Thrombus feature analysis method and system based on deep learning of lower limb vein data

Abstract

The application discloses a thrombus feature analysis method and system based on deep learning of lower limb vein data, solves the problems of insufficient lower limb thrombus feature recognition and attention focusing deviation. The method comprises the following steps: acquiring an ex vivo lower limb vein data set containing original data and a thrombus region segmentation mask; preprocessing and extracting key layer data covering the thrombus region as analysis data, and constructing a double-channel tensor of the data and the mask; balancing the data and dividing the data into training, verification and test sets; inputting an improved EfficientNet-B3 (additional fine feature convolution layer) and a mask guided multi-head self-attention module to extract thrombus features and focus on the thrombus region, and outputting a classification identifier representing thrombus attributes; generating a thrombus region visual attention distribution map based on the identifier, and synchronously outputting thrombus bounding box, area and other parameters. The system comprises preprocessing, data balancing, model training and reasoning, and result output modules. The application improves the thrombus feature analysis accuracy and result interpretability, and is suitable for automatic analysis of lower limb vein thrombus features.

Description

Technical Field

[0001] This invention relates to the field of medical data processing and analysis technology, specifically to a method and system for thrombosis feature analysis based on deep learning of lower extremity venous data. Background Technology

[0002] Pulmonary thromboembolism (PTE) is a common and highly fatal vascular emergency in clinical practice, with thrombi primarily originating from deep vein thrombosis (DVT) in the lower extremities. Lower extremity vein examination, as a non-invasive and real-time diagnostic technique, is a crucial tool for analyzing thrombus characteristics. The accurate identification and quantification of thrombus features directly impacts subsequent medical decisions. In recent years, deep learning technology has been widely applied in the field of medical data analysis; however, existing lower extremity vein diagnostic equipment generally lacks embedded dedicated intelligent analysis systems, resulting in significant deficiencies in its ability to automatically analyze thrombus features.

[0003] Lower extremity venous data inherently exhibits low signal-to-noise ratios and poor contrast between thrombi and surrounding tissues, making the features of microthrombi easily masked by background noise. In existing technologies, fixed-thickness data extraction during preprocessing often misses the thrombus region, and standardization further weakens the feature information of low-contrast thrombi. Simultaneously, the model's attention mechanism largely relies on data-driven automatic learning, which struggles to actively focus on the thrombus region under the interference of complex factors such as vessel wall texture and blood flow signals. Furthermore, the output results are often single probability values, lacking interpretable representation of the judgment criteria and quantitative parameters of the thrombus, making it difficult to assess the reliability of the analysis results. These problems collectively limit the system's ability to accurately analyze thrombus features, reducing its clinical auxiliary value. Therefore, it is urgent to improve existing technologies to address the core issues of insufficient thrombus feature recognition, biased focus areas, and unquantifiable analysis results. Summary of the Invention

[0004] To address the shortcomings of existing technologies in lower extremity venous thrombosis feature analysis, such as insufficient identification capabilities, bias in the model's region of interest, and lack of interpretability of analysis results, this application discloses a thrombosis feature analysis method and system based on deep learning of lower extremity venous data.

[0005] The technical solution of this application is as follows:

[0006] Firstly, this application discloses a thrombosis feature analysis method based on deep learning of lower limb vein data, comprising: acquiring a lower limb vein dataset, wherein the file contains original lower limb vein data and a thrombosis region segmentation mask registered therewith, the lower limb vein dataset being an ex vivo digital file stored on a computer storage medium; performing preprocessing on the lower limb vein dataset, extracting key layer data covering thrombosis regions from the original lower limb vein data as data to be analyzed, performing size normalization and feature optimization on the data to be analyzed and the thrombosis region segmentation mask respectively, concatenating the processed data to be analyzed and the thrombosis region segmentation mask in the channel dimension to form a two-channel tensor, the two-channel tensor being the preprocessed two-dimensional input; performing data balancing processing on the two-dimensional input, constructing a sample subset with a consistent ratio of negative samples to positive samples, and performing data balancing according to the preprocessing parameters. The sample subset is divided into training, validation, and test sets according to a set ratio. These sets are then used for training, validation, and testing of the deep learning classification model. The dual-channel tensor is input into the deep learning classification model, which uses EfficientNet-B3 with added convolutional layers for subtle feature extraction as its backbone network. During the classification stage, a multi-head self-attention module with a mask-guided mechanism and a thrombus region segmentation mask as a priori is introduced. Thrombus features are extracted from the lower limb vein data through the backbone network, and the self-attention module focuses on the thrombus region, outputting a classification identifier to characterize the thrombus attributes. Based on the classification identifier, a visual attention distribution map of the thrombus region that the model focuses on is generated, and the quantitative parameters of the thrombus region, including the bounding box and area, are output simultaneously.

[0007] Through the above technical solution, this application constructs a dual-channel input containing processed data to be analyzed and a thrombus region segmentation mask, and combines it with an improved deep learning network architecture to effectively extract thrombus features from lower extremity venous data, significantly improving the accuracy of thrombus feature analysis.

[0008] Furthermore, in the preprocessing step, the key layer data consists of intermediate slices of lower limb vein data containing thrombi and their adjacent slices; the size normalization scales the data to be analyzed to a uniform size using bilinear interpolation, and scales the thrombus region segmentation mask to the same size as the data to be analyzed using nearest neighbor interpolation; the feature optimization includes performing grayscale normalization on the data to be analyzed, and performing morphological optimization and binarization on the thrombus region segmentation mask; the two channels of the dual-channel tensor are the optimized data to be analyzed and the optimized thrombus region segmentation mask, respectively, with each dimension matching the other.

[0009] This technical solution ensures the quality and consistency of the input data, laying a solid foundation for the effective training and accurate analysis of subsequent deep learning models.

[0010] Furthermore, the data balancing process includes: matching positive samples with corresponding negative samples from the raw lower limb vein data, and randomly pairing them to construct multiple sample subsets; performing data augmentation on the lower limb vein samples in each sample subset, with augmentation methods including random horizontal flipping, rotation within an angle range, and Gaussian noise addition; merging all augmented sample subsets, and dividing them into training set, validation set, and test set according to a preset ratio to ensure that the ratio of negative samples to positive samples is consistent in each dataset.

[0011] This technical solution effectively addresses the common class imbalance problem in medical image data through a scientific data balancing strategy, thereby improving the stability and generalization ability of model training.

[0012] Furthermore, the construction of the improved EfficientNet-B3 backbone network includes: adjusting the number of input channels of the first convolutional layer of the backbone network to adapt it to the dual-channel tensor; adding a convolutional layer after the penultimate convolutional block of the backbone network, wherein the added convolutional layer is used to enhance thrombus feature extraction; initializing the backbone network using pre-trained model weights, and adjusting the weights of the first convolutional layer for adaptation.

[0013] Through this technical solution, this application makes targeted improvements to the classic network architecture, making it particularly suitable for processing thrombosis features in lower extremity venous data, and significantly improving the ability and efficiency of feature extraction.

[0014] Furthermore, the multi-head self-attention module with mask guidance mechanism includes: setting multiple attention heads, dividing the attention heads into different groups according to the resolution of the lower extremity vein feature data, and capturing feature information at different scales respectively; configuring network layers, adaptive pooling layers, and fixed-length feature sequences to suppress overfitting and improve the model's generalization ability; using the normalized thrombus region segmentation mask as prior information, and forcing the model to focus on the thrombus region in the lower extremity vein data through the attention weight correction mechanism to reduce background interference; performing convolutional dimensionality reduction on the high-dimensional features output by the backbone network, converting them into a one-dimensional sequence through spatial compression, calculating feature correlation in parallel by multiple attention heads, concatenating the outputs of each attention head and mapping them to the classification dimension to complete accurate feature association.

[0015] Through this technical solution, this application introduces a mask-guided attention mechanism, enabling the model to actively focus on thrombus-related areas, effectively reducing background interference and improving the accuracy of thrombus feature analysis.

[0016] Furthermore, the training strategy of the deep learning classification model includes: employing an improved focusing loss function to assign higher weights to positive samples whose prediction results are close to the classification threshold, thereby enhancing the model's learning of thrombosis samples in lower limb vein data; selecting an adaptive optimizer with weight decay to suppress model overfitting; adopting a dynamic learning rate scheduling strategy based on validation set loss, automatically reducing the learning rate when the validation set loss does not decrease continuously, thereby improving training stability; employing a gradient pruning strategy to avoid gradient explosion during training; and using adaptive one-dimensional average pooling to uniformly represent feature sequences, ensuring consistent input of sequences of different lengths.

[0017] Through this technical solution, this application comprehensively utilizes a variety of advanced training strategies to ensure the stability, convergence speed, and generalization performance of the model training, thereby obtaining a high-performance thrombosis feature analysis model.

[0018] Furthermore, the step of generating the visualization attention distribution map includes: obtaining the lower limb vein feature data and corresponding category gradients output by the last convolutional layer of the deep learning classification model; calculating the weights of each feature data channel, where the weight values ​​are positively correlated with the contribution of the channel to the classification results, highlighting the role of thrombosis-related feature channels; weighting and summing the channel weights with the corresponding feature data, processing them with an activation function, and then upsampling them to the same size as the data to be analyzed; and overlaying the processed feature data with the data to be analyzed to generate an attention distribution map. The attention distribution map clearly presents the key attention areas for model decision-making, and simultaneously labels the bounding box, area, and diameter parameters of thrombi, providing an intuitive basis for clinical validation.

[0019] This technical solution greatly enhances the interpretability of model analysis results by generating intuitive and visual attention distribution maps and accurate thrombosis parameters, providing medical personnel with a reliable reference for thrombosis feature analysis.

[0020] Secondly, this application also discloses a thrombosis feature analysis system based on deep learning of lower extremity vein data, comprising: a data preprocessing module, which receives digital files of isolated lower extremity vein medical images stored on a computer storage medium, performs key layer data extraction, feature optimization of the data to be analyzed and the mask, and dual-channel tensor construction, and outputs the preprocessed two-dimensional input; a data balancing module, which performs data balancing processing on the two-dimensional input, constructs a sample subset and divides it into a training set, a validation set and a test set; a model training and inference module, which incorporates a deep learning classification model, the deep learning classification model using an improved EfficientNet-B3 with added convolutional layers for subtle feature extraction combined with multi-head self-attention with a mask guidance mechanism as its core architecture, the model training and inference module receiving the training set, validation set and test set for model training and validation, receiving the dual-channel tensor for thrombosis feature analysis and inference, and outputting a classification identifier representing thrombosis attributes; and a result output module, which generates a visual attention distribution map of the thrombosis region based on the classification identifier, and simultaneously outputs the quantitative parameters of the thrombosis region.

[0021] Through this technical solution, this application achieves the systematic deployment of thrombosis feature analysis methods through modular system design, providing an efficient and reliable solution for clinical applications.

[0022] Thirdly, this application also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the thrombosis feature analysis method based on deep learning of lower limb vein data as described above.

[0023] This invention offers the following advantages: By introducing a thrombus region segmentation mask as prior information, the model can identify suspected thrombus regions in advance. A dual-channel tensor is constructed, and combined with an improved EfficientNet-B3 backbone network and a mask-guided multi-head self-attention mechanism, it can more accurately identify thrombus features from complex lower extremity venous data, effectively reducing background interference and misjudgments, and significantly improving the ability to identify and analyze thrombus features in lower extremity venous data. Simultaneously, a visual attention distribution map is generated using gradient-weighted class activation mapping, clearly showing the region on which the model makes its judgments. Furthermore, outputting quantitative parameters such as thrombus size and location makes the analysis results objective and verifiable, greatly increasing doctors' trust in the intelligent analysis results, enhancing the model's interpretability and clinical applicability, and effectively solving the problems of low identification sensitivity and unreliable results in existing technologies. Attached Figure Description

[0024] Figure 1This is a flowchart illustrating the thrombosis feature analysis method based on deep learning of lower extremity venous data proposed in this application.

[0025] Figure 2 This is a schematic diagram of the system structure of the thrombosis feature analysis method based on deep learning of lower extremity vein data in this application.

[0026] Figure 3 This is a schematic diagram comparing the training loss of different models (the model in this application and the comparative model).

[0027] Figure 4 This is an example illustration of the attention distribution map of the thrombus region generated using Gradient Weighted Class Activation Mapping (Grad-CAM).

[0028] In the diagram: 1. Data preprocessing module; 2. Data balancing module; 3. Model training and inference module; 4. Result output module. Detailed Implementation

[0029] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0030] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0031] Reference Figure 1 This application proposes a thrombosis feature analysis method based on deep learning of lower extremity venous data, including:

[0032] S100: Obtain a lower limb vein dataset, which includes raw lower limb vein data and a thrombus region segmentation mask registered with it; the lower limb vein dataset is an ex vivo digital file stored on a computer storage medium.

[0033] S200: Perform preprocessing on the lower limb vein dataset, extract key layer data covering the thrombus region from the original lower limb vein data as data to be analyzed, perform size normalization and feature optimization on the data to be analyzed and the thrombus region segmentation mask respectively, and concatenate the processed data to be analyzed and the thrombus region segmentation mask in the channel dimension to form a dual-channel tensor, which is the preprocessed two-dimensional input;

[0034] S300: Perform data balancing processing on the preprocessed two-dimensional input to construct a sample subset with the same ratio of negative samples to positive samples. Divide the sample subset into a training set, a validation set, and a test set according to a preset ratio. Use the divided training set, validation set, and test set for training, validation, and testing of the deep learning classification model.

[0035] S400: The dual-channel tensor is input into the deep learning classification model, which uses EfficientNet-B3 with added convolutional layers for subtle feature extraction as the backbone network. In the classification stage, a multi-head self-attention module with a mask-guided mechanism based on thrombus region segmentation mask is introduced. Thrombus features in the lower limb vein data are extracted through the backbone network, and the self-attention module is used to focus on the thrombus region, outputting a classification identifier to characterize the thrombus attributes.

[0036] S500: Based on the classification identifier, generate a visual attention distribution map of the thrombus region that the model is interested in, and simultaneously output the quantitative parameters of the thrombus region, including the bounding box and area.

[0037] The lower extremity vein dataset originates from digital signal sequences obtained by scanning the lower extremity vein regions of patients. These sequences are digitized and sampled to form a two-dimensional data matrix containing spatial structural information. The thrombus region segmentation mask refers to the binarized labeled data of the thrombus region obtained through data segmentation algorithms or expert annotation. The key layer data refers to the representative two-dimensional data matrix covering the thrombus region in the original lower extremity vein data, typically the intermediate layer data matrix containing the thrombus and its adjacent data matrices. The dual-channel tensor is a multi-dimensional array formed by concatenating the processed data to be analyzed and the thrombus region segmentation mask along the channel dimension, used as input to the deep learning model. The improved EfficientNet-B3 backbone network is a convolutional neural network with structural adjustments and functional enhancements based on the original EfficientNet-B3 network. The mask-guided multi-head self-attention module refers to a multi-head self-attention mechanism that uses the thrombus region segmentation mask as prior information to guide the calculation of attention weights. The classification identifier representing the intrinsic properties of thrombi refers to a set of quantifiable and standardized symbolic identifiers output by a deep learning classification model after extracting and analyzing the thrombus features in lower extremity venous data. Its core function is to objectively describe the key inherent properties of thrombi, rather than directly providing a disease diagnosis. It includes morphological and pathological joint identifiers and risk association identifiers, with each dimension's identifier generated based on the model's feature parsing of a two-channel tensor. The visualized attention distribution map refers to the distribution data generated through gradient-weighted class activation mapping to represent the model's attention intensity, used to display the key areas of focus for the model.

[0038] This method is primarily applied to the automated analysis of lower extremity venous thrombosis characteristics. In this scenario, it is necessary to accurately identify and quantify thrombosis features from lower extremity venous data. Lower extremity venous data is characterized by high noise and low contrast, and thrombosis features, especially thrombosis features, are often not obvious, making them difficult to accurately identify using traditional methods.

[0039] In the process of thrombosis feature analysis based on lower extremity vein data, the first step is to obtain a lower extremity vein dataset. This can be achieved by scanning the patient's lower extremity veins with equipment to obtain key cross-sections. Thrombosis region segmentation masks can be generated using data segmentation algorithms or manual annotation by radiologists. The binary classification label for pulmonary thromboembolism can be determined based on the patient's clinical diagnosis, such as by using CT pulmonary angiography (CTPA) results to determine whether pulmonary thromboembolism has occurred.

[0040] Secondly, preprocessing is performed on the acquired lower extremity vein dataset. This step includes extracting key cross-sectional slices, typically selecting representative slices containing the core region of the thrombus. Then, the data to be analyzed and the thrombus region segmentation mask are scaled to a uniform size of 224×224 using bilinear interpolation. Bilinear interpolation preserves data details by weighted averaging of surrounding pixel values. The thrombus region segmentation mask is scaled to a size of 224×224 using nearest neighbor interpolation to ensure that the spatial location of the mask corresponds completely to the data (e.g., the "1" value at coordinate (100, 100) in the mask still corresponds to the thrombus region at the same location in the data to be analyzed). Next, feature optimization is performed. Gray-level normalization is performed on the data to be analyzed, and the pixel value x is mapped to the [0, 1] interval using formula (1) to enhance data contrast and reduce interference from illumination differences.

[0041] (1)

[0042] in, x min The minimum pixel value of the data to be analyzed. x max The maximum pixel value of the data to be analyzed. x norm The values ​​are normalized pixel values. Morphological optimization (using a 3×3 convolution kernel for "erosion-dilation" to remove edge noise and then dilate to restore the complete outline of the thrombus region) and binarization (automatically determining the threshold using the Otsu thresholding method to convert mask pixel values ​​to 0 or 1, clearly defining the boundary between thrombus and non-thrombus regions) are performed on the thrombus region segmentation mask. Finally, the processed data to be analyzed and the thrombus region segmentation mask are concatenated along the channel dimension to form a 224×224×2 dual-channel tensor as model input. Channel 1 is the optimized data to be analyzed, and channel 2 is the optimized thrombus region segmentation mask.

[0043] Next, data balancing is performed on the preprocessed two-dimensional input. Since positive samples (samples with significant thrombotic features or meeting specific clinical classification criteria) are often far fewer than negative samples in medical data, data balancing is necessary to prevent the model from being biased towards the majority class. Assuming the original dataset contains 100 positive samples and 800 negative samples (i.e., samples without significant thrombotic features or not meeting specific clinical classification criteria), first, negative samples are matched according to the number of positive samples. Then, 100 negative samples are randomly selected and paired with 100 positive samples to construct 10 sample subsets (each containing 10 positive and 10 negative samples). Data augmentation is performed on the samples within each subset. Specific augmentation methods include: random horizontal flipping (flipping probability set to 0.5, simulating data mirroring caused by horizontal probe movement during examination), rotation within an angular range (rotation angle range of -10° to +10°, step size of 1°, adapting to data angle differences under different examination positions), and Gaussian noise addition (noise mean μ=0, standard deviation σ=0.01, improving the model's robustness to inherent data noise). Finally, all enhanced sample subsets are merged and divided into training, validation, and test sets according to a preset ratio of 7:1.5:1.5. This ensures that the ratio of negative to positive samples in each dataset remains 1:1, avoiding the "biased towards negative samples" problem caused by sample imbalance during model training.

[0044] Then, the dual-channel tensor is input into the deep learning classification model for training and prediction. This classification model uses an improved EfficientNet-B3 as its backbone network, and the improvements specifically include:

[0045] 1) Adjust the number of input channels of the first convolutional layer from the default 3 channels adapted to RGB data to 2 channels, so that it can adapt to a 224×224×2 dual-channel tensor input;

[0046] 2) Add a 1×1 convolutional layer after the penultimate convolutional block of the backbone network (i.e., the MBConv6 block of EfficientNet-B3). The number of convolutional kernels can be set to 128 to enhance the extraction of subtle features of thrombi, such as the low-contrast texture of thrombus edges.

[0047] 3) In the classification stage, a multi-head self-attention module with a mask guidance mechanism is introduced. The module is specifically configured as follows: 8 attention heads are set and divided into 2 groups according to the feature data resolution. For example, 4 attention heads process high-resolution feature data of 256×256 to capture the details of the thrombus edge; 4 attention heads process low-resolution feature data of 64×64 to capture the overall morphology of the thrombus; a Dropout layer (Dropout probability = 0.3) is configured to suppress overfitting, and an adaptive pooling layer with a fixed output size converts feature data of different scales into feature sequences of fixed length; the thrombus region segmentation mask is normalized by formula (1) and used as prior information. The model is forced to focus on the thrombus region by the attention weight correction formula (2).

[0048] (2)

[0049] in, Attention raw The original attention weight matrix calculated for the model. M norm Here is the normalized thrombus region segmentation mask matrix, and ⊗ represents element-wise multiplication. Attention final This represents the corrected final attention weights.

[0050] The high-dimensional features output by the backbone network are reduced by 1×1 convolution, converted into a one-dimensional sequence by global average pooling, and then the feature correlation is calculated in parallel by 8 attention heads. The outputs of each attention head are concatenated and mapped to the classification dimension representing thrombosis attributes through a linear layer to complete accurate feature association and thrombosis attribute classification.

[0051] The model training employs an improved focusing loss function, optimized for thrombus samples (difficult-to-distinguish samples), as shown in the following formula:

[0052] (3)

[0053] in, p t For the model to class t The predicted probability, Class weights (positive samples) =2.0, negative samples Error! No bookmark name given.=1.0), Error! No bookmark name given. is the difficulty adjustment coefficient. The Error! No bookmark name given. term assigns higher weights to positive samples whose predicted probability is close to the classification threshold, strengthening the model's learning of thrombosis samples. At the same time, the AdamW adaptive optimizer with weight decay, dynamic learning rate scheduling (ReduceLROnPlateau strategy, patience=5, that is, if the validation set loss does not decrease for 5 consecutive rounds, the learning rate decays by factor=0.5) and gradient pruning (pruning threshold=1.0, to avoid gradient explosion during training) are used to ensure training stability.

[0054] Finally, a visual attention distribution map of the thrombus region is generated based on the classification identifier, and the quantitative parameters of the thrombus region are output simultaneously. Specific steps include:

[0055] 1) Obtain the lower limb vein feature data and corresponding class gradients output by the last convolutional layer of the model (the output layer of the improved EfficientNet-B3, with a dimension of 7×7×1536) (calculated through backpropagation).

[0056] 2) Calculate the weights of each feature data channel, and perform global average pooling on the gradient using formula (4). The weight values ​​are positively correlated with the contribution of the channel to the classification result:

[0057] (4)

[0058] in, For the first k The weight of each channel, For feature data size, For the first k One channel in ( i， j The gradient value at the position.

[0059] 3) The channel weights are weighted and summed with the corresponding feature data (5). After processing with the ReLU activation function and filtering out negative contribution features, bilinear interpolation is used to upsample the data to the same size as the data to be analyzed.

[0060] (5)

[0061] in, For the feature data of the attention distribution map, For the first k Feature data for each channel.

[0062] 4) The processed feature data is overlaid with the data to be analyzed to generate a attention distribution map, and the bounding box, area and diameter parameters of the thrombus are simultaneously labeled to provide intuitive evidence for clinical validation.

[0063] The proposed thrombosis feature analysis method based on deep learning of lower extremity vein data works by using a dual-channel input and mask-guided mechanism to enable the deep learning model to effectively focus on the thrombosis region and accurately extract thrombosis features, thus achieving precise analysis of thrombosis features. Specifically, the method first acquires a lower extremity vein dataset, providing a comprehensive data foundation for subsequent analysis. This data includes the data to be analyzed, thrombosis region segmentation masks, and classification labels, providing the necessary conditions for supervised learning. Subsequently, preprocessing is performed on the lower extremity vein dataset to extract key layer data and construct a dual-channel tensor. This step ensures the quality and consistency of the input data, providing optimized input for model training. Next, data balancing is performed on the preprocessed two-dimensional input. Data augmentation and sample resampling address the class imbalance problem in medical data, improving the model's generalization ability. Then, the dual-channel tensor is input into a deep learning classification model. Through an improved network architecture and a mask-guided attention mechanism, the model can accurately extract thrombosis features and reduce background interference. Finally, the "output of pulmonary thromboembolism prediction results and generation of a visual attention distribution map" not only provides prediction results, but also enhances the interpretability of the results through attention distribution maps and thrombosis parameters, providing clinicians with a reliable diagnostic reference.

[0064] Through the above technical solution, this application can effectively solve the problem of difficulty in identifying thrombi in lower extremity venous data, improve the accuracy and reliability of thrombus feature analysis, and provide strong technical support for clinical evaluation.

[0065] In another embodiment of this application, S200 specifically includes:

[0066] S210: Extract the intermediate slices and adjacent slices of the three-dimensional volume data of lower limb veins along the Z-axis as key layer data;

[0067] S220: The data to be analyzed is scaled to a uniform size by bilinear interpolation, and the thrombus region segmentation mask is scaled to the same size by nearest neighbor interpolation;

[0068] S230: Perform grayscale normalization on the data to be analyzed, and perform morphological optimization and binarization on the thrombus region segmentation mask;

[0069] S240: The optimized data to be analyzed is concatenated with the thrombus region segmentation mask in the channel dimension to form a dual-channel tensor with matched dimensions in each dimension.

[0070] Specifically, when extracting key layer data, a representative section is selected, which includes a complete display of the long diameter of the vessel containing thrombus.

[0071] In the size normalization process, the data to be analyzed is scaled using bilinear interpolation to maintain its smoothness and detail. The thrombus region segmentation mask is scaled using nearest-neighbor interpolation to preserve its boundary clarity and binary properties. In the feature optimization stage, grayscale normalization of the data to be analyzed eliminates the influence of illumination differences and enhances data contrast. Morphological optimization of the thrombus region segmentation mask eliminates noise and small holes, while binarization ensures a clear mask definition. The resulting dual-channel tensor simultaneously contains information from the data to be analyzed and information from the thrombus region segmentation mask, providing rich input features for subsequent deep learning models.

[0072] In another embodiment of this application, S300 further includes:

[0073] S310: Match the corresponding negative samples from the raw lower extremity vein data according to the number of positive samples, and randomly pair them to construct multiple sample subsets;

[0074] S320: Perform data augmentation on samples within each subset of samples, including random horizontal flipping, rotation within an angular range, and Gaussian noise addition;

[0075] S330: Merge all enhanced sample subsets and divide them into training, validation, and test sets according to a preset ratio to ensure that the ratio of negative to positive samples is consistent in each dataset.

[0076] When matching negative samples with positive samples from the original data, a balanced subset can be constructed by randomly selecting a number of negative samples that are roughly equal to the number of positive samples. For example, if the original data contains 50 positive samples and 500 negative samples, 50 negative samples can be randomly selected and paired with 50 positive samples to construct 5 sample subsets (each containing 10 positive and 10 negative samples). This avoids insufficient diversity in the augmented data caused by an excessively small sample size in a single subset.

[0077] Randomly pairing samples to construct multiple subsets increases data diversity and improves model robustness. Data augmentation operations include random horizontal flipping to create mirrored versions of the data, rotation within a certain angle range to simulate data from different acquisition angles, and adding Gaussian noise to enhance the model's robustness to noise. These augmentations increase data diversity without altering its semantic information. Finally, all augmented subsets are merged and divided into training, validation, and test sets according to a predefined ratio to ensure a consistent proportion of negative and positive samples in each dataset, avoiding bias during training. These techniques significantly improve the model's learning ability on minority class samples (positive samples), enhance its generalization performance, and prevent overfitting.

[0078] In another embodiment of this application, the construction of the improved EfficientNet-B3 backbone network includes:

[0079] Adjust the number of input channels in the first convolutional layer to adapt it to dual-channel tensor input;

[0080] An additional convolutional layer is added after the penultimate convolutional block to enhance thrombus feature extraction;

[0081] The network is initialized using the weights of the pre-trained model, and the weights of the first convolutional layer are adjusted for adaptability.

[0082] Specifically, when adjusting the number of input channels in the first convolutional layer, the original 3-channel input is changed to 2-channel input to adapt to the two-channel tensor. For example, the original EfficientNet-B3's first convolutional layer is Conv2d(3, 40, kernel_size=3, stride=2, padding=1), and after adjustment, it becomes Conv2d(2, 40, kernel_size=3, stride=2, padding=1), ensuring that it matches the input dimension of the two-channel tensor.

[0083] The added convolutional layers typically include operations such as convolution, normalization, and activation functions to further extract and enhance thrombus features. For example, the added 1×1 convolutional layer is configured as Conv2d(1536, 128, kernel_size=1, stride=1, padding=0), followed by BatchNorm2d(128) and the Swish activation function to further extract and enhance thrombus features.

[0084] The use of pre-trained weights can accelerate model convergence and improve training efficiency. Adaptive adjustments to the weights of the first convolutional layer are typically made using random initialization or interpolation based on the existing weights to ensure the network can effectively handle dual-channel input.

[0085] Through the above technical solution, this application can significantly improve the network's ability to extract thrombus features from lower extremity venous data, laying the foundation for accurate analysis of thrombus features.

[0086] In another embodiment of this application, the multi-head self-attention module with masked guidance mechanism further includes:

[0087] Multiple attention heads are set up and divided into different groups according to the feature data resolution to capture feature information at different scales. Network layers to suppress overfitting, adaptive pooling layers, and fixed-length feature sequences are configured to improve the model's generalization ability. The thrombus region segmentation mask is normalized and used as prior information. Through the attention weight correction mechanism, the model is forced to focus on the thrombus region. The high-dimensional features output by the backbone network are subjected to convolutional dimensionality reduction, which is converted into a one-dimensional sequence through spatial compression. After parallel computation by multiple attention heads, the sequence is concatenated and output and mapped to the classification dimension.

[0088] Specifically, the attention head setup typically includes 4-8 attention heads, grouped according to the resolution of the feature data, to process feature information at different scales. For example, with 8 attention heads, 4 handle high-resolution feature data (e.g., 256×256), focusing on capturing the subtle textures of the thrombus edges; and 4 handle low-resolution feature data (e.g., 64×64), focusing on capturing the overall shape and location of the thrombus.

[0089] Measures to suppress overfitting include adding Dropout layers, LayerNorm layers, etc., while adaptive pooling layers are used to convert feature data of different sizes into fixed-length feature sequences. For example, the output size of an adaptive average pooling layer is fixed at 1×1024, ensuring that the output sequence length is 1024 regardless of the size of the input feature data.

[0090] Normalization of the thrombus region segmentation mask typically employs min-max normalization, converting it into values ​​between 0 and 1 as prior weights. The attention weight correction mechanism strengthens the attention weights for the thrombus region by multiplying the mask weights by the calculated attention weights. Convolutional dimensionality reduction uses 1x1 convolutions to reduce the number of channels, and spatial compression transforms the two-dimensional feature data into a one-dimensional sequence through a reshaping operation. After parallel computation of multiple attention heads, the outputs of each head are concatenated and mapped to the classification dimension through a linear layer.

[0091] Through the above technical solutions, this application can significantly improve the model's focus on the thrombus region, reduce background interference, and improve the accuracy of feature representation, thereby improving the accuracy of thrombus feature analysis.

[0092] In another embodiment of this application, the training strategy for the deep learning classification model is further proposed to include:

[0093] An improved focusing loss function is used to assign higher weights to positive samples that are close to the classification threshold;

[0094] An adaptive optimizer with weight decay is selected to suppress model overfitting;

[0095] A dynamic learning rate scheduling strategy based on validation set loss is adopted, which automatically reduces the learning rate when the validation set loss does not decrease continuously.

[0096] A gradient clipping strategy is adopted to avoid gradient explosion during training;

[0097] Feature sequences are uniformly represented by adaptive one-dimensional average pooling.

[0098] Specifically, the improved focusing loss function, based on the standard cross-entropy loss, adds higher weight coefficients to positive samples near the decision boundary, strengthening the model's learning of these difficult samples. Adaptive optimizers with weight decay (such as AdamW) suppress overfitting and improve model generalization ability through weight decay terms. Dynamic learning rate scheduling strategies monitor changes in validation set loss; when the loss no longer decreases for several consecutive epochs, the learning rate is reduced proportionally to improve training stability. Gradient pruning strategies avoid gradient explosion during training by setting gradient thresholds. Adaptive one-dimensional average pooling converts feature sequences of different lengths into fixed-length representations, ensuring consistency in subsequent processing.

[0099] Through the above technical solution, this application can achieve stable and efficient model training and obtain a high-performance thrombosis feature analysis model.

[0100] In another embodiment of this application, the step of generating the visualized attention distribution map includes:

[0101] Obtain the feature data and corresponding class gradients output from the last convolutional layer of the model;

[0102] Calculate the weight of each feature data channel. The weight is positively correlated with the channel's contribution to the classification result.

[0103] The channel weights are weighted and summed with the corresponding feature data, and then upsampled to the size of the data to be analyzed after processing by the activation function.

[0104] The processed feature data is overlaid with the data to be analyzed to generate an attention distribution map, and the bounding box, area and diameter parameters of the thrombus are labeled simultaneously.

[0105] Specifically, feature data and category gradients are obtained through backpropagation, reflecting the contribution of each feature data channel to the final classification decision. Channel weights are typically calculated using global average pooling to aggregate gradients, yielding the importance weight for each channel. A weighted summation operation multiplies each feature data point by its corresponding weight and sums the results to obtain a preliminary attention distribution map. Activation function processing typically uses the ReLU function to preserve features that positively contribute to classification. An upsampling operation uses bilinear interpolation to restore the attention distribution map size to match the data to be analyzed. The final attention distribution map is overlaid on the data to be analyzed, using color intensity to represent the importance of regions. Thrombosis parameter annotation automatically calculates parameters such as bounding boxes, area, and diameter by analyzing the geometric characteristics of the highlighted areas in the attention distribution map.

[0106] Through the above technical solutions, this application not only provides accurate prediction results, but also enhances the interpretability of the results through visualization and parameter output, providing reliable analytical references for medical personnel.

[0107] The following describes the method described in this application, using a real-world clinical application scenario: Data from one patient was obtained in a lower extremity vein screening project at a hospital. In the preprocessing stage, representative data of the long diameter of thrombus-containing vessels were selected. The data to be analyzed was 512×512 pixels, normalized to 224×224 pixels; the thrombus region segmentation mask was also scaled to 224×224 pixels. Gray-level normalization was performed on the data to be analyzed, mapping pixel values ​​to the 0-1 range; morphological closure was performed on the thrombus region segmentation mask to eliminate small holes, and binarization was performed. Finally, the processed data and the mask were concatenated along the channel dimension to form a 224×224×2 two-channel tensor.

[0108] In the data balancing process, it is assumed that the original data contains 100 positive samples and 1000 negative samples. 100 negative samples are randomly paired with 100 positive samples to construct 10 sample subsets. Data augmentation is performed on each subset, including random horizontal flipping (probability 0.5), rotation within an angular range (-10° to +10°), and Gaussian noise addition (μ=0, σ=0.01). All augmented samples are merged, resulting in 2000 samples (1000 positive + 1000 negative), which are then divided into a training set (1400 samples), a validation set (300 samples), and a test set (300 samples) in a 7:1.5:1.5 ratio.

[0109] The model was trained using an improved EfficientNet-B3 backbone network, with the number of input channels in the first convolutional layer reduced to 2. A convolutional layer with 128 filters was added after the penultimate convolutional block. The network was initialized using weights pre-trained on ImageNet, and the weights in the first layer were adaptively adjusted. An improved focusing loss function was used during training, assigning a 3x weight to positive samples with predicted probabilities between 0.4 and 0.6. The AdamW optimizer (weight decay of 0.01) was used with an initial learning rate of 0.001, which was halved when the validation set loss did not decrease for 5 consecutive epochs. The gradient clipping threshold was 1.0. After 100 epochs of training, the model with the best performance on the validation set was selected.

[0110] During the inference phase, the patient's two-channel tensor is input, and the model outputs a classification identifier representing the thrombus attributes. A focus distribution map is generated using Grad-CAM, showing that the model primarily focuses on a hypoechoic region within the femoral vein. The bounding box of this region is automatically calculated to be [120:150, 80:110], with an area of ​​approximately 900 pixels and a diameter of approximately 34 pixels.

[0111] To comprehensively evaluate the effectiveness and advancement of the proposed method, a systematic experiment was conducted based on a clinically collected lower extremity vein dataset. The dataset contained several three-dimensional samples, all of which underwent ethical review and anonymization, and were divided into training, validation, and test sets in a 7:1.5:1.5 ratio. Multiple metrics were used for comprehensive evaluation, including accuracy, precision, recall, F1 score, specificity, AUCROC, and AUPRC, with particular emphasis on recall and AUCROC to demonstrate the sensitive identification capability for thrombi.

[0112] The experiment first compared the performance data of the method described in this application (EfficientNet-B3+MHSA) with that of various mainstream CNN models (including ResNet-50, VGG-16, and DenseNet-121). As shown in Table 1, EfficientNet-B3 itself has shown superior performance (accuracy 80.82% and recall 87.50%), verifying its potential as a backbone network.

[0113] Table 1: Comparison of thrombosis feature analysis classification performance based on different deep learning models.

[0114]

[0115] Furthermore, to verify the effectiveness of the multi-head self-attention mechanism (MHSA), an ablation experiment was conducted by introducing an MHSA module on top of EfficientNet-B3 and DenseNet-121. Table 2 shows that EfficientNet-B3+MHSA further improved across all metrics, achieving the best overall performance (accuracy 83.06%, AUCROC 88.55%), and exhibiting a more balanced recall and specificity, demonstrating its ability to effectively suppress background interference while focusing on the thrombus area.

[0116] Table 2: Performance comparison of baseline model and model incorporating MHSA mechanism in thrombosis characterization task

[0117]

[0118] Please refer to Figure 3 Training process analysis shows that the model converges more stably and the validation loss is significantly reduced after introducing MHSA, indicating stronger generalization ability. Additionally, please refer to... Figure 4 The attention distribution map generated by Grad-CAM shows the original data on the left and the thrombus localization results with the attention distribution map superimposed on the right. It intuitively shows that the model's attention area is highly consistent with the thrombus location marked by the physician, especially in low-contrast thrombi, which can still be accurately located, significantly improving the interpretability and clinical credibility of the results.

[0119] Experimental results show that the proposed model (EfficientNet-B3+MHSA) outperforms the comparative model in multiple metrics (see Tables 1 and 2). The introduction of multi-head self-attention (MHSA) further enhances the balance between recall and specificity (see Table 2), demonstrating its ability to effectively focus on the thrombus region and suppress background interference. The training process is also more stable (see...). Figure 2 Furthermore, the attention distribution map generated by Grad-CAM shows that the model's attention area highly matches the physician's annotations (see...). Figure 3 , 4 This significantly improves the interpretability of the results and has good clinical applicability.

[0120] Please refer to Figure 2 The specific implementation of this application also discloses a thrombosis feature analysis system based on deep learning of lower extremity venous data, including:

[0121] Data preprocessing module 1 is used to receive digital medical images of ex vivo lower limb veins stored in a computer storage medium, perform key layer data extraction, feature optimization of the data to be analyzed and the mask, and dual-channel tensor construction, and output the preprocessed two-dimensional input.

[0122] Data balancing module 2 is used to perform data balancing processing on two-dimensional input, construct sample subsets and divide them into training set, validation set and test set;

[0123] Model training and inference module 3 has a built-in deep learning classification model with an improved EfficientNet-B3 combined with mask-guided multi-head self-attention as its core architecture. It receives training, validation and test sets for model training and validation, receives dual-channel tensors for thrombosis feature analysis and inference, and outputs classification identifiers that characterize thrombosis attributes.

[0124] Result output module 4 is used to generate a visual attention distribution map of the thrombus region based on the classification identifier, and simultaneously output the quantitative parameters of the thrombus region. This system aims to achieve intelligent analysis of thrombus characteristics through a modular approach, effectively solving the problems of difficult thrombus identification, model attention region bias, and uninterpretable results in existing technologies. The data preprocessing module ensures input data quality, the data balancing module solves class imbalance problems, the model training and inference module achieves accurate prediction, and the result output module provides interpretable result displays. This systematic design ensures the accuracy, reliability, and interpretability of the prediction process, providing strong support for clinical diagnosis.

[0125] The data preprocessing module 1 is used to receive and process the lower limb vein 3D dataset. This module typically includes functional units such as data reading, key layer data selection, size normalization, feature optimization, and dual-channel construction. In actual deployment, this module can integrate data processing libraries (such as OpenCV) and scientific computing libraries (such as NumPy) to implement related functions. The module's input is a DICOM format data file and its corresponding annotation file, and its output is the processed dual-channel tensor.

[0126] The data balancing module 2 addresses class imbalance issues. This module includes functional units for sample pairing, data augmentation, and dataset partitioning. The sample pairing unit randomly selects matching samples from the negative samples based on the number of positive samples; the data augmentation unit implements various augmentation transformations; and the dataset partitioning unit divides the data into training, validation, and test sets according to a preset ratio. This module is typically implemented using Python data processing libraries (such as Pandas) and other data processing libraries.

[0127] Model Training and Inference Module 3 is the core of the system, containing the training and inference functions for the deep learning model. This module uses a deep learning framework (PyTorch or TensorFlow) to implement an improved EfficientNet-B3 backbone network and a multi-head self-attention mechanism. Training functions include loss calculation, backpropagation, and optimizer updates; inference functions include forward propagation and classification identifier output. This module requires GPU acceleration to ensure training and inference efficiency.

[0128] Module 4, the Results Output module, generates the final analysis results and visualizations. This module includes functional units such as classification identifier parsing, attention distribution map generation, and parameter calculation. The classification identifier parsing unit converts the model output into a thrombus attribute representation; the attention distribution map generation unit generates a visual attention distribution map using the Grad-CAM algorithm; and the parameter calculation unit automatically calculates quantitative parameters such as the bounding box, area, and diameter of the thrombus region. This module typically integrates visualization libraries (such as Matplotlib) and data processing libraries to implement these functions, and the output results can be displayed through a graphical interface or used to generate report files.

[0129] This application further proposes that the data preprocessing module 1 also includes a quality detection submodule, which is used to automatically identify and filter low-quality lower extremity vein data; the criteria for judging low-quality data include data signal-to-noise ratio lower than a preset threshold, ambiguity higher than a preset threshold, and thrombus region mask integrity lower than a preset threshold.

[0130] The quality inspection submodule is an automated screening unit integrated into the data preprocessing workflow. Specifically, it can be implemented using data quality assessment algorithms combined with preset threshold judgment rules. By calculating the signal-to-noise ratio (SNR), ambiguity index, and mask integrity parameters, samples that do not meet quality requirements are filtered out. A SNR below the preset threshold means the energy ratio of effective signal to noise in the data does not meet the analyzable standard. This can be achieved using frequency domain energy analysis methods to eliminate invalid images with excessive noise interference. Ambiguity above the preset threshold means the edge sharpness of the data does not meet the feature extraction requirements. This can be achieved by calculating the energy proportion of high-frequency components using the Laplacian gradient operator to eliminate motion artifacts or blurry data caused by device defocusing. Thrombosis region mask integrity below the preset threshold means the effective pixel proportion of the thrombus region in the segmentation result does not meet the standard of covering the actual lesion. This can be achieved by calculating the connected component area proportion method to filter out incomplete masks caused by segmentation algorithm failure.

[0131] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A thrombus feature analysis method based on deep learning of lower extremity vein data, characterized in that, The method comprises the following steps: obtaining a lower extremity vein data set, including original lower extremity vein data and a thrombus area segmentation mask registered therewith; performing preprocessing on the lower extremity vein data set, extracting key layer data covering the thrombus area in the original lower extremity vein data as to-be-analyzed data, performing size normalization and feature optimization on the to-be-analyzed data and the thrombus area segmentation mask respectively, splicing the processed to-be-analyzed data and the thrombus area segmentation mask in the channel dimension to form a dual-channel tensor, and the dual-channel tensor is a two-dimensional input after preprocessing; performing data balancing processing on the two-dimensional input, constructing a sample subset with consistent negative sample and positive sample proportions, dividing the sample subset into a training set, a validation set and a test set according to a preset proportion, and using the divided training set, validation set and test set for training, validation and testing of a deep learning classification model; inputting the dual-channel tensor into the deep learning classification model, wherein the deep learning classification model takes an EfficientNet-B3 with added subtle feature extraction convolutional layers as a backbone network, and introduces a multi-head self-attention module with a mask guided mechanism based on the thrombus area segmentation mask in the classification stage; the construction of the EfficientNet-B3 backbone network comprises: adjusting the input channel number of the first convolutional layer of the backbone network to adapt to the dual-channel tensor; adding a convolutional layer after the second-to-last convolutional block of the backbone network, wherein the added convolutional layer is used to strengthen thrombus feature extraction; and initializing the backbone network using pre-trained model weights and adaptively adjusting the weights of the first convolutional layer; the multi-head self-attention module with the mask guided mechanism comprises: setting multiple attention heads, dividing the attention heads into different groups according to the lower extremity vein feature data resolution, and capturing feature information of different scales respectively; configuring a network layer for suppressing overfitting, an adaptive pooling layer and a fixed-length feature sequence to improve the model generalization capability; normalizing the thrombus area segmentation mask as prior information, and through an attention weight correction mechanism, forcing the model to focus on the thrombus area in the lower extremity vein data and reducing the background interference degree; performing convolution dimension reduction on the high-dimensional features output by the backbone network, converting them into one-dimensional sequences through spatial compression, calculating the feature correlation through multiple attention heads in parallel, splicing the outputs of each attention head and mapping them to the classification dimension to complete the precise feature correlation; extracting thrombus features in the lower extremity vein data through the backbone network, focusing on the thrombus area through the self-attention module, and outputting a classification identifier representing the thrombus attributes; based on the classification identifier, generating a visual attention distribution map of the thrombus area focused on by the model, and synchronously outputting a quantitative parameter of the thrombus area, wherein the quantitative parameter comprises a bounding box and an area.

2. The method of claim 1, wherein, In the preprocessing step, the key layer data is the middle layer slice and its adjacent slices of the lower extremity vein data containing thrombus; the size normalization scales the to-be-analyzed data to a uniform size through bilinear interpolation and scales the thrombus area segmentation mask to the same size as the to-be-analyzed data through nearest neighbor interpolation; The feature optimization includes performing grayscale normalization on the data to be analyzed, performing morphological optimization and binarization processing on the thrombus region segmentation mask; The two channels of the double-channel tensor are the optimized data to be analyzed and the optimized thrombus region segmentation mask, respectively, and the dimension sizes are matched with each other.

3. The method of claim 1, wherein, The data balancing processing step includes: Match the corresponding negative samples with the positive samples from the lower extremity vein original data, and randomly pair to construct multiple sample subsets; Perform data enhancement on the lower extremity vein samples in each sample subset, and the enhancement methods include random horizontal flip, rotation within an angle range, and Gaussian noise addition; Merge all the enhanced sample subsets, and divide them into training set, validation set and test set according to the preset proportion, to ensure that the proportion of negative samples and positive samples in each data set is consistent.

4. The method of claim 1, wherein, The training strategy of the deep learning classification model includes: Using an improved focal loss function, the positive samples close to the classification threshold are given higher weights to strengthen the model's learning of thrombus features; Selecting a self-adaptive optimizer with weight decay to suppress model overfitting; Using a dynamic learning rate scheduling strategy based on the validation set loss, automatically reducing the learning rate when the validation set loss does not decrease continuously, to improve the training stability; Using a gradient clipping strategy to avoid gradient explosion during training; Through adaptive one-dimensional average pooling, the feature sequence is uniformly represented to ensure consistent input of different length sequences.

5. The method of claim 1, wherein, The visualization attention distribution map generation step includes: Obtain the lower extremity vein feature data and the corresponding class gradient output by the last convolution layer of the deep learning classification model; Calculate the weight of each feature data channel, and the weight value is positively correlated with the contribution of the channel to the classification result, highlighting the role of thrombus-related feature channels; Weighted sum of channel weight and corresponding feature data, and then upsampled to the same size as the data to be analyzed after activation function processing; Overlay the processed feature data and the data to be analyzed to generate an attention distribution map, which presents the key attention area of the model decision through the attention distribution map, and simultaneously labels the bounding box, area and diameter parameters of the thrombus, providing intuitive evidence for clinical verification.

6. A thrombus feature analysis system based on deep learning of lower extremity venous data, characterized by, It includes: A data preprocessing module is used to receive ex vivo lower extremity vein medical image digital files stored in a computer storage medium, perform key layer data extraction, feature optimization of data to be analyzed and masks, and construct a double-channel tensor, and output the preprocessed two-dimensional input; A data balancing module is used to perform data balancing processing on the two-dimensional input, construct sample subsets, and divide them into training set, validation set and test set; A model training and reasoning module is built-in with a deep learning classification model, which uses an improved EfficientNet-B3 with additional fine feature extraction convolution layers combined with a multi-head self-attention mechanism with mask guidance as the core architecture, and receives the training set, validation set and test set for model training and validation, and receives the double-channel tensor for thrombus feature analysis and reasoning, and outputs a classification identifier representing the thrombus attributes. The EfficientNet-B3 backbone network is constructed by adjusting the input channel number of the first convolutional layer of the backbone network to adapt to the double-channel tensor, adding a convolutional layer after the second-to-last convolutional block of the backbone network, and using the pre-trained model weight to initialize the backbone network and adaptively adjusting the weight of the first convolutional layer. The multi-head self-attention module with a mask guide mechanism includes: setting multiple attention heads, dividing the attention heads into different groups according to the resolution of the lower limb vein feature data, and capturing feature information of different scales respectively; configuring a network layer for suppressing overfitting, an adaptive pooling layer and a fixed-length feature sequence to improve the model generalization ability; normalizing the thrombus region segmentation mask as prior information, and through an attention weight correction mechanism, forcing the model to focus on the thrombus region in the lower limb vein data and reducing the background interference degree; performing convolution dimension reduction on the high-dimensional features output by the backbone network, converting them into one-dimensional sequences through spatial compression, calculating the feature correlation through multiple attention heads in parallel, splicing the outputs of each attention head and mapping them to the classification dimension to complete the precise feature correlation; The result output module is configured to generate a thrombus region visual attention distribution map based on the classification identifier and output the quantitative parameters of the thrombus region synchronously.

7. An electronic device, comprising: The computer program stored in the memory and executable on the processor implements the deep learning-based thrombus feature analysis method for lower limb vein data according to any one of claims 1-5 when the processor executes the computer program.

Images