A method for detecting fractured wrist area images in children

Abstract

The invention discloses a method for detecting an image of a child's wrist fracture region, and relates to the field of computer vision. The method constructs an improved YOLOv8 model, which is based on a YOLOv8 network, adds an LSKNet module for dynamically adjusting the spatial receptive field in front of the SPPF module at the end of the Backbone in the YOLOv8 network, adds a channel feature fusion CFFNet module between the Backbone and the Neck, and replaces the regression box loss function with the WIoU v3 loss function at the Head end. After training the improved YOLOv8 model, an optimal child's wrist fracture detection model is obtained, and the optimal child's wrist fracture detection model can be used to detect the child's wrist fracture region in the image. The present invention can accurately identify fracture regions of different sizes, and improves the detection accuracy.

Description

Technical Field

[0001] The invention relates to the field of computer vision, and in particular to a method for detecting images of a children's wrist fracture area. Background Art

[0002] Fractures are an important global public health issue that greatly affects people's quality of life. X-Ray, Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) are three common devices used for fracture detection. Compared with the other two devices, X-Ray is relatively low in cost, simple to operate and has a fast film output speed, making it the most widely used fracture detection device. At present, the interpretation of X-Ray images still adopts a manual method that combines the skills of radiologists and surgeons, but this requires radiologists who have undergone long-term professional training in this field. According to surveys, the proportion of X-Ray images that are misunderstood has reached 26%. In addition, manual interpretation of X-Ray images is also very time-consuming.

[0003] The pediatric department of a hospital often receives a large number of fracture cases, most of which are wrist fractures in children. Children are still in the growth and development period, which will affect the diagnosis of wrist fractures. The paper "DeepLOC: Deep Learning-based Bone Pathology Localization and Classification in Wrist X-ray" (Razan Dibo, Andrey Galichin, Pavel Astashev, et al, arXiv preprint arXiv:2308.12727) points out that the missed diagnosis rate of wrist fractures in children is the highest, accounting for 32.4% of cases.

[0004] In recent years, computer-aided diagnosis systems have shown great potential in helping radiologists perform accurate and effective medical image analysis. With the development of deep learning technology, researchers have begun to apply deep neural network models to the automatic detection of fractures, which is a research hotspot in the field of computer vision. However, research on automatic detection technology for wrist fractures is limited. F,et al,Scientific Data,2022,9(1):222) released a children's wrist trauma X-Ray dataset for researchers to use and study. In 2023, Ultralytics proposed the YOLOv8 model, which can achieve a good balance between accuracy and inference speed and is more suitable for real-time detection tasks.

[0005] In children's wrist X-ray fracture images, some fracture instances are small in size and difficult to identify based on their external features alone. Accurate detection of these targets usually requires extensive contextual information. At the same time, annotated fracture datasets often contain low-quality annotated samples, resulting in low detection accuracy. Summary of the invention

[0006] The purpose of the present invention is to provide a method for detecting images of fractured wrist regions of children, which can accurately identify fractured regions of different sizes and improve detection accuracy.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A method for detecting a child's wrist fracture area image, the method comprising:

[0009] Construct an improved YOLOv8 model; the improved YOLOv8 model is based on the YOLOv8 network, and an LSKNet module for dynamically adjusting the spatial receptive field is added in front of the SPPF module at the end of the Backbone in the YOLOv8 network, a channel feature fusion CFFNet module is added between the Backbone and the Neck, and the regression box loss function is replaced with the WIoU v3 loss function at the Head end;

[0010] Establishing a children's wrist trauma image dataset; the children's wrist trauma image dataset includes multiple children's wrist trauma X-Ray images and fracture marks;

[0011] Using the children's wrist trauma image dataset to train an improved YOLOv8 model, an optimal children's wrist fracture detection model is obtained;

[0012] The X-Ray image of the child's wrist to be detected is input into the optimal child's wrist fracture detection model to determine the child's wrist fracture area in the X-Ray image of the child's wrist to be detected.

[0013] Optionally, the LSKNet module for dynamically adjusting the spatial receptive field includes: a first Conv2d layer, a first BatchNorm2d layer, a first GELU, a LSKBlock module, and a second Conv2d layer connected in sequence;

[0014] The first Conv2d layer is used to perform convolution dimensionality reduction on the output of the fifth C2f feature extraction module in the Backbone module;

[0015] The first BatchNorm2d layer is used to batch normalize the feature maps after convolutional dimension reduction in the Conv2d layer;

[0016] The first GELU is used to perform nonlinear operations on the feature maps after batch normalization in the first BatchNorm2d layer;

[0017] The LSKBlock module is used to dynamically adjust the receptive field of the feature extraction backbone in the feature map after the first GELU nonlinear operation;

[0018] The second Conv2d layer is used to further extract features from the feature map output by the LSKBlock module and then input it into the SPPF module.

[0019] Optionally, the LSKBlock module includes: large kernel convolution and spatial kernel selection mechanism;

[0020] The large kernel convolution uses deep convolution with different receptive fields to extract features from the feature map after the first GELU nonlinear operation, and each extracted feature map is transformed into a 1×1 convolution layer. According to the formula Further extract features; among them, Represents a 1×1 convolutional layer The feature map of the i-th depth convolution is further extracted, represents the convolution layer corresponding to the i-th depth convolution, M i represents the feature map extracted by the i-th depth convolution, i∈[1,N], N represents the number of decomposed kernels;

[0021] The spatial kernel selection mechanism is implemented through the convolutional layer According to the formula Get the attention feature S; where, Represents the spatial attention map corresponding to the i-th depth convolution;

[0022] The feature map after the first GELU nonlinear operation and the attention feature S are calculated according to the formula Perform element-by-element multiplication to obtain the final output feature map Y; where X represents the feature map after the first GELU nonlinear operation, Represents element-wise product.

[0023] Optionally, the CFFNet module includes a second BatchNorm2d layer, a LightConv module, a second GELU and a third Conv2d layer connected in sequence;

[0024] The second BatchNorm2d layer is used to batch normalize the feature maps output by Backbone;

[0025] The LightConv module is used to fuse the channel information and compression parameters of the feature map after batch normalization;

[0026] The second GELU is used to perform nonlinear operations on the fused feature map;

[0027] The third Conv2d layer is used to perform convolution dimensionality reduction on the feature map after nonlinear operation.

[0028] Optionally, the LightConv module includes: a fourth Conv2d, a third BatchNorm2d layer, a ReLU and a DWConv connected in sequence.

[0029] Optionally, the WIoU v3 loss function is:

[0030]

[0031]

[0032]

[0033]

[0034] in, is the WIoU v3 loss function; is the WIoU v1 loss function; β is the outlier used to describe the quality of the bounding box; α and δ are hyperparameters; x and y are the coordinates of the predicted bounding box; x gt and gt is the coordinate of the real bounding box; W g and H g is the width and height of the minimum enclosed bounding box; * indicates separation from the computational graph; is the intersection-over-union loss function; W i and H i is the cross width and cross height between the predicted bounding box and the true bounding box; w and h are the width and height of the predicted bounding box, w gt and h gt are the width and height of the true bounding box; γ is the non-monotonic focusing coefficient.

[0035] Optionally, the bounding box regression loss function in each decoupled head of the improved YOLOv8 model is for:

[0036]

[0037] Where DFL stands for distributed focal loss.

[0038] Optionally, a children's wrist trauma image dataset is established, specifically including:

[0039] The GRAZPEDWRI-DX dataset consisting of multiple X-Ray images of children’s wrist injuries was selected;

[0040] Annotate images in the GRAZPEDWRI-DX dataset; the annotated categories include: fracture, bone abnormality, bone lesion, foreign body, metal, periosteal reaction, pronation sign, soft tissue, and text;

[0041] All X-Ray images and annotations of children's wrist trauma are combined to form a children's wrist trauma image dataset.

[0042] Optionally, the improved YOLOv8 model is trained using the children's wrist trauma image dataset to obtain an optimal children's wrist fracture detection model, specifically including:

[0043] Dividing the children's wrist trauma image dataset into a training set, a validation set and a test set according to a preset ratio;

[0044] Use the training set and validation set to train the improved YOLOv8 model;

[0045] The trained YOLOv8 model was tested using the test set to obtain the optimal model for detecting wrist fractures in children.

[0046] Optionally, during the training of the improved YOLOv8 model, data enhancement is performed on the training set; the data enhancement includes HSV hue, HSV saturation, HSV value, conversion, scaling, flipping, mosaic and mixed enhancement.

[0047] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0048] A method for detecting a child's wrist fracture area image in an embodiment of the present invention optimizes the backbone network by adding an LSKNet module in front of the SPPF module at the end of the Backbone in the YOLOv8 network to dynamically adjust the receptive field of the feature extraction backbone, more effectively process the change range of the detected object, and enable the model to effectively learn local and global context information; a channel feature fusion CFFNet module is added between the Backbone and the Neck for channel information fusion and feature refinement; the regression box loss function of the original YOLOv8 model is replaced with the WIoU v3 loss function at the Head end, while reducing the competitiveness of the high-quality anchor box, reducing the harmful gradient generated by the low-quality samples; after training the improved YOLOv8 model, the optimal child wrist fracture detection model can be used to accurately identify fracture areas of different sizes, thereby improving the detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 A schematic diagram of a flow chart of a method for detecting a fractured wrist region in children provided by an embodiment of the present invention;

[0051] Figure 2 A schematic diagram of the improved YOLOv8 model structure provided by an embodiment of the present invention;

[0052] Figure 3 A schematic diagram of the network structure of the LSKNet module provided in an embodiment of the present invention;

[0053] Figure 4 A schematic diagram of the LSKBlock module provided in an embodiment of the present invention;

[0054] Figure 5 A schematic diagram of a CFFNet module provided in an embodiment of the present invention;

[0055] Figure 6 A schematic diagram of a LightConv module provided in an embodiment of the present invention;

[0056] Figure 7 This is a diagram of the detection results of a child's wrist fracture provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] like Figure 1 As shown, a method for detecting a child's wrist fracture region image in this embodiment includes the following steps:

[0060] Step 1: Build an improved YOLOv8 model; the improved YOLOv8 model is based on the YOLOv8 network as the basic architecture, and adds an LSKNet module that dynamically adjusts the spatial receptive field in front of the SPPF module at the end of the Backbone in the YOLOv8 network, adds a channel feature fusion CFFNet module between the Backbone and the Neck, and replaces the regression box loss function with the WIoU v3 loss function at the Head end.

[0061] Compared with the original YOLOv8, the improved YOLOv8 model adds the following: an LSKNet module is added in front of the SPPF module at the end of the Backbone to dynamically adjust the receptive field of the feature extraction backbone, which can more effectively handle the range of changes in the detected object, so that the model can effectively learn local and global context information; a CFFNet module is added between the Backbone and the Neck for channel information fusion and feature refinement; the regression box loss of the original model is replaced with WIoU v3 at the Head end, which reduces the harmful gradients generated by low-quality samples.

[0062] like Figure 2 As shown in the figure, the improved YOLOv8 network model specifically includes:

[0063] Input X-Ray images of children's wrist fractures with a size of 640×640×3. Feature extraction is performed through the Backbone network with the LSKNet module. Figure 3 As shown in Figure 1, the network structure of the LSKNet module includes Conv2d, BatchNorm2d, GELU, LSKBlock, and Conv2d operations.

[0064] In the Backbone module, the output of the 5th C2f feature extraction module first passes through a 1×1 Conv2d layer for feature extraction, then is input into the BatchNorm2d layer for batch normalization, and is nonlinearly operated through GELU, and then passes through the LSKBlock module, which can dynamically adjust the receptive field of the feature extraction backbone and can effectively handle the range of changes in the detected object. The feature map output by LSKBlock is input into another 1×1 Conv2d layer for further feature extraction, and finally the output feature map is input into the SPPF module for feature extraction.

[0065] The LSKNet module can more effectively handle the changing range of detected objects by dynamically adjusting the receptive field of the feature extraction backbone. This is achieved through a spatial selection mechanism that effectively weights the features processed by a series of large deep kernels and then merges them spatially. The weights of these kernels are dynamically determined based on the input and adjust the receptive field of each target in space as needed, allowing the model to adaptively use different large kernels to more effectively learn local and global contextual information.

[0066] The network structure of the LSKBlock module is as follows:

[0067] The LSKBlock module consists of a series of large kernel convolutions and spatial kernel selection mechanisms, such as Figure 4 shown.

[0068] Large kernel convolution applies a series of decomposed deep convolutions with different receptive fields to obtain features with rich contextual information from different ranges of the input X:

[0069]

[0070] in is a depthwise convolution. Assume there are N decomposed kernels, each of which consists of a 1×1 convolution layer. Further processing:

[0071]

[0072] This allows channel mixing for each spatial feature vector. Then, a spatial kernel selection mechanism is used to dynamically select kernels for various objects based on the obtained multi-scale features. The features from the decomposed large kernel sequence are weighted by their corresponding spatial selection masks and passed through the convolutional layer. To obtain the attention feature S:

[0073]

[0074] in represents the i-th spatial attention map. The final output of LSKBlock is the element-wise product between the input features X and S:

[0075]

[0076] Add the CFFNet module between Backbone and Neck, such as Figure 5 As shown in the figure. The feature map output by the feature extraction layer SPPF is used as input, and the channel information is fused, the features are refined, and the parameters are compressed through the CFFNet network. The CFFNet network structure includes BatchNorm2d, LightConv, GELU, and 1×1 Conv2d. BatchNorm2d is used for batch normalization operations, LightConv is used to fuse the channel information and compress parameters of the feature map, GELU is a nonlinear activation function, and 1×1 Conv2d is used for convolution dimensionality reduction to reduce the number of parameters.

[0077] like Figure 6 As shown, the LightConv module includes 1×1 Conv2d, BatchNorm2d, ReLU, and 1×1 Depth-wise convolution DWConv.

[0078] The Backbone network inputs feature images of three different resolutions, 80×80×256, 40×40×512, and 20×20×512, into the Neck network. The Neck network inputs feature images of three different sizes, 20×20×512, 40×40×512, and 80×80×256, into the Head network. After feature processing by the Head network, object detection results of different sizes can be obtained. The loss function of the original YOLOv8 object detection algorithm consists of two parts: the classification loss function and the bounding box regression loss function. The bounding box regression loss function is the sum of the regression box loss and the distribution focus loss DFL.

[0079] At the head end, the regression box loss of the original model is replaced by WIoU v3. The WIoU v3 loss function is defined as WIoU v1 with a non-monotonic focusing coefficient expressed as follows:

[0080]

[0081]

[0082]

[0083]

[0084] The hyperparameters α=1.9, δ=3 are set; β is the outlier used to describe the quality of the bounding box; x and y are the coordinates of the predicted bounding box; x gt and gt is the coordinate of the true bounding box; IoU is the intersection over union loss function; W i and H i is the intersection width and intersection height between the predicted bounding box and the true bounding box; W g and H g is the width and height of the minimum enclosed bounding box; * indicates separation from the computational graph; w and h are the width and height of the predicted bounding box, w gt and h gt are the width and height of the true bounding box; γ is the non-monotonic focusing coefficient.

[0085] In the early stage of training, in order to prevent low-quality anchor boxes from falling, a Given a small momentum, where t is the period when the average precision improvement slows down significantly, and n is the number of batches. In the later stages of training, when β is large, small gradient gains are assigned to low-quality anchor boxes to minimize harmful gradients. Therefore, WIoU v3 allows the model to mask the impact of low-quality samples while paying attention to normal-quality anchor boxes.

[0086] Finally, the bounding box regression loss function in each decoupled head of the improved YOLOv8 network can be expressed as: Where DFL stands for distributed focal loss.

[0087] Step 2: Establish a children's wrist trauma image dataset; the children's wrist trauma image dataset includes multiple children's wrist trauma X-Ray images and fracture marks.

[0088] Select the GRAZPEDWRI-DX dataset:

[0089] The GRAZPEDWRI-DX dataset consists of 20,327 X-Ray images of pediatric wrist trauma from 6,091 patients treated at the Department of Pediatric Surgery at the University Hospital of Graz. Several pediatric radiologists annotated the dataset images by placing lines, bounding boxes, or polygons to mark pathologies such as fractures or periosteal reactions. The images were annotated into 9 different categories: fractures, bone abnormalities, bone lesions, foreign bodies, metal, periosteal reactions, pronation signs, soft tissue, and text. The dataset has a total of 74,459 image labels and 67,771 labeled objects.

[0090] Step 3: Use the children's wrist trauma image dataset to train an improved YOLOv8 model to obtain the optimal children's wrist fracture detection model.

[0091] Divide the dataset: Divide the dataset into training set, validation set and test set according to the preset ratio.

[0092] The improved YOLOv8 model is trained using the training set and validation set. The optimal model for detecting wrist fractures in children is obtained through training.

[0093] During the training process, data enhancement was applied to the training set, including HSV hue, HSV saturation, HSV value, translation, scaling, flipping, mosaic, and mixing enhancement. The corresponding values ​​were set as: HSV-Hue = 0.015, HSV-Saturation = 0.7, HSV-Value = 0.4, Translate = 0.2, Scale = 0.7, Fliplr = 0.5, mosaic = 1.0, mixup = 0.1. The optimal model for detecting wrist fractures in children was obtained through training.

[0094] The divided test set images are input into the trained YOLOv8 model, and the detected wrist fracture area is output. The detection results are as follows: Figure 7 As shown in Figure 2, the performance of the trained YOLOv8 model is tested.

[0095] Step 4: Input the child wrist X-Ray image to be detected into the optimal child wrist fracture detection model to determine the child wrist fracture area in the child wrist X-Ray image to be detected.

[0096] The present invention constructs an LSKNet module for dynamically adjusting the spatial receptive field and a channel feature fusion CFFNet module on the basis of the YOLOv8 model and modifies the bounding box loss function to WIoU v3. The child wrist fracture detection method based on the improved YOLOv8 can accurately identify fracture areas of different sizes and improve the detection accuracy.

[0097] The beneficial effects of the present invention are as follows:

[0098] The present invention optimizes the backbone network by adding an LSKNet module in front of the SPPF module at the end of the Backbone of the original YOLOv8 model, so as to dynamically adjust the receptive field of the feature extraction backbone, more effectively process the range of changes in the detected object, and enable the model to effectively learn local and global context information; add a CFFNet module between the Backbone and the Neck for channel information fusion and feature refinement; replace the regression box loss of the original YOLOv8 model with WIoU v3 at the Head end, while reducing the competitiveness of the high-quality anchor frame, reducing the harmful gradient generated by the low-quality samples; use the training set to train the improved YOLOv8 model, and use the online data enhancement technology on the training set to improve the model performance. Experiments show that the X-Ray image children's wrist fracture detection method based on the improved YOLOv8 proposed by the present invention can accurately detect fracture areas of various sizes and improve the detection accuracy.

[0099] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for detecting a child's wrist fracture area image, characterized in that: The method comprises: Construct an improved YOLOv8 model; the improved YOLOv8 model is based on the YOLOv8 network, and an LSKNet module for dynamically adjusting the spatial receptive field is added in front of the SPPF module at the end of the Backbone in the YOLOv8 network, a channel feature fusion CFFNet module is added between the Backbone and the Neck, and the regression box loss function is replaced with the WIoU v3 loss function at the Head end; Establishing a children's wrist trauma image dataset; the children's wrist trauma image dataset includes multiple children's wrist trauma X-Ray images and fracture marks; Using the children's wrist trauma image dataset to train an improved YOLOv8 model, an optimal children's wrist fracture detection model is obtained; Inputting the child wrist X-Ray image to be detected into the optimal child wrist fracture detection model to determine the child wrist fracture area in the child wrist X-Ray image to be detected; The CFFNet module includes a second BatchNorm2d layer, a LightConv module, a second GELU and a third Conv2d layer connected in sequence; The second BatchNorm2d layer is used to batch normalize the feature maps output by Backbone; The LightConv module is used to fuse the channel information and compression parameters of the feature map after batch normalization; The second GELU is used to perform nonlinear operations on the fused feature map; The third Conv2d layer is used to perform convolution dimensionality reduction on the feature map after the nonlinear operation; the third Conv2d layer is a 1×1 Conv2d layer; The LightConv module includes: a fourth Conv2d, a third BatchNorm2d layer, a ReLU and a DWConv connected in sequence.

2. The method for detecting the regional image of wrist fracture in children according to claim 1, characterized in that: The LSKNet module for dynamically adjusting the spatial receptive field includes: a first Conv2d layer, a first BatchNorm2d layer, a first GELU, a LSKBlock module, and a second Conv2d layer connected in sequence; The first Conv2d layer is used to perform convolution dimensionality reduction on the output of the fifth C2f feature extraction module in the Backbone module; the first Conv2d layer is a 1×1 Conv2d layer; The first BatchNorm2d layer is used to batch normalize the feature maps after convolutional dimension reduction in the Conv2d layer; The first GELU is used to perform nonlinear operations on the feature maps after batch normalization in the first BatchNorm2d layer; The LSKBlock module is used to dynamically adjust the receptive field of the feature extraction backbone in the feature map after the first GELU nonlinear operation; The second Conv2d layer is used to further extract features from the feature map output by the LSKBlock module and then input it into the SPPF module.

3. The method for detecting the regional image of the wrist fracture of children according to claim 2, characterized in that: The LSKBlock module includes: large kernel convolution and spatial kernel selection mechanism; The large kernel convolution uses deep convolution with different receptive fields to extract features from the feature map after the first GELU nonlinear operation, and each extracted feature map is transformed into a 1×1 convolution layer. According to the formula Further extract features; among them, Represents a 1×1 convolutional layer The feature map of the i-th depth convolution is further extracted, represents the convolution layer corresponding to the i-th depth convolution, M i represents the feature map extracted by the i-th depth convolution, i∈[1,N], N represents the number of decomposed kernels; Spatial kernel selection mechanism through convolutional layer According to the formula Get the attention feature S; where, Represents the spatial attention map corresponding to the i-th depth convolution; The feature map after the first GELU nonlinear operation and the attention feature S are calculated according to the formula Perform element-by-element multiplication to obtain the final output feature map Y; where X represents the feature map after the first GELU nonlinear operation, Represents element-wise product.

4. The method for detecting the fractured wrist region of children according to claim 1, characterized in that: The WIoU v3 loss function is: in, is the WIoU v3 loss function; is the WIoU v1 loss function; β is the outlier used to describe the quality of the bounding box; α and δ are hyperparameters; x and y are the coordinates of the predicted bounding box; x gt and gt is the coordinate of the real bounding box; W g and H g is the width and height of the minimum enclosed bounding box; * indicates separation from the computational graph; is the intersection-over-union loss function; W i and H i is the cross width and cross height between the predicted bounding box and the true bounding box; w and h are the width and height of the predicted bounding box, w gt and h gt are the width and height of the true bounding box; γ is the non-monotonic focusing coefficient.

5. The method for detecting the regional image of the wrist fracture of children according to claim 4, characterized in that: Bounding box regression loss function in each decoupled head of the improved YOLOv8 model for: Where DFL stands for distributed focal loss.

6. The method for detecting the regional image of the wrist fracture of children according to claim 1, characterized in that: Establish a children's wrist trauma image dataset, including: The GRAZPEDWRI-DX dataset consisting of multiple X-Ray images of children’s wrist injuries was selected; Annotate images in the GRAZPEDWRI-DX dataset; the annotated categories include: fracture, bone abnormality, bone lesion, foreign body, metal, periosteal reaction, pronation sign, soft tissue, and text; All X-Ray images and annotations of children's wrist trauma are combined to form a children's wrist trauma image dataset.

7. The method for detecting the regional image of a child's wrist fracture according to claim 1, characterized in that: The improved YOLOv8 model is trained using the children's wrist trauma image dataset to obtain the optimal children's wrist fracture detection model, which specifically includes: Dividing the children's wrist trauma image dataset into a training set, a validation set and a test set according to a preset ratio; Use the training set and validation set to train the improved YOLOv8 model; The trained YOLOv8 model was tested using the test set to obtain the optimal model for detecting wrist fractures in children.

8. The method for detecting the regional image of the wrist fracture of children according to claim 7, characterized in that: In the process of training the improved YOLOv8 model, data enhancement is performed on the training set; the data enhancement includes HSV hue, HSV saturation, HSV value, conversion, scaling, flipping, mosaic and mixed enhancement.

Images