A knee joint localization method based on deep learning

Abstract

The present invention discloses a knee joint positioning method based on deep learning. The invention only needs to input the T1 image of the corresponding knee joint, and can output the corresponding scanning surface parameters to achieve positioning. For the knee joint image, the femur and tibia of the corresponding knee joint are first segmented by Unet; then, based on the segmentation results of the femur and tibia, special points are calculated and positioning lines are drawn; then, based on the type of surface to be positioned, the scanning surface parameters are calculated according to the parameters of the positioning line; finally, the parameters are input into the nuclear magnetic resonance scanning machine to complete the positioning of the knee joint. This method is an end-to-end method that can greatly reduce the workload of doctors. At the same time, experimental results show that this method has good knee joint positioning accuracy.

Description

Technical Field

[0001] The present invention relates to a knee joint method based on deep learning, belonging to the field of computer image processing. Background Art

[0002] The knee joint is not only the largest joint in the human body, but also the most biomechanically complex. It connects the longest bones at both ends and has the largest joint cavity, the largest sesamoid bones (bone), and the largest joint capsule. Accurate knee positioning facilitates efficient diagnosis and treatment of disease. This is a fundamental skill that practitioners should possess, crucial for diagnosis and treatment, yet mastering it fully is challenging. Therapists frequently need to treat this joint in their daily work. Post-traumatic and post-operative knee treatment is a component of nearly every rehabilitation clinic or physical therapy practice.

[0003] Currently, doctors mainly use the "five-finger positioning method" to locate the knee joint. The specific method is as follows: the doctor stands on the patient's affected side and uses the ipsilateral hand to make five fingers to position. If the lesion is in the right knee joint, the doctor uses the right hand to locate it, and if the lesion is in the left knee joint, the doctor uses the left hand to locate it. The palm faces the center of the patella, and the five fingers are spread as far as possible, with the fingers semi-flexed. The middle finger faces the middle part of the patellar ligament, the index finger and ring finger correspond to the inner and outer knee sockets respectively, the big toe faces the starting point of the tibial collateral ligament and the lower part of the vastus medialis muscle, the little finger faces the meridian line of the iliotibial band, and the base of the palm faces the suprapatellar bursa. In addition, 4 cm below the index finger and 3 cm inward is the insertion point of the anserine bursa.

[0004] Manual positioning requires extensive experience and is not very accurate. The present invention can help doctors perform end-to-end knee joint positioning. By simply inputting the T1 image of the knee joint, the scanning surface parameters can be output to assist doctors in knee joint treatment. Summary of the Invention

[0005] In response to the problems existing in manual positioning methods, the present invention proposes a knee joint positioning method based on deep learning. After locating the knee joint, it can reduce the workload of doctors and improve the efficiency of knee joint scanning.

[0006] In order to achieve the above objectives, the technical solution of the present invention is as follows: The present invention provides a knee joint positioning method based on deep learning, the positioning range is as follows: Figure 1 As shown in the black and white box in the figure, the specific steps are as follows:

[0007] The specific steps are divided into two parts: image segmentation and knee joint positioning.

[0008] Step 1: The specific steps of image segmentation are as follows Figure 2 As shown:

[0009] Step 1-1, obtain training sample data, which specifically consists of 48 knee joint T1 coronal images, 48 ​​knee joint T1 sagittal images, and 48 knee joint T1 cross-sectional images provided by Xingaoyi Company.

[0010] Steps 1-2: Preprocess the training data. The preprocessing steps are as follows: ① Manually clean the dataset. For the cross-section, extract the required femur labels. For the coronal and sagittal planes, extract the required femur and tibia labels. ② Resample the data to (512, 512) and normalize it to obtain the target dataset. The resampling of the knee joint image uses the bilinear interpolation algorithm, and the resampling of the label uses the nearest neighbor interpolation algorithm. The normalization formula is:

[0011]

[0012] Among them, Mean and Std represent the mean and standard deviation of the image grayscale, X normalization is the normalized pixel grayscale value.

[0013] Steps 1-3, train the Unet model, such as Figure 4 As shown, the left side of Unet is the feature extraction network (encoder), and the right side is the feature fusion network (decoder), including the downsampling path and the upsampling path. The jump connection in the middle is used to transmit the detailed information of the image. In the knee joint segmentation task, the foreground is the femur and tibia we need, and the background is the pixels other than these. During the training process, we used data enhancement methods such as random addition of Gaussian blur and random translation to prevent network overfitting and enhance the generalization ability of the network. The loss function is the sum of the binary cross entropy and the Dice coefficient, and the optimizer chooses stochastic gradient descent. At the same time, due to the different properties of the three sections, it is necessary to train the coronal plane, sagittal plane and transverse plane separately to obtain the corresponding model parameters.

[0014] a. Data enhancement uses the random addition of Gaussian blur to enable the network to correctly segment some relatively blurry images. Let σ 2 is the variance of the image grayscale value, then the formula is as follows:

[0015]

[0016] where σ 2 is the variance of the image grayscale value, x, y correspond to the x-axis coordinate and y-axis coordinate of the image respectively, and g(x, y) is the grayscale value of the image after Gaussian blur.

[0017] b. Data augmentation uses random translation because sometimes the spines of different people are misaligned, and the network needs to learn the corresponding misalignment. The formula is as follows:

[0018] g(x,y)=g(xi,yj)

[0019] Where i and j are the horizontal and vertical translation distances respectively, x and y correspond to the x-axis and y-axis coordinates of the image respectively, and g(x,y) represents the grayscale value at (x,y) in the image.

[0020] c. The optimization method uses SGD stochastic gradient descent. Stochastic gradient descent randomly selects a data point to calculate in the direction of fastest descent, rather than scanning the entire training data set, which speeds up the iteration speed. Its formula is:

[0021] θ j+1 =θ j +(y i -h θ (x i ))x i

[0022] Among them, θ j Represents the parameters of the network at the jth iteration, y i Indicates the label of the i-th image, h θ (x i ) represents the prediction result of the current network, x i Represents the i-th knee joint image.

[0023] d. The loss uses the binary Dice loss function, where X is the label of the input image and Y is the predicted segmentation result. It is calculated as follows:

[0024]

[0025]

[0026] Among them, X represents the label of the knee joint, Y represents the predicted value of the current network, N represents the size of batch_size in this experiment, and y i Represents the label value of the i-th image, p i Represents the predicted value of the i-th image.

[0027] Steps 1-4: predict the knee joint image that needs to be positioned.

[0028] Step 2: Specific steps for knee joint positioning Figure 3 As shown:

[0029] Step 2-1, take the maximum connected graph of the segmentation graph. Specifically, when segmenting the cross section, take the maximum connected graph; when segmenting the coronal plane and the sagittal plane, take the two largest connected graphs.

[0030] Step 2-2, calculate the position of the positioning line, such as Figure 5 , the calculation steps are as follows:

[0031] a. For the positioning of cross sections, such as Figure 1 As shown in (a), the sagittal and coronal positioning lines are parallel to the black positioning line from top to bottom. The positioning range is within the box, covering the entire knee joint. We need to find two special points at the lower left and right corners of the femur and two special points at the upper left and right corners of the tibia. Using the idea of ​​linear programming, first, establish a plane rectangular coordinate system with the center of gravity of the segmented femur as the origin, that is, to find:

[0032] (x1,y1)=argmin(a1y-b1x)

[0033] (x²,y²)=argmin(a²y+b²x)

[0034] Next, find the two special points at the upper left and upper right corners of the tibia, that is, find:

[0035] (x3,y3)=argmax(a3y-b3x)

[0036] (x4,y4)=argmax(a4y-b4x)

[0037] Where (x i ,y i ) represents the coordinates of a point; (a i ,b i ) represents a set of hyperparameters greater than 0 that can be adjusted based on prior information about femoral and tibial deflection; argmin(·) and argmax(·) represent the values ​​of the variable at the maximum and minimum of the objective function in the parentheses. After obtaining these four vertices, the corresponding upper and lower vertices are connected (i.e., the special points at the lower left corner of the femur and the upper left corner of the tibia are connected, and the special points at the lower right corner of the femur and the upper right corner of the tibia are connected). The center points of these two connecting lines are connected to obtain the central positioning line. Next, the other positioning lines are obtained by translating them up and down according to the predetermined standard positioning line distance.

[0038] b. For sagittal plane positioning, such as Figure 1 As shown in (b), we need to draw the positioning line for the transverse position and the coronal position. For the positioning line for the transverse position, we only need to take two special points at the lower left and lower right corners of the femur, connect these two special points, and translate downward until they are tangent to the femur. At this time, we get the tangent line at the bottom of the femur. That is,

[0039] (x1,y1)=argmin(a1y-b1x)

[0040] (x²,y²)=argmin(a²y+b²x)

[0041] Then draw its perpendicular line, and translate it left and right according to the predetermined standard positioning line distance to obtain other positioning lines.

[0042] For the coronal positioning line, the positioning line is parallel to the black positioning line from top to bottom, and the positioning range is within the box, covering the entire knee joint. The same linear programming concept is used. First, the center of gravity of the segmented femur is used as the origin to establish a plane rectangular coordinate system, that is, to find:

[0043] (x1,y1)=argmin(a1y-b1x)

[0044] (x²,y²)=argmin(a²y+b²x)

[0045] Next, find the two special points at the upper left and upper right corners of the tibia, that is, find:

[0046] (x3,y3)=argmax(a3y-b3x)

[0047] (x4,y4)=argmax(a4y-b4x)

[0048] After obtaining these four vertices, connect the corresponding upper and lower vertices (i.e., connect the special points at the lower left corner of the femur and the upper left corner of the tibia, and connect the special points at the lower right corner of the femur and the upper right corner of the tibia). Connect the center points of these two connecting lines to obtain the central positioning line. Then, draw a perpendicular line to it and translate it left and right according to the predetermined standard positioning line distance to obtain the other positioning lines.

[0049] c. For coronal positioning, such as Figure 1 As shown in (c), we need to draw the positioning lines for the transverse and sagittal planes. For the positioning lines for the transverse plane, we only need to take two special points, the lower left and lower right corners of the femur, and connect these two special points to obtain the corresponding positioning lines, that is, to find:

[0050] (x1,y1)=argmin(a1y-b1x)

[0051] (x²,y²)=argmin(a²y+b²x)

[0052] Then, according to the predetermined standard positioning line distance, other positioning lines can be obtained by moving up and down respectively.

[0053] For the sagittal positioning line, the positioning line is parallel to the black positioning line from top to bottom. The positioning range is within the box, covering the entire knee joint. The idea of ​​linear programming can be used. First, the center of gravity of the segmented femur is used as the origin to establish a plane rectangular coordinate system, that is, to find:

[0054] (x1,y1)=argmin(a1y-b1x)

[0055] (x²,y²)=argmin(a²y+b²x)

[0056] Next, find the two special points at the upper left and upper right corners of the tibia, that is, find:

[0057] (x3,y3)=argmax(a3y-b3x)

[0058] (x4,y4)=argmax(a4y-b4x)

[0059] After obtaining these four vertices, connect the corresponding upper and lower vertices (i.e., connect the special points at the lower left corner of the femur and the upper left corner of the tibia, and connect the special points at the lower right corner of the femur and the upper right corner of the tibia). Connect the center points of these two connecting lines to obtain the central positioning line. Then, draw a perpendicular line to it and translate it left and right according to the predetermined standard positioning line distance to obtain the other positioning lines.

[0060] Step 2-3: Calculate the parameters of the knee joint scanning surface. The specific steps are:

[0061] a) When performing cross-sectional positioning, obtain two positioning lines of the coronal and sagittal planes. Take the midpoint P of the sagittal positioning line and the two third points Q and R on the coronal positioning line. Calculate the world coordinates of P, Q, and R based on the Origin, Orientation, and Spacing parameters in the DICOM header file. Calculate the plane formed by these three points and record it as Ax+By+Cz+D=0. This plane intersects the central axis of the coronal plane at point O, and record the normal vector of this plane as The parameter calculation formula of the target scanning surface is as follows:

[0062]

[0063] Among them, NORM() function represents vector normalization, LT is the coordinate of the upper left corner of the target scanning surface of the nuclear magnetic resonance, FOV is the side length of the target scanning surface, is the unit vector in the direction of the horizontal axis slope, is the unit vector in the direction of the vertical axis slope. These parameters can uniquely determine the target scanning surface.

[0064] b) When performing coronal plane positioning, obtain two positioning lines for the transverse and sagittal planes. Take the midpoint P of the sagittal plane positioning line and the two third points Q and R on the transverse plane positioning line. Calculate the world coordinates of P, Q, and R based on the Origin, Orientation, and Spacing parameters in the DICOM header file. Calculate the plane formed by these three points and record it as Ax+By+Cz+D=0. This plane intersects the midline of the sagittal plane at point O, and record the normal vector of this plane as The parameter calculation formula of the target scanning surface is as follows:

[0065]

[0066] FOV=FOV

[0067] Among them, NORM() function represents vector normalization, LT is the coordinate of the upper left corner of the target scanning surface of the nuclear magnetic resonance, FOV is the side length of the target scanning surface, is the unit vector in the direction of the horizontal axis slope, is the unit vector in the direction of the vertical axis slope. These parameters can uniquely determine the target scanning surface.

[0068] c) When performing sagittal plane positioning, obtain two positioning lines for the transverse and coronal planes. Take the midpoint P of the coronal plane positioning line and the two third points Q and R on the transverse plane positioning line. Calculate the world coordinates of P, Q, and R based on the Origin, Orientation, and Spacing parameters in the DICOM header file. Calculate the plane formed by these three points and record it as Ax+By+Cz+D=0. This plane intersects the midline of the transverse plane at point O, and record the normal vector of this plane as The parameter calculation formula of the target scanning surface is as follows:

[0069]

[0070]

[0071] FOV=FOV

[0072] Among them, NORM() function represents vector normalization, LT is the coordinate of the upper left corner of the target scanning surface of the nuclear magnetic resonance, FOV is the side length of the target scanning surface, is the unit vector in the direction of the horizontal axis slope, is the unit vector in the direction of the vertical axis slope. These parameters can uniquely determine the target scanning surface.

[0073] Step 2-4: Input the scanning surface parameters into the machine to complete the knee joint positioning

[0074] Compared with the existing technology, the advantages of the present invention are as follows: (1) Based on the deep learning method, the segmentation results are very accurate; (2) Due to the use of corresponding data enhancement methods, it can cope with various situations such as image blur and offset; (3) The amount of data required is not large, and it is a lightweight network; (4) The present invention is an automated, end-to-end method, such as Figure 6 As shown, after integrating this method into software, the doctor only needs to click to import the DICOM sequence and wait for a moment to obtain the knee joint positioning result. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 For example,

[0076] Figure 2 Schematic diagram of the image segmentation process.

[0077] Figure 3 This is a flowchart of knee joint positioning.

[0078] Figure 4 This is a schematic diagram of the Unet network structure used in the present invention.

[0079] Figure 5 Schematic diagram of special points in sagittal, coronal and transverse planes.

[0080] Figure 6 This is the software running result diagram. DETAILED DESCRIPTION

[0081] The present invention is further illustrated below with reference to specific examples. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0082] Example: See Figure 2 The present invention provides a knee joint positioning method based on deep learning, the positioning range is as follows Figure 1 As shown in the black and white box in the figure, the specific steps are as follows:

[0083] The specific steps are divided into two parts: image segmentation and knee joint positioning:

[0084] Step 1: The specific steps of image segmentation are as follows Figure 2 As shown:

[0085] Step 1-1, acquiring data, that is, scanning the coronal plane, sagittal plane or transverse plane of the knee joint as a positioning image for knee joint scanning.

[0086] In steps 1-2, the training data is preprocessed, including data normalization, and the image is resized to (512, 512) using a bilinear interpolation algorithm, and the label is resized to (512, 512) using a nearest neighbor interpolation algorithm.

[0087] Steps 1-3, train the Unet model, the network structure used is as follows Figure 4 At the same time, due to the different properties of the three sections, it is necessary to train the coronal plane, sagittal plane and transverse plane separately to obtain the corresponding model parameters.

[0088] In steps 1-4, according to the specific slice type (cross section / coronal plane / sagittal plane), the pre-processed positioning image is input into the corresponding trained Unet segmentation network to obtain the segmentation map of the femur and tibia.

[0089] Step 2: Specific steps for knee joint positioning Figure 3 As shown:

[0090] Step 2-1, take the connected graph, that is: when splitting the cross section, take the maximum connected graph; when splitting the coronal plane and the sagittal plane, take the two largest connected graphs.

[0091] Step 2-2, calculating the position of the positioning line according to the specific slice type to be positioned;

[0092] Step 2-3, calculating the parameters of the magnetic resonance scanning surface according to the positioning line obtained in step 2-2.

[0093] a) When performing cross-sectional positioning, obtain two positioning lines of the coronal and sagittal planes. Take the midpoint P of the sagittal positioning line and the two third points Q and R on the coronal positioning line. Calculate the world coordinates of P, Q, and R based on the Origin, Orientation, and Spacing parameters in the DICOM header file. Calculate the plane formed by these three points and record it as Ax+By+Cz+D=0. This plane intersects the central axis of the coronal plane at point O, and record the normal vector of this plane as The coordinate of the upper left corner of the target scanning surface of the MRI is LT, the side length of the target scanning surface is FOV, and the unit vector in the direction of the horizontal axis slope is The unit vector in the direction of the slope of the vertical axis is The parameter calculation formula of the target scanning surface is as follows:

[0094]

[0095] FOV=FOV

[0096] Among them, the NORM() function represents vector normalization.

[0097] b) When performing coronal plane positioning, obtain two positioning lines for the transverse and sagittal planes. Take the midpoint P of the sagittal plane positioning line and the two third points Q and R on the transverse plane positioning line. Calculate the world coordinates of P, Q, and R based on the Origin, Orientation, and Spacing parameters in the DICOM header file. Calculate the plane formed by these three points and record it as Ax+By+Cz+D=0. This plane intersects the midline of the sagittal plane at point O, and record the normal vector of this plane as The coordinate of the upper left corner of the target scanning surface of the MRI is LT, the side length of the target scanning surface is FOV, and the unit vector in the direction of the horizontal axis slope is The unit vector in the direction of the slope of the vertical axis is The parameter calculation formula of the target scanning surface is as follows:

[0098]

[0099]

[0100] FOV=FOV

[0101] Among them, the NORM() function represents vector normalization.

[0102] c) When performing sagittal plane positioning, obtain two positioning lines for the transverse and coronal planes. Take the midpoint P of the coronal plane positioning line and the two third points Q and R on the transverse plane positioning line. Calculate the world coordinates of P, Q, and R based on the Origin, Orientation, and Spacing parameters in the DICOM header file. Calculate the plane formed by these three points and record it as Ax+By+Cz+D=0. This plane intersects the midline of the transverse plane at point O, and record the normal vector of this plane as The coordinate of the upper left corner of the target scanning surface of the MRI is LT, the side length of the target scanning surface is FOV, and the unit vector in the direction of the horizontal axis slope is The unit vector in the direction of the slope of the vertical axis is The parameter calculation formula of the target scanning surface is as follows:

[0103]

[0104] Among them, the NORM() function represents vector normalization.

[0105] Steps 2-4: Input the scanning surface parameters into the MRI scanner to complete the knee joint positioning.

[0106] Effect evaluation:

[0107] This paper proposes a knee joint method based on deep learning. The test results of this method are as follows: Figure 6 As shown in the figure, the present invention is integrated into a software. When clicking the "Load Image" button during the test, the software will automatically read the image. Clicking "Calculate Positioning Line" will cause the software to segment the image and calculate the positioning line. Finally, clicking "Save ini" will save the corresponding positioning results and scanning surface parameters to the ini file. The test results show that this method can quickly and accurately locate the knee joint, providing doctors with an efficient solution for knee joint treatment.

[0108] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person familiar with the technology can understand and think of any changes or replacements within the technical scope disclosed by the present invention, which should be included in the scope of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A knee joint positioning method based on deep learning, characterized in that: The specific steps are as follows: Step 1: Image segmentation, as follows: Step 1-1, obtaining training sample data, that is, scanning the coronal plane, sagittal plane and transverse section of the knee joint as positioning images for knee joint scanning; Step 1-2: Preprocess the training data, including data normalization, and use the bilinear interpolation algorithm to resample the image to a size of (512, 512). The label is resampled to a size of (512, 512) using the nearest neighbor interpolation algorithm. Steps 1-3: Train the Unet model. It is necessary to train the coronal plane, sagittal plane, and transverse plane separately to obtain the corresponding model parameters. Steps 1-4, based on the specific slice type, i.e., transverse, coronal, and sagittal planes, use the corresponding model parameters to predict the knee joint to be located and obtain the segmentation map. Step 2: Position the knee joint as follows: Step 2-1: Take the largest connected graph for the segmented graph: when segmenting the cross section, take the largest connected graph, i.e., the femur; when segmenting the coronal and sagittal planes, take the two largest connected graphs, i.e., the femur and tibia; Step 2-2, calculate the position of the positioning line according to the type of plane to be positioned, i.e., transverse plane, coronal plane, or sagittal plane; Step 2-3, calculate the parameters of the knee joint scanning surface, Steps 2-4: Input the scan surface parameters into the MRI machine to complete knee joint positioning; The parameters of the knee joint scanning surface calculated in steps 2-3 are as follows: a) When performing cross-sectional positioning, obtain two positioning lines for the coronal and sagittal planes, take the midpoint P of the sagittal positioning line, and take the two tripoints Q and R on the coronal positioning line; calculate the world coordinates of P, Q, and R based on the image parameters in the DICOM header file, namely the coordinate Origin of the upper left corner of the image, the orientation vectors of the X and Y axes, and the distance Spacing between unit pixels, and calculate the plane formed by these three points, denoted as Ax+By+Cz+D=0. This plane intersects the central axis of the coronal plane at point O, and the normal vector of this plane is denoted as The parameter calculation formula of the target scanning surface is as follows: FOV=FOV Among them, NORM() function represents vector normalization, LT is the coordinate of the upper left corner of the target scanning surface of the nuclear magnetic resonance, FOV is the side length of the target scanning surface, is the unit vector in the direction of the horizontal axis slope, is the unit vector in the direction of the slope of the vertical axis, b) When performing coronal plane positioning, obtain two positioning lines for the transverse and sagittal planes. Take the midpoint P of the sagittal plane positioning line and the two third points Q and R on the transverse plane positioning line. Based on the image parameters in the DICOM header file, namely the coordinate Origin of the upper left corner of the image, the orientation vectors of the X and Y axes, and the distance Spacing between unit pixels, the world coordinates of P, Q, and R are calculated. The plane formed by these three points is calculated and recorded as Ax+By+Cz+D=0. This plane intersects the midline of the sagittal plane at point O, and the normal vector of this plane is recorded as The parameter calculation formula of the target scanning surface is as follows: FOV=FOV Among them, NORM() function represents vector normalization, LT is the coordinate of the upper left corner of the target scanning surface of the nuclear magnetic resonance, FOV is the side length of the target scanning surface, is the unit vector in the direction of the horizontal axis slope, is the unit vector in the direction of the slope of the vertical axis, c) When performing sagittal plane positioning, obtain two positioning lines for the transverse and coronal planes. Take the midpoint P of the coronal plane positioning line and the two third points Q and R on the transverse plane positioning line. Based on the image parameters in the DICOM header file, namely the coordinate Origin of the upper left corner of the image, the direction vectors Orientation of the X and Y axes, and the distance Spacing between unit pixels, calculate the world coordinates of P, Q, and R. Calculate the plane formed by these three points and record it as Ax+By+Cz+D=0. This plane intersects the midline of the transverse plane at point O, and record the normal vector of this plane as The parameter calculation formula of the target scanning surface is as follows: FOV=FOV Among them, NORM() function represents vector normalization, LT is the coordinate of the upper left corner of the target scanning surface of the nuclear magnetic resonance, FOV is the side length of the target scanning surface, is the unit vector in the direction of the horizontal axis slope, is the unit vector in the direction of the slope of the vertical axis.

2. A knee joint positioning method based on deep learning according to claim 1, characterized in that: The training sample data in step 1-1 is 48 T1 coronal images of the knee joint, 48 T1 sagittal images of the knee joint, and 48 T1 transverse images of the knee joint. The positions of the femur and tibia are manually marked on each knee joint.

3. The deep learning knee joint positioning method according to claim 1, characterized in that: In steps 1-2, the training samples are preprocessed. The preprocessing steps are as follows: ① manually clean the data set to extract the required knee joints, including the femur and tibia; ② normalize the cleaned data to obtain the target data set.

4. The deep learning knee joint positioning method according to claim 1, characterized in that: In steps 1-3, the Unet model is trained using image enhancement methods such as random Gaussian blur, random Gaussian noise, contrast enhancement, gamma transform, and translation. The loss function is set to the sum of binary cross entropy and Dice coefficient, and the optimizer uses stochastic gradient descent (SGD) to update the network parameters.

5. The deep learning knee joint positioning method according to claim 1, characterized in that: In step 2-1, the largest connected graph is obtained as follows: first find all connected graphs in the segmented graph, calculate the area of ​​each connected graph, find the connected graphs with the largest area, and set the other connected graphs as the background.

6. The deep learning knee joint positioning method according to claim 1, characterized in that: The calculation of the positioning line position mentioned in step 2-2 includes the following components: calculating the center of gravity of the femur and tibia; finding two special points at the lower left and lower right corners of the femur and four special points at the upper left and upper right corners of the tibia through linear programming; drawing positioning lines based on these four special points; and finally drawing the remaining positioning lines based on the positioning line distance parameters.

Images