Orthopedic image automatic adjustment method

Abstract

The present application relates to the technical field of image processing, in particular to a kind of orthopedic image automatic adjustment method, the method is carried out skeleton region segmentation and peak classification by deep learning, determines inflection point using gradient-distance double criterion, dynamically adjusts parameter according to image entropy and standard deviation, fuses multimodal feature optimization threshold, realizes the automatic adjustment of orthopedic image.Compared with traditional method, the skeleton boundary recognition accuracy is significantly improved, adaptability and processing efficiency are enhanced, can effectively assist doctor to carry out orthopedic disease diagnosis and surgery planning.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for automatic adjustment of orthopedic images, which aims to improve the quality of orthopedic images and assist doctors in making more accurate disease diagnoses and surgical plans. Background Technology

[0002] CT imaging plays a crucial role in orthopedic clinical diagnosis and treatment. Doctors need clear and accurate CT images to observe bone structures, assess lesions, and formulate treatment plans. However, raw CT images often have various problems and cannot directly meet clinical needs.

[0003] On the one hand, CT images acquired under different equipment and scanning conditions vary significantly in grayscale and contrast. For example, some low-dose CT scans produce images with high noise levels and blurred boundaries between bone and soft tissue; while high-resolution CT images, although rich in detail, have a more discrete voxel distribution, making it difficult for traditional fixed-parameter adjustment methods to adapt to this variation. This makes it difficult for doctors to clearly distinguish the fine structures of bones and lesion features when observing the images, easily leading to misdiagnosis or missed diagnosis.

[0004] On the other hand, existing image adjustment methods have many limitations. Some traditional methods rely on manually adjusting window width and level, which not only consumes a lot of time and energy for doctors, but also the adjustment effect depends on the doctor's experience and subjective judgment, and the adjustment results may vary greatly among different doctors. In addition, some automatic adjustment methods are based on simple histogram analysis, using fixed mathematical conditions to screen peaks and using empirical parameters to determine inflection points and thresholds. In an automatic CT image adjustment method published in CN115568873A, the peak judgment condition relies on a fixed formula, which does not consider different bone types and complex clinical situations, and is prone to missing atypical peaks; the method for determining the geometric convexity inflection point is too simple and is prone to misjudgment when there are complex fluctuations in the soft tissue-bone boundary; and the fixed empirical parameters (such as m=9, setting a threshold of 0.02) lack universality and may fail in images from different devices and different scanning sites. These problems seriously affect the accuracy and efficiency of image adjustment and cannot meet the diverse needs of clinical practice.

[0005] With the continuous development of medical technology, the requirements for the quality of orthopedic imaging are becoming increasingly stringent. Developing an efficient, accurate, and adaptive automatic adjustment method for orthopedic images has become an urgent problem to be solved. It is of great significance for improving the diagnostic accuracy of orthopedic diseases, optimizing surgical planning, and enhancing the quality of medical services. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic adjustment method for orthopedic images based on multimodal fusion and adaptive parameters, in order to solve the problems of poor accuracy, weak adaptability and low efficiency in the adjustment of orthopedic images in the prior art, improve the accuracy of bone boundary recognition, enhance the adaptability to different equipment, different scanning sites and normal and diseased images, shorten the image adjustment time, and assist doctors to diagnose orthopedic diseases and plan surgeries more efficiently and accurately.

[0007] The technical solution adopted by this invention to solve its technical problem is: an automatic adjustment method for orthopedic images, comprising:

[0008] The input DICOM format CT images are preprocessed to standardize the CT value range to a specific interval, and a pre-trained bone segmentation model is used to extract the regions of interest (ROI) of cortical bone and cancellous bone.

[0009] The histogram of the ROI is smoothed with 5 levels of Gaussian smoothing. A density-based peak detection algorithm is used to identify all local maxima. Then, a convolutional neural network classifier is used to classify the peaks and select high-density skeletal feature peaks with confidence > 0.9 and CT value > 200 HU.

[0010] Calculate the first and second derivatives of the histogram interval to the right of the target peak, and mark candidate points in the descending region where the slope changes from positive to negative, the derivative is less than 0, and the second derivative is less than 0. Calculate the distance from each candidate point to the line connecting the peak point and the end point, as well as the absolute value of the local gradient. Based on an adaptive weight dynamically adjusted by the image entropy value of the bone region of interest, select points that simultaneously meet the distance threshold and gradient threshold conditions as convex inflection points of the bone-soft tissue boundary. The selection criteria for convex inflection points are: the distance from the candidate point to the line connecting the peak point and the end point is greater than 80% of the maximum distance of all candidate points, and the absolute value of the local gradient of the candidate point is greater than the average of the absolute values ​​of the local gradients of all candidate points.

[0011] Based on the location of the convexity inflection point, the minimum and maximum thresholds are dynamically calculated by combining the image entropy value and the standard deviation of the voxel distribution, and then the window level and window width are calculated to complete the automatic optimization of image grayscale display.

[0012] Specifically, the formula for standardizing the CT value range is: ,in These are the original CT values. This is the standardized CT value.

[0013] Specifically, the density-based peak detection algorithm is the DBSCAN algorithm.

[0014] Specifically, the adaptive weights are dynamically adjusted based on the entropy value of the skeletal ROI, and the formula for calculating the entropy value is as follows: ,in For the first The probability of each gray level appearing This represents the total number of gray levels.

[0015] Specifically, when dynamically calculating the maximum threshold, the search window The range is ,and Automatic adjustment based on image entropy value; the higher the image entropy value, the better. The larger.

[0016] Specifically, when calculating the maximum threshold, a threshold is set. The formula is: ,in The average value of the voxel distribution. The standard deviation of the voxel distribution. This is the adjustment coefficient.

[0017] Specifically, the pre-trained skeleton segmentation model is a 3D-ResNet model.

[0018] Specifically, it also includes using the Canny operator to extract bone edges before calculating the threshold, calculating the CT value distribution of edge points, so that the threshold range includes at least 95% of the edge points.

[0019] Specifically, it also includes calculating ROI texture uniformity based on Local Binary Pattern (LBP), and if the texture variance is greater than the threshold, the window width is automatically reduced to ±20% of the peak range.

[0020] Specifically, the method can process normal and pathological (such as fractures, bone tumors) orthopedic CT images of adults and children.

[0021] The beneficial effects of this invention are:

[0022] 1. Improved accuracy: By using deep learning for skeletal ROI segmentation and peak classification, combined with multimodal feature fusion, the bone boundaries and features can be identified more accurately. In a test of 500 orthopedic CT images, the accuracy of bone boundary recognition increased from 89.2% to 98.7% compared to traditional methods. In particular, the error was less than 5% in low-contrast images (noise standard deviation > 100), effectively reducing misdiagnosis and missed diagnosis.

[0023] 2. Enhanced adaptability: The designed adaptive parameter adjustment mechanism dynamically adjusts parameters based on image entropy and standard deviation, enabling the method to automatically adapt to the image characteristics of different devices (such as 64-slice / 320-slice CT), different scanning sites (spine / joint), and adults / children, normal / lesion (such as fractures, bone tumors), without the need for manual parameter calibration.

[0024] 3. Improved efficiency: The processing time for a single image is less than 0.3 seconds, which is 98.8% faster than manual adjustment (average 2.5 minutes). It also supports real-time adjustment of 3D volume data, which greatly saves doctors' time and improves clinical work efficiency.

[0025] 4. Optimized detail display: By fusing edge and texture features to optimize thresholds, the details of complex anatomical structures (such as articular surfaces and trabeculae) can be better displayed, improving image quality and providing doctors with richer diagnostic information.

[0026] 5. High level of intelligence: The introduction of deep learning preprocessing and data-driven adaptive algorithms reduces human intervention, lowers the reliance on doctors' experience, and improves the system's intelligence level, enabling it to better meet diverse clinical needs. Attached Figure Description

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] Figure 1 The flowchart illustrates an automatic adjustment method for orthopedic images provided by this invention. Detailed Implementation

[0029] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0030] like Figure 1 As shown, the present invention provides the following technical solution:

[0031] Preprocessing and tissue ROI extraction: Input DICOM format CT images and use formulas. The CT value range was standardized to a specific interval to avoid zero-value interference. A pre-trained 3D-ResNet skeleton segmentation model was used to extract cortical and cancellous bone ROIs, excluding soft tissue interference areas and narrowing the scope of subsequent histogram analysis.

[0032] Multi-peak feature selection and dynamic sorting: The ROI histogram is smoothed with 5 levels of Gaussian smoothing, and the density-based DBSCAN peak detection algorithm is used to identify all local maxima. Peak points are classified by a convolutional neural network (CNN) classifier and labeled as effective skeletal peaks or invalid noise peaks. Peaks with a confidence score > 0.9 and a CT value > 200 HU are selected as target peaks and sorted from high to low CT values, prioritizing high-density bones.

[0033] Inflection point localization based on gradient-distance dual criteria: Calculate the first derivative sequence of the histogram interval to the right of the target peak, and mark all "descending inflection point candidate points" where the slope changes from positive to negative (i.e., derivative less than 0 and second derivative less than 0). For each candidate point, calculate the distance D from it to the line connecting the peak point and the endpoint, using the formula: (where the equation of the straight line is) Point coordinates And the absolute value of the local gradient G. Choose the option that simultaneously satisfies ( (the maximum distance among all candidate points), and ( The point where the average of the absolute values ​​of the local gradients of the candidate points is taken as the final convexity inflection point. Adaptive weights (such as distance weights and gradient weights) are dynamically adjusted by the entropy value of the skeletal ROI, calculated using the formula: ,in For the first The probability of each gray level appearing (This represents the total number of gray levels).

[0034] Adaptive threshold calculation and window width / level adjustment: minimum threshold Take the left boundary value of the interval corresponding to the convexity inflection point. Dynamically search for the maximum threshold. ,initialization After calculating the current interval Group voxel average ,in Automatic adjustment based on image entropy value (the higher the image entropy value, the more dispersed the voxel distribution). The larger, the range is The threshold formula is set as follows: (in, The average value of the voxel distribution. The standard deviation of the voxel distribution. (This is the adjustment coefficient). If the condition is met, then... Take the first Right boundary of the group; otherwise, search to the end of the histogram. Calculate window level. Apply window functions to adjust the grayscale display of the image.

[0035] Multimodal feature fusion optimization: The Canny operator is used to extract bone edges, and the CT value distribution of edge points is calculated, with a constraint threshold range covering at least 95% of the edge points. ROI texture uniformity is calculated based on Local Binary Pattern (LBP). If the texture variance > threshold, the window width is automatically reduced to ±20% of the peak range to enhance detail contrast.

[0036] Example 1: Adjustment of normal adult spinal CT images

[0037] A DICOM format CT image of a normal adult spine is acquired, with raw CT values ​​ranging from -1000 to 3000. This image is then input into the automatic image adjustment system of this invention. First, the system adjusts the image according to the formula... Standardizing the CT values ​​and mapping them to the [0,1] interval can effectively eliminate the influence of differences in CT values ​​from different devices, laying the foundation for subsequent accurate analysis.

[0038] Next, the system calls the pre-trained 3D-ResNet model to extract the skeletal region from the image, accurately separating the cortical and cancellous bone ROIs of the spine, eliminating interference from surrounding soft tissues, narrowing the scope of subsequent analysis, and improving processing efficiency and accuracy.

[0039] The extracted ROI histogram is smoothed using a 5-level Gaussian smoothing process, which effectively reduces data noise and makes the histogram curve smoother, facilitating accurate peak identification. Subsequently, the DBSCAN algorithm is used to identify local maxima in the smoothed histogram. After identifying numerous maxima, the system uses a CNN classifier to filter out target peaks that meet the set conditions—confidence > 0.9 and CT value > 200 HU.

[0040] To determine the convexity inflection point, the system calculates the first derivative of the histogram interval to the right of the target peak and marks candidate points in the descending region where both the first and second derivatives are less than 0. For each candidate point, the system applies the formula... (where the equation of the straight line is) Point coordinates Calculate the distance D from the peak point to the line connecting the peak and the terminal points, and simultaneously calculate the absolute value G of the local gradient. This is achieved through a set dual criterion, namely... ( (the maximum distance among all candidate points), and ( The point that is the average of the absolute values ​​of the local gradients of the candidate points is taken as the final convexity inflection point.

[0041] The system then calculates the entropy value of the image and automatically adjusts the search window based on the entropy value. A higher entropy value indicates a more dispersed voxel distribution. Increase accordingly ( According to the adjusted After calculating the current interval Group voxel average Combined with the standard deviation of voxel distribution and adjustment coefficient Through formula Calculate the maximum threshold minimum threshold Take the left boundary value of the interval corresponding to the convexity inflection point.

[0042] Furthermore, according to the formula Calculate the window level. Calculate the window width.

[0043] Finally, the system applies window functions to adjust the grayscale display of the image. At this point, the adjusted image clearly displays the vertebral bodies, intervertebral spaces, and other structures of the spine. The bone boundaries are clear and sharp, and details such as internal trabeculae are also clearly discernible, providing doctors with high-quality imaging evidence for accurately observing the morphology of the spine and determining whether there are any lesions.

[0044] Example 2: Adjustment of normal limb bone CT images in children

[0045] CT images of normal limb bones in children were selected, with CT values ​​ranging from -500 to 2000. After the images were input into the system, CT values ​​were standardized and mapped to the interval [0,1].

[0046] Next, the ROIs of cortical and cancellous bone in the limbs were accurately extracted using a 3D-ResNet model. Since children's bones are still in the growth and development stage, with a higher cartilage component and lower CT values ​​compared to adults, peak selection presented a challenge. However, the CNN classifier in this system, with its powerful feature recognition capabilities, was still able to accurately identify peaks that met the criteria.

[0047] When determining the inflection point, the system also calculates the first derivative of the histogram interval to the right of the target peak, marks candidate points in the descent region, and then determines the convexity inflection point based on the dual criteria of distance and gradient. During this process, adaptive weights are dynamically adjusted according to the entropy value of the ROI of the child's limb bones to ensure the accuracy of inflection point localization.

[0048] By combining image entropy and standard deviation, the system dynamically adjusts parameters to calculate the minimum threshold suitable for children's limb bone images. and maximum threshold The window level and width are then determined. The adjusted images clearly show the growth plates of the child's limb bones, with the low-density areas clearly demarcated from the surrounding bone. The cortical and cancellous bone structures of the diaphysis are also clearly visible, with fine and clear bone texture. This effectively avoids the image blurring problem caused by parameter incompatibility, providing reliable imaging support for doctors to assess the child's bone growth and development.

[0049] Example 3: Adjustment of CT images of fracture sites in adults

[0050] For CT images of fracture sites in adults (CT value range [-800, 3500]), the system first performs standardization processing, mapping the CT values ​​to the [0, 1] interval, and then uses a 3D-ResNet model to extract the ROI of the fracture site and surrounding bones.

[0051] The histogram features of the fracture area are complex, with multiple peaks that contain information about the fracture line, bone fragments, and normal bone tissue. The system utilizes DBSCAN and a CNN classifier to accurately identify peaks related to the fracture, based on a confidence level > 0.9 and a CT value > 200 HU.

[0052] When determining the inflection point, due to the unique characteristics of the fracture area, its boundary may exhibit irregular fluctuations. The system calculates the first derivative of the histogram interval to the right of the target peak, marks candidate points in the descending region, and determines the inflection point based on dual criteria. During this process, adaptive weights are adjusted according to the entropy value of the ROI of the fracture area, resulting in more accurate positioning of the fracture boundary.

[0053] By combining multimodal feature fusion and threshold optimization, the system uses the Canny operator to extract bone edges, ensuring that the threshold range includes at least 95% of edge points. Simultaneously, it calculates ROI texture uniformity based on the LBP algorithm. If the texture variance exceeds the threshold, the window width is automatically reduced to ±20% of the peak range to highlight details in the fracture area. The adjusted image clearly displays the fracture line, its shape and direction are readily apparent; the location and size of bone fragments are also clearly visible, providing crucial imaging information for doctors to formulate fracture treatment plans, such as selecting appropriate reduction methods and fixation devices.

[0054] Example 4: Adjustment of CT images for adult bone tumors

[0055] CT images of adults with bone tumors (CT value range [-600, 4000]) are acquired. After the images are input into the system, they are first standardized to map the CT values ​​to the [0, 1] interval. Then, the ROI of the tumor area and the surrounding normal bone tissue is extracted using a 3D-ResNet model.

[0056] During peak selection, the CNN classifier fully leverages its powerful classification capabilities to effectively distinguish peak features between tumor tissue and normal bone tissue. Peak values ​​in tumor tissue may exhibit different CT value ranges and distribution characteristics compared to normal bone; through the CNN classifier's selection process, tumor-related peak values ​​can be accurately identified.

[0057] When determining the inflection point based on dual criteria, the system adjusts adaptive weights according to the entropy value of the tumor region's area of ​​interest (ROI), making the determination of the inflection point more accurate and thus clearly defining the tumor boundary. Combining image entropy and standard deviation, the system dynamically adjusts parameters to calculate a suitable minimum threshold. and maximum threshold This allows us to obtain the window position and window width.

[0058] Simultaneously, edge and texture features are fused using the Canny operator and the LBP algorithm. The bone edge information extracted by the Canny operator helps determine the threshold range, ensuring that the tumor boundary is clearly presented in the image; the texture uniformity calculated by the LBP algorithm further highlights the morphology and texture features of the tumor. The adjusted image makes the tumor boundary and surrounding bone structure clearly distinguishable, and the internal texture details of the tumor, such as the presence of necrotic areas and calcifications, are also clearly displayed. This helps doctors determine the nature and extent of the tumor, providing important evidence for developing personalized treatment plans.

[0059] Example 5: Comparison of adjustments to orthopedic images scanned by different devices

[0060] Orthopedic images of the same patient were acquired using 64-slice and 320-slice CT scanners, respectively. Images from the 64-slice CT scanner had relatively higher noise levels and sparser voxel distribution; images from the 320-slice CT scanner had a more discrete voxel distribution and higher resolution, but may also exhibit different noise patterns.

[0061] Both sets of images were input into the system of this invention, and both were first subjected to CT value standardization, mapping the CT values ​​to the [0,1] interval. Then, a 3D-ResNet model was used to extract the skeletal ROI. In subsequent processing, due to the differences in characteristics between the two sets of images, the adaptive parameter adjustment mechanism played a crucial role.

[0062] For 64-slice CT images, the system automatically increases the search window based on their higher noise level and image entropy value. At the same time, adjust the threshold. The calculation parameters are optimized to better adapt to noisy environments and avoid peak misjudgment and inaccurate thresholding caused by noise interference. For 320-slice CT images, due to their more discrete voxel distribution, the system adjusts the search window accordingly based on the image entropy value and standard deviation. and threshold This ensures that the fine details of the bones can be accurately captured.

[0063] Ultimately, after processing by this system, both sets of images clearly displayed the skeletal structure. The 64-slice CT images, while suppressing noise, preserved key skeletal features; the 320-slice CT images, fully utilizing their high resolution, clearly presented minute details of the bones, such as the fine structure of trabeculae. This demonstrates that this method has good adaptability to images scanned by different devices, effectively improving the quality of images from various equipment and providing reliable support for clinical diagnosis.

[0064] Example 6: Comparison of Orthopedic Image Adjustments at Different Scanning Sites

[0065] CT images of the spine and hip joint from the same patient were selected. The spine has a complex structure, including multiple vertebrae, intervertebral discs, and surrounding ligaments and other tissues; the hip joint has special structures such as the femoral head and acetabulum, and its anatomical structure and tissue density distribution differ significantly from those of the spine.

[0066] CT images of the spine and hip joint are input into the system separately. First, CT values ​​are standardized and mapped to the [0,1] interval. Then, the corresponding cortical and cancellous bone ROIs are accurately extracted for different locations using a 3D-ResNet model.

[0067] During the peak selection phase, due to the complex peak shapes of the histograms in spinal images, there are multiple peaks reflecting different vertebral bodies and intervertebral disc structures. The system's CNN classifier accurately selects the target peaks based on the characteristics of the spinal ROI. The peaks in hip joint images are mainly related to the structures of the femoral head and acetabulum, and the CNN classifier can also accurately identify them.

[0068] When determining inflection points, the adaptive weights of spinal images are adjusted based on their complex peak shape and high entropy value. The gradient-distance dual criterion is used to more accurately determine the boundary inflection points between vertebrae and between vertebrae and intervertebral discs. Hip joint images, on the other hand, adjust adaptive weights based on their own structural characteristics, such as the circular outline of the femoral head and the concave structure of the acetabulum, to accurately determine the boundary inflection points of various structures in the hip joint.

[0069] During threshold calculation, the system dynamically adjusts parameters based on the entropy and standard deviation of images from different locations. For example, spinal images may require more refined threshold adjustments to distinguish the structures of different vertebrae and intervertebral discs; hip joint images optimize threshold calculation based on the density differences between the femoral head and acetabulum.

[0070] The final adjusted images clearly show the shape of the vertebral bodies, the width of the intervertebral spaces, and the position and shape of the intervertebral discs, which helps doctors observe the stability of the spine and whether there are lesions such as disc herniation. The hip joint images clearly show the integrity of the femoral head, the shape of the acetabulum, and the condition of the hip joint space, providing doctors with clear and accurate imaging evidence for diagnosing hip joint diseases such as femoral head necrosis and hip dysplasia.

[0071] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic adjustment of orthopedic images, characterized in that, include: The input DICOM format CT images are preprocessed to standardize the CT value range to a specific interval, and a pre-trained bone segmentation model is used to extract the regions of interest for cortical bone and cancellous bone. The histogram of the region of interest is smoothed by five levels of Gaussian smoothing. A density-based peak detection algorithm is used to identify all local maxima. Then, a convolutional neural network classifier is used to classify the peaks and select high-density skeletal feature peaks with confidence > 0.9 and CT value > 200 HU. Calculate the first and second derivatives of the histogram interval to the right of the target peak, and mark candidate points in the descending region where the slope changes from positive to negative, the derivative is less than 0, and the second derivative is less than 0. Calculate the distance from each candidate point to the line connecting the peak point and the end point, as well as the absolute value of the local gradient. Based on an adaptive weight dynamically adjusted by the image entropy value of the bone region of interest, select points that simultaneously meet the distance threshold and gradient threshold conditions as convex inflection points of the bone-soft tissue boundary. The selection criteria for convex inflection points are: the distance from the candidate point to the line connecting the peak point and the end point is greater than 80% of the maximum distance of all candidate points, and the absolute value of the local gradient of the candidate point is greater than the average of the absolute values ​​of the local gradients of all candidate points. Based on the location of the convexity inflection point, the minimum and maximum thresholds are dynamically calculated by combining the image entropy value and the standard deviation of the voxel distribution, and then the window level and window width are calculated to complete the automatic optimization of image grayscale display.

2. The method for automatic adjustment of orthopedic images according to claim 1, characterized in that: The formula for standardizing the CT value range is: ,in These are the original CT values. This is the standardized CT value.

3. The method for automatic adjustment of orthopedic images according to claim 1, characterized in that: The density-based peak detection algorithm is the DBSCAN algorithm.

4. The method for automatic adjustment of orthopedic images according to claim 1, characterized in that: The adaptive weights are dynamically adjusted based on the entropy value of the region of interest in the skeleton. The formula for calculating the entropy value is as follows: ,in For the first The probability of each gray level appearing This represents the total number of gray levels.

5. The method for automatic adjustment of orthopedic images according to claim 1, characterized in that: When dynamically calculating the maximum threshold, the search window... The range is ,and Automatic adjustment based on image entropy value; the higher the image entropy value, the better. The larger.

6. The method for automatic adjustment of orthopedic images according to claim 1, characterized in that: When calculating the maximum threshold, set a threshold. The formula is: ,in The average value of the voxel distribution. The standard deviation of the voxel distribution. This is the adjustment coefficient.

7. The method for automatic adjustment of orthopedic images according to claim 1, characterized in that: The pre-trained skeleton segmentation model is a 3D-ResNet model.

8. The method for automatic adjustment of orthopedic images according to claim 1, characterized in that: It also includes using the Canny operator to extract bone edges and calculate the CT value distribution of edge points before calculating the threshold, so that the threshold range includes at least 95% of the edge points.

9. The method for automatic adjustment of orthopedic images according to claim 1, characterized in that: It also includes calculating the texture uniformity of the region of interest based on the local binary pattern. If the texture variance is greater than the threshold, the window width is automatically reduced to ±20% of the peak range.

10. The method for automatic adjustment of orthopedic images according to claim 1, characterized in that: The method can process normal and pathological orthopedic CT images of adults and children.

Images