Flat-scanning CT image aortic valve calcification segmentation system based on space-time prior

Abstract

A plain-scan CT image aortic valve calcification segmentation system based on space-time priori comprises a position priori information extraction module, an encoding module and a decoding module, after an aortic image is automatically segmented in an off-line stage to obtain a segmentation mask, attention weights are generated through preprocessing and encoding, and the aortic valve calcification is subjected to image segmentation; and inputting the aorta image serving as a training set into a segmentation model comprising a position prior information extraction module, a coding module and a decoding module, and performing real-time image segmentation through the trained segmentation model in an online stage. According to the method, calcification point features of different sizes and forms are captured through multi-scale prediction; the aorta segmentation prior is introduced, the sensitivity and positioning precision of the model to aortic valve calcification are improved, coronary artery opening sequence time sequence information is introduced, a double-branch structure is adopted to adapt to structural differences, accurate positioning and recognition of aortic valve calcification lesions are achieved, and the false detection rate and the omission ratio are effectively reduced.

Description

Technical Field

[0001] This invention relates to a technology in the field of image processing, specifically a system for segmenting aortic valve calcification in plain CT images based on spatiotemporal priors. Background Technology

[0002] Aortic valve calcification is associated with a variety of cardiovascular diseases, and its automatic assessment depends on the accurate segmentation of calcification. Existing automatic segmentation technology has the following defects: (1) segmentation is easily affected by structures such as aortic wall calcification and coronary artery calcification, resulting in missed segmentation and missegmentation; (2) segmentation performance is poor for microcalcifications and calcification edges. Summary of the Invention

[0003] To address the aforementioned deficiencies in existing technologies, this invention proposes a spatiotemporal prior-based aortic valve calcification segmentation system for plain CT images. It captures calcification features of varying sizes and shapes using a Feature Pyramid Network (FPN) architecture, reducing information loss during segmentation. By introducing aortic segmentation priors, the system enhances the model's sensitivity and localization accuracy for aortic valve calcification. Furthermore, by incorporating coronary ostial temporal information and employing a dual-branch decoding structure to adapt to anatomical differences, it achieves accurate identification and segmentation of aortic valve calcification lesions, effectively reducing both false positive and false negative rates.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a spatiotemporal prior-based aortic valve calcification segmentation system for plain CT images, comprising: a location prior information extraction module, an encoding module, and a decoding module, wherein: the location prior information extraction module preprocesses and encodes the initial aortic segmentation results obtained from the original CT images into attention weights; the encoding module extracts multi-scale image features from the original CT images and combines them with attention weights to generate weighted images that guide the network to focus on the aortic region; the decoding module restores the resolution of the weighted images through convolution and upsampling operations, and uses a branching structure to achieve more accurate segmentation of different images in the sequence.

[0006] The location prior information extraction module includes an automatic segmentation unit, a mask preprocessing unit, a downsampling unit, and a mask encoding unit. Specifically: the automatic segmentation unit inputs the original CT image obtained from instrument scanning into a general segmentation model to obtain an initial aortic segmentation result; the mask preprocessing unit filters and dilates the initial aortic segmentation result to obtain a usable aortic segmentation mask; the downsampling unit performs multiple downsampling operations on the aortic segmentation mask to obtain a set of aortic segmentation masks at different scales; and the mask encoding unit uses multiple convolutions to generate a spatial attention map from the aortic segmentation mask, outputting it as a spatial attention weight to the encoding module for fusion with the feature map extracted from the original image, thereby enhancing the model's ability to distinguish the aortic valve and its adjacent structures.

[0007] The encoding module includes four encoding layers, each comprising a residual convolutional unit and a prior fusion unit. Each residual convolutional unit includes two convolutional blocks based on residual connections, used to extract semantic information from the raw CT image obtained from the input instrument scan or the feature map output by the previous encoding layer, output a feature map, and pass it to the decoding module through skip connections. The prior fusion unit weights the feature map output by each residual convolutional unit according to attention weights to generate a weighted image used to enhance the model's perception ability of the aortic-related region.

[0008] The decoding module specifically includes: four decoding layers composed of residual convolutional units, a branch selection unit located between the third and fourth decoding layers, and a multi-scale output unit connected to the output of each decoding layer. Each decoding layer progressively amplifies the deep feature map through upsampling and uses skip connections between the encoding and decoding layers at corresponding scales to reduce information loss. The branch selection unit activates the corresponding decoding branch based on the temporal label of the original CT image, achieving differential segmentation of restored images with different structural characteristics. The multi-scale output unit receives the restored images output by multiple residual convolutional units, upsamples them uniformly to the original image size, and generates the final segmentation result.

[0009] The residual convolutional unit includes two convolutional blocks based on residual connections. It concatenates the weighted image or the magnified deep feature map output from the previous decoding layer with the feature map from the skip connection of the encoding module. The spatial resolution of the image is gradually restored through convolution to generate a restored image, thereby achieving pixel-level segmentation of the aortic calcification region.

[0010] Technical effect

[0011] This invention encodes the initial aortic segmentation mask generated by a fully automated segmentation tool into multi-scale spatial attention weights. This guides the network to focus on the ascending aorta and adjacent structures at multiple resolutions, while simultaneously using coronary ostial structures as temporal information to control the decoder to automatically select branches. This achieves differentiated processing of images with different anatomical states. A spatial weight function based on the aortic centroid angle is constructed to address the spatial distribution of aortic valve calcifications. Loss is calculated using multi-scale outputs, and a warm-up strategy is employed, using only standard IoU loss during the initial training phase to ensure stable model convergence. Compared to existing technologies, this invention significantly improves the detection sensitivity and segmentation accuracy for small-scale, morphologically diverse, and poorly defined aortic valve calcification lesions. It outperforms comparative methods in evaluation metrics such as Dice coefficient and IoU, providing more complete and finely segmented data for both micro-calcifications and larger calcification areas. Attached Figure Description

[0012] Figure 1 This is a flowchart of the present invention;

[0013] Figure 2 This is a schematic diagram of the segmentation model of the present invention;

[0014] Figure 3 This is a schematic diagram of the location prior coding module;

[0015] Figure 4 This is a schematic diagram of the encoding module;

[0016] Figure 5 This is a schematic diagram of a residual convolutional unit;

[0017] Figure 6 This is a schematic diagram of the decoding module;

[0018] Figure 7 To compare the segmentation results of the experiment. Detailed Implementation

[0019] like Figure 1 As shown, this embodiment relates to a method for segmenting aortic valve calcification in plain CT images based on the spatiotemporal prior of the above system, including:

[0020] Step 1: Obtain experimental data, specifically including:

[0021] 1.1 Acquiring plain chest CT images and corresponding aortic valve calcification segmentation labels. The experimental data in this embodiment came from Peking University Third Hospital and was acquired using both GE Revolution CT and SOMATOM Definition Flash devices, including 355 cases of plain chest CT sequences, totaling 27,852 images, of which 1,264 were positive images containing aortic valve calcification. The CT sequence is denoted as... The corresponding aortic valve calcification label is ,in: This indicates the sequence length, which is the number of images contained in a CT sequence. , The height and width of each frame of the image.

[0022] 1.2 For each CT sequence, an image index is provided that first appears in the scanning direction from bottom to top, containing the right coronary artery ostium. Based on this index, a time label for the right coronary artery ostium is automatically assigned to each frame of the sequence. , where t is the image index.

[0023] Step 2, construct as follows Figure 2 The aortic valve calcification segmentation model shown uses a location prior information extraction module to encode the initial aortic segmentation result obtained from the original CT image into attention weights. The encoding module extracts multi-scale image features from the original CT image and combines them with the attention weights to generate a weighted image that guides the network to focus on the aortic region. The decoding module performs convolution and upsampling operations to restore the resolution of the weighted image and employs a branching structure to achieve more accurate segmentation of different images in the sequence. Specifically, this includes:

[0024] 2.1 As Figure 3 As shown, after obtaining the initial aortic segmentation result for each image from the original CT image through general segmentation, the initial aortic segmentation result is preprocessed and encoded.

[0025] The general segmentation mentioned above is implemented using, but is not limited to, the fully automatic general segmentation tool (Total Segmentator).

[0026] The preprocessing specifically includes: since the aortic valve is located in the ascending aorta, only the segmentation results of the ascending aorta region where calcification may occur need to be retained; connected components in the original aortic segmentation mask are detected, and only the ascending aorta region closer to the image center is retained. Considering that the segmentation by the automatic segmentation tool is relatively coarse, a 13x13 dilation kernel is used to morphologically expand the ascending aorta region to further cover the candidate aorta region, finally obtaining the aortic segmentation mask. , , This represents the height and width of the image.

[0027] The encoding specifically includes: aortic segmentation mask. Perform separately Subsampling with a second kernel of 2 yields a set of aortic segmentation masks. Then, for each mask, use two concatenated... After extracting spatial information using convolution, a 1×1 convolution is used to compress the number of channels to 1, followed by normalization using the Sigmoid function. The final output is the multi-scale spatial attention weights. ,in: The size is , This represents the Sigmoid function. Indicates the kernel size as The convolution operation.

[0028] The multi-scale spatial attention weights This information, used as prior information about aortic anatomy, is fused with features output from each coding layer in the coding module.

[0029] 2.2 such as Figure 4 As shown and Figure 5 As shown, in each residual convolutional unit of the encoding module, the residual convolutional block sequentially performs 3×3 convolution, batch normalization, and Rectified Linear Unit (ReLU) to extract features from the input features. Then, cascaded channel attention and spatial attention are used to capture key features. A 1×1 convolution is then performed to adjust the number of channels in the processed feature map to match the input features. Finally, the input feature map and the processed feature map are superimposed. Specifically... : , Given the input feature map, The output feature map, For channel attention, For spatial attention, For the Sigmoid function, and For weight parameters, To correct the linear unit, For batch normalization operations, and These are average pooling and max pooling operations along the channel dimension, respectively. and These are the summation and maximum value operations performed along the channel dimension, respectively.

[0030] 2.3 Each prior fusion unit in the encoding module injects the aortic anatomical prior information at different resolutions captured by the location prior information extraction module into the encoding network, and combines it with the output features of each residual convolutional unit. Multiplication and superposition are used to guide the network to focus on the aortic region, enhance the specificity of the target location, and suppress interference from the anatomical neighborhood. Specifically: ,in: The fused feature map is the output of the prior fusion unit. This represents the output of the residual convolution unit. For Hadama accumulation.

[0031] 2.4 such as Figure 6 As shown, the branch selection unit in the decoding module selects based on the label of the input image. The system automatically selects between two residual convolutional units to achieve targeted predictions for images with different anatomical states; the multi-scale output unit converts the output of each decoding layer... After 1×1 convolution, the segmentation prediction results of each layer are obtained. Then The image is uniformly upsampled to the original image size. The upsampled results are then processed by channel stitching and convolution to obtain the final segmentation prediction result. Specifically: ), ,in: This represents the feature map output by the decoding layer. This represents the prediction result for the i-th decoding layer. proportionally Perform upsampling operation. This indicates splicing along the channel dimension. .

[0032] Step 3: Using the CT images obtained in Step 1, the corresponding aortic valve calcification labels, and the right coronary ostium temporal labels, train the aortic valve calcification segmentation model constructed in Step 2. This specifically includes:

[0033] 3.1 Data sampling and class balance, specifically including:

[0034] A) Screening of raw CT images: All images containing aortic valve calcification were included in the training set to ensure that the model made full use of the data and learned the calcification features and their diversity;

[0035] B) For images without aortic valve calcification, given the high similarity of image information between consecutive frames, in order to reduce training redundancy and balance the distribution of training data, only the last frame is extracted from every 4 consecutive frames of images without aortic valve calcification and added to the training set to improve training efficiency while ensuring the representativeness of negative samples.

[0036] 3.2 Constructing Weighted Scale-Sensitive and Location-Sensitive Loss Functions for Model Training: Based on the scale-sensitive loss function, and considering the spatial distribution characteristics of aortic valve calcification and the directionality of anatomical structures, different loss weights are constructed for different locations in the image to enhance the model's attention to error-prone areas. During model training, the original segmentation labels are downsampled to construct labels of corresponding sizes, and then the prediction results obtained from each layer of the decoding module are compared. and final segmentation prediction results The losses are calculated separately, and the final total loss function is the average of the sum of the losses of all prediction results. Specifically: Total Loss Function ,in: This represents the prediction loss of each decoding layer. This represents the loss in the final segmentation result. , This indicates the warm-up iteration rounds, meaning that only standard IoU loss is used in the first ten training rounds. Train the system until it stabilizes, then use the weighted scale-sensitive loss. and location-sensitive loss train.

[0037] The scale-sensitive loss function and location-sensitive loss function mentioned herein employ, but are not limited to, the techniques described in Liu Q, Liu R, Zheng B, et al. Infrared small target detection with scale and locationsensitivity[C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2024: 17490-17499.

[0038] The standard IoU loss function ,in: , .

[0039] The weighted scale-sensitive loss function is as follows: ,in: , , For the predicted results, For real labels, Weighting function , , The centroid of the aortic region is the center of mass. The angle between the aortic valve calcification segmentation result and the centroid. This is a preset angle value. Determine the focus area and set the weight of the pixels in that area to... The factor is multiplied to emphasize its contribution to the loss function, guiding the network to learn error-prone calcifications on the coronary ostium and aortic wall.

[0040] In this embodiment, Set to 1.2. Set as .

[0041] Step 4, Segmentation Stage: Acquire test set CT images, input the segmentation model trained in Step 3, and output the aortic valve calcification segmentation results.

[0042] Through specific practical experiments, using a single NVIDIA A100 graphics card, in a PyTorch 2.6.0 and Python 3.12.9 environment, the initial learning rate was set to 0.001, the Adam optimizer was used, the weight decay was set to 1e-5, the batch size was set to 4, the total epochs were set to 250, and the experimental data was divided into training and test sets in an 8:2 ratio to carry out the experiment.

[0043] The segmentation evaluation metrics involved in the experiment included the Dice similarity coefficient (Dice), intersection over union (IoU), recall, precision, and 95% Hausdorff distance (HD95). Specifically: , , , , , where: set and These are segmentation labels and segmentation predictions, respectively. For positive samples that are predicted to be of the positive class, For negative samples that are predicted to be positive, For negative samples that are predicted to be negative, and Sets All points to set Maximum distance to nearest point and set All points to set The maximum value of the nearest point distance. The Dice coefficient and intersection-union ratio reflect the degree of overlap between the segmentation prediction and the segmentation label, the recall reflects the model's ability to detect positive samples, the precision reflects the proportion of all samples predicted as positive that are actually positive, tends to penalize false positives, and HD95 reflects the degree of overlap of the segmentation boundary.

[0044] The effectiveness of this invention was further verified through ablation experiments: Dice and IoU were used as evaluation indicators, and the results are shown in Table 1. Higher Dice and IoU indicate a greater overlap between the segmentation result and the segmentation label, resulting in more accurate segmentation. It can be seen that by incorporating prior location information encoding, based on the right coronary artery ostium temporal label... By employing a branch decoding structure and introducing a weighted IoU loss function into the loss function, the segmentation performance of the model is improved. Compared with the baseline model, the Dice coefficient is improved by 3.7% and the IoU is improved by 6.8%, demonstrating the effectiveness of the present invention in accurately segmenting aortic valve calcification.

[0045] Table 1

[0046] To demonstrate the superior performance of our proposed method in aortic valve calcification segmentation, we compared it with several other segmentation methods. The quantitative results of the comparative experiments are shown in Table 2. Compared to existing techniques, our method exhibits higher segmentation accuracy and robustness. Compared to MSHNet, the Dice coefficient and IoU improved by 1.2% and 2.2%, respectively, while HD95 decreased by 0.07, demonstrating that our method provides more accurate segmentation and better edge segmentation. Compared to U-Net, Recall improved by 1.9% and Precision by 7.3%, proving that our method segments more targets while maintaining a low false positive rate.

[0047] Table 2

[0048] In addition, the qualitative results of the comparative experiment are as follows Figure 6 As shown in the figure, comparing the results of the four segmentation methods, the present invention and the MSHNet method demonstrate stronger capabilities in capturing small calcified lesions, achieving fine segmentation of blurred calcified edges. Simultaneously, the present invention provides more complete segmentation results for larger calcified regions, outperforming other methods. These comparative results demonstrate that the present invention effectively integrates and utilizes the spatial prior information of the aortic structure and the temporal features of CT sequences, enhancing the segmentation ability of calcified lesions within complex anatomical regions and resulting in superior network performance.

[0049] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A system for segmenting aortic valve calcifications in plain CT images based on spatiotemporal priors, characterized in that, include: The system includes a location prior information extraction module, an encoding module, and a decoding module. The location prior information extraction module encodes the initial aortic segmentation results obtained from the original CT image into attention weights. The encoding module extracts multi-scale image features from the original CT images and combines them with attention weights to generate weighted images that guide the network to focus on the aortic region. The decoding module restores the resolution of the weighted images through convolution and upsampling operations and uses a branching structure to achieve more accurate segmentation of different images in the sequence.

2. The aortic valve calcification segmentation system based on spatiotemporal prior in plain CT images according to claim 1, characterized in that, The location prior information extraction module includes an automatic segmentation unit, a mask preprocessing unit, a downsampling unit, and a mask encoding unit. Specifically: the automatic segmentation unit inputs the original CT image obtained from instrument scanning into a general segmentation model to obtain an initial aortic segmentation result; the mask preprocessing unit filters and dilates the aortic segmentation result to obtain a usable aortic segmentation mask; the downsampling unit performs multiple downsampling operations on the aortic segmentation mask to obtain a set of aortic segmentation masks at different scales; and the mask encoding unit uses multiple convolutions to generate a spatial attention map from the aortic segmentation mask, outputting it as a spatial attention weight to the encoding module for fusion with the feature map extracted from the original image, thereby enhancing the model's ability to distinguish the aortic valve and its adjacent structures.

3. The aortic valve calcification segmentation system based on spatiotemporal prior in plain CT images according to claim 1, characterized in that, The encoding module includes four encoding layers, each comprising a residual convolutional unit and a prior fusion unit. Each residual convolutional unit includes two convolutional blocks based on residual connections, used to extract semantic information from the raw CT image obtained from the input instrument scan or the feature map output by the previous encoding layer, output a feature map, and pass it to the decoding module through skip connections. The prior fusion unit weights the feature map output by each residual convolutional unit according to attention weights to generate a weighted image used to enhance the model's perception ability of the aortic-related region.

4. The aortic valve calcification segmentation system based on spatiotemporal prior in plain CT images according to claim 1, characterized in that, The decoding module specifically includes: four decoding layers composed of residual convolutional units, a branch selection unit located between the third and fourth decoding layers, and a multi-scale output unit connected to the output of each decoding layer. Each decoding layer progressively amplifies the deep feature map through upsampling and uses skip connections between the encoding and decoding layers at corresponding scales to reduce information loss. The branch selection unit activates the corresponding decoding branch based on the temporal label of the original CT image, achieving differential segmentation of restored images with different structural characteristics. The multi-scale output unit receives the restored images output by multiple residual convolutional units, upsamples them uniformly to the original image size, and generates the final segmentation result.

5. The aortic valve calcification segmentation system based on spatiotemporal prior in plain CT images according to claim 4, characterized in that, The residual convolutional unit includes two convolutional blocks based on residual connections. It concatenates the weighted image or the magnified deep feature map output from the previous decoding layer with the feature map from the skip connection of the encoding module. The spatial resolution of the image is gradually restored through convolution to generate a restored image, thereby achieving pixel-level segmentation of the aortic calcification region.

6. A method for segmenting aortic valve calcification in plain CT images based on spatiotemporal priors of the system described in any one of claims 1-5, characterized in that, include: Step 1: Obtain experimental data; Step 2: Construct an aortic valve calcification segmentation model. The initial aortic segmentation result obtained from the original CT image is encoded into attention weights by the location prior information extraction module. Multi-scale image features are extracted from the original CT image by the encoding module and combined with the attention weights to generate a weighted image that can guide the network to focus on the aortic region. The resolution of the weighted image is restored by performing convolution and upsampling operations by the decoding module. A branch structure is used to achieve more accurate segmentation of different images in the sequence. Step 3: Using the CT images obtained in Step 1, the corresponding aortic valve calcification labels, and the right coronary artery ostium temporal labels, train the aortic valve calcification segmentation model constructed in Step 2. Step 4, Segmentation Stage: Acquire test set CT images, input the segmentation model trained in Step 3, and output the aortic valve calcification segmentation results.

7. The method for segmenting aortic valve calcification in plain CT images based on spatiotemporal priors according to claim 6, characterized in that, Step 2 specifically includes: 2.1 After obtaining the initial aortic segmentation result for each image from the original CT images through general segmentation, the initial aortic segmentation result is preprocessed and encoded; The preprocessing specifically includes: since the aortic valve is located in the ascending aorta, only the segmentation results of the ascending aorta region where calcification may occur need to be retained; connected components in the original aortic segmentation mask are detected, and only the ascending aorta region closer to the image center is retained; considering that the segmentation of the automatic segmentation tool is relatively coarse, a 13x13 dilation kernel is used to morphologically expand the ascending aorta region to further cover the candidate aorta region, finally obtaining the aortic segmentation mask. , , The height and width of the image; The encoding specifically includes: aortic segmentation mask. Perform separately Subsampling with a second kernel of 2 yields a set of aortic segmentation masks. Then, for each mask, use two concatenated... After extracting spatial information using convolution, a 1×1 convolution is used to compress the number of channels to 1, followed by normalization using the Sigmoid function. The final output is the multi-scale spatial attention weights. ,in: The size is , This represents the Sigmoid function. Indicates the kernel size as Convolutional operations, multi-scale spatial attention weights This information, used as prior information about aortic anatomy, will be fused with the features output from each coding layer in the coding module. In the 2.2 encoding module, each residual convolutional unit's residual convolutional block sequentially performs 3×3 convolution, batch normalization, and modified linear units to extract features from the input features. Then, cascaded channel attention and spatial attention are used to capture key features. A 1×1 convolution is then performed to adjust the number of channels in the processed feature map to match the input features. Finally, the input feature map and the processed feature map are superimposed. Specifically... : , Given the input feature map, The output feature map, For channel attention, For spatial attention, For the Sigmoid function, and For weight parameters, To correct the linear unit, For batch normalization operations, and These are average pooling and max pooling operations along the channel dimension, respectively. and These are respectively the summation and maximum value operations performed along the channel dimension; 2.3 Each prior fusion unit in the encoding module injects the aortic anatomical prior information at different resolutions captured by the location prior information extraction module into the encoding network, and combines it with the output features of each residual convolutional unit. Multiplication and superposition are used to guide the network to focus on the aortic region, enhance the specificity of the target location, and suppress interference from the anatomical neighborhood. Specifically: ,in: The fused feature map is the output of the prior fusion unit. This represents the output of the residual convolution unit. For Hadamah accumulation; 2.4 The branch selection unit in the decoding module selects based on the label of the input image. The system automatically selects between two residual convolutional units to achieve targeted predictions for images with different anatomical states; the multi-scale output unit converts the output of each decoding layer... After 1×1 convolution, the segmentation prediction results of each layer are obtained. Then The image is uniformly upsampled to the original image size. The upsampled results are then processed by channel stitching and convolution to obtain the final segmentation prediction result. Specifically: ), ,in: This represents the feature map output by the decoding layer. This represents the prediction result for the i-th decoding layer. proportionally Perform an upsampling operation. This indicates splicing along the channel dimension. .

8. The method for segmenting aortic valve calcification in plain CT images based on spatiotemporal priors according to claim 6, characterized in that, Step 3 specifically includes: 3.1 Data sampling and class balance, specifically including: A) Screening of raw CT images: All images containing aortic valve calcification were included in the training set to ensure that the model makes full use of the data and learns the calcification features and their diversity; B) For images without aortic valve calcification, given the high similarity of image information between consecutive frames, in order to reduce training redundancy and balance the distribution of training data, only the last frame is extracted from every 4 consecutive frames of images without aortic valve calcification and added to the training set to improve training efficiency while ensuring the representativeness of negative samples. 3.2 Constructing Weighted Scale-Sensitive and Location-Sensitive Loss Functions for Model Training: Based on the scale-sensitive loss function, and combining the spatial distribution characteristics of aortic valve calcification and the directionality of anatomical structures, different loss weights are constructed for different locations in the image to enhance the model's attention to error-prone areas. During model training, the original segmentation labels are downsampled to construct labels of corresponding sizes, and then the prediction results obtained from each layer of the decoding module are compared. and final segmentation prediction results The losses are calculated separately, and the final total loss function is the average of the sum of the losses of all prediction results. Specifically: Total Loss Function ,in: This represents the prediction loss of each decoding layer. This represents the loss in the final segmentation result. , This indicates the warm-up iteration rounds, meaning that only standard IoU loss is used in the first ten training rounds. Train the system until it stabilizes, then use the weighted scale-sensitive loss. and location-sensitive loss train.

9. The method for segmenting aortic valve calcification in plain CT images based on spatiotemporal priors according to claim 8, characterized in that, The standard IoU loss function ,in: , ; The weighted scale-sensitive loss function is as follows: ,in: , , For the predicted results, For real labels, Weighting function , , The centroid of the aortic region is the center of mass. The angle between the aortic valve calcification segmentation result and the centroid. For the preset angle value, according to Determine the focus area and set the weight of the pixels in that area to... The factor is multiplied to emphasize its contribution to the loss function, guiding the network to learn error-prone calcifications on the coronary ostium and aortic wall.

Images