A rib rapid diagnosis method and system

Abstract

The application provides a rib rapid diagnosis method and system, which comprises the following steps: receiving a rib CT scan image, embedding a convolution block attention module and a global attention mechanism on a YOLOv8n skeleton to enhance multi-scale channel-space feature expression, and realizing extraction of a rib candidate box in the scan image; tracking the input rib candidate box through an improved ByteTrack algorithm, and generating a three-dimensional rib point cloud trajectory, wherein the ByteTrack algorithm integrates a deformable convolution and a space-time interaction mechanism; after coordinate system correction and axial scaling of the three-dimensional rib point cloud trajectory, generating a rib MPR image through a least square method fitting section reconstruction algorithm, and performing fracture diagnosis according to the rib MPR image. Through the scheme, the accuracy of rib MPR reconstruction and the image processing efficiency can be considered, and the accuracy of fracture diagnosis is significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, and in particular relates to a rapid diagnostic method and system for ribs. Background Technology

[0002] Rib fractures are a common high-risk trauma in clinical practice, easily leading to fatal complications such as pneumothorax or organ damage. Early and accurate diagnosis is directly related to patient prognosis. Currently, clinical diagnosis relies on computed tomography (CT) scans, but this method requires physicians to analyze hundreds of cross-sectional images layer by layer, which is time-consuming and prone to missed diagnoses due to visual fatigue, especially in low-contrast areas such as the junction of the rib and cartilage.

[0003] Despite the increasing application of artificial intelligence (AI) technology in assisted diagnosis, existing solutions still face a dual bottleneck: On the one hand, while mainstream 3D segmentation models (such as deep learning-based rib reconstruction algorithms) can reconstruct skeletal structures, their computational complexity is too high, making it difficult to meet the real-time requirements of emergency scenarios. On the other hand, while 2D detection models offer significant speed improvements, they frequently produce false positives due to neglecting the spatial continuity of the ribs and are extremely sensitive to morphological variations in the paravertebral region, leading to severe identity jumps during target tracking. Furthermore, in image post-processing, while surface reconstruction techniques can unfold the rib structure, they inevitably introduce artifacts that interfere with diagnosis. Multiplanar reconstruction, while possessing superior sensitivity, suffers from the persistent bottleneck of manual process adjustment efficiency. Therefore, current rib image processing methods struggle to balance detection accuracy and processing efficiency. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a rapid rib diagnosis method and system to solve the problem that current rib image processing is difficult to balance detection accuracy and processing efficiency.

[0005] In a first aspect of the present invention, a rapid rib diagnosis method is provided, comprising: The system receives rib CT scan images and extracts rib candidate boxes from the scan images based on an improved YOLOv8 network with embedded convolutional block attention modules and a global attention mechanism. The improved ByteTrack algorithm tracks the input rib candidate boxes and generates a 3D rib point cloud trajectory. The ByteTrack algorithm integrates deformable convolution and spatiotemporal interaction mechanisms. After coordinate system correction and axial scaling of the three-dimensional rib point cloud trajectory, a cross-section reconstruction algorithm based on least squares fitting is used to generate rib MPR images, and fracture diagnosis is performed based on the rib MPR images.

[0006] In a second aspect of the present invention, a system for rapid rib diagnosis is provided, comprising: The target detection module is used to receive rib CT scan images and extract rib candidate boxes from the scan images based on an improved YOLOv8 network with embedded convolutional block attention modules and a global attention mechanism. The target tracking module is used to track the input rib candidate boxes using the improved ByteTrack algorithm and generate a three-dimensional rib point cloud trajectory. The ByteTrack algorithm integrates deformable convolution and spatiotemporal interaction mechanisms. The planar reconstruction module is used to perform coordinate system correction and axial scaling on the three-dimensional rib point cloud trajectory, generate rib MPR images based on the least squares fitting cross-section reconstruction algorithm, and perform fracture diagnosis based on the rib MPR images.

[0007] In a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor executes the computer program to implement the steps of the method as described in the first aspect of the present invention.

[0008] In a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method provided in the first aspect of the present invention.

[0009] In this embodiment of the invention, an improved YOLOv8 network with embedded convolutional block attention modules and a global attention mechanism is used for rib target detection, and an improved ByteTrack algorithm is employed to track the rib targets. A section reconstruction algorithm using least squares fitting is then used to generate rib MPR images. This not only effectively improves the quality of generated MPR images but also balances the accuracy of rib MPR reconstruction with image processing efficiency, thereby enhancing the efficiency and accuracy of fracture detection. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating a rapid rib diagnosis method according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the MPR reconstruction process provided in one embodiment of the present invention; Figure 3This is a schematic diagram of a system for rapid rib diagnosis provided in one embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0013] It should be understood that the terms "comprising" and other similar expressions in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, or apparatus that includes a series of steps or units and is not limited to the listed steps or units. Furthermore, "first" and "second" are used to distinguish different objects and are not intended to describe a specific order.

[0014] Please see Figure 1 A flowchart illustrating a rapid rib diagnosis method provided in this embodiment of the invention includes: S101. Receive rib CT scan images and extract rib candidate boxes from the scan images based on an improved YOLOv8 network with embedded convolutional block attention modules and a global attention mechanism. CT scans are medical images that use X-rays and computer technology to generate three-dimensional images of the internal structures of the human body. During CT imaging, a selected slice is divided into many small cubes of equal volume (i.e., voxels). The X-ray attenuation coefficient or absorption coefficient of each voxel is calculated by computer and arranged into a digital matrix. This matrix is ​​then converted into pixels of different gray levels using a digital-to-analog converter, and finally arranged according to the matrix to form the CT image. High-speed CT scans can acquire multiple CT images in a short time (e.g., 40ms), and MPR reconstruction can be achieved based on these multiple CT images.

[0015] The improved YOLOv8 network is an object detection network model that includes a backbone network, a feature enhancement network, and a detection head. The Convolutional Block Attention Module (CBAM) can enhance features by using attention maps in both channel and spatial dimensions, while the Global Attention Mechanism (GMA) dynamically adjusts the model's attention weights on different parts of the input data, thereby improving its sensitivity to key information and feature extraction capabilities.

[0016] An improved YOLOv8 network, which embeds a convolutional block attention module and a global attention mechanism, can track rib targets in CT scan images and output target detection boxes, i.e. rib candidate boxes.

[0017] In this embodiment, YOLOv8 is used as the basic detection framework, and convolutional attention and global attention mechanisms are integrated to deeply optimize it, so as to enhance its robustness and accuracy in rib fracture detection tasks.

[0018] To address the scale adaptability issues of traditional detectors when handling ribs of varying sizes and minute fractures, a Global Attention Mechanism (GAM) and a Convolutional Block Attention Module (CBAM) were introduced into the YOLOv8 backbone and feature fusion network. CBAM and GAM were prioritized because they offer significant advantages in local detail enhancement (CBAM) and global context awareness (GAM), respectively. Rib fracture detection requires attention to both the local details at the fracture site and an understanding of the rib's overall position and morphology within the thoracic cavity; combining these two approaches provides more comprehensive information. While CA and ECA are also effective attention mechanisms, the combination of CBAM and GAM demonstrated superior synergistic effects in experimental validation, better adapting to the complexity of rib fracture detection.

[0019] S102. The improved ByteTrack algorithm is used to track the input rib candidate boxes and generate a three-dimensional rib point cloud trajectory. The ByteTrack algorithm integrates deformable convolution and spatiotemporal interaction mechanisms. The ByteTrack algorithm is a detection-based multi-target tracking algorithm. It makes full use of all detection boxes through a two-stage data association strategy (high-confidence detection boxes and low-confidence detection boxes), which can improve the tracking stability and accuracy in occluded scenarios.

[0020] The improved ByteTrack algorithm is used to track the detected target and output a stable 3D point cloud trajectory.

[0021] It is understandable that CT images are three-dimensional data composed of a series of two-dimensional slices. Relying solely on the detection results of a single frame cannot fully capture the continuity and overall morphology of the ribs in three-dimensional space, nor can it solve the problems of accurate localization and three-dimensional reconstruction of rib fractures. Therefore, it is necessary to introduce a cross-frame tracking algorithm (ByteTrack) to establish the continuous trajectory of the ribs in three-dimensional space. At the same time, based on the ByteTrack framework, the ByteTrack algorithm can be adaptively improved to capture the deformation and motion patterns of the ribs in three-dimensional space more precisely, and to integrate the topological relationships between the ribs, thereby outputting a stable and accurate three-dimensional point cloud trajectory.

[0022] This embodiment improves the ByteTrack algorithm by introducing deformable convolution to enhance feature extraction and by introducing a spatiotemporal interaction mechanism to enhance correlation robustness.

[0023] S103. After coordinate system correction and axial scaling of the three-dimensional rib point cloud trajectory, a rib MPR image is generated based on the least squares fitting section reconstruction algorithm, and fracture diagnosis is performed based on the rib MPR image.

[0024] The section reconstruction algorithm is a technique used in medical image processing to generate arbitrary path sections from 3D CT images. Its core objective is to achieve accurate visualization of complex anatomical structures through interactive path planning and mathematical calculations.

[0025] The temporal point cloud obtained by stable tracking is transformed into a multiplane with topological consistency, and the rib MPR image is generated by fitting and reconstructing based on the least squares method.

[0026] For example, the image reconstruction process is as follows: Figure 2 As shown in the figure, B8 and B9 are backbone layers, H12 and H15 are head layers, C2f represents the feature extraction module, Conv represents the convolution module, Concat represents the feature tensor concatenation module, SPPF represents fast spatial pyramid pooling, Upsample represents upsampling, Detect represents the output layer, CBAM represents the convolutional block attention module, and GAM represents the global attention module. L01-L12 represent the detection results of the left rib, and R01-R12 represent the detection results of the right rib.

[0027] In the figure, after target detection is performed on the CT image using the improved YOLOv8 network, the target trajectory can be tracked through cascade matching and bounding box matching, and the tangent plane can be reconstructed based on the least squares method.

[0028] In one embodiment, based on gradient-weighted class activation mapping technology, the attention distribution of the YOLOv8 network model at different levels is visualized and analyzed, and the levels where convolutional block attention modules and global attention mechanisms are inserted are selected respectively. The convolutional block attention module is integrated into the output features of the model backbone layer, and the global attention module is integrated into the output features of the model head layer. The extracted features of the convolutional block attention module and the global attention mechanism are fused based on the hierarchical attention fusion strategy. Construct an improved loss function for the YOLOv8 network, which includes a bounding box regression loss based on Distance IoU and a classification loss based on Focal Loss. Set a cross-frame semantic consistency strategy for detection boxes to fuse detection boxes in adjacent frames.

[0029] Gradient-weighted class activation mapping (Grad-CAM) is used to determine the optimal insertion position of the attention module, enabling visualization analysis of the attention distribution of the YOLOv8n model at different layers. This is achieved by training on a training set containing a large number of rib fracture samples and visualizing different layers, for example... Figure 2 As shown, Grad-CAM heatmaps of Backbone8 (B8) and Backbone9 (B9) layers in the backbone network, and Head12 (H12) and Head (H15) layers in the detection head are used to identify the regions that the model is most interested in when processing rib features.

[0030] The Grad-CAM heatmap shows that layer B8 exhibits high activation in the rib distribution area, with some activation also observed in areas such as the lungs, liver, and blood vessels, but the main focus is on the skeletal structure. This indicates that layer B8 has a certain ability to recognize ribs and can capture the overall distribution information of the ribs. Layer B9 shows a similar rib attention pattern to B8, with activation more concentrated on the thoracic skeletal structure and reduced activation in surrounding non-rib areas. This suggests that layer B9 is better at learning more refined rib morphological features and can better suppress background noise. The heatmap of layer H12 shows that its attention is mainly focused on the spinal region, while the activation in the rib region is low and scattered, exhibiting fragmented characteristics. This suggests that layer H12 may focus more on the structure of the spine or related features and is not suitable for precise rib localization. The heatmap of layer H15 shows extremely high specificity, with its attention almost entirely focused on the rib structure, and almost no interference from non-rib areas such as the lungs and blood vessels was observed. This indicates that layer H15 can utilize features from different layers for refined localization and can effectively distinguish ribs from other structures.

[0031] Preferably, attention modules are inserted into layers B8 and B9 of the backbone network and layer H15 of the detector head. These layers represent different stages in the YOLOv8 network, from high-level semantic features to fine local features. Placing the attention mechanism at these key nodes can achieve semantic preservation and feature enhancement across layers.

[0032] Based on a hierarchical attention fusion strategy, a combination of H15GAM + B9CBAM is designed for the YOLOv8v8n model, combining CBAM and GAM.

[0033] Layer B9 was chosen instead of B8 because it demonstrated better rib focus and lower background interference in the experiments. Placing CBAM (focusing on local details) in layer B9 provides more refined local information for subsequent high-level feature fusion. GAM (focusing on global interaction) was placed in layer H15 to leverage its long-range dependency capture capability and globally understand the relationship between the fracture area and the surrounding ribs.

[0034] In the YOLOv8 backbone network, layer B9 is typically a C2f module. This embodiment integrates the CBAM module into the output features of layer B9 as an enhancement process for those features. Specifically, if the output feature map of layer B9 is... The feature map after CBAM processing The calculation is as follows:

[0035] in, It is a channel attention mapping. It is a spatial attention mapping. ⊕ represents element-wise multiplication, and ⊕ represents element-wise addition.

[0036] CBAM first generates channel attention, then performs element-wise multiplication between the channel attention and the original feature map to obtain a channel-weighted feature map. Next, the channel-weighted feature map is input into the spatial attention module, and element-wise multiplication is performed again to obtain the final spatially and channel-weighted feature map. .

[0037] The YOLOv8 detection head typically employs the PA-FPN (Path Aggregation Feature Pyramid Network) concept, with some simplifications. Layer H15 receives features from multi-scale fusion. The GAM module is integrated after the output features of layer H15. Let the output feature map of layer H15 be... GAM module output The calculation is as follows:

[0038] in, This is the global attention map generated by GAM. GAM generates attention through two parallel branches: channel attention branch and channel attention branch.

[0039] Where GAP is global average pooling, GMP is global max pooling, Conv is a convolutional layer, and σ is the activation function. Spatial attention branch:

[0040] AvgPool is average pooling, and MaxPool is max pooling.

[0041] GAM combines these two attention maps and introduces a rotation transformation to enhance their interactivity, ultimately generating a global attention map that interacts with the input feature map. Multiply to obtain the enhanced features .

[0042] It is understandable that a hierarchical attention fusion strategy, combined with convolutional block attention modules and global attention mechanisms, can enhance the recognition of local details. Applying CBAM to layer B9 allows the model to capture the texture and subtle discontinuities at the rib fracture site more precisely, highlighting these key features even in low-contrast areas and reducing misidentification caused by image noise or artifacts. It can also improve the understanding of global morphology. Applying GAM to layer H15 helps the model understand the positional relationship of the currently detected rib fragment in the entire thoracic cavity and its overall continuity. This is crucial for distinguishing real fractures from anatomical variations or handling the loss of local information due to CT slice intervals. Simultaneously, it can promote the stability of cross-frame features. When the attention mechanism can more accurately identify the core features of the rib (whether it's local fracture details or global morphology), the model's detection results for the same rib on different slices become more consistent. For example, even if the fracture is not obvious on a certain slice, the model can still associate it with known fracture areas in adjacent slices through GAM's perception of the global structure, thereby maintaining the stability of the detection box and reducing drift. This consistency of cross-frame semantics makes the features output by the model more stable, providing a solid foundation for subsequent temporal tracking algorithms.

[0043] The loss function of YOLOv8 is optimized by expressing it as the bounding box loss of DIoU and the classification loss of FocalLoss.

[0044] To further improve the model's performance in rib fracture detection, especially in handling small targets and class imbalance, this embodiment optimizes the loss function of YOLOv8. YOLOv8 originally used CIoU (CompleteIntersection over Union) loss as the bounding box regression loss and combined it with Binary Cross-Entropy (BCE) loss as the classification loss.

[0045] CIoU loss can take into account the overlapping areas of bounding boxes, the distance between center points, and the aspect ratio, which improves the positioning accuracy to a certain extent. However, it may still result in inaccurate positioning for slender ribs or minor fractures. This embodiment considers introducing DIoU (Distance IoU) loss on the basis of CIoU loss to more effectively handle the position and shape information of the bounding boxes.

[0046] The DIoU loss takes into account the distance between the center points of the bounding box, which can make the bounding box converge faster. Its formula is:

[0047] in, and These are the center points of the predicted bounding box and the ground truth bounding box, respectively. It is Euclidean distance. It is the diagonal length of the smallest closed box that can contain two bounding boxes.

[0048] Using DIoU as a base and combining it with the aspect ratio penalty term in CIoU, a more robust loss function can be obtained:

[0049] in, and It is a balancing term used to control the impact of differences in width and height.

[0050] In rib fracture detection, there are two categories: "rib-normal" and "rib-fractured." The number of normal ribs is far greater than that of fractured ribs, resulting in a severe class imbalance. To mitigate this issue, this embodiment uses Focal Loss as the classification loss function. Focal Loss introduces a modulation factor to reduce the weight of easily distinguishable samples, allowing the model to focus more on learning difficult-to-distinguish samples (i.e., fracture samples). The formula for Focal Loss is:

[0051] in, It is the probability that the model predicts it to be a positive class. This is the focusing parameter, used to smoothly reduce the weight of easily classified samples. When At that time, Focal Loss degenerates into standard cross-entropy loss. This is a class weighting factor used to address class imbalance. For the positive class, For negative classes, . It is a hyperparameter, set to 0.25.

[0052] In this embodiment, by combining the bounding box loss of DIoU and the classification loss of Focal Loss, the YOLOv8 network model is improved, which can more accurately locate ribs and more effectively identify a small number of fracture samples, thereby outputting stable detection features that are more beneficial to subsequent temporal tracking.

[0053] Optionally, the cross-frame semantic consistency strategy for detection boxes includes: For CT scan images of adjacent frames, if the overlap between the detection box of the previous frame and the detection box of the current frame in the same rib region exceeds a preset value, the confidence scores of the two detection boxes are weighted averaged, or the box with higher confidence score is taken as the final detection box of the current rib region. IoU matching is used to perform inter-frame association of similar rib candidate boxes, and Kalman filtering is used to smooth the center point and size of the candidate boxes in the current frame.

[0054] For adjacent CT slices, if there is some overlap between the detection bounding box B_t in frame t and the detection bounding box B_(t+1) in frame t+1 for the same rib region (e.g., IoU > 0.6), then the confidence scores of the two detection bounding boxes are weighted and averaged, or the box with the higher confidence score is used as the final detection bounding box for that region. ;

[0055] in, This is the ground truth bounding box for the current frame (during the training phase). and This represents a threshold. It is not used during the inference phase. Instead, it compares and merges the detection boxes of consecutive frames.

[0056] By fusing the detection frames, the accuracy of target detection can be effectively improved, while ensuring the stability of the detection results.

[0057] By performing a simple inter-frame correlation on the detection results (candidate boxes), similar detection boxes are linked together, and the center point and size of the detection boxes are smoothed. Through a simple IoU matching and a variation of Kalman filtering, the trajectory of the detection boxes can be initially estimated and smoothed, thereby reducing noise in single-frame detection.

[0058] Specifically, let the detection box of the previous frame be... The detection box of the current frame is .if If they are the same target, then the two are considered to likely belong to the same target. In this case, a new detection box can be added. Parameters (such as center point coordinates) Update: ; Where λ∈[0,1] is the smoothing factor.

[0059] It is understandable that improving the YOLOv8 detection network not only enhances the detection accuracy and robustness of ribs and their fractures in single-frame images, especially in handling subtle fractures and scale variations, but more importantly, by introducing hierarchical attention fusion and cross-frame semantic consistency strategies (detection box fusion, association, and smoothing, etc.), it can output more stable and reliable detection features, laying a solid foundation for subsequent complex 3D temporal tracking tasks. These stable detection features, including accurate bounding boxes, high confidence scores, and attention-enhanced feature vectors, provide high-quality input for the subsequent cross-frame tracking in step S102, greatly reducing the tracking failure rate caused by unstable single-frame detection.

[0060] In one embodiment, step S102 includes: Deformable convolutions are integrated into the target appearance feature extraction network of the ByteTrack algorithm, replacing some or all of the standard convolutional layers; In the ByteTrack algorithm, a spatiotemporal graph is constructed to represent the rib detection results on the slice and the temporal relationship between adjacent slices. The neighborhood node information of the spatiotemporal graph is extracted based on the graph neural network to realize spatiotemporal interaction.

[0061] The ByteTrack algorithm integrates deformable convolution and a spatiotemporal interaction mechanism: deformable convolution is integrated into the target appearance feature extraction network in ByteTrack; in ByteTrack, a spatiotemporal graph is constructed to represent the rib detection results on the slice and the temporal relationship between adjacent slices respectively, and the neighborhood node information of the spatiotemporal graph is aggregated based on graph neural network to realize the spatiotemporal interaction mechanism.

[0062] To capture the deformation of ribs across different slices more precisely, the ByteTrack algorithm integrates deformable convolutions into the appearance feature extraction module for target tracking. Deformable convolutions allow the kernel to sample around the original sampling location with learnable offsets, thus enabling adaptive matching of the object's shape and scale.

[0063] The principle of deformable convolution is that standard convolution samples within a fixed-size receptive field, while deformable convolution samples at each kernel position. , will learn a set of relative offset And perform convolution operations on these offset sampling points. For a position Input feature map at the location Output feature map The calculation method is as follows:

[0064] in, It is the sampling grid of standard convolution. It is the learned offset. These are learnable convolutional weights. This is the number of sampling points. Offset. and convolution weights It is typically predicted by two parallel convolutional layers.

[0065] In ByteTrack, a separate feature extraction network (such as ResNet or Transformer) is typically used to generate appearance feature descriptors for each detected target. This embodiment replaces some or all of the standard convolutional layers with deformable convolutional layers to enhance the feature extractor's ability to perceive rib deformation and subtle structural changes.

[0066] A deformable convolutional layer is inserted in the middle layer of the feature extraction network. Let the input feature map be... Intermediate features are obtained through a standard convolutional layer. Then, for The parameters for deformable convolutions are generated using two parallel convolutional layers (one for predicting offsets and one for predicting weights):

[0067] in, Predict the offset (2D) for each sampling point. Predict the weight for each sampling point. Then, based on the learned offset... and weight For input features Sampling and weighted summation are performed to obtain the output features. :

[0068] In this way, the feature extractor can more flexibly capture the local deformation of the ribs in three-dimensional space. For example, the ribs deform when the chest cavity expands or contracts. Deformable convolution can better adapt to this change and generate more distinctive appearance features.

[0069] This embodiment introduces a spatiotemporal interaction mechanism, which can more effectively utilize the spatiotemporal information of ribs in three-dimensional CT sequences and overcome the problems of topological mismatch and identity drift between slices.

[0070] The goal of this mechanism is to integrate 2D detection results into 3D space and establish a continuous 3D trajectory. First, the 2D detection boxes and their corresponding feature vectors are elevated to 3D space. Assume there are... CT slices, each slice have One detected rib instance Each instance Includes bounding box information Confidence level and the extracted appearance feature vector .

[0071] The spacing of CT slices is usually known and can be used as dimension information along the Z-axis. Assuming the slices... Corresponding to three-dimensional space Coordinates. Therefore, a two-dimensional detection box... It can be initially regarded as a three-dimensional "columnar" region, and its range on the Z-axis can be estimated based on the spacing between adjacent slices.

[0072] To achieve spatiotemporal interaction, a spatiotemporal graph is constructed. Each node in the graph represents the detection result of a rib instance on a certain slice. The connections between nodes are based on the temporal relationship between adjacent slices, connecting the detection results of the same rib on adjacent slices. For example, if slices detection box With slices detection box A high IoU indicates that the rib is represented on different slices, and a connection can be established.

[0073] Based on this, a graph neural network (GNN) is used to process the spatiotemporal graph. By aggregating the information of neighboring nodes through the GNN, a globally consistent node representation is learned.

[0074] Specifically, for nodes In its first Hidden representation after round message passing You can update it in the following ways:

[0075] in, It is a node The set of neighboring nodes (including nodes inside and outside the same slice). From node To the node Message passing function, It is an aggregate function.

[0076] Through multiple rounds of message passing, each node Hidden representation It will integrate information from its spatiotemporal neighborhood, thereby enabling the capture of the continuity of the ribs, their movement trends, and their relative positional relationships with other ribs.

[0077] In one embodiment, the steps preceding step S102 include: Set topological constraints for ribs. If the rib candidate box is located within the spinal region and the rib IDs and spatial relationships of the ribs that have been tracked around the candidate box do not match the labeled rib sequence, then reduce the matching priority of the candidate box or mark the candidate box as pending. A short-term forgetting mechanism is set up so that when a rib target is lost, the rib target is marked as to be recovered. When a rib object with similar appearance features or the same trajectory is detected again in a subsequent frame, the rib object in the subsequent frame is matched with the rib target marked as to be recovered.

[0078] A topological constraint is constructed for the ribs, with predefined rib numbers (L01-L12, R01-R12). When the tracker encounters ambiguity in the near-spine region (e.g., L01, L02, R01, R02), it can assist in the judgment by querying the tracking status of its neighboring ribs (e.g., L03, R03). If a detection box is in the near-spine region, and the IDs and spatial relationships of the tracked ribs around it do not match the labeled conventional rib sequence, its matching priority can be reduced, or it can be marked as pending.

[0079] When a rib instance cannot be detected for several consecutive frames (e.g., due to depth occlusion), a traditional tracker might terminate its trajectory and assign a new ID once the occlusion is resolved. This embodiment introduces a "brief forgetting" mechanism: when a target is lost, it is not immediately deleted, but its status is marked as "pending recovery." In subsequent frames, if an object with similar appearance features and a possible trajectory is detected, an attempt is made to match it with the "pending recovery" target.

[0080] By using a detection box in the near-spine region and employing a brief forgetting mechanism, the accuracy and reliability of target tracking can be improved, avoiding the influence of factors such as occlusion, thereby improving the accuracy of the tracking trajectory.

[0081] The ByteTrack algorithm incorporates near-spine mismatch detection and a brief forgetting mechanism. Combined with the role of deformable convolution in occlusion recovery, it can capture subtle deformations in the visible area, even when the occluded region and ribs are visible, generating more distinctive features. This makes it easier to match based on appearance features even after the occlusion is removed, provided the target has shifted or deformed within the visible area.

[0082] By combining the spatiotemporal graph GNN processing in the ByteTrack algorithm, even if a node (rib instance) is lost on a slice, the representations of its neighboring nodes (instances of other ribs) still contain information about that rib. When the rib reappears, the rich contextual information of its surrounding nodes helps to reassociate it with its previous ID, thus avoiding ID redistribution.

[0083] Based on the deformable temporal and spatiotemporal interaction mechanism in the ByteTrack algorithm, it can effectively solve the problems of occlusion and mismatch encountered by traditional tracking algorithms in rib CT sequences, and output stable, reliable and topologically consistent 3D point cloud trajectories.

[0084] In one embodiment, step S103 includes: The three-dimensional rib point cloud trajectory is used as the parameterized curve of the rib, the direction of the rib's long axis is defined as the tangent vector of the curve, the cross section perpendicular to the long axis of the rib is used as the MPR plane, and the tangent vector is set as the normal vector of the MPR plane. Define a series of MPR planes parallel to the MPR plane normal vector, slice along the long axis of the rib, and sample each MPR plane to generate an MPR image; The process involves constructing a coordinate transformation from the original CT 3D voxels to the MPR plane. Based on this coordinate transformation, each pixel in the MPR image is mapped to a point in 3D space. By interpolating the voxel values ​​sampled in the original CT scan image, the grayscale values ​​of each pixel in the MPR image are obtained.

[0085] Before performing MPR reconstruction, the major axis direction of each rib needs to be determined using the least squares method to serve as the reference for the MPR plane. The 3D rib point cloud trajectory is then used. The smoothed 3D point cloud trajectory, considered as a ribcage. This can be viewed as a parametric curve of the ribs, where the major axis of the ribs can be defined as the tangent vector of the curve. In the parameter... At that point, the tangent vector is:

[0086] The tangent vectors along the entire rib trajectory are averaged to obtain the direction of the rib's major axis. This determines the direction of the rib's major axis (centerline). This allows us to define the MPR plane. A typical MPR plane can be a cross-section perpendicular to the long axis of the rib. If... If the vector is the unit tangent vector to the major axis of the rib, then the normal vector of the MPR plane is... It can be set to Any point on the MPR plane satisfy:

[0087] in, It is a reference point on the plane, and you can choose the center point or the starting point of the rib trajectory.

[0088] For a specific rib, along its long axis (parameter) Take a series of points, and define a series of parallel lines centered on these points. The MPR planes are defined. These planes share a common normal vector, but their positions in space differ, thus "slicing" along the long axis of the ribs to form the MPR image. With the MPR planes defined, CT data can then be sampled onto these planes to generate images.

[0089] When the error is concentrated in a specific direction, the least squares method can be used to solve the regression coefficients to calibrate the error.

[0090] The equation of the MPR plane is expressed as: Then, using the known 3D point cloud data of the ribs, the predicted values ​​are minimized. Value and reality The regression coefficients are calculated using the squared error between the values. :

[0091] In the formula, It includes and The design matrix of coordinates, It is corresponding Coordinate vector.

[0092] The generation process of the MPR image includes coordinate transformation from three-dimensional voxels to the MPR plane and interpolation of the MPR image.

[0093] Assume the coordinates of the voxels in the original CT data are The normal vector of the MPR plane is A reference point on the plane is Calculate the voxel point. Distance to the MPR plane:

[0094] The voxel points are then projected onto the MPR plane. Assume the MPR plane defines a local coordinate system. , The axis is parallel to the direction of the long axis of the rib (or its perpendicular direction). Axis perpendicular to If the axis is in the MPR plane, then Therefore, Coordinates to coordinate.

[0095] For each pixel on the MPR image Map it back to a point in three-dimensional space Then, the grayscale value of each pixel is obtained by interpolating (e.g., bilinear interpolation) the original CT voxel values: ;

[0096] In the formula, This represents the original CT voxel value.

[0097] In one embodiment, a bidirectional error feedback mechanism is constructed to compare the rib contour in the MPR image with the projection of the rib candidate box on the MPR image plane, calculate the model detection error, and feed the detection error as a loss term back to the YOLOv8 network, and update the YOLOv8 network parameters through backpropagation. By analyzing the local continuity and morphology of MPR images, the model tracking error is evaluated, and the tracking error is fed back to the ByteTrack algorithm to fine-tune the deformable convolution and graph neural network in the ByteTrack algorithm.

[0098] It is understandable that the accuracy of the 3D rib point cloud trajectory directly affects the quality of the MPR (Mean Correction Profile), and the quality of the MPR, in turn, can be used to evaluate the accuracy of the target detection network. The MPR view should theoretically represent the true shape of the ribs. If the rib morphology in a certain region of the MPR image is abnormal (e.g., discontinuous, deformed, or even inconsistent with the original CT voxel values), this may indicate that the detection box is inaccurate in that region, or that the tracking trajectory is deviated in that region.

[0099] Specifically, the rib contours on the MPR image are compared with the projections of the detected 2D bounding boxes onto the MPR plane. If the actual shape of the ribs in the MPR view does not perfectly match the boundaries of the detection boxes, it indicates that there is an error in the detection boxes.

[0100] Assuming that on the MPR plane, the true contour of the rib is composed of a series of points This indicates that the projection of the detected rib bounding box onto the MPR plane is... .

[0101] Optionally, the error can be quantified by calculating the Hausdorff distance between the MPR image and the rib detection box:

[0102] In the formula, It is a set of real contour points. It is the set of projection points of the detection box. It is the Euclidean distance between points.

[0103] The calculated error is used as a loss term and backpropagated to the YOLOv8 detection network, achieved by modifying the training process. During training, after generating the MPR and evaluating the error, this error can be added to the original detection loss function, DIoU + Focal Loss:

[0104] In the formula, This represents the loss calculated based on the MPR error. This is the weighting factor.

[0105] For example, The calculated Hausdorff distance is used to learn how to adjust the network's parameters through gradient descent to minimize the MPR error, thereby outputting more accurate detection boxes.

[0106] The quality of MPR (Multiplanar Reconstruction) is not only related to detection, but also closely related to the trajectory of target tracking. For example, if MPR shows unnatural distortion at the rib fracture site, it may mean that the long axis direction is inaccurately estimated when traversing the fracture area, or that the trajectory is overly smoothed.

[0107] The MPR process essentially projects a 3D point cloud onto a 2D plane. If the MPR plane itself is inaccurate (e.g., defined by an incorrect rib trajectory), the rib morphology in the MPR image will naturally be distorted. The quality of the tracking trajectory is evaluated by analyzing the local continuity and morphology of the MPR image. For example, in the MPR view, the ribs should exhibit relatively smooth curves, and fracture sites should have clear indications, rather than sudden bends or breaks. If the rib morphology in the MPR does not match the expected morphology based on the MPR plane defined by the trajectory, the tracking trajectory of the ByteTrack algorithm is considered problematic, and error feedback is needed: Trajectory shape evaluation calculates the curvature of the rib contour in the MPR view. If the curvature is abnormally high in some areas (indicating sharp bends) or there are abrupt changes in smooth areas, it indicates that the tracking trajectory needs adjustment. Appearance feature feedback examines the visualization of the rib fracture area in the MPR image. If the fracture features are unclear in the MPR, it may mean that the appearance features were not fully utilized during target tracking training or that the matching of appearance features was not accurate enough. Topological constraint feedback: If the MPR shows that the topological relationships of the ribs (e.g., relative positions with adjacent ribs) do not match expectations, the spatiotemporal interaction part of the GNN in target tracking can be further adjusted to strengthen the topological constraints.

[0108] These MPR-based evaluation results can be transformed into loss functions for fine-tuning deformable convolutional models, GNN models, or tracking correlation algorithms in the ByteTrack algorithm.

[0109] By constructing a bidirectional error feedback loop, a collaborative closed loop of target detection, tracking, and feedback is achieved. A trained YOLOv8 model is used for rib detection to obtain preliminary detection boxes and features. An improved ByteTrack algorithm, combined with deformable convolution and spatiotemporal interaction, is used to generate preliminary 3D rib point cloud trajectories. Using a least-squares geometric fitting method, MPR images are generated based on the point cloud trajectories. The quality of the MPR images is then evaluated and compared with the original CT data to calculate the detection and tracking errors. The MPR errors are fed back to the detection network and tracking algorithm, fine-tuning their model parameters until the errors converge or a preset number of iterations is reached. This continuously improves the accuracy of target detection, tracking, and MPR images.

[0110] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0111] Figure 3 This invention provides a schematic diagram of a system for rapid rib diagnosis, comprising: The target detection module 310 is used to receive rib CT scan images and extract rib candidate boxes in the scan images based on an improved YOLOv8 network with embedded convolutional block attention modules and global attention mechanisms. The target tracking module 320 is used to track the input rib candidate box using the improved ByteTrack algorithm and generate a three-dimensional rib point cloud trajectory. The ByteTrack algorithm integrates deformable convolution and spatiotemporal interaction mechanisms. The planar reconstruction module 330 is used to generate rib MPR images based on the least squares fitting section reconstruction algorithm after coordinate system correction and axial scaling of the three-dimensional rib point cloud trajectory, and to diagnose fractures based on the rib MPR images.

[0112] In one embodiment, the improved YOLOv8 network based on embedded convolutional block attention modules and a global attention mechanism includes the following steps before extracting rib candidate boxes from the scanned image: Based on gradient-weighted class activation mapping technology, the attention distribution of the YOLOv8 network model at different levels is visualized and analyzed, and the levels where convolutional block attention modules and global attention mechanisms are inserted are selected respectively. The convolutional block attention module is integrated into the output features of the model backbone layer, and the global attention module is integrated into the output features of the model head layer. The extracted features of the convolutional block attention module and the global attention mechanism are fused based on the hierarchical attention fusion strategy. Construct an improved loss function for the YOLOv8 network, which includes a bounding box regression loss based on Distance IoU and a classification loss based on Focal Loss. Set a cross-frame semantic consistency strategy for detection boxes to fuse detection boxes in adjacent frames.

[0113] Optionally, the cross-frame semantic consistency strategy for setting detection boxes, which fuses the detection boxes of adjacent frames, includes: For CT scan images of adjacent frames, if the overlap between the detection box of the previous frame and the detection box of the current frame in the same rib region exceeds a preset value, the confidence scores of the two detection boxes are weighted averaged, or the box with higher confidence score is taken as the final detection box of the current rib region. IoU matching is used to perform inter-frame association of similar rib candidate boxes, and Kalman filtering is used to smooth the center point and size of the candidate boxes in the current frame.

[0114] In one embodiment, deformable convolutions are integrated into the target appearance feature extraction network of the ByteTrack algorithm, replacing some or all of the standard convolutional layers. In the ByteTrack algorithm, a spatiotemporal graph is constructed to represent the rib detection results on the slice and the temporal relationship between adjacent slices. The neighborhood node information of the spatiotemporal graph is extracted based on the graph neural network to realize spatiotemporal interaction.

[0115] In one embodiment, topological constraints are set for ribs. If the rib candidate box is located within the spinal region and the tracked rib IDs and spatial relationships of the ribs around the candidate box do not match the calibrated rib sequence, the matching priority of the candidate box is reduced or the candidate box is marked as pending. A short-term forgetting mechanism is set up so that when a rib target is lost, the rib target is marked as to be recovered. When a rib object with similar appearance features or the same trajectory is detected again in a subsequent frame, the rib object in the subsequent frame is matched with the rib target marked as to be recovered.

[0116] In one embodiment, the planar reconstruction module 330 includes: Planar construction unit is used to treat the 3D rib point cloud trajectory as the parametric curve of the rib, define the rib major axis direction as the tangent vector of the curve, treat the cross section perpendicular to the rib major axis as the MPR plane, and set the tangent vector as the normal vector of the MPR plane. The image generation unit is used to define a series of MPR planes parallel to the MPR plane normal vector, slice along the long axis of the rib, and sample each MPR plane to generate an MPR image; The process involves constructing a coordinate transformation from the original CT 3D voxels to the MPR plane. Based on this coordinate transformation, each pixel in the MPR image is mapped to a point in 3D space. By interpolating the voxel values ​​sampled in the original CT scan image, the grayscale values ​​of each pixel in the MPR image are obtained.

[0117] In one embodiment, the system for rapid rib diagnosis further includes: The error feedback module is used to construct a two-way error feedback mechanism. It compares the rib contour in the MPR image with the projection of the rib candidate box onto the MPR image plane, calculates the model detection error, and feeds the detection error as a loss term back to the YOLOv8 network. The YOLOv8 network parameters are updated through backpropagation. By analyzing the local continuity and morphology of the MPR image, the model tracking error is evaluated, and the tracking error is fed back to the ByteTrack algorithm to fine-tune the deformable convolution and graph neural network in the ByteTrack algorithm.

[0118] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and modules described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0119] Figure 4 This is a schematic diagram of an electronic device according to an embodiment of the present invention. The electronic device is used for multiplanar reconstruction of ribs. Figure 4 As shown, the electronic device 40 of this embodiment includes: a memory 410, a processor 420, and a system bus 430. The memory 410 includes an executable program 4101 stored thereon. As those skilled in the art will understand, Figure 4 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0120] The following is combined with Figure 4 A detailed description of each component of the electronic device 40 is provided below: The memory 410 can be used to store software programs and modules. The processor 420 executes various functional applications and data processing of the electronic device by running the software programs and modules stored in the memory 410. The memory 410 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device (such as cached data), etc. In addition, the memory 410 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0121] The memory 410 contains an executable program 4101 for image reconstruction methods. This executable program 4101 can be divided into one or more modules / units, which are stored in the memory 410 and executed by the processor 420 to achieve rapid rib diagnosis, etc. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, describing the execution process of the executable program 4101 in the electronic device 40. For example, the executable program 4101 can be divided into functional modules such as a target detection module, a target tracking module, and a planar reconstruction module.

[0122] The processor 420 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines. By running or executing software programs and / or modules stored in the memory 410, and by calling data stored in the memory 410, it performs various functions and processes data, thereby monitoring the overall status of the electronic device. Optionally, the processor 420 may include one or more processing units; preferably, the processor 420 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, application programs, etc., and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor 420.

[0123] The system bus 430 is used to connect various functional components inside the computer, transmitting data, address, and control information. Its type can be, for example, a PCI bus, an ISA bus, or a CAN bus. Instructions from the processor 420 are transmitted to the memory 410 via the bus, and the memory 410 sends data back to the processor 420. The system bus 430 is responsible for data and instruction exchange between the processor 420 and the memory 410. Of course, the system bus 430 can also connect to other devices, such as network interfaces and display devices.

[0124] In this embodiment of the invention, the executable program executed by the processor 420 included in the electronic device includes: The system receives rib CT scan images and extracts rib candidate boxes from the scan images based on an improved YOLOv8 network with embedded convolutional block attention modules and a global attention mechanism. The improved ByteTrack algorithm tracks the input rib candidate boxes and generates a 3D rib point cloud trajectory. The ByteTrack algorithm integrates deformable convolution and spatiotemporal interaction mechanisms. After coordinate system correction and axial scaling of the three-dimensional rib point cloud trajectory, a cross-section reconstruction algorithm based on least squares fitting is used to generate rib MPR images, and fracture diagnosis is performed based on the rib MPR images.

[0125] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0126] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0127] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rapid diagnostic method for ribs, characterized in that, include: The system receives rib CT scan images and extracts rib candidate boxes from the scan images based on an improved YOLOv8 network with embedded convolutional block attention modules and a global attention mechanism. The improved ByteTrack algorithm tracks the input rib candidate boxes and generates a 3D rib point cloud trajectory. The ByteTrack algorithm integrates deformable convolution and spatiotemporal interaction mechanisms. After coordinate system correction and axial scaling of the three-dimensional rib point cloud trajectory, a cross-section reconstruction algorithm based on least squares fitting is used to generate rib MPR images, and fracture diagnosis is performed based on the rib MPR images.

2. The method according to claim 1, characterized in that, The improved YOLOv8 network based on embedded convolutional block attention modules and a global attention mechanism includes the following steps before extracting rib candidate boxes from the scanned image: Based on gradient-weighted class activation mapping technology, the attention distribution of the YOLOv8 network model at different levels is visualized and analyzed, and the levels where convolutional block attention modules and global attention mechanisms are inserted are selected respectively. The convolutional block attention module is integrated into the output features of the model backbone layer, and the global attention module is integrated into the output features of the model head layer. The extracted features of the convolutional block attention module and the global attention mechanism are fused based on the hierarchical attention fusion strategy. Construct an improved loss function for the YOLOv8 network, which includes a bounding box regression loss based on Distance IoU and a classification loss based on Focal Loss. Set a cross-frame semantic consistency strategy for detection boxes to fuse detection boxes in adjacent frames.

3. The method according to claim 2, characterized in that, The cross-frame semantic consistency strategy for setting detection boxes includes fusing detection boxes from adjacent frames: For CT scan images of adjacent frames, if the overlap between the detection box of the previous frame and the detection box of the current frame in the same rib region exceeds a preset value, the confidence scores of the two detection boxes are weighted averaged, or the box with higher confidence score is taken as the final detection box of the current rib region. IoU matching is used to perform inter-frame association of similar rib candidate boxes, and Kalman filtering is used to smooth the center point and size of the candidate boxes in the current frame.

4. The method according to claim 1, characterized in that, The process of tracking the input rib candidate boxes using the improved ByteTrack algorithm includes: Deformable convolutions are integrated into the target appearance feature extraction network of the ByteTrack algorithm, replacing some or all of the standard convolutional layers; In the ByteTrack algorithm, a spatiotemporal graph is constructed to represent the rib detection results on the slice and the temporal relationship between adjacent slices. The neighborhood node information of the spatiotemporal graph is extracted based on the graph neural network to realize spatiotemporal interaction.

5. The method according to claim 1, characterized in that, Before tracking the input rib candidate boxes using the improved ByteTrack algorithm, the following steps are also included: Set topological constraints for ribs. If the rib candidate box is located within the spinal region and the rib IDs and spatial relationships of the ribs that have been tracked around the candidate box do not match the labeled rib sequence, then reduce the matching priority of the candidate box or mark the candidate box as pending. A short-term forgetting mechanism is set up so that when a rib target is lost, the rib target is marked as to be recovered. When a rib object with similar appearance features or the same trajectory is detected again in a subsequent frame, the rib object in the subsequent frame is matched with the rib target marked as to be recovered.

6. The method according to claim 1, characterized in that, The rib MPR image generated by the least squares fitting-based section reconstruction algorithm includes: The three-dimensional rib point cloud trajectory is used as the parameterized curve of the rib, the direction of the rib's long axis is defined as the tangent vector of the curve, the cross section perpendicular to the long axis of the rib is used as the MPR plane, and the tangent vector is set as the normal vector of the MPR plane. Define a series of MPR planes parallel to the MPR plane normal vector, slice along the long axis of the rib, and sample each MPR plane to generate an MPR image; The process involves constructing a coordinate transformation from the original CT 3D voxels to the MPR plane. Based on this coordinate transformation, each pixel in the MPR image is mapped to a point in 3D space. By interpolating the voxel values ​​sampled in the original CT scan image, the grayscale values ​​of each pixel in the MPR image are obtained.

7. The method according to claim 1, characterized in that, The rib MPR image generation algorithm based on least squares fitting also includes: A bidirectional error feedback mechanism is constructed to compare the rib contour in the MPR image with the projection of the rib candidate box on the MPR image plane, calculate the model detection error, and feed the detection error as a loss term back to the YOLOv8 network. The YOLOv8 network parameters are updated through backpropagation. By analyzing the local continuity and morphology of MPR images, the model tracking error is evaluated, and the tracking error is fed back to the ByteTrack algorithm to fine-tune the deformable convolution and graph neural network in the ByteTrack algorithm.

8. A system for rapid diagnosis of ribs, characterized in that, include: The target detection module is used to receive rib CT scan images and extract rib candidate boxes from the scan images based on an improved YOLOv8 network with embedded convolutional block attention modules and a global attention mechanism. The target tracking module is used to track the input rib candidate boxes using the improved ByteTrack algorithm and generate three-dimensional rib point cloud trajectories. The ByteTrack algorithm integrates deformable convolution and spatiotemporal interaction mechanisms. The planar reconstruction module is used to perform coordinate system correction and axial scaling on the three-dimensional rib point cloud trajectory, generate rib MPR images based on the least squares fitting sectional reconstruction algorithm, and perform fracture diagnosis based on the rib MPR images.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of a rapid rib diagnosis method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed, it implements the steps of a rapid rib diagnosis method as described in any one of claims 1 to 7.

Images