Method and system for judging stability of ankle joint

Abstract

The invention discloses a method and a system for judging the stability of an ankle joint. The system comprises an image acquisition module, a network construction module, a skeleton edge point positioning module and a joint stability judgment module, the image acquisition module performs frame extraction and denoising on the ankle ultrasonic video, and marks tibia and fibula edge points; the network construction module generates a high-precision target positioning network; the skeleton edge point positioning module is used for positioning skeleton edge points in real time through the trained network; the joint stability judgment module is used for sequentially calculating the bone gaps when the ankle joints of the patient face the outer side and the neutral position and calculating the final ratio, and when the extorsion position distance is obviously larger than the neutral position, namely, the stretch ratio is larger than 1, it is considered that the joints are unstable. According to the invention, joint problems can be quickly responded during operation, and doctors can be assisted in diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing and artificial intelligence, and in particular to a method and system for judging ankle joint stability. BACKGROUND

[0002] Ankle fracture is one of the most common fractures in clinical practice, accounting for about 3.9% of all fractures in the human body. Due to the presence of numerous complex and interlaced ligaments in the ankle joint of the human body, damage to the ankle joint often involves the destruction of the ligament complex. Among all the ligament sprains caused by ankle joint injury, tibiofibular ligament injury is the most common. Failure to diagnose or ignore such injuries in a timely manner can lead to various joint abnormalities and trigger post-traumatic arthritis.

[0003] Current research on tibiofibular joint injury is mainly preoperative, which cannot reflect the stability of the ankle joint after surgery and cannot guide fracture fixation surgery. In the latest medical research, a new method for confirming intraoperative ankle joint stability is proposed. This method compares the joint space length ratio of the patient's foot when it is laterally outward and when it is not under external force to determine joint stability. Joint reduction is a complex, precise and common clinical procedure. Ignoring the stability of the patient's joint during surgery will lead to various postoperative complications. Traditional intraoperative diagnosis is highly dependent on the experience of doctors, requiring high levels of workers, which puts a lot of pressure on practitioners, and the accuracy is limited. If artificial intelligence technology is introduced to assist doctors in diagnosis, it can improve accuracy while reducing the requirement for doctors' professional knowledge, facilitating the operation.

[0004] At the same time, ultrasound imaging is a low-cost, low-radiation medical diagnostic method that does not cause radiation to patients like CT and X-ray examinations, nor does it cause invasive wounds like arthroscopy. In addition, ultrasound acquisition devices come in a variety of sizes and can meet a variety of application needs.

[0005] Current deep learning technology has been widely used in the medical field. However, most existing artificial intelligence assisted diagnosis methods choose to segment the entire bone as the recognition means. This task of recognizing complete tissue structures often encounters various problems, such as low resolution of ultrasound images, background speckles and blurred boundaries of other structures caused by complex human tissues, and other challenges that reduce detection rates. In addition, the calculation complexity of segmenting the entire structure is high, which cannot meet the real-time requirements of intraoperative diagnosis. At this time, the medical community has also proposed a new intraoperative diagnosis scheme that can obtain the stability of the patient's joint through simple calculations. This scheme has low computational requirements and the results are more concise and easy to understand. However, this newly proposed scheme does not have a matching deep learning assisted technology. SUMMARY

[0006] The application provides a method and system for judging ankle joint stability, which can quickly respond to joint problems during surgery and assist doctors in making diagnoses.

[0007] The application provides a method for judging ankle joint stability, which comprises the following steps:

[0008] Step 1, data acquisition and preprocessing; frame extraction and denoising are performed on the ankle ultrasound video, and tibia and fibula edge points are labeled;

[0009] Step 2, network training; a multi-stage deep learning network is constructed, a background filtering module uses RPN as a backbone network to realize background rough screening, a semantic feature extraction module and a frequency domain feature extraction module complete multi-dimensional feature extraction, a self-attention modeling module based on the Transformer optimizes feature weights, and finally a high-precision target positioning network is generated;

[0010] Step 3, bone edge point positioning; the trained network is used to position the bone edge points in real time;

[0011] Step 4, joint stability judgment; the bone gaps of the patient's ankle joint when the ankle joint is outward and in a neutral position are calculated in sequence, and the final ratio is calculated, and when the distance in the outward rotation position is significantly greater than that in the neutral position, that is, the stretching ratio is greater than 1, it is considered that the joint is unstable.

[0012] Further, in step 1, the ankle ultrasound video is subjected to frame extraction and denoising, and the tibia and fibula edge points are labeled, the speckle noise in the ultrasound image is removed, the CLAHE algorithm is used for adaptive histogram equalization to enhance the contrast, and the motion amplitude detection algorithm is used to select clear frames containing the tibiofibular lower gap, and the displacement amount greater than a preset threshold (≥5 pixels) is reserved.

[0013] Step 11, the stable feature point set of the tibiofibular edge is obtained through a SIFT feature point extractor, and a RANSAC algorithm is used to calculate an affine transformation matrix between adjacent frames;

[0014] Step 12, the translation vector is obtained after the transformation matrix is decomposed, and the mean value of the displacement amount of the feature point cluster is calculated as a motion amplitude quantitative index;

[0015] Step 13, a dynamic threshold adjustment mechanism based on image resolution is introduced, when the ultrasound probe frequency is 10MHz, the corresponding pixel size is 0.1mm, the basic threshold is set to 5 pixels, corresponding to 0.5mm anatomical displacement, and a threshold automatic conversion formula is established according to the probe parameters;

[0016] Step 14, to solve the problem of feature point loss caused by large displacement, a multi-scale analysis module is designed to construct a three-level resolution pyramid through down-sampling processing. After rough matching is completed in the low resolution layer, it is transferred to the original resolution layer for fine calculation. Compared with the traditional frame difference method, the motion detection accuracy is improved by 32.6% (p<0.01), which can effectively screen out high-quality frames with displacement ≥5 pixels and clear tibiofibular space anatomic structure; manual labeling, the anatomical feature points of the distal tibia and the distal fibula are distinguished during labeling, and the accurate coordinates of the two edge points are given, and the Ground Truth image with a size not exceeding 2x2 is generated by the deep learning model trainer.

[0017] Further, in step 2, the background filtering module uses RPN to realize background rough screening, uses a lightweight residual network as the backbone network, selects the second stage of Resnet18 as the network main body, filters the background and outputs the candidate region.

[0018] Further, in step 2, the semantic feature extraction module and the frequency domain feature extraction module fuse semantic and frequency domain information, and use a full convolution network without down-sampling to extract the spatial features of each micro pixel point. The extraction result is denoted as FS, and the calculation formula is as follows:

[0019] FS = ReLU(BN(Wn x (ReLU(BN(Wn-1(...ReLU(BN(W1 x FI)))...)))))

[0020] Where, ReLU represents the activation function, BN represents the normalization operation, Wi represents the i-th layer convolution, and n is set to 8.

[0021] Further, an octave convolution network is used to extract frequency domain features for distinguishing edges and main bodies. YH and YL represent high-frequency and low-frequency features respectively, and the calculation formula is as follows:

[0022] Y H =F(X H ;W H→H )+Upsample(F(X L ;W L→H ),2)

[0023] Y L =F(X H ;W H→H )+F(pool(X H ,2);W H→L )

[0024] wherein, wherein F(X; W) represents a convolution with parameter W, which varies with the image, pool(x, k) is an average pooling operation with a step of k and a k x k kernel, and Upsample(F(X; K)) is an up-sampling operation by nearest interpolation with parameter K;

[0025] wherein, FP represents the fusion of the extracted high-frequency features and low-frequency features, and the calculation formula is as follows:

[0026]

[0027] wherein, Resize represents adjusting the features to a fixed size, represents element-wise addition.

[0028] Further, in step 2, the Transformer-based self-attention modeling module optimizes the feature weights, inputs the double-branch features into the Transformer encoder, and models the relevance between the target pixels through the multi-head attention mechanism, and the formula is as follows:

[0029]

[0030] wherein, A represents an attention matrix, X represents the i-th layer feature, k represents a scaling factor, m = 1, …, M represents an M-head multi-head attention module, i represents an encoding layer, Q, V, and K are learnable parameters, Contact represents a series operation, and the mapping matrix W and the M-head attention module run in parallel.

[0031] Further, in step 3, the edge points of the skeleton are located in real time by the trained network, the Resnet is processed by lightweight, and the anchor box fusion detection method is used to quickly filter the background. The image detection rate should be greater than or equal to 24 frames per second, which can meet the real-time processing of the input ultrasonic image in the operation and output the coordinates of the tibia and fibula edge points in real time. The output result should reach sub-millimeter positioning accuracy (error <0.2mm). The double-flow network structure is based on the double-branch topology structure decoupled by anatomical features, which is divided into tibia special branch and fibula special branch. The tibia special branch integrates lightweight ResNet-18 as the backbone network, reduces the computational complexity through channel pruning (compression rate 65%) and depth separable convolution, while retaining the morphological feature capture ability of the wide joint surface of the distal tibia. The fibula special branch adopts the U-Net variant architecture, enhances the edge response of the narrow fibula joint through the multi-scale hollow convolution module (expansion rate = [2, 4, 6]), and designs a cross-layer attention gate mechanism to suppress the interference of tendon artifacts. In the decoding stage, a deformable convolution pyramid (Deformable FPN) is introduced to dynamically fuse the double-flow feature maps through learnable weights, focusing on enhancing the cross-skeletal feature expression of the tibiofibular joint space (<1mm), and overcoming the pain point of blurred bone edges in traditional ultrasonic images. At the same time of outputting the bone edge point coordinates, a range anchor box with confidence is visualized in the image. If the image is blurred due to intraoperative bleeding and other problems, a low confidence warning will appear, reminding the doctor to verify the model output result again to ensure the safety of the operation.

[0032] Further, in step 4, the edge point coordinates of the tibia and fibula output in step 3 are accepted, recorded as the tibia edge point (t x ,t y ), and the fibula edge point (f x .f y ), and substituted into the Euclidean distance calculation formula, The minimum distance d between the tibia and fibula is calculated. In order to eliminate the influence of individual differences, a relative index, the stretch ratio, is introduced in the method. The gap width at the neutral position is denoted as N, and the gap width at the external rotation position is denoted as E. The stretch ratio is E / N. If the calculated stretch ratio is greater than 1, the joint is considered unstable, otherwise the joint is considered stable. When the calculated ratio is greater than 1, a warning is given, and when the ratio is less than 1, only the calculation result is output. Through this simple diagnostic standard, the joint stability of the patient during the operation can be quickly and real-time evaluated, thereby guiding the fixation of the tibiofibular joint and reducing the risk of postoperative traumatic arthritis.

[0033] Correspondingly, a system for judging ankle joint stability comprises an image acquisition module, a network construction module, a bone edge point positioning module, and a joint stability judgment module; the image acquisition module performs frame extraction and denoising on ankle ultrasound video, and labels tibia and fibula edge points; the network construction module generates a high-precision target positioning network; the bone edge point positioning module positions bone edge points in real time through the trained network; and the joint stability judgment module calculates the bone gap of the patient's ankle joint when the ankle joint is outward and at a neutral position, and calculates the final ratio, and when the distance at the outward rotation position is significantly greater than that at the neutral position, that is, the stretching ratio is greater than 1, it is considered that the joint is unstable.

[0034] Further, the network construction module comprises a background filtering module, a semantic feature extraction module, a frequency domain feature extraction module, and a self-attention modeling module; the region proposal module realizes background rough screening, the multi-dimensional feature extraction module fuses semantic and frequency domain information, the self-attention modeling module based on the Transformer optimizes feature weights, and finally a high-precision target positioning network is generated.

[0035] Advantages: Compared with the prior art, the present application has the following remarkable advantages: the present application judges the joint stability by measuring the distance between two bones, the trained model has faster response speed due to a large amount of background clutter filtering, the speed of processing 24 frames of pictures per second can adapt to the ultrasound video collected by most devices, and real-time processing can be achieved for use in surgery; the tibia-fibula double-branch network structure adopted can further reduce errors after the addition of the multi-dimensional extraction module of semantic and frequency domain features, and finally the self-attention modeling module is introduced to make the output result more condensed, and the final error can be kept at about 0.1 mm, which is much lower than the performance of other target detection models. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The structure diagram of the network trained by the present application.

[0037] Figure 2 The operation schematic diagram of the ankle joint stability tension test of the present application.

[0038] Figure 3 The method flowchart of the present application. DETAILED DESCRIPTION

[0039] As shown in Figure 3 A method for judging ankle joint stability comprises the following steps:

[0040] Step 1, data acquisition and preprocessing; frame extraction and denoising are performed on ankle ultrasound video, and tibia and fibula edge points are labeled.

[0041] The ankle ultrasound video is frame extracted and denoised, the tibia and fibula edge points are labeled, the speckle noise in the ultrasound image is removed, and the CLAHE algorithm is used for adaptive histogram equalization to enhance the contrast; Based on the motion amplitude detection algorithm, the clear frame containing the tibiofibular space is screened, and the displacement greater than the preset threshold (≥5 pixels) is retained; manual labeling, distinguishing the anatomical feature points of the distal tibia and fibula, giving the accurate coordinates of the two edge points, and the deep learning model trainer generates a Ground Truth image with a size of not more than 2×2.

[0042] The measurement point is to place the ultrasound probe 2 cm above the lateral malleolus, as shown in Figure 2 , and the range of the target point is shown by the yellow box, where T is the tibia edge point, F is the fibula edge point, and the dashed line represents the joint space length.

[0043] Step 2, network training; construct a multi-stage deep learning network, the background filtering module RPN realizes the rough screening of the background, the semantic feature extraction module and the frequency domain feature extraction module fuse the semantic and frequency domain information, the self-attention modeling module based on Transformer optimizes the feature weight, and finally generates a high-precision target positioning network.

[0044] Use the SGD optimizer for training, set the training batch to 30, and set the initial learning rate to 10 -4 . Training is performed on a computer with a 3.40GHz Intel Core i7-13700kf and a 12G GPU configured with an RTX4070, as shown in Figure 1 .

[0045] The background filtering module uses RPN to realize the rough screening of the background, uses a lightweight residual network as the backbone network, selects Resnet18 as the second stage as the network main body, filters the background and outputs the candidate region.

[0046] The semantic feature extraction module and the frequency domain feature extraction module fuse the semantic and frequency domain information, use a non-downsampling full convolution network to extract the spatial features of each micro pixel point, and the extraction result is denoted as FS, and the calculation formula is as follows:

[0047] FS = ReLU(BN(Wn×(ReLU(BN(Wn-1(...ReLU(BN(W1×FI)))...)))))

[0048] Where ReLU represents the activation function, BN represents the normalization operation, Wi represents the i-th layer convolution, and n here is set to 8.

[0049] Frequency domain features: the frequency domain features are extracted using an octave convolution network, which is used to distinguish the edge and the main body, and YH and YL represent the high-frequency and low-frequency features respectively, and the calculation formula is as follows:

[0050] Y H = F(X H ; W H→H ) + Upsample (F(X L ; W L→H ), 2)

[0051] Y L = F(X H ; W H→H ) + F(pool(X H , 2); W H→L )

[0052] Wherein, wherein F(X; W) represents a convolution with parameters W, which can vary with the change of the image, pool(x, k) has an average pooling operation with a step of k and a k*k kernel, and Upsample(F(X; K)) is an up-sampling operation by nearest interpolation, with parameter K.

[0053] Let FP represent the fusion of the extracted high-frequency features and low-frequency features, and the calculation formula is as follows:

[0054]

[0055] Wherein, Resize represents adjusting the feature to a fixed size, Represents element-wise addition.

[0056] The self-attention modeling module based on the Transformer optimizes the feature weight, inputs the double-branch feature into the Transformer encoder, and models the relevance between the target pixels through the multi-head attention mechanism, and the formula is as follows:

[0057]

[0058] Wherein, A represents the attention matrix, X represents the i-th layer feature, k represents the scaling factor, m = 1, …, M represents the M-head multi-head attention module, i represents the encoding layer, Q, V, K are learnable parameters, Contact represents the series operation, and the mapping matrix W and the M-head attention module run in parallel.

[0059] Step 3, locate the bone edge point; locate the bone edge point in real time through the trained network.

[0060] The trained network is used to locate the bone edge points in real time, the Resnet is lightened, and the anchor box fusion detection method is used to quickly filter the background. The image detection rate should be greater than or equal to 24 frames per second, which can meet the real-time processing of input ultrasound images during surgery and output the coordinates of the tibia and fibula edge points in real time. The output result should reach sub-millimeter positioning accuracy (error <0.2mm). The specially designed double-stream network structure processes the edge features of the tibia and fibula respectively, and enhances the recognition ability of small bone joints through the feature pyramid fusion layer to overcome the pain point of blurred bone edge in traditional ultrasound images. For the specialized diagnosis and analysis process of ankle fracture reduction surgery, a 0-50N full-metal force gauge should be used to achieve standardization, and an "F" type tool should be used to fix the direction of force. When measuring, the patient's leg should be straightened and placed on the measuring bed, the foot dorsum should be flexed 90°, the measuring point should be set 2cm above the lateral malleolus at the tibiofibular interval, and a 7.2N·m torque should be applied to the patient's ankle when recording the external rotation position. The model will detect and output the tibia and fibula edge points at the neutral position and external rotation position respectively. At the same time of outputting the bone edge point coordinates, a range anchor box containing confidence will be visualized in the image, if the image is blurred due to intraoperative bleeding and other problems, a low confidence warning will appear, reminding the doctor to verify the model output result again to ensure the safety of the surgery.

[0061] Step 4, joint stability judgment; the bone space of the patient's ankle joint to the outside and the neutral position is calculated in turn, and the final ratio is calculated. If the distance at the external rotation position is significantly greater than that at the neutral position, i.e. the stretch ratio is greater than 1, it is considered that the joint is unstable.

[0062] Based on the bone edge position obtained in step 3, the shortest distance of the tibiofibular interval is output using the Euclidean distance calculation formula. In order to eliminate the influence of individual differences, a relative index, the stretch ratio, is introduced in the method. The interval width at the neutral position is denoted as N, and the interval width at the external rotation position is denoted as E, then the stretch ratio is E / N; if the calculated stretch ratio is greater than 1, it is considered that the joint is unstable, otherwise it is considered that the joint is stable. Therefore, when the calculated ratio is greater than 1, a warning is given, and when the ratio is less than 1, only the calculation result is output. Through this simple diagnostic standard, the joint stability of the patient during surgery can be quickly and real-time evaluated, thereby guiding the fixation of the tibiofibular joint and reducing the risk of postoperative traumatic arthritis.

Claims

1. A method for determining ankle joint stability, characterized in that, Includes the following steps: Step 1: Data Acquisition and Preprocessing; Frame extraction and noise reduction are performed on the ankle ultrasound video, and the edge points of the tibia and fibula are marked; Step 2: Network Training; Construct a multi-stage deep learning network. The background filtering module uses RPN as the backbone network to achieve coarse background screening. The semantic feature extraction module and the frequency domain feature extraction module complete multi-dimensional feature extraction. The Transformer-based self-attention modeling module optimizes feature weights and finally generates a high-precision target localization network. Step 3: Bone edge point localization; real-time localization of bone edge points using the trained network; Step 4: Joint stability assessment; calculate the interosseous gaps of the patient's ankle joint in the lateral and neutral positions, and calculate the final ratio. If the gap in the external rotation position is significantly greater than that in the neutral position, i.e., the stretch ratio is greater than 1, then the joint is considered unstable.

2. The method for determining ankle joint stability as described in claim 1, characterized in that, In step 1, frame extraction and denoising are performed on the ankle ultrasound video, and the edge points of the tibia and fibula are marked. The specific steps include the following: Step 11: Obtain a stable set of feature points on the edge of the tibia and fibula using the SIFT feature point extractor, and calculate the affine transformation matrix between adjacent frames using the RANSAC algorithm. Step 12: After decomposing the transformation matrix, obtain the translation vector and calculate the average displacement of the feature point cluster as a quantitative index of motion amplitude. Step 13: Introduce a dynamic threshold adjustment mechanism based on image resolution. When the ultrasound probe frequency is 10MHz, the corresponding pixel size is 0.1mm. Set the basic threshold to 5 pixels, which corresponds to 0.5mm anatomical displacement. Establish an automatic threshold conversion formula based on probe parameters. Step 14: Design a multi-scale analysis module. Construct a three-level resolution pyramid through downsampling. After coarse matching is completed in the low-resolution layer, the pyramid is passed to the original resolution layer for fine calculation. Manually annotate the anatomical feature points of the distal tibia and distal fibula. The annotation provides the precise coordinates of the edge points on both sides. Generate a Ground Truth image with a size not exceeding 2×2 for the annotation points.

3. The method for determining ankle joint stability as described in claim 1, characterized in that, In step 2, the background filtering module uses RPN to perform background coarse screening, adopts a lightweight residual network as the backbone network, selects the second stage of ResNet18 as the main body of the network, filters the background and outputs candidate regions.

4. The method for determining ankle joint stability as described in claim 1, characterized in that, In step 2, the semantic feature extraction module and the frequency domain feature extraction module fuse semantic and frequency domain information, and use a fully convolutional network without downsampling to extract the spatial features of each tiny pixel. The extraction result is represented by FS, and the calculation formula is as follows: FS=ReLU(BN(Wn×(ReLU(BN(Wn-1(...ReLU(BN(W1×FI))))...))))) Where ReLU represents the activation function, BN represents the normalization operation, Wi represents the i-th layer convolution, and n is set to 8.

5. The method for determining ankle joint stability as described in claim 4, characterized in that, An octave convolutional network is used to extract frequency domain features to distinguish between edges and the main body. YH and YL represent high-frequency and low-frequency features, respectively. The calculation formula is as follows: Y H =F(X H ;W H→H )+Upsample(F(X L ;W L→H ),2) Y L =F(X H ;W H→H )+F(pool(X H ,2);W H→L ) Where F(X;W) represents a convolution with parameter W, which varies with the image, pool(x,k) is an average pooling operation with stride k and k×k kernel, and Upsample(F(X;K)) is an upsampling operation performed by the most recent interpolation with parameter K. Let FP represent the high-frequency and low-frequency features extracted by fusion, and the calculation formula is as follows: Resize indicates that the feature is resized to a fixed size. This indicates element-wise addition.

6. The method for determining ankle joint stability as described in claim 1, characterized in that, In step 2, the Transformer-based self-attention modeling module optimizes the feature weights, inputs the dual-branch features into the Transformer encoder, and models the correlation between target pixels through a multi-head attention mechanism, as shown in the formula: Where A represents the attention matrix, X represents the features of the i-th layer, k represents the scaling factor, m = 1, ..., M represents the M-head multi-head attention module, i represents the encoding layer, Q, V, and K are learnable parameters, Contact represents the concatenation operation, and the mapping matrix W and the M-head attention module run in parallel.

7. The method for determining ankle joint stability as described in claim 1, characterized in that, In step 3, the trained network is used to locate the bone edge points in real time. The ResNet is lightweighted and the anchor box fusion detection method is used to quickly filter the background. The image detection rate should be greater than or equal to 24 frames / second. The coordinates of the tibia and fibula edge points are output in real time. The two-stream network structure is based on a two-branch topology structure decoupled from anatomical features. It is divided into a tibia-specific branch and a fibula-specific branch. The tibia-specific branch integrates a lightweight ResNet-18 as the backbone network. The computational complexity is reduced by channel pruning and depthwise separable convolution, while retaining the ability to capture the morphological features of the wide articular surface of the distal tibia. The fibula-specific branch adopts a U-Net variant architecture. The edge response of the narrow fibular suture is enhanced by multi-scale dilated convolution modules. A cross-layer attention gating mechanism is designed to suppress tendon artifact interference. In the decoding stage, a deformable convolutional pyramid (Deformable FPN) is introduced. The two-stream feature maps are dynamically fused by learnable weights to enhance the cross-bone association feature expression of the infratibiofibular syndesmosis. While outputting the coordinates of the bone edge points, a range anchor box with confidence level will be visualized on the image simultaneously, reminding users to perform secondary verification of the model output to ensure surgical safety.

8. The method for determining ankle joint stability as described in claim 1, characterized in that, In step 4, the coordinates of the edge points of the tibia and fibula output in step 3 are received and recorded as the tibial edge point (t). x ,t y ), and the edge point of the fibula (f x .f y Substitute this into the Euclidean distance calculation formula. The minimum distance d between the tibia and fibula is calculated, and a relative exponential stretch ratio is introduced. The gap width in the neutral position is denoted as N, and that in the external rotation position is denoted as E. The stretch ratio is E / N. If the calculated stretch ratio is greater than 1, the joint is considered unstable, and vice versa. An early warning is issued when the calculated ratio is greater than 1, and only the calculation result is output when the ratio is less than 1.

9. A system based on the method for determining ankle joint stability as described in claim 1, characterized in that, include: Image acquisition module, network construction module, bone edge point localization module, and joint stability assessment module; The image acquisition module extracts and denoises the ankle ultrasound video, and marks the edge points of the tibia and fibula; the network construction module generates a high-precision target localization network; the bone edge point localization module locates the bone edge points in real time through the trained network; the joint stability judgment module calculates the bone gaps of the patient's ankle joint in the lateral and neutral positions, and calculates the final ratio. When the gap in the external rotation position is significantly greater than that in the neutral position, that is, the stretch ratio is greater than 1, the joint is considered unstable.

10. The system for determining ankle joint stability as described in claim 9, characterized in that, The network construction modules include a background filtering module, a semantic feature extraction module, a frequency domain feature extraction module, and a self-attention modeling module; the region proposal module performs coarse background screening, the multi-dimensional feature extraction module integrates semantic and frequency domain information, and the Transformer-based self-attention modeling module optimizes feature weights, ultimately generating a high-precision target localization network.

Images