A Fully Automated Quality Control Method for Knee Joint Radiographs Based on Multi-Task Deep Learning and Geometric Quantization

Abstract

This invention discloses a fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization. The method includes: using DICOM format knee joint DR images as input data and performing data preprocessing; constructing a multi-task model architecture comprising a knee joint anteroposterior and lateral bone segmentation model, a knee joint anteroposterior and lateral keypoint detection model, a foreign object detection model, and a left / right marker detection model, defining the network infrastructure and input / output of each model; calculating quality control-related indicators based on the bone mask, keypoints, foreign object detection results, and left / right marker detection results output by the four models; formulating scoring rules to score the quality of knee joint DR images, and performing automatic quality control of knee joint DR images. This invention achieves objective, efficient, and quantifiable knee joint DR image quality assessment by automatically extracting anatomical structures, calculating geometric parameters, and executing scoring rules, providing a reusable technical paradigm.

Description

Technical Field

[0001] This invention relates to the field of image quality control technology, specifically to a fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization. Background Technology

[0002] In orthopedic clinical practice, plain X-rays of the knee joint are a fundamental imaging tool for assessing conditions such as osteoarthritis, fractures, and malalignments. The image quality directly affects diagnostic accuracy and treatment decisions. High-quality DR images must meet requirements such as standard patient positioning, structural integrity, clear contrast, and absence of foreign body interference. Currently, image quality control mainly relies on the visual inspection and experience of radiology technicians, which is inefficient, inconsistent, and lacks quantitative data, making it difficult to meet the increasing volume of examinations and the demands of precision medicine.

[0003] In recent years, deep learning technology has rapidly developed in medical image analysis, with keypoint localization, semantic segmentation, and object detection models now capable of accurately identifying anatomical structures and abnormal objects. However, existing research, such as fracture detection or joint space measurement, largely focuses on single tasks and has not yet formed a multi-model collaborative reasoning framework for "image quality control." Especially in the knee joint quality control scenario, it is necessary to simultaneously evaluate multiple dimensions such as body position angle, center offset, field of view coverage, sharpness, and foreign object interference, which places higher demands on the geometric fusion and regularized scoring of the model output. Summary of the Invention

[0004] The purpose of this invention is to provide a fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization, so as to solve the problems mentioned in the background art.

[0005] The specific technical solution provided by this invention is as follows: A fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization, comprising:

[0006] The following are the steps to run:

[0007] Step S1: Use DICOM format knee joint DR images as input data and perform data preprocessing.

[0008] Preferably, data preprocessing includes:

[0009] Step S11: Read the window width and window level, as well as its anteroposterior and lateral views, from the header of the DICOM format knee joint DR image file;

[0010] Step S12: Convert DICOM pixel values ​​to window width and window level. grayscale value;

[0011] Step S13: Convert the image from pixel values ​​to grayscale values, convert it to a single-channel grayscale image, and the resulting matrix contains only one channel for subsequent input;

[0012] Step S14: Normalize the image pixel values ​​to The range is then converted to float32 to adapt to the deep learning model inference engine, outputting the input data for the multi-task model.

[0013] Step S2: Construct a multi-task model architecture including a knee joint anterolateral bone segmentation model, a knee joint anterolateral keypoint detection model, a foreign object detection model, and a left and right label detection model, and define the network infrastructure and input / output of each model.

[0014] Preferably, the knee joint anterolateral bone segmentation model, the knee joint anterolateral key point detection model, the foreign object detection model, and the left and right marker detection model are all trained using the mature deep learning training framework PyTorch.

[0015] Preferably, the network infrastructure and input / output of each model are defined as follows:

[0016] Step S21: UNet is used as the network architecture for the knee joint anteroposterior and lateral bone segmentation model. The input of this network architecture is defined as a 512×512×1 grayscale image. The output of the anteroposterior model is defined as a 4-channel probability matrix and the output of the lateral model is defined as a 5-channel probability matrix. The threshold of each channel is set to 0.5. In the probability matrix, the value of the point greater than the threshold is set to 1, and otherwise it is set to 0, generating a binarized bone mask.

[0017] Step S22: Use SCNet as the network infrastructure for the knee joint key point detection model. Define the input of this network infrastructure as a grayscale image of 640×640×1. Define the output of the frontal model as a 17-channel heatmap and the output of the lateral model as a 21-channel heatmap. The heatmap of each channel represents the probability of the predicted coordinates of the key point.

[0018] Step S23: Use Yolov9 as the network architecture for the foreign object detection model. Define the input of this network architecture as a grayscale image of 640×640×1. After inference and post-processing, the coordinate box information of the foreign object is obtained, and the foreign object detection result is output.

[0019] Step S24: Faster R-CNN is used as the network infrastructure for the left and right label detection model. The input of this network infrastructure is defined as a 512×512×1 grayscale image. When the character "L" or "R" is detected, the output is the corresponding category of the character and the bounding box coordinate information. When no "L" or "R" character is detected, None is returned. After inference and post-processing, the left and right label detection results are output.

[0020] Step S3: Calculate the quality control related indicators based on the skeleton mask, key points, foreign object detection results, and left and right marker detection results output by the four models.

[0021] Preferably, the quality control related indicators include: positioning angle, distance between the knee joint space and the center of the image, bilateral shadows of the knee joint lateral articular surface, patellar-femur space in the lateral view of the knee joint, field of view, signal-to-noise ratio, sharpness, overall image sharpness, number of foreign objects, and left and right markings.

[0022] Step S4: Develop scoring rules, score the quality of knee DR images, and automatically control the quality of knee DR images based on the scoring results.

[0023] Preferably, the established scoring rules include:

[0024] A score of 1 is awarded if the placement angle is within the set angle threshold, otherwise a score of 0 is awarded; the placement angle includes the angle in both upright and side positions.

[0025] One point is awarded if the distance between the knee joint space and the center of the image is less than a set threshold; otherwise, zero points are awarded.

[0026] A medial epicondyle area of ​​0 indicates no bilateral shadow on the lateral view of the knee joint, earning 1 point; otherwise, 0 points are awarded.

[0027] A score of 1 point is awarded for a clear separation between the patella and femur when the area of ​​the patella and femur mask intersecting is 0 in the lateral view of the knee joint; otherwise, a score of 0 is awarded.

[0028] One point is awarded if the field of vision is within the set threshold range; otherwise, zero points are awarded.

[0029] A score of 1 is awarded if the signal-to-noise ratio is within the set threshold range; otherwise, a score of 0 is awarded.

[0030] If the image sharpness is greater than the set threshold, the image is clear and sharp and scores 1 point; otherwise, it scores 0 points.

[0031] The number of foreign objects is 1 point is awarded, otherwise 0 points are awarded.

[0032] Left and right signs are represented as 1 point is awarded, otherwise 0 points are awarded.

[0033] For information regarding the medical institution's examination: 1 point is awarded for complete information including the name of the medical institution, the time of the equipment examination, the patient's examination number, the patient's information, and the imaging examination parameters; otherwise, 0 points are awarded.

[0034] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention automatically extracts anatomical structures, uses DICOM format knee joint DR images as input data, and performs data preprocessing; it constructs a multi-task model architecture including a knee joint anteroposterior and lateral bone segmentation model, a knee joint anteroposterior and lateral keypoint detection model, a foreign body detection model, and a left and right label detection model, defining the network infrastructure and input / output of each model; it calculates quality control-related indicators based on the bone mask, keypoints, foreign body detection results, and left and right label detection results output by the four models; it formulates scoring rules to score the quality of knee joint DR images, and performs automatic control of knee joint DR image quality, realizing objective, efficient, and quantifiable knee joint DR image quality assessment, and providing a reusable technical paradigm for the quality control of plain films of other parts such as the hand and foot. Attached Figure Description

[0035] Figure 1 This is a logical schematic diagram of a fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization provided in an embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0037] Example 1:

[0038] Combination Figure 1 As shown in the figure, the fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization described in this embodiment includes the following steps:

[0039] Step S1: Use DICOM format knee joint DR images as input data and perform data preprocessing.

[0040] In this embodiment, the data preprocessing process includes the following steps:

[0041] Step S11: Read the window width (WW) and window level (WL), as well as its front and side position information, from the DICOM file header;

[0042] Step S12: Convert DICOM pixel values ​​to window width (WW) and window level (WL). Grayscale value:

[0043]

[0044] in, This represents the pixel value after mapping is complete. Represents the original pixel value. Represents a numerical decision function that restricts the result to... Within the range, values ​​less than 0 are set to 0, and values ​​greater than 255 are set to 255;

[0045] Step S13: Convert the obtained image into a single-channel grayscale image. , ,in and These represent the height and width of the image, respectively. The resulting matrix contains only one channel and is used as subsequent input.

[0046] Step S14: Normalize the image pixel values ​​to The range is then converted to float32 to adapt to the deep learning model inference engine. The conversion formula is as follows: , This is the input data used for the multi-task model after transformation.

[0047] Step S2: Construct a multi-task model architecture including a knee joint anterolateral bone segmentation model, a knee joint anterolateral keypoint detection model, a foreign object detection model, and a left and right marker detection model, and define the network infrastructure and input / output mode of each model.

[0048] In this embodiment, the present invention adopts a multi-task model architecture, mainly using four models: a knee joint anterolateral bone segmentation model, a knee joint anterolateral keypoint detection model, a foreign object detection model, and a left / right marker detection model. All models are trained using mature deep learning training frameworks such as PyTorch. The network infrastructure and input / output modes of each model are defined as follows:

[0049] Step S21: UNet is used as the network architecture for the knee joint anteroposterior and lateral bone segmentation model. The input is a 512×512×1 grayscale image. The anteroposterior model outputs a 4-channel probability matrix, with each channel representing the mask probability of the femur, tibia, fibula, and patella. The lateral model outputs a 5-channel probability matrix, with each channel representing the mask probability of the femur, tibia, fibula, patella, and medial epicondyle. Finally, a value of 0.5 is applied to each channel to generate a binarized bone mask. ;

[0050] Step S22: SCNet is used as the network infrastructure of the knee joint key point detection model. The input is a grayscale image of 640×640×1. The frontal model outputs a 17-channel heatmap, and the lateral model outputs a 21-channel heatmap. The heatmap of each channel represents the probability of the predicted coordinates of the key point.

[0051] For example, the coordinates of the keypoint can be obtained by taking the global maximum response point. as follows:

[0052]

[0053] in, Represents the first The heatmap of each key point is obtained, and the coordinates of each key point are mapped back to the original image. The key points obtained specifically include: anteroposterior and lateral views, as detailed below:

[0054] Anteroposterior view: Lateral end of femoral shaft, medial prominence of femoral condyle, lowest point of patella, midpoint of femoral condyle at the apex of intercondylar fossa, lowest point of the center of intercondylar crest of tibia, midpoint of lateral / medial tibial plateau, lateral / medial epicondyle of femur, medial end of femoral shaft, lateral / medial end of tibial shaft, lateral / medial femoral condyle, lateral / medial end of tibial plateau, lateral prominence of femoral condyle;

[0055] Lateral view: superior posterior end of femoral shaft, inferior anterior end of tibial shaft, posterior / anterolateral end of tibial plateau, center of osteoporotic zone in intercondylar fossa of femur, superior / middle / inferior end of posterior border of patella, junction of tibia and fibula, intercondylar fossa of femur, bottom of intercondylar groove of femur, superior anterior end of femoral shaft, inferior anterior end of patella, anterior tibial tuberosity, inferior posterior / anterolateral end of femoral shaft, posterior / anterior end of long axis of femoral condyle, superior / inferior posterior / anterolateral end of tibial shaft.

[0056] Step S23: Using Yolov9 as the network architecture for the foreign object detection model, the input is a 640×640×1 grayscale image. After inference and post-processing, the coordinate bounding box information of the foreign object is obtained, and the output is the foreign object detection result, which is denoted as [missing information]. ,Right now ,in , This represents the minimum and maximum values ​​of the foreign object's coordinate frame on the x-axis. , Similarly.

[0057] Step S24: Faster R-CNN is used as the network architecture for the left and right label detection model. The input is a 512×512×1 grayscale image. If the character "L" or "R" is detected, the output is the category and bounding box coordinate information; otherwise, it returns None. After inference and post-processing, the left and right label detection results are output. , ,in This indicates that the detected character category is "L" or "R".

[0058] Step S3: Calculate the quality control related indicators based on the skeleton mask, key points, foreign object detection results, and left and right marker detection results output by the four models.

[0059] In this embodiment, the preprocessed image is inferred by the above model to obtain the knee joint bone mask, key point information, foreign object detection results and left and right label detection results. Based on these results, quality control related indicators are calculated one by one, and some quality control indicators are slightly different in the frontal and lateral views.

[0060] In this embodiment, the calculation process for the quality control indicator "positioning angle" in the upright and lateral positions is as follows:

[0061] In the anteroposterior position: the degree of verticality of the lower limb is calculated, that is, the lateral endpoint of the femoral shaft is calculated. - Lateral end of the tibial shaft The angle between the line connecting the two points and the vertical direction Medial end point of the femoral shaft -Medial end point of the tibial shaft The angle between the line connecting the two points and the vertical direction .

[0062] The defined vector is represented as follows:

[0063]

[0064]

[0065]

[0066] in, Represents the lateral end of the femoral shaft Pointing to the lateral end of the tibial shaft The vector, This represents the point from the medial end of the femoral shaft. Pointing to the medial end of the tibial shaft The vector, It represents a unit vector pointing in the positive y-axis direction in the vertical direction.

[0067] In the lateral view: by calculating the knee flexion angle. That is, the angle between the center line of the femoral shaft and the center line of the tibia. The femoral shaft centerline is defined as: the anterior superior end point of the femoral shaft. -Posterior superior end of the femoral shaft The midpoint of the line connecting the two points is the lower end of the anterior aspect of the femoral shaft. -Posterior inferior end of the femoral shaft The midpoint of the connecting line forms the central line of the tibia. The central line of the tibia is then the upper anterior end point of the tibia. - Upper end of the posterior aspect of the tibial shaft The midpoint of the line connecting the two points is the lower end of the anterior aspect of the tibial shaft. The lower end of the posterior aspect of the tibial shaft The midpoint of the line forms the line.

[0068]

[0069]

[0070]

[0071] in, Representative point and The midpoint coordinates, That is the point x-coordinate It is a point The ordinate, and This represents the point number; It is the center line of the femoral shaft. That is the center line of the tibial shaft.

[0072] In this embodiment, the calculation process for the distance between the knee joint space and the center of the image in the anteroposterior and lateral views is as follows:

[0073] In the neutral position: the lowest point of the patella. Euclidean distance from the center of the image The calculation formula is as follows:

[0074]

[0075]

[0076] in, This represents the physical coordinates of the image's center point in the world coordinate system. Represents the height and width of an image in pixels. This represents the length of each pixel in the real world, in mm. Indicates the lowest point of the patella. The corresponding x and y coordinates, , The x and y coordinates of the image center point are represented in pixels.

[0077] In the lateral view: the lower end of the anterior border of the patella. - Posterior end point of the tibial plateau Connect the line to the posterior end of the long axis of the femoral condyle. -Anterior end point of the tibial plateau The intersection point is taken when two line segments of a line intersect. Find the Euclidean distance between the intersection point and the center of the image. .

[0078] For example, the intersection point is at a Euclidean distance from the center of the image. The calculation steps are as follows:

[0079] a1. Define the direction vector of the line segment as follows:

[0080]

[0081]

[0082]

[0083]

[0084] in For point The x and y coordinates, The corresponding serial numbers for each endpoint are, such as 14, 15, 16, 17, etc. Representative line segment of Directional increment; Representative line segment of Directional increment.

[0085] a2. Calculate the determinant The calculation formula is as follows:

[0086]

[0087] like If the two line segments intersect, then the intersection point can be calculated further.

[0088] a3. Find the focus parameter :

[0089]

[0090] in For the section along the line The normalized scaling factor.

[0091] a4. Calculate the coordinates of the intersection point:

[0092]

[0093]

[0094] in, Represents the x-coordinate and y-coordinate of the intersection point.

[0095] a5. Finally, calculate the Euclidean distance between the intersection point and the image center. :

[0096]

[0097] in, This represents the length of each pixel in the real world, measured in mm.

[0098] In this embodiment, the area of ​​the medial epicondyle mask of the knee joint is calculated based on the bilateral images of the lateral articular surfaces. as follows:

[0099]

[0100] In this embodiment, the area of ​​intersection between the patella and femur in the lateral view of the knee joint is calculated. as follows:

[0101]

[0102] in, This indicates the patella in a lateral position at the knee joint. This refers to the femoral mask.

[0103] In this embodiment, the Field of View (FOV) is defined as the ratio of the total imaging size to the image size, and the imaging size is determined based on the boundaries of the entire skeleton mask.

[0104] For example, the specific calculation process for the field of view is as follows:

[0105] b1. Define the set of all pixels representing the skeleton as... , means as follows:

[0106]

[0107] in, The binary mask representing the skeleton. Image height, Image width, This represents the horizontal and vertical coordinate indices of that pixel. This indicates that a pixel value of 1 at this index represents a skeleton. That serves as the background.

[0108] b2. Find the smallest x and y coordinate indices corresponding to the pixel. and the largest x and y coordinate index , means as follows:

[0109]

[0110]

[0111] get These are the minimum and maximum x and y coordinates of the Mask representing the skeleton.

[0112] b3. Calculate FOV:

[0113]

[0114] In this embodiment, the signal-to-noise ratio (SNR) of the knee joint imaging is calculated, and the pixel set of the original image is defined as... , This represents the original image coordinates as Pixel values; This represents the coordinates within the skeleton mask. The pixel values ​​are used to obtain the original pixel set of the skeletal region. and the original image pixel set in the noise region They are represented as follows:

[0115]

[0116]

[0117]

[0118]

[0119]

[0120] in, This represents the mean signal in the skeletal region. This represents the standard deviation of background noise (noise std).

[0121] In this embodiment, the sharpness, a quality control index, is calculated. This invention uses a method based on second-order gray-level difference to characterize edge slope changes, thereby calculating image sharpness. The specific process is as follows:

[0122] c1. Calculate the local second difference ( );

[0123] For example, the image is first subjected to median filtering to remove noise and preserve edges, and the second-order difference of pixel gray levels is calculated in the horizontal (x) and vertical (y) directions. The calculation formula is as follows:

[0124]

[0125] in, The grayscale value is the result of median filtering. , These represent the positions of the image after median filtering. and The grayscale value at that location.

[0126] c2. Calculate the local sharpness index;

[0127] For example, in the edge neighborhood window Inner summation to estimate the total change in slope :

[0128]

[0129] in, This represents the row index along the horizontal direction, from arrive .

[0130] Then calculate the estimated edge contrast within the same window. :

[0131]

[0132] in, Represents the image intensity function, i.e., at the row index. Column index Pixel intensity value at that location That is, in row index Column index The pixel intensity value at that location, i.e., with The adjacent pixel in the previous row.

[0133] Finally, normalization is used to obtain the horizontal sharpness of a single pixel. :

[0134]

[0135] vertical Similarly.

[0136] c3. Calculate the overall image sharpness;

[0137] For example, setting an empirical threshold by using a threshold. To determine whether a single pixel is sharp, the ratio of sharp edge pixels to all edge pixels is calculated, yielding the ratio of sharp edge pixels in the horizontal and vertical directions. Finally, image sharpness is calculated based on this ratio. :

[0138]

[0139] in, This indicates the pixel ratio of sharp edges in the horizontal direction. This indicates the pixel ratio of sharp edges in the vertical direction.

[0140] In this embodiment, the number of foreign objects obtained by the statistical foreign object detection model is... .

[0141] In this embodiment, the left and right labels are... The system checks for the existence and correctness of the left and right markers. If the left and right markers are present and correctly positioned, they are then... Recorded as Otherwise remember .

[0142] Step S4: Develop scoring rules to score and control the quality of knee joint DR images.

[0143] In this embodiment, based on the results of geometric quantization parameter calculation, the present invention formulates a series of scoring rules to statistically score the input image. These scoring rules are formulated based on the statistical distribution of a large amount of data measurement results, thereby achieving quality control of knee joint DR images.

[0144] For example, the scoring rules are as follows:

[0145] (1) 1 point for proper placement, 0 points for otherwise; proper placement means in the correct orientation. In the case of lateral positioning, .

[0146] (2) If the distance between the knee joint space and the center of the image is less than 40mm, you get 1 point; otherwise, you get 0 points.

[0147] (3) 1 point is awarded if there is no bilateral shadow on the lateral view of the knee joint; otherwise, 0 points are awarded. The absence of bilateral shadow indicates the area of ​​the medial epicondyle. .

[0148] (4) 1 point is awarded if the patellar-femoral space is clearly separated; otherwise, 0 points are awarded. This refers to the area where the patella and femoral mask intersect in a lateral view of the knee joint. .

[0149] (5) 1 point for a suitable field of vision, 0 points otherwise. .

[0150] (6) A suitable signal-to-noise ratio earns 1 point; otherwise, 0 points are awarded. .

[0151] (7) A clear and sharp image earns 1 point; otherwise, it earns 0 points. .

[0152] (8) Number of foreign objects You get 1 point, otherwise you get 0 points.

[0153] (9) Left and right signs You get 1 point, otherwise you get 0 points.

[0154] (10) Patient and institution examination information: name of medical institution, time of equipment examination, patient examination number, patient information, and imaging examination parameters. 1 point is awarded for complete information above, otherwise 0 points are awarded.

[0155] Based on the scoring rules established above, the scores of the input knee joint images are calculated to help determine the image quality. The results of the specific index calculations can be saved to a JSON file for further review later.

[0156] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0157] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization, characterized by: The following steps are included: Step S1: Use DICOM format knee joint DR images as input data and perform data preprocessing; Step S2: Construct a multi-task model architecture including a knee joint anterolateral bone segmentation model, a knee joint anterolateral keypoint detection model, a foreign object detection model, and a left and right label detection model, and define the network infrastructure and input / output of each model; The network infrastructure and input / output of each model are defined as follows: Step S21: UNet is used as the network architecture for the knee joint anteroposterior and lateral bone segmentation model. The input of this network architecture is defined as a 512×512×1 grayscale image. The output of the anteroposterior model is defined as a 4-channel probability matrix and the output of the lateral model is defined as a 5-channel probability matrix. A threshold is set for each channel. In the probability matrix, the value of the point greater than the threshold is set to 1, and otherwise it is set to 0. A binarized bone mask is generated. Step S22: Use SCNet as the network infrastructure for the knee joint key point detection model. Define the input of this network infrastructure as a grayscale image of 640×640×1. Define the output of the frontal model as a 17-channel heatmap and the output of the lateral model as a 21-channel heatmap. The heatmap of each channel represents the probability of the predicted coordinates of the key point. Step S23: Use Yolov9 as the network architecture for the foreign object detection model. Define the input of this network architecture as a grayscale image of 640×640×1. After inference and post-processing, the coordinate box information of the foreign object is obtained, and the foreign object detection result is output. Step S24: Faster R-CNN is used as the network architecture for the left and right label detection model. The input of this network architecture is defined as a grayscale image of 512×512×1. When the character "L" or "R" is detected, the output is the corresponding category of the character and the bounding box coordinate information; when no "L" or "R" character is detected, None is returned. After inference and post-processing, the left and right label detection results are output. Step S3: Calculate the quality control related indicators based on the skeleton mask, key points, foreign object detection results, and left and right marker detection results output by the four models; Step S4: Develop scoring rules, score the quality of knee DR images, and automatically control the quality of knee DR images based on the scoring results.

2. The fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization as described in claim 1, characterized in that: Data preprocessing includes: Step S11: Read the window width and window level, as well as its anteroposterior and lateral views, from the header of the DICOM format knee joint DR image file; Step S12: Convert DICOM pixel values ​​to window width and window level. grayscale value; Step S13: Convert the image from pixel values ​​to grayscale values, convert it to a single-channel grayscale image, and the resulting matrix contains only one channel for subsequent input; Step S14: Normalize the image pixel values ​​to The range is then converted to float32 to adapt to the deep learning model inference engine, outputting the input data for the multi-task model.

3. The fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization as described in claim 2, characterized in that: The knee joint anteroposterior and lateral skeletal segmentation model, the knee joint anteroposterior and lateral key point detection model, the foreign object detection model, and the left and right marker detection model were all trained using the deep learning training framework PyTorch.

4. The fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization as described in claim 3, characterized in that: In step S21, the 4-channel probability matrix represents the mask probability of the femur, tibia, fibula and patella respectively; the 5-channel probability matrix represents the mask probability of the femur, tibia, fibula, patella and medial epicondyle respectively.

5. The fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization as described in claim 4, characterized in that: Step S22 also includes: obtaining the coordinates of each keypoint by taking the global maximum response point as the keypoint coordinates, and then mapping it back to the original image to obtain the keypoint coordinates; the keypoints corresponding to the keypoint coordinates include: Key points in the anteroposterior view: lateral end of the femoral shaft, medial prominence of the femoral condyle, lowest point of the patella, midpoint of the femoral condyle at the apex of the intercondylar fossa, lowest point of the center of the intercondylar crest of the tibia, midpoint of the lateral / medial tibial plateau, lateral / medial epicondyle of the femur, medial end of the femoral shaft, lateral / medial end of the tibial shaft, lateral / medial femoral condyle, lateral / medial end of the tibial plateau, lateral prominence of the femoral condyle; Key points in the lateral view: superior posterior end of the femoral shaft, inferior anterior end of the tibial shaft, posterior / anterolateral end of the tibial plateau, center of the osteoporotic zone of the intercondylar fossa of the femur, superior / middle / inferior end of the posterior border of the patella, junction of the tibia and fibula, intercondylar fossa of the femur, bottom of the intercondylar groove of the femur, superior anterior end of the femoral shaft, inferior anterior end of the patella, anterior tibial tuberosity, inferior posterior / anterolateral end of the femoral shaft, posterior / anterior end of the long axis of the femoral condyle, superior / inferior posterior / anterolateral end of the tibial shaft.

6. The fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization as described in claim 5, characterized in that: The quality control indicators in step S3 include: positioning angle, distance between the knee joint space and the center of the image, bilateral shadows of the knee joint lateral articular surface, patellar-femur space in the lateral view of the knee joint, field of view, signal-to-noise ratio, sharpness, overall image sharpness, number of foreign objects, and left / right markings.

7. The fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization as described in claim 6, characterized in that: The calculation of the placement angle includes: In the anteroposterior position: calculate the degree of verticality of the lower limb, that is, calculate the angle between the line connecting the lateral endpoint of the femoral shaft and the lateral endpoint of the tibia and the vertical direction, and the angle between the line connecting the medial endpoint of the femoral shaft and the medial endpoint of the tibia and the vertical direction. In the lateral view: Calculate the knee flexion angle, which is the angle between the center line of the femoral shaft and the center line of the tibia. The center line of the femoral shaft is the line formed by the midpoint of the line connecting the upper anterior end point of the femoral shaft and the upper posterior end point of the femoral shaft, and the midpoint of the line connecting the lower anterior end point of the femoral shaft and the lower posterior end point of the femoral shaft. The center line of the tibia is the line formed by the midpoint of the line connecting the upper anterior end point of the tibia and the upper posterior end point of the tibia, and the midpoint of the line connecting the lower anterior end point of the tibia and the lower posterior end point of the tibia. The calculation of the distance between the knee joint space and the center of the image includes: In the anteroposterior view: the Euclidean distance between the lowest point of the patella and the center of the image. Perform calculations; In the lateral view: Intersect the two line segments connecting the lower end of the anterior border of the patella to the posterior end of the tibial plateau with the line connecting the posterior end of the long axis of the femoral condyle to the anterior end of the tibial plateau. Calculate the Euclidean distance between this intersection point and the center of the image. ; The Euclidean distance between the intersection point and the center of the image The calculations include: a1. Define the direction vector of the line segment; a2. Calculate the determinant. If the calculated determinant is not equal to 0, determine whether the two line segments intersect. a3. Further calculate the intersection parameters; a4. Calculate the coordinates of the intersection point; a5. Finally, calculate the Euclidean distance between the intersection point and the image center. .

8. The fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization as described in claim 7, characterized in that: The calculation of the field of view includes: b1. Define the set of all pixels representing the skeleton; b2. Find the smallest and largest x and y coordinate indices corresponding to the pixel; b3. Calculate the field of view; Sharpness calculation includes: c1. Perform median filtering on the image to remove noise and preserve edges, and calculate the second difference of pixel gray levels in the horizontal and vertical directions; c2. Sum within the edge neighborhood window to estimate the total slope change, calculate the estimated edge contrast within the same window, and finally normalize to obtain the horizontal sharpness of a single pixel. c3. Determine whether a single pixel is sharp by setting an empirical threshold, calculate the ratio of sharp edge pixels to all edge pixels, obtain the ratio of sharp edge pixels in the horizontal direction and the ratio of sharp edge pixels in the vertical direction, and calculate the image sharpness.

9. The fully automated quality control method for knee joint DR radiographs based on multi-task deep learning and geometric quantization as described in claim 8, characterized in that: The established scoring rules include: A score of 1 is awarded if the placement angle is within the set angle threshold, otherwise a score of 0 is awarded; the placement angle includes the angle in both upright and side positions. One point is awarded if the distance between the knee joint space and the center of the image is less than a set threshold; otherwise, zero points are awarded. A medial epicondyle area of ​​0 indicates no bilateral shadow on the lateral view of the knee joint, earning 1 point; otherwise, 0 points are awarded. A score of 1 point is awarded for a clear separation between the patella and femur when the area of ​​the patella and femur mask intersecting is 0 in the lateral view of the knee joint; otherwise, a score of 0 is awarded. One point is awarded if the field of vision is within the set threshold range; otherwise, zero points are awarded. A score of 1 is awarded if the signal-to-noise ratio is within the set threshold range; otherwise, a score of 0 is awarded. If the image sharpness is greater than the set threshold, the image is clear and sharp and scores 1 point; otherwise, it scores 0 points. The number of foreign objects is 1 point is awarded, otherwise 0 points are awarded. Left and right signs are represented as 1 point is awarded, otherwise 0 points are awarded. For information regarding the medical institution's examination: 1 point is awarded for complete information including the name of the medical institution, the time of the equipment examination, the patient's examination number, the patient's information, and the imaging examination parameters; otherwise, 0 points are awarded.

Images