A key point recognition method, device and storage medium

Abstract

This application provides a keypoint recognition method, device, and storage medium, relating to the field of computer science. The method includes: performing multi-resolution processing on a 3D target image to be recognized to obtain a first-resolution image and a second-resolution image; inputting the first-resolution image into a first probability prediction model to obtain first probability distribution information for each keypoint in the first-resolution image; determining the first spatial location of each keypoint based on the first probability distribution information; extracting local region information corresponding to each keypoint from the second-resolution image based on the first spatial location; and inputting the local region information into a second probability prediction model to obtain second probability distribution information for each keypoint, thereby determining the second spatial location of each keypoint based on the second probability distribution information. This solution requires no manual intervention throughout the entire process, has minimal subjective error, is simple to operate, and is time-efficient.

Description

Technical Field

[0001] This application relates to the field of computers, and more specifically, to a key point identification method, apparatus, and storage medium in the field of computers. Background Technology

[0002] Cone-beam computed tomography (CBCT) has been widely used in orthodontics, maxillofacial surgery, and other fields due to its advantages such as low radiation dose, low equipment cost, and fast imaging speed. CBCT images can provide high-resolution three-dimensional anatomical information, which is an important basis for cranial measurement and analysis, deformity assessment, and surgical simulation.

[0003] In existing technologies, doctors typically need to manually identify multiple key points in soft and hard tissues (such as the skull base, nasal tip, mandible, and soft tissue contours) on CBCT images, and then quantitatively measure deformity indicators such as distance and angle based on these key points. This process requires doctors to have a high level of radiological experience and three-dimensional spatial understanding. Different doctors will produce different results in identifying key points, resulting in large subjective errors and making it impossible to consistently and accurately identify key points. Furthermore, the entire process is cumbersome and time-consuming. Summary of the Invention

[0004] This application provides a key point identification method, device, and storage medium. The entire process is fully automatic without human intervention, with small subjective errors, simple operation, and short time consumption.

[0005] Firstly, this application provides a key point identification method, the method comprising: The three-dimensional target image to be identified is subjected to multi-resolution processing to obtain a first resolution image and a second resolution image, wherein the resolution of the first resolution image is lower than that of the second resolution image; The first resolution image is input into the first probability prediction model to obtain the first probability distribution information of each key point in the first resolution image, wherein the first probability prediction model is used to predict the probability distribution of key points from the global low resolution image. The first spatial location of each key point is determined based on the first probability distribution information. Based on the first spatial location, local region information corresponding to each key point is extracted from the second resolution image; The local region information is input into the second probability prediction model to obtain the second probability distribution information of each key point, so as to determine the second spatial location of each key point based on the second probability distribution information. The second probability prediction model is used to predict the probability distribution of key points from the local high-resolution image.

[0006] Secondly, this application provides a key point recognition device, which includes: The processing module is used to perform multi-resolution processing on the three-dimensional target image to be identified, to obtain a first resolution image and a second resolution image, wherein the resolution of the first resolution image is lower than the resolution of the second resolution image; The input module is used to input the first resolution image into the first probability prediction model to obtain the first probability distribution information of each key point in the first resolution image, wherein the first probability prediction model is used to predict the probability distribution of key points from the global low resolution image. The determination module is used to determine the first spatial location of each key point based on the first probability distribution information, and to input the local region information into the second probability prediction model to obtain the second probability distribution information of each key point, so as to determine the second spatial location of each key point based on the second probability distribution information, wherein the second probability prediction model is used to predict the probability distribution of key points from the local high-resolution image.

[0007] The extraction module is used to extract local region information corresponding to each key point from the second resolution image based on the first spatial location.

[0008] Thirdly, this application provides a model training method, the method comprising: Obtain a training sample set, which contains multiple 3D sample images labeled with the real spatial locations of each key point; The multi-resolution processing is performed on each 3D sample image to obtain a first-resolution sample image and a second-resolution sample image; First supervision information is generated based on the real spatial location of each key point in the first resolution sample image; The first resolution sample image is input into the first probability prediction model to be trained to obtain the first prediction probability distribution information, and the parameters of the first probability prediction model are updated based on the first prediction probability distribution information and the first supervision information. Based on the spatial location output by the first probability prediction model, a local sample block is cropped from the second resolution sample image, and second supervision information is generated based on the true spatial location of each key point in the local sample block. The local sample block is input into the second probability prediction model to be trained to obtain the second prediction probability distribution information, and the parameters of the second probability prediction model are updated based on the second prediction probability distribution information and the second supervision information.

[0009] Fourthly, this application provides a model training apparatus, which includes: The acquisition module is used to acquire a training sample set, which includes multiple three-dimensional sample images labeled with the real spatial locations of each key point. The processing module is used to perform the multi-resolution processing on each three-dimensional sample image to obtain a first resolution sample image and a second resolution sample image. The generation module is used to generate first supervision information based on the real spatial location of each key point in the first resolution sample image; The update module is used to input the first resolution sample image into the first probability prediction model to be trained, obtain the first prediction probability distribution information, and update the parameters of the first probability prediction model based on the first prediction probability distribution information and the first supervision information; and to input the local sample block into the second probability prediction model to be trained, obtain the second prediction probability distribution information, and update the parameters of the second probability prediction model based on the second prediction probability distribution information and the second supervision information. The cropping module is used to crop local sample blocks from the second resolution sample image based on the spatial location output by the first probability prediction model, and generate second supervision information based on the real spatial location of each key point in the local sample block.

[0010] Fifthly, embodiments of this application provide an electronic device, including: a memory, a processor, and a communication interface; wherein, the memory stores executable code, and when the executable code is executed by the processor, the processor performs the method as described in the first or third aspect.

[0011] In a sixth aspect, embodiments of this application provide a computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the methods described in the first or third aspect.

[0012] In a seventh aspect, embodiments of this application provide a computer program product, the computer program product including a computer program, which, when executed by a processor, can implement the method described in the first or third aspect.

[0013] The keypoint recognition scheme provided in this application performs multi-resolution processing on the 3D target image to be recognized, obtaining a low-resolution first-resolution image and a high-resolution second-resolution image, balancing global view and local details, computational efficiency, and accuracy limits. By inputting the first-resolution image into a first probability prediction model, the first probability distribution information of each keypoint is obtained, transforming the localization problem into probability distribution learning, thus improving robustness to fuzzy soft tissues. The first spatial position of each keypoint is determined based on the first probability distribution information, achieving sub-voxel-level coarse localization and providing accurate priors for subsequent local cropping. Then, based on the first spatial position, local region information corresponding to each keypoint is extracted from the second-resolution image, focusing the model's attention on the high-resolution key region, significantly reducing computation and eliminating background interference. Finally, by inputting the local region information into the second probability prediction model, the second probability distribution information of each keypoint is obtained to determine the final spatial position, achieving precise keypoint localization. The entire process is fully automated without manual intervention, with minimal subjective error, simple operation, and short processing time. Attached Figure Description

[0014] Figure 1 This is a flowchart of a key point identification method provided in an embodiment of this application; Figure 2 This is a flowchart of a model training method provided in an embodiment of this application; Figure 3 This is a comparative diagram of the key point identification method provided in the embodiments of this application and the traditional annotation method for key point identification; Figure 4 This is a schematic diagram of the structure of a key point recognition device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a model training device provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0015] The technical solutions in this application will be clearly and thoroughly described below with reference to the accompanying drawings. In the description of the embodiments of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. "And / or" in the text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, in the description of the embodiments of this application, "multiple" refers to two or more than two.

[0016] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0017] The following is a description of the terms used in the embodiments of this application: Voxel spacing: refers to the physical distance between the centers of two adjacent voxels, which determines the spatial resolution of the image. A voxel is the smallest basic unit in a 3D image; one voxel represents the image information within a small cube (or cuboid).

[0018] Cone-beam computed tomography (CBCT) has been widely used in orthodontics, maxillofacial surgery, and other fields due to its advantages such as low radiation dose, low equipment cost, and fast imaging speed. CBCT images provide high-resolution three-dimensional anatomical information, which is an important basis for cranial measurement analysis, deformity assessment, and surgical simulation. Currently, doctors usually need to manually mark multiple key points of soft and hard tissues (such as the skull base, nasal tip, mandible, soft tissue contours, etc.) on CBCT images and perform quantitative measurements of deformity indicators such as distance and angle based on these key points. This process requires doctors to have a high level of radiological experience and three-dimensional spatial understanding. Different doctors will have different results in marking key points, resulting in large subjective errors and making it impossible to identify key points stably and accurately. Moreover, the entire process is cumbersome and time-consuming. In view of this, this application provides a key point identification method.

[0019] Figure 1 This is a flowchart illustrating a key point identification method provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps: Step 101: Perform multi-resolution processing on the 3D target image to be identified to obtain a first resolution image and a second resolution image. The resolution of the first resolution image is lower than that of the second resolution image.

[0020] Step 102: Input the first resolution image into the first probability prediction model to obtain the first probability distribution information of each key point in the first resolution image, wherein the first probability prediction model is used to predict the probability distribution of key points from the global low resolution image.

[0021] Step 103: Determine the first spatial location of each key point based on the first probability distribution information.

[0022] Step 104: Based on the first spatial location, extract the local region information corresponding to each key point from the second resolution image.

[0023] Step 105: Input the local region information into the second probability prediction model to obtain the second probability distribution information of each key point, so as to determine the second spatial location of each key point based on the second probability distribution information. The second probability prediction model is used to predict the probability distribution of key points from the local high-resolution image.

[0024] In practical applications, the first step is to acquire a 3D target image (which can be a head CBCT image, hand CBCT image, foot CBCT image, etc.) of the key points to be identified. Then, multi-resolution processing is performed on the 3D target image to obtain a first-resolution image and a second-resolution image. The specific processing procedure is as follows: The 3D target image is resampled to an isotropic first voxel spacing to obtain the first-resolution image; the original resolution at a second voxel spacing is retained to obtain the second-resolution image, where the first voxel spacing is greater than the second voxel spacing. For ease of understanding, the following explanation uses a head CBCT image as an example of a 3D target image: In practice, the anisotropic high-resolution image of the 3D target image (e.g., a voxel spacing of 0.22×0.22×0.3125 mm) can be resampled into an isotropic image (i.e., an image with consistent voxel spacing) with a larger voxel spacing (e.g., a voxel spacing of 1.4×1.4×1.4 mm). It should be understood that the image size will be significantly reduced after resampling. For example, assuming the original 3D target image has a voxel grid size of 1024×1024×1000, the resampled voxel grid size could be 160×160×160. Simultaneously, the original CBCT image's voxel spacing and size are completely preserved without any resampling to obtain the second-resolution image. In summary, the first-resolution image refers to the low-resolution image used for global coarse localization, typically resampled to isotropic to reduce computational load and expand the receptive field. The second-resolution image, on the other hand, is used to preserve the original fine voxel spacing for subsequent local refinement to capture minute features.

[0025] By resampling the 3D target image to an isotropic first voxel spacing to obtain a first resolution image, the spatial scale inconsistency caused by the anisotropy of the original CBCT image due to different acquisition devices is eliminated. Furthermore, a larger first voxel spacing can significantly reduce the image data volume (e.g., reducing the voxel grid size from 1024×1024×1000 to 160×160×160), thereby significantly reducing computational resources and video memory usage, enabling the processing of large-field-of-view 3D images under limited hardware conditions. Simultaneously, this low-resolution image (i.e., the first resolution image) has a large receptive field, capable of capturing overall facial contextual information, providing stable global spatial constraints for subsequent coarse localization of key points. By retaining the original resolution at the second voxel spacing of the 3D target image, a second resolution image is obtained, which can fully preserve the fine texture of soft and hard tissue boundaries, minute anatomical structures, and sub-voxel-level spatial details (e.g., the slight undulations of the nasal tip in soft tissue, the precise location of tooth root apexes, etc.). These details are smoothed out or even lost in low-resolution images, while high-resolution images provide a crucial information foundation for the second-stage local refinement, improving the accuracy of key points subsequently obtained based on the second probability prediction model.

[0026] In summary, by adopting a multi-resolution processing approach of "large-pitch low-resolution global coarse localization + small-pitch high-resolution local fine localization", it not only ensures computational efficiency and global perception capabilities, but also breaks through the bottleneck of single-resolution models in detail localization, achieving high-precision and fully automatic recognition of key points of soft and hard tissues in the head.

[0027] Next, the first-resolution image is input into the first probabilistic prediction model to obtain the first probability distribution information of each keypoint. The first probabilistic prediction model can be understood as a first-level 3D U-Net (also called a 3D U-Net). During training, it takes a low-resolution image as input and outputs a 3D Gaussian heatmap (each keypoint corresponds to a channel heatmap). The value of each voxel in the 3D Gaussian heatmap represents the probability of the keypoint being located at the current voxel position (the probability is highest at the peak and decreases Gaussian outwards). For ease of understanding, for example, assuming the voxel grid size of the first-resolution image input to the first probabilistic prediction model is 160×160×160, the tensor shape of the 3D Gaussian heatmap output by the first probabilistic prediction model will be 74×160×160×160, where 74 represents the number of channels. For the key point of "soft tissue nasal tip", in the corresponding three-dimensional Gaussian heat map channel, the value of the voxel where the nasal tip is located (assuming the coordinates are (80, 90, 75)) is 0.98, and the value of the surrounding voxels gradually decreases to below 0.1, which forms the first probability distribution information.

[0028] After obtaining the first probability distribution information, the value of each voxel in the corresponding three-dimensional Gaussian heatmap can be normalized to obtain normalized probability distribution data. Using the normalized probability distribution data as weights, the weighted average value of each voxel position is calculated, and the weighted average value is used as the coordinates of the spatial position.

[0029] In practice, all voxel values ​​in the 3D Gaussian heatmap are normalized, which can be achieved using the Softmax function.

[0030] in, It is the raw response value of the 3D Gaussian heatmap at location (i, j, k). This is the normalized probability. After processing with the Softmax function, the sum of the probabilities of all voxels is 1, forming a true three-dimensional probability distribution. The final normalized probability distribution data can be denoted as: ,and

[0031] Here, D, H, and W represent three spatial dimension parameters: depth, height, and width, respectively.

[0032] Next, the spatial coordinates of each voxel (which can be the voxel's index (i, j, k) or its actual physical coordinates) are multiplied by its corresponding normalized probability, and then summed over all voxels to obtain the desired coordinates (i.e., the coordinates of the first spatial position):

[0033] It should be noted that this weighted average is the mathematical expectation of the probability distribution, i.e., the center of the "probability mass" in the 3D Gaussian heatmap. When the probability mass exhibits a unimodal Gaussian shape, the expected position approximates the peak position. Compared to the traditional method of directly taking the maximum value in the 3D Gaussian heatmap, this method has advantages such as sub-voxel accuracy and differentiability.

[0034] After determining the coordinates of the first spatial location based on the above method, the local region information corresponding to each key point can be extracted from the second resolution image based on the first spatial location. The specific process is as follows: map the first spatial location to the voxel coordinate system of the second resolution image to obtain the high-resolution center coordinates; with the high-resolution center coordinates as the center, crop out a three-dimensional image block of a preset size from the second resolution image as local region information.

[0035] In practical applications, it should be noted that the voxel coordinate systems of the first-resolution image and the second-resolution image are different. The coordinates predicted in the first-resolution image are in voxel units of the low-resolution image and cannot be directly used to index the high-resolution image. Therefore, embodiments of this application use coordinate mapping to convert low-resolution coordinates into physical coordinates or voxel coordinates in the high-resolution image. The mapping formula can be:

[0036] in, It is the voxel spacing vector of the low-resolution image relative to the original physical space. These are low-resolution voxel coordinates. After calculating the physical coordinates, divide them by the voxel spacing of the high-resolution image to obtain the high-resolution voxel coordinates. .

[0037] Subsequently, in the high-resolution image (i.e., the second-resolution image), the center coordinates obtained by mapping (i.e., high-resolution voxel coordinates) are used. Centered on a keypoint, a fixed-size 3D cube region is cropped out, which is a local, high-resolution 3D image patch. This patch contains only a limited area around the keypoint, but retains the fine texture and edge information of the original image. It is important to note that image boundaries must be considered during the cropping process; if the center is close to the edge, appropriate padding can be applied.

[0038] By mapping the first spatial location to the voxel coordinate system of the second-resolution image, high-resolution center coordinates are obtained. Using these high-resolution center coordinates as the center, a 3D image patch of a preset size is cropped from the second-resolution image as local region information. This achieves a "global-to-local" attention transfer, reducing computational load. Furthermore, the mapping is based on physical coordinates rather than simple pixel multiples, ensuring geometric consistency between images of different resolutions and avoiding error accumulation.

[0039] After obtaining the above local area information, the local area information can be input into the second probability prediction model to obtain the second probability distribution information of each key point (for the specific process, please refer to the process of obtaining the first probability distribution information above, which will not be repeated here). The second spatial position of each key point can be determined based on the second probability distribution information (for the specific process, please refer to the process of obtaining the first spatial position above, which will not be repeated here).

[0040] To facilitate understanding, a specific embodiment is provided below to illustrate this solution in detail: The 3D target image to be identified (voxel grid size is 1024×1024×1000, voxel spacing is...) Multi-resolution processing is performed. Specifically, the 3D target image is resampled to an isotropic low-resolution space, and the voxel spacing is uniformly adjusted to 1.4×1.4×1.4 mm. After resampling, the voxel grid size becomes 160×160×160, which yields the first-resolution image. Simultaneously, the original high-resolution data is retained, i.e., the voxel spacing remains constant. Without downsampling, a second-resolution image is obtained.

[0041] Standardization preprocessing is performed on the resampled first-resolution image: First, it is cropped to a fixed voxel grid size of 192×192×192 (by mirroring or boundary expansion, all sample dimensions are unified). Then, pixel value normalization is performed: the window width and window level truncation threshold are set to ±1024, and pixel values ​​greater than 1024 are set to 1024, and those less than -1024 are set to -1024; then, the truncated pixel values ​​are divided by the mean of the dataset (e.g., the mean is approximately 500), and finally the data distribution is normalized to the interval [0, 1].

[0042] Next, the standardized preprocessed image at the first resolution is input into the first probabilistic prediction model. This first probabilistic prediction model adopts a 3D U-Net architecture, and its specific structure is as follows: The encoder consists of 6 downsampling layers. Each encoder layer comprises two 3×3×3 3D convolutional layers connected in series with a ReLU activation function, followed by a 2×2×2 max pooling layer. After 6 pooling operations, the feature map size of the deepest layer is reduced to 3×3×3.

[0043] Decoder: Contains 6 upsampling layers. Each decoder layer is upsampled by a transposed convolution, then concatenated with the feature map of the corresponding encoder layer along the channel dimension via skip connections. This is followed by feature fusion through two 3×3×3 convolutional layers and a ReLU activation function. This skip connection structure allows the fusion of shallow high-frequency detail information (such as edges and textures) extracted by the encoder with deep low-frequency semantic information from the decoder, thus solving the gradient vanishing problem in deep networks.

[0044] Output layer: The last layer uses 1×1×1 convolutions to map the number of channels to the number of keypoints (C=74). The output tensor dimension is 74 × 192 × 192 × 192.

[0045] This first probabilistic prediction model does not directly output coordinates, but instead outputs a three-dimensional Gaussian heatmap for each keypoint. For example, for the keypoint "soft tissue nose tip", in its corresponding output channel 1×192×192×192, a high-response region appears near the approximate location of the nose tip (e.g., voxel coordinates (96, 100, 80)), with a central peak value of approximately 0.85, gradually decreasing outwards to 0.05, forming a "halo"-shaped probability distribution.

[0046] Subsequently, the 74 three-dimensional Gaussian heatmaps output by the first probabilistic prediction model were decoded using a soft-argmax operation to obtain the coarse three-dimensional coordinates of each keypoint. Taking the "soft tissue tip of the nose" as an example, assuming that it is normalized and weighted averaged using the formula in the above embodiment, the coarse coordinates of the "soft tissue tip of the nose" (coordinates of the first spatial position in low-resolution voxels) are (96.3, 100.1, 80.4), which are sub-voxel precision.

[0047] After obtaining the coordinates of the first spatial location, local region information is extracted from the second resolution image, which is specifically divided into two sub-steps: coordinate mapping and cropping. For the coordinate mapping part: First, the low-resolution voxel coordinates are converted into physical coordinates (with a voxel spacing of 1.4 × 1.4 × 1.4 mm): x = 96.3 × 1.4 = 134.82 mm; y = 100.1 × 1.4 = 140.14 mm; z = 80.4 × 1.4 = 112.56 mm.

[0048] Next, the physical coordinates are converted to high-resolution voxel coordinates (with a voxel spacing of 1). mm): X=134.82 / ≈612.82; Y=140.14 / ≈637.00; Z=112.56 / ≈360.19.

[0049] That is, the final high-resolution center coordinates are (612.82, 637.00, 360.19).

[0050] For the cropped portion: Using the high-resolution center coordinates (613, 637, 360) as the center (rounded to the nearest integer), a 3D image patch of a preset size of 64 × 64 × 64 voxels is cropped from the second-resolution image. Cropping boundary calculation: X direction: from 613 - 32 = 581 to 613 + 31 = 644 (64 voxels in total); Y direction: from 637 - 32 = 605 to 637 + 31 = 668; Z direction: from 360 - 32 = 328 to 360 + 31 = 391.

[0051] If the cropping boundary exceeds the range of the high-resolution image (original high-resolution image dimensions 909 × 909 × 480), zero-padding is applied to the excess portion. This results in a 64 × 64 × 64 high-resolution local image patch, where the "soft tissue nasal tip" is roughly located at the center of the patch. The high-resolution local image patch contains detailed anatomical information such as the skin texture around the nasal tip, the edges of the nostrils, and the soft tissue contours.

[0052] Next, high-resolution local image patches are input into the second probabilistic prediction model. This second probabilistic prediction model shares the same 3D U-Net architecture as the first probabilistic prediction model, but has a smaller input size (64³), and the network depth can be adjusted appropriately (e.g., the encoder part only needs 4 downsampling layers, resulting in a final feature map size of 4×4×4). The output of the second probabilistic prediction model is also a three-dimensional Gaussian heatmap with dimensions of 74 × 64 × 64 × 64.

[0053] For the "soft tissue nasal tip," the corresponding output heatmap exhibits a sharp, unimodal distribution within a local block. Using the same method described above, fine coordinates within the local block are decoded, for example (32.1, 31.8, 32.4) (in local block voxels). These local coordinates are then transformed back to the global high-resolution voxel coordinate system. X final =(581+32.1)=613.1; Y final =(605+31.8)=636.8; Z final =(328+32.4)=360.4.

[0054] The (613.1, 636.8, 360.4) is the final precise spatial location of the "soft tissue tip of the nose".

[0055] Repeat the above operations to perform local cropping and fine regression on all 74 key points (including the skull base, orbit, nose, maxilla, mandible, zygomatic bone, teeth and soft tissue contours, etc.) to obtain the second spatial position of all key points.

[0056] Based on the above, the key point recognition method provided in this application performs multi-resolution processing on the 3D target image to be recognized, obtaining a low-resolution first-resolution image and a high-resolution second-resolution image, balancing global field of view with local details, computational efficiency, and accuracy limits. By inputting the first-resolution image into a first probability prediction model, the first probability distribution information of each key point is obtained, transforming the localization problem into probability distribution learning, thus improving robustness to fuzzy soft tissues. The first spatial position of each key point is determined based on the first probability distribution information, achieving sub-voxel-level coarse localization and providing accurate priors for subsequent local cropping. Then, based on the first spatial position, local region information corresponding to each key point is extracted from the second-resolution image, focusing the model's attention on the high-resolution key region, significantly reducing computational load and eliminating background interference. Finally, by inputting the local region information into the second probability prediction model, the second probability distribution information of each key point is obtained to determine the final spatial position, achieving precise key point localization. The entire process is fully automatic without manual intervention, with minimal subjective error, simple operation, and short processing time.

[0057] Figure 2 This is a flowchart illustrating a model training method provided in an embodiment of this application. This model training method is used to train the aforementioned first probability prediction model and second probability prediction model. Figure 2 As shown, the model training method includes the following steps: Step 201: Obtain the training sample set, which contains multiple 3D sample images labeled with the real spatial locations of each key point.

[0058] Step 202: Perform multi-resolution processing on each 3D sample image to obtain a first-resolution sample image and a second-resolution sample image.

[0059] Step 203: Generate first supervision information based on the real spatial location of each key point in the first resolution sample image.

[0060] Step 204: Input the first resolution sample image into the first probability prediction model to be trained to obtain the first prediction probability distribution information, and update the parameters of the first probability prediction model based on the first prediction probability distribution information and the first supervision information.

[0061] Step 205: Based on the spatial location output by the first probability prediction model, crop out local sample blocks from the second resolution sample image, and generate second supervision information based on the real spatial location of each key point in the local sample block.

[0062] Step 206: Input the local sample block into the second probability prediction model to be trained to obtain the second prediction probability distribution information, and update the parameters of the second probability prediction model based on the second prediction probability distribution information and the second supervision information.

[0063] In practical applications, the training sample set is the foundation of deep learning models. To facilitate understanding, the process of obtaining the training sample set is illustrated below with an example: Assume the 3D sample image is a head CBCT scan, and the true spatial locations of all key points have been manually annotated by experienced clinicians or anatomists. These key points total 74, covering anatomical structures crucial for diagnosis, including the skull base, orbit, nose, maxilla, mandible, zygomatic bone, teeth (crown and apex), and soft tissue contours. During annotation, the physician typically clicks precisely point by point in the original high-resolution image, with the coordinates of each key point accurate to the sub-voxel level. After annotation, these coordinates are used as supervisory information for subsequent training.

[0064] Next, the same multi-resolution processing as in the inference phase is performed on each 3D sample image to ensure consistent data distribution between training and inference. After obtaining the first-resolution sample image and the second-resolution sample image through multi-resolution processing, supervision information for training the first probabilistic prediction model is generated based on the true spatial location of each keypoint in the first-resolution sample image (i.e., the coordinates marked by the doctor in the original high-resolution image). Since the output of the first-level model is a 3D Gaussian heatmap, the supervision information also needs to be a Gaussian heatmap of the same size. In practice, the true coordinates of each keypoint (first transformed to the voxel coordinate system of the low-resolution image) can be mapped to a 3D Gaussian heatmap. The 3D Gaussian heatmap has a maximum response value at the true coordinates (usually set to 1), which decays with increasing distance according to a Gaussian function.

[0065] Then, the first-resolution sample images are input into the first probabilistic prediction model to be trained to obtain the first predicted probability distribution information, namely, the predicted 74 3D Gaussian heatmaps. The loss function between the predicted 3D Gaussian heatmaps and the supervision information is calculated. Specifically, based on the difference between the predicted probability distribution information and the supervision information, adaptive wing loss and mean squared error loss are calculated. The adaptive wing loss measures the difference between the predicted probability distribution and the true probability distribution at each voxel position in the 3D Gaussian heatmap, and the mean squared error loss is the average of the squares of the differences between the predicted and true values ​​at all voxel positions. Based on the adaptive wing loss and the mean squared error loss, the model parameters are updated.

[0066] In practice, the total loss function is calculated according to the following formula, and the model parameters are updated based on the total loss function:

[0067] Where y represents the supervisory information, For the predicted 3D Gaussian heatmap, To set a balance factor (assumed to be 0.5). This represents the Adaptive Wing Loss (AWing). This represents the mean squared error loss. Specifically, The calculation formula is as follows:

[0068] Where w is the scaling factor, The segmentation threshold, , for , for All the parameters mentioned above are known data. This adaptive wing loss provides a smooth and sensitive gradient in the small error region, employs a linear penalty in the large error region, and is robust to noise and outliers.

[0069] It's important to note that the adaptive wing loss can adaptively adjust the loss weights, focusing on boundary regions and important features, thereby enhancing the model's accuracy and stability. It also provides a smooth gradient, which is beneficial for stable convergence in complex data, especially in scenarios with noise and outliers. Each marker point has a corresponding Gaussian heatmap label and a predicted Gaussian heatmap. The adaptive wing loss and mean squared error loss are calculated between these two. After calculating both losses, the gradient of the learnable parameters is calculated using backpropagation, and then the learnable parameters are updated using a stochastic gradient descent algorithm. The adaptive wing loss is a loss function specifically designed for heatmap regression tasks (such as facial landmark detection and fine segmentation). It automatically changes the shape of the loss curve based on whether a pixel is a label, resulting in faster convergence and robustness to noise and outliers. Addressing the characteristics of blurred soft tissue boundaries and low signal-to-noise ratio in CBCT images, the adaptive wing loss function can automatically adjust the weights based on the difference between the true and predicted values. It provides a smooth gradient when the error is small and focuses on boundary regions when the error is large, effectively solving the problem of traditional mean squared error loss's sensitivity to outliers.

[0070] After calculating the total loss function based on the above formula, the gradients of the parameters of each layer are calculated through backpropagation, and the weights of the first probabilistic prediction model are updated using the optimizer. This process is repeated until the first probabilistic prediction model converges on the validation set.

[0071] Next, the trained first probabilistic prediction model is used to infer the predicted spatial locations (i.e., coarse coordinates) of each keypoint in the first-resolution sample image, obtaining these coarse coordinates. Then, these coarse coordinates are mapped to the voxel coordinate system of the second-resolution sample image, and high-resolution local sample blocks of fixed size are cropped with each mapped coordinate as the center. Simultaneously, second-level supervision information is generated based on the true spatial locations of each keypoint in these local sample blocks (original labeled coordinates transformed to the local block coordinate system). The second-level supervision information also takes the form of a 3D Gaussian heatmap, but the heatmap size is consistent with the local block size, and the relative position of the true coordinates within the local block determines the center of the 3D Gaussian heatmap.

[0072] Then, the cropped local sample blocks are input into the second probabilistic prediction model to be trained. This model also uses a 3D U-Net architecture, but with a smaller input size, allowing for a reduction in network depth (e.g., a 4-layer encoder-decoder). The second probabilistic prediction model outputs a second predicted probability distribution, i.e., a predicted heatmap for each key point in the local block. Then, the difference between the predicted heatmap and the second supervised heatmap is calculated using the loss function method described above. Since the second probabilistic prediction model focuses on local regions, its learning objective is more refined: precise localization within millimeter or even sub-millimeter ranges is required. After loss calculation, the parameters of the second probabilistic prediction model are updated through backpropagation. During training, the first probabilistic prediction model is typically trained until convergence, and then its parameters are fixed, and the local sample blocks generated are used to train the second probabilistic prediction model. A joint training strategy can also be used, but to avoid training instability, staged training is more commonly used in practice.

[0073] By using the above model training method, a two-stage deep learning model of "global coarse localization → local fine-tuning" is essentially constructed. First, the first-level model (i.e., the first probabilistic prediction model) is trained using low-resolution images, enabling it to efficiently and robustly detect global keypoint candidate regions. Then, based on the prediction results of the first-level model, detailed local regions are automatically cropped from the high-resolution image to train the second-level model (i.e., the second probabilistic prediction model), focusing on high-precision localization. This training strategy not only fully utilizes the advantages of multi-resolution data but also avoids the enormous computational overhead of training directly on the entire high-resolution image.

[0074] To facilitate understanding, a specific example is provided below to illustrate the model training process: Because the original input 3D sample image has a large dimension, it is usually in Left and right, with obvious anisotropy, and the voxel spacing is approximately To accommodate subsequent algorithm processing and eliminate differences introduced by device acquisition, the 3D sample images were first resampled. Specifically, the resampling process generated two datasets with different resolutions: first, the 3D sample images were resampled to an isotropic low-resolution space, with the voxel spacing uniformly adjusted. First, it is used for subsequent global feature extraction and coarse localization; second, it preserves the original high-resolution data, that is, the voxel spacing remains at a certain value. This is used for subsequent refined regression.

[0075] For the resampled data, a standardization preprocessing operation is performed. First, to reduce computation and focus on the target region, the image is cropped to a fixed dimension. Then, pixel value normalization is performed: a truncation threshold is set, and pixel values ​​greater than 1024 are set to 1024, and those less than -1024 are set to -1024. Then, the truncated pixel values ​​are divided by the mean of the dataset, and finally the data distribution is normalized to the interval [0, 1].

[0076] For both the first and second probability prediction models to be trained, 3D U-Net can be used. This network specifically consists of a 6-layer encoder and a 6-layer decoder. Each encoder layer comprises two 3×3 convolutional layers connected in series with a ReLU activation function, followed by a 2×2 max-pooling layer for downsampling. Each decoder layer consists of an upsampling convolutional layer (Deconvolution / Transpose Conv) that first performs upsampling, then concatenates the feature maps of the corresponding encoder layer along the channel dimension, and finally fuses the features through two 3×3 convolutional layers and a ReLU activation function.

[0077] The input to the first probabilistic prediction model is the preprocessed image. The output is the prediction result. ,in, The number of marker points is denoted by . During the training phase, both adaptive wing loss and mean squared error loss are used as the loss function to enhance the fitting ability to the target region. The AdamW optimizer is selected. After calculating both loss components in each iteration, the gradients of the learnable parameters are calculated using the backpropagation algorithm, and the parameters are updated using the AdamW optimizer until the model converges.

[0078] The second probabilistic prediction model does not directly output (x, y, z) coordinates. Instead, it outputs a tensor of dimension C×D×H×W, where C is the number of keypoint categories (e.g., 74). Each channel in this tensor corresponds to an independent 3D heatmap, intuitively reflecting the probability distribution of the keypoint's presence at its current spatial location, denoted as . Where n represents the number of markers to be detected. Each output channel... A 3D Gaussian heatmap corresponds to a marker point, with peaks representing the probability distribution of that point. By performing an extremum search on this heatmap, a rough 3D coordinate of the marker point is obtained. Through this mechanism, the network no longer performs rigid numerical fitting but learns "where this point is most likely to be found." Compared to directly regressing coordinates, 3D Gaussian heatmap regression provides richer gradient information, especially in regions with blurred soft tissue boundaries. The probability distribution effectively smooths out noise interference, providing a more stable input for subsequent extremum point searches.

[0079] To provide a more intuitive understanding, the following provides a comparison of application data in real-world scenarios between a standard 3DUnet and the cascaded model provided in this solution (i.e., a first probability prediction model as the first-level model + a second probability prediction model as the second-level model):

[0080] Wherein, "MRE±SD" represents the mean ± standard deviation of the average radial error calculated on all test samples. It can be seen that the cascaded model provided in this application has an MRE of nearly 3 mm higher than that of ordinary 3D U-Net for direct keypoint prediction across all keypoints, and shows significant improvements in both MRE value and successful detection rate in both hard and soft tissues.

[0081] Furthermore, Figure 3 This diagram illustrates a comparison between the key point identification method provided in this application and traditional annotation methods for key point identification. Figure 3 In the diagram, green represents the annotations of clinicians, and red represents the identification results of this application. It can be seen that the key point identification results of this application are more accurate.

[0082] The following is a summary of the beneficial effects of this plan: 1. By introducing an adaptive feature enhancement mechanism and a multi-stage keypoint localization strategy, the accuracy and robustness of keypoint detection in low-contrast soft tissue environments are significantly improved. Traditional methods are prone to localization errors or missed detections due to insufficient feature response when facing soft tissue regions with weak gray-level differences and blurred boundaries. This scheme effectively enhances the identifiability of keypoints by fusing local gradient information and contextual semantic features, and exhibits higher stability, especially in the identification of keypoints in soft tissue.

[0083] 2. Improved noise resistance of the algorithm on low-quality CBCT images, effectively addressing common clinical noise interference such as metal artifacts, motion blur, and low-dose imaging. To address the prevalent stripe artifacts and shot noise in CBCT images, this scheme employs a composite denoising module combining anisotropic diffusion and frequency domain filtering in the preprocessing stage. During training, the model actively learns noise distribution characteristics and suppresses their interference with keypoint prediction. This mechanism not only preserves high-frequency details in the image but also avoids edge blurring issues caused by traditional denoising methods.

[0084] 3. This scheme adopts an improved cascaded U-Net architecture, combining cascaded inference results and heatmap regression strategies. It preserves high-resolution spatial information in shallow networks, enhances semantic representation in deep networks, and achieves adaptive fusion of multi-scale features through skip connections. This design enables the network to simultaneously respond to both high-contrast hard tissue keypoints and low-contrast soft tissue keypoints.

[0085] The key point identification apparatus of one or more embodiments of this application will be described in detail below. Those skilled in the art will understand that these apparatuses can all be configured using commercially available hardware components through the steps taught in this solution.

[0086] Figure 4 This is a schematic diagram of the structure of a key point recognition device provided in an embodiment of this application, as shown below. Figure 4 As shown, the device includes: a processing module 41, an input module 42, a determination module 43, and an extraction module 44.

[0087] The processing module 41 is used to perform multi-resolution processing on the three-dimensional target image to be identified, to obtain a first resolution image and a second resolution image, wherein the resolution of the first resolution image is lower than the resolution of the second resolution image.

[0088] The input module 42 is used to input the first resolution image into the first probability prediction model to obtain the first probability distribution information of each key point in the first resolution image, wherein the first probability prediction model is used to predict the probability distribution of key points from the global low resolution image.

[0089] The determining module 43 is used to determine the first spatial location of each key point based on the first probability distribution information, and to input the local region information into the second probability prediction model to obtain the second probability distribution information of each key point, so as to determine the second spatial location of each key point based on the second probability distribution information, wherein the second probability prediction model is used to predict the probability distribution of key points from the local high-resolution image.

[0090] Extraction module 44 is used to extract local region information corresponding to each key point from the second resolution image based on the first spatial location.

[0091] In one possible implementation, the processing module 41 is specifically used to: resample the three-dimensional target image to an isotropic first voxel spacing to obtain the first resolution image; retain the original resolution of the three-dimensional target image at a second voxel spacing to obtain the second resolution image, wherein the first voxel spacing is greater than the second voxel spacing.

[0092] In one possible implementation, the extraction module 44 is specifically used to: map the first spatial position to the voxel coordinate system of the second resolution image to obtain high-resolution center coordinates; and crop out a three-dimensional image block of a preset size from the second resolution image as local region information, centered on the high-resolution center coordinates.

[0093] In one possible implementation, the first probability distribution information and the second probability distribution information are represented as a three-dimensional Gaussian heatmap, where the value of each voxel in the three-dimensional Gaussian heatmap represents the probability that the key point is located at the current voxel position.

[0094] In one possible implementation, the determining module 43 is specifically used to: normalize the value of each voxel in the three-dimensional Gaussian heatmap corresponding to the probability distribution information to obtain normalized probability distribution data; calculate the weighted average value of each voxel position using the normalized probability distribution data as weights, and use the weighted average value as the coordinates of the spatial position.

[0095] Figure 4 The device shown can perform the steps in the key point recognition method in the foregoing embodiments. For detailed execution process and technical effects, please refer to the description in the foregoing embodiments, which will not be repeated here.

[0096] Figure 5 This is a schematic diagram of the structure of a model training device provided in an embodiment of this application, as shown below. Figure 5 As shown, the device includes: an acquisition module 51, a processing module 52, a generation module 53, an update module 54, and a trimming module 55.

[0097] The acquisition module 51 is used to acquire a training sample set, which includes multiple three-dimensional sample images labeled with the real spatial locations of each key point.

[0098] The processing module 52 is used to perform the multi-resolution processing on each three-dimensional sample image to obtain a first resolution sample image and a second resolution sample image.

[0099] The generation module 53 is used to generate first supervision information based on the real spatial location of each key point in the first resolution sample image.

[0100] The update module 54 is used to input the first resolution sample image into the first probability prediction model to be trained, obtain the first prediction probability distribution information, and update the parameters of the first probability prediction model based on the first prediction probability distribution information and the first supervision information; and to input the local sample block into the second probability prediction model to be trained, obtain the second prediction probability distribution information, and update the parameters of the second probability prediction model based on the second prediction probability distribution information and the second supervision information.

[0101] The cropping module 55 is used to crop out local sample blocks from the second resolution sample image based on the spatial location output by the first probability prediction model, and generate second supervision information based on the real spatial location of each key point in the local sample block.

[0102] In one possible implementation, the update module 54 is specifically used to: calculate the adaptive wing loss and the mean squared error loss based on the difference between the predicted probability distribution information and the supervision information, wherein the adaptive wing loss is used to measure the difference between the predicted probability distribution and the true probability distribution at each voxel position in the three-dimensional Gaussian heatmap, and the mean squared error loss is the average of the squares of the differences between the predicted value and the true value at all voxel positions; and update the model parameters based on the adaptive wing loss and the mean squared error loss.

[0103] In one possible implementation, the first probabilistic prediction model and / or the second probabilistic prediction model is a three-dimensional U-shaped network, specifically comprising: an encoder for performing multi-scale feature extraction on the input three-dimensional sample image, generating a multi-scale feature map from high-resolution detail features to low-resolution semantic features through layer-by-layer downsampling; a decoder for performing layer-by-layer upsampling on the deepest feature map output by the encoder, restoring the spatial size of the feature map to be the same as that of the three-dimensional sample image; and a skip connection structure for concatenating the feature maps output by each layer of the encoder with the feature maps of the corresponding layers of the decoder in the channel dimension.

[0104] Figure 5The device shown can perform the steps in the model training method in the foregoing embodiments. For detailed execution process and technical effects, please refer to the description in the foregoing embodiments, which will not be repeated here.

[0105] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, in practice, this electronic device includes a memory 21 and a processor 22.

[0106] Memory 21 is used to store computer programs and can be configured to store various other data to support operation on the electronic device. Examples of this data include instructions for any application or method used to operate on the electronic device, data structures, contact data, phone book data, messages, pictures, videos, etc.

[0107] The processor 22, coupled to the memory 21, is used to execute the computer program in the memory 21 to implement the key point recognition method provided in the foregoing embodiments.

[0108] Furthermore, such as Figure 6 As shown, the electronic device also includes other components such as a communication component 23, a display 24, a power supply component 25, and an audio component 26. Figure 6 The diagram only shows some components and does not mean that the electronic device includes only these components. Figure 6 The components shown are as follows. The electronic device in this embodiment can be a terminal device such as a desktop computer, laptop computer, smartphone, or IoT device, or a server device such as a conventional server, cloud server, or server array.

[0109] The aforementioned memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0110] The aforementioned communication component is configured to facilitate wired or wireless communication between the device containing the communication component and other devices. The device containing the communication component can access wireless networks based on communication standards, such as 2G, 3G, 4G / LTE, 5G, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel.

[0111] The aforementioned display includes a screen, which may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a Touch Panel, the screen can be implemented as a touchscreen to receive input signals from the user. The Touch Panel includes one or more touch sensors to sense touches, swipes, and gestures on the Touch Panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0112] The aforementioned power supply components provide power to various components within the device in which they reside. These power supply components may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device in which they reside.

[0113] The aforementioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0114] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile components, or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium.

[0115] Accordingly, this application also provides a computer program product, which includes a computer program or instructions that, when executed by a processor, cause the processor to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. Furthermore, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, enabling the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to function as an apparatus for implementing the corresponding functions in the above method embodiments.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. 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. Such 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.

Claims

1. A key point recognition method, characterized by, The method includes: The three-dimensional target image to be identified is subjected to multi-resolution processing to obtain a first resolution image and a second resolution image, wherein the resolution of the first resolution image is lower than that of the second resolution image; The first resolution image is input into the first probability prediction model to obtain the first probability distribution information of each key point in the first resolution image, wherein the first probability prediction model is used to predict the probability distribution of key points from the global low resolution image. The first spatial location of each key point is determined based on the first probability distribution information. Based on the first spatial location, local region information corresponding to each key point is extracted from the second resolution image; The local region information is input into the second probability prediction model to obtain the second probability distribution information of each key point, so as to determine the second spatial location of each key point based on the second probability distribution information. The second probability prediction model is used to predict the probability distribution of key points from the local high-resolution image.

2. The method of claim 1, wherein, The process of performing multi-resolution processing on the 3D target image to be identified to obtain a first resolution image and a second resolution image includes: The three-dimensional target image is resampled to an isotropic first voxel spacing to obtain the first resolution image; The original resolution of the three-dimensional target image at the second voxel spacing is preserved to obtain the second resolution image, wherein the first voxel spacing is greater than the second voxel spacing.

3. The method of claim 1, wherein, The step of extracting local region information corresponding to each key point from the second resolution image based on the first spatial location includes: The first spatial location is mapped to the voxel coordinate system of the second resolution image to obtain the high-resolution center coordinates; Using the high-resolution center coordinates as the center, a three-dimensional image block of a preset size is cropped from the second-resolution image as local region information.

4. The method according to any one of claims 1 to 3, characterized in that, The first probability distribution information and the second probability distribution information are represented as a three-dimensional Gaussian heatmap, where the value of each voxel in the three-dimensional Gaussian heatmap represents the probability that the key point is located at the current voxel position.

5. The method of claim 4, wherein, The steps for determining the spatial location of each key point based on probability distribution information include: The value of each voxel in the three-dimensional Gaussian heatmap corresponding to the probability distribution information is normalized to obtain normalized probability distribution data; Using the normalized probability distribution data as weights, a weighted average value is calculated for each voxel position, and the weighted average value is used as the coordinates of the spatial position.

6. A model training method, comprising: The method includes: Obtain a training sample set, which contains multiple 3D sample images labeled with the real spatial locations of each key point; Each 3D sample image is processed using multi-resolution methods to obtain a first-resolution sample image and a second-resolution sample image. First supervision information is generated based on the real spatial location of each key point in the first resolution sample image; The first resolution sample image is input into the first probability prediction model to be trained to obtain the first prediction probability distribution information, and the parameters of the first probability prediction model are updated based on the first prediction probability distribution information and the first supervision information. Based on the spatial location output by the first probability prediction model, a local sample block is cropped from the second resolution sample image, and second supervision information is generated based on the true spatial location of each key point in the local sample block. The local sample block is input into the second probability prediction model to be trained to obtain the second prediction probability distribution information, and the parameters of the second probability prediction model are updated based on the second prediction probability distribution information and the second supervision information.

7. The method of claim 6, wherein, The first and second predicted probability distribution information are represented as three-dimensional Gaussian heatmaps. The steps for updating the model parameters based on the predicted probability distribution information and the supervision information include: Based on the difference between the predicted probability distribution information and the supervision information, the adaptive wing loss and the mean squared error loss are calculated. The adaptive wing loss is used to measure the difference between the predicted probability distribution and the true probability distribution at each voxel position in the three-dimensional Gaussian heatmap. The mean squared error loss is the average of the squares of the differences between the predicted value and the true value at all voxel positions. The model parameters are updated based on adaptive wing loss and mean squared error loss.

8. The method according to claim 6 or 7, characterized in that, The first probabilistic prediction model and / or the second probabilistic prediction model are three-dimensional U-shaped networks, specifically including: The encoder is used to extract multi-scale features from the input 3D sample image, generating a multi-scale feature map from high-resolution detail features to low-resolution semantic features through layer-by-layer downsampling; The decoder is used to upsample the deepest feature map output by the encoder layer by layer to restore the spatial size of the feature map to be the same as that of the three-dimensional sample image. A skip connection structure is used to concatenate the feature maps output by each layer of the encoder with the feature maps of the corresponding layers of the decoder along the channel dimension.

9. An electronic device, characterized in that, include: A processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to execute the steps of the method as claimed in any one of claims 1-8.

10. A computer storage medium, characterized in that, The computer storage medium stores a plurality of instructions adapted for loading by a processor and executing the steps of the method as described in any one of claims 1-8.

Images