Fetal doppler blood flow spectrum anomaly recognition method based on axial attention mechanism

Abstract

This application relates to a method, apparatus, computer device, and storage medium for identifying fetal Doppler blood flow spectrum anomalies based on an axial attention mechanism. It involves acquiring a fetal Doppler spectrum dataset from early pregnancy, preprocessing the fetal Doppler spectrum images to obtain preprocessed grayscale spectrograms, constructing a backbone network based on the axial attention mechanism, and a multi-task anomaly recognition head targeting three types of target vessels. The model is then jointly trained using the preprocessed grayscale spectrograms to obtain a trained model. The preprocessed fetal Doppler spectrum images from early pregnancy are fed into the trained model for forward inference, outputting the recognition result. By employing a dual-axis decoupled axial attention mechanism, the time axis and frequency axis are strictly decoupled and modeled independently, adaptively amplifying specific weak hemodynamic features in the Doppler blood flow spectrum and suppressing background noise, thereby improving feature extraction efficiency and robustness.

Description

Technical Field

[0001] This application relates to the field of ultrasound image detection technology, and in particular to a method for identifying abnormalities in fetal Doppler blood flow spectrum based on axial attention mechanism. Background Technology

[0002] Early pregnancy screening is currently the core of the prenatal diagnostic system. Authoritative institutions such as ISUOG and the China Maternal and Child Health Association have clearly included ductal venous Doppler, tricuspid valve Doppler, and umbilical artery Doppler in the core components of early pregnancy prenatal screening. Combined screening of these three abnormal indicators can provide early warning of life-threatening adverse pregnancy conditions such as severe congenital heart disease, early-onset preeclampsia, and fetal growth restriction. Traditional fetal Doppler blood flow spectrum analysis relies heavily on the doctor's clinical experience for manual feature extraction and morphological analysis, and is extremely sensitive to speckle noise and signal aliasing commonly found in ultrasound images, resulting in low feature extraction efficiency. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, device, computer equipment, and storage medium for identifying fetal Doppler blood flow spectrum abnormalities based on axial attention mechanism that can improve feature extraction efficiency, in order to address the above problems.

[0004] The first aspect of this application provides a method for identifying fetal Doppler blood flow spectrum abnormalities based on an axial attention mechanism, including: Obtain a fetal Doppler spectrum dataset from early pregnancy; the fetal Doppler spectrum dataset from early pregnancy includes fetal Doppler spectrum images from early pregnancy and annotated datasets. The fetal Doppler spectrum images from early pregnancy cover three types of target vessels: ductus venosus, tricuspid valve, and umbilical artery. The annotated dataset includes vessel type, abnormality type, and abnormal region coordinates. The fetal Doppler spectrum image in early pregnancy is preprocessed to obtain a preprocessed grayscale spectrum image; A backbone network with an axial attention mechanism and a multi-task anomaly recognition head for three types of target blood vessels are constructed. The model is then trained jointly for multiple tasks using preprocessed grayscale spectrograms to obtain a trained model. The backbone network is used to perform independent decoupling of the time axis and frequency axis. The trained model represents the correspondence between the fetal Doppler spectrum image in early pregnancy and the recognition results of the three types of target blood vessels. After preprocessing the Doppler spectrum image of the fetus in early pregnancy to be identified, it is fed into the trained model for forward inference and the identification result is output.

[0005] In one embodiment, the early pregnancy fetal Doppler spectrum image is preprocessed to obtain a preprocessed grayscale spectrum image, including: The fetal Doppler spectrum images from early pregnancy are cropped, scaled to a preset size, and normalized. Data augmentation is performed on the normalized image using random scaling, random Gaussian noise injection, and random contrast adjustment to simulate artifacts and signal fluctuations, resulting in a pre-processed grayscale spectrum.

[0006] In one embodiment, the backbone network constructing the axial attention mechanism and the multi-task anomaly recognition head targeting three types of target blood vessels are combined with preprocessed grayscale spectrograms for multi-task joint training to obtain a trained model, including: Based on the preprocessed grayscale spectrogram, a backbone network with an axial attention mechanism is constructed to decouple the time axis and frequency axis. The decoupling results are then fused to generate a high-level semantic feature map. Based on the high-level semantic feature map, a multi-task anomaly recognition head is constructed for three types of target blood vessels. The output is the probability of occurrence of each type of anomaly. Multi-task joint training is performed by combining a multi-task joint loss function and a cosine annealing learning rate decay strategy to achieve dynamic adjustment of the learning rate, thus obtaining the trained model.

[0007] In one embodiment, based on the preprocessed grayscale spectrogram, a backbone network with an axial attention mechanism is constructed to decouple the time and frequency axes. Feature fusion is then performed on the decoupling results to generate a high-level semantic feature map, including: Based on the preprocessed grayscale spectrum, shallow feature extraction is performed through convolutional layers to preserve the basic edge and texture features of the spectrum and generate a shallow feature map. Based on the shallow feature map, perform temporal axis attention calculation, frequency axis attention calculation, and local detail convolution to obtain the output results of the three branches; Feature fusion is performed based on the outputs of the three branches to generate a high-level semantic feature map.

[0008] In one embodiment, feature fusion is performed based on the outputs of the three branches to generate a high-level semantic feature map, including: The outputs of the three branches are fused, and the temporal attention, frequency attention and local detail features are added element by element. After normalizing the fused features, nonlinearity is introduced through the GELU activation function to obtain a high-level semantic feature map.

[0009] In one embodiment, the multi-task joint loss function is: ; in, The focus loss for the main classification branch, The cross-entropy loss is the branching loss of blood vessel types. The multi-label focus loss is used for the abnormal type branch, where λ1, λ2, and λ3 are weights.

[0010] In one embodiment, the model outputs the identification results including blood vessel type, abnormality type, and corresponding abnormality probability value.

[0011] A second aspect of this application provides a fetal Doppler blood flow spectrum abnormality identification device based on an axial attention mechanism, comprising: The data acquisition module is used to acquire a fetal Doppler spectrum dataset in early pregnancy. The fetal Doppler spectrum dataset in early pregnancy includes fetal Doppler spectrum images and labeled datasets. The fetal Doppler spectrum images in early pregnancy cover three types of target vessels: ductus venosus, tricuspid valve, and umbilical artery. The labeled dataset includes vessel type, abnormality type, and abnormal region coordinates. The data processing module is used to preprocess the fetal Doppler spectrum image in early pregnancy to obtain a preprocessed grayscale spectrum image. The model training module is used to construct the backbone network of the axial attention mechanism and the multi-task anomaly recognition head for three types of target blood vessels. The model is jointly trained by multi-task using the preprocessed grayscale spectrogram to obtain the trained model. The backbone network is used to perform independent decoupling of the time axis and frequency axis. The trained model represents the correspondence between the fetal Doppler spectrum image in early pregnancy and the recognition results of the three types of target blood vessels. The anomaly detection module is used to preprocess the Doppler spectrum image of the fetus in early pregnancy to be identified, and then feed it into the trained model for forward inference to output the recognition result.

[0012] A third aspect of this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0013] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0014] The aforementioned method, device, computer equipment, and storage medium for fetal Doppler blood flow spectrum anomaly recognition based on the axial attention mechanism acquire a fetal Doppler spectrum dataset from early pregnancy, preprocess the fetal Doppler spectrum images from early pregnancy to obtain preprocessed grayscale spectrograms, construct a backbone network based on the axial attention mechanism and a multi-task anomaly recognition head for three types of target vessels, and perform multi-task joint training of the model in conjunction with the preprocessed grayscale spectrograms to obtain a trained model. The preprocessed fetal Doppler spectrum images from early pregnancy to be identified are then fed into the trained model for forward inference, outputting the recognition results. By using a dual-axis decoupled axial attention mechanism, the time axis and frequency axis are strictly decoupled and modeled independently, adaptively amplifying specific weak hemodynamic features in the Doppler blood flow spectrum and suppressing background noise, thereby achieving efficient, robust, and high-precision end-to-end automated fine-grained classification, improving feature extraction efficiency and robustness. Attached Figure Description

[0015] Figure 1 This is a flowchart of a method for identifying abnormal fetal Doppler blood flow spectrum based on axial attention mechanism in one embodiment; Figure 2 This is a schematic diagram of the backbone network structure of the axial attention mechanism in one embodiment; Figure 3 This is a structural block diagram of a fetal Doppler blood flow spectrum abnormality identification device based on axial attention mechanism in one embodiment; Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0018] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0019] In clinical practice, Doppler flow spectrograms are crucial for assessing the hemodynamic status of the fetus and mother. Traditionally, sonographers manually locate the target vessel, acquire at least three consecutive stable cardiac cycles of spectral images, visually observe the spectral morphology, manually mark abnormal areas, and rely on experience to determine whether the abnormality is pathological. To automate spectral analysis, early research primarily employed a combination of traditional signal processing and machine learning. Such approaches typically begin by generating a time-frequency graph using short-time Fourier transform, then extract the maximum velocity envelope of blood flow using predefined mathematical thresholds, active contour models, or morphological operators, and finally calculate the statistical characteristics of the envelope and input them into a support vector machine or random forest for classification.

[0020] In recent years, with the development of deep learning technology, directly using convolutional neural networks (CNNs) for end-to-end feature extraction and classification of spectrograms has become the most similar alternative in both industry and academia, and is currently the mainstream alternative in industry. This approach treats the Doppler spectrogram as a standard two-dimensional image, employs classic image classification backbone networks such as ResNet, VGG, and EfficientNet, completes feature extraction through stacked convolutional layers, and finally outputs anomaly classification results through fully connected layers, achieving end-to-end recognition. Some improved solutions introduce channel attention modules such as CBAM and SE to enhance the ability to focus on key regions.

[0021] While current conventional deep learning methods have improved automation, they often treat spectrograms merely as ordinary two-dimensional images for local convolution processing, failing to effectively mine and utilize the specific physical laws inherent in the horizontal axis (time / cardiac cycle) and vertical axis (Doppler shift / blood flow velocity) of the spectrogram. Therefore, existing models struggle to accurately capture the subtle physiological signals that dynamically change with the cardiac cycle, and also fail to effectively distinguish extremely subtle morphological differences between different spectral envelopes. This results in severely insufficient accuracy and specificity in anomaly identification under complex clinical interference, failing to meet the clinical requirements of high accuracy and low false positive rates for prenatal screening in early pregnancy.

[0022] Existing analytical methods suffer from two main drawbacks: 1. Traditional manual analysis is time-consuming, labor-intensive, and highly subjective. Early automated solutions based on signal processing and machine learning (such as using active contour models to extract envelopes combined with support vector machine classification) heavily rely on tedious manual feature engineering. These pre-defined operators have extremely poor generalization ability to spectra acquired by different ultrasound devices, and feature extraction is prone to failure when encountering complex tissue motion artifacts or speckle noise. 2. The most similar alternative is to use conventional convolutional neural networks (CNNs) for end-to-end classification. However, CNNs directly treat spectrograms as ordinary two-dimensional natural images and perform local convolution processing. The smoothing of time-frequency features during convolution operations directly loses subtle pathological features in early pregnancy (such as subtle inversion of the a wave in the ductus venosus and minor pathological regurgitation of the tricuspid valve). Furthermore, conventional CNNs have limited receptive fields, making it difficult to effectively capture global temporal rhythms spanning the entire cardiac cycle, and difficult to distinguish between physiological variations and pathological abnormalities, resulting in insufficient specificity and sensitivity, and a persistently high false positive rate. Existing solutions introduce channel attention and spatial attention modules, which can only achieve coarse-grained regional weight allocation and cannot independently model "long-term temporal features of cardiac cycle" and "subtle abrupt changes in local blood flow velocity". Although the global Transformer solution can capture long-term dependencies, the computational cost is too high and it is severely overfitted to small sample data in early pregnancy.

[0023] To address the shortcomings of existing technologies, such as time-frequency feature confusion, limited local receptive field, and easy loss of fine-grained features, this application provides a method for identifying fetal Doppler blood flow spectrum anomalies in early pregnancy based on an axial attention mechanism. This method aims to overcome the limitations of conventional spatial convolution by achieving independent decoupling and deep logical association between the "time axis" and "frequency axis" features of the spectrogram through an axial attention mechanism. This method can accurately focus on long-term cardiac temporal rhythms and abrupt changes in axial local frequency shifts, adaptively amplify specific weak hemodynamic features in the Doppler blood flow spectrum, and suppress background noise. This enables efficient, robust, and high-precision end-to-end automated fine-grained classification, effectively meeting the needs of precise clinical ultrasound-assisted diagnosis.

[0024] In one embodiment, a method for identifying fetal Doppler blood flow spectrum abnormalities based on an axial attention mechanism is provided, applicable to the identification of fetal Doppler blood flow spectrum abnormalities in early pregnancy. Figure 1 As shown, the method includes: Step S110: Obtain the fetal Doppler spectrum dataset for early pregnancy. The fetal Doppler spectrum dataset for early pregnancy includes fetal Doppler spectrum images and annotated datasets. The fetal Doppler spectrum images for early pregnancy cover three types of target vessels: ductus venosus, tricuspid valve, and umbilical artery. The annotated dataset includes vessel type, abnormality type, and abnormal region coordinates.

[0025] Doppler flow spectrogram (DFS) is a medical signal visualization atlas based on the Doppler effect of ultrasound, used to record and analyze the dynamic changes in blood flow status within blood vessels during the heartbeat cycle. Essentially, it is a multidimensional unfolding of blood flow velocity: its horizontal axis (X-axis) represents time, reflecting the progression of the cardiac cycle, such as systole and diastole; the vertical axis (Y-axis) represents Doppler frequency shift or blood flow velocity, typically distinguished by the zero baseline to indicate whether blood flow is moving towards or away from the ultrasound probe; and the image brightness or grayscale (Z-axis) characterizes the signal energy intensity at a specific time and flow velocity, i.e., the relative number of red blood cells at that flow velocity.

[0026] Specifically, fetal Doppler spectral images during early pregnancy conforming to ISUOG guidelines were collected, covering three target vessels: ductus venosus, tricuspid valve, and umbilical artery. Each image contained at least three consecutive and stable fetal cardiac cycles. The images were annotated by two or more associate chief physicians with more than five years of experience in prenatal ultrasound diagnosis. The annotations included vessel type, abnormality type, and abnormal region coordinates. An annotated dataset containing normal and abnormal samples was constructed and divided into training, validation, and test sets in a 7:2:1 ratio for subsequent model training.

[0027] Step S120: Preprocess the fetal Doppler spectrum image in early pregnancy to obtain a preprocessed grayscale spectrum.

[0028] The specific method for preprocessing fetal Doppler spectrum images in early pregnancy is not unique and can be set according to actual needs. Specifically, step S120 includes: cropping the fetal Doppler spectrum image in early pregnancy, uniformly scaling it to a preset size and performing normalization processing; performing data enhancement on the normalized image, using enhancement methods such as random scaling, random Gaussian noise injection, and random contrast adjustment to simulate artifacts and signal fluctuations, to obtain a preprocessed grayscale spectrum image.

[0029] In this step, the images undergo preprocessing. First, the acquired early pregnancy fetal Doppler spectrum images are cropped and uniformly scaled to [size missing]. And perform normalization: Where x is the original pixel value, x min x max The minimum and maximum values ​​of the image pixels are defined as follows, and the image pixel values ​​are normalized to the [0,1] interval. Finally, data augmentation is performed on the clinical characteristics of the spectrum in early pregnancy using random scaling, random Gaussian noise injection, and random contrast adjustment to simulate artifacts and signal fluctuations in clinical scanning, avoiding model overfitting. At the same time, flipping operations are strictly prohibited to prevent damage to the physical meaning of blood flow direction.

[0030] Step S130: Construct a backbone network with an axial attention mechanism and a multi-task anomaly recognition head for three types of target vessels. Combine the preprocessed grayscale spectrograms to perform multi-task joint training of the model, resulting in a trained model. The backbone network is used for independent decoupling of the time and frequency axes. The trained model represents the correspondence between the fetal Doppler spectral images in early pregnancy and the recognition results of the three types of target vessels.

[0031] Attention mechanism (AM) is a core algorithmic component in deep learning, inspired by the human visual and cognitive ability to selectively focus on key areas while ignoring background noise when faced with a large amount of information. Specifically, when a neural network processes input data, the attention mechanism dynamically assigns different weights to different dimensions of the data. Regions with high weight scores are "amplified" and processed with emphasis by the network, while regions with low scores are suppressed.

[0032] In one embodiment, step S130 includes: constructing a backbone network with an axial attention mechanism based on the preprocessed grayscale spectrogram, performing dual-axis decoupling of the time axis and frequency axis, fusing features on the decoupling results to generate a high-level semantic feature map; constructing a multi-task anomaly recognition head for three types of target blood vessels based on the high-level semantic feature map, outputting the probability of occurrence of each type of anomaly, performing multi-task joint training in conjunction with a multi-task joint loss function, and dynamically adjusting the learning rate using a cosine annealing learning rate decay strategy to obtain a trained model.

[0033] The process involves extracting shallow features from the preprocessed grayscale spectrogram using convolutional layers, preserving the basic edge and texture features of the spectrum to generate a shallow feature map. Then, temporal attention, frequency-axis attention, and local detail convolution are performed on the shallow feature map to obtain the outputs of three branches. Finally, feature fusion is performed based on the outputs of the three branches to generate a high-level semantic feature map.

[0034] Furthermore, during feature fusion, the outputs of the three branches are fused, and the temporal attention, frequency attention, and local detail features are added element by element. After normalizing the fused features, nonlinearity is introduced through the GELU activation function to obtain a high-level semantic feature map.

[0035] This step involves constructing an axial attention mechanism to decouple the Doppler spectral data along two axes. For example... Figure 2 As shown, the entire backbone network adopts a hierarchical structure, divided into four feature extraction stages. Each stage consists of one convolutional block and two axial attention modules stacked together, as detailed below: 1. Shallow feature convolution embedding: The input preprocessed grayscale spectrum image is first processed by two 3×3 convolutional layers to extract shallow features, expanding the number of channels to 64 and generating a shallow feature map. This preserves the basic edge and texture features of the spectrum while reducing the complexity of subsequent attention calculations.

[0036] 2. The input to the axial attention module is a shallow feature map. H corresponds to the frequency axis of the spectrum, W corresponds to the time axis of the spectrum (horizontal axis), and C is the number of channels. The axial attention module is divided into 3 core branches, and the calculation process is completely decoupled to avoid the mixing of time and frequency features. Branch 1: Temporal Axis Attention Calculation. This branch is primarily responsible for capturing global dependencies along the temporal dimension. First, the input shallow feature map is globally averaged along the frequency axis, compressing it into one-dimensional sequence features: ; Then, a scaled dot product self-attention mechanism is applied along the time axis (W axis) to calculate the attention weights between time steps: ; In this calculation, And the attention head dimension is set to the number of feature channels ( Finally, the attention feature map is restored to its original spatial dimension through a broadcasting mechanism to obtain the temporal attention output: ; Branch 2: Frequency Axis Attention Calculation. This branch extracts relevance features along the frequency dimension. Similar to Branch 1, it first performs global average pooling along the time axis on the shallow feature map to extract sequence information along the frequency dimension. ; Next, self-attention computation is performed along the frequency axis (H-axis) to establish long-range dependencies between different frequency bands: ; in, ,and After the calculation is complete, the result is expanded back to the full feature map size using the broadcast operation to generate the frequency attention output: ; Branch 3: Local Detail Convolution. The third branch aims to preserve local spatial details in the original shallow feature maps and achieve cross-channel feature transformation. This is accomplished through a standard one-dimensional convolution operation: ; 3. Feature Fusion and Output: After obtaining the independent outputs of the three branches, the axial attention module enters the feature fusion stage. First, the temporal attention, frequency attention, and local detail features are added element-wise: ; To stabilize training and accelerate convergence, the fused features are normalized using a LayerNorm layer. The mean, For variance, For smoothing terms, , (for learnable parameters) ; Finally, by introducing nonlinearity through the GELU activation function (where erf is the error function), the final high-level semantic feature map is obtained: .

[0037] To address the identification needs of three core abnormalities in early pregnancy, a multi-task joint learning identification head is constructed to achieve end-to-end synchronous identification, avoiding feature redundancy in single-task models. First, the 7×7×512 high-level semantic feature map output by the backbone network is... A 512-dimensional global feature vector is generated through the fusion of global average pooling and global max pooling. A multi-task branching design divides the task into a main classification task, a vessel type branch, and an anomaly type branch. Each task combines the global feature vector to identify spectral anomalies. The main task is a binary classification task, determining whether the input spectrum contains pathological abnormalities and outputting the anomaly probability. The vessel type branch is a three-class classification task, identifying the vessel type (ductus venosus / tricuspid valve / umbilical artery) corresponding to the input spectrum, providing prior constraints for anomaly type identification. The anomaly type branch is a multi-label classification task, simultaneously identifying three core anomalies: ductus venosus a-wave abnormalities, tricuspid regurgitation, and umbilical artery end-diastolic flow loss, outputting the probability of occurrence for each type of anomaly.

[0038] After constructing the backbone network with the axial attention mechanism and the multi-task anomaly detection head, multi-task joint training of the model can be performed. To address the imbalanced nature of the early pregnancy dataset, a dedicated loss function and training strategy are designed, as follows: 1. Multi-task joint loss function. The total loss (Loss) is the weighted sum of the losses from the three branches, as shown in the formula: ; in, Focal Loss, used for the main classification branch, addresses the imbalance between normal and abnormal samples. The cross-entropy loss is for branching based on blood vessel type. The multi-label focus loss is used for the anomaly type branch; λ1, λ2, and λ3 are weights, set to 0.4, 0.2, and 0.4 respectively, to prioritize the accuracy of anomaly identification.

[0039] 2. Training Strategy. The optimizer used is AdamW, with an initial learning rate set to 1×10⁻⁶. -4 The learning rate is dynamically adjusted using a cosine annealing learning rate decay strategy. The training batch size is set to 16, and the total training rounds are 100. An early stopping mechanism is also introduced—training is terminated immediately when the validation set loss does not decrease for 10 consecutive rounds, effectively avoiding the risk of model overfitting. To address the limited sample size of clinical data in early pregnancy, the backbone network is initialized with weights pre-trained on a large-scale medical ultrasound image dataset. Fine-tuning training is then carried out on this basis, which not only ensures the model's convergence speed but also improves the model's adaptability to the Doppler spectral features of early pregnancy.

[0040] Step S140: After preprocessing the Doppler spectrum image of the fetus in early pregnancy to be identified, it is fed into the trained model for forward inference and the identification result is output. The identification result output by the model includes the blood vessel type, abnormality type, and corresponding abnormality probability value. The identification result can be output to display, stored, or used for further processing.

[0041] Specifically, the Doppler spectrum images of the fetus in early pregnancy, acquired in real time by ultrasound equipment, are input into the system. Standardized preprocessing operations, consistent with step S120 of the training phase, are then performed. The preprocessed images are then fed into the trained model for forward inference. The model can simultaneously output the vessel type, abnormality type, and corresponding abnormality probability value. Furthermore, the model's output recognition results can be combined to generate clinically readable diagnostic results, automatically marking the coordinates of abnormal areas and providing abnormality type hints and risk classifications, offering auxiliary diagnostic references for ultrasound physicians. Simultaneously, the inference results are archived and saved to the prenatal screening information system.

[0042] The aforementioned method for identifying fetal Doppler blood flow spectrum abnormalities based on the axial attention mechanism innovatively introduces and applies the biaxially decoupled axial attention mechanism to the physical properties of the Doppler spectrum in a deeply isomorphic manner to the specific clinical scenario of "early pregnancy Doppler spectrum abnormality identification." Its core points include: 1. For the preprocessing steps of fetal Doppler spectrum in early pregnancy, strictly preserve the orthogonal physical correspondence between the spectral time axis (horizontal axis) and the frequency axis (vertical axis) to complete the standardized input construction.

[0043] 2. Based on the feature extraction steps of the axially decoupled axial attention module, fully decoupled self-attention calculations are performed on the time axis and frequency axis of the spectrum, respectively. The time axis attention is constrained to the horizontal axis dimension to model the long-term temporal dependence of the cardiac cycle, while the frequency axis attention acts independently on the vertical axis dimension to capture subtle abrupt changes in blood flow velocity.

[0044] 3. Based on multi-task joint learning, the anomaly identification steps simultaneously complete the end-to-end identification of three core anomalies: abnormal a-wave in the ductus venosus during early pregnancy, tricuspid regurgitation, and loss of end-diastolic blood flow in the umbilical artery.

[0045] 4. Model training and inference output steps: Adapting to the clinical scenario of early pregnancy with a dedicated loss function and training strategy, outputting clinically readable auxiliary diagnostic results.

[0046] Compared with current 2D CNN-based schemes for identifying Doppler spectral anomalies in early pregnancy, this application has the following outstanding advantages: the feature extraction logic is deeply isomorphic to the physical properties of the spectrum, solving the core defect of easily losing weak features from the bottom layer. Existing CNN schemes treat the spectrum as ordinary natural images, indiscriminately mixing time-frequency features, which easily loses the weak pathological features in early pregnancy; this application innovatively proposes a dual-axis decoupled axial attention mechanism, strictly decoupling and modeling the time axis and frequency axis independently, which is completely consistent with the orthogonal physical properties of the Doppler spectrum. It can capture the global temporal rhythm across the complete cardiac cycle and accurately locate the key phases of abnormality, and can also focus on subtle changes in local frequency shifts, completely preserving the core abnormal features of ductus venosus a-wave abnormality, tricuspid regurgitation, and umbilical artery end-diastolic blood flow loss.

[0047] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0048] Based on the same inventive concept, this application also provides a fetal Doppler flow spectrum abnormality identification device based on the axial attention mechanism for implementing the aforementioned method for identifying fetal Doppler flow spectrum abnormalities. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the fetal Doppler flow spectrum abnormality identification device based on the axial attention mechanism provided below can be found in the limitations of the fetal Doppler flow spectrum abnormality identification method based on the axial attention mechanism described above, and will not be repeated here.

[0049] In one embodiment, such as Figure 3 As shown, a fetal Doppler blood flow spectrum abnormality identification device based on axial attention mechanism is provided, comprising: The data acquisition module 110 is used to acquire the fetal Doppler spectrum dataset in early pregnancy. The fetal Doppler spectrum dataset in early pregnancy includes fetal Doppler spectrum images and labeled datasets. The fetal Doppler spectrum images in early pregnancy cover three types of target vessels: ductus venosus, tricuspid valve, and umbilical artery. The labeled dataset includes vessel type, abnormality type, and abnormal region coordinates.

[0050] The data processing module 120 is used to preprocess the fetal Doppler spectrum image in early pregnancy to obtain the preprocessed grayscale spectrum.

[0051] The model training module 130 is used to construct the backbone network of the axial attention mechanism and the multi-task anomaly recognition head for three types of target vessels. It combines preprocessed grayscale spectrograms for multi-task joint training of the model to obtain the trained model. The backbone network is used for independent decoupling of the time and frequency axes. The trained model represents the correspondence between the fetal Doppler spectral images in early pregnancy and the recognition results of the three types of target vessels.

[0052] The anomaly recognition module 140 is used to preprocess the Doppler spectrum image of the fetus in early pregnancy to be identified, and then feed it into the trained model for forward inference and output the recognition result.

[0053] In one embodiment, the data processing module 120 is used to crop the fetal Doppler spectrum image in early pregnancy, uniformly scale it to a preset size and perform normalization processing: the normalized image is then subjected to data enhancement, and enhancement methods such as random scaling, random Gaussian noise injection and random contrast adjustment are used to simulate artifacts and signal fluctuations to obtain a pre-processed grayscale spectrum.

[0054] In one embodiment, the model training module 130 is used to construct a backbone network with an axial attention mechanism based on the preprocessed grayscale spectrogram, decouple the time axis and frequency axis, perform feature fusion on the decoupling result, and generate a high-level semantic feature map; based on the high-level semantic feature map, a multi-task abnormality recognition head for three types of target blood vessels is constructed, the output is the probability of occurrence of each type of abnormality, and multi-task joint training is performed in combination with a multi-task joint loss function, and the learning rate is dynamically adjusted by a cosine annealing learning rate decay strategy to obtain the trained model.

[0055] In one embodiment, the model training module 130 is used to perform shallow feature extraction through convolutional layers based on the preprocessed grayscale spectrogram, preserving the basic edge and texture features of the spectrum to generate a shallow feature map; perform time-axis attention calculation, frequency-axis attention calculation, and local detail convolution based on the shallow feature map to obtain the output results of the three branches; and perform feature fusion based on the output results of the three branches to generate a high-level semantic feature map.

[0056] In one embodiment, the model training module 130 is used to perform feature fusion on the output results of the three branches, add the temporal attention, frequency attention and local detail features element by element, normalize the fused features, and introduce nonlinearity through the GELU activation function to obtain a high-level semantic feature map.

[0057] Each module in the aforementioned fetal Doppler blood flow spectrum abnormality identification device based on the axial attention mechanism can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0058] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a method for identifying fetal Doppler blood flow spectrum abnormalities based on an axial attention mechanism. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0059] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0060] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0061] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above method.

[0062] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method described above.

[0063] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0064] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The processors involved in the embodiments provided in this application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited thereto.

[0065] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0066] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for identifying fetal Doppler blood flow spectrum abnormalities based on axial attention mechanism, characterized in that, include: Obtain a fetal Doppler spectrum dataset from early pregnancy; the fetal Doppler spectrum dataset from early pregnancy includes fetal Doppler spectrum images from early pregnancy and annotated datasets. The fetal Doppler spectrum images from early pregnancy cover three types of target vessels: ductus venosus, tricuspid valve, and umbilical artery. The annotated dataset includes vessel type, abnormality type, and abnormal region coordinates. The fetal Doppler spectrum image in early pregnancy is preprocessed to obtain a preprocessed grayscale spectrum image; A backbone network with an axial attention mechanism and a multi-task anomaly recognition head for three types of target blood vessels are constructed. The model is then trained jointly for multiple tasks using preprocessed grayscale spectrograms to obtain a trained model. The backbone network is used to perform independent decoupling of the time axis and frequency axis. The trained model represents the correspondence between the fetal Doppler spectrum image in early pregnancy and the recognition results of the three types of target blood vessels. After preprocessing the Doppler spectrum image of the fetus in early pregnancy to be identified, it is fed into the trained model for forward inference and the identification result is output.

2. The method according to claim 1, characterized in that, The fetal Doppler spectrum image in early pregnancy is preprocessed to obtain a preprocessed grayscale spectrum image, including: The fetal Doppler spectrum images from early pregnancy are cropped, scaled to a preset size, and normalized. Data augmentation is performed on the normalized image using random scaling, random Gaussian noise injection, and random contrast adjustment to simulate artifacts and signal fluctuations, resulting in a pre-processed grayscale spectrum.

3. The method according to claim 1, characterized in that, The backbone network constructing the axial attention mechanism and the multi-task anomaly recognition head targeting three types of target blood vessels are combined with preprocessed grayscale spectrograms for multi-task joint training to obtain the trained model, including: Based on the preprocessed grayscale spectrogram, a backbone network with an axial attention mechanism is constructed to decouple the time axis and frequency axis. The decoupling results are then fused to generate a high-level semantic feature map. Based on the high-level semantic feature map, a multi-task anomaly recognition head is constructed for three types of target blood vessels. The output is the probability of occurrence of each type of anomaly. Multi-task joint training is performed by combining a multi-task joint loss function and a cosine annealing learning rate decay strategy to achieve dynamic adjustment of the learning rate, thus obtaining the trained model.

4. The method according to claim 3, characterized in that, Based on the preprocessed grayscale spectrogram, a backbone network with an axial attention mechanism is constructed to decouple the time and frequency axes. Feature fusion is then performed on the decoupling results to generate a high-level semantic feature map, including: Based on the preprocessed grayscale spectrum, shallow feature extraction is performed through convolutional layers to preserve the basic edge and texture features of the spectrum and generate a shallow feature map. Based on the shallow feature map, perform temporal axis attention calculation, frequency axis attention calculation, and local detail convolution to obtain the output results of the three branches; Feature fusion is performed based on the outputs of the three branches to generate a high-level semantic feature map.

5. The method according to claim 4, characterized in that, Based on the outputs of the three branches, feature fusion is performed to generate a high-level semantic feature map, including: The outputs of the three branches are fused, and the temporal attention, frequency attention and local detail features are added element by element. After normalizing the fused features, nonlinearity is introduced through the GELU activation function to obtain a high-level semantic feature map.

6. The method according to claim 3, characterized in that, The multi-task joint loss function is: ； in, The focus loss for the main classification branch, The cross-entropy loss is the branching loss of blood vessel types. λ1, λ2, and λ3 are the weights for the multi-label focus loss of the abnormal type branch.

7. The method according to any one of claims 1 to 6, characterized in that, The model outputs the identification results, including blood vessel type, abnormality type, and corresponding abnormality probability value.

8. A device for identifying fetal Doppler blood flow spectrum abnormalities based on axial attention mechanism, characterized in that, include: The data acquisition module is used to acquire a fetal Doppler spectrum dataset in early pregnancy. The fetal Doppler spectrum dataset in early pregnancy includes fetal Doppler spectrum images and labeled datasets. The fetal Doppler spectrum images in early pregnancy cover three types of target vessels: ductus venosus, tricuspid valve, and umbilical artery. The labeled dataset includes vessel type, abnormality type, and abnormal region coordinates. The data processing module is used to preprocess the fetal Doppler spectrum image in early pregnancy to obtain a preprocessed grayscale spectrum image. The model training module is used to construct the backbone network of the axial attention mechanism and the multi-task anomaly recognition head for three types of target blood vessels. The model is jointly trained by multi-task using the preprocessed grayscale spectrogram to obtain the trained model. The backbone network is used to perform independent decoupling of the time axis and frequency axis. The trained model represents the correspondence between the fetal Doppler spectrum image in early pregnancy and the recognition results of the three types of target blood vessels. The anomaly detection module is used to preprocess the Doppler spectrum image of the fetus in early pregnancy to be identified, and then feed it into the trained model for forward inference to output the recognition result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images