Method and system for detecting and locating bone metastasis based on SPECT bone imaging

Abstract

The application relates to the field of image processing and discloses a bone metastasis detection and positioning system and method based on SPECT bone imaging. The system extracts and processes multi-scale features of a to-be-detected image through a feature pyramid network to obtain a multi-scale feature map; then extracts metabolic features, morphological features and uptake mode features from the multi-scale feature map through a bone-specific feature model, and gradually fuses the metabolic features, the morphological features and the uptake mode features to obtain a bone-specific enhanced feature map; then extracts and fuses features of the bone-specific enhanced feature map through a mixed feature fusion model to obtain mixed fusion features; finally, the mixed fusion features are subjected to probability reasoning and double-uncertainty quantization processing through a probability reasoning network to determine a bone metastasis detection result, high-precision detection and positioning of bone metastasis lesions are realized, and the problems that existing technologies cannot effectively identify different size lesions and lack global perception ability are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and relates to, but is not limited to, a system and method for detecting and locating bone metastases based on SPECT bone imaging. Background Technology

[0002] Single-photon emission computed tomography (SPECT) bone scintigraphy, using a 99mTc-MDP tracer to detect changes in bone metabolic activity, is the standard imaging method for diagnosing bone metastases. Traditional SPECT diagnosis of bone metastases relies heavily on the subjective interpretation of radiologists, resulting in significant inter-observer variability. Although SPECT bone scintigraphy has a sensitivity of over 95%, its specificity is only 60-70%, leading to a high false-positive rate. The diagnostic consistency among different physicians for the same case is only 60-75%, particularly in identifying early, small bone metastases (<5mm in diameter), easily leading to misdiagnosis and missed diagnosis.

[0003] Currently, SPECT bone metastases can be detected using image classification network architectures. However, existing deep learning methods mostly employ general image classification network architectures (such as ResNet and DenseNet), which cannot effectively extract specific features of bone metastases, resulting in limited recognition capabilities and poor detection accuracy. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a system and method for detecting and locating bone metastases based on SPECT bone imaging, which can achieve high-precision detection and location of bone metastases, solving the problems of existing technologies being unable to effectively identify lesions of different sizes and lacking global perception capabilities.

[0005] The specific technical solutions of this invention are as follows:

[0006] The first aspect of this application provides a system for detecting and locating bone metastases based on SPECT bone imaging, comprising:

[0007] The preprocessing module is configured to acquire raw bone scan images and preprocess the raw bone scan images to obtain images to be detected;

[0008] The feature pyramid network, including a small lesion detection branch, a local lesion detection branch, a diffuse lesion detection branch, and a multi-scale feature fusion module, is configured to perform multi-scale feature extraction and feature processing on the image to be detected, and obtain a multi-scale feature map.

[0009] The bone-specific feature model, including a metabolic abnormality detection branch, a morphological analysis branch, an extraction pattern recognition branch, and a bone-specific feature fusion module, is configured to extract metabolic features, morphological features, and uptake pattern features from the multi-scale feature map, and progressively fuse the metabolic features, the morphological features, and the uptake pattern features to obtain a bone-specific enhanced feature map.

[0010] The hybrid feature fusion model, including a CNN local feature branch, a Transformer global feature branch, and a hybrid feature fusion module, is configured to perform local-global feature extraction and feature fusion on the bone-specific enhanced feature map to obtain hybrid fused features;

[0011] A probabilistic inference network is configured to perform probabilistic inference and dual uncertainty quantification on the hybrid fusion features to obtain the bone metastasis probability and the overall confidence level, and to determine the bone metastasis detection result based on the bone metastasis probability and the overall confidence level.

[0012] In some embodiments, the preprocessing module includes:

[0013] The bone tissue mask construction unit is configured to determine the optimal segmentation threshold of bone tissue in the original bone scan image using a multi-threshold maximum inter-class variance algorithm, and construct a bone tissue mask based on the optimal segmentation threshold, as well as the gradient information and morphological constraints corresponding to the original bone scan image, to obtain a first bone scan image; wherein, the multi-threshold maximum inter-class variance algorithm is used to maximize the weighted inter-class variance of bone tissue, soft tissue and background in the original bone scan image;

[0014] The regional standardization unit is configured to standardize the bone tissue region, soft tissue region, and background region in the first bone scan image based on the mean and standard deviation of the bone tissue region, the statistical parameters of the soft tissue region, and the statistical parameters of the background region, to obtain the second bone scan image.

[0015] The multi-view registration and fusion unit is configured to perform multi-scale registration processing on the feature points corresponding to the reference image and the floating image in the second bone scan image based on preset optimal transformation parameters, and to perform adaptive weighted fusion on the reference image and the floating image after multi-scale registration processing based on local image quality to obtain the image to be detected.

[0016] In some embodiments, the small lesion detection branch is constructed based on dense blocks and feature reuse blocks and is configured to extract a first-scale feature map from the image to be detected;

[0017] The local lesion detection branch is constructed based on the first residual block and the first max pooling layer, and is configured to extract the second scale feature map from the image to be detected through residual connection, group normalization and pooling downsampling.

[0018] The global detection branch for diffuse lesions is constructed based on a first multi-scale dilated convolutional layer and a second max pooling layer, and is configured to extract a third-scale feature map from the image to be detected; wherein, the pooling window of the second max pooling layer is larger than the pooling window of the first max pooling layer, and the resolutions of the first-scale feature map, the second-scale feature map, and the third-scale feature map are all different.

[0019] The multi-scale feature fusion module is configured to perform feature fusion on the first-scale feature map, the second-scale feature map, and the third-scale feature map through bilinear interpolation upsampling, channel concatenation, and convolution operations to obtain the multi-scale feature map.

[0020] In some embodiments, the metabolic anomaly detection branch is constructed based on a multi-path pooling layer and configured to extract the metabolic features from the multi-scale feature map;

[0021] The morphological analysis branch is constructed based on deformable convolutional layers and depthwise separable convolutional layers and is configured to extract the morphological features from the multi-scale feature map.

[0022] The extraction pattern recognition branch is constructed based on a second multi-scale dilated convolutional layer and a CBAM module, and is configured to extract the ingestion pattern features from the multi-scale feature map.

[0023] The bone-specific feature fusion module includes a first-stage fusion unit and a second-stage fusion unit; the first-stage fusion unit is configured to perform a first-stage fusion of the metabolic features and the morphological features to obtain a first fusion feature; the second-stage fusion unit is configured to perform a second-stage fusion of the first fusion feature and the uptake pattern features to obtain the bone-specific enhancement feature map.

[0024] In some embodiments, the second-stage fusion is constructed based on a second residual block and a self-attention block, and is configured to fuse the first fusion feature and the uptake pattern feature through residual connections and a self-attention mechanism to obtain the bone-specific enhancement feature map.

[0025] In some embodiments, the CNN local feature branch is constructed based on multi-scale residual blocks and configured to extract local features from the bone-specific enhancement feature map;

[0026] The Transformer global feature branch is configured to reconstruct the bone-specific enhanced feature map to obtain a feature sequence, and to model the global dependency relationship of the feature sequence through a multi-head self-attention mechanism to obtain global features;

[0027] The hybrid feature fusion module is configured to fuse the local features and the global features through a cross-attention mechanism to obtain the hybrid fused features.

[0028] In some embodiments, the probabilistic inference network includes:

[0029] The probability determination module is configured to convert the hybrid fusion features into bone transfer probabilities using a Bayesian deep learning framework.

[0030] The comprehensive confidence level determination module is configured to convert the hybrid fusion features into a comprehensive confidence level;

[0031] The classification decision module is configured to determine the bone metastasis detection result based on the bone metastasis probability and the overall confidence level.

[0032] In some embodiments, the comprehensive confidence determination module includes:

[0033] The cognitive uncertainty estimation unit is configured to determine the prediction distribution corresponding to the hybrid fusion features through Monte Carlo Dropout, and to determine the cognitive uncertainty based on the prediction distribution;

[0034] The random uncertainty estimation unit is configured to determine the predicted fluctuation of the hybrid fusion feature through a heteroscedastic regression network, and to determine the random uncertainty based on the predicted fluctuation;

[0035] An uncertainty fusion unit is configured to fuse the cognitive uncertainty and the accidental uncertainty to obtain the comprehensive confidence level.

[0036] In some embodiments, the classification decision module is further configured to:

[0037] When the probability of bone metastasis is greater than a first probability threshold and the overall confidence level is greater than a first confidence threshold, the bone metastasis detection result is determined to be high-risk bone metastasis.

[0038] When the probability of bone metastasis is less than the second probability threshold and the overall confidence level is greater than the second confidence threshold, the bone metastasis detection result is determined to be low-risk bone metastasis.

[0039] When the probability of bone metastasis is less than or equal to a first probability threshold and greater than or equal to a second probability threshold, the bone metastasis detection result is determined to be medium-risk bone metastasis.

[0040] When the overall confidence level is less than or equal to the second confidence threshold, the bone metastasis detection result is determined to be an indeterminate case.

[0041] A second aspect of this application provides a method for detecting and locating bone metastases based on SPECT bone scintigraphy, applied to the SPECT bone scintigraphy-based bone metastases detection and location system as described in any one of the first aspects, the method comprising:

[0042] Acquire raw bone scan images and preprocess them to obtain images to be detected;

[0043] Multi-scale feature extraction and feature processing are performed on the image to be detected to obtain a multi-scale feature map;

[0044] Metabolic features, morphological features, and uptake pattern features are extracted from the multi-scale feature map, and the metabolic features, morphological features, and uptake pattern features are progressively fused to obtain a bone-specific enhancement feature map.

[0045] Local-global feature extraction and feature fusion are performed on the bone-specific enhancement feature map to obtain hybrid fused features;

[0046] The hybrid fusion features are subjected to probabilistic reasoning and dual uncertainty quantification to obtain the bone metastasis probability and comprehensive confidence level. Based on the bone metastasis probability and the comprehensive confidence level, the bone metastasis detection result is determined.

[0047] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0048] In this embodiment of the invention, a preprocessing module first acquires the original bone scan image and preprocesses it to obtain the image to be detected. Then, a feature pyramid network is used to extract and process multi-scale features from the image to be detected, resulting in a multi-scale feature map. Next, a bone-specific feature model extracts metabolic, morphological, and uptake pattern features from the multi-scale feature map, and progressively fuses these features to obtain a bone-specific enhanced feature map. Then, a hybrid feature fusion model is used to extract and fuse local and global features from the bone-specific enhanced feature map, resulting in a hybrid fused feature map. Finally, a probabilistic inference network is used to perform probabilistic inference and dual uncertainty quantification on the hybrid fused feature map to obtain the bone metastasis probability and overall confidence level. Based on the bone metastasis probability and overall confidence level, the bone metastasis detection result is determined. This achieves high-precision detection and localization of bone metastatic lesions, providing doctors with accurate diagnostic decision support and effectively solving the problems of existing technologies being unable to effectively identify lesions of different sizes and lacking global perception capabilities. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0050] Figure 1 A schematic diagram of the structure of the SPECT bone imaging-based bone transfer detection and localization system provided in an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of the feature pyramid network provided in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the structure of the bone-specific feature model provided in an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the structure of the hybrid feature fusion model provided in an embodiment of the present invention;

[0054] Figure 5 This is a schematic diagram of the structure of a probabilistic inference network provided in an embodiment of the present invention;

[0055] Figure 6 This is a flowchart illustrating the method for detecting and locating bone metastases based on SPECT bone imaging provided in an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0058] It should be noted that the terms "first, second, and third" used in the embodiments of the present invention are only used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described herein can be implemented in an order other than that illustrated or described herein.

[0059] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which these embodiments of the invention pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0060] Figure 1 This is a schematic diagram of a bone transfer detection and localization system based on SPECT bone imaging, provided as an embodiment of the present invention. Figure 1 As shown, the bone transfer detection and localization system based on SPECT bone imaging provided in this embodiment of the invention includes a preprocessing module, a feature pyramid network, a bone-specific feature model, a hybrid feature fusion model, and a probabilistic inference network.

[0061] Because raw bone scan images contain multiple components such as bone tissue, soft tissue, blood pool, and background, and because the imaging agents used for radionuclide bone imaging (such as 99mTc-MDP) have different uptake characteristics in different tissues, in some embodiments, a preprocessing module is configured to acquire the raw bone scan image and preprocess it to obtain the image to be detected, in order to eliminate imaging noise and enhance bone transfer contrast. The raw bone scan image is the image obtained by performing a bone scan on a patient using single-photon emission computed tomography (SPECT); it can also be called the raw SPECT image. The raw bone scan image after adaptive preprocessing of bone tissue is the image to be detected; that is, the image to be detected is the preprocessed raw bone scan image.

[0062] In the original bone scan images, early micrometastatic lesions are 2-3 mm in diameter, localized patchy lesions are 1-2 cm in diameter, and diffuse multiple lesions involve multiple anatomical regions. To effectively identify and locate bone metastatic lesions of different sizes, in some embodiments, a feature pyramid network is configured to perform multi-scale feature extraction and processing on the image to be detected, resulting in a multi-scale feature map.

[0063] Bone metastases, due to increased osteoblast activity, accelerated angiogenesis, and bone remodeling, lead to significantly increased imaging agent uptake. Different types of bone metastases (e.g., osteolytic, osteoblastic, and mixed) exhibit distinct morphological patterns, including punctate, patchy, and diffuse uptake patterns, each corresponding to different pathological mechanisms and clinical significance. To maintain the independence of features across different dimensions while achieving effective information integration, in some embodiments, a bone-specific feature model is configured to extract metabolic, morphological, and uptake pattern features from a multi-scale feature map, and progressively fuse these features to obtain a bone-specific enhanced feature map. In some embodiments, a hybrid feature fusion model is configured to perform local-global feature extraction and feature fusion on the bone-specific enhanced feature map to obtain a hybrid fused feature. This hybrid feature fusion model includes a CNN local feature branch, a Transformer global feature branch, and a hybrid feature fusion module. The CNN local feature branch extracts local detail features, while the Transformer global feature branch extracts global features. The hybrid feature fusion module is also involved.

[0064] To provide a quantitative assessment of dual uncertainties (i.e., cognitive uncertainty and stochastic uncertainty) along with the test results, enabling physicians to understand the confidence level of the test results, thereby improving the reliability and credibility of the test results and enhancing the accuracy and safety of clinical diagnosis, in some embodiments, a probabilistic inference network is configured to perform probabilistic inference and dual uncertainty quantification on the hybrid fusion features to obtain the probability of bone metastasis and the overall confidence level. Based on the probability of bone metastasis and the overall confidence level, the bone metastasis detection result is determined; wherein, the bone metastasis detection result is used to indicate the specific type of bone metastasis, such as high-risk metastasis, intermediate-risk bone metastasis, low-risk bone metastasis, and indeterminate case.

[0065] In some embodiments, the preprocessing module includes a bone tissue mask construction unit, a region normalization unit, and a multi-view registration and fusion unit.

[0066] The bone tissue mask construction unit is used to adaptively segment the original bone scan image based on the intake intensity distribution characteristics, which is the first step in preprocessing. In some embodiments, the bone tissue mask construction unit is configured to determine the optimal segmentation threshold of bone tissue in the original bone scan image using a multi-threshold maximum inter-class variance algorithm (also known as the multi-threshold Otsu algorithm), and construct a bone tissue mask based on the optimal segmentation threshold, as well as the gradient information and morphological constraints corresponding to the original bone scan image, to obtain a first bone scan image. The multi-threshold maximum inter-class variance algorithm is used to maximize the weighted inter-class variance of bone tissue, soft tissue, and background in the original bone scan image.

[0067] For example, the multi-threshold maximum inter-class variance algorithm can be represented by Equation 1, that is, the optimal segmentation threshold of bone tissue in the original bone scan image can be determined by Equation 1, as shown below:

[0068] (Formula 1);

[0069] in, This represents the optimal segmentation threshold. For the number of categories, Let i be the weight of the i-th class. The variance is between classes.

[0070] Understandably, the multi-threshold maximum inter-class variance algorithm shown in Formula 1 takes into account the multi-tissue characteristics of the original bone scan image, and can more accurately separate bone tissue, soft tissue and background.

[0071] In some embodiments, the process of constructing a bone tissue mask based on the optimal segmentation threshold and the gradient information and morphological constraints corresponding to the original bone scan image to obtain the first bone scan image can be as follows: First, construct an initial bone tissue mask based on the optimal segmentation threshold; then, combine the gradient information and morphological constraints of the original bone scan image to remove non-bone tissue high uptake areas such as blood pools and kidneys from the initial bone tissue mask to obtain the target bone tissue mask, thereby obtaining the first bone scan image; that is, the first bone scan image is a bone tissue image with the target bone tissue mask.

[0072] For example, the bone tissue mask construction unit can binarize the original bone scan image, and then, based on the optimal segmentation threshold, determine the pixels in the binarized original bone scan image whose grayscale values ​​are greater than or equal to the optimal segmentation threshold as bone tissue regions, and construct an initial bone tissue mask based on these pixels. Then, based on the gradient magnitude of the pixels in the initial bone tissue mask and a preset gradient threshold, it is determined whether the pixels in the initial bone tissue mask belong to bone tissue regions. When the gradient magnitude of a pixel is less than or equal to the gradient threshold, it is determined that the pixel does not belong to a bone tissue region and is removed from the initial bone tissue mask; when the gradient magnitude of a pixel is greater than the gradient threshold, it is determined that the pixel belongs to a bone tissue region and is retained. Since the original bone scan image may contain small noise areas after binarization, in order to ensure the accuracy of subsequent feature extraction, morphological operations are used to further remove artifacts to obtain the target bone tissue mask, thereby obtaining the first bone scan image.

[0073] It should be noted that the methods for obtaining gradient information and morphological constraints from the original bone scan image are all existing technologies. For example, after binarizing and performing dilation and erosion operations on the original bone scan image, the gradient image corresponding to the original bone scan image (equivalent to the aforementioned gradient information) can be obtained by subtracting the eroded bone scan image from the dilated bone scan image. Another example is that, based on the characteristics of the original bone scan image, structuring elements of appropriate shape and size can be selected, and then the morphological constraints of the original bone scan image can be determined through comparison of multiple structuring elements. Further details are omitted here.

[0074] In some embodiments, the partitioning and standardization unit is used to standardize different tissues according to their uptake characteristics. In some embodiments, the partitioning and standardization unit is configured to standardize the bone tissue region, soft tissue region, and background region in the first bone scan image based on the mean and standard deviation of the bone tissue region, statistical parameters of the soft tissue region, and statistical parameters of the background region to obtain a second bone scan image.

[0075] For example, the bone tissue region, soft tissue region, and background region in the first bone scan image can be standardized using Formula 2, as shown below:

[0076] (Formula 2);

[0077] in, This represents the second bone scan image; This represents the grayscale value at pixel (x,y) of the original bone scan image; , These represent the mean and standard deviation of the bone tissue region, respectively. , These are the statistical parameters for the soft tissue region; , These are the statistical parameters for the background region; It is a numerically stable term; This represents the mask corresponding to the bone tissue region. This represents the mask corresponding to the soft tissue region. This represents the mask corresponding to the background area.

[0078] In some embodiments, the multi-view registration and fusion unit is used to solve the spatial consistency and information complementarity problems of multi-view data in bone scan images (i.e., SPECT images) based on a multi-scale registration strategy based on feature point matching and an adaptive weighted fusion algorithm based on local image quality. In some embodiments, the view registration and fusion unit is configured to perform multi-scale registration processing on the feature points corresponding to the reference image and the floating image in the second bone scan image based on preset optimal transformation parameters, and to perform adaptive weighted fusion on the multi-scale registered reference image and the floating image based on local image quality to obtain the image to be detected.

[0079] For example, the multi-scale registration strategy based on feature point matching can be represented by Equation 3, that is, the view registration fusion unit can perform multi-scale registration processing on the feature points corresponding to the reference image and the floating image in the second bone scan image through Equation 3, as shown below:

[0080] (Formula 3);

[0081] in, The optimal transformation parameters; and These are the corresponding feature points of the reference image and the floating image in the second bone scan image; For feature point weights, , , These are the rotation, translation, and scaling parameters, respectively.

[0082] The adaptive weighted fusion algorithm based on local image quality can be expressed by Equation 4. That is, the view registration fusion unit can adaptively weight and fuse the reference image and the floating image after multi-scale registration based on local image quality using Equation 4, as shown below:

[0083] (Formula 4);

[0084] in, This represents the pixel value of the second bone scan image at coordinates (x, y), i.e., the adaptive weighted fusion result; Indicates the number of views; For the v-th view at position Local quality score, , For hyperparameters, This indicates the position of the v-th view. Structural similarity, This indicates the position of the v-th view. The gradient magnitude; The pixel values ​​registered for the v-th view.

[0085] Understandably, adaptive weighted fusion of the reference image and the floating image after multi-scale registration based on local image quality can ensure that the fused second bone scan image fully utilizes the high-quality information of each view, thereby improving the accuracy of the second bone scan image.

[0086] Figure 2 A schematic diagram of a feature pyramid network structure is shown. Figure 2 As shown, the feature pyramid network includes a small lesion detection branch, a local lesion detection branch, a diffuse lesion detection branch, and a multi-scale feature fusion module.

[0087] In some embodiments, the small lesion detection branch is constructed based on dense blocks and feature reuse blocks and is configured to extract a first-scale feature map from the image to be detected.

[0088] For example, such as Figure 2 As shown, the small lesion detection branch can include a dense block connected in sequence ( Figure 2 The model consists of two blocks: a Dense Block and a Feature Reuse Block. The Dense Block comprises a batch normalization layer, a Swish activation layer, and multiple densely connected convolutional layers, each using a small convolutional kernel (e.g., 3×3). The Dense Block extracts feature information from different levels of the image to be detected layer by layer. In practice, the Dense Block normalizes the input data (the image to be detected) using a batch normalization layer, then introduces nonlinearity through a Swish activation layer to enhance the model's expressive power. Following this, multiple densely connected convolutional layers with small convolutional kernels reduce the number of parameters while maintaining the receptive field, thus improving the efficiency of feature extraction and ultimately outputting an initial first-scale feature map. The Feature Reuse Block further processes the initial first-scale feature map output by the Dense Block. In practice, the Feature Reuse Block reuses and fuses features in the initial first-scale feature map through feature reuse and direct gradient propagation to extract more discriminative and robust features, thereby enhancing the expressive power of small features and obtaining the first-scale feature map.

[0089] In some embodiments, the local lesion detection branch is constructed based on a first residual block and a first max pooling layer, and is configured to extract a second-scale feature map from the image to be detected through residual connections, group normalization, and pooling downsampling.

[0090] For example, such as Figure 2 As shown, the local lesion detection branch may include a first residual block connected in sequence ( Figure 2The first residual block consists of a ResBlock and a first max-pooling layer. The first residual block includes a group normalization layer, a swish activation layer, and multiple convolutional layers with residual connections, each using a medium-sized kernel (e.g., 5×5). In practical applications, the local lesion detection branch can standardize the image to be detected using the group normalization layer in the first residual block to make the data distribution more stable; then, a nonlinear factor is introduced through the swish activation layer to enhance the model's expressive power; next, the convolutional layers in the first residual block capture the medium-range spatial dependencies in the standardized image to be detected, thereby completing the extraction of patchy lesion features; finally, the feature map output by the first residual block can be downsampled using the first max-pooling layer to reduce the feature map resolution while retaining the main feature information, thus obtaining the second-scale feature map.

[0091] In some embodiments, the global detection branch for diffuse lesions is constructed based on a first multi-scale dilated convolutional layer and a second max-pooling layer, and is configured to extract a third-scale feature map from the image to be detected. The pooling window of the second max-pooling layer is larger than the pooling window of the first max-pooling layer; for example, the pooling window of the first max-pooling layer can be 2×2, and the pooling window of the second max-pooling layer can be 4×4.

[0092] For example, such as Figure 2 As shown, the global detection branch for diffuse lesions may include a first multi-scale porous convolutional layer connected in sequence ( Figure 2 The system consists of a multi-scale dilated convolutional layer and a second max-pooling layer. The dilation rate of the first-scale dilated convolutional layer can be (2, 4, 8). In practical applications, the global detection branch for diffuse lesions can capture the spatial context of different ranges in the image to be detected using the first multi-scale dilated convolution with a dilation rate of (2, 4, 8), thereby expanding the receptive field without increasing the number of parameters. Then, the second max-pooling layer can downsample the feature map output by the first multi-scale dilated convolutional layer to further extract and integrate feature information, resulting in a third-scale feature map.

[0093] In some embodiments, the multi-scale feature fusion module is composed of a feature alignment block, a feature concatenation layer, and a first convolutional layer. It is configured to fuse the first-scale feature map, the second-scale feature map, and the third-scale feature map through bilinear interpolation upsampling, channel concatenation, and convolution operations to obtain a multi-scale feature map. The first-scale feature map, the second-scale feature map, and the third-scale feature map all have different resolutions. For example, the resolution of the first-scale feature map may be 512×512×64, the resolution of the second-scale feature map may be 256×256×64, and the resolution of the third-scale feature map may be 128×128×64. Thus, the resolution of the multi-scale feature map obtained based on the above feature map fusion can be 512×512×64.

[0094] For example, such as Figure 2 As shown, the multi-scale feature fusion module may include a feature alignment block, a feature concatenation layer, and a first convolutional layer connected in sequence. Figure 2 The first convolutional layer uses a 1×1 kernel. In practical applications, the multi-scale feature fusion module can align the first, second, and third scale feature maps to a preset size using bilinear interpolation upsampling through feature alignment blocks. Then, a feature stitching layer stitches the unified-size first, second, and third scale feature maps along the channel dimension to obtain a stitched feature map. Finally, the first convolutional layer performs a convolution operation on the stitched feature map to effectively fuse feature information from different scales, reduce the number of channels in the feature map, lower computational complexity, and obtain a multi-scale feature map.

[0095] It is understood that the feature pyramid network provided in this application embodiment can effectively identify and locate bone metastases of different sizes. Compared with traditional multi-scale feature pyramid networks, the feature pyramid network provided in this application embodiment has higher detection sensitivity and specificity, effectively improving the accuracy of detecting small lesions and locating large lesions. The resulting multi-scale feature map contains complete scale information from small lesions to diffuse lesions. Thus, it can provide doctors with auxiliary treatment planning and efficacy evaluation.

[0096] Figure 3 A schematic diagram of the structure of a bone-specific feature model is shown. Figure 3 As shown, the bone-specific feature model includes a metabolic abnormality detection branch, a morphological analysis branch, an extraction pattern recognition branch, and a bone-specific feature fusion module.

[0097] In some embodiments, the metabolic anomaly detection branch is constructed based on a multi-path pooling layer and a second convolutional layer and is configured to extract metabolic features from a multi-scale feature map.

[0098] For example, such as Figure 3 As shown, the metabolic abnormality detection branch includes a multipath pooling layer and a second convolutional layer connected in sequence. Figure 3 The second convolutional layer uses a 3×3 kernel. Multi-path pooling layers can be constructed using max pooling, average pooling, and adaptive pooling layers. In practical applications, the metabolic anomaly detection branch can use a max pooling layer to select the maximum value within each pooling window as the output, preserving the most significant metabolic features in the multi-scale feature map; an average pooling layer calculates the average value within each pooling window as the output, obtaining a smoother distribution of metabolic features in the multi-scale feature map; and an adaptive pooling layer adjusts the pooling window size and stride according to the size of the multi-scale feature map and the preset output size, performing downsampling operations on the multi-scale feature map to adapt to the extraction needs of metabolic features at different scales. Thus, combining max pooling, average pooling, and adaptive pooling layers allows for comprehensive and effective extraction of metabolic features from the multi-scale feature map. The metabolic anomaly detection branch can then use the second convolutional layer to perform convolution operations on the feature maps output by the max pooling, average pooling, and adaptive pooling layers, further extracting and fusing metabolic features to detect metabolic abnormalities in bone metastases and obtain metabolic characteristics.

[0099] In some application scenarios, the metabolic abnormality detection branch can be represented by Equation 5, which is shown below:

[0100] (Formula 5);

[0101] in, This represents a k×k max pooling layer. This represents a k×k average pooling layer. This indicates an adaptive pooling layer. Represents multi-scale feature maps, Indicates metabolic characteristics, This indicates a feature concatenation operation. This indicates a convolutional layer with 32 channels and a 3×3 kernel.

[0102] In some embodiments, the adaptive weights in the adaptive pooling layer are learned through global average pooling and fully connected layers, enabling automatic adjustment of the ratios of max pooling and average pooling according to the characteristics of different images. This allows the metabolic abnormality detection branch to simultaneously capture local peak uptake, regional average uptake, and adaptive uptake patterns, enhancing the expressive power of metabolic abnormality features.

[0103] In some embodiments, the morphological analysis branch is constructed based on deformable convolutional layers and depthwise separable convolutional layers and is configured to extract morphological features from multi-scale feature maps.

[0104] For example, such as Figure 3 As shown, the morphological analysis branch comprises a depthwise separable convolutional layer, a deformable convolutional layer, and a pointwise convolutional layer connected in sequence (not shown in the figure). In practical applications, the morphological analysis branch can standardize multi-scale feature maps using a layer normalization layer to eliminate differences between different samples; then, a Mish activation function is used to perform a nonlinear transformation on the standardized feature maps; next, a depthwise separable convolutional layer extracts depth and spatial dimensions features from the nonlinearly transformed multi-scale feature maps; then, a deformable convolutional layer dynamically adjusts the shape and position of the convolutional kernel based on the extracted features to better capture morphological features; finally, a pointwise convolutional layer integrates the features output by the deformable convolutional layer to obtain the morphological features.

[0105] In some application scenarios, the morphological analysis branch can be represented by Equation 6, which is shown below:

[0106] (Formula 6);

[0107] in, Represents multi-scale feature maps; Presentation layer normalization operation; This represents the Mish activation function; This represents a 3×3 depth-separable convolution; This represents a 3×3 deformable convolutional layer; This represents a 1×1 pointwise convolutional layer; Indicates morphological characteristics.

[0108] In some embodiments, the extraction pattern recognition branch is constructed based on a second multi-scale dilated convolutional layer and a CBAM module and is configured to extract ingested pattern features from a multi-scale feature map.

[0109] For example, such as Figure 3 As shown, the pattern recognition extraction branch includes a second multi-scale dilated convolutional layer connected in sequence ( Figure 3 The diagram shows a multi-scale spatial convolutional layer and a convolutional block attention module (CBAM). The dilation rate of the second-scale dilated convolutional layer can be (2, 4, 8). In practical applications, the pattern recognition extraction branch can capture initial ingested pattern features from different ranges in the multi-scale feature map through the second multi-scale dilated convolutional layer using different dilation rates. Then, the CBAM uses channel attention and spatial attention mechanisms to weight the initial ingested pattern features from the channel dimension and spatial dimension, respectively, to adaptively adjust the importance of different feature channels and spatial locations, thus obtaining the ingested pattern features.

[0110] In some embodiments, the bone-specific feature fusion module includes a first-stage fusion unit and a second-stage fusion unit. The first-stage fusion unit is configured to perform a first-stage fusion of metabolic and morphological features to obtain a first fused feature; the second-stage fusion unit is configured to perform a second-stage fusion of the first fused feature and uptake pattern features to obtain a bone-specific enhancement feature map. The second-stage fusion is constructed based on a second residual block and a self-attention block, and is configured to fuse the first fused feature and uptake pattern features through residual connections, combined with a self-attention mechanism, to obtain the bone-specific enhancement feature map.

[0111] For example, such as Figure 3 As shown, the bone-specific feature fusion module includes a first-stage fusion unit and a third convolutional layer ( Figure 3 The 3×3 convolutional layer shown), the fourth convolutional layer ( Figure 3 The 1×1 convolutional layer shown), the second residual block ( Figure 3 The diagram shows a residual block and a self-attention block. The first-stage fusion unit is connected to the third convolutional layer, the third and fourth convolutional layers are connected to the second residual block, and the second residual block is connected to the self-attention block. In practical applications, the bone-specific feature fusion module performs a first-stage fusion of metabolic and morphological features through the first-stage fusion unit to generate a first fused feature. The first-stage fusion unit includes a feature splicing layer, which splices metabolic and morphological features to achieve feature fusion and obtain the first fused feature. Then, the third and fourth convolutional layers perform feature transformation and dimensionality reduction on the first fused feature and the uptake pattern feature, respectively. Next, the transformed and dimensionality-reduced first fused feature and uptake pattern feature are input into the second residual block for second-stage fusion. Finally, the self-attention block performs self-attention calculation on the features output by the second residual block, dynamically adjusting the weights of features at different locations through the self-attention mechanism to obtain a bone-specific enhanced feature map.

[0112] Understandably, the bone-specific feature fusion module can enhance the three key pathological features of bone metastasis: abnormal metabolic activity, pathological morphological changes, and uptake pattern characteristics. This allows for the representation of abnormally high uptake characteristics of isomorphic enhancement imaging agents, the pathological morphological features of bone metastasis, and the characteristics of different uptake patterns.

[0113] Figure 4 A schematic diagram of a hybrid feature fusion model is shown. Figure 4 As shown, the hybrid feature fusion model includes a CNN local feature branch, a Transformer global feature branch, and a hybrid feature fusion module.

[0114] In some embodiments, the CNN local feature branch is constructed based on multi-scale residual blocks and is configured to extract local features from bone-specific enhanced feature maps.

[0115] For example, such as Figure 4 As shown, the CNN local feature branch includes sequentially connected multi-scale residual blocks and feature concatenation blocks. The multi-scale residual blocks consist of multiple residual blocks connected by residuals. In practical applications, the CNN local feature branch can perform convolution operations at different scales on the input features through each residual block in the multi-scale residual blocks, and fuse feature information at different scales through residual connections to achieve multi-level feature extraction of bone-specific enhancement feature maps. Then, the feature concatenation blocks concatenate the features extracted from the multi-scale residual blocks to form local features.

[0116] In some embodiments, the Transformer global feature branch is configured to reorganize the bone-specific enhanced feature map to obtain a feature sequence, and to model the global dependency relationship of the feature sequence through a multi-head self-attention mechanism to obtain global features.

[0117] For example, such as Figure 4 As shown, the Transformer global feature branch consists of sequentially connected image patch partitioning blocks, a positional encoding layer, a multi-head self-attention block, and a layer normalization layer. In practical applications, the Transformer global feature branch divides the bone-specific enhancement feature map into multiple image patches using the image patch partitioning blocks. These multiple image patches are then reassembled into a sequence and input into the positional encoding layer. The positional encoding layer adds positional information to each image patch in the sequence to preserve the spatial structure information in the feature map. Then, the multi-head self-attention block processes the image patches with added positional encoding, calculating the attention weights between different image patches to capture global dependencies. Finally, the layer normalization layer normalizes the output of the multi-head self-attention block to obtain the global features.

[0118] In some embodiments, the hybrid feature fusion module is configured to fuse local and global features through a cross-attention mechanism to obtain hybrid fused features.

[0119] For example, such as Figure 4 As shown, the hybrid feature fusion module includes a cross-attention block, a residual block, and a layer normalization layer connected in sequence. In practical applications, the hybrid feature fusion module uses the cross-attention block to fuse local and global features based on the cross-attention mechanism; the residual block is used to add the local features to the features processed by the cross-attention mechanism by introducing residual connections, and then the layer normalization layer normalizes the features output by the residual block to obtain the hybrid fused features.

[0120] Figure 5 A schematic diagram of a probabilistic reasoning network is shown. For example... Figure 5 As shown, the probabilistic reasoning network includes a probability determination module, a comprehensive confidence determination module, and a classification decision module.

[0121] In some embodiments, the probability determination module is configured to convert hybrid fusion features into bone metastasis probabilities using a Bayesian deep learning framework.

[0122] For example, the probability determination module may include multiple fully connected layers, multiple Dropout layers, and a Sigmoid activation layer. Wherein, as... Figure 5 As shown, multiple fully connected layers can include a first fully connected layer, a second fully connected layer, and a third fully connected layer; multiple dropout layers can include a first dropout layer and a second dropout layer. The first fully connected layer is 256-dimensional, the second fully connected layer is 128-dimensional, and the third fully connected layer is 1-dimensional. The retention rate of the first dropout layer is 0.3, and the retention rate of the second dropout layer is 0.5. In practical applications, the probability determination module performs preliminary feature transformation on the hybrid fusion features through the first fully connected layer, mapping the hybrid fusion features to a 256-dimensional feature space to extract more representative feature information. Subsequently, the first dropout layer randomly discards some neurons with a retention rate of 0.3 to prevent overfitting. Next, the second fully connected layer further reduces the dimensionality of the features processed by the first dropout layer, mapping them to a 128-dimensional feature space to further refine key features. Afterward, the second dropout layer again randomly discards some neurons with a retention rate of 0.5 to enhance the model's generalization ability. Finally, the third fully connected layer reduces the feature dimension to 1 dimension and maps the output to a probability value between 0 and 1 through a Sigmoid activation layer. This probability value is the bone metastasis probability.

[0123] In some embodiments, the comprehensive confidence level determination module is configured to convert hybrid fusion features into a comprehensive confidence level. The comprehensive confidence level determination module includes a cognitive uncertainty estimation unit, a random uncertainty estimation unit, and an uncertainty fusion unit.

[0124] In some embodiments, the cognitive uncertainty estimation unit is configured to determine the prediction distribution corresponding to the hybrid fusion features by Monte Carlo Dropout, and to determine the cognitive uncertainty based on the prediction distribution.

[0125] For example, such as Figure 5 As shown, the cognitive uncertainty estimation unit includes a Monte Carlo Dropout module. Figure 5The cognitive uncertainty estimation unit (CUE) uses the Monte Carlo Dropout module to perform multiple (e.g., 100) forward propagations of the hybrid fusion features. During each propagation, the Dropout layer randomly deactivates a portion of neurons, resulting in different prediction results. The variance of the prediction results is then calculated to determine the prediction distribution. Finally, the prediction distribution is accurately quantified into variance to obtain the cognitive uncertainty.

[0126] In some application scenarios, the cognitive uncertainty estimation unit can be represented by Equation 7, which is shown below:

[0127] (Formula 7);

[0128] in, Indicates cognitive uncertainty, Indicates the number of predictions. This represents the result of the t-th prediction. This represents the average of the T prediction results.

[0129] In some embodiments, the random uncertainty estimation unit is configured to determine the predicted volatility of the hybrid fusion features through a heteroscedastic regression network and to determine the random uncertainty based on the predicted volatility.

[0130] For example, such as Figure 5 As shown, the random uncertainty estimation unit comprises a heteroscedastic regression network, a fully connected layer, and a Softplus activation layer connected in sequence. In practical applications, the random uncertainty estimation unit processes the hybrid fusion features through the heteroscedastic regression network to obtain preliminary predicted fluctuation information. Then, the fully connected layer integrates and transforms the preliminary predicted fluctuation information, mapping the high-dimensional feature information to a more suitable dimensional space. Finally, the Softplus activation layer performs a nonlinear transformation on the output of the fully connected layer, converting the numerical values ​​within a certain range of the fully connected layer output into positive values ​​that conform to variance statistics, thereby representing the random uncertainty in the form of variance.

[0131] In some embodiments, the uncertainty fusion unit is configured to fuse cognitive uncertainty and accidental uncertainty to obtain a comprehensive confidence level.

[0132] In some application scenarios, the uncertainty fusion unit can be used as shown in Formula 8, that is, Formula 8 can be used to fuse cognitive uncertainty and random uncertainty to obtain the comprehensive confidence level. Formula 8 is shown below:

[0133] (Formula 8);

[0134] in, Indicates cognitive uncertainty, Indicates random uncertainty. Indicates the overall confidence level. This represents the reference value for the maximum uncertainty.

[0135] In some embodiments, the classification decision module is configured to determine the bone metastasis detection result based on the bone metastasis probability and the overall confidence level.

[0136] For example, the classification decision module is further configured to determine the bone metastasis detection result as high-risk bone metastasis when the probability of bone metastasis is greater than a first probability threshold (e.g., 0.8) and the overall confidence level is greater than a first confidence threshold (e.g., 0.8); to determine the bone metastasis detection result as low-risk bone metastasis when the probability of bone metastasis is less than a second probability threshold (e.g., 0.2) and the overall confidence level is greater than a second confidence threshold (e.g., 0.7); to determine the bone metastasis detection result as medium-risk bone metastasis when the probability of bone metastasis is less than or equal to the first probability threshold and greater than or equal to the second probability threshold; and to determine the bone metastasis detection result as an indeterminate case when the overall confidence level is less than or equal to the second confidence threshold.

[0137] It is understood that the embodiments of this application, through feature pyramid networks and bone-specific feature models, achieve high-precision detection and localization of bone metastatic lesions, providing doctors with accurate diagnostic decision support and effectively solving the problems of existing technologies being unable to effectively identify lesions of different sizes and lacking global perception capabilities. Simultaneously, the use of uncertainty quantification technology significantly improves the reliability and interpretability of the detection results, enhances the confidence level of clinical diagnosis, and enables the detection results to truly reflect the actual state of the lesions and the confidence level of the detection.

[0138] Based on the same inventive concept, this application also provides a method for detecting and locating bone metastases based on SPECT bone imaging, applied to a system for detecting and locating bone metastases based on SPECT bone imaging. The solution provided by this method is similar to the implementation scheme described in the above system; therefore, please refer to the limitations of the system for detecting and locating bone metastases based on SPECT bone imaging described above, and will not be repeated here. Specifically, Figure 6 This is a schematic flowchart illustrating a method for detecting and locating bone metastases based on SPECT bone imaging, as described in an embodiment of this application. Figure 6 As shown, the methods for detecting and locating bone metastases based on SPECT bone imaging include:

[0139] S601. Acquire the original bone scan image and preprocess it to obtain the image to be detected.

[0140] S602. Perform multi-scale feature extraction and feature processing on the image to be detected to obtain a multi-scale feature map.

[0141] S603. Extract metabolic features, morphological features, and uptake pattern features from the multi-scale feature map, and progressively fuse the metabolic features, morphological features, and uptake pattern features to obtain a bone-specific enhancement feature map.

[0142] S604. Perform local-global feature extraction and feature fusion on the bone-specific enhancement feature map to obtain hybrid fused features.

[0143] S605. Perform probabilistic reasoning and dual uncertainty quantification on the hybrid fusion features to obtain the bone metastasis probability and comprehensive confidence level, and determine the bone metastasis detection result based on the bone metastasis probability and comprehensive confidence level.

[0144] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of the invention, the sequence numbers of the above-described processes do not imply a sequential 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 the invention. The sequence numbers of the above-described embodiments of the invention are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0145] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0146] In the several embodiments provided by this invention, it should be understood that the disclosed methods can be implemented in other ways. The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments. The features disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0147] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A system for detecting and locating bone metastases based on SPECT bone imaging, characterized in that, include: The preprocessing module is configured to acquire raw bone scan images and preprocess the raw bone scan images to obtain images to be detected; The feature pyramid network, including a small lesion detection branch, a local lesion detection branch, a diffuse lesion detection branch, and a multi-scale feature fusion module, is configured to perform multi-scale feature extraction and feature processing on the image to be detected, and obtain a multi-scale feature map. The bone-specific feature model, including a metabolic abnormality detection branch, a morphological analysis branch, an extraction pattern recognition branch, and a bone-specific feature fusion module, is configured to extract metabolic features, morphological features, and uptake pattern features from the multi-scale feature map, and progressively fuse the metabolic features, the morphological features, and the uptake pattern features to obtain a bone-specific enhanced feature map. The hybrid feature fusion model, including a CNN local feature branch, a Transformer global feature branch, and a hybrid feature fusion module, is configured to perform local-global feature extraction and feature fusion on the bone-specific enhanced feature map to obtain hybrid fused features; A probabilistic inference network is configured to perform probabilistic inference and dual uncertainty quantification on the hybrid fusion features to obtain the bone metastasis probability and the overall confidence level, and to determine the bone metastasis detection result based on the bone metastasis probability and the overall confidence level.

2. The system according to claim 1, characterized in that, The preprocessing module includes: The bone tissue mask construction unit is configured to determine the optimal segmentation threshold of bone tissue in the original bone scan image using a multi-threshold maximum inter-class variance algorithm, and construct a bone tissue mask based on the optimal segmentation threshold, as well as the gradient information and morphological constraints corresponding to the original bone scan image, to obtain a first bone scan image; wherein, the multi-threshold maximum inter-class variance algorithm is used to maximize the weighted inter-class variance of bone tissue, soft tissue and background in the original bone scan image; The regional standardization unit is configured to standardize the bone tissue region, soft tissue region, and background region in the first bone scan image based on the mean and standard deviation of the bone tissue region, the statistical parameters of the soft tissue region, and the statistical parameters of the background region, to obtain the second bone scan image. The multi-view registration and fusion unit is configured to perform multi-scale registration processing on the feature points corresponding to the reference image and the floating image in the second bone scan image based on preset optimal transformation parameters, and to perform adaptive weighted fusion on the reference image and the floating image after multi-scale registration processing based on local image quality to obtain the image to be detected.

3. The system according to claim 1, characterized in that, The micro lesion detection branch is constructed based on dense blocks and feature reuse blocks and is configured to extract a first-scale feature map from the image to be detected. The local lesion detection branch is constructed based on the first residual block and the first max pooling layer, and is configured to extract the second scale feature map from the image to be detected through residual connection, group normalization and pooling downsampling. The global detection branch for diffuse lesions is constructed based on a first multi-scale dilated convolutional layer and a second max pooling layer, and is configured to extract a third-scale feature map from the image to be detected; wherein, the pooling window of the second max pooling layer is larger than the pooling window of the first max pooling layer, and the resolutions of the first-scale feature map, the second-scale feature map, and the third-scale feature map are all different. The multi-scale feature fusion module is configured to perform feature fusion on the first-scale feature map, the second-scale feature map, and the third-scale feature map through bilinear interpolation upsampling, channel concatenation, and convolution operations to obtain the multi-scale feature map.

4. The system according to claim 1, characterized in that, The metabolic anomaly detection branch is constructed based on a multi-path pooling layer and is configured to extract the metabolic features from the multi-scale feature map; The morphological analysis branch is constructed based on deformable convolutional layers and depthwise separable convolutional layers and is configured to extract the morphological features from the multi-scale feature map. The extraction pattern recognition branch is constructed based on a second multi-scale dilated convolutional layer and a CBAM module, and is configured to extract the ingestion pattern features from the multi-scale feature map. The bone-specific feature fusion module includes a first-stage fusion unit and a second-stage fusion unit; The first-stage fusion unit is configured to perform a first-stage fusion of the metabolic feature and the morphological feature to obtain a first fused feature; The second-stage fusion unit is configured to perform a second-stage fusion of the first fusion feature and the uptake pattern feature to obtain the bone-specific enhancement feature map.

5. The system according to claim 4, characterized in that, The second-stage fusion is constructed based on the second residual block and the self-attention block, and is configured to fuse the first fusion feature and the uptake pattern feature through residual connections and in combination with the self-attention mechanism to obtain the bone-specific enhancement feature map.

6. The system according to claim 1, characterized in that, The CNN local feature branch is constructed based on multi-scale residual blocks and is configured to extract local features from the bone-specific enhancement feature map; The Transformer global feature branch is configured to reconstruct the bone-specific enhanced feature map to obtain a feature sequence, and to model the global dependency relationship of the feature sequence through a multi-head self-attention mechanism to obtain global features; The hybrid feature fusion module is configured to fuse the local features and the global features through a cross-attention mechanism to obtain the hybrid fused features.

7. The system according to claim 1, characterized in that, The probabilistic inference network includes: The probability determination module is configured to convert the hybrid fusion features into bone transfer probabilities using a Bayesian deep learning framework. The comprehensive confidence level determination module is configured to convert the hybrid fusion features into a comprehensive confidence level; The classification decision module is configured to determine the bone metastasis detection result based on the bone metastasis probability and the overall confidence level.

8. The system according to claim 7, characterized in that, The comprehensive confidence level determination module includes: The cognitive uncertainty estimation unit is configured to determine the prediction distribution corresponding to the hybrid fusion features through Monte Carlo Dropout, and to determine the cognitive uncertainty based on the prediction distribution; The random uncertainty estimation unit is configured to determine the predicted fluctuation of the hybrid fusion feature through a heteroscedastic regression network, and to determine the random uncertainty based on the predicted fluctuation; An uncertainty fusion unit is configured to fuse the cognitive uncertainty and the accidental uncertainty to obtain the comprehensive confidence level.

9. The system according to claim 7, characterized in that, The classification decision module is further configured as follows: When the probability of bone metastasis is greater than a first probability threshold and the overall confidence level is greater than a first confidence threshold, the bone metastasis detection result is determined to be high-risk bone metastasis. When the probability of bone metastasis is less than the second probability threshold and the overall confidence level is greater than the second confidence threshold, the bone metastasis detection result is determined to be low-risk bone metastasis. When the probability of bone metastasis is less than or equal to a first probability threshold and greater than or equal to a second probability threshold, the bone metastasis detection result is determined to be medium-risk bone metastasis. When the overall confidence level is less than or equal to the second confidence threshold, the bone metastasis detection result is determined to be an indeterminate case.

10. A method for detecting and locating bone metastases based on SPECT bone scintigraphy, characterized in that, The method, applied to the SPECT bone scintigraphy-based bone metastasis detection and localization system as described in any one of claims 1-9, comprises: Acquire raw bone scan images and preprocess them to obtain images to be detected; Multi-scale feature extraction and feature processing are performed on the image to be detected to obtain a multi-scale feature map; Metabolic features, morphological features, and uptake pattern features are extracted from the multi-scale feature map, and the metabolic features, morphological features, and uptake pattern features are progressively fused to obtain a bone-specific enhancement feature map. Local-global feature extraction and feature fusion are performed on the bone-specific enhancement feature map to obtain hybrid fused features; The hybrid fusion features are subjected to probabilistic reasoning and dual uncertainty quantification to obtain the bone metastasis probability and comprehensive confidence level. Based on the bone metastasis probability and the comprehensive confidence level, the bone metastasis detection result is determined.

Images