Vertebral compression fracture recognition method based on x-ray film

By introducing a deformable attention mechanism and a large separable nuclear attention module in the YOLOv8 framework, combined with a dynamic detection head, the problem of accuracy and computational complexity of fracture detection methods is solved, and the low data volume and complexity of the background Efficient recognition effect.

CN119006792AActive Publication Date: 2024-11-22NANCHANG UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411137481.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-11-22
Estimated Expiration
2044-08-19

AI Technical Summary

Technical Problem

Existing fracture detection methods perform poorly in recognition accuracy and computational complexity, making it difficult to meet the needs of clinical applications, especially when data is limited and when processing fracture images at different scales and complex backgrounds.

Method used

Add a deformable attention mechanism in the YOLOv8 framework, improve the SPPF module to reduce the computational complexity through the large separable core attention module, and replace the detection head with a dynamic detection head, adopting the self-attention mechanism of scale perception, spatial perception and task perception To improve the representation ability of target detection.

Benefits of technology

Significantly improve the accuracy of the model, reduce the computational complexity and memory usage, can effectively identify lumbar fractures on a smaller data set, reduce the doctor's image acquisition time and processing workload, and enhance the robustness and generalization of the model. ability to transform.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119006792A_ABST
    Figure CN119006792A_ABST
Patent Text Reader

Abstract

Aiming at the problems of high labor intensity of clinical doctors, low working efficiency of orthopedists and the like, the deep learning method is applied to medical image analysis gradually and widely. The invention provides a centrum compression fracture recognition method based on an x-ray film, and aims to improve the recognition precision and efficiency of a model and reduce the calculation complexity and memory occupation by improving a YOLOv8 model structure and introducing a C2f-DAtention module, an LSKA module and a dynamic detection head. Therefore, the automatic detection method for the vertebral compression fracture is rapid, accurate and low in cost, and clinical doctors are assisted to improve the diagnosis efficiency and accuracy. The mAP50 of the improved experimental model is improved by 9.1% and the mAP50-95 is improved by 16.6% on a data set compared with an original model yov8m, and the mAP50 and the mAP50-95 are respectively 89.6% and 71%. Therefore, the algorithm precision of the compression fracture is improved while the calculation complexity of the model is not increased, and an effective means is provided for realizing accurate vertebral compression fracture diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of medical image analysis, in particular to an X-ray-based vertebral compression fracture identification method. Background Art

[0002] When clinicians identify compression fractures, they usually conduct a preliminary assessment and determine the location and extent of the fracture through physical examination and observation of the patient's posture, gait, and local tenderness. Imaging assessment methods are precise methods for determining the location of compression fractures. They use imaging examinations such as X-rays, CT scans, and MRI to accurately display the compression and deformation of the fracture site, which helps to confirm and assess the severity of the fracture. At present, X-rays are generally used to diagnose vertebral compression fractures. Its detection cost is significantly lower than MRI and CT, the detection speed is faster, the radiation to the human body is minimal, and the equipment is widely used in medical institutions. Therefore, our series of studies are also based on X-rays.

[0003] With the application of deep learning in medical imaging, a research boom has gradually emerged to explore deep learning methods to solve fracture diagnosis tasks. From the development of computer-aided detection and diagnosis methods to the application of various deep learning networks in compression fracture recognition, the recognition effect is getting better and better. However, there is still room for improvement in its accuracy. If you want to truly apply these models to clinical diagnosis, such accuracy is obviously not enough.

[0004] In the traditional diagnosis process, doctors need to find abnormal lesions such as lung nodules or inflammation from a large number of X-rays, which often relies on the doctor's experience and naked eye observation, which is not only time-consuming and labor-intensive, but also may have the risk of misjudgment. The YOLOv8 algorithm can realize automatic detection of X-rays, which can improve the diagnosis speed and reduce the risk of misjudgment to a certain extent. Due to its high precision and fast detection capabilities, the YOLOv8 algorithm is highly favored in the field of medical image recognition and has become an ideal choice. Medical image processing has extremely high requirements for accuracy and detection efficiency, and YOLOv8 perfectly meets these requirements with its powerful performance. The algorithm framework of YOLOv8 has good scalability. Using this algorithm framework, researchers can identify lesions and abnormal structures in medical images more quickly and accurately, providing important support for the early detection and treatment of diseases.

[0005] In the process of implementing the technical method of the embodiment of the present invention, the inventor of this patent found at least the following technical problems in the prior art: 1. Existing fracture detection methods perform poorly in terms of recognition accuracy and average detection accuracy, and are difficult to meet clinical application needs.

[0006] 2. Although some advanced deep learning methods have improved detection accuracy, they are accompanied by high computational complexity and memory usage, which limits their application in actual medical environments.

[0007] 3. Existing models are highly dependent on large-scale data sets and perform poorly when the amount of data is small, making them difficult to apply to data-limited scenarios in clinical practice.

[0008] 4. Existing methods have limitations in capturing subtle features of compression fractures and fusing multi-scale features, which leads to unstable performance of the model when processing fracture images of different scales and complex backgrounds.

[0009] In summary, existing fracture detection methods cannot meet actual clinical diagnosis needs. Summary of the invention

[0010] The embodiment of the present invention provides a method for identifying vertebral compression fractures based on X-ray films that combines efficiency and accuracy, solving the technical problem that existing fracture detection methods cannot meet actual clinical needs.

[0011] The present invention provides a method for identifying vertebral compression fractures based on X-ray films, comprising: Add a deformable attention mechanism to the last C2f module in the yolov8 framework backbone network. The positions of key and value pairs are selected in a data-dependent manner. This flexible scheme enables the self-attention module to focus on relevant areas and capture more informative features. The SPPF module in the backbone network is improved by adding a large separable kernel attention module to decompose the two-dimensional convolution kernel of the deep convolution layer into cascaded horizontal and vertical one-dimensional kernels, reducing the computational complexity of the model; Replacing the detection head with a dynamic detection head significantly improves the representation capability of the object detection head without increasing any computational overhead.

[0012] One or more technical solutions provided in the embodiments of the present invention have at least the following technical effects or advantages: 1. The model of the present invention can greatly improve the accuracy of the model.

[0013] 2. The model of the present invention has lower computational complexity and less memory usage.

[0014] 3. Only less than 500 lumbar fracture datasets are needed, which saves doctors’ image acquisition time and reduces the workload of image processing.

[0015] 4. This method significantly improves the representation capability of the object detection head by coherently combining multiple self-attention mechanisms between feature levels for scale-awareness, between spatial locations for space-awareness, and within output channels for task-awareness, without adding any computational overhead. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A diagram of a Yolov8m-CF model in one embodiment of the present invention; Figure 2 An illustrative diagram of a deformable attention mechanism in one embodiment of the present invention; Figure 3 This is a C2f structure diagram in which a DAT mechanism is added in one embodiment of the present invention; Figure 4 This is a structural diagram of an SPPF with an LSKA mechanism added in one embodiment of the present invention; Figure 5 A diagram of a dynamic header module in one embodiment of the present invention; Figure 6 A diagram showing the layout and style of an interface in an embodiment of the present invention; Figure 7 A diagram showing the parameter column on the right side in one embodiment of the present invention; Figure 8 A model display diagram is introduced in one embodiment of the present invention; Fig. 9 This is a picture display diagram of uploading pictures in one embodiment of the present invention; Fig.10 A successful display diagram is shown in an embodiment of the present invention; Fig.11 This is a diagram showing the prediction results in one embodiment of the present invention. DETAILED DESCRIPTION

[0017] This technical solution mainly makes three innovative modifications to the yolov8 network, which are introduced in detail as follows.

[0018] Improve the network structure of yolov8: Although Yolov8 has achieved certain results in detection accuracy and speed, it cannot meet our expectations in the small sample compression fracture recognition task. Therefore, in order to make the model pay more attention to the characteristics of compression fractures and thus improve the accuracy of fracture recognition, this paper proposes the Yolov8m-CF model, please refer to Figure 1 .

[0019] Improved C2f structure: The C2f structure in Yolov8 is an improved convolutional module, which is more complex than the traditional convolutional module and contains more layers and parameters, which significantly increases the training and inference time when training with high-resolution images and large-scale datasets. Therefore, a deformable attention vision transformer (Vision Transformer with DeformableAttention, DAT) is proposed. The traditional Transformer uses a standard attention mechanism, which processes all pixels in the image, resulting in a large amount of calculation. DAT introduces a deformable attention mechanism, which only focuses on a small part of the key area in the image. We apply this deformable mechanism to the C2f structure of yolov8.

[0020] See also Figure 2 , which is an illustration of the deformable attention mechanism. On the left part, a set of reference points are evenly placed on the feature map, and the offsets of these reference points are learned from the offset network. The deformed keys and values ​​are then projected from the sampled features according to the deformation points. Finally, standard multi-head attention is applied to focus on queries on the sampled keys and aggregate features from the deformed values. In addition, the positions of the deformation points provide a stronger relative position bias to facilitate the learning of deformable attention. For clarity of presentation, only 4 reference points are shown here, and there are actually more reference points in the actual implementation.

[0021] See also Figure 3 As shown, the present invention introduces the DAT mechanism in the C2f structure. In X-rays, the features of compression fractures may be subtle and difficult to capture. The DAT mechanism can dynamically adjust the attention weight, highlight the features of the fracture area, and improve the model's sensitivity to these important areas, thereby enhancing the recognition ability of compression fractures. It enables the model to better adapt to X-rays of different patients and enhances the generalization ability and robustness of the model in different scenarios. X-rays may contain a lot of background information and noise, and the DAT mechanism can effectively suppress these irrelevant features, allowing the model to focus more on the fracture area. The reduction in the number of parameters will not affect the detection performance.

[0022] Improved SPPF structure: The SPPF (Spatial Pyramid Pooling Fast) structure in YOLOv8 is an improved spatial pyramid pooling (SPP) module designed to improve feature extraction efficiency and computational performance. The SPPF structure enhances feature representation capabilities by fusing multi-scale features while maintaining high computational efficiency. Although SPPF helps with multi-scale feature fusion, compression fractures are relatively small features, and the structure has limited performance in small target detection. Since the SPPF structure increases the complexity of the model, the model may learn noise and details in the training data, resulting in overfitting. Therefore, the LSKA (Large Separable Kernel Attention Module) is introduced in the SPPF module to decompose the two-dimensional convolution kernel of the deep convolution layer into cascaded horizontal and vertical one-dimensional kernels, reduce the complexity of the model, and use deeper convolution layers to extract compression fracture features.

[0023] See also Figure 4 As shown in the figure, the LSKA mechanism is added to the SPPF structure. It is an attention mechanism that focuses on capturing large-scale spatial information. Compression fractures may be very subtle and hidden in X-rays, and the surrounding bone structure and background information will have an important impact on their detection. LSKA captures a wider range of contextual information through large-size convolution kernels, which helps the model to more accurately identify fracture features. Since LSKA can capture a wider range of spatial features, it is very beneficial for the detection of detailed features such as fracture edges in X-rays, and can improve the model's sensitivity to details and recognition accuracy. Small fractures are often difficult to detect in medical images. The LSKA mechanism can better focus on small targets through large-size convolution kernels and attention mechanisms, thereby improving the detection effect of small targets.

[0024] Improved dynamic detection head: The detection head in YOLOv8 is one of the key parts of the network, responsible for generating bounding boxes and classification predictions. Although its detection head has significant improvements in speed and accuracy compared to the previous YOLO series, there are still some problems in fracture detection: the features of compression fractures may be diluted or ignored in high-level feature maps, resulting in a decrease in detection rate. Detection of unclear fracture images will also be difficult. Therefore, a new dynamic head framework is proposed to solve these problems. It uses attention mechanisms in three dimensions to capture subtle fracture features more accurately. The present invention replaces the original detection head of yolov8 with it.

[0025] Given a feature tensor , apply self-attention The general expression of is:

[0026] in is the attention function. This solution of attention function is implemented by fully connected layers. But directly learning attention functions over all dimensions is computationally expensive and practically unaffordable due to the high dimensionality of tensors.

[0027] In contrast, the present invention transforms the attention function into three consecutive attentions, each focusing on only one angle:

[0028] In the formula , , They are three different attention functions acting on dimensions L, S, and C respectively.

[0029] First, we introduce scale-aware attention ,dynamically fuse features according to the semantic importance at different scales.

[0030]

[0031] in, It is a linear function approximated by a 1*1 convolutional layer. is the hard sigmoid function.

[0032] The present invention adopts another spatial perception attention module based on fusion features To focus on discriminative regions that consistently coexist between spatial locations and feature levels. Considering the high dimensionality of S, this module is decomposed into two steps: first, using deformable convolution to make attention learning sparse, and then aggregating features across layers at the same spatial location:

[0033] Among them, L is the number of pyramid layers, K is the number of sparse sampling positions, w is the width of the feature layer, c is the number of channels, and the number of layers is the self-learning spatial offset The position of the shift to focus on a discrimination area, Position for self-learning Both are learned from the input features at the median level of F.

[0034] To achieve joint learning and summarizing different representations of objects, task-aware attention is deployed at the end It dynamically switches the ON and OFF channels of functionality to support different tasks:

[0035] in, for The characteristic slice of the channel, A hyperfunction for learning to control the activation threshold. The implementation is first dimensionality to reduce the dimensionality, then two fully connected layers and a normalization layer are used, and finally the output is normalized to using the shifted sigmoid function .

[0036] Finally, since the above three attention mechanisms are applied sequentially, the formula can be nested multiple times, effectively combining multiple , , Blocks are stacked together. Figure 5 , which shows the detailed configuration of the dynamic head (ie, DyHead simplified) module of the present invention.

[0037] The dynamic detection head can adaptively adjust according to the features of the input image through dynamic convolution kernels and attention mechanisms. This adaptive feature enables the model to better extract the features of the fracture area when processing X-rays. Multi-scale feature fusion and dynamic attention mechanisms can better focus on small targets of compression fractures and improve detection accuracy.

[0038] The following is a flow chart of the present invention: The layout and style of the display interface are implemented using Pyside, see Figure 6 ; Open the parameter bar on the right, see Figure 7 ; Import the trained weight model best.pt, conf is the confidence, delay is the delay in seconds to display the image, please refer to Figure 8 ; Open the image directory to be tested, click open in the upper right corner, and upload the image. Fig. 9 ; The lower left corner shows load file 657.jpg, indicating that the image has been imported successfully. Fig.10 ; Click the Start button to display the prediction results. Fig.11 .

[0039] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for identifying vertebral compression fractures based on X-ray films, characterized in that: include: Add a deformable attention mechanism to the last C2f module in the yolov8 framework backbone network. The positions of key and value pairs are selected in a data-dependent manner. This flexible scheme enables the self-attention module to focus on relevant areas and capture more informative features. The SPPF module in the backbone network is improved by adding a large separable kernel attention module to decompose the two-dimensional convolution kernel of the deep convolution layer into cascaded horizontal and vertical one-dimensional kernels, reducing the computational complexity of the model; Replacing the detection head with a dynamic detection head significantly improves the representation capability of the object detection head without increasing any computational overhead.

Citation Information

Patent Citations

  • Fall detection method based on lightweight YOLOv8 network

    CN117912111A

  • Lightweight target detection network design method for indoor semantic SLAM system

    CN118097365A