Femoral head region bone feature segmentation method using SAM-Med3d

Abstract

The invention discloses a femoral head region bone feature segmentation method using SAM-Med3d, and belongs to the field of medical image segmentation optimization. According to a preoperative three-dimensional CT image of an SCFE patient, disease test CT is obtained, data labeling of thighbone, femoral head epiphyseal edge, femoral head and subchondral bone is carried out on part of disease test CT, a labeling result is obtained, data resampling and data scale adjustment are carried out according to the space proportion of a target area, a resampling image of each labeling label is obtained, and the resampling image of each labeling label is obtained. The method comprises the following steps of: putting the four tags and a resampling image of an original image corresponding to the four tags into SAM-Med3d for training to obtain segmentation results of the four tags, performing resampling to restore to an original scale, and combining the four tags together to obtain a final bone feature segmentation result. According to the method, the femur, the femoral head epiphyseal edge, the femoral head and the subchondral bone area in the preoperative three-dimensional CT data of the SCFE patient can be automatically segmented, and compared with original SAM-Med3d, the segmentation accuracy can be improved, and the method has important significance for subsequent operation scheme design.

Description

Technical Field

[0001] The present invention relates to a method for segmenting bone features in the femoral head region using SAM-Med3d, belonging to the field of optimization of medical image segmentation.

Background Art

[0002] With the continuous progress of medical imaging technology, three-dimensional CT data plays an increasingly important role in the diagnosis and treatment of orthopedic diseases. For patients with slipped capital femoral epiphysis (SCFE), preoperative three-dimensional CT data not only provides an intuitive basis for the early diagnosis of the disease, but also provides key structural information for surgical planning and prognosis assessment. However, the bone structure in the femoral region, especially in the femoral head region, is complex, involving multiple structures such as the femoral shaft, the epiphyseal margin of the femoral head, the femoral head body, and the subchondral bone. Traditional manual or semi-automatic segmentation methods are often time-consuming and laborious, and are easily affected by the subjective factors of operators, resulting in large fluctuations in the consistency and accuracy of the segmentation results. Therefore, how to overcome the deficiencies of traditional segmentation techniques in dealing with complex bone structures and improve the accuracy and efficiency of segmentation is an urgent problem to be solved in this field.

Summary of the Invention

[0003] In view of this, the present invention provides a method for segmenting bone features in the femoral head region using SAM-Med3d to achieve automatic segmentation of the femur, the epiphyseal margin of the femoral head, the femoral head, and the subchondral bone region in the preoperative three-dimensional CT data of SCFE patients undergoing surgery.

[0004] An embodiment of the present invention provides a method for segmenting bone features in the femoral head region using SAM-Med3d, including:

[0005] According to the original preoperative three-dimensional CT images of SCFE patients, each image is divided into left and right sides along the median sagittal plane, and the image of the diseased side is retained to obtain the CT of the diseased side of SCFE patients. And data annotation of the femur, the epiphyseal margin of the femoral head, the femoral head, and the subchondral bone is performed on a part of the CT of the diseased side to obtain an annotation result;

[0006] According to the original preoperative three-dimensional CT images of SCFE patients and the annotation results of each image, resampling of the data is performed according to the spatial proportion of the target region to adjust the data scale, and resampled images of each annotation label and resampled images of the corresponding CT images are obtained;

[0007] The resampled images of each annotation label and the resampled images of the corresponding original images are put into SAM-Med3d for training using pre-trained parameters to obtain fine-tuned parameters capable of segmenting the femur, the epiphyseal margin of the femoral head, the femoral head, and the subchondral bone;

[0008] Based on the parameters obtained after training and the pre-operative CT scans of all SCFE patients, the segmentation results of the femur, femoral head physeal margin, femoral head, and subchondral bone of all SCFE patients are obtained. The segmentation results are resampled, and after restoring to the original scale, the four labels (femur, femoral head physeal margin, femoral head, and subchondral bone) are merged together to obtain the final bone feature segmentation result.

[0009] In the above method, based on the original pre-operative three-dimensional CT images of SCFE patients, each image is divided into left and right sides along the median sagittal plane, and the image of the affected side is retained to obtain the CT scans of the affected sides of SCFE patients. Data annotation of the femur, femoral head physeal margin, femoral head, and subchondral bone is performed on some of the CT scans of the affected sides to obtain the annotation results, including

[0010] The DICOM format CT images are converted to NIFTI format using ITK-SNAP software, and the images are divided into left and right sides along the median sagittal plane. The images of bilateral slip patients are split into two cases of data, and the CT scans of the affected sides of unilateral slip patients are used as one case of data. Under the guidance of professional orthopedic doctors, the femur, femoral head physeal margin, femoral head, and subchondral bone on the CT images are labeled with different tags.

[0011] In the above method, based on the original pre-operative three-dimensional CT images of SCFE patients and the annotation results of each image, data resampling is performed according to the spatial proportion of the target region to adjust the data scale, and resampled images of each annotation label and the corresponding resampled images of the CT images are obtained, including

[0012] Based on the original pre-operative three-dimensional CT images of SCFE patients and the annotation results of each image, for target regions with a small spatial proportion and high requirements for segmentation accuracy (subchondral bone, femoral head physeal margin, and femoral head), a resampling method of reducing the voxel spacing of the original image is adopted to improve the spatial resolution of the image and enhance the model's perception ability for small feature regions. For larger regions (femur), a resampling method of enlarging the voxel spacing of the original image is adopted to reduce the computational complexity and retain the overall structural information, optimizing the segmentation accuracy of the neural network for target regions with significant scale differences.

[0013] For the input image I(x), x ∈ R 3 representing the three-dimensional coordinate points in space, the resampling operation R of the target region is set Δs as a scale transformation on the spatial coordinate x, where Δs is the target scale factor, representing the voxel spacing after scale transformation. Through the variable-scale resampling operation, the new image I′(x) is obtained as:

[0014]

[0015] where soriginal and s target represent the voxel spacings of the original image and the target image, respectively;

[0016] Let V(I) be the volume of a certain label region in the image, and Ω(I) be the total volume of the image input into the network. The volume ratio α of the label region can be defined as

[0017]

[0018] For a hollow spherical region, calculate the ratio of the convex hull volume S(I) of this region to the total volume Ω(I) of the image to reflect the size of the label region. Let the boundary region size ratio β be

[0019]

[0020] Combining the actual voxel ratio and the boundary region size ratio, define the comprehensive metric γ as follows:

[0021] γ = λ1α + λ2β

[0022] where λ1 and λ2 are the weight coefficients of the features;

[0023] The resampling strategy can be achieved through the decision condition of the following formula:

[0024]

[0025] where γ min_threshold and γ max_threshold are preset thresholds. The label regions within the interval [γ min_threshold , γ max_threshold do not need to be resampled. The label regions smaller than this threshold are sampled using the resolution s high , and vice versa, using the resolution s low for sampling to obtain the resampled original image and the annotated image.

[0026] In the above method, the resampled images of each annotated label and the resampled images of their corresponding original images are put into SAM - Med3d for training using the pre - trained parameters, and the fine - tuned parameters that can segment the femur, femoral head epiphyseal edge, femoral head, and subchondral bone are obtained, including

[0027] Taking the resampled images of each annotated label and the resampled images of their corresponding original images as a group, put them into SAM - Med3d for training using the pre - trained parameter file sam_med3d_turbo.pth. Train for the four labels of the femur, femoral head epiphyseal edge, femoral head, and subchondral bone to obtain the fine - tuned parameter file that can segment the femur, femoral head epiphyseal edge, femoral head, and subchondral bone.

[0028] In the above method, according to the parameters obtained after training and the CT scans of all SCFE patients, the segmentation results of the femur, the epiphyseal margin of the femoral head, the femoral head, and the subchondral bone of all SCFE patients are obtained. The segmentation results are resampled, and after restoring to the original scale, the four labels (femur, epiphyseal margin of the femoral head, femoral head, and subchondral bone) are merged together to obtain the final bone feature segmentation result, including

[0029] According to the parameter file obtained after training and the CT scans of all SCFE patients, the segmentation results of the femur, the epiphyseal margin of the femoral head, the femoral head, and the subchondral bone of all SCFE patients are obtained. The segmentation results are resampled. For the input image I(x), x ∈ R 3 represents the three-dimensional coordinate points in space, and the resampling operation R 1 / Δs of the target region is to perform a scale transformation on the spatial coordinate x, where Δs is the target scale factor calculated above. Through the variable-scale resampling operation, a new image I″(x) can be obtained as follows:

[0030] I″(x) = R 1 / Δs (I(x))

[0031] After restoring to the original scale, the four labels are merged together. First, record the coordinates of the label region in the annotation file with the label "epiphyseal margin of the femoral head", and modify the coordinates in the annotation file with the label "femoral head" to the label "epiphyseal margin of the femoral head", and record this file as "File A"; then record the coordinates of the label region in the annotation file with the label "subchondral bone", and modify these coordinates in "File A" to the label "subchondral bone", and record this file as "File B"; finally, record the coordinates of the label region in the annotation file with the label "femur", and modify these coordinates in "File B" to the label "femur" to obtain the final bone feature segmentation result.

[0032] The technical solution provided by the embodiments of the present invention can automatically segment the femur, the epiphyseal margin of the femoral head, the femoral head, and the subchondral bone regions on the basis of the preoperative three-dimensional CT images of SCFE patients, which is of great significance for the subsequent surgical plan design.

BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, without creative and laborious efforts, other drawings can also be obtained based on these drawings.

[0034] Figure 1It is a schematic flowchart of a method for segmenting bone features in the femoral head region using SAM-Med3d provided by an embodiment of the present invention;

[0035] Figure 2 It is a data annotation diagram of an embodiment of the present invention;

[0036] Figure 3 It is a segmentation effect diagram of an embodiment of the present invention;

[0037] Figure 4 It is the segmented result of the combined bone features and its three-dimensional display diagram of an embodiment of the present invention;

Specific Embodiment

[0038] To better understand the technical solution of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0039] It should be clear that the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0040] An embodiment of the present invention provides a method for segmenting bone features in the femoral head region using SAM-Med3d. Please refer to Figure 1 , which is a schematic flowchart of a method for segmenting bone features in the femoral head region using SAM-Med3d provided by an example of the present invention. As Figure 1 shown, the method includes the following steps:

[0041] Step 101: According to the original preoperative three-dimensional CT images of SCFE patients, each image is divided into left and right sides along the median sagittal plane, and the image of the affected side is retained to obtain the CT of the affected side of SCFE patients. Then, data annotation is performed on the femur, femoral head epiphyseal margin, femoral head, and subchondral bone of part of the CT of the affected side to obtain the annotation result.

[0042] The present invention uses ITK-SNAP software to convert DICOM format CT images into NIFTI format, and divides the images into left and right sides along the median sagittal plane, retaining the image of the affected side. That is, the images of bilateral slipped patients are split into two cases of data, and the CT of the affected side of unilateral slipped patients is used as one case of data. Under the guidance of professional orthopedic doctors, the present invention labels the femur, femoral head epiphyseal margin, femoral head, and subchondral bone on the CT image with different tags. As Figure 2 shown, where (a-c) are the displays on the transverse, sagittal, and coronal planes respectively, and (d) is the three-dimensional display.

[0043] Step 102: According to the original preoperative three-dimensional CT images of SCFE patients and the annotation results of each image, resample the data based on the spatial proportion of the target area, adjust the data scale, and obtain the resampled images of each annotation label and the resampled images of their corresponding original images.

[0044] According to the original preoperative three-dimensional CT images of SCFE patients and the annotation results of each image, for target areas with a small spatial proportion and high requirements for segmentation accuracy (such as small areas like subchondral bone, femoral head epiphyseal margin, and femoral head), the present invention adopts a high-precision resampling method, that is, reducing the voxel spacing of the original image, thereby improving the spatial resolution of the image and enhancing the model's perception ability for small feature areas; for larger areas (such as femoral labels), a lower voxel spacing is used to reduce the computational complexity and retain the overall structural information, optimizing the segmentation accuracy of the neural network for target areas with significant scale differences.

[0045] For the input image I(x), x ∈ R 3 represents the three-dimensional coordinate points in space; to optimize the feature representation at different scales, the resampling operation R of the target area is set Δs to perform a scale transformation on the spatial coordinate x, where Δs is the target scale factor, representing the new voxel spacing; through the variable-scale resampling operation, the present invention can obtain the new image I′(x) as follows:

[0046] I′(x) = R Δs (I(x)),

[0047] where s original and s target respectively represent the voxel spacings of the original image and the target image;

[0048] ① Actual voxel proportion

[0049] For each label area, the present invention selects an appropriate target voxel spacing s according to its spatial proportion and importance in the image target ; set V(I) as the volume of a certain label area in the image, Ω(I) as the total volume of the image input to the network, and the volume proportion α of the label area can be defined as

[0050]

[0051] ② Proportion of boundary area size

[0052] For complex-shaped regions (such as hollow spherical regions), in addition to volume calculation, the boundary size of the region is also an important indicator to measure the "scope" of its space occupation; the present invention calculates the ratio of the convex hull volume S(I) of the region to the total volume of the Ω(I) image to reflect the size of the labeled region, and the size ratio β of the labeled region can be defined as

[0053]

[0054] Taking into account multiple features such as volume, boundary size, and convex hull volume, a comprehensive metric γ is defined as follows:

[0055] γ = λ1α + λ2β

[0056] where λ1 and λ2 are the weight coefficients of the features;

[0057] For labeled regions with a relatively small proportion (such as subchondral bone, femoral head epiphyseal margin, etc.), the present invention improves the ability to capture details by increasing the spatial resolution of the image (reducing the voxel spacing); conversely, for labeled regions with a relatively large proportion (such as femoral labels), the present invention uses a lower resolution to avoid wasting computing resources; the specific resampling strategy can be achieved through the decision condition of the following formula:

[0058]

[0059] where γ min_threshold and γ max_threshold are preset thresholds. Labeled regions within the interval [γ min_threshold , γ max_threshold do not require resampling. Labeled regions smaller than this threshold are sampled using a higher resolution s high , and conversely, a lower resolution s low is used for sampling. Different resamplings are performed for different labels to obtain the resampled original image and labeled image.

[0060] Step 103: Put the resampled image of each labeled label and the resampled image of its corresponding original image into SAM-Med3d for training using the pre-trained parameter file sam_med3d_turbo.pth to obtain a fine-tuned parameter file that can segment the femur, femoral head epiphyseal margin, femoral head, and subchondral bone.

[0061] Take the resampled image of each labeled label and the resampled image of its corresponding original image as a group, put them into SAM-Med3d for training using the pre-trained parameter file sam_med3d_turbo.pth, and train for four labels: the femur, femoral head epiphyseal margin, femoral head, and subchondral bone, to obtain a fine-tuned parameter file that can segment the femur, femoral head epiphyseal margin, femoral head, and subchondral bone.

[0062] Step 104: According to the parameter file obtained after training and the CT scans of all SCFE patients with the disease, obtain the segmentation results of the femur, femoral head physeal margin, femoral head, and subchondral bone of all SCFE patients. Resample the segmentation results according to the previous resampling strategy. After restoring to the original scale, merge the four labels together to obtain the final bone feature segmentation result.

[0063] According to the parameter file obtained after training and the CT scans of all SCFE patients with the disease, obtain the segmentation results of the femur, femoral head physeal margin, femoral head, and subchondral bone of all SCFE patients. Resample the segmentation results. For the input image I(x), x ∈ R 3 represents the three-dimensional coordinate points in space, and the resampling operation R of the target region 1 / Δs is to perform a scale transformation on the spatial coordinate x, where Δs is the target scale factor calculated above. Through the variable-scale resampling operation, a new image I″(x) can be obtained as follows:

[0064] I″(x) = R 1 / Δs (I(x))

[0065] After restoring to the original scale, merge the four labels together. First, record the coordinates of the label regions in the annotation file with the label "femoral head physeal margin", and modify the coordinates in the annotation file with the label "femoral head" to the label "femoral head physeal margin". Denote this file as "File A"; then record the coordinates of the label regions in the annotation file with the label "subchondral bone", and modify these coordinates in "File A" to the label "subchondral bone". Denote this file as "File B"; finally, record the coordinates of the label regions in the annotation file with the label "femur", and modify these coordinates in "File B" to the label "femur" to obtain the final bone feature segmentation result.

[0066] The technical solution of the embodiment of the present invention has the following beneficial effects:

[0067] Based on the original preoperative three-dimensional CT images of SCFE patients, each image was divided into left and right sides by the median sagittal plane, and the images of the affected side were retained to obtain the CT of the affected side of SCFE patients. Data annotation was performed on the femur, femoral head epiphyseal margin, femoral head, and subchondral bone of some CT scans of the affected side to obtain the annotation results. According to the original preoperative three-dimensional CT images of SCFE patients and the annotation results of each image, data resampling was performed according to the spatial proportion of the target area to adjust the data scale, obtaining the resampled images of each annotation label and the resampled images of their corresponding original images. The resampled images of each annotation label and the resampled images of their corresponding original images were put into SAM-Med3d and trained using the pre-trained parameter file sam_med3d_turbo.pth to obtain a fine-tuned parameter file that can segment the femur, femoral head epiphyseal margin, femoral head, and subchondral bone after training. According to the parameter file obtained after training and all CT scans of the affected side of SCFE patients, the segmentation results of the femur, femoral head epiphyseal margin, femoral head, and subchondral bone of all SCFE patients were obtained. The segmentation results were resampled according to the previous resampling strategy, and after restoring to the original scale, the four labels were merged together to obtain the final bone feature segmentation result. This method can achieve the automatic segmentation of the femur, femoral head epiphyseal margin, femoral head, and subchondral bone regions in the preoperative three-dimensional CT data of SCFE surgical patients, and can improve the segmentation accuracy compared with the original SAM-Med3d, which is of great significance for the subsequent surgical plan design.

[0068] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0069] The content not detailedly described in the specification of the present invention belongs to the well-known technology of those skilled in the art.

Claims

1. A femoral head region bone feature segmentation method using SAM-Med3d, characterized in that: The method comprises: Based on the original preoperative 3D CT images of SCFE patients, each image was divided into left and right sides in the midsagittal plane, and the image of the affected side was retained to obtain the affected CT images of SCFE patients. The data of the femur, femoral epiphysis edge, femoral head and subchondral bone were annotated on some of the affected CT images to obtain the annotation results. Based on the original preoperative 3D CT images of SCFE patients and the annotation results of each image, the data is resampled according to the spatial proportion of the target area, and the data scale is adjusted to obtain the resampled image of each annotated label and the resampled image of the corresponding CT image; The resampled image of each labeled label and its corresponding resampled image of the original image are put into SAM-Med3d and trained using pre-trained parameters to obtain fine-tuned parameters that can segment the femur, femoral epiphysis edge, femoral head and subchondral bone after training; According to the parameters obtained after training and the CT scans of all SCFE patients, the segmentation results of the femur, femoral epiphysis, femoral head and subchondral bone of all SCFE patients were obtained. The segmentation results were resampled and restored to the original scale. Then the four labels (femur, femoral epiphysis, femoral head and subchondral bone) were merged together to obtain the final bone feature segmentation result.

2. The femoral head region bone feature segmentation method using SAM-Med3d according to claim 1, characterized in that: According to the original preoperative 3D CT images of SCFE patients, each image was divided into left and right sides in the midsagittal plane, and the image of the affected side was retained to obtain the affected CT images of SCFE patients. The data of the femur, femoral epiphysis edge, femoral head and subchondral bone were annotated for some affected CT images to obtain the annotation results. The specific process is as follows: ITK-SNAP software was used to convert DICOM format CT images into NIFTI format, and the images were divided into left and right sides in the midsagittal plane. The images of patients with bilateral slippage were split into two data sets, and the diseased CT scans of patients with unilateral slippage were taken as one data set. Under the guidance of professional orthopedic surgeons, the femur, femoral epiphysis edge, femoral head and subchondral bone on the CT images were annotated with different labels.

3. The femoral head region bone feature segmentation method using SAM-Med3d according to claim 1, characterized in that: According to the original preoperative 3D CT images of SCFE patients and the annotation results of each image, the data is resampled according to the spatial proportion of the target area, and the data scale is adjusted to obtain the resampled image of each annotated label and the resampled image of the corresponding CT image. The specific process is as follows: According to the original preoperative 3D CT images of SCFE patients and the annotation results of each image, for target areas with a small spatial occupancy and high segmentation accuracy requirements (subchondral bone, femoral epiphysis edge and femoral head), the resampling method of reducing the voxel spacing of the original image is adopted to improve the spatial resolution of the image and enhance the model's perception of small feature areas; for larger areas (femur), the resampling method of enlarging the voxel spacing of the original image is adopted to reduce the computational complexity and retain the overall structural information, thereby optimizing the segmentation accuracy of the neural network for target areas with significant scale differences; For an input image I(x), x∈R 3 Represents the three-dimensional coordinate point in space and sets the resampling operation R of the target area Δs To perform a scale transformation on the spatial coordinate x, Δs is the target scale factor, which represents the voxel spacing after the scale transformation; through the scale resampling operation, the new image I′(x) is obtained as follows: Among them, s original and target Represent the voxel spacing of the original image and the target image respectively; Let V(I) be the volume of a label region in the image, Ω(I) be the total volume of the image input to the network, and the volume fraction α of the label region can be defined as For the hollow spherical region, the ratio of the convex hull volume S(I) of the region to the total volume of the Ω(I) image is calculated to reflect the size of the label region, and the boundary region size ratio β is set to Taking into account the actual voxel ratio and the boundary area size ratio, the comprehensive metric γ is defined as follows: γ=λ1α+λ2β Among them, λ1 and λ2 are the weight coefficients of the features; The resampling strategy can be implemented through the following judgment conditions: where γ min_threshold and γ max_threshold is the preset threshold, located in the interval [γ min_threshold ,γ max_threshold The label area within ] does not need to be resampled, and the label area smaller than this threshold uses the resolution s high Sampling is performed, otherwise resolution s is used low Sampling, obtain the resampled original image and labeled image.

4. The femoral head region bone feature segmentation method using SAM-Med3d according to claim 1, characterized in that: The resampled image of each labeled label and its corresponding resampled image of the original image are put into SAM-Med3d and trained using the pre-trained parameters to obtain the fine-tuned parameters that can segment the femur, femoral epiphysis edge, femoral head and subchondral bone after training. The specific process is as follows: The resampled image of each labeled label and its corresponding resampled image of the original image are taken as a group and put into SAM-Med3d using the pre-trained parameter file sam_med3d_turbo.pth for training. The training is performed on four labels: femur, femoral epiphysis edge, femoral head and subchondral bone, and the fine-tuned parameter file that can segment the femur, femoral epiphysis edge, femoral head and subchondral bone is obtained after training.

5. The femoral head region bone feature segmentation method using SAM-Med3d according to claim 1, characterized in that: According to the parameters obtained after training and the CT scans of all SCFE patients, the segmentation results of the femur, femoral epiphysis, femoral head and subchondral bone of all SCFE patients were obtained. The segmentation results were resampled and restored to the original scale. The four labels (femur, femoral epiphysis, femoral head and subchondral bone) were merged together to obtain the final bone feature segmentation results. The specific process is as follows: According to the parameter file obtained after training and the CT scans of all SCFE patients, the segmentation results of the femur, femoral epiphysis, femoral head and subchondral bone of all SCFE patients are obtained. The segmentation results are resampled. For the input image I(x), x∈R 3 Represents the three-dimensional coordinate point in space, the resampling operation R of the target area 1 / Δs To perform a scale transformation on the spatial coordinate x, where Δs is the target scale factor calculated above, the new image I″(x) can be obtained through the scale resampling operation as shown below: I″(x)=R 1 / Δs (I(x)) After restoring to the original scale, the four labels are merged together. First, the coordinates of the label area in the annotation file labeled "femoral head epiphysis edge" are recorded, and these coordinates in the annotation file labeled "femoral head" are modified to the "femoral head epiphysis edge" label, and this file is recorded as "file A"; then the coordinates of the label area in the annotation file labeled "subchondral bone" are recorded, and these coordinates in "file A" are modified to the "subchondral bone" label, and this file is recorded as "file B"; finally, the coordinates of the label area in the annotation file labeled "femur" are recorded, and these coordinates in "file B" are modified to the "femur" label to obtain the final bone feature segmentation result.

Images