A method and device for constructing a fracture recognition model based on spinal DR images

Abstract

The present invention relates to the technical field of chest DR medical image processing, and specifically to a method for constructing a fracture recognition model based on spinal DR images, comprising: obtaining an original data set, wherein the original data set includes spinal DR images of fracture patients and spinal DR images of normal human bodies; preprocessing the original data set; resampling the preprocessed original data set to obtain data to be trained; and constructing a fracture recognition model based on the data to be trained. By obtaining an original data set including spinal DR images of fracture patients and spinal DR images of normal human bodies, preprocessing and resampling the original data set to obtain data to be trained, and constructing a fracture recognition model based on the data to be trained, a fracture recognition model for the spinal portion of DR images is established, thereby improving the accuracy of fracture diagnosis of the spinal portion in DR images. The present invention also provides a device for constructing a fracture recognition model based on spinal DR images.

Description

Technical Field

[0001] The present invention relates to the technical field of chest DR medical image processing, and in particular to a method and device for constructing a fracture recognition model based on spinal DR images. Background Art

[0002] Fractures are a common ailment among hospital patients. They refer to a complete or partial disruption of the continuity of the bone structure. They are most common in children and the elderly, but can also occur in young and middle-aged individuals. Patients often suffer a single fracture, with a few experiencing multiple fractures. With timely and appropriate treatment, most patients can regain their original function, but a few may experience varying degrees of residual symptoms. Therefore, timely diagnosis and early treatment are crucial.

[0003] In the existing technology, fractures are usually diagnosed based on clinical symptoms and the results of imaging examinations. Traditional imaging examinations require doctors to view patient images and visually determine whether a fracture exists and the area of ​​the fracture. This places high demands on the doctor's experience and ability. At the same time, compared to CT imaging, DR imaging may become a common means of medical diagnosis due to its low cost and low-dose radiation characteristics. However, due to the obstruction of the spine by the body, DR images of the spine are not clear, making spinal fractures more difficult for doctors to identify. Therefore, there is an urgent need for a fracture recognition model for the spine in DR images to assist doctors in identifying spinal fractures and improve the accuracy of diagnosis of spinal fractures. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention proposes a method and device for constructing a fracture recognition model based on spinal DR images, which improves the accuracy of fracture diagnosis of the spinal part in DR images.

[0005] In a first aspect, the present invention provides a method for constructing a fracture recognition model based on spinal DR images.

[0006] In a first possible implementation, a method for constructing a fracture recognition model based on spinal DR images includes:

[0007] In combination with the first feasible method, in the second feasible method, obtaining the original data set includes: obtaining spinal DR images of several fracture patients, and labeling the spinal DR image of each fracture patient as positive, and marking a rectangular box in the fracture area; obtaining spinal DR images of several normal human bodies, and labeling the spinal DR image of each normal human body as negative.

[0008] In combination with the first implementable manner, in a third implementable manner, the original data set is preprocessed, including: performing noise reduction processing on the original data set; and performing image enhancement processing on the original data set after noise reduction.

[0009] In combination with the third implementable manner, in a fourth implementable manner, the original data set is subjected to noise reduction processing, including: using guided filtering to perform noise reduction processing on each spinal DR image in the original data set.

[0010] In combination with the fourth possible implementation method, in a fifth possible implementation method, image enhancement processing is performed on the original data set after noise reduction, including: using gamma transform to enhance the image of each spinal DR image after noise reduction to obtain each spinal DR enhanced image.

[0011] In combination with the fifth feasible method, in the sixth feasible method, the preprocessed original data set is resampled to obtain the data to be trained, including: enlarging each spinal DR enhanced image by one time in the X and Y dimensions to obtain each spinal DR enlarged image; reducing each spinal DR enhanced image by one time in the X and Y dimensions to obtain each spinal DR reduced image; sampling each spinal DR enlarged image and each spinal DR reduced image respectively; and determining each spinal DR enhanced image, the sampled spinal DR enlarged image and each spinal DR reduced image as the data to be trained.

[0012] In combination with the first feasible method, in the seventh feasible method, a fracture recognition model is constructed according to the data to be trained, including: improving the YOLOv5 network; and training the improved YOLOv5 network using the data to be trained to obtain the fracture recognition model.

[0013] In combination with the seventh implementable method, in the eighth implementable method, the YOLOv5 network is improved, including: changing the ordinary convolution in the last two layers of the Backbone module in the YOLOv5 network to dilated convolution, and dilate=2; changing the ordinary convolution in the three output CSP2_1 of the Neck module in the YOLOv5 network to dilated convolution, and dilate=2.

[0014] In a second aspect, the present invention provides a device for constructing a fracture recognition model based on spinal DR images.

[0015] In the ninth possible implementation method, the original data set acquisition module is configured to acquire the original data set, which includes DR images of the spine of fracture patients and DR images of the spine of normal people; the preprocessing module is configured to preprocess the original data set; the training data acquisition module is configured to resample the preprocessed original data set to obtain the training data; the fracture recognition model construction module is configured to construct a fracture recognition model based on the training data.

[0016] It can be seen from the above technical solution that the beneficial technical effects of the present invention are as follows:

[0017] By obtaining the original data set including the spinal DR images of fracture patients and the spinal DR images of normal people, the original data set is preprocessed and resampled to obtain the data to be trained, and then a fracture recognition model is constructed based on the data to be trained, thereby establishing a fracture recognition model for the spinal part of DR images, assisting doctors in identifying fractures in spinal DR images, and improving the efficiency and accuracy of fracture diagnosis of the spinal part in DR images. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly describes the drawings required for the specific embodiments or the description of the prior art. Similar elements or parts are generally identified by similar reference numerals throughout the drawings. Elements or parts in the drawings are not necessarily drawn to scale.

[0019] Figure 1 A schematic diagram of a method for constructing a fracture recognition model based on spinal DR images provided by the present invention;

[0020] Figure 2 This is a schematic diagram of a device for constructing a fracture recognition model based on spinal DR images provided by the present invention. DETAILED DESCRIPTION

[0021] The following embodiments of the technical solution of the present invention will be described in detail with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore only examples and are not intended to limit the scope of protection of the present invention.

[0022] It should be noted that, unless otherwise specified, the technical or scientific terms used in this application should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0023] Combine Figure 1 As shown, this embodiment provides a method for constructing a fracture recognition model based on spinal DR images, including:

[0024] Step S01: obtaining an original data set, wherein the original data set includes DR images of the spine of a fracture patient and DR images of the spine of a normal person;

[0025] Step S02: preprocessing the original data set;

[0026] Step S03: resampling the preprocessed original data set to obtain data to be trained;

[0027] Step S04: construct a fracture recognition model based on the data to be trained.

[0028] Optionally, obtaining the original data set includes: obtaining spinal DR images of several fracture patients, and labeling the spinal DR image of each fracture patient as positive, and marking a rectangular box in the fracture area; obtaining spinal DR images of several normal human bodies, and labeling the spinal DR image of each normal human body as negative.

[0029] Optionally, the original dataset is preprocessed, including: performing noise reduction on the original dataset; and performing image enhancement on the noise-reduced original dataset. Performing noise reduction and image enhancement on each spinal DR image in the original dataset reduces the difficulty of deep learning network training.

[0030] Optionally, the original data set is subjected to noise reduction processing, including: using guided filtering to perform noise reduction processing on each spinal DR image in the original data set. The noise reduction processing on each spinal DR image by guided filtering better preserves the characteristic contours in the image and enhances the details of the features.

[0031] Optionally, image enhancement processing is performed on the original denoised dataset, including: performing image enhancement on each denoised spinal DR image using a gamma transform to obtain each enhanced spinal DR image. The gamma transform is used to enhance the guided filtered spinal DR image, thereby widening the grayscale range of high grayscale values ​​in the spinal DR image to make the bone shape area more distinct.

[0032] By preprocessing the spinal DR images through guided filtering and gamma transformation, the impact of the unclear imaging of the spinal part of the DR image on fracture detection is reduced. The image quality of the spinal bone structure part of the DR image can be optimized to the greatest extent, reducing the false positive of the model while increasing the detection rate of the fracture area.

[0033] Optionally, the preprocessed original data set is resampled to obtain data to be trained, including: doubling the X and Y dimensions of each spinal DR enhanced image to obtain each spinal DR magnified image; doubling the X and Y dimensions of each spinal DR enhanced image to obtain each spinal DR reduced image; sampling each spinal DR magnified image and each spinal DR reduced image respectively; and determining each spinal DR enhanced image, the sampled spinal DR magnified image, and each spinal DR reduced image as data to be trained. By performing image resampling operations on each spinal DR enhanced image, and then using each spinal DR enhanced image, the sampled spinal DR magnified image, and each spinal DR reduced image as training data, the improved YOLOv5 network is trained so that the improved YOLOv5 network can better learn the overall features and local subtle features of the fracture, thereby improving the accuracy of the constructed fracture recognition model. At the same time, the input training mode of spinal DR enhanced images of three scales is adopted, so that the model can learn the potential features of the fracture area from small features and global features, which provides great guarantees for the overall accuracy of the model and the good generalization ability of the model.

[0034] Optionally, constructing a fracture recognition model based on the data to be trained includes: improving the YOLOv5 network; and training the improved YOLOv5 network using the data to be trained to obtain the fracture recognition model.

[0035] Optionally, the YOLOv5 network is improved, including: replacing the ordinary convolutions in the last two layers of the Backbone module in the YOLOv5 network with dilated convolutions with a dilate value of 2; and replacing the ordinary convolutions in the CSP2_1 of the three outputs of the Neck module in the YOLOv5 network with dilated convolutions with a dilate value of 2. By replacing the ordinary convolutions in the last two layers of the Backbone module and the CSP2_1 of the three outputs of the Neck module in the YOLOv5 network with dilated convolutions, the network's receptive field is increased, allowing feature extraction to retain more global information and multi-level features, enhancing the improved YOLOv5's ability to recognize fracture area features and reducing the network's missed detection rate.

[0036] In some embodiments, the improved YOLOv5 network (R-YOLO-v5) includes an input end, an image feature extraction part, namely the Backbone module, an image feature extraction refinement part, namely the Neck module, and a prediction part, namely the Prediction module. Each spinal DR enhanced image, each sampled spinal DR magnified image, and each spinal DR reduced image are input into R-YOLO-v5 through the input end and transmitted to the Backbone module for image feature extraction. The last two layers of the Backbone module in R-YOLO-v5 are dilated convolutions (Dilated Conv), and dilate=2, which doubles the receptive field and reduces the network's missed detection rate. The Backbone module also includes a residual network convolution block. The Backbone module transmits the extracted image features to the Neck module, and the Neck module performs in-depth and refined feature extraction on the extracted image features. The Neck module includes a dilated convolution module and some short-circuit connection modules. The three outputs of the Neck module, CSP2_1, are dilated convolutions, and dilate=2. The Neck module transmits the extracted in-depth and refined features to the Prediction module. The Prediction module predicts the fracture score of each spinal DR image based on the in-depth and refined features through the softmax activation function. It determines whether there is a fracture based on the score. If there is a fracture, it outputs 1 and predicts the coordinates of the regression box of the fracture area (X1, Y1; X2, Y2). If it is normal, it outputs 0.

[0037] In some embodiments, the data to be trained includes training samples and test samples. The spinal DR images in the training samples are used as the input of the model, and the labels corresponding to the spinal DR images in the training samples are used as the output of the model, and the improved YOLOv5 network is subjected to deep learning training. The spinal DR images in the test samples are input into the trained network to obtain output values; the output values ​​of each spinal DR image are compared with the corresponding labels respectively. If they are consistent, the result is correct, and if they are inconsistent, the result is wrong. The total number of test samples and the number of correct results are counted, and the number of correct results is divided by the total number of test samples to obtain the accuracy of the trained network. When the accuracy is greater than the preset threshold, the training is stopped and the network corresponding to the accuracy is determined as the final fracture recognition model.

[0038] Combine Figure 2As shown, this embodiment provides a device for constructing a fracture recognition model based on spinal DR images, comprising: an original data set acquisition module 101, a preprocessing module 102, a module for acquiring data to be trained 103, and a fracture recognition model construction module 104. The original data set acquisition module 101 is configured to acquire an original data set, which includes spinal DR images of patients with fractures and spinal DR images of normal people; the preprocessing module 102 is configured to preprocess the original data set; the module for acquiring data to be trained 103 is configured to resample the preprocessed original data set to obtain data to be trained; and the fracture recognition model construction module 104 is configured to construct a fracture recognition model based on the data to be trained.

[0039] By obtaining the original data set including the spinal DR images of fracture patients and the spinal DR images of normal people, the original data set is preprocessed and resampled to obtain the data to be trained, and then a fracture recognition model is constructed based on the data to be trained, thereby establishing a fracture recognition model for the spinal part of DR images, assisting doctors in identifying fractures in spinal DR images, and improving the efficiency and accuracy of fracture diagnosis of the spinal part in DR images.

[0040] Optionally, the original data set acquisition module obtains the original data set in the following manner, including: obtaining spinal DR images of several fracture patients, and labeling the spinal DR image of each fracture patient as positive, and marking a rectangular box in the fracture area; obtaining spinal DR images of several normal human bodies, and labeling the spinal DR image of each normal human body as negative.

[0041] Optionally, the preprocessing module preprocesses the original data set in the following manner, including: performing noise reduction processing on the original data set; and performing image enhancement processing on the original data set after noise reduction.

[0042] Optionally, the preprocessing module performs noise reduction on the original dataset by using guided filtering to reduce the noise of each spinal DR image in the original dataset. Guided filtering can effectively protect the feature contours in the image and enhance the feature details.

[0043] Optionally, the preprocessing module performs image enhancement processing on the original data set after noise reduction by: performing image enhancement on each denoised spinal DR image using a gamma transform to obtain each enhanced spinal DR image. The gamma transform is used to enhance the spinal DR image after guided filtering, thereby widening the grayscale range of high grayscale values ​​in the spinal DR image to make the bone shape area more distinct.

[0044] Optionally, the module for acquiring data to be trained resamples the preprocessed original data set to obtain data to be trained, including: doubling each spinal DR enhanced image in the X and Y dimensions to obtain each spinal DR magnified image; halving each spinal DR enhanced image in the X and Y dimensions to obtain each spinal DR reduced image; sampling each spinal DR magnified image and each spinal DR reduced image respectively; and determining each spinal DR enhanced image, each sampled spinal DR magnified image, and each spinal DR reduced image as data to be trained. By performing image resampling operations on each spinal DR enhanced image, and then using each spinal DR enhanced image, each sampled spinal DR magnified image, and each spinal DR reduced image as data to be trained, the improved YOLOv5 network is trained so that the improved YOLOv5 network can better learn the overall characteristics and local subtle features of the fracture, thereby improving the accuracy of the constructed fracture recognition model.

[0045] Optionally, the fracture recognition model is constructed according to the data to be trained by the following method, including: improving the YOLOv5 network; training the improved YOLOv5 network using the data to be trained to obtain the fracture recognition model.

[0046] Optionally, the fracture recognition model construction model improves the YOLOv5 network by: replacing the ordinary convolutions in the last two layers of the Backbone module in the YOLOv5 network with dilated convolutions with a dilate value of 2; and replacing the ordinary convolutions in the CSP2_1 of the three outputs of the Neck module in the YOLOv5 network with dilated convolutions with a dilate value of 2. By replacing the ordinary convolutions in the last two layers of the Backbone module and the CSP2_1 of the three outputs of the Neck module in the YOLOv5 network with dilated convolutions, the network's receptive field is increased and the network's missed detection rate is reduced.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A method for constructing a fracture recognition model based on spinal DR images, characterized in that: include: Acquiring an original data set, wherein the original data set includes DR images of the spine of a fracture patient and DR images of the spine of a normal person; The original data set is preprocessed, including: Using guided filtering to perform noise reduction on each spinal DR image in the original data set; Gamma transform was used to enhance the denoised spinal DR images to obtain enhanced spinal DR images. Resampling the preprocessed original data set to obtain data to be trained, including: doubling each of the spinal DR enhanced images in the X and Y dimensions to obtain each spinal DR enlarged image; doubling each of the spinal DR enhanced images in the X and Y dimensions to obtain each spinal DR reduced image; sampling each of the spinal DR enlarged images and each of the spinal DR reduced images respectively; and determining each of the spinal DR enhanced images, the sampled spinal DR enlarged images, and each of the spinal DR reduced images as data to be trained; Constructing a fracture recognition model according to the training data, comprising: improving a YOLOv5 network, comprising: changing ordinary convolutions in the last two layers of a Backbone module in the YOLOv5 network to dilated convolutions with a dilate value of 2; changing ordinary convolutions in CSP2_1 of three outputs of a Neck module in the YOLOv5 network to dilated convolutions with a dilate value of 2; The improved YOLOv5 network is trained using the training data to obtain a fracture recognition model.

2. The method according to claim 1, characterized in that The obtaining of the original data set comprises: Obtaining spinal DR images of several fracture patients, and marking each spinal DR image of the fracture patient as positive, and marking a rectangular frame on the fracture area; A plurality of normal human spine DR images are obtained, and each normal human spine DR image is labeled as negative.

3. A device for constructing a fracture recognition model based on spinal DR images, characterized in that: include: an original data set acquisition module configured to acquire an original data set, wherein the original data set includes a DR image of the spine of a fracture patient and a DR image of the spine of a normal person; The preprocessing module is configured to preprocess the original data set, including: Using guided filtering to perform noise reduction on each spinal DR image in the original data set; Gamma transform was used to enhance the denoised spinal DR images to obtain enhanced spinal DR images. The training data acquisition module is configured to resample the preprocessed original data set to obtain the training data, including: magnifying each of the spinal DR enhanced images by one time in the X and Y dimensions to obtain each spinal DR magnified image; reducing each of the spinal DR enhanced images by one time in the X and Y dimensions to obtain each spinal DR reduced image; sampling each of the spinal DR magnified images and each of the spinal DR reduced images respectively; and determining each of the spinal DR enhanced images, the sampled spinal DR magnified images, and the sampled spinal DR reduced images as the training data; A fracture recognition model construction module is configured to construct a fracture recognition model based on the training data, including: improving the YOLOv5 network, including: changing the ordinary convolution in the last two layers of the Backbone module in the YOLOv5 network to dilated convolution with dilate = 2; changing the ordinary convolution in the CSP2_1 of the three outputs of the Neck module in the YOLOv5 network to dilated convolution with dilate = 2; The improved YOLOv5 network is trained using the training data to obtain a fracture recognition model.

Images