框将发生骨折的椎骨框出。
步骤S3中,所述分类分支通过交叉熵损失函数来进行反向传播,更新网络参数。
步骤S3中,所述回归分支使用GIoU损失函数来进行反向传播,并更新网络参数。
步骤S3中,当输出X线图像中具有骨质疏松性压缩性骨折时,对存在骨质疏松性压缩性骨折的椎体进行提示,在X射线图像中对骨质疏松性压缩性骨折的位置进行标注。
一种计算机可读存储介质,所述计算机可读存储介质存储有计算机程 序,所述计算机程序由处理器运行时,执行如下步骤:
S1、对原始的脊柱X射线图像进行图像预处理;
S2、将预处理后的X射线图像输入卷积神经网络CNNs进行处理,其中,所述卷积神经网络经配置使用残差模块提取图像特征,并使用类金字塔结构的网络模块提取语义特征,以便实现X射线图像中的骨质疏松性压缩性骨折位置信息的提取;
S3、将提取的特征经过预测输出部分进行分类和回归,输出骨质疏松性压缩性骨折的检测结果。
与现有技术相比,本发明具有如下有益效果:
本发明提出一种检测脊柱骨质疏松性压缩性骨折的装置,能够使用在临床中非常常用的胸腹部正位X射线图像(这些胸腹部正位X射线图像可在进行其他疾病或者常规检查时获得),通过人工智能算法提示可能发生脊柱压缩性骨折的位置,自动帮助医生及早发现有可能发生脊柱骨质疏松性压缩性骨折的位置,可以及时提醒医生,并提醒患者及早进行检查和治疗。这样既可以大大的减轻了医生的工作负担,又可以避免其他科室的医生由于专业性不够无法做出诊断的情况发生。在本发明的一个实施例中,用正位X线片作为输入,利用目标检测的卷积神经网络,如果输出X射线图像中具有OVF,则在X射线图像中进行标注。本发明实施例的测试结果为,按照OVF椎体个数检测正确的准确率为0.944,按照有无OVF的病人个数计假阳性率为0.09。
相比较于传统的临床诊断方法,本发明的装置使用AI算法可以大大的减少临床医生的工作量,并且可以应用的其他门诊部门,从而避免了因图像质量问题、医生关注区域不同以及其他疾病临床医生专业度不够所造成的漏诊,从而达到一个提早发现提早治疗的目的,大大减少了患者病情加重的情况。
本发明首次提出的利用胸腹部正位X射线图像进行脊椎压缩性骨折诊断的思想,并提出人工智能算法,辅助医生进行诊断,具有重要的临床意义和市场潜力。使用本发明的装置,可以辅助医生在进行支气管炎、肺炎等其他疾病诊断时,使用胸腹部X射线图像对脊柱骨质疏松性压缩性骨折自动进行检测,并及时提醒医生,患者有无发生脊椎压缩性骨折,以便患者进一步诊疗。
附图说明
图1为本发明一种实施例的处理流程示意图。
图2为本发明一种实施例的整体结构示意图。
图3为本发明一种实施例的图像预处理效果示意图。
图4为本发明一种实施例的残差网络结构原理图。
图5为本发明一种实施例的具体残差结构示意图。
图6为本发明一种实施例的网络模型结构示意图。
具体实施方式
以下对本发明的实施方式作详细说明。应该强调的是,下述说明仅仅是示例性的,而不是为了限制本发明的范围及其应用。
参阅图1、图2和图6,在第一方面的实施例中,本发明提供一种检测脊柱骨质疏松性压缩性骨折的装置,包括处理器,所述处理器运行由计算机可读存储介质存储的计算机程序时,执行如下步骤:
S1、对原始的胸腹部正位X射线图像正位X线片进行图像预处理;
S2、将预处理后的X射线图像输入卷积神经网络CNNs进行处理,其中,所述卷积神经网络经配置使用残差模块提取图像特征,并使用类金字塔结构的网络模块提取语义特征,以便实现X射线图像中的骨质疏松性压缩性骨折位置信息的提取;
S3、将提取的特征经过预测输出部分进行分类和回归,输出骨质疏松性压缩性骨折的检测结果。
参阅图1、图2和图6,在第二方面的实施例中,本发明提供一种计算机可读存储介质,所述计算机可读存储介质存储有计算机程序,所述计算机程序由处理器运行时,执行如下步骤:
S1、对原始的脊柱X射线图像进行图像预处理;
S2、将预处理后的X射线图像输入卷积神经网络CNNs进行处理,其中,所述卷积神经网络经配置使用残差模块提取图像特征,并使用类金字塔结构的网络模块提取语义特征,以便实现X射线图像中的骨质疏松性压缩性骨折位置信息的提取;
S3、将提取的特征经过预测输出部分进行分类和回归,输出骨质疏松性压缩性骨折的检测结果。
利用临床上广泛使用的胸腹部X射线图像来进行压缩性骨折的检测已经在本发明之前的研究中验证了其可行性。本发明提出的检测脊柱骨质疏松性压缩性骨折的装置,使用在临床中非常常用的胸腹部正位X线片(这些胸腹部正位X线片可在进行其他疾病或者常规检查时获得),通过人工 智能算法,自动帮助医生及早发现有可能发生脊柱压缩性骨折的位置,并提醒患者及早进行检查和治疗。这样既可以大大的减轻了医生的工作负担,又可以避免其他科室的医生由于专业性不够无法做出诊断的情况发生。
以下进一步举例描述本发明具体实施例。
参阅图1至图6,本发明的实现过程包括以下几个部分:图像预处理部分,用于提取图像中的OVF位置信息的CNNs(Convolutional Neural Network)以及预测输出部分。
(1)图像预处理部分:由于X射线图像具有低分辨率、低对比度和椎骨目标较小等特点,本发明优选实施例提出一种数据增强的方式,将4张图片进行随机缩放、随机裁剪、随机排布的方式进行拼接,如图3所示,这样可以减少GPU的使用。
数据增强的方法不但丰富了检测数据集,特别是随机缩放增加了很多小目标,让网络具有更好的鲁棒性;而且减少了GPU算力需求,在使用该数据增强的方法进行训练时,可以直接计算4张图片的数据,降低了Mini-batch大小,单张GPU就可以使网络损失函数收敛,得到较好的模型训练效果。
(2)CNN模型:该部分主要是卷积神经网络。在该部分中,除了使用残差模块更好的提取图像特征以外,还构建类似于金字塔结构的网络模块来更好的提取语义特征。
在一些实施例中,CNN模型部分主要由不同大小的卷积核构成卷积层、池化层、上采样层和特征合并层组成。卷积层、池化层、批归一化和激活函数构成残差模块。使用残差模块与卷积层和池化层组合成网络,能更好地提取图像特征。图像批归一化使模型具有良好的泛化鲁棒性能,同时残差模块可以解决网络加深带来的模型退化问题。其结构如图4所示。
残差模块利用中间的非线性层得到F(x)=H(x)–x的映射关系,则网络层的输出可以表示为H(x)=F(x)+x。由图4右边所示,相比H(x)=F(x)+x函数,F(x)=H(x)-x函数更容易收敛,网络参数将更容易优化。本发明优选实施例的残差模块结构如图5所示。
同时,本发明实施例可以采用类似于金字塔结构的网络模块来更好的提取语义特征,保证模型的OVF检测具有良好的性能,其整体网络结构如图6示。
(3)预测输出部分:该部分主要包含分类分支和回归分支两个部分。 分类分支提供检测到的椎骨发生骨折的概率,并且在训练的过程中,在通过全连接层之后,对于二分类问题会得到每一个类别的概率;优选实施例通过交叉熵损失函数来进行反向传播,更新并优化网络参数。回归分支则提供有可能发生骨折的椎骨的位置,并用检测框将发生OVF骨折的椎骨框出。在训练的过程中,由于IoU损失函数的局限性,优选实施例使用GIoU损失函数来进行反向传播,并更新网络的训练参数,以得到最好的检测结果(Bounding Box检测结果)。
图1示出了OVF检测流程,其中,先输入原始的胸腹部正位X射线图像正位X线片,经过预处理和卷积神经网络模型处理之后,最终输出OVF的检测结果。
图2示出本发明优选实施例在预处理之后的X射线图像输入包括下采样残差卷积层、提取OVF位置特征的金字塔结构的网络模型中,最后将图像特征经过分类和回归,对于存在OVF的椎体可用红色边界框进行提示。其中,Conv1x1表示1X1的卷积核,Conv3x3表示3x3的卷积核。
图3示出本发明优选实施例的图像预处理,将4张图片进行随机缩放、随机裁剪、随机排布的方式进行拼接,作为卷积神经网络的输入。
图4示出本发明优选实施例的残差网络结构原理,残差模块利用中间的非线性层得到F(x)=H(x)–x的映射关系,使得网络更容易收敛,参数更容易优化。
图5示出本发明优选实施例的残差模块的具体结构,包括两组卷积层、批处理层、激活函数,和特征融合层。
图6示出本发明优选实施例的网络模型结构,左边部分为下采样残差卷积层,其作用是用于提取强语义特征;而虚线框中结构为特征金字塔模块。下采样残差卷积层将高层的特征信息通过下采样的方式进行传递融合,得到进行预测的特征图,自顶向下传达强语义特征;在下采样残差卷积层的后面添加了一个自底向上的特征金字塔结构,自底向上传达强定位特征,从不同的主干层对不同的检测层进行参数聚合,达到提取强定位特征的目的,保证模型的OVF检测具有良好的性能。
在本发明的一个具体实施例中,用正位X线片作为输入,利用目标检测的卷积神经网络,输出X线图像中如果具有OVF,其在X射线图像中进行标注。使用本发明具体实施例的测试结果为:按照OVF椎体个数检测正确的准确率为0.944,按照有无OVF的病人个数计假阳性率为0.09。
本发明首次将人工智能算法用于胸腹部X射线图像压缩性骨折的检测,并使用大量临床数据验证了算法的可行性。发明人使用来自不同地区、不同医院的临床数据验证了算法的可行性,并取得了较好的检测结果,其准确率(按椎体个数计,Based on Vertebrae)和假阳性率(按患者个数计,Based on Patients)分别达到了0.944和0.09。
根据本发明,可将算法编写成可视化的操作界面,便于临床医生进行操作。
相比较于传统的临床诊断方法,使用人工智能算法可以大大的减少临床医生的工作量,从而避免了因图像质量问题、医生关注区域不同以及其他疾病临床医生专业度不够所造成的漏诊,从而达到提早发现提早治疗的目的。
本发明的背景部分可以包含关于本发明的问题或环境的背景信息,而不一定是描述现有技术。因此,在背景技术部分中包含的内容并不是申请人对现有技术的承认。
以上内容是结合具体/优选的实施方式对本发明所作的进一步详细说明,不能认定本发明的具体实施只局限于这些说明。对于本发明所属技术领域的普通技术人员来说,在不脱离本发明构思的前提下,其还可以对这些已描述的实施方式做出若干替代或变型,而这些替代或变型方式都应当视为属于本发明的保护范围。在本说明书的描述中,参考术语“一种实施例”、“一些实施例”、“优选实施例”、“示例”、“具体示例”、或“一些示例”等的描述意指结合该实施例或示例描述的具体特征、结构、材料或者特点包含于本发明的至少一个实施例或示例中。在本说明书中,对上述术语的示意性表述不必须针对的是相同的实施例或示例。而且,描述的具体特征、结构、材料或者特点可以在任一个或多个实施例或示例中以合适的方式结合。在不相互矛盾的情况下,本领域的技术人员可以将本说明书中描述的不同实施例或示例以及不同实施例或示例的特征进行结合和组合。尽管已经详细描述了本发明的实施例及其优点,但应当理解,在不脱离专利申请的保护范围的情况下,可以在本文中进行各种改变、替换和变更。

Claims (10)

  1. 一种利用胸腹部正位X线片检测脊柱骨质疏松性压缩性骨折的装置,包括处理器,其特征在于,所述处理器运行由计算机可读存储介质存储的计算机程序时,执行如下步骤:
    S1、对原始的胸腹部正位X射线图像FR进行图像预处理;
    S2、将预处理后的X射线图像输入卷积神经网络CNNs进行处理,其中,所述卷积神经网络经配置使用残差模块提取图像特征,并使用类金字塔结构的网络模块提取语义特征,以便实现X射线图像中的骨质疏松性压缩性骨折位置信息的提取;
    S3、将提取的特征经过预测输出部分进行分类和回归,输出骨质疏松性压缩性骨折的检测结果。
  2. 如权利要求1所述的利用胸腹部正位X线片检测脊柱骨质疏松性压缩性骨折的装置,其特征在于,步骤S1中,将4张图片进行随机缩放、随机裁剪、随机排布的方式进行拼接,以作为卷积神经网络的输入。
  3. 如权利要求1所述的利用胸腹部正位X线片检测脊柱骨质疏松性压缩性骨折的装置,其特征在于,步骤S2中,所述卷积神经网络的网络模型结构包括:
    下采样残差卷积层,用于提取强语义特征,所述下采样残差卷积层将高层的特征信息通过下采样的方式进行传递融合,得到进行预测的特征图,自顶向下传达强语义特征;
    在所述下采样残差卷积层后面的特征金字塔结构,所述特征金字塔结构自底向上传达强定位特征,从不同的主干层对不同的检测层进行参数聚合,以提取强定位特征。
  4. 如权利要求1所述的利用胸腹部正位X线片检测脊柱骨质疏松性压缩性骨折的装置,其特征在于,步骤S2中,所述残差模块经配置利用中间的非线性层得到F(x)=H(x)–x的映射关系,以便使得网络收敛和参数优化。
  5. 如权利要求4所述的利用胸腹部正位X线片检测脊柱骨质疏松性压缩性骨折的装置,其特征在于,步骤S2中,残差网络结构包括两组卷积层、批处理层、激活函数,以及特征融合层。
  6. 如权利要求1至5任一项所述的利用胸腹部正位X线片检测脊柱骨质疏松性压缩性骨折的装置,其特征在于,步骤S3中,预测输出部分 包含分类分支和回归分支,所述分类分支经配置以确定检测到的椎骨发生骨折的概率,并且在训练的过程中,在通过全连接层之后,对于二分类问题得到每一个类别的概率;所述回归分支经配置以确定有可能发生骨折的椎骨的位置,并用检测框将发生骨折的椎骨框出。
  7. 如权利要求6所述的利用胸腹部正位X线片检测脊柱骨质疏松性压缩性骨折的装置,其特征在于,步骤S3中,所述分类分支通过交叉熵损失函数来进行反向传播,更新网络参数。
  8. 如权利要求6所述的利用胸腹部正位X线片检测脊柱骨质疏松性压缩性骨折的装置,其特征在于,步骤S3中,所述回归分支使用GIoU损失函数来进行反向传播,并更新网络参数。
  9. 如权利要求1至5任一项所述的利用胸腹部正位X线片检测脊柱骨质疏松性压缩性骨折的装置,其特征在于,步骤S3中,当输出X线图像中具有骨质疏松性压缩性骨折时,对存在骨质疏松性压缩性骨折的椎体进行提示,在X射线图像中对骨质疏松性压缩性骨折的位置进行标注。
  10. 一种计算机可读存储介质,所述计算机可读存储介质存储有计算机程序,其特征在于,所述计算机程序由处理器运行时,执行如下步骤:
    S1、对原始的脊柱X射线图像进行图像预处理;
    S2、将预处理后的X射线图像输入卷积神经网络CNNs进行处理,其中,所述卷积神经网络经配置使用残差模块提取图像特征,并使用类金字塔结构的网络模块提取语义特征,以便实现X射线图像中的骨质疏松性压缩性骨折位置信息的提取;
    S3、将提取的特征经过预测输出部分进行分类和回归,输出骨质疏松性压缩性骨折的检测结果。
PCT/CN2021/110109 2021-08-02 2021-08-02 利用胸腹部正位x线片检测脊柱骨质疏松性压缩性骨折的装置 WO2023010248A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/110109 WO2023010248A1 (zh) 2021-08-02 2021-08-02 利用胸腹部正位x线片检测脊柱骨质疏松性压缩性骨折的装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/110109 WO2023010248A1 (zh) 2021-08-02 2021-08-02 利用胸腹部正位x线片检测脊柱骨质疏松性压缩性骨折的装置

Publications (1)

Publication Number Publication Date
WO2023010248A1 true WO2023010248A1 (zh) 2023-02-09

Family

ID=85154883

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/110109 WO2023010248A1 (zh) 2021-08-02 2021-08-02 利用胸腹部正位x线片检测脊柱骨质疏松性压缩性骨折的装置

Country Status (1)

Country Link
WO (1) WO2023010248A1 (zh)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160113612A1 (en) * 2014-10-28 2016-04-28 Siemens Aktiengesellschaft Method for the fully automatic detection and assessment of damaged vertebrae
CN110097550A (zh) * 2019-05-05 2019-08-06 电子科技大学 一种基于深度学习的医学图像分割方法及系统
US20190336097A1 (en) * 2014-07-21 2019-11-07 Zebra Medical Vision Ltd. Systems and methods for prediction of osteoporotic fracture risk
CN111417980A (zh) * 2017-12-01 2020-07-14 Ucb生物制药有限责任公司 用于椎骨骨折的识别的三维医学图像分析方法和系统
CN111899880A (zh) * 2020-08-03 2020-11-06 暨南大学附属第一医院(广州华侨医院) 腰椎骨小梁负载应力改变及隐匿骨折人工风险评估方法
CN112541900A (zh) * 2020-12-15 2021-03-23 平安科技(深圳)有限公司 基于卷积神经网络的检测方法、装置、计算机设备及存储介质

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190336097A1 (en) * 2014-07-21 2019-11-07 Zebra Medical Vision Ltd. Systems and methods for prediction of osteoporotic fracture risk
US20160113612A1 (en) * 2014-10-28 2016-04-28 Siemens Aktiengesellschaft Method for the fully automatic detection and assessment of damaged vertebrae
CN111417980A (zh) * 2017-12-01 2020-07-14 Ucb生物制药有限责任公司 用于椎骨骨折的识别的三维医学图像分析方法和系统
CN110097550A (zh) * 2019-05-05 2019-08-06 电子科技大学 一种基于深度学习的医学图像分割方法及系统
CN111899880A (zh) * 2020-08-03 2020-11-06 暨南大学附属第一医院(广州华侨医院) 腰椎骨小梁负载应力改变及隐匿骨折人工风险评估方法
CN112541900A (zh) * 2020-12-15 2021-03-23 平安科技(深圳)有限公司 基于卷积神经网络的检测方法、装置、计算机设备及存储介质

Similar Documents

Publication Publication Date Title
Liang et al. A transfer learning method with deep residual network for pediatric pneumonia diagnosis
Hadj Saïd et al. Development of an Artificial Intelligence Model to Identify a Dental Implant from a Radiograph.
US10593041B1 (en) Methods and apparatus for the application of machine learning to radiographic images of animals
Goyal et al. A refined deep learning architecture for diabetic foot ulcers detection
Ver Berne et al. A deep learning approach for radiological detection and classification of radicular cysts and periapical granulomas
Sonavane et al. Dental cavity classification of using convolutional neural network
Lin et al. Tooth numbering and condition recognition on dental panoramic radiograph images using CNNs
Nguyen et al. A novel approach for evaluating bone mineral density of hips based on Sobel gradient-based map of radiographs utilizing convolutional neural network
Ryu et al. Potential role of artificial intelligence in craniofacial surgery
Liu et al. Artificial intelligence to diagnose tibial plateau fractures: an intelligent assistant for orthopedic physicians
Varçin et al. Diagnosis of lumbar spondylolisthesis via convolutional neural networks
Shen et al. Structure-Oriented Transformer for retinal diseases grading from OCT images
WO2023010248A1 (zh) 利用胸腹部正位x线片检测脊柱骨质疏松性压缩性骨折的装置
Yau et al. An adaptive region growing method to segment inferior alveolar nerve canal from 3D medical images for dental implant surgery
Xue et al. Region-of-interest aware 3D ResNet for classification of COVID-19 chest computerised tomography scans
Zhang et al. Deep learning system assisted detection and localization of lumbar spondylolisthesis
Rifa'i et al. Analysis for diagnosis of pneumonia symptoms using chest X-Ray based on Resnet-50 models with different epoch
CN115701344A (zh) 利用胸腹部正位x线片检测脊柱骨质疏松性压缩性骨折的装置
Mu et al. Automatic calcaneus fracture identification and segmentation using a multi-task U-Net
Luo et al. Dual-source computed tomography image information under deep learning algorithm in evaluation of coronary artery lesion in children with Kawasaki disease
Bui et al. Tooth localization using YOLOv3 for dental diagnosis on panoramic radiographs
Musatian et al. Medical images segmentation operations
Dos Santos Employing Advanced Deep Learning Technology for the Detection of Kidney Stones in Unenhanced Computed Tomography (CT) Imaging: A Model-Based Approach
Kulkarni et al. Deep learning approaches for detection of COVID 19 from CT image: a review
US20240156337A1 (en) Image inpainting of in-vivo images

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21952158

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE