Method for automatically identifying osteolysis around prosthesis after total hip replacement

Abstract

The invention discloses an automatic identification method for osteolysis around a prosthesis after total hip replacement, and belongs to the technical field of medical image processing.The method is achieved through an osteolysis focus identification and prediction model, in the model, image data of a patient after total hip replacement is converted and mapped, and an image matrix is obtained; obtaining an initial partition map of the image matrix through the anatomical partition template; determining a weight matrix of the initial partition map according to preset weights of different anatomical partitions, multiplying the weight matrix by the image matrix to obtain a partition guide enhanced image matrix, inputting the partition guide enhanced image matrix into a feature extraction module, inputting extracted micro texture features and macro features into a Sigmoid gating dynamic fusion module, and outputting the Sigmoid gating dynamic fusion module. And a target bounding box and target confidence are obtained through the fusion features, the intermediate layer feature map and the YOLOv8 recognition module, a staging result is obtained based on the fusion features, clinical data and a staging prediction module, and visual recognition information is further obtained.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, and is a method for automatic identification of periprosthetic osteolysis after total hip arthroplasty, but is not limited to one. Background Technology

[0002] Total hip arthroplasty is a core treatment for end-stage hip joint diseases and is widely used in clinical practice. However, postoperative osteolysis around the prosthesis is one of the main complications leading to prosthesis loosening, failure, and ultimately requiring a second revision. Therefore, identifying whether osteolysis has occurred around the prosthesis is of paramount importance.

[0003] In related technologies, physicians manually interpret images to diagnose osteolysis or use general orthopedic models to identify whether bone dissolution has occurred. However, this method suffers from several drawbacks. Firstly, it is highly subjective. Secondly, because early lesions are only 2-5 mm in size, it is difficult to capture the subtle textural features and spatial distribution patterns of osteolysis, resulting in a low recognition rate.

[0004] Therefore, how to efficiently and accurately identify micro-bone resorption around the prosthesis has become an urgent problem to be solved. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide an automatic identification method for periprosthetic osteolysis after total hip arthroplasty, which at least solves the problem that related technologies cannot efficiently and accurately identify minute osteolysis around the prosthesis.

[0006] According to a first aspect of the present invention, an automatic identification method for periprosthetic osteolysis after total hip arthroplasty is provided. The osteolysis lesion identification and prediction model includes a feature extraction module, a Sigmoid-gated dynamic fusion module, a YOLOv8 identification module, and a staging prediction module, comprising: Obtain imaging and clinical data of patients after total hip arthroplasty; The image data is transformed and mapped to obtain the corresponding first image matrix; and the initial partition map corresponding to the preprocessed first image matrix is ​​obtained based on the anatomical partition template. The weight matrix corresponding to the initial partition map is determined based on the preset weights of each anatomical partition in the anatomical partition template. The first image matrix and the weight matrix are then multiplied to obtain the second image matrix for partition-guided enhancement. The preset weights are set based on the osteolysis incidence rate. The second image matrix is ​​input into the feature extraction module to obtain micro-texture features, macro-features and intermediate layer feature maps. The micro-texture features and macro-features are then input into the Sigmoid-gated dynamic fusion module to obtain fused features. The target bounding box and target confidence of osteolysis lesions are obtained based on fusion features, intermediate layer feature maps and YOLOv8 recognition module. The staging results of osteolysis lesions are obtained based on fusion features, clinical data and staging prediction module. The loss function of osteolysis lesion recognition and prediction model is constructed based on staging loss and localization loss, with the weight of staging loss being greater than the weight of localization loss. The target bounding box, target confidence, and staging results are overlaid on the image data to obtain visualized recognition information.

[0007] Optionally, the feature extraction module includes a micro-texture feature extraction layer and a macro-feature extraction layer; The step of inputting the second image matrix into the feature extraction module to obtain micro-texture features, macro-features, and intermediate layer feature maps includes: The second image matrix is ​​input into the micro-texture feature extraction layer, decomposed by Haar wavelet to obtain decomposition vectors, and feature extraction is performed based on the decomposition vectors to obtain micro-texture features; The second image matrix is ​​input into the macroscopic feature extraction layer to obtain macroscopic features and intermediate layer feature maps.

[0008] Optionally, the YOLOv8 recognition module includes three detection layers at different scales; The acquisition of the target bounding box and target confidence score of osteolysis lesions based on fusion features, intermediate layer feature maps, and the YOLOv8 recognition module includes: Based on the fusion features and intermediate layer feature maps, multiple fused feature maps of different scales are obtained, and the multiple fused feature maps of different scales are input into the detection layer of the corresponding scale to obtain the initial bounding box and initial confidence of the osteolysis lesion. The target bounding box and target confidence are obtained by filtering the initial confidence using a pre-set confidence threshold and a non-maximum suppression algorithm.

[0009] Optionally, the phased prediction module includes a feature splicing layer and multiple fully connected layers; The staging results of osteolytic lesions, based on fusion features, clinical data, and staging prediction modules, are obtained, including: The fusion features, normalized postoperative time and friction interface type from clinical data are input into the feature splicing layer to obtain splicing features. The splicing features are then input into multiple fully connected layers to obtain the staging results of osteolytic lesions.

[0010] Optionally, the osteolysis lesion identification and prediction model is trained through the following process: Acquire imaging sample data and clinical sample data; the imaging sample data carries manually annotated bounding box samples of osteolytic lesions, staging labels, and anatomical partition labels; Training sample data is obtained from the image sample data; and the corresponding partition-guided enhancement third image matrix is ​​obtained based on the anatomical partition labels and the training sample data. Based on the third image matrix, clinical sample data, the feature extraction module to be trained, the Sigmoid gated dynamic fusion module to be trained, the YOLOv8 recognition module to be trained, and the staging prediction module to be trained, the predicted staging results and the first bounding box of the osteolysis lesion are obtained respectively. The phase loss is calculated based on the predicted phase results and phase labels, and the positioning loss is calculated using the first bounding box and bounding box samples. The total loss is calculated based on the stage loss and the localization loss. The parameters of the osteolysis lesion identification and prediction model are then adjusted using the total loss until the trained osteolysis lesion identification and prediction model is obtained.

[0011] Optionally, the formula for the total loss is as follows: ; In the above formula, For the total loss, To pinpoint the loss, The losses are calculated in installments, with weights of 0.4 and 0.6 respectively.

[0012] According to a second aspect of the present invention, an electronic device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; the memory is used to store at least one executable instruction, wherein the executable instruction causes the processor to perform an operation corresponding to the method described in the first aspect.

[0013] According to the solution provided in the embodiments of the present invention, imaging data and clinical data of a patient after total hip arthroplasty are acquired; the imaging data are converted and mapped to obtain a corresponding first image matrix; and an initial partition map corresponding to the preprocessed first image matrix is ​​obtained based on an anatomical partition template; a weight matrix corresponding to the initial partition map is determined according to the preset weights corresponding to each anatomical partition in the anatomical partition template, and the first image matrix and the weight matrix are multiplied to obtain a partition-guided enhancement second image matrix; wherein, the preset weights are set based on the osteolysis incidence rate; the second image matrix is ​​input to a feature extraction module to obtain micro-texture features and macro-features. The system first obtains the fused features by inputting the intermediate layer feature map, micro-texture features, and macro-features into a Sigmoid-gated dynamic fusion module. Based on the fused features, intermediate layer feature map, and YOLOv8 recognition module, the target bounding box and target confidence of the osteolytic lesion are obtained. Then, based on the fused features, clinical data, and staging prediction module, the staging result of the osteolytic lesion is obtained. The loss function of the osteolytic lesion recognition and prediction model is constructed based on staging loss and localization loss, with the weight of staging loss being greater than that of localization loss. The target bounding box, target confidence, and staging result are superimposed on the image data to obtain visualized recognition information. In this process, a weight matrix is ​​generated for the initial partition map, and a partition-guided enhancement image matrix is ​​generated based on the weight matrix, which strengthens the grayscale values ​​of pixels in high-risk partitions and weakens those in low-risk partitions. Then, the Sigmoid-gated dynamic fusion module fuses the microscopic texture features and macroscopic features to accurately capture minute osteolytic lesions. These lesions are then processed by the YOLOv8 recognition module, allowing them to be captured by a dedicated detection layer for small targets. Finally, the target bounding box, target confidence score, and staging results are obtained. In summary, this method can efficiently and accurately identify minute osteolytic lesions around prostheses. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart illustrating an automatic identification method for periprosthetic osteolysis after total hip arthroplasty, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0015] Figure 1This is a flowchart illustrating an automatic identification method for periprosthetic osteolysis after total hip arthroplasty provided by an embodiment of the present invention. The automatic identification method for periprosthetic osteolysis after total hip arthroplasty provided by an embodiment of the present invention can be executed by an electronic device, such as a computer or server.

[0016] like Figure 1 As shown, the automatic identification method for periprosthetic osteolysis after total hip arthroplasty includes: S101. Obtain imaging and clinical data of the patient after total hip arthroplasty.

[0017] In embodiments of the present invention, total hip arthroplasty is a surgical treatment method that replaces the diseased or damaged femoral head and acetabulum of the hip joint with artificial prostheses to restore joint function and relieve pain. After a patient undergoes total hip arthroplasty, imaging data and clinical data are collected. The imaging data can be anteroposterior X-ray images or computed tomography (CT) images, and the clinical data can include the patient's surgical procedure, postoperative time, baseline time, and friction interface type, etc.

[0018] S102. The image data is transformed and mapped to obtain the corresponding first image matrix; and the initial partition map corresponding to the preprocessed first image matrix is ​​obtained based on the anatomical partition template.

[0019] In an embodiment of the present invention, pixel values ​​in the image data are converted into true grayscale values, and the converted true grayscale values ​​are mapped to the range [0, 255] to obtain a first image matrix corresponding to the image data. Next, the first image matrix undergoes preprocessing based on the type of image data, including contrast enhancement, denoising, artifact removal, and grayscale normalization, to obtain a preprocessed first image matrix. A recognition model is trained and used to identify the preprocessed first image matrix, finding the acetabular rotation center and the lower edge of the lesser trochanter, which are then used as reference points. A preset standard anatomical partition template is obtained. This anatomical partition template is a structured reference model based on standard anatomical definitions. Based on the reference points, the preset standard anatomical partition template is mapped onto the preprocessed first image matrix through geometric calculations to obtain an initial partition map of the image data. This initial partition map is an anatomical structure base map precisely aligned with the image data, marking all standard anatomical partitions such as acetabular regions I, II, and III, and femoral regions 1-7.

[0020] For example, when using positive X-ray image data, the pixel values ​​are first converted to true grayscale values ​​and then mapped to [0, 255] to obtain the corresponding first image matrix. Then, the window size is dynamically adjusted based on the standard deviation of the pixel grayscale values ​​in the first image matrix to enhance contrast. Adaptive median filtering smooths soft tissue noise, and Otsu thresholding generates a metal mask, resulting in a binary mask matrix. Finally, this binary mask matrix is ​​multiplied inversely with the first image matrix to temporarily mask the prosthesis region, yielding the artifact-free first image matrix, i.e., the preprocessed first image matrix. When using CT image data, a coronal plane reconstruction image is first extracted from the CT image data. The pixel values ​​in the coronal plane reconstruction image are converted to HU values, and the converted HU values ​​are truncated to [-100, 1000], retaining only bone tissue and the prosthesis, and then mapped to [0, 255] to obtain the corresponding first image matrix. An iterative reconstruction artifact correction method is used to correct beam hardening artifacts on this first image matrix, followed by denoising, to obtain the final preprocessed first image matrix.

[0021] S103. Determine the weight matrix corresponding to the initial partition map according to the preset weight of each anatomical partition in the anatomical partition template, and multiply the first image matrix and the weight matrix to obtain the second image matrix of partition-guided enhancement; wherein, the preset weight is set based on the osteolysis incidence rate.

[0022] In embodiments of the present invention, based on existing medical knowledge and clinical statistical data, a standard anatomical partition template can be pre-created. This template includes different regions, and the probability of osteolysis in each region is statistically analyzed. Weights are assigned to each region based on the statistical results, resulting in an anatomical partition weight assignment standard table, as shown in Table 1. Table 1 includes different anatomical partitions, each with a corresponding estimated incidence of osteolysis lesions and a set weight. The preset weights are assigned to different regions of the corresponding initial partition map based on Table 1, ultimately obtaining the weight matrix corresponding to the initial partition map. Finally, the first image matrix and the weight matrix are multiplied element-wise to obtain the second image matrix for partition-guided enhancement.

[0023] Table 1. Anatomical Region Weight Assignment Standards ; S104. Input the second image matrix into the feature extraction module to obtain micro-texture features, macro-features and intermediate layer feature maps, and input the micro-texture features and macro-features into the Sigmoid-gated dynamic fusion module to obtain fused features.

[0024] In an embodiment of the present invention, the osteolysis lesion identification and prediction model includes a feature extraction layer and a Sigmoid-gated dynamic fusion layer. A partition-guided enhanced second image matrix is ​​input to the feature extraction layer, which focuses on the high-frequency texture details corresponding to the second image matrix to obtain micro-texture features. Simultaneously, macro-features and intermediate layer feature maps are extracted from the partition-guided enhanced second image matrix. Finally, the micro-texture features and macro-features are input to the Sigmoid-gated dynamic fusion layer and fused according to a fusion formula to obtain fused features. The fusion formula is as follows: ; In the above formula, As a feature of fusion, Microscopic texture features, As a macroscopic feature, To calculate the weighting coefficients.

[0025] The weight coefficients W are calculated using the Sigmoid activation function, as follows: ; In the above formula, , For learnable weights, For bias vectors, It is the Sigmoid activation function.

[0026] S105. Based on fusion features, intermediate layer feature maps, and the YOLOv8 recognition module, the target bounding box and target confidence of osteolysis lesions are obtained. Based on fusion features, clinical data, and the staging prediction module, the staging results of osteolysis lesions are obtained. The loss function of the osteolysis lesion recognition and prediction model is constructed based on staging loss and localization loss, with the weight of staging loss being greater than the weight of localization loss.

[0027] In an embodiment of the present invention, the YOLOv8 recognition module is the main task branch, comprising three detection layers with different scales. These three detection layers process the fused features and intermediate layer feature maps to obtain the target bounding box and target confidence score of the osteolytic lesion. The staging prediction module is a sub-task branch, which processes the fused features and clinical data to obtain the predicted staging result of the osteolytic lesion, including early, intermediate, or late stages.

[0028] Specifically, the staging prediction module and the YOLOv8 recognition module output staging and location results. Based on the predicted staging and location results, the corresponding losses are calculated. Weights are set for the staging loss and the location loss, with the weight of the staging loss being greater than that of the location loss. The staging loss and the location loss are summed based on the set weights to obtain the total loss. The osteolysis lesion recognition and prediction model is then trained to obtain the trained osteolysis lesion recognition and prediction model.

[0029] S106. Overlay the target bounding box, target confidence score, and staging results onto the image data to obtain visualized recognition information.

[0030] In an embodiment of the present invention, annotations are overlaid on the image data, wherein red bounding boxes mark the target bounding boxes of osteolytic lesions, white text marks the staging results, yellow text marks the target confidence level, and yellow dashed boxes mark high-risk zones. Finally, visualized image data and clinical reports are generated for subsequent diagnosis and strategy formulation.

[0031] It is understood that, in the embodiments of the present invention, the osteolysis lesion identification and prediction model includes a feature extraction module, a Sigmoid-gated dynamic fusion module, a YOLOv8 identification module, and a staging prediction module. The process involves acquiring imaging and clinical data of the patient after total hip arthroplasty; converting and mapping the imaging data to obtain a corresponding first image matrix; obtaining an initial partition map corresponding to the preprocessed first image matrix based on an anatomical partition template; determining the weight matrix corresponding to the initial partition map according to the preset weights of each anatomical partition in the anatomical partition template; and multiplying the first image matrix and the weight matrix to obtain a partition-guided enhanced second image matrix; wherein the preset weights are set based on the osteolysis incidence rate; and inputting the second image matrix into the feature extraction module to obtain micro-texture features, macro-features, and intermediate layer feature maps. The micro-texture features and macro-features are input into a Sigmoid-gated dynamic fusion module to obtain fused features. Based on the fused features, intermediate layer feature maps, and a YOLOv8 recognition module, the target bounding box and target confidence of the osteolysis lesions are obtained. Based on the fused features, clinical data, and a staging prediction module, the staging results of the osteolysis lesions are obtained. The loss function of the osteolysis lesion recognition and prediction model is constructed based on staging loss and localization loss, with the weight of staging loss being greater than that of localization loss. The target bounding box, target confidence, and staging results are superimposed on the image data to obtain visualized recognition information. In this process, a weight matrix is ​​generated for the initial partition map, and a partition-guided enhancement image matrix is ​​generated based on the weight matrix, which strengthens the grayscale values ​​of pixels in high-risk partitions and weakens those in low-risk partitions. Then, the Sigmoid-gated dynamic fusion module fuses the microscopic texture features and macroscopic features to accurately capture minute osteolytic lesions. These lesions are then processed by the YOLOv8 recognition module, allowing them to be captured by a dedicated detection layer for small targets. Subsequently, the target bounding box, target confidence score, and staging results are obtained. In summary, this method can efficiently and accurately identify minute osteolytic lesions around prostheses.

[0032] In some embodiments of the present invention, inputting the second image matrix into the feature extraction module in S104 to obtain micro-texture features, macro-features and intermediate layer feature maps can be achieved through S1041 to S1042, as described in the following steps.

[0033] S1041. Input the second image matrix into the micro-texture feature extraction layer, decompose it using Haar wavelet to obtain the decomposition vector, and extract features based on the decomposition vector to obtain the micro-texture features.

[0034] S1042. Input the second image matrix into the macro feature extraction layer to obtain macro features and intermediate layer feature maps.

[0035] In some embodiments of the present invention, the feature extraction module includes a micro-texture feature extraction layer and a macro-feature extraction layer. The second image matrix with partitioned guidance enhancement is input into the micro-texture feature extraction layer and decomposed by Haar wavelet to obtain multiple high-frequency decomposition vectors. The corresponding gray-level entropy, contrast and energy are calculated for each of the multiple high-frequency decomposition components, and statistical features are extracted from the gray-level entropy, contrast and energy to obtain micro-texture features. At the same time, the second image matrix with partitioned guidance enhancement is input into the macro-feature extraction layer and processed by the ResNet50 network to obtain macro-features and intermediate layer feature maps of different levels.

[0036] In some embodiments of the present invention, the acquisition of the target bounding box and target confidence of the osteolysis lesion based on fusion features, intermediate layer feature maps and YOLOv8 recognition module in S105 can be implemented through S1051 to S1052, as described in the following steps.

[0037] S1051. Based on the fusion features and intermediate layer feature maps, obtain multiple fused feature maps at different scales, and input the multiple fused feature maps at different scales into the detection layer at the corresponding scale to obtain the initial bounding box and initial confidence of the osteolysis lesion.

[0038] In an embodiment of the present invention, after obtaining the fused features, the fused features and corresponding intermediate layer feature maps of different levels can be input into the feature pyramid network to obtain multiple fused feature maps of different scales. Since the three detection layers have different scales, the multiple fused feature maps of different scales are input into the corresponding scale detection layers to obtain the initial bounding box and initial recognition confidence of the osteolytic lesion.

[0039] S1052. Using a pre-set confidence threshold and a non-maximum suppression algorithm, the target bounding box and target confidence are obtained by filtering the initial confidence.

[0040] In some embodiments of the present invention, since some initial bounding boxes are background boxes, duplicate boxes, or erroneous boxes, they are first filtered from the initial confidence levels using a preset confidence threshold to obtain the bounding boxes of the osteolytic lesions and the filtered confidence levels. The filtered confidence levels are then further processed using a non-maximum suppression algorithm. Through sorting the filtered confidence levels, calculating the overlap of the bounding boxes, and removing redundancy, the final target bounding box and target confidence level of the osteolytic lesions are obtained.

[0041] In some embodiments of the present invention, the staging result of osteolysis lesions obtained by the fusion feature, clinical data and staging prediction module in S105 can be achieved through S105A, which will be explained through the following steps.

[0042] S105A: Input the fusion features, normalized postoperative time and friction interface type from clinical data into the feature splicing layer to obtain splicing features, and input the splicing features into multiple fully connected layers to obtain the staging results of osteolytic lesions.

[0043] In some embodiments of the present invention, the staging prediction module includes a feature splicing layer and multiple fully connected layers, such as three fully connected layers. The friction interface type can be obtained from clinical data and can be ceramic-ceramic, metal-polyethylene, or ceramic-polyethylene. The fusion features, normalized postoperative time from the clinical data, and friction interface type are input into the feature splicing layer. The normalized postoperative time and friction interface type are data expanded in spatial dimension and spliced ​​with the fusion features in the channel dimension to obtain the spliced ​​features. The spliced ​​features are then input into multiple fully connected layers to obtain the staging results of the osteolytic lesions. The multiple fully connected layers are sequentially connected.

[0044] In some embodiments of the present invention, the osteolysis lesion identification and prediction model can be implemented through the following steps: S201. Obtain imaging sample data and clinical sample data; the imaging sample data carries manually annotated bounding box samples of osteolytic lesions, staging labels, and anatomical partition labels.

[0045] In some embodiments of the present invention, patient image sample data and clinical sample data are acquired. The image sample data can be anteroposterior X-ray image sample data and CT image sample data. If it is CT image sample data, it is reconstructed in three dimensions. Then, the reconstructed image sample data can be filtered according to factors such as postoperative follow-up time ≥3 months, absence of motion artifacts or metal artifacts ≤30%, clear prosthesis type, and complete surgical records. Finally, image sample data containing multiple valid cases is selected, covering the early, middle, and late postoperative follow-up periods. The corresponding clinical sample data covers the patient's surgical records, including prosthesis type, friction interface type, surgical procedure, etc.

[0046] Furthermore, after acquiring the imaging and clinical sample data, the imaging sample data is manually labeled. Doctors mark the bounding boxes of osteolytic lesions on the imaging sample data, and then select a corresponding anatomical partition label for each lesion according to a standard anatomical partition template. Next, doctors determine a staging label by combining the visual size of the bounding boxes of the osteolytic lesions with the existing manual staging in the clinical sample data. If the bounding boxes of osteolytic lesions marked by multiple doctors have low overlap or different staging labels, senior doctors are consulted for judgment. Finally, a consistent boundary label and staging label are obtained, and each imaging sample data ultimately carries a staging label, bounding box, and anatomical partition label.

[0047] S202. Obtain training sample data from the image sample data; and obtain the corresponding partition-guided enhancement third image matrix based on the anatomical partition labels and the training sample data.

[0048] S203. Based on the third image matrix, clinical sample data, the feature extraction module to be trained, the Sigmoid gated dynamic fusion module to be trained, the YOLOv8 recognition module to be trained, and the staging prediction module to be trained, the predicted staging results and the first bounding box of the osteolysis lesion are obtained respectively.

[0049] In some embodiments of the present invention, the image sample data is divided to obtain training sample data, test sample data and validation sample data; the training sample data is processed by the above steps S102 to S105 to obtain the predicted staging results and the first bounding box of the osteolysis lesion.

[0050] S204. Calculate the phase loss based on the predicted phase results and phase labels, and calculate the positioning loss using the first bounding box and bounding box samples.

[0051] In some embodiments of the present invention, a loss function, such as the cross-entropy loss function, is set, and the predicted staging results and staging labels are substituted into the cross-entropy loss function to obtain the staging loss. Then, another loss function is set, which can be the full intersection-union loss function, and the first bounding box and the bounding box samples are substituted into the full intersection-union loss function to obtain the localization loss.

[0052] S205. Calculate the total loss based on the stage loss and the location loss, and adjust the parameters of the osteolysis lesion identification and prediction model through the total loss until the trained osteolysis lesion identification and prediction model is obtained.

[0053] In some embodiments of the present invention, the formula for the total loss is as follows: ; In the above formula, For the total loss, To pinpoint the loss, The losses are calculated in installments, with weights of 0.4 and 0.6 respectively.

[0054] Furthermore, after substituting the staged loss and localization loss into the above loss function to calculate the total loss, the parameters of the osteolysis lesion identification and prediction model are adjusted using the total loss to obtain the initial osteolysis lesion identification and prediction model. Then, the initial osteolysis lesion identification and prediction model is tested and verified using test sample data and validation sample data until the trained osteolysis lesion identification and prediction model is obtained.

[0055] In the training, testing, and validation processes, the performance metrics for evaluating the osteolysis lesion identification and prediction model can be accuracy, recall, precision, and intersection over union (IoU). During the validation phase, if the early lesion recall rate is <85%, the confidence threshold during inference is lowered to 0.4 and the model is re-evaluated without retraining.

[0056] Reference Figure 2 The diagram shows a structural schematic of an electronic device according to an embodiment of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the electronic device.

[0057] like Figure 2 As shown, the electronic device may include: a processor 502, a communication interface 504, a memory 506, and a communication bus 508.

[0058] in: The processor 502, communication interface 504, and memory 506 communicate with each other via communication bus 508.

[0059] Communication interface 504 is used to communicate with other electronic devices or servers.

[0060] The processor 502 is used to execute program 510, specifically the relevant steps in the above method embodiments.

[0061] Specifically, program 510 may include program code that includes computer operation instructions.

[0062] Processor 502 may be a central processing unit (CPU), a specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The smart device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0063] Memory 506 is used to store program 510. Memory 506 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0064] Specifically, program 510 can be used to cause processor 502 to perform the operations corresponding to the methods described in the above method embodiments.

[0065] The specific implementation of each step in program 510 can be found in the corresponding descriptions of the steps and units in the above method embodiments, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments, and will not be repeated here.

[0066] It should be noted that, depending on the implementation needs, the various components / steps described in the embodiments of the present invention can be broken down into more components / steps, or two or more components / steps or parts of the operation of components / steps can be combined into new components / steps to achieve the purpose of the embodiments of the present invention.

[0067] The methods described above according to embodiments of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

Claims

1. A method for automatically identifying periprosthetic osteolysis after total hip arthroplasty, characterized in that, The method is implemented through a bone resorption lesion identification and prediction model, which includes a feature extraction module, a Sigmoid-gated dynamic fusion module, a YOLOv8 identification module, and a staging prediction module. The method includes: Obtain imaging and clinical data of patients after total hip arthroplasty; The image data is transformed and mapped to obtain the corresponding first image matrix; and the initial partition map corresponding to the preprocessed first image matrix is ​​obtained based on the anatomical partition template. The weight matrix corresponding to the initial partition map is determined based on the preset weights of each anatomical partition in the anatomical partition template. The first image matrix and the weight matrix are then multiplied to obtain the second image matrix for partition-guided enhancement. The preset weights are set based on the osteolysis incidence rate. The second image matrix is ​​input into the feature extraction module to obtain micro-texture features, macro-features and intermediate layer feature maps. The micro-texture features and macro-features are then input into the Sigmoid-gated dynamic fusion module to obtain fused features. The target bounding box and target confidence of osteolysis lesions are obtained based on fusion features, intermediate layer feature maps and YOLOv8 recognition module. The staging results of osteolysis lesions are obtained based on fusion features, clinical data and staging prediction module. The loss function of osteolysis lesion recognition and prediction model is constructed based on staging loss and localization loss, with the weight of staging loss being greater than the weight of localization loss. The target bounding box, target confidence, and staging results are overlaid on the image data to obtain visualized recognition information.

2. The method according to claim 1, characterized in that, The feature extraction module includes a micro-texture feature extraction layer and a macro-feature extraction layer; The step of inputting the second image matrix into the feature extraction module to obtain micro-texture features, macro-features, and intermediate layer feature maps includes: The second image matrix is ​​input into the micro-texture feature extraction layer, decomposed by Haar wavelet to obtain decomposition vectors, and feature extraction is performed based on the decomposition vectors to obtain micro-texture features; The second image matrix is ​​input into the macroscopic feature extraction layer to obtain macroscopic features and intermediate layer feature maps.

3. The method according to claim 1, characterized in that, The YOLOv8 recognition module contains three detection layers of different scales; The acquisition of the target bounding box and target confidence score of osteolysis lesions based on fusion features, intermediate layer feature maps, and the YOLOv8 recognition module includes: Based on the fusion features and intermediate layer feature maps, multiple fused feature maps of different scales are obtained, and the multiple fused feature maps of different scales are input into the detection layer of the corresponding scale to obtain the initial bounding box and initial confidence of the osteolysis lesion. The target bounding box and target confidence are obtained by filtering the initial confidence using a pre-set confidence threshold and a non-maximum suppression algorithm.

4. The method according to claim 1, characterized in that, The phased prediction module includes a feature splicing layer and multiple fully connected layers; The staging results of osteolytic lesions, based on fusion features, clinical data, and staging prediction modules, are obtained, including: The fusion features, normalized postoperative time and friction interface type from clinical data are input into the feature splicing layer to obtain splicing features. The splicing features are then input into multiple fully connected layers to obtain the staging results of osteolytic lesions.

5. The method according to any one of claims 1 to 4, characterized in that, The osteolysis lesion identification and prediction model was trained through the following process: Acquire imaging sample data and clinical sample data; the imaging sample data carries manually annotated bounding box samples of osteolytic lesions, staging labels, and anatomical partition labels; Training sample data is obtained from the image sample data; and the corresponding partition-guided enhancement third image matrix is ​​obtained based on the anatomical partition labels and the training sample data. Based on the third image matrix, clinical sample data, the feature extraction module to be trained, the Sigmoid gated dynamic fusion module to be trained, the YOLOv8 recognition module to be trained, and the staging prediction module to be trained, the predicted staging results and the first bounding box of the osteolysis lesion are obtained respectively. The phase loss is calculated based on the predicted phase results and phase labels, and the positioning loss is calculated using the first bounding box and bounding box samples. The total loss is calculated based on the stage loss and the localization loss. The parameters of the osteolysis lesion identification and prediction model are then adjusted using the total loss until the trained osteolysis lesion identification and prediction model is obtained.

6. The method according to claim 5, characterized in that, The formula for the total loss is as follows: ； In the above formula, For the total loss, To pinpoint the loss, The losses are calculated in installments, with weights of 0.4 and 0.6 respectively.

Images