A mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing

Abstract

The present invention discloses a mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing, which adopts a constructed mandibular fracture detection model; the mandibular fracture detection model includes a backbone network, a feature fusion network and a detection head part; the backbone network is responsible for efficiently extracting low-level features in the image, and then further feature extraction is performed. Through the C3K2_WT module, global and local feature information is obtained to further enhance the detailed expression of the mandibular fracture line, and then the spatial attention of the feature map is enhanced through the C2PSA_EMA module. The feature fusion network aggregates feature maps from different resolutions and passes them to the detection head, and the C3K2_WT module is used again to improve the model's ability to extract the most important part of the image; the prediction results generated by the three detection heads are merged through non-maximum suppression to obtain the final result, and the position of the mandibular fracture line is accurately located. By using the present invention, the detection accuracy of the mandibular fracture line and the robustness of the model can be significantly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision, and in particular relates to a mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing. Background Art

[0002] Although maxillofacial fractures in children account for only 1%-5% of all age groups, mandibular fractures comprise approximately 40%-60% of these fractures, highlighting their importance in clinical research. The mandible is a vital, independent bony structure in the lower third of the face, defining both facial contours and supporting masticatory function. As the only mobile bone protruding from the skull base, the mandible lacks adequate protection, making it highly susceptible to fractures under high-impact forces, such as those encountered in traffic accidents. Studies have shown that mandibular fractures are a common type of fracture in humans, with common fracture sites including the median symphysis, left and right mental foramina, mandibular angle, and condylar neck. When the force is low, the fracture area typically remains stable due to muscle attachment. However, when the force is high, the mandibular condyle fractures first, absorbing the impact and protecting intracranial organs from severe injury. Mandibular fractures are often associated with complex complications, including dislocation, compression, anteversion, or retraction at the fracture site, further complicating diagnostic imaging. In CT images of mandibular fractures, the fracture area varies greatly in size and morphology, and associated symptoms such as bleeding can cause local changes in X-ray absorption, making fracture line identification more difficult. These factors directly prolong the radiologist's diagnostic time and can easily lead to misdiagnosis or missed diagnosis, hindering timely treatment for patients.

[0003] At the same time, compared with adults, treating fractures in children is more complex and challenging because they are in a growth and developmental stage with high metabolic activity and robust tissue healing capabilities. Treatment plans must balance functional recovery with the long-term effects of growth and development, placing higher demands on diagnostic accuracy and timely treatment. However, existing research primarily focuses on the imaging identification, diagnosis, and treatment of mandibular fractures in adults, with relatively insufficient research on the pediatric population.

[0004] With the rapid development of medical imaging technology, such as the continuous maturity of image fusion, target detection, and image segmentation, its role in clinical diagnosis is becoming increasingly prominent. Computer-aided diagnosis systems based on high-performance deep learning are gradually being applied to the detection and analysis of mandibular fractures, hoping to improve the consistency and accuracy of diagnostic results. However, although deep learning combined with medical imaging has made significant progress in the field of fracture detection, existing methods still face many challenges:

[0005] 1. Difficulty in data collection: Due to concerns about patient privacy and imaging differences between different devices, it is difficult to obtain a large-scale and standardized mandibular fracture dataset, which limits the training effect of the deep learning model.

[0006] 2. Significant differences in fracture areas: In CT images, the size and morphology of the fracture area vary greatly due to differences in anatomical position (coronal, sagittal, and axial), fracture site (median symphysis, mental foramen, mandibular angle, and condylar neck), and external force intensity. This makes the existing model inadequate in regional recognition accuracy.

[0007] 3. Lack of multi-scale information: Existing algorithms pay little attention to multi-scale features, which limits the model's ability to identify complex fracture areas, especially when dealing with small or atypical fractures.

[0008] 4. Identifying Interference Factors: Fractures are often accompanied by localized bleeding, and changes in X-ray absorption caused by bleeding make it more difficult to identify fracture lines on CT images. Furthermore, fractures in the same location can vary in size and morphology, further complicating algorithmic detection. Summary of the Invention

[0009] In response to the technical difficulties in mandibular fracture line detection, such as insufficient ability to recognize small fracture lines in complex fracture images, easy missed detection and false detection under the influence of interference factors, and poor performance in small target detection, the present invention provides a mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing. By introducing the wavelet convolution module and multi-scale attention mechanism, the accuracy and efficiency of the model in fracture line detection are significantly improved.

[0010] A mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing includes a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor. The computer memory stores a trained mandibular fracture detection model; the mandibular fracture detection model includes a backbone network, a feature fusion network, and a detection head.

[0011] The backbone network is used to extract feature maps P3, P4, and P5 of different scales. The backbone network first downsamples the input image using two convolutional layers, and then iteratively extracts the P3 feature map through two layers of C3k2_WT modules. Based on the P3 feature map, the P4 feature map is further extracted through the third layer of C3k2_WT modules. Based on the P4 feature map, the P5 feature map is finally generated with low resolution and high-level semantic representation capabilities through the fourth layer of C3k2_WT modules, the spatial pyramid fast pooling module SPPF, and the C2PSA_EMA modules.

[0012] The feature fusion network adopts a bidirectional fusion strategy, first performing top-down semantic enhancement, upsampling the P5 feature map by bilinear interpolation to align its size with the P4 feature map, and then splicing it along the channel dimension to generate the fused feature map P4. new , and then P4 new After upsampling, it is concatenated with the P3 feature map to generate P3 new ; Then transfer the details from bottom to top, and make the spliced ​​P3 new A C3k2_WT module is introduced to perform feature compression and nonlinear enhancement, complete downsampling, and compare it with the original P4 and P5 features. Figure 2 Second splicing to form a closed-loop fusion path and generate P4 new ', P5 new ';

[0013] The detection head part consists of three independent detection heads, with multi-scale anchor frames preset for feature maps of different resolutions; the fused P3 new 、P4 new ', P5 new 'The feature map is passed to three independent detection heads, and the small-scale anchor boxes densely cover P3 new High-resolution grid, medium-scale anchor box adapted to P4 new 'Medium-resolution grid, large-scale anchor box adapted to P5 new ' low-resolution grid; each detection head simultaneously outputs bounding box coordinates, category probability and confidence, and the outputs of the three detection heads are fused to ultimately retain high-precision predictions;

[0014] When the computer processor executes the computer program, the following steps are implemented:

[0015] The image to be detected is input into the trained mandibular fracture detection model to obtain the prediction result of mandibular fracture.

[0016] Furthermore, the backbone network first uses two convolutional layers to downsample the input image, and the formula is as follows:

[0017] O1=Conv(I,64,3,2)

[0018] O2=Conv(O1,128,3,2)

[0019] Among them, O1 represents the output feature map after the first convolution layer, and O2 represents the output feature map after the second convolution layer; the input data I performs the first convolution operation, using 64 convolution kernels of size 3×3 and a stride of 2 to obtain the output feature map O1; then the first convolution operation is performed again, using 128 convolution kernels of size 3×3 and a stride of 2 to obtain the new output feature map O2.

[0020] Furthermore, the working process of the C3k2_WT module is as follows:

[0021] First, perform feature decomposition on the feature map of the input C3k2_WT module to obtain feature maps X1 and X2;

[0022] Perform lightweight convolution feature extraction on the feature map X1 to obtain X light ;

[0023] The wavelet transform decomposes the feature map X2 into a low-frequency component F through four filters. LL and three high frequency components F LH 、F HL and F HH , and then use the inverse wavelet transform to fuse the low-frequency component and the three high-frequency components to reconstruct the feature map X2 with the same spatial resolution. IWT ;

[0024] The obtained X light and X IWT Perform fusion to obtain feature fusion results for subsequent operations.

[0025] Furthermore, lightweight convolution feature extraction is performed on the feature map X1, and the formula is as follows:

[0026] X light =Conv light (X1)=Pointwise(Depthwise(X1,k=2))

[0027] In the formula, Depthwise means using depth convolution with a 2×2 small kernel, Pointwise means using 1×1 convolution for channel fusion, and k=2 means that the size of the convolution kernel is 2×2.

[0028] Furthermore, a low-frequency component is F LL , reflecting the overall structural information of the input; the three high-frequency components are F LH 、F HL and F HH , where F LH Represents the high-frequency component in the horizontal direction, which is used to capture the edge details in the horizontal direction; F HL Represents the high-frequency component in the vertical direction, which is used to capture the edge details in the vertical direction; F HH Represents the high-frequency component in the diagonal direction and is used to extract the texture features in the diagonal direction.

[0029] Furthermore, the working process of the C2PSA_EMA module is as follows:

[0030] After the feature map is input into the C2PSA_EMA module, it passes through a 1×1 convolutional layer to compress the number of channels and decompose it into retained feature A and feature B for further processing. Feature A directly participates in the final fusion as a global feature, while feature B enters the multi-scale attention path to extract spatial and contextual information.

[0031] Apply average pooling along the height and width directions to feature B to generate B respectively. h and B w ; Then B h and B w The attention weight matrix is ​​generated by splicing and 1×1 convolution, and the attention of the salient area is further strengthened by the sigmoid activation function σ to obtain the final attention weight matrix W EMA Finally, the attention weight is applied to the feature map B through weighted operation to highlight the target area and obtain the attention weighted feature B EMA ;

[0032] To B EMA 3×3 convolution and 1×1 convolution are applied in parallel to extract local spatial correlation and global channel context dependency respectively: 3×3 convolution captures texture and edge details in the local neighborhood by expanding the spatial receptive field, and 1×1 convolution models the semantic dependency between global channels through cross-channel linear combination; then the convolution result B of the two is concatenated. FNN Through the two-layer feedforward network (including nonlinear activation function), nonlinear transformation is performed to further enhance the feature expression ability, and the enhanced result is compared with the original B EMA Add, retain the original information through residual connection, and finally obtain the enhanced feature B′;

[0033] The retained feature A and the enhanced feature B′ are concatenated in the channel dimension, and the number of output channels is restored through a 1×1 convolution to obtain the final feature map for subsequent operations.

[0034] Preferably, in the detection head part, the outputs of the three detection heads are fused across scales through non-maximum suppression (NMS).

[0035] Furthermore, the dataset for training the mandibular fracture detection model is constructed as follows:

[0036] Images were extracted from maxillofacial CT images in DICOM format and a dataset was constructed. The fracture area was annotated, including length, width, and center coordinates. The images were resized and formatted to ensure the consistency and adaptability of the input data. Finally, the dataset was expanded through data enhancement.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. Wavelet Convolution: This paper introduces wavelet convolution (C3k2_WT) into the backbone network to enhance feature extraction of fracture area details. During detection, the model preserves the structural information of the original input and captures local details at multiple scales and directions, reducing unnecessary loss in feature extraction, thereby improving fracture line detection and localization accuracy.

[0039] 2. Multi-Scale Attention Mechanism: This paper innovatively introduces a multi-scale attention mechanism (C2PSA_EMA) during the feature extraction phase. Combined with the rich, detailed features obtained through wavelet convolution, this allows the model to focus on specific areas of the image most relevant for detection, effectively zeroing in on key information in the mandibular region. Next, a cross-stage network (CSP) integrates low-level and high-level features to capture subtle variations in the fracture line at different scales, thereby improving fracture line recognition capabilities despite complex interference and fracture morphology.

[0040] 3. Multi-scale feature fusion optimization: This invention optimizes the feature fusion stage of the model by integrating the above-mentioned wavelet convolution (C3k2_WT) and multi-scale attention mechanism (C2PSA_EMA) in the feature fusion network (Neck). This module maps high-level semantic information to a finer pixel space by increasing the resolution to facilitate the detection of small objects. With the support of wavelet convolution and multi-scale attention mechanism, the key details of the image are accurately preserved. The upsampled feature map is then spliced ​​with the adjacent high-resolution feature map in the channel dimension, effectively aggregating multi-scale feature information, thereby helping the model better identify and locate complex fracture lines, especially when the fracture position changes little or is partially occluded, enhancing the ability to capture details.

[0041] 4. In specific experiments, the proposed method for mandibular fracture line detection outperformed existing fracture detection methods across various performance metrics, with significant improvements in detection accuracy, recall rate, and F1 score. Detection accuracy in the sagittal, axial, and coronal planes reached 87.97%, 89.31%, and 92.66%, respectively. This method provides new insights for mandibular fracture detection and effectively supports fracture diagnosis in clinical practice, with the potential for further application in other medical imaging diagnostics. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is the overall architecture diagram of the mandibular fracture detection model in the present invention.

[0043] Figure 2 This is the core architecture diagram of the C3K2_WT module in the present invention.

[0044] Figure 3This is the core architecture diagram of the C2PSA_EMA module in the present invention. DETAILED DESCRIPTION

[0045] The present invention will be described in further detail below with reference to the accompanying drawings and examples. It should be noted that the following examples are intended to facilitate understanding of the present invention and do not have any limiting effect on the present invention.

[0046] A mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing includes a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor. The computer memory stores a trained mandibular fracture detection model; the mandibular fracture detection model includes a backbone network, a feature fusion network, and a detection head. The overall architecture of the mandibular fracture detection model is as follows: Figure 1 shown.

[0047] The backbone network extracts multi-scale feature maps P3, P4, and P5 from the input image through layer-by-layer convolution operations. It employs progressive downsampling to reduce spatial resolution and increase channel depth, providing foundational information for subsequent feature fusion. The backbone network first downsamples the input image through two convolutional layers: the first convolutional layer uses 64 3×3 convolution kernels (stride 2) to generate feature map O1; the second convolutional layer uses 128 3×3 convolution kernels (stride 2) to generate feature map O2. O2 then iteratively passes through two layers of C3k2_WT modules to extract the high-resolution feature map P3. The output of each C3k2_WT module undergoes a 1×1 standard convolution to adjust the number of channels. After being processed by the third layer of the C3k2_WT module, the P3 feature map is downsampled through a 3×3 convolutional layer (stride 2) to generate a medium-resolution feature map P4. The P4 feature map is further processed by the fourth layer of the C3k2_WT module, the Spatial Pyramid Fast Pooling module (SPPF), and the C2PSA_EMA module to generate the low-resolution, high-level semantic feature map P5. The parameters of the convolutional layer following each C3k2_WT module are dynamically adjusted based on the feature map hierarchy.

[0048] The feature fusion network adopts a closed-loop bidirectional strategy to achieve multi-scale feature complementarity. First, perform top-down semantic enhancement, perform bilinear interpolation upsampling on the P5 feature map to align its size with the P4 feature map, and then splice it along the channel dimension to generate the fused feature map P4. new , and then P4 new After upsampling, it is concatenated with the P3 feature map to generate P3 new ; Then transfer the details from bottom to top, and make the spliced ​​P3 new A C3k2_WT module is introduced to perform feature compression and nonlinear enhancement, complete downsampling, and compare it with the original P4 and P5 features. Figure 2 Second splicing to form a closed-loop fusion path and generate P4 new', P5 new ';

[0049] The detection head part consists of three independent detection heads, with multi-scale anchor frames preset for feature maps of different resolutions; the fused P3 new 、P4 new ', P5 new 'The feature map is passed to three independent detection heads, and the small-scale anchor boxes densely cover P3 new High-resolution grid, medium-scale anchor box adapted to P4 new 'Medium-resolution grid, large-scale anchor box adapted to P5 new ' low-resolution grid; each detection head synchronously outputs bounding box coordinates, category probability and confidence, and the outputs of the three detection heads are fused to finally retain high-precision predictions.

[0050] The present invention uses a backbone network to complete layer-by-layer convolution operations to extract basic features from multi-scale input images; a gradual downsampling method is used to reduce spatial resolution and increase the depth of feature maps, providing basic information for subsequent advanced feature extraction.

[0051] The backbone network first uses two convolutional layers to downsample the input image, and the formula is as follows:

[0052] O1=Conv(I,64,3,2)

[0053] O2=Conv(O1,128,3,2)

[0054] Among them, O1 represents the output feature map after the first convolution layer, and O2 represents the output feature map after the second convolution layer; the input data I performs the first convolution operation, using 64 convolution kernels of size 3×3 and a stride of 2 to obtain the output feature map O1; then the first convolution operation is performed again, using 128 convolution kernels of size 3×3 and a stride of 2 to obtain the new output feature map O2.

[0055] The present invention is improved based on YOLOv11, and the C3K2_WT (wavelet convolution) module based on the cross-stage part (CSP) network and wavelet transform is introduced into the model to enhance the capture of local details and reduce unnecessary losses in feature extraction. The module consists of lightweight convolution and wavelet convolution, and extends the one-dimensional wavelet transform (WT) to two dimensions, so that features of different scales and directions can be effectively modeled. Specifically, the module is based on the cross-stage part (CSP) network, which divides the input feature map into two parts, one of which is subjected to feature extraction through lightweight convolution and directly fused with the output features, and the other part is subjected to wavelet transform. By constructing four wavelet transform filters, it is used to decompose the input features into low-frequency and high-frequency components in three directions. Subsequently, the CSP structure splices and fuses the two parts of the features, so that the model not only retains the structural information of the original input, but also captures local details at multiple scales and directions, which helps to identify fracture lines in complex backgrounds and improve the accuracy of the model.

[0056] like Figure 2 As shown, the working process of the C3K2_WT module is as follows:

[0057] 1.1 Input eigendecomposition:

[0058] X split =split(X)={X1,X2}

[0059] The input feature map X is decomposed into two paths: the main path: X1, for lightweight convolution; the wavelet path: X2, for wavelet transform expansion.

[0060] 1.2 Lightweight convolution feature extraction

[0061] X light =Conv light (X1)=Pointwise(Depthwise(X1,k))

[0062] Use depthwise convolution and pointwise convolution to process X1 to extract local features and efficiently fuse channel information. Depthwise means using depthwise convolution with a small 2×2 kernel, and pointwise means using 1×1 convolution for channel fusion.

[0063] 1.3 Two-dimensional wavelet transform extension:

[0064]

[0065] The formula shows that the wavelet transform decomposes the input feature map X into low-frequency and high-frequency components through four filters. LL Represents the low-frequency component, reflecting the overall structural information of the input; F LHRepresents the high-frequency component in the horizontal direction, which is used to capture the edge details in the horizontal direction; F HL Represents the high-frequency component in the vertical direction, which is used to capture the edge details in the vertical direction; F HH Represents the high-frequency component in the diagonal direction and is used to extract the texture features in the diagonal direction.

[0066] 1.4 Wavelet reconstruction (inverse wavelet transform):

[0067] X IWT =Conv transposed ([F LL F LH F HL F HH ],[X LL X LH X HL X HH ])

[0068] The formula indicates that the feature map X is reconstructed back to the original space using inverse convolution (transposed convolution) and the high and low frequency components of the above decomposition.

[0069] 1.5 Output feature fusion:

[0070] X out =Concat(X light ,X IWT )

[0071] The formula represents the output X of the lightweight convolution path light and the output X of the wavelet transform path IWT to integrate.

[0072] The specific execution process of the module is as follows:

[0073] 1. Feature Decomposition:

[0074] The input feature map X is decomposed into two parts: the first part X1 is used for lightweight convolution processing to extract local features; the second part X2 is used for wavelet transform processing and decomposed into multi-scale high and low frequency features.

[0075] 2. Lightweight convolution feature extraction:

[0076] Apply depthwise convolution (channel-by-channel processing) and point-by-point convolution (channel information fusion) to X1 to generate lightweight features X light , efficiently capture local details of the input.

[0077] 3. Wavelet transform feature expansion

[0078] Apply a two-dimensional wavelet transform to X2 and pass it through the filter F LL 、F LH 、FHL 、F HH The features are decomposed into a low-frequency component (global structure) and three high-frequency components (detail information in different directions).

[0079] 4. Wavelet Reconstruction

[0080] Use inverse wavelet transform to transform the above components X LL 、X LH 、X HL 、X HH Fusion to reconstruct feature X IWT , recovering multi-scale details and global features.

[0081] 5. Output feature fusion

[0082] The lightweight convolution output X light and wavelet reconstruction feature X IWT Perform channel concatenation (Concat) to generate the final output feature X of the module out .

[0083] This invention is based on the improvement of YOLOv11. The model introduces a multi-scale attention mechanism and enhances the spatial attention of the feature map through the C2PSA_EMA module. This helps the model focus on the specific areas in the image that are most relevant to the detection, thereby improving the performance of small objects such as (mandibular fracture lines). The module uses a cross-space learning method to handle short-term and long-term dependencies. Specifically, the model divides the input feature map into two parts, takes one part to strengthen the attention of the salient area through spatial attention generation (EMA module) and context enhancement (FNN module), and extracts local and global context information. Finally, it is spliced ​​with the other part for features, so that the model can retain more accurate location information, while weakening irrelevant information in the background area, thereby improving the accuracy of detection.

[0084] like Figure 3 As shown, the working process of the C2PSA_EMA module is as follows:

[0085] 2.1 Eigendecomposition

[0086] A,B=Split(Conv 1×1 X))

[0087] The formula indicates that through a 1×1 convolution layer, the number of channels is compressed and decomposed into two parts: the retained part A and the part B for further processing, where A directly participates in the final fusion as a global feature; B enters the multi-scale attention path to extract spatial and contextual information.

[0088] 2.2 Spatial Attention Generation (EMA Module)

[0089] Bh =Pool h (B),B w =Pool w (B)

[0090] W EMA =σ(Conv 1×1 (Concat(B h ,B w )))

[0091] B EMA =W EMA ·B

[0092] Where σ represents the sigmoid activation function, EMA represents the exponential moving average technique, and the formula represents the average pooling applied to B along the height and width directions to generate B respectively. h and B w , extracting horizontal and vertical global context information; then B h and B w Generate the attention weight matrix W by concatenation and 1×1 convolution EMA , further strengthen the attention of the salient area through the sigmoid activation function; finally, through weighted operation, the attention weight is applied to the feature map to highlight the target area.

[0093] 2.3 Context Enhancement (FNN Module)

[0094] B FNN =Conv 1×1 (Conv 3×3 (B EMA ))

[0095] B′=B EMA +B FNN

[0096] The formula represents B EMA 3×3 convolution and 1×1 convolution are applied in parallel to extract local and global context information, respectively. The convolution results are then transformed nonlinearly through a two-layer feedforward network to further enhance feature expression capabilities.

[0097] 2.4 Feature Fusion

[0098] X out =Conv 1×1 (Concat(A,B'))

[0099] The formula indicates that the initial retained A and the enhanced B′ are spliced ​​in the channel dimension, and the number of output channels is restored through a 1×1 convolution to obtain the final feature map X out .

[0100] The specific execution process of the module is as follows:

[0101] 1. Feature Decomposition:

[0102] Apply 1×1 convolution to the input feature X for channel compression and decompose it into two parts:

[0103] A: Global features, directly used for final splicing; B: Used for subsequent multi-scale attention calculation and enhancement.

[0104] 2. Spatial Attention Generation (EMA Module):

[0105] Average pooling of B: Generate B along the height direction h , generate B along the width direction w . Splicing B h and B w , generate attention weight W through 1×1 convolution EMA , and highlight the target area through sigmoid activation. EMA Applied to B, generating attention-weighted feature B EMA .

[0106] 3. Context Enhancement (FNN Module)

[0107] To B EMA Perform local (3×3 convolution) and global (1×1 convolution) feature extraction. The 3×3 convolution captures the edge details and morphological changes of the fracture line through the local spatial convolution kernel, while the 1×1 convolution strengthens the feature response related to the fracture semantics through cross-channel weight adjustment. The parallel processing of the two realizes the synergistic enhancement of spatial details and channel context, and then performs nonlinear enhancement through the feedforward network. The enhanced result is compared with the original B EMA Add them together, retain the original information through residual connection, and finally obtain the enhanced feature B′.

[0108] 4. Feature Fusion

[0109] The retained feature A and the enhanced feature B′ are concatenated, and the number of output channels is restored through 1×1 convolution to obtain the final feature map X out .

[0110] Through the above calculations, the C2PSA_EMA module achieves efficient fusion of global and local features. The module utilizes the attention weight matrix generated by the spatial attention mechanism (EMA module) to highlight salient regions, suppress background interference, and enhance the global expressiveness of the feature map. Through contextual enhancement (FNN module), it fuses local and global information, further improving the distinguishability and expressive richness of features. Finally, through channel compression and concatenation operations, it fully utilizes the global information of the initial features and the enhanced detail information, outputting a high-quality feature map that combines both global and local features.

[0111] The present invention integrates a wavelet convolution module in the feature fusion network (Neck), adopts a multi-level feature extraction architecture and a bidirectional fusion strategy to aggregate feature maps from different resolutions and pass them to the corresponding detection head. Specifically, the C3k2_WT module is integrated into the neck network to improve the speed and performance of feature aggregation. Upsampling and splicing layers are applied to combine feature maps of different scales. The C3k2_WT module is used again after splicing to ensure efficient feature aggregation, which is very important for improving accuracy in relatively blurred CT images.

[0112] At the same time, based on the above feature extraction and fusion, the model uses the detection head part (Head) of YOLOv11 to perform the positioning prediction task of the mandibular fracture line. The detection head part will output the bounding box, class probability and confidence score. Specifically, the model has three detection heads, called shallow detection head, middle detection head and deep detection head, which make predictions on feature maps of different scales respectively, and each detection head is responsible for targets of different sizes. During training, each detection head will participate in the loss calculation. During inference, all three detection heads will generate prediction results, and then merge the results through non-maximum suppression (NMS). The detection of small targets mainly relies on the shallow detection head, and the middle and deep detection heads will make a small amount of supplementary predictions for small targets (especially when small targets are in complex backgrounds).

[0113] The specific working process of the feature fusion network and detection head is as follows:

[0114] Feature maps of multiple scales (P3, P4, P5) are extracted from the backbone network (Backbone). The P3 feature map is iteratively extracted by two layers of C3k2_WT modules to output a high-resolution feature map, retaining rich spatial detail information, which is suitable for small target detection. The P4 feature map is further extracted by the third layer of C3k2_WT module on the basis of P3, with reduced resolution and balanced details and semantic information. The P5 feature map is based on P4 and passes through the fourth layer of C3k2_WT module and the spatial pyramid fast pooling module (SPPF) in turn to compress the spatial dimension, and introduce the C2PSA_EMA module to enhance channel attention, finally generating a low-resolution feature map with high-level semantic representation capabilities. The P4 and P5 feature maps will be used for secondary splicing later. To achieve cross-scale feature complementarity, this scheme adopts a bidirectional fusion strategy: first, perform top-down semantic enhancement, upsample the low-resolution feature maps (such as P4, P5) by bilinear interpolation to match their size with the high-resolution feature maps, thereby mapping high-level semantic information to a finer pixel space. Specifically, perform bilinear interpolation upsampling on P5 to align its size with P4, and splice along the channel dimension to generate the fused feature map P4. new , and then P4 new After upsampling, concatenate with P3 to generate P3 new ; Then transfer the details from bottom to top, and make the spliced ​​P3 new The C3k2_WT module is introduced to perform feature compression and nonlinear enhancement, complete downsampling, and compare it with the original P4 and P5 features. Figure 2 Second splicing to form a closed-loop fusion path and generate P4 new ', P5 new ', ensuring the two-way interaction between shallow details and deep semantics, while improving the speed and effect of multi-scale feature aggregation. new 、P4 new ', P5 new The feature maps are passed to three independent detection heads respectively, and multi-scale anchor frames are preset for feature maps of different resolutions. The small-scale anchor frames densely cover P3. new High-resolution grid and large-scale anchor box adapted to P5 new Each detection head outputs bounding box coordinates, category probability and confidence simultaneously, and the results are fused across scales through non-maximum suppression (NMS), ultimately retaining high-precision predictions.

[0115] The mandibular fracture detection system proposed in this paper significantly improves the accuracy and robustness of fracture line detection by combining lightweight feature extraction, spatial attention optimization, and multi-scale feature fusion. The C3k2_WT module and C2PSA_EMA module were introduced into the model design to achieve efficient modeling of global and local features. An optimized feature fusion strategy was used to enhance the model's ability to express complex fracture morphologies and subtle changes. The optimized feature extraction and target detection strategies further improved detection accuracy and positioning performance, providing an efficient and reliable solution for automatic mandibular fracture detection, with significant clinical application value and promotion potential.

[0116] Experimental results show that the mandibular fracture line detection system of the present invention outperforms existing mainstream methods in key performance indicators such as detection accuracy, recall rate, and F1 score. Through the multi-scale feature capture capability of wavelet transform and the significant regional enhancement of the spatial attention mechanism, the present invention can achieve richer feature expression in complex mandibular fracture detection tasks, especially showing superior performance in the detection of subtle fracture lines and small targets. The improved model shows strong robustness and adaptability in dealing with challenging fracture morphologies and complex background structures. It can effectively cope with CT imaging data of different patients and significantly reduce false detections and missed detections.

[0117] To verify the effect of the present invention, the present invention was tested below.

[0118] S1. Dataset

[0119] The raw data used in this paper comes from CT images from the Children's Hospital of Zhejiang University School of Medicine. It includes 284 fracture specimens and 300 non-fracture (no fracture or skull fracture) control specimens. This paper focuses on the automated detection of mandibular fracture lines in minors. The 284 fracture specimens range in age from 1 month to 15 years and 5 months, with 182 males and 102 females.

[0120] The dataset is large, with 156,806 images in the training set and 39,202 images in the test set, with 24,008 and 5,989 annotations, respectively. The dataset covers coronal, sagittal, and axial slice orientations, effectively capturing the multi-angle characteristics of mandibular fracture lines. Each image is accompanied by an annotation file. Medical experts annotate the fracture lines using YOLO format on Labelimg, generating a txt annotation file (with the label, the x and y coordinates of the image center, and the image's length and width). A label of 0 indicates the presence of a fracture line.

[0121] S2. Baseline and Evaluation Metrics

[0122] In order to comprehensively evaluate the effectiveness of the mandibular fracture detection model (WE-YOLO) constructed by the present invention in the target detection task, the present invention sets up a series of baseline models and systematically compares and verifies their performance through multiple evaluation indicators.

[0123] Baseline model settings: To verify the improvement effect of the WE-YOLO model in fracture line detection, this study selected the following baseline models for comparative experiments:

[0124] (1) YOLOv5 model: We use the lightweight YOLOv5 model as a comparison baseline and focus on evaluating its performance in the fracture line detection task, especially in terms of the balance between inference speed and model accuracy.

[0125] (2) YOLOv8 model: The YOLOv8 standard model is selected as the strong baseline model, focusing on comparing the differences between the WE-YOLO model and the previous generation detection model in terms of accuracy, robustness, and small target detection capabilities.

[0126] (4) YOLOv11 model: Without adding the improved module, the YOLOv11 standard model was used alone to predict the fracture line as an important baseline for comparative experiments to evaluate the improvement effect of the C3k2_WT module and the C2PSA_EMA module on model performance.

[0127] (5) RT-DETR, EfficientDet, and Faster R-CNN: We introduce mainstream target detection models as baselines, including RT-DETR, EfficientDet, and Faster R-CNN, to comprehensively evaluate the advantages of the improved YOLOv11 model in detection efficiency and accuracy, especially its adaptability when processing complex medical images.

[0128] These baseline models cover a variety of detection structures, from lightweight to large-scale, and can systematically verify the innovation and effectiveness of the WE-YOLO model. Through comparative analysis, we can comprehensively evaluate the advantages of this invention in improving small object detection capabilities, enhancing adaptability to complex backgrounds, and optimizing multi-scale feature representation.

[0129] Evaluation indicators: This paper evaluates the performance of the model using the following indicators:

[0130] (1) True Positive (TP): The number of instances that the model correctly predicts as positive. This means instances that are actually positive and correctly identified as positive by the model.

[0131] (2) False Positive (FP): The number of instances where the model mistakenly predicts the negative class as the positive class. This means that the instances actually belong to the negative class but are mistakenly labeled as positive.

[0132] (3) True Negative (TN): The number of instances that the model correctly predicts as negative. This refers to instances that are actually negative and are correctly identified as negative by the model.

[0133] (4) False Negative (FN): The number of instances where the model incorrectly predicts the positive class as the negative class. This means that the instances actually belong to the positive class but are incorrectly labeled as negative.

[0134] (5) Accuracy: It is the ratio of all correctly predicted samples (TP+TN) to the total number of samples. The calculation formula is: It measures the overall correctness of the model.

[0135] (6) Precision: The proportion of samples that are actually positive among all samples predicted by the model to be positive. The calculation formula is: This reflects how accurate the model is in predicting the positive class.

[0136] (7) Recall: Also known as sensitivity or detection rate, it is the proportion of samples that are actually positive that are correctly identified by the model. The calculation formula is: A higher Recall value indicates that the model has a low missed detection rate and can cover most of the actual fracture lines, which is crucial for reducing the risk of missed diagnosis in medical applications.

[0137] (8) F1 Score: It is the harmonic mean of precision and recall, used to comprehensively evaluate the performance of the model. It is used when both precision and recall are important. The calculation formula is: A higher F1 score indicates better overall model performance.

[0138] (9) Average Precision (AP): For binary classification problems, AP is equivalent to Precision. This metric reflects the model's ability to detect fracture lines at different confidence levels by calculating the area under the Precision-Recall curve. A higher AP50 value indicates that the model has better localization and classification performance for the fracture line area and can accurately capture the fracture line features.

[0139] S3. Implementation details

[0140] The experiments in this paper were implemented on the PyTorch 12.4 framework, using the Python 3.10 programming language, and trained on an RTX3090 GPU. To ensure the convergence and stability of the model, a unified training strategy and hyperparameter configuration were set up in the experiment. The specific implementation details are as follows:

[0141] (1) Training Epochs: The training epochs are set to 200 to ensure that the model can fully learn the data features and capture the complex patterns of fracture lines, while avoiding overfitting caused by too many training epochs. An appropriate number of training epochs helps the model gradually optimize its weights and improve detection accuracy.

[0142] (2) Batch Size: The batch size is set to 64. A moderate batch size can balance memory usage and computational efficiency while ensuring the stability of gradient calculations. When using a larger GPU, a larger batch size helps to increase training speed and also allows the optimizer to have more data support for each parameter update, thereby improving training stability.

[0143] (3) Optimizer: Adopting Adam optimizer, it can quickly find the direction in the early iteration in high-dimensional space.

[0144] (4) Learning Rate: The initial learning rate is set to 0.01, which is used to control the convergence speed of the model and prevent it from converging too fast or too slow.

[0145] The above implementation details ensure the stability and convergence of the WE-YOLO mandibular fracture detection model of the present invention during the training process, providing a reliable experimental basis for the final fracture detection accuracy.

[0146] S4. Experimental Results

[0147] 4.4.1 Quantitative Analysis

[0148] This paper systematically and quantitatively analyzes the predictive effectiveness of our mandibular fracture detection model (WE-YOLO model) by comparing the performance metrics of different models. The experimental results show that WE-YOLO significantly outperforms the YOLOv11 baseline model across all key metrics, surpassing other comparison models, verifying the model's effectiveness and robustness. The following are the main analysis results:

[0149] (1) Axial surface:

[0150] The WE-YOLO model outperforms other models in terms of AP, Precision, Recall, and F1 score on the axial plane, with AP: 0.89309, Precision: 0.880, and F1 score: 0.85.

[0151] Compared with the YOLOv11 baseline model (AP: 0.81737, Precision: 0.770, F1: 0.76), the AP is increased by about 7.6%, the Precision is increased by 11.4%, and the F1 score is increased by 9.2%, significantly enhancing the ability to capture axial surface features and detect small targets.

[0152] Yolov8+ContextAggregation performs second best in AP and F1 scores (AP: 0.883, F1: 0.831), but still lower than the improved YOLOv11.

[0153] The AP of YOLOv5, Faster R-CNN, and EfficientDet are 0.733, 0.736, and 0.715, respectively. The Precision and Recall performances are significantly behind, indicating that they are difficult to adapt to complex fracture detection scenarios.

[0154] (2) Sagittal plane:

[0155] The WE-YOLO model far outperforms other models in terms of AP, Precision, Recall, and F1 score in the sagittal plane, with AP: 0.8797, Precision: 0.830, and F1 score: 0.83.

[0156] Compared with YOLOv11 (AP: 0.88519, Precision: 0.830, F1: 0.84), the performance is slightly reduced, but its Precision and Recall remain at a high level, indicating that wavelet convolution and attention mechanisms still have strong adaptability in complex backgrounds.

[0157] Yolov8+ContextAggregation maintains high performance in the sagittal plane (AP: 0.882, F1: 0.839), but fails to surpass the improved models of the YOLOv11 series.

[0158] Compared with other models, the values ​​of other models on the sagittal plane have dropped significantly, especially the DETR model, whose mAP50 on the sagittal plane has dropped by 0.2-0.3 compared with other planes, indicating that the adaptability of the model is limited. However, the WE-YOLO model still maintains high mAP50, precision and recall values ​​on the sagittal plane, which shows that it has adapted well to complex scenes.

[0159] (3) Coronal plane:

[0160] The WE-YOLO model outperforms other models in AP, Precision, Recall, and F1 score on the coronal plane. Its AP reaches 0.92662, Precision is 0.860, and F1 score is 0.88, indicating that our model performs best on the coronal plane.

[0161] Compared with the YOLOv11 baseline model (AP: 0.86335, Precision: 0.830, F1: 0.82), the AP is increased by 7.3%, the Precision is increased by 3.6%, and the F1 score is increased by 6.8%, showing the superiority of multi-scale feature fusion and spatial attention mechanism.

[0162] Yolov8+ContextAggregation achieved an AP of 0.913, but failed to surpass WE-YOLO in terms of Precision and F1 scores. The quantitative analysis results of this paper validate the superiority of the WE-YOLO mandibular fracture detection model in fracture feature extraction and accurate detection, providing strong technical support for its application in medical imaging.

[0163] The AP of RT-DETR in the coronal plane is close to WE-YOLO (0.926), but the F1 score is slightly lower, which further verifies the strong adaptability of the improved model to small target detection.

[0164] The quantitative analysis results of the present invention verify the superiority of the WE-YOLO mandibular fracture detection model in fracture feature extraction and accurate detection, providing strong data support for the application of the model in medical imaging.

[0165] 4.4.2 Qualitative Analysis

[0166] This paper further validates the effectiveness of the mandibular object detection method based on wavelet convolution and multi-scale attention-enhanced YOLOv11 through qualitative analysis, highlighting its advantages in feature extraction, attention optimization, and multi-scale feature fusion. Visual analysis of the detection results and feature attention areas of specific samples clearly demonstrates how the model leverages the improved YOLOv11 architecture to improve the accuracy and robustness of mandibular fracture detection.

[0167] (1) Effectiveness of wavelet convolution in enhancing local detail extraction:

[0168] In the WE-YOLO model's mandibular fracture detection process, the C3k2_WT module captures multi-scale features through wavelet convolution, significantly enhancing the model's ability to focus on local details. Visual analysis of the detection results shows that the model effectively retains detailed information, laying the foundation for subsequent feature extraction, thereby significantly improving the accuracy of fracture line detection. This demonstrates that wavelet convolution can effectively improve feature expression capabilities in fracture detection tasks.

[0169] (2) Effectiveness of multi-scale attention focusing on key parts

[0170] In the WE-YOLO model's mandibular fracture detection process, the C2PSA_EMA module uses a spatial attention mechanism to focus on key areas, further enhancing the model's sensitivity to mandibular fracture lines. It receives detailed information from the C3k2_WT module, highlighting the significance of the fracture area while suppressing background noise, significantly improving the model's detection robustness in complex backgrounds.

[0171] (3) Complementarity of multi-scale feature fusion:

[0172] The WE-YOLO model combines feature maps at different scales, enhancing its comprehensive ability to detect fracture lines of varying sizes through a multi-scale feature fusion mechanism. When processing complex fracture structures, the model is able to capture subtle boundary information from high-resolution features while simultaneously obtaining global context from low-resolution features. Even when fracture lines are minute or boundaries are blurred, the model can still accurately locate the fracture area, significantly reducing the probability of missed detections and false positives. This complementary nature of multi-scale features effectively enhances the model's adaptability and detection accuracy.

[0173] In general, the automatic detection system for mandibular fractures of the present invention significantly improves the detection accuracy and robustness of the model through wavelet convolution optimization, multi-scale attention mechanism and multi-scale feature fusion, and has broad clinical application prospects.

[0174] Tables 1 to 3 show the experimental results.

[0175] Table 1

[0176]

[0177] Table 2

[0178]

[0179]

[0180] Table 3

[0181]

[0182] In summary, the qualitative analysis of the present invention further demonstrates the significant advantages of the present invention in feature extraction, multi-scale fusion and fracture line localization, proving the good performance and high accuracy of the model in the mandibular fracture detection task.

[0183] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing, characterized by: The invention comprises a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein the computer memory stores a trained mandibular fracture detection model; the mandibular fracture detection model comprises a backbone network, a feature fusion network, and a detection head part; The backbone network is used to extract feature maps P3, P4, and P5 of different scales. The backbone network first downsamples the input image using two convolutional layers, and then iteratively extracts the P3 feature map through two layers of C3k2_WT modules. Based on the P3 feature map, the P4 feature map is further extracted through the third layer of C3k2_WT modules. Based on the P4 feature map, the P5 feature map is finally generated with low resolution and high-level semantic representation capabilities through the fourth layer of C3k2_WT modules, the spatial pyramid fast pooling module SPPF, and the C2PSA_EMA modules. The feature fusion network adopts a bidirectional fusion strategy, first performing top-down semantic enhancement, upsampling the P5 feature map by bilinear interpolation to align its size with the P4 feature map, and then splicing it along the channel dimension to generate the fused feature map P4. new , and then P4 new After upsampling, it is concatenated with the P3 feature map to generate P3 new ; Then transfer the details from bottom to top, and make the spliced ​​P3 new A C3k2_WT module is introduced to perform feature compression and nonlinear enhancement, complete downsampling, and secondary splicing with the original P4 and P5 feature maps to form a closed-loop fusion path to generate P4 new ', P5 new '; The detection head part consists of three independent detection heads, with multi-scale anchor frames preset for feature maps of different resolutions; the fused P3 new 、P4 new ', P5 new 'The feature map is passed to three independent detection heads, and the small-scale anchor boxes densely cover P3 new High-resolution grid, medium-scale anchor box adapted to P4 new 'Medium-resolution grid, large-scale anchor box adapted to P5 new ' low-resolution grid; each detection head simultaneously outputs bounding box coordinates, category probability and confidence, and the outputs of the three detection heads are fused to ultimately retain high-precision predictions; When the computer processor executes the computer program, the following steps are implemented: The image to be detected is input into the trained mandibular fracture detection model to obtain the prediction result of mandibular fracture.

2. The mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing according to claim 1 is characterized in that: The backbone network first uses two convolutional layers to downsample the input image, and the formula is as follows: O1=Conv(I,64,3,2) O2=Conv(O1,128,3,2) Among them, O1 represents the output feature map after the first convolution layer, and O2 represents the output feature map after the second convolution layer; the input data I performs the first convolution operation, using 64 convolution kernels of size 3×3 and a stride of 2 to obtain the output feature map O1; then the first convolution operation is performed again, using 128 convolution kernels of size 3×3 and a stride of 2 to obtain the new output feature map O2.

3. The mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing according to claim 1 is characterized in that: The working process of the C3k2_WT module is as follows: First, perform feature decomposition on the feature map of the input C3k2_WT module to obtain feature maps X1 and X2; Perform lightweight convolution feature extraction on the feature map X1 to obtain X light ; The wavelet transform decomposes the feature map X2 into a low-frequency component F through four filters. LL and three high frequency components F LH 、F HL and F HH , and then use the inverse wavelet transform to fuse the low-frequency component and the three high-frequency components to reconstruct the feature map X2 with the same spatial resolution. IWT ; The obtained X light and X IWT Perform fusion to obtain feature fusion results for subsequent operations.

4. The mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing according to claim 3 is characterized in that: Perform lightweight convolution feature extraction on the feature map X1. The formula is as follows: X light =Conv light (X1)=Pointwise(Depthwise(X1,k=2)) In the formula, Depthwise means using depth convolution with a 2×2 small kernel, Pointwise means using 1×1 convolution for channel fusion, and k=2 means that the size of the convolution kernel is 2×2.

5. The mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing according to claim 3 is characterized in that: A low frequency component is F LL , reflecting the overall structural information of the input; the three high-frequency components are F LH 、F HL and F HH ; F LH Represents the high-frequency component in the horizontal direction, which is used to capture the edge details in the horizontal direction; F HL Represents the high-frequency component in the vertical direction, which is used to capture the edge details in the vertical direction; F HH Represents the high-frequency component in the diagonal direction and is used to extract the texture features in the diagonal direction.

6. The mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing according to claim 1 is characterized in that: The working process of the C2PSA_EMA module is as follows: After the feature map is input into the C2PSA_EMA module, it passes through a 1×1 convolutional layer to compress the number of channels and decompose it into retained feature A and feature B for further processing. Feature A directly participates in the final fusion as a global feature, while feature B enters the multi-scale attention path to extract spatial and contextual information. Apply average pooling along the height and width directions to feature B to generate B respectively. h and B w ; Then B h and B w The attention weight matrix is ​​generated by splicing and 1×1 convolution, and the attention of the salient area is further strengthened by the sigmoid activation function σ to obtain the final attention weight matrix W EMA Finally, the attention weight is applied to the feature map B through weighted operation to highlight the target area and obtain the attention weighted feature B EMA ; To B EMA 3×3 convolution and 1×1 convolution are applied in parallel to extract local spatial correlation and global channel context dependency respectively: 3×3 convolution captures texture and edge details in the local neighborhood by expanding the spatial receptive field, and 1×1 convolution models the semantic dependency between global channels through cross-channel linear combination; then the convolution result B of the two is concatenated. FNN Through the two-layer feedforward network, nonlinear transformation is performed to further enhance the feature expression ability, and the enhanced result is compared with the original B EMA Add, retain the original information through residual connection, and finally obtain the enhanced feature B′; The retained feature A and the enhanced feature B′ are concatenated in the channel dimension, and the number of output channels is restored through a 1×1 convolution to obtain the final feature map for subsequent operations.

7. The mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing according to claim 1 is characterized in that: In the detection head part, the outputs of the three detection heads are fused across scales through non-maximum suppression (NMS).

8. According to the mandibular fracture detection system based on wavelet transform feature enhancement and multi-scale attention focusing in claim 1, the dataset for training the mandibular fracture detection model is constructed as follows: Images were extracted from maxillofacial CT images in DICOM format and a dataset was constructed. The fracture area was annotated, including length, width, and center coordinates. The images were resized and formatted to ensure the consistency and adaptability of the input data. Finally, the dataset was expanded through data enhancement.

Images