Femoral head center identification method and device and computer equipment

Abstract

The embodiment of the invention is suitable for the technical field of computer-aided medical treatment and image processing, and provides a femoral head center identification method and device and computer equipment, and the method comprises the steps: obtaining to-be-identified femoral head image data; identifying the femoral head image data by using a femoral head center identification model to obtain model prediction coordinates; wherein the femoral head center identification model is obtained by training a plurality of sample image data marked with a femoral head center point; and determining the caput femoris center coordinates in the caput femoris image data according to the model prediction coordinates. By adopting the method, the femoral head center in the femoral head image data can be automatically identified based on the trained artificial intelligence model, and the efficiency and accuracy of femoral head center identification are improved.

Description

Technical Field

[0001] This application relates to the fields of computer-aided medical technology and image processing technology, and in particular to a method, device and computer equipment for identifying the center of the femoral head. Background Technology

[0002] Identifying landmarks at human joints has important applications in medical image analysis, orthopedic surgical planning, and auxiliary diagnosis. For example, the femoral head center is a key reference point for preoperative planning, biomechanical analysis, and diagnosis of related diseases in hip replacement surgery, and is of great significance in orthopedic clinical diagnosis and surgical planning.

[0003] Traditional methods for identifying the femoral head center rely primarily on the doctor's experience and manual measurement. This approach is not only time-consuming and labor-intensive, but also inefficient, and susceptible to the influence of the doctor's subjective factors, making it difficult to guarantee the accuracy and consistency of the identification results. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method, apparatus, and computer device for femoral head center recognition, which can automatically recognize femoral head image data by training an artificial intelligence (AI) model, quickly output accurate femoral head center coordinates, and improve the efficiency and accuracy of femoral head center recognition.

[0005] The first aspect of this application provides a method for identifying the center of the femoral head, including: Obtain the femoral head imaging data to be identified; The femoral head center recognition model is used to identify the femoral head image data to obtain the model's predicted coordinates; wherein, the femoral head center recognition model is trained using multiple sample image data labeled with the femoral head center point; Based on the coordinates predicted by the model, the coordinates of the femoral head center in the femoral head image data are determined.

[0006] A second aspect of this application provides a femoral head center identification device, comprising: The acquisition module is used to acquire the femoral head image data to be identified; The recognition module is used to identify the femoral head image data using a femoral head center recognition model to obtain the model's predicted coordinates; wherein, the femoral head center recognition model is trained using multiple sample image data labeled with the femoral head center point; The determination module is used to determine the coordinates of the femoral head center in the femoral head image data based on the coordinates predicted by the model.

[0007] A third aspect of this application provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the computer device performs the method described in the first aspect above.

[0008] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a computer, implements the method described in the first aspect above.

[0009] A fifth aspect of this application provides a computer program product, including a computer program that, when run, causes the method described in the first aspect to be executed.

[0010] Compared with the prior art, the embodiments of this application have the following beneficial effects: In this embodiment, an AI model capable of automatically identifying the femoral head center can be trained using multiple sample image data labeled with the femoral head center point. Thus, during the preoperative or intraoperative phase, when femoral head image data to be identified is acquired, the trained AI model can automatically process and identify the femoral head image data, outputting corresponding model-predicted coordinates. Based on these model-predicted coordinates, the coordinates of the center point of the femoral head center in the currently acquired original image data of the femoral head can be obtained, achieving automatic identification of the femoral head center and improving identification efficiency and accuracy. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of a femoral head center identification method provided in an embodiment of this application; Figure 2 This is a schematic diagram of another method for identifying the femoral head center provided in an embodiment of this application; Figure 3 This is a schematic diagram of another femoral head center identification method provided in the embodiments of this application; Figure 4 This is a schematic diagram of a femoral head center identification process provided in an embodiment of this application; Figure 5 This is a schematic diagram of a femoral head center recognition device provided in an embodiment of this application; Figure 6 This is a schematic diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0014] The technical solution of this application will be described below through specific embodiments.

[0015] Reference Figure 1 The diagram illustrates a method for identifying the center of the femoral head according to an embodiment of this application, which may specifically include the following steps: S101. Obtain the femoral head image data to be identified.

[0016] It should be noted that the embodiments of this application can be applied to computer devices, that is, the executing entity of this method can be a computer device. By executing the various steps of the method provided in the embodiments of this application, the computer device can automatically process femoral head image data using an AI model and output accurate coordinates of the femoral head center point. The aforementioned computer device can be a desktop computer, a cloud server, or other similar devices; the embodiments of this application do not limit the type of computer device. The aforementioned AI model can be a femoral head center recognition model trained using multiple sample image data labeled with the femoral head center point.

[0017] The femoral head image data to be identified can be obtained from medical imaging equipment that performs imaging scans of the patient's femoral head region; that is, the femoral head image data to be identified is image data containing the patient's femoral head region.

[0018] In one example, the femoral head imaging data can be two-dimensional (2D) imaging data, such as X-ray images obtained by irradiating the femoral head region using an X-ray device; in another example, the femoral head imaging data can also be three-dimensional (3D) imaging data, such as computed tomography (CT) imaging data or magnetic resonance imaging (MRI) imaging data. This application does not limit the type of femoral head imaging data.

[0019] The aforementioned femoral head imaging data to be identified can be imaging data obtained from preoperative scanning for preoperative planning, or imaging data obtained from real-time scanning during the intraoperative stage that can be used to guide the execution of the surgical procedure. This application embodiment does not limit this.

[0020] In one possible implementation of this application embodiment, the femoral head image data to be identified may include multiple types of image data. For example, during the intraoperative stage, X-ray images are obtained through X-ray scanning, and the task at this time is to identify the femoral head center in the X-ray images. If the patient's CT images have also been acquired through a CT scan during the preoperative stage, then multiple types of image data are included when performing the current identification task. At this time, multiple types of image data can be registered. Thus, when executing S102, the femoral head image data input to the femoral head center identification model for femoral head center identification can be the registered image data. For example, in the aforementioned intraoperative stage, by registering the aforementioned two-dimensional X-ray image and three-dimensional CT image, the registered X-ray image can be used as the femoral head image data to be identified, thereby further improving the accuracy of femoral head center identification.

[0021] S102. The femoral head image data is identified using a femoral head center recognition model to obtain the model-predicted coordinates.

[0022] The femoral head center recognition model, or AI model, in this embodiment is pre-trained. Computer equipment can obtain this AI model by processing and training multiple sample image data.

[0023] The sample image data can be a large and diverse collection of femoral head images collected in advance, with the femoral head boundary (mask) and center point precisely marked in the sample image data. For three-dimensional sample image data, the center of the femoral head is usually the geometric center of the femoral head sphere, while for two-dimensional sample image data, the center may be the projection center.

[0024] Sample image data can be obtained through legal and compliant means such as hospital image archiving and communication systems (PACS), public datasets (such as PelvisXray, TotalSegmentator for CT / MRI), and related collaborative research. This application does not limit the method or source of obtaining sample image data.

[0025] In this embodiment of the application, for sample image data, the computer device can preprocess it and train the AI ​​model based on the preprocessed sample image data, thereby improving the accuracy and efficiency of model training.

[0026] The preprocessing of sample image data can include steps such as data standardization and data augmentation. Specifically, in the data standardization stage, computer equipment can resample multiple sample image data to the same resolution. For example, for three-dimensional CT or MRI data, the resampled resolution can be 1×1×1 mm. 3 For two-dimensional X-ray images, the resampled resolution can be 0.2 × 0.2 mm. 2 After data resampling, the computer can perform intensity normalization on multiple resampled image data with the same resolution, such as scaling the data to the [0,1] range. In the data augmentation stage, the computer can obtain a larger amount of image data through spatial transformation, intensity transformation, and modality-specific augmentation, thereby alleviating the problem of insufficient data volume and improving the generalization of the image data. The computer can also resample and normalize the augmented image data as described above.

[0027] Based on this, the computer device can divide the intensity-normalized sample image data into multiple datasets. For example, these datasets may include a training set, a validation set, and a test set. The data volume of the training set, validation set, and test set can be divided in a ratio of 70%, 15%, and 15%, respectively. The computer device can use these datasets to train an AI model, resulting in a femoral head center recognition model that can be used for automatic identification of the femoral head center.

[0028] In one possible implementation of this application, the sample image data can be divided according to patient ID. That is, any sample image data may include patient ID information. When dividing the dataset, sample image data collected from any patient should only exist in the same dataset, avoiding different image data of the same patient appearing in different datasets. For example, if patient A's image data can be entirely allocated to the training set, it should be ensured that patient A's image data does not appear in the validation set or test set. Similarly, if patient B's image data is entirely allocated to the validation set, it should be ensured that patient B's image data does not appear in the training set or test set.

[0029] Using the pre-defined datasets, computer devices can perform subsequent model training to obtain AI models.

[0030] In one possible implementation of this application embodiment, for two-dimensional sample image data, the AI ​​model trained on it will be mainly used to process the two-dimensional image data to be identified, and output the coordinates of the femoral head center point in the two-dimensional image, such as the coordinates of the femoral head center point in an X-ray image, i.e., the two-dimensional coordinates (x, y) of the femoral head center; for three-dimensional sample image data, the AI ​​model trained on it will be mainly used to process the three-dimensional image data to be identified, and output the coordinates of the femoral head center point in the three-dimensional image, such as the coordinates of the femoral head center point in a CT image or MRI image, i.e., the three-dimensional coordinates (x, y, z) of the femoral head center.

[0031] In another possible implementation of this application, the aforementioned AI model can be a cascaded model or a multi-task model. A cascaded model can consist of multiple models and is therefore also called a joint model. Each model in the cascaded model is used to perform different functions. For example, the femoral head segmentation model in the cascaded model can be used to segment the input femoral head image data and output the segmentation result; the femoral head center point localization model in the cascaded model can be used to further process the segmentation result output by the femoral head segmentation model to achieve key point detection and output predicted coordinates. A multi-task model can be a single model, which can simultaneously output image segmentation results and predicted coordinates. This application does not limit the specific architecture of the AI ​​model.

[0032] In this embodiment, after the femoral head image data to be identified is input into the AI ​​model, the AI ​​model can output model-predicted coordinates by processing and recognizing the image data. These model-predicted coordinates can be the center of the femoral head predicted by the AI ​​model in the segmented femoral head region. Typically, the model-predicted coordinates are coordinates on an image patch or a downsampled image. The computer device also needs to execute S103 to convert these coordinates into physical coordinates in the original image.

[0033] S103. Determine the coordinates of the femoral head center in the femoral head image data based on the coordinates predicted by the model.

[0034] In this embodiment, determining the femoral head center coordinates in the femoral head image data based on the model-predicted coordinates can mean converting the model-predicted coordinates output by the AI ​​model into coordinates in the original image of the femoral head. If the femoral head image data to be identified is a two-dimensional image, the converted femoral head center coordinates are also two-dimensional coordinates (x, y); if the femoral head image data to be identified is a three-dimensional image, the converted femoral head center coordinates are three-dimensional coordinates (x, y, z).

[0035] In the process of converting the model-predicted coordinates into the physical coordinates in the original image of the femoral head, it is necessary to combine the transformation matrix obtained by resampling, cropping, filling and other operations, which will not be elaborated in this application.

[0036] In this embodiment, an AI model capable of automatically identifying the femoral head center can be trained using multiple sample image data labeled with the femoral head center point. Thus, during the preoperative or intraoperative phase, when femoral head image data to be identified is acquired, the trained AI model can automatically process and identify the femoral head image data, outputting corresponding model-predicted coordinates. Based on these model-predicted coordinates, the coordinates of the center point of the femoral head center in the currently acquired original image data of the femoral head can be obtained, achieving automatic identification of the femoral head center and improving identification efficiency and accuracy.

[0037] Based on the foregoing embodiments, the femoral head center recognition method provided in this application will be described below using two different model architectures: a cascaded model and a multi-task model.

[0038] Reference Figure 2 This diagram illustrates another method for identifying the femoral head center provided in an embodiment of this application. Figure 2 The example shown uses a femoral head center recognition model as a cascaded model to illustrate this method. The method specifically includes the following steps: S201. Obtain the femoral head image data to be identified.

[0039] The femoral head image data to be identified can be two-dimensional or three-dimensional; or, when the acquired image data to be identified includes multiple types of image data, the image data can also include both two-dimensional and three-dimensional image data. The embodiments of this application do not limit the type and source of the femoral head image data to be identified.

[0040] S202. The femoral head image data is processed using the femoral head segmentation model to obtain a femoral head segmentation probability map.

[0041] The AI ​​model in this embodiment is a cascaded model, meaning the femoral head center recognition model in this embodiment is composed of multiple cascaded models, such as a femoral head segmentation model and a femoral head center point localization model. The femoral head segmentation model is used to implement image segmentation, while the femoral head center point localization model is used to solve the problem of key point detection.

[0042] In one possible implementation of this application embodiment, the femoral head segmentation model can be trained based on a U-Net, V-Net, or nnU-Net architecture. For example, a femoral head segmentation model mainly used for processing two-dimensional image segmentation tasks can be trained based on a 2D U-Net architecture, and a femoral head segmentation model mainly used for processing three-dimensional image segmentation tasks can be trained based on a 3D U-Net architecture.

[0043] After acquiring the femoral head image data to be identified, the femoral head image data can be input into the femoral head segmentation model, which will then perform image segmentation processing to obtain a femoral head segmentation probability map. This femoral head segmentation probability map can be used to indicate the probability that each pixel or voxel in the original image data belongs to the femoral head.

[0044] S203. Input the original image of the femoral head image data and the femoral head segmentation probability map into the femoral head center point localization model to obtain the model predicted coordinates output by the femoral head center point localization model.

[0045] In this embodiment, a femoral head center point localization model is cascaded with the femoral head segmentation model. This femoral head center point localization model can be a heatmap regression model. The input data of the aforementioned femoral head center point localization model includes the original image of the femoral head imaging data and the femoral head segmentation probability map output by the aforementioned femoral head segmentation model. After processing by the femoral head center point localization model, the femoral head center point annotation can be transformed into a Gaussian heatmap with the annotation point as the center and a standard deviation σ. By performing model learning and prediction on this Gaussian heatmap, the corresponding model-predicted coordinates can be obtained.

[0046] Specifically, the computer equipment can use a femoral head center point localization model to process the original image and femoral head segmentation probability map of the input femoral head image data, output a predicted heatmap, and determine the model's predicted coordinates based on the predicted heatmap. The model's predicted coordinates can be the coordinates of the peak position point (global maximum point) in the output predicted heatmap, which represents the femoral head center predicted by the model.

[0047] In one possible implementation of this application, when determining the model prediction coordinates based on the prediction heatmap, the computer device can first determine the global maximum point in the prediction heatmap and calculate the weighted centroid in the neighborhood of the global maximum point. This can obtain model prediction coordinates with sub-pixel or sub-voxel accuracy. That is, based on the initial femoral head center determined by the peak position of the prediction heatmap, the computer device can also use the average value of each pixel or voxel point in the prediction heatmap as the weight to calculate the weighted centroid in a small neighborhood near the peak position, thereby correcting the initially determined femoral head center and obtaining more accurate prediction coordinates of the femoral head center point.

[0048] S204. Based on the coordinates predicted by the model, determine the coordinates of the femoral head center in the femoral head image data.

[0049] S204 is similar to S103 in the previous embodiment, which is to convert the model prediction coordinates output by the AI ​​model into physical coordinates in the original image data. This is one of the key steps in clinical applications.

[0050] In this embodiment, by training a cascaded model, when acquiring femoral head image data to be identified, the femoral head segmentation model and the femoral head center point localization model in the cascaded model can respectively solve the problems of image segmentation and key point detection, thereby achieving automatic identification of the femoral head center point in the image data and improving the efficiency of femoral head center point identification. Specifically, the femoral head center point identification model can be trained using a heatmap regression model, which is more beneficial for improving localization accuracy and helps to obtain accurate femoral head center point identification results, thus improving the robustness of the model processing.

[0051] Reference Figure 3 This diagram illustrates yet another method for identifying the femoral head center provided in an embodiment of this application. Figure 3 The following is an introduction to this method using the femoral head center recognition model as an example of a multi-task model. The method specifically includes the following steps: S301. Obtain the femoral head image data to be identified.

[0052] S301 is similar to S101 and S102 in the previous embodiments, and will not be described again here.

[0053] S302. The encoder in the multi-task model is used to extract the image features from the femoral head image data.

[0054] The AI ​​model in this embodiment is a multi-task model, which can consist of an encoder and a decoder. Thus, a single multi-task model can simultaneously output image segmentation results and model prediction results.

[0055] In one possible implementation of this application embodiment, the encoder in the multi-task model can be a shared encoder, which can be used to extract common image features from the femoral head image data.

[0056] In one example, the shared encoder described above can be trained based on a convolutional neural network (CNN), a Transformer, or a hybrid CNN+Transformer architecture.

[0057] S303. The image features are processed using the decoder in the multi-task model, and the model-predicted coordinates are output.

[0058] In this embodiment, the decoder in the multi-task model can be a dual decoder, meaning that decoder-related functions are implemented by at least two decoders. For example, the dual decoder may include a segmentation decoder and a keypoint decoder. The segmentation decoder outputs the image segmentation result, and the keypoint decoder processes the image segmentation result to output a predicted heatmap.

[0059] Specifically, after extracting image features from the femoral head image data using a shared encoder, these image features can be transmitted to a segmentation decoder. The segmentation decoder performs feature segmentation on the image features to obtain a femoral head mask. Based on this, a keypoint decoder can process the femoral head mask and output a predicted heatmap. Computer equipment can then determine the model's predicted coordinates based on this predicted heatmap.

[0060] Similar to the processing of prediction heatmaps in the aforementioned cascaded models, when determining model prediction coordinates based on the prediction heatmap, the computer device can use the global maximum point in the prediction heatmap as the center of the femoral head and output the corresponding prediction coordinates. Alternatively, the computer device can also calculate the weighted centroid within a small neighborhood of the global maximum point to obtain model prediction coordinates with sub-pixel or sub-voxel accuracy.

[0061] S304. Based on the coordinates predicted by the model, determine the coordinates of the femoral head center in the femoral head image data.

[0062] S304 is similar to S103 and S204 in the previous embodiments, that is, converting the model prediction coordinates output by the AI ​​model into the physical coordinates in the original image of the image data, which will not be described in detail here.

[0063] This application embodiment trains a multi-task model, which can directly process the image data to be identified using a single model and output the model's predicted coordinates. Compared to cascaded models, the training process of the multi-task model is more complex, requiring a balance between the losses of image segmentation and keypoint detection tasks. However, in the process of automatically identifying the femoral head center, the multi-task model has higher computational efficiency. The image segmentation and keypoint detection tasks can share image features and mutually promote each other, further improving the accuracy of the final output model's predicted coordinates.

[0064] It should be noted that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0065] To facilitate understanding, a complete example will be used to introduce the femoral head center identification method provided in the embodiments of this application.

[0066] like Figure 4 The diagram shown is a schematic representation of the femoral head center identification process provided in an embodiment of this application. Figure 4 The process shown includes the entire process from data preparation and preprocessing, task definition and model architecture selection to model training and optimization, post-processing and result generation, and finally model evaluation and deployment. Figure 4 Each step of the process shown can be implemented using computer equipment. The following sections will describe each step in detail.

[0067] S401, Data Preparation and Preprocessing This step may include data collection, data standardization, data augmentation, data format processing, and dataset partitioning. Among these: During data collection, a large number of diverse and labeled CT, MRI, and X-ray images can be collected. For the collected images, the femoral head boundary and center point need to be accurately delineated during annotation. Collected images can come from hospital PACS systems, public datasets, and related research collaborations. All collected images can be used as sample image data for subsequent training.

[0068] During data standardization, computer equipment can resample all images to a uniform resolution. For example, for CT or MRI images, the resampled resolution can be 1×1×1 mm. 3 For X-ray images, the resampled resolution can be 0.2 × 0.2 mm. 2Then, the computer equipment can perform intensity normalization on each of the resampled images. For example, for X-ray images, after contrast stretching, they can be scaled to the [0,1] range; for MRI images, off-field correction can be applied, followed by Z-score normalization or histogram matching for normalization.

[0069] In data augmentation, spatial transformation, intensity transformation, and modality-specific enhancement can be used to increase the amount of sample image data and improve its generalization ability. Spatial transformation can be achieved through rotation, translation, scaling, elastic deformation, and mirror flipping; intensity transformation can be achieved by adding Gaussian noise, randomly adjusting brightness / contrast, and simulating blur; modality-specific enhancement can simulate different slice thicknesses and partial volume effects for CT or MRI images, and different projection angles for X-ray images.

[0070] During data format processing, CT or MRI images can be processed into 3D data blocks with a size of 64×64×64 or 128×128×128; X-ray images can be processed into 2D data blocks with a size of 512×512 or 1024×1024.

[0071] During the dataset partitioning process, the sample image data obtained from the aforementioned processing can be divided into training set, validation set, and test set according to the patient ID. The data volume ratio of each dataset can be 70%:15%:15%. In addition, it should be ensured that different images of the same patient do not appear in different sets.

[0072] S402, Task Definition and Model Architecture Selection The main purpose of this step is to select a suitable model architecture to solve the image segmentation and keypoint detection problems. In practical applications, cascaded models or multi-task model architectures can be selected for training. Specifically: (a) Cascade Model The cascaded model in this application embodiment may include a femoral head segmentation model and a femoral head center point localization model.

[0073] The femoral head segmentation model is used for image segmentation tasks. Its task is to process the input data (the image data to be identified) and output a binary mask of the femoral head region. The femoral head segmentation model can be trained using a 3D U-Net, V-Net, or nnU-Net architecture (CT / MRI images) or a 2D U-Net, DeepLabv3+, Attention U-Net, or U-Net++ architecture (X-ray images) as its basic architecture. In practical applications, the femoral head segmentation model can output a probability map (segmentation probability map) showing the probability that each pixel / voxel in each sample image belongs to the femoral head.

[0074] The task of the femoral head center point localization model is to take the original image and segmentation probability map as input and output the corresponding model-predicted coordinates. The femoral head center point localization model can be trained using a 3D CNN, 3D Heatmap Regression architecture (CT / MRI images) or a 2D CNN, 2D Heatmap Regression architecture (X-ray images) as the basic architecture. After processing by the femoral head center point localization model, the femoral head center point annotation can be transformed into a Gaussian heatmap with a standard deviation σ centered on the annotation point. By learning and predicting from this Gaussian heatmap, the corresponding model-predicted coordinates can be obtained.

[0075] (ii) Multi-task model The multi-task model in this embodiment needs to perform the task of simultaneously outputting a segmentation mask and model-predicted coordinates using a single model. This multi-task model can consist of a shared encoder and dual decoders. The shared encoder can be used to extract general image features, and its architecture can be implemented using 3D CNN, Transformer+CNN (CT / MRI images), or 2D CNN, Transformer (X-ray images). The dual decoders can include a segmentation decoder and a keypoint decoder. The segmentation decoder outputs the segmentation mask, and the keypoint decoder outputs the predicted heatmap or directly outputs the regression coordinates representing the femoral head center. Furthermore, during different processing stages, the decoders can fuse the features extracted by the shared encoder to achieve feature fusion.

[0076] Compared to cascaded models, the training process of multi-task models is more complex, requiring a balance between the losses of image segmentation and keypoint detection tasks. However, in the process of automatically identifying the femoral head center, multi-task models are more computationally efficient. Image segmentation and keypoint detection tasks can share image features and promote each other, further improving the accuracy of the model's predicted coordinates in the final output.

[0077] Furthermore, this application also provides a method for obtaining the coordinates of the femoral head center based on geometric calculations of the segmentation results. When the segmentation accuracy is high, direct geometric calculations can achieve automatic identification of the femoral head center without the need for additional training of a keypoint model.

[0078] Based on the above introduction, when pursuing the best accuracy and robustness, a cascaded model (segmentation + heatmap regression) can be used; when pursuing deployment efficiency, a multi-task model (segmentation + heatmap regression) can be used; and in scenarios where accuracy requirements are not high, direct geometric calculations can also be used to obtain the coordinates of the femoral head center.

[0079] S403, Model Training and Optimization In this step, the first step is to define the loss function and then use an optimizer to train the selected model architecture through learning rate scheduling. When defining the loss function, different loss functions can be designed for cascaded models and multi-task models. The corresponding loss functions can also differ for image segmentation and keypoint detection tasks during the training process of cascaded models.

[0080] The following training techniques can be used during model training: Batch normalization (BatchNorm): Widely used in 2D / 3D CNNs.

[0081] Deep supervision: Add auxiliary loss to the intermediate layers of U-Net-like networks to help gradient propagation.

[0082] Early Stopping: Monitor the validation set loss / metrics to prevent overfitting.

[0083] Weight initialization: Use pre-trained weights (ImageNet for 2D CNN / X-ray, MedicalNet / Kinetics for 3D CNN / CT / MRI) or He / Kaiming for initialization.

[0084] Mixed precision training: Use AMP to accelerate training and save GPU memory.

[0085] S404, Post-processing and Result Generation This step may include post-segmentation processing, keypoint post-processing, and coordinate transformation. Specifically: During post-segmentation processing, a threshold can be set to convert the segmentation probability map into a binary mask, which can then be processed using morphological operations. For example, small holes can be filled using closing operations, and isolated small points can be removed using opening operations. Furthermore, connected component analysis can be used to retain the largest connected region as the identified femoral head and remove other small regions.

[0086] During keypoint post-processing, Gaussian smoothing can be applied to the predicted heatmap, and the global maximum point in the heatmap can be found as the model prediction coordinates. Alternatively, based on determining the global maximum point in the heatmap, a weighted centroid can be calculated within its small neighborhood to further obtain sub-pixel / sub-voxel precision prediction coordinates.

[0087] During the coordinate transformation process, the model-predicted coordinates obtained above need to be converted into physical coordinates in the original image space of the image data to be identified. This process needs to be implemented in conjunction with the transformation matrices from preprocessing stages such as resampling, cropping, and padding.

[0088] S405, Model Evaluation and Deployment Application During the model evaluation phase, the performance of the trained AI model can be assessed using validation and test sets from sample image data. Specifically, evaluations can be conducted separately for image segmentation and keypoint detection tasks. In image segmentation evaluation, the overlap of segmented regions can be measured; if the overlap is ≥90%, the image segmentation result is considered excellent. Furthermore, the performance of image segmentation tasks can be evaluated by measuring metrics such as the furthest error of the segmentation boundary and surface consistency within a certain distance tolerance. In keypoint detection evaluation, the physical spatial distance between the center point predicted by the model and the actual center point labeled in the sample image data can be used as the core evaluation metric.

[0089] Furthermore, the trained AI model can be comprehensively evaluated by validating various key metrics on independent test sets. AI models that pass the evaluation can then be deployed for practical applications in medical settings.

[0090] In practical applications, when receiving femoral head image data to be detected, the previously trained AI model can be called. That is, the femoral head image data is input into the AI ​​model, and the automatic recognition function of the AI ​​model is used to efficiently output the accurate coordinates of the femoral head center point, thereby achieving fully automatic and accurate recognition of the femoral head center.

[0091] Reference Figure 5 The diagram illustrates a femoral head center identification device according to an embodiment of this application, which may specifically include an acquisition module 501, an identification module 502, and a determination module 503, wherein: The acquisition module 501 is used to acquire the femoral head image data to be identified; The recognition module 502 is used to recognize the femoral head image data using a femoral head center recognition model to obtain the model's predicted coordinates; wherein, the femoral head center recognition model is trained using multiple sample image data labeled with the femoral head center point; The determination module 503 is used to determine the coordinates of the femoral head center in the femoral head image data based on the coordinates predicted by the model.

[0092] In one possible implementation of this application embodiment, the femoral head center recognition model can be a cascaded model, which may include a femoral head segmentation model and a femoral head center point localization model. Specifically, the recognition module 502 can be used for: The femoral head segmentation model is used to process the femoral head image data to obtain a femoral head segmentation probability map. The femoral head segmentation probability map is used to indicate the probability that each pixel or each voxel in the original image of the femoral head image data belongs to the femoral head. The original image of the femoral head image data and the femoral head segmentation probability map are input into the femoral head center point localization model to obtain the model predicted coordinates output by the femoral head center point localization model.

[0093] In this embodiment of the application, the femoral head center point localization model can be a heatmap regression model, and the recognition module 502 can also be used for: The femoral head center point localization model is used to process the original image of the femoral head image data and the femoral head segmentation probability map to output a predicted heatmap; The model's predicted coordinates are determined based on the predicted heatmap.

[0094] In another possible implementation of this application embodiment, the femoral head center recognition model can be a multi-task model, which can be composed of an encoder and a decoder. The recognition module 502 can also be used for: The encoder in the multi-task model is used to extract image features from the femoral head image data; The image features are processed using the decoder in the multi-task model, and the model-predicted coordinates are output.

[0095] In this embodiment, the decoder can be a dual decoder, which may include a segmentation decoder and a keypoint decoder. The recognition module 502 can also be used for: The image features are segmented using the segmentation decoder to obtain a femoral head mask, and the femoral head mask is processed using the keypoint decoder to output a predicted heatmap. The model's predicted coordinates are determined based on the predicted heatmap.

[0096] In this embodiment of the application, the identification module 502 can also be used for: Determine the global maximum point in the predicted heatmap; The weighted centroid is calculated in the neighborhood of the global maximum point to obtain the model prediction coordinates with sub-pixel or sub-voxel accuracy.

[0097] In one possible implementation of this application embodiment, the apparatus may further include the following modules: The preprocessing module is used to resample multiple sample image data to the same resolution and perform intensity normalization processing on the multiple sample image data with the same resolution after resampling. The data partitioning module is used to divide the multiple sample image data after intensity normalization into multiple data sets, the multiple data sets including a training set, a validation set and a test set; wherein, each of the sample image data includes patient ID information, and the sample image data collected based on any patient exists only in the same data set; The model training module is used to train the femoral head center recognition model using multiple datasets.

[0098] In another possible implementation of this application embodiment, the femoral head imaging data may include various imaging data, and the device may further include the following modules: The registration module is used to register various types of image data; wherein the femoral head image data input to the femoral head center recognition model for recognition is the registered image data.

[0099] This application provides a femoral head center recognition device. This device can be the computer device described in the foregoing embodiments, or other devices or modules within devices capable of performing the relevant functions of the aforementioned computer device. Using this device, the steps in the foregoing method embodiments can be implemented.

[0100] As the apparatus embodiments are basically similar to the method embodiments, they are described in a relatively simple manner. For relevant details, please refer to the description in the method embodiment section.

[0101] Reference Figure 6 The diagram illustrates a computer device provided in an embodiment of this application. Figure 6As shown, the computer device 600 in this embodiment includes: a processor 610, a memory 620, and a computer program 621 stored in the memory 620 and executable on the processor 610. When the processor 610 executes the computer program 621, it implements the steps in the various embodiments of the above-described femoral head center recognition method, for example... Figure 1 Steps S101 to S103 are shown. Alternatively, when the processor 610 executes the computer program 621, it implements the functions of each module / unit in the above-described device embodiments, for example... Figure 5 The functions of modules 501 to 503 are shown.

[0102] For example, the computer program 621 can be divided into one or more modules / units, which are stored in the memory 620 and executed by the processor 610 to complete this application. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, which can be used to describe the execution process of the computer program 621 in the computer device 600. For example, the computer program 621 can be divided into an acquisition module, an identification module, and a determination module, with the specific functions of each module as follows: The acquisition module is used to acquire the femoral head image data to be identified; The recognition module is used to identify the femoral head image data using a femoral head center recognition model to obtain the model's predicted coordinates; wherein, the femoral head center recognition model is trained using multiple sample image data labeled with the femoral head center point; The determination module is used to determine the coordinates of the femoral head center in the femoral head image data based on the coordinates predicted by the model.

[0103] The computer device 600 may be a device capable of executing the steps in the foregoing method embodiments or realizing related functions. For example, the computer device 600 can train an AI model on sample image data marked with the femoral head center, which may be the femoral head center recognition model in the foregoing embodiments; and / or, the computer device 600 can automatically process the femoral head image data and output the corresponding femoral head center point coordinates. The aforementioned computer device 600 may be a desktop computer, a cloud server, or other devices. The computer device 600 may include, but is not limited to, a processor 610 and a memory 620. Those skilled in the art will understand that... Figure 6 This is merely one example of computer device 600 and does not constitute a limitation on computer device 600. It may include more or fewer components than shown, or combine certain components, or different components. For example, computer device 600 may also include input / output devices, network access devices, buses, etc.

[0104] The processor 610 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0105] The memory 620 can be an internal storage unit of the computer device 600, such as a hard disk or RAM of the computer device 600. The memory 620 can also be an external storage device of the computer device 600, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc., equipped on the computer device 600. Furthermore, the memory 620 can include both internal and external storage units of the computer device 600. The memory 620 is used to store the computer program 621 and other programs and data required by the computer device 600. The memory 620 can also be used to temporarily store data that has been output or will be output.

[0106] This application also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the femoral head center recognition method as described in the foregoing embodiments.

[0107] This application also discloses a computer-readable storage medium storing a computer program that, when executed by a computer, implements the femoral head center identification method as described in the foregoing embodiments.

[0108] This application also discloses a computer program product, including a computer program that, when run on a computer, causes the computer to execute the femoral head center recognition method described in the foregoing embodiments.

[0109] The embodiments described above are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for identifying the center of the femoral head, characterized in that, include: Obtain the femoral head imaging data to be identified; The femoral head center recognition model is used to identify the femoral head image data to obtain the model's predicted coordinates; wherein, the femoral head center recognition model is trained using multiple sample image data labeled with the femoral head center point; Based on the coordinates predicted by the model, the coordinates of the femoral head center in the femoral head image data are determined.

2. The method according to claim 1, characterized in that, The femoral head center recognition model is a cascaded model, which includes a femoral head segmentation model and a femoral head center point localization model. The step of using the femoral head center recognition model to identify the femoral head image data and obtain the model-predicted coordinates includes: The femoral head segmentation model is used to process the femoral head image data to obtain a femoral head segmentation probability map. The femoral head segmentation probability map is used to indicate the probability that each pixel or each voxel in the original image of the femoral head image data belongs to the femoral head. The original image of the femoral head image data and the femoral head segmentation probability map are input into the femoral head center point localization model to obtain the model predicted coordinates output by the femoral head center point localization model.

3. The method according to claim 2, characterized in that, The femoral head center point localization model is a heatmap regression model. The step of inputting the original image of the femoral head image data and the femoral head segmentation probability map into the femoral head center point localization model to obtain the model-predicted coordinates output by the femoral head center point localization model includes: The femoral head center point localization model is used to process the original image of the femoral head image data and the femoral head segmentation probability map to output a predicted heatmap; The model's predicted coordinates are determined based on the predicted heatmap.

4. The method according to claim 1, characterized in that, The femoral head center recognition model is a multi-task model, which consists of an encoder and a decoder. The step of using the femoral head center recognition model to identify the femoral head image data and obtain the model's predicted coordinates includes: The encoder in the multi-task model is used to extract image features from the femoral head image data; The image features are processed using the decoder in the multi-task model, and the model-predicted coordinates are output.

5. The method according to claim 4, characterized in that, The decoder is a dual decoder, comprising a segmentation decoder and a keypoint decoder. The process of using the decoder in the multi-task model to process the image features and output the model-predicted coordinates includes: The image features are segmented using the segmentation decoder to obtain a femoral head mask, and the femoral head mask is processed using the keypoint decoder to output a predicted heatmap. The model's predicted coordinates are determined based on the predicted heatmap.

6. The method according to claim 3 or 5, characterized in that, Determining the model prediction coordinates based on the predicted heatmap includes: Determine the global maximum point in the predicted heatmap; The weighted centroid is calculated in the neighborhood of the global maximum point to obtain the model prediction coordinates with sub-pixel or sub-voxel accuracy.

7. The method according to any one of claims 1 to 5, characterized in that, Before acquiring the femoral head imaging data to be identified, the following steps are also included: Multiple sample image data are resampled to the same resolution, and intensity normalization is performed on the multiple sample image data with the same resolution after resampling. The intensity-normalized sample image data are divided into multiple datasets, including a training set, a validation set, and a test set; wherein, each sample image data includes patient ID information, and the sample image data collected based on any patient exists only in the same dataset; The femoral head center recognition model is trained using multiple datasets.

8. The method according to any one of claims 1-5, characterized in that, The femoral head image data includes various types of image data. Before using a femoral head center recognition model to identify the femoral head image data and obtain the model's predicted coordinates, the following steps are also included: The various image data are registered; wherein the femoral head image data input to the femoral head center recognition model for recognition is the registered image data.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it causes the computer device to implement the method as described in any one of claims 1 to 8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is run, the method as described in any one of claims 1 to 8 is performed.

Images