A CT image pulmonary embolism segmentation and classification method combined with quality evaluation

Abstract

The application discloses a CT image pulmonary embolism segmentation and classification method combined with quality evaluation, relates to the technical field of image processing, and comprises the following steps: inputting a 256*256 pulmonary embolism CT image and a quality score thereof into a quality score guided encoder, expanding the quality score dimension through linear transformation, point multiplying and fusing with feature maps extracted by ResNet34 layer by layer, and generating multi-scale coding features; carrying out wavelet domain decomposition and reconstruction on the coding features through a wavelet transform fusion skip link module, and optimizing feature transmission; adopting a multi-scale cross enhancement decoder to perform multi-scale deconvolution fusion on the features, outputting a segmentation result in combination with an efficient channel attention mechanism, and outputting a pulmonary embolism existence judgment through a classification head, so that the method provides strong support for early diagnosis of pulmonary embolism, development of an image assisted diagnosis system and clinical application, and has wide application prospect and far-reaching social significance.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a CT image pulmonary embolism segmentation and classification method that incorporates quality assessment. Background Technology

[0002] As a life-threatening disease, early diagnosis and accurate segmentation of pulmonary embolism are key to improving cure rates and reducing mortality.

[0003] However, existing methods for detecting pulmonary embolism still have limitations when dealing with complex images, especially in terms of image quality assessment and accurate segmentation.

[0004] To address the aforementioned issues, we propose a CT image pulmonary embolism segmentation and classification method that incorporates quality assessment. Summary of the Invention

[0005] To address the problems mentioned in the background section, this invention provides a CT image pulmonary embolism segmentation and classification method that incorporates quality assessment.

[0006] The present invention provides a CT image pulmonary embolism segmentation and classification method that combines quality assessment, and adopts the following technical solution:

[0007] A CT image segmentation and classification method for pulmonary embolism that incorporates quality assessment includes the following steps:

[0008] Step 1: Input the 256×256 pulmonary embolism CT image and its quality score into the quality score-guided encoder. Expand the quality score dimension through linear transformation and fuse it with the feature map extracted by ResNet34 layer by layer to generate multi-scale encoded features.

[0009] Step 2: Perform wavelet domain decomposition and reconstruction of the encoded features by fusing the skip link module with wavelet transform, and optimize feature transmission;

[0010] Step 3: Use a multi-scale cross-enhanced decoder to perform multi-scale deconvolution fusion on the features, and output the segmentation results by combining an efficient channel attention mechanism. At the same time, output the pulmonary embolism presence judgment through the classification head.

[0011] Preferably, in step one, the quality scoring guided encoder includes a Stem module, four encoding modules, and a classification head network architecture;

[0012] The Stem module includes a 7×7 convolutional layer with a stride of 2, a batch normalization layer, and a ReLU activation function. The four encoding modules are stacked in 3, 4, 6, and 3 Res-Blocks respectively. Each Res-Block contains two 3×3 convolutional operations, along with a ReLU activation function and a BN operation.

[0013] Each of the aforementioned encoding modules incorporates a 2×2 convolutional layer with a stride of 2, combined with a batch normalization layer, to perform downsampling operations;

[0014] The quality score guided encoder also includes an image quality score module. The image quality score module expands the quality score S through five different linear layers, gradually increasing its dimension, which are: [b, 3], [b, 64], [b, 128], [b, 256], [b, 512]. The dimension of the quality score matches the dimension of the feature map extracted by ResNet34.

[0015] The quality score is aligned spatially with image features through a data broadcasting mechanism, generating quality guidance information with the same dimension as each feature map layer. The quality score is then expanded to... The corresponding dimensions are: [b, 3, 1, 1], [b, 64, 1, 1], [b, 128, 1, 1], [b, 256, 1, 1], [b, 512, 1, 1].

[0016] Preferably, the input image is By comparing image X with the initial quality score Perform a dot product to obtain weighted features, then input these weighted features into the Stem module to obtain preliminary features. ,feature Will be related to the second phase of quality scoring Perform a dot product to obtain the features after quality score weighting. This data is then fed into a subsequent Res-Block for feature extraction. will with The dot product is performed and then fed into the subsequent Res-Block. After continuous processing, five encoder features are obtained. The formula is as follows:

[0017] ;

[0018] in, For the first One Res-Block function; Use for each floor The quantity is [3, 4, 6, 3].

[0019] Preferably, the classification head network architecture includes an adaptive global average pooling layer, a linear layer, a ReLU activation function, another linear layer, and a Sigmoid activation function;

[0020] An adaptive global average pooling layer performs global pooling on all encoded features, aggregating information from various spatial locations to generate a fixed-dimensional feature vector. This feature vector is then transformed by a linear layer, introducing non-linearity using the ReLU activation function, as shown in the following formula:

[0021] .

[0022] Preferably, the wavelet transform fusion jump link module performs a one-dimensional discrete wavelet transform on the input signal. Given a wavelet function , As a scale factor, For time factor and scaling function Input signal exist The decomposition formula is as follows:

[0023] ;

[0024] in, Represents high-frequency components. This indicates the low-frequency components; the input signal is gradually decomposed into wavelet domain signals.

[0025] Introducing vectors Let represent the high-frequency and low-frequency filters, respectively. The decomposition of the original signal on the wavelet basis can be regarded as a recursive convolution of the original signal with a specific filter at a stride of 2, as shown in the formula:

[0026] ,

[0027] ,

[0028] The inverse wavelet transform is obtained through two composite vectors. This is achieved through transposed convolution. Axial high-frequency information is extracted using two-dimensional discrete wavelet transform, extending the data to two dimensions. A learnable method is designed to adapt to data distribution and feature layers. The formula for constructing the two-dimensional forward wavelet convolution kernel is as follows: ,

[0029] ,

[0030] ,

[0031] Set as a learnable filter, These are low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency convolution operators obtained through vector outer products, which are then concatenated into a wavelet kernel. The learnable wavelet transform is implemented as grouped convolution, given a set of input feature maps. Its projection in the wavelet domain is Generated via wavelet convolution;

[0032] For complex numbers Given a transform filter for an application, its transform can be expressed as: Then get Corresponding transformation And satisfy the condition formula:

[0033] ;

[0034] The input wavelet domain components are separated into separate dimensions. Wavelet domain features are extracted and transformed through depthwise separable convolution with an expansion factor. Channel expansion and scaling are performed through convolution. The wavelet domain feature map is restored to the spatial domain for output through learnable wavelet inverse transform, thus forming a learnable wavelet network with an expansion factor of a, where a is 2 or 6.

[0035] Preferably, the multi-scale cross-enhanced decoder, for the last feature output by the encoder. The feature is divided into four identical feature blocks along the channel dimension. Each feature block is upsampled using deconvolution modules of different sizes, with kernel sizes of 2×2, 4×4, 6×6, and 8×8, to obtain the upsampled features. ;

[0036] Will The features are concatenated along the channel dimension to obtain features that have undergone parallel multi-scale deconvolution upsampling. Using features processed by WF-SC Then, the Sigmoid function is used to... and We perform weighting to obtain their respective weights. and ;

[0037] These weights are cross-multiplied with the corresponding features and then summed to obtain the fused features. These features are then fed into the channel attention mechanism for attention feature enhancement to obtain the output of the decoder module, as shown in the following formula:

[0038] ;

[0039] in, This represents an efficient channel attention mechanism. It uses a global average pooling layer and dimension transformation to obtain a 1×C feature map, and then calculates the size of the adaptive convolution kernel in the one-dimensional convolution operation. The formula is:

[0040] ;

[0041] In the formula, The number of channels in the feature map. and These are all hyperparameters, and their default settings are used in this paper. The operation of finding the nearest odd number can be represented as: ;

[0042] In the formula, As input features, For output features, The kernel size is The adaptive one-dimensional convolution obtains channel weights through the sigmoid activation function, and multiplies the input feature map with the weights to obtain the output of the efficient channel attention mechanism, i.e., the output of the multi-scale cross-enhanced decoder. The subsequent decoder module sequentially fuses the features from the skip connections with the features from the lower-level decoder, performing feature upsampling. The specific formula is as follows:

[0043] ;

[0044] In the formula This indicates that multi-scale crossover adds decoding operations.

[0045] Preferably, the quality score is assessed by a radiologist on a 5-point scale, and calculated in conjunction with objective indicators such as SNR and CNR.

[0046] The noise value is the standard deviation of the background area.

[0047] Preferably, the network training uses the Adam optimizer with an initial learning rate of 0.0001, a weight decay of 0.1, a decay every 25 rounds, 100 rounds of five-fold cross-validation, and a batch size of 16.

[0048] In summary, the present invention has the following beneficial technical effects:

[0049] First, a quality score-guided encoder is used. By fusing the image quality score with the image data and combining it with a ResNet34 network for feature extraction, the guiding role of image quality in classification and segmentation tasks is enhanced. Second, a wavelet transform fusion skip link module is used to further refine image features through wavelet transform, enhancing the model's ability to capture details. Finally, a multi-scale cross-enhanced decoder module is employed. By fusing multi-scale information to optimize the decoder output, the segmentation accuracy is improved. This not only enables efficient detection of pulmonary embolism but also precise segmentation of lesion areas. Through quality score guidance and feature fusion techniques, this system demonstrates high robustness on images of varying quality, addressing the problem of insufficient adaptability of existing methods to low-quality images.

[0050] This method provides strong support for the early diagnosis of pulmonary embolism, the development of image-assisted diagnostic systems, and clinical applications, and has broad application prospects and profound social significance. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of a CT image pulmonary embolism segmentation and classification method that combines quality assessment in an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of a module of a CT image pulmonary embolism segmentation and classification method that combines quality assessment in an embodiment of the present invention. Detailed Implementation

[0053] The following is in conjunction with the appendix Figures 1 to 2 The present invention will be described in further detail below.

[0054] This invention discloses a CT image pulmonary embolism segmentation and classification method that incorporates quality assessment. (Refer to...) Figures 1 to 2 A CT image segmentation and classification method for pulmonary embolism that incorporates quality assessment includes the following steps:

[0055] Step 1: Input the 256×256 pulmonary embolism CT image and its quality score into the quality score-guided encoder. Expand the quality score dimension through linear transformation and fuse it with the feature map extracted by ResNet34 layer by layer to generate multi-scale encoded features.

[0056] Step 2: Perform wavelet domain decomposition and reconstruction of the encoded features by fusing the skip link module with wavelet transform, and optimize feature transmission;

[0057] Step 3: Use a multi-scale cross-enhanced decoder to perform multi-scale deconvolution fusion on the features, and output the segmentation results by combining an efficient channel attention mechanism. At the same time, output the pulmonary embolism presence judgment through the classification head.

[0058] By leveraging ResNet34's deep feature extraction capabilities and optimization strategies, multi-level information in images can be effectively processed, thereby improving the performance of pulmonary embolism classification and segmentation tasks.

[0059] Specifically, the quality score-guided encoder includes a Stem module, four encoding modules, and a classification head network architecture;

[0060] The Stem module includes a 7×7 convolutional layer with a stride of 2, a batch normalization layer, and a ReLU activation function. The four encoding modules are stacked with 3, 4, 6, and 3 Res-Blocks respectively. Each Res-Block contains two 3×3 convolutional operations, along with a ReLU activation function and a BN operation.

[0061] Each encoding module incorporates a 2×2 convolutional layer with a stride of 2, combined with a batch normalization layer, to perform downsampling operations;

[0062] Specifically, the introduction of 7×7 large-kernel convolutional layers not only slightly increases the parameters but also expands the receptive field, helping the network capture global features in the image. This method helps extract rich low-level features in the initial stage of the network and provides a solid foundation for subsequent deep feature extraction. After the convolutional layers, the use of batch normalization layers can accelerate network training, reduce internal covariance shift, and enhance model stability. The four subsequent encoding modules are stacked with 3, 4, 6, and 3 Res-Blocks respectively (e.g., ...). Figure 1 As shown, each Res-Block contains two 3×3 convolutional operations, combined with ReLU activation function and BN operation to enhance the non-linear representation of features and the stability of training. Through this structure, the network can gradually extract and strengthen higher-level features in the image, thereby enabling the model to have stronger representation capabilities in subsequent tasks. In order to further optimize the spatial complexity and feature learning efficiency of the network, a 2×2 convolutional layer with a stride of 2 is added to each encoding module, combined with a batch normalization layer to perform downsampling operation. Each downsampling operation can not only halve the size of the feature map, but also effectively reduce the number of channels, thereby significantly reducing the feature dimension and network parameters and enhancing the computational efficiency of the network.

[0063] The quality score guided encoder also includes an image quality score module. The image quality score module expands the quality score S through five different linear layers, gradually increasing its dimension: [b, 3], [b, 64], [b, 128], [b, 256], [b, 512]. The dimension of the quality score matches the dimension of the feature map extracted by ResNet34.

[0064] The quality score is aligned spatially with image features through a data broadcasting mechanism, generating quality guidance information with the same dimension as each feature map layer. The quality score is then expanded to... The corresponding dimensions are: [b, 3, 1, 1], [b, 64, 1, 1], [b, 128, 1, 1], [b, 256, 1, 1], [b, 512, 1, 1].

[0065] Specifically, the input image is By comparing image X with the initial quality score Perform a dot product to obtain weighted features, then input these weighted features into the Stem module to obtain preliminary features. ,feature Will be related to the second phase of quality scoring Perform a dot product to obtain the features after quality score weighting. This data is then fed into a subsequent Res-Block for feature extraction. will with The dot product is performed and then fed into the subsequent Res-Block. After continuous processing, five encoder features are obtained. The formula is as follows:

[0066] ;

[0067] in, For the first One Res-Block function; Use for each floor The quantity is [3, 4, 6, 3].

[0068] Specifically, the classification head network architecture includes an adaptive global average pooling layer, a linear layer, a ReLU activation function, another linear layer, and a Sigmoid activation function;

[0069] An adaptive global average pooling layer performs global pooling on all encoded features, aggregating information from various spatial locations to generate a fixed-dimensional feature vector. This feature vector is then transformed by a linear layer, introducing non-linearity using the ReLU activation function, as shown in the following formula:

[0070] .

[0071] By dynamically integrating image quality scores into the feature extraction process, the model's segmentation and classification performance for pulmonary embolism in CT images is improved. The core of this approach lies in constructing a hierarchical, cascaded quality guidance mechanism: firstly, low-dimensional quality scores are expanded to dimensions matching the feature maps of each layer of ResNet34 through multi-level linear transformations. Then, a broadcast mechanism is used to achieve automatic spatial alignment, allowing quality information to be fused with image features layer by layer through dot-multiplication, thereby guiding the network to dynamically adjust feature weight allocation based on image quality. This design not only preserves the spatial integrity of the original features but also enhances the model's adaptability to low-quality images through quality-aware weight modulation, effectively solving the feature shift problem caused by differences in medical image quality. In the feature decoding stage, a multi-scale classification head is introduced to aggregate multi-level features through global average pooling. Combined with nonlinear activation functions and linear transformations, this achieves efficient mapping from complex feature spaces to classification decisions. This hierarchical quality fusion strategy enables the network to simultaneously capture local details and global semantic information, exhibiting higher sensitivity and robustness in pulmonary embolism detection tasks, providing a reliable quality-aware feature foundation for subsequent accurate segmentation.

[0072] Specifically, the wavelet transform fusion jump link module performs a one-dimensional discrete wavelet transform on the input signal. Given a wavelet function , As a scale factor, For time factor and scaling function Input signal exist The decomposition formula is as follows:

[0073] ;

[0074] in, Represents high-frequency components. This indicates the low-frequency components; the input signal is gradually decomposed into wavelet domain signals.

[0075] Introducing vectors Let represent the high-frequency and low-frequency filters, respectively. The decomposition of the original signal on the wavelet basis can be regarded as a recursive convolution of the original signal with a specific filter at a stride of 2, as shown in the formula:

[0076] ,

[0077] ,

[0078] The inverse wavelet transform is obtained through two composite vectors. This is achieved through transposed convolution. Axial high-frequency information is extracted using two-dimensional discrete wavelet transform, extending the data to two dimensions. A learnable method is designed to adapt to data distribution and feature layers. The formula for constructing the two-dimensional forward wavelet convolution kernel is as follows:

[0079] ,

[0080] ,

[0081] ,

[0082] Set as a learnable filter, These are low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency convolution operators obtained through vector outer products, which are then concatenated into a wavelet kernel. The learnable wavelet transform is implemented as grouped convolution, given a set of input feature maps. Its projection in the wavelet domain is Generated via wavelet convolution; for complex numbers Given a transform filter for an application, its transform can be expressed as: Then get Corresponding transformation And satisfy the condition formula:

[0083] ,

[0084] The input wavelet domain components are separated into separate dimensions. Wavelet domain features are extracted and transformed through depthwise separable convolution with an expansion factor. Channel expansion and scaling are performed through convolution. The wavelet domain feature map is restored to the spatial domain for output through learnable wavelet inverse transform, thus forming a learnable wavelet network with an expansion factor of a, where a is 2 or 6.

[0085] By introducing a learnable wavelet transform fusion skip connection mechanism, the ability of multi-scale networks to recover detailed features of medical images is improved. Its core innovation lies in combining the time-frequency adaptive characteristics of traditional wavelet transform with deep neural networks: learnable low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency convolutional kernels are constructed based on two-dimensional discrete wavelet transform. Multi-directional filter banks are generated through outer product operations, achieving bidirectional mapping of image features between the spatial and wavelet domains. The filter parameters are constrained through the principle of perfect reconstruction, preserving the mathematical completeness of wavelet transform while endowing the network with the ability to adaptively adjust its frequency response range. In the feature decoding stage, the collaborative design of deep separable convolution and inverse wavelet transform effectively decouples the expression space of high-frequency details from low-frequency semantic information. Through multi-layered cascaded wavelet head modules and fusion modules, refined reconstruction of frequency features across scales is achieved. This mechanism fully utilizes the sensitivity of wavelet transform to abrupt signals, supplementing the high-frequency edge and texture information lost in traditional convolution through skip connections. While maintaining the spatial resolution of the feature map, it significantly enhances the localization accuracy of micro-embolic lesions. This differentiable wavelet domain feature interaction framework enables the network to automatically balance the representation weights of global semantics and local details during end-to-end training, providing a more robust multi-scale feature fusion scheme for complex medical image segmentation tasks.

[0086] Specifically, the multi-scale cross-enhanced decoder, for the last feature of the encoder output... The feature is divided into four identical feature blocks along the channel dimension. Each feature block is upsampled using deconvolution modules of different sizes, with kernel sizes of 2×2, 4×4, 6×6, and 8×8, to obtain the upsampled features. ;

[0087] Will The features are concatenated along the channel dimension to obtain features that have undergone parallel multi-scale deconvolution upsampling. Using features processed by WF-SC Then, the Sigmoid function is used to... and We perform weighting to obtain their respective weights. and ;

[0088] These weights are cross-multiplied with the corresponding features and then summed to obtain the fused features. These features are then fed into the channel attention mechanism for attention feature enhancement to obtain the output of the decoder module, as shown in the following formula:

[0089] ;

[0090] in, This represents an efficient channel attention mechanism. It uses a global average pooling layer and dimension transformation to obtain a 1×C feature map, and then calculates the size of the adaptive convolution kernel in the one-dimensional convolution operation. The formula is:

[0091] ;

[0092] In the formula, The number of channels in the feature map. and These are all hyperparameters, and their default settings are used in this paper. The operation of finding the nearest odd number can be represented as: ;

[0093] In the formula, As input features, For output features, The kernel size is The adaptive one-dimensional convolution obtains channel weights through the sigmoid activation function, and multiplies the input feature map with the weights to obtain the output of the efficient channel attention mechanism, i.e., the output of the multi-scale cross-enhanced decoder. The subsequent decoder module sequentially fuses the features from the skip connections with the features from the lower-level decoder, performing feature upsampling. The specific formula is as follows:

[0094] ;

[0095] In the formula This indicates that multi-scale crossover adds decoding operations.

[0096] The design of a multi-scale cross-enhanced decoder (MC-Decoder) improves the accuracy of localization and feature recovery of small lesions in medical image segmentation tasks. Its core advantage lies in the integration of multi-scale large-kernel deconvolution and attention mechanisms: by dividing the encoded features into four parallel branches and upsampling them with deconvolution kernels of different sizes from 2×2 to 8×8, it reduces computational complexity while achieving differentiated extraction of multi-resolution features, effectively alleviating the problem of high-frequency detail loss caused by single upsampling operations. The introduction of channel-dimensional concatenation and sigmoid weighted cross-multiplication strategies enables dynamic interaction between low-level edge features and deep semantic features through skip connections, strengthening the complementarity of cross-level features. Combined with an efficient channel attention (ECA) mechanism, it uses adaptive one-dimensional convolution kernels to automatically calibrate the weight distribution of each channel, accurately focusing on key feature regions while reducing the number of parameters. This decoder expands the receptive field coverage through large-kernel deconvolution and, combined with the multi-scale context awareness capability of dilated convolution, solves the problem of traditional methods struggling to capture small embolic lesions on low-resolution feature maps. Its divide-and-conquer fusion architecture not only ensures the integrity of multi-scale feature representation, but also optimizes the continuity of segmentation boundaries through an attention-guided feature enhancement mechanism, ultimately achieving better detail preservation and spatial consistency in the pulmonary embolism detection task.

[0097] Specifically, the quality score is assessed by radiologists on a 5-point scale, and calculated in conjunction with objective indicators such as SNR and CNR.

[0098] The noise value is the standard deviation of the background area.

[0099] Specifically, the network training uses the Adam optimizer with an initial learning rate of 0.0001, a weight decay of 0.1, a decay every 25 rounds, 100 rounds of five-fold cross-validation, and a batch size of 16.

[0100] experiment:

[0101] After the image data was exported from the server in DICOM format, a physician with over three years of experience in radiology was responsible for delineating regions of interest (ROIs) on a GE Advanced Workstation post-processing workstation. To ensure measurement accuracy, all ROIs were set to circular shapes, and the largest possible area was selected for each measurement to avoid interference from factors such as thrombi, calcifications, and plaques on the vessel walls. ROIs included the main pulmonary artery, left pulmonary artery trunk, right pulmonary artery trunk, and latissimus dorsi muscle. For each ROI, its CT and SD values ​​were recorded for subsequent analysis.

[0102] Two radiologists with over 5 years of experience in CTPA diagnosis were responsible for independently conducting randomized, anonymous reviews of images on standardized display equipment, and providing subjective scores based on image quality. The subjective evaluation criteria for CTPA images were as follows: 5 points - Excellent image quality, sharp and clear details, no artifacts; 4 points - Good image quality, clear boundaries, only a few artifacts, fully meeting clinical diagnostic requirements; 3 points - Moderate image quality, although some quality issues exist, it still meets basic diagnostic requirements; 2 points - Poor image quality, with obvious artifacts or blurring, reducing the confidence in the diagnostic results; 1 point - Very poor image quality, almost unrecognizable, containing no information suitable for diagnosis. Only images with a subjective score of 3 or higher were considered to meet the basic requirements for clinical diagnosis. The average subjective score of all images participating in the evaluation was used as the mean opinion score (MOS). This scoring mechanism ensured the objectivity and accuracy of the evaluation, providing an important basis for subsequent analysis.

[0103] All images showing pulmonary embolism (PE) were manually delineated layer by layer in axial views using 3D Slicer software (version 5.4.0) by two radiologists with three years of experience. The segmentation results were then reviewed and confirmed by a radiologist with seven years of experience. To achieve accurate segmentation, all CT slice data for each PE case, from its first appearance to complete disappearance, were recorded and saved, with slices without PE used as negative training samples. In the data preprocessing stage, the window width and window level of all CTPA images were first set to 700 and 100, respectively. Then, a min-max normalization method was used to map all pixel values ​​in the images to the range [0, 1]. The image pixel size used in the experiment was adjusted to 512 x 512, and the segmentation area was enlarged by 50% to better fit the image. To further improve the clarity of the segmented area, a contrast-limited adaptive histogram equalization technique was applied, ultimately forming a three-channel image.

[0104] Considering the unique characteristics of CTPA images, the data augmentation strategy did not include brightness adjustment or random cropping to avoid affecting the accuracy of embolus location identification and segmentation. Accordingly, geometric transformations such as vertical rotation, horizontal rotation, and horizontal flipping were employed to quadruple the sample size, thereby improving the model's generalization ability. The dataset was randomly divided into five equal parts, with four parts used for training and the remaining part for testing. Since the division was based on case-specific data, there was no case-level data leakage issue. Five-fold cross-validation was used to fully evaluate the model's performance and stability.

[0105] The model designed in this invention is implemented using Python 3.7 and PyTorch 1.12. The image processing workstation is equipped with an Intel i9-13900K CPU and two NVIDIA RTX 4090 GPUs with 24GB of video memory. The initial parameters during model training are obtained using PyTorch's default initialization method, and the Adam optimizer is used to update the network parameters. Specifically, the initial learning rate is set to 0.0001, the weight decay coefficient is 0.1, the learning rate decays every 25 iterations, and the model is trained for a total of 100 iterations.

[0106] When training the model, the Dice function is used as the loss function for the segmentation results. The formula is as follows:

[0107] ;

[0108] In the formula, This serves as a true mask for pulmonary embolism on CT images. This serves as a mask for the model's predictions. The cross-entropy function is used as the loss function for the classification results. The formula is as follows:

[0109] ;

[0110] In the formula, The label value indicates the presence of pulmonary embolism in the image. This indicates the result of the model classification. , The image indicates the presence of a pulmonary embolism. This indicates that there is no pulmonary embolism in the image.

[0111] like Figure 1 As shown, the model's learning objectives include the detection and segmentation of pulmonary embolism; therefore, the final loss function is expressed as follows:

[0112] ;

[0113] In the formula, This represents the weighting factor, which is set to 0.6 in this experiment.

[0114] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A CT image pulmonary embolism segmentation and classification method incorporating quality assessment, characterized in that, Includes the following steps: Step 1: Input the 256×256 pulmonary embolism CT image and its quality score into the quality score-guided encoder. Expand the quality score dimension through linear transformation and fuse it with the feature map extracted by ResNet34 layer by layer to generate multi-scale encoded features. Step 2: Perform wavelet domain decomposition and reconstruction of the encoded features by fusing the skip link module with wavelet transform, and optimize feature transmission; Step 3: Use a multi-scale cross-enhanced decoder to perform multi-scale deconvolution fusion on the features, and output the segmentation results by combining an efficient channel attention mechanism. At the same time, output the pulmonary embolism presence judgment through the classification head. In step one, the quality scoring guided encoder includes a Stem module, four encoding modules, and a classification head network architecture; The Stem module includes a 7×7 convolutional layer with a stride of 2, a batch normalization layer, and a ReLU activation function. The four encoding modules are stacked in 3, 4, 6, and 3 Res-Blocks respectively. Each Res-Block contains two 3×3 convolutional operations, along with a ReLU activation function and a BN operation. Each of the aforementioned encoding modules incorporates a 2×2 convolutional layer with a stride of 2, combined with a batch normalization layer, to perform downsampling operations; The quality score guided encoder also includes an image quality score module. The image quality score module expands the quality score S through 5 different linear layers, gradually increasing its dimension, namely: [b,3], [b, 64], [b, 128], [b, 256], [b, 512]. The dimension of the quality score matches the dimension of the feature map extracted by ResNet34. The quality score is aligned spatially with image features through a data broadcasting mechanism, generating quality guidance information with the same dimension as each feature map layer. The quality score is then expanded to... The corresponding dimensions are: [b, 3, 1, 1], [b, 64, 1, 1], [b, 128, 1, 1], [b, 256, 1, 1], [b, 512, 1, 1].

2. The CT image pulmonary embolism segmentation and classification method combined with quality assessment according to claim 1, characterized in that: The input image is By comparing image X with the initial quality score Perform a dot product to obtain weighted features, then input these weighted features into the Stem module to obtain preliminary features. ,feature Will be related to the second phase of quality scoring Perform a dot product to obtain the features after quality score weighting. This data is then fed into a subsequent Res-Block for feature extraction. will with The dot product is performed and then fed into the subsequent Res-Block. After continuous processing, five encoder features are obtained. The formula is as follows: in, For the first One Res-Block function; Use for each floor The quantity is [3, 4, 6, 3].

3. The CT image pulmonary embolism segmentation and classification method combined with quality assessment according to claim 1, characterized in that: The classification head network architecture includes an adaptive global average pooling layer, a linear layer, a ReLU activation function, another linear layer, and a Sigmoid activation function; An adaptive global average pooling layer performs global pooling on all encoded features, aggregating information from various spatial locations to generate a fixed-dimensional feature vector. This feature vector is then transformed by a linear layer, introducing non-linearity using the ReLU activation function, as shown in the following formula: 。 4. The CT image pulmonary embolism segmentation and classification method combined with quality assessment according to claim 1, characterized in that: The wavelet transform fusion jump link module, on the one-dimensional discrete wavelet transform, for the input signal... Given a wavelet function , As a scale factor, For time factor and scaling function Input signal exist The decomposition formula is as follows: ； in, Represents high-frequency components. This indicates the low-frequency components; the input signal is gradually decomposed into wavelet domain signals. Introducing vectors , Let represent the high-frequency and low-frequency filters, respectively. The decomposition of the original signal on the wavelet basis can be regarded as a recursive convolution of the original signal with a specific filter at a stride of 2, as shown in the formula: ； ； The inverse wavelet transform is obtained through two composite vectors. This is achieved through transposed convolution. Axial high-frequency information is extracted using two-dimensional discrete wavelet transform, extending the data to two dimensions. A learnable method is designed to adapt to data distribution and feature layers. The formula for constructing the two-dimensional forward wavelet convolution kernel is as follows: ； ； ； Set as a learnable filter, These are low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency convolution operators obtained through vector outer products, which are then concatenated into a wavelet kernel. The learnable wavelet transform is implemented as grouped convolution, given a set of input feature maps. Its projection in the wavelet domain is Generated via wavelet convolution; For complex numbers Given a transform filter for an application, its transform can be expressed as: Then get Corresponding transformation And satisfy the condition formula: ； ； The input wavelet domain components are separated into separate dimensions. Wavelet domain features are extracted and transformed through depthwise separable convolution with an expansion factor. Channel expansion and scaling are performed through convolution. The wavelet domain feature map is restored to the spatial domain for output through a learnable wavelet inverse transform, thus forming a learnable wavelet network with an expansion factor of a, where a is 2 or 6.

5. The CT image pulmonary embolism segmentation and classification method combined with quality assessment according to claim 1, characterized in that: The multi-scale cross-enhanced decoder, for the last feature output by the encoder. The feature is divided into four identical feature blocks along the channel dimension. Each feature block is upsampled using deconvolution modules of different sizes, with kernel sizes of 2×2, 4×4, 6×6, and 8×8, to obtain the upsampled features. ; Will The features are concatenated along the channel dimension to obtain features that have undergone parallel multi-scale deconvolution upsampling. Using features processed by WF-SC Then, the Sigmoid function is used to... and We perform weighting to obtain their respective weights. and ; These weights are cross-multiplied with the corresponding features and then summed to obtain the fused features. These features are then fed into the channel attention mechanism for attention feature enhancement to obtain the output of the decoder module, as shown in the following formula: ； in, This represents an efficient channel attention mechanism. It uses a global average pooling layer and dimension transformation to obtain a 1×C feature map, and then calculates the size of the adaptive convolution kernel in the one-dimensional convolution operation. The formula is: ； In the formula, The number of channels in the feature map. and These are all hyperparameters, and their default settings are used in this paper. The operation of finding the nearest odd number can be represented as: ; In the formula, As input features, For output features, The kernel size is The adaptive one-dimensional convolution obtains channel weights through the sigmoid activation function, and multiplies the input feature map with the weights to obtain the output of the efficient channel attention mechanism, i.e., the output of the multi-scale cross-enhanced decoder. The subsequent decoder module sequentially fuses the features from the skip connections with the features from the lower-level decoder, performing feature upsampling. The specific formula is as follows: ； In the formula This indicates that multi-scale crossover adds decoding operations.

6. The CT image pulmonary embolism segmentation and classification method combined with quality assessment according to claim 1, characterized in that: The quality score is calculated by radiologists on a 5-point scale, combined with objective indicators such as SNR and CNR. , The noise value is the standard deviation of the background area.

7. The CT image pulmonary embolism segmentation and classification method combined with quality assessment according to claim 1, characterized in that: The network was trained using the Adam optimizer with an initial learning rate of 0.0001, a weight decay of 0.1, and a decay every 25 rounds. Five-fold cross-validation was used for 100 rounds of training, and the batch size was 16.

Images