CN115689987A - DR image-based dual-view vertebral fracture characteristic detection method - Google Patents
DR image-based dual-view vertebral fracture characteristic detection method Download PDFInfo
- Publication number
- CN115689987A CN115689987A CN202211021620.4A CN202211021620A CN115689987A CN 115689987 A CN115689987 A CN 115689987A CN 202211021620 A CN202211021620 A CN 202211021620A CN 115689987 A CN115689987 A CN 115689987A
- Authority
- CN
- China
- Prior art keywords
- image
- fracture
- vertebral
- window
- transform
- Prior art date
- Legal status (The legal status 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 status listed.)
- Pending
Links
Images
Landscapes
- Apparatus For Radiation Diagnosis (AREA)
Abstract
The invention discloses a processing method based on a double-visual-angle DR image, which is used for detecting fracture characteristics of patients suffering from pathological and osteoporosis vertebral compression. The invention realizes the acquisition of images from DR equipment and the image data preprocessing, determines the positive lateral vertebra interested area, designs a neural network to extract image characteristics, automatically positions and grades the severity of the vertebra fracture, and provides an effective method for detecting the vertebra fracture characteristics for orthopedics doctors to carry out clinical screening and early diagnosis and treatment scheme formulation of patients.
Description
Technical Field
The invention relates to a double-view angle spine fracture characteristic detection method based on DR images, belonging to the field of computer image processing.
Background
Vertebral fracture is a common fracture type in clinic, and Vertebral Compression Fractures (VCF) are easily caused due to pathological causes or osteoporosis, so that potential destructive sequelae can be caused, the life quality of a patient is reduced, the death rate is increased, and high social and economic costs are generated. In recent years, with the aging of the population of China accelerating, the incidence of diseases rises year by year, about 140 million newly-increased people are affected by the disease every year, and the disease becomes an important health problem of wide attention. Although the imaging technology of the X-ray medical imaging equipment is continuously improved, the image still has the problems of noise artifact and uneven gray scale, deviation is inevitably generated when a doctor visually identifies a fracture part, and the accuracy rate of fracture detection is influenced. The accurate and effective fracture characteristic detection method is adopted, the treatment and prognosis effects of the patient can be obviously improved, the survival rate of the patient is improved, and the service life of the patient is prolonged.
A semi-quantitative assessment method currently commonly used to analyze vertebral fractures determines the fracture location based on the shape of the vertebral body and the percentage of anterior, posterior and medial height reduction and endplate surface loss of the vertebral body based on visual observation by the physician, and classifies the severity of vertebral fractures as Grand 1 (mild) to Grand 3 (severe) according to the height reduction and surface loss levels, where Grand 1 is a height loss of 20-25% or a surface loss of 10-20%, grand 2 is a height loss of 25-40% or a surface loss of 20-40%, and Grand 3 is a height loss or a surface loss of more than 40%. This is a practical and reproducible spinal fracture assessment method, easy to implement in clinical practice, suitable for epidemiological studies and clinical efficacy trials, but this method requires a professional and experienced doctor.
The X-ray DR plain film is a quick, widely-used and low-cost technology, the radiation dose is relatively low, the X-ray DR plain film is usually used as a preferred checking mode of the vertebral fracture, and due to the composition and the characteristics of DR imaging equipment, system intrinsic noise and random noise exist in a DR image, the noise reduces the contrast of the image and influences the characteristic detection of the vertebral fracture. Clinically, patients are screened for fractures by observing DR images through naked eyes, different doctors may give different conclusions, and the semi-quantitative evaluation method requires large calculation workload, so that the pressure of radiologists on processing image data is increased, and deviation can occur in the severity classification of vertebral fractures. Aiming at the existing problems, the invention provides a double-view angle spine fracture characteristic detection method based on DR images by simulating the process of observing two visual angle plain films at the positive side of DR, which can be used for positioning and grading the severity of spine fracture, improving the fracture characteristic detection accuracy of radiologists and fracture doctors and making a more reasonable treatment scheme for patients.
Disclosure of Invention
The invention aims to provide a double-view spine fracture characteristic detection method based on DR images, which provides a reference for fracture characteristic detection and treatment scheme formulation of radiologists and fracture doctors.
The invention adopts the following technical scheme for solving the problems:
the invention provides a double-view vertebral fracture characteristic detection method based on DR images, which comprises the following specific steps:
And denoising random noise in the image acquired from the DR imaging system by a median filtering method and an anisotropic diffusion filtering method, and filtering noise points causing the fuzzy spine vertebral body edge in the DR image. And adjusting and setting the window width and the window level of the DR image according to the DICOM format data labels or the maximum value and the minimum value of the original pixel, and mapping the original pixel to the 8-bit image. Adopts self-adaptive histogram equalization for image enhancement and enhances the contrast of DR image and inhibits noise at the same time, so that the distinction of vertebra and soft tissue is more obvious,
step 2, obtaining the Region of interest (ROI) of different centrums in the positive side double-angle DR plain film by using the preprocessed image and the label of a radiologist with abundant clinical experience,
cutting out image blocks (ROI patches) containing vertebral fracture ROI according to the DR image and corresponding labeling information, simultaneously obtaining image blocks of normal vertebral body ROI from adjacent positions, adjusting the sizes of the image blocks to be 224 x 224, endowing the image blocks with fracture and normal classification labels y,
further, the step 3 specifically includes the following steps: the ROI of the front and back positive position sheets and the lateral position sheet comprises a normal centrum image block I N And fractured vertebral body image block I F From CNNmodel, the feature map f (I) is obtained N ,I F ) And predictive classification
The classification loss function (Class loss) of normal or fractured bones can be calculated, i.e.
Where N is the total number of training samples. In the neural network training process, the classification loss function is used as a constraint term to optimize the CNNmodel, and parameters in the pre-training network model are finely adjusted, so that the characterization capability of image features is kept, and the statistical distribution of spine fracture data can be adapted.
Feature map f (I) extracted from ROI image blocks by the first stage CNNmodel N ,I F ) In the second stage, f (I) of dimension H multiplied by W multiplied by C is mixed through the characteristic image blocks N ,I F ) And converting the characteristic into H/4 xW/4 x2C dimensions, and respectively inputting the dimensions into a Multi-head transform and a Shifted-window transform to avoid the loss of useful characteristic information due to downsampling, wherein H, W and C respectively represent the height, width and channel number of the characteristic diagram. In a Multi-head transform, three 1X 1 convolutions would input the feature map
Is projected to
Flattened and transposed into a sequence of size n x d, where d is the dimension of the embedding, Q, K, V flattened and transposed into a sequence of size n x d, where n = H x W, whereby the output is obtained
This allows the model to jointly infer attention from different representation subspaces for fracture localization. In the Shifted-window transform, the characteristic diagram calculates the self-attention in a Shifted local window, the window is uniformly divided in a non-overlapping mode, the huge calculation amount of the global attention is reduced, and the output is
Where B is the shift bias. Is prepared from O MHT And O SWT Performing characteristic splicing to obtain f cat ＝Cancate(O MHT ,O SWT ) Further, the fracture position t = (t) predicted by the neural network can be known x ,t y ,t w ,t h ) And a localization loss function
Wherein (x, y, w, h) is the horizontal and vertical coordinates, width and height of the prediction frame, smooth L1 Is a smooth L1 loss function.
The method is characterized in that a positive lateral double-view DR image is adopted, and a Multi-head transform and a Shifted-window transform parallel network module are designed, and the prediction results of different severity grading probabilities of fracture and possible lesion positions in the input ROI image block can be obtained through the neural network training and optimizing process.
And 4, according to the fracture severity grade output by the neural network and the probability value of the prediction box, the method realizes fracture positioning and severity grade of the double-visual-angle DR image.
And 5, determining the fracture condition of the patient and formulating an effective treatment scheme by a clinical radiologist and an orthopedist through observing the actual DR image and detecting and analyzing the vertebral fracture characteristics and combining the positioning and severity grading results of the vertebral fracture.
The whole flow diagram of the invention is shown in figure 1.
Compared with the prior art, the neural network of the CNN model and the transform network module is constructed by utilizing the double-view DR image and combining the scoliography information of the positive side position, the advantages of the image representation of the neural network are fully exerted, the detection rate of the vertebral fracture is favorably improved, and the beneficial effects are as follows: the invention realizes automatic vertebra fracture positioning and severity grading in the double-visual-angle DR image, and provides an effective vertebra fracture characteristic detection method for clinical screening and treatment scheme formulation of pathological and osteoporosis type vertebra compression fracture patients.
Drawings
FIG. 1 is a schematic diagram of a vertebral fracture characterization detection process according to the present invention;
figure 2 is a block diagram of a neural network in a method,
FIG. 3 shows the results of neural network derived spine fracture location and severity grading in the method;
FIG. 4 is a schematic view of the overall process of the present invention.
Detailed Description
The present invention is further illustrated by the following examples, which are intended to be purely exemplary and are not intended to limit the scope of the invention, as various equivalent modifications of the invention will occur to those skilled in the art upon reading the present disclosure and fall within the scope of the appended claims.
Example 1: referring to fig. 1-4, the invention provides a dual-view vertebral fracture characteristic detection method based on a DR image, which comprises the following specific steps:
And denoising random noise in the image acquired from the DR imaging system by a median filtering method and an anisotropic diffusion filtering method, and filtering noise points causing the fuzzy spine vertebral body edge in the DR image. And adjusting and setting the window width and the window level of the DR image according to the DICOM format data labels or the maximum value and the minimum value of the original pixel, and mapping the original pixel to the 8-bit image. Adopts self-adaptive histogram equalization for image enhancement and enhances the contrast of DR image and inhibits noise at the same time, so that the distinction of vertebra and soft tissue is more obvious,
step 2, obtaining the Region of interest (ROI) of different centrums in the positive side double-angle DR plain film by using the preprocessed image and the label of a radiologist with abundant clinical experience,
cutting out image blocks (ROI patches) containing vertebral fracture ROI according to the DR image and corresponding labeling information, simultaneously obtaining image blocks of normal vertebral body ROI from adjacent positions, adjusting the sizes of the image blocks to be 224 x 224, endowing the image blocks with fracture and normal classification labels y,
further, the step 3 specifically includes the following steps: the ROI of the front and back positive position sheets and the lateral position sheet comprises a normal centrum image block I N And fractured vertebral body image block I F From CNNmodel, the feature map f (I) is obtained N ,I F ) And predictive classification
Can be used forCalculating the Classification loss function (Class loss) of normal or fractured bones, i.e.
Where N is the total number of training samples. In the neural network training process, the classification loss function is used as a constraint term to optimize the CNNmodel, and parameters in the pre-training network model are finely adjusted, so that the characterization capability of image features is kept, and the statistical distribution of spine fracture data can be adapted.
Feature map f (I) extracted from ROI image blocks by the first stage CNNmodel N ,I F ) In the second stage, f (I) of dimension H multiplied by W multiplied by C is mixed through the characteristic image blocks N ,I F ) And converting the characteristic into H/4 xW/4 x2C dimensions, and respectively inputting the dimensions into a Multi-head transform and a Shifted-window transform to avoid the loss of useful characteristic information due to downsampling, wherein H, W and C respectively represent the height, width and channel number of the characteristic diagram. In a Multi-head transform, three 1X 1 convolutions would input the feature map
Is projected to
Flattened and transposed into a sequence of size n x d, where d is the embedded dimension, Q, K, V flattened and transposed into a sequence of size n x d, where n = H x W, whereby the output can be derived
This allows the model to jointly infer attention from the different representation subspaces for fracture localization. In the Shifted-window transform, the feature map calculates the self-attention in a Shifted local window, the window is uniformly divided in a non-overlapping way, the huge calculation amount of the global attention is reduced, and the output is
Where B is the shift bias. Is prepared from O MHT And O SWT Performing characteristic splicing to obtain f cat ＝Cancate(O MHT ,O SWT )，Further knowing the fracture position predicted by the neural network t = (t) x ,t y ,t w ,t h ) And a localization loss function
Wherein (x, y, w, h) is the horizontal and vertical coordinates, width and height of the prediction frame, smooth L1 Is a smooth L1 loss function.
The method is characterized in that a positive lateral double-view DR image is adopted, and a Multi-head transform and a Shifted-window transform parallel network module are designed, and the prediction results of different severity grading probabilities of fracture and possible lesion positions in the input ROI image block can be obtained through the neural network training and optimizing process.
And 4, according to the fracture severity grade output by the neural network and the probability value of the prediction box, the method realizes fracture positioning and severity grade of the double-visual-angle DR image.
And 5, determining the fracture condition of the patient and formulating an effective treatment scheme by a clinical radiologist and an orthopedist through observing the actual DR image and detecting and analyzing the vertebral fracture characteristics and combining the positioning and severity grading results of the vertebral fracture.
The whole flow diagram of the invention is shown in figure 1.
Example 2: referring to fig. 1 to 3, a dual-view spine fracture characteristic detection method based on a DR image includes the following specific steps:
Specifically, the step 1 comprises: and filtering noise points in the DR image by using a median filtering method and an anisotropic diffusion filtering method. And setting the window width and the window level of the DR image and carrying out image enhancement by adaptive histogram equalization for limiting the contrast.
And 2, obtaining the ROI of the spine in the double-angle DR image at the front side position through the preprocessed image and the label.
Specifically, step 2 cuts out image blocks containing the vertebral fracture ROI according to the DR image and corresponding labeling information, obtains image blocks of the normal vertebral ROI from adjacent positions, and gives fracture and normal classification labels y to the image blocks.
And 3, referring to fig. 2, inputting the ROI image block into a designed neural network (neural network), training a pre-training model (CNNmodel) obtained by ImageNet on a data set, mixing the image block features (hybrid patch), respectively inputting the image block features into a Multi-head transform module and a Shifted-window transform network module, performing feature splicing, and obtaining the prediction results of the different severity classification probabilities of the fracture and the possible lesion positions in the input ROI image block by the method through the neural network training and optimization process.
And 4, according to the fracture severity grading and the probability value of the prediction box output by the neural network, realizing fracture positioning and severity grading of the double-visual-angle DR image by the method.
And 5, the clinical radiologist and the orthopedist determine an effective treatment scheme after determining the fracture condition of the patient by observing the actual DR image and analyzing the detection result of the vertebral fracture characteristics to position and grade the severity of the vertebral fracture focus.
FIG. 2 is a block diagram of a neural network in the method, which mainly includes network modules such as CNNmodel, multi-head transform and Shifted-window transform, and main links such as Classification, hybrid patch, fracturedirection, etc. According to the DR image and corresponding labeling information, image blocks (Fractpatches) containing vertebral Fracture are cut out, normal vertebral body image blocks (Normalpatches) in adjacent positions are used as input of a CNNmodel, the obtained prediction classification result and the real Class calculation classification loss (Class loss) are obtained, output feature image blocks are mixed and then spliced after passing through a Multi-head transform and a Shifted-window transform parallel branch network module, the position and severity grading of a vertebral Fracture focus is predicted, and a positioning loss function (Location) is calculated through positioning coordinates and real labels.
FIG. 3 shows the results of the neural network-derived spine fracture location and severity ranking in the method.
The left image is a front and back positive DR image: grade 1 represents mild vertebral fracture with surface loss of 10% -20%; grade 3 represents a severe vertebral fracture with surface losses of more than 40%.
The right image is a lateral DR image: grade 1 represents mild vertebral fracture with 20% -25% loss of height; grade 3 represents a severe vertebral fracture with a height loss of more than 40%.
And (3) effect evaluation:
the method is based on a DR image dual-view vertebral fracture characteristic detection method, and provides an effective vertebral fracture characteristic detection method for clinical screening and early treatment of patients with pathological and osteoporosis vertebral compression fractures.
It should be noted that the above-mentioned embodiments are only preferred embodiments of the present invention, and are not intended to limit the scope of the present invention, and all equivalent substitutions or substitutions made on the above-mentioned technical solutions belong to the scope of the present invention.
Claims (4)
1. A double-view-angle spine fracture characteristic detection method based on DR images is characterized by comprising the following specific steps:
step 1, collecting an image of a vertebral fracture patient from DR equipment, and preprocessing image data;
step 2, obtaining interested areas of different vertebral bodies in the positive side double-angle DR plain film by using the preprocessed image and the label of a radiologist with abundant clinical experience;
step 3, inputting the ROI image block into a designed neural network, mixing the characteristics of the image block after a pre-training model CNNmodel obtained by training on an ImageNet data set, then respectively inputting the image block into a Multi-head transform module and a Shifted-window transform network module, and then performing characteristic splicing to finally obtain a prediction result of the vertebral fracture;
step 4, according to the fracture severity grade output by the neural network and the probability value of the prediction box, fracture positioning and severity grade of the double-visual-angle DR image are achieved;
and 5, positioning and grading the severity of the fracture focus by a clinical radiologist or an orthopedist by observing the actual DR image and combining the detection result of the vertebral fracture characteristics.
2. The method as claimed in claim 1, wherein the method for detecting the folding characteristics of the dual-view spine based on the DR image is characterized in that random noise in the DR image is denoised by a median filtering method and an anisotropic diffusion filtering method, noise points causing blurring of the spine and vertebral body edges in the DR image are filtered, the window width and window position of the DR image are set according to DICOM format data labels or original pixel values, the original pixels are mapped to an 8-bit image, and image enhancement is performed by adaptive histogram equalization for limiting the contrast, so that the contrast of the DR image is enhanced while the noise is suppressed, and the spine and soft tissue are distinguished more obviously.
3. The method for detecting characteristics of vertebral fracture based on two visual angles of DR image as claimed in claim 1, wherein according to DR image and corresponding labeling information, cutting out image blocks containing vertebral fracture ROI, obtaining image blocks of normal vertebral ROI from adjacent positions, adjusting size of image blocks, and giving fracture and normal classification labels to image blocks.
4. The method for detecting folding features of spine with two viewing angles based on DR image as claimed in claim 1, wherein said step 3 is specifically as follows: the ROI of the front and back positive position films and the lateral position film comprises a normal centrum image block I N And fractured vertebral body image block I F First stage feature map f (I) extracted from ROI image blocks by CNNmodel N ,I F ) And predictive classification
In the second stage, f (I) of dimension H multiplied by W multiplied by C is mixed through characteristic image blocks N ,I F ) Converting into H/4 xW/4 x2C dimension, inputting into Multi-head transform and Shifted-window transform, wherein H, W, C represent height, width and channel number of feature map, and inputting feature map by three 1 x 1 convolutions in Multi-head transform
Is projected to
Flattened and transposed into a sequence of size n x d, where d is the embedded dimension, Q, K, V flattened and transposed into a sequence of size n x d, where n = H x W, whereby the output can be derived
This allows the model to jointly infer attention from different representation subspaces for fracture localization, in the Shifted-window transform, the feature map calculates self-attention within a Shifted local window, the window is uniformly divided in a non-overlapping manner, and the output is
Wherein B is a shift bias, and MHT and O SWT Performing characteristic splicing to obtain f cat ＝Cancate(O MHT ,O SWT ) Further, it can be seen that the fracture position t = (t) predicted by the neural network x ,t y ,t w ,t h ) And a localization loss function
Wherein (x, y, w, h) is the horizontal and vertical coordinates, width and height of the prediction frame, smooth L1 In order to smooth the L1 loss function,
the method is characterized in that a double-view DR image at the front side is adopted, and a Multi-head transform and a Shifted-window transform parallel network module are designed, and through the neural network training and optimizing process, the method can detect the fracture characteristics in the DR image to obtain the lesion position and severity grading of the fracture.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211021620.4A CN115689987A (en) | 2022-08-24 | 2022-08-24 | DR image-based dual-view vertebral fracture characteristic detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211021620.4A CN115689987A (en) | 2022-08-24 | 2022-08-24 | DR image-based dual-view vertebral fracture characteristic detection method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115689987A true CN115689987A (en) | 2023-02-03 |
Family
ID=85060996
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211021620.4A Pending CN115689987A (en) | 2022-08-24 | 2022-08-24 | DR image-based dual-view vertebral fracture characteristic detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115689987A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116403728A (en) * | 2023-06-09 | 2023-07-07 | 吉林大学第一医院 | Data processing device for medical treatment data and related equipment |
| CN117274418A (en) * | 2023-10-08 | 2023-12-22 | 北京长木谷医疗科技股份有限公司 | CT image generation method, device and equipment based on positive side X-ray image |
-
2022
- 2022-08-24 CN CN202211021620.4A patent/CN115689987A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116403728A (en) * | 2023-06-09 | 2023-07-07 | 吉林大学第一医院 | Data processing device for medical treatment data and related equipment |
| CN116403728B (en) * | 2023-06-09 | 2023-08-29 | 吉林大学第一医院 | Data processing device for medical treatment data and related equipment |
| CN117274418A (en) * | 2023-10-08 | 2023-12-22 | 北京长木谷医疗科技股份有限公司 | CT image generation method, device and equipment based on positive side X-ray image |
| CN117274418B (en) * | 2023-10-08 | 2024-04-02 | 北京长木谷医疗科技股份有限公司 | CT image generation method, device and equipment based on positive side X-ray image |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110599508B (en) | Artificial intelligence-based spine image processing method and related equipment | |
| CA2188394C (en) | Automated method and system for computerized detection of masses and parenchymal distortions in medical images | |
| CN105513077B (en) | A kind of system for diabetic retinopathy screening | |
| CN115689987A (en) | DR image-based dual-view vertebral fracture characteristic detection method | |
| CA2744690C (en) | Image analysis | |
| CN111951221A (en) | Glomerular cell image identification method based on deep neural network | |
| He et al. | Breast image pre-processing for mammographic tissue segmentation | |
| CN109919254B (en) | Breast density classification method, system, readable storage medium and computer device | |
| CN115661149A (en) | Lung image processing system based on lung tissue data | |
| CN111784704A (en) | MRI coxitis disease segmentation and classification automatic quantitative grading sequential method | |
| CN114693682B (en) | Spine feature identification method based on image processing | |
| CN106780492B (en) | Method for extracting key frame of CT pelvic image | |
| Zhang et al. | Automatic background recognition and removal (ABRR) in computed radiography images | |
| TWI743969B (en) | System and method for detection of thoracic and lumbar vertebral fractures | |
| Sergeant et al. | Same task, same observers, different values: the problem with visual assessment of breast density | |
| WO2022153100A1 (en) | A method for detecting breast cancer using artificial neural network | |
| Tatjana et al. | Computer-aided Analysis and Interpretation of HRCT Images of the Lung | |
| Srinivasan et al. | Segmentation and classification of cervix lesions by pattern and texture analysis | |
| CN116797597B (en) | Three-stage full-network-based full detection method and system for degenerative spinal diseases | |
| CN117876402B (en) | Intelligent segmentation method for temporomandibular joint disorder image | |
| CN117911716B (en) | Arthritis CT image feature extraction method based on machine vision | |
| Sulaiman et al. | A Convolutional Neural Network Model for Image Enhancement of Extremely Dense Breast Tissue in Digital Breast Tomosynthesis Images | |
| KR20230008279A (en) | Method, apparatus and computer program for analysis of spondylitis | |
| Hajnal et al. | Classifying mammograms by density: rationale and preliminary results | |
| Ganeson | Detection of Distrustful Masses in Digital Mammograms |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |