Ultrasound echocardiogram quality evaluation method based on nonlinear causal geometric learning

Abstract

The present application belongs to the technical field of echocardiogram quality evaluation, and specifically discloses an echocardiogram quality evaluation method based on nonlinear causal geometric learning. In view of the deficiencies of the existing echocardiogram quality evaluation method, the present application proposes an echocardiogram quality evaluation model, wherein a causal geometric learning module and a nonlinear relationship aggregation module are designed at the model architecture level, so that the model can simultaneously explicitly model the directed causal relationship between visual factors and their geometric topological structure, and further aggregate the relationship structure and local quality features differently, so that the quality determining factors can be strengthened in the relationship space, and the co-occurring but non-determining visual factors are suppressed, to obtain more stable, topologically aware and discriminative quality representation, thereby improving the accuracy, stability and consistency of the evaluation result with the true quality label of the echocardiogram quality evaluation under complex imaging conditions.

Description

Technical Field

[0001] This invention belongs to the field of echocardiography quality assessment technology, and specifically relates to an echocardiography quality assessment method based on nonlinear causal geometric learning. Background Technology

[0002] Echocardiographic quality assessment is a crucial foundation for ensuring the reliability of cardiac structure identification, functional analysis, and clinical auxiliary diagnosis. The imaging quality of echocardiograms is easily affected by various factors, including equipment performance, imaging parameters, operator experience, acoustic window conditions, and individual patient differences, leading to quality degradation phenomena such as blurred boundaries, incomplete chamber visualization, structural distortion, increased noise, and changes in contrast. Different quality degradation factors often coexist and interact, making the process of echocardiographic quality formation highly complex. Therefore, extracting stable and discriminative quality representations from complex visual changes is a critical issue in automated echocardiographic quality assessment. Existing deep learning-based quality assessment methods typically achieve quality prediction by learning the mapping relationship between image visual features and quality labels.

[0003] However, various visual factors in echocardiography have different effects on quality assessment results: some visual factors directly affect the effective display of key anatomical structures such as heart chambers, myocardium, and valves, and are important factors determining image quality; other visual changes, although often occurring simultaneously with low-quality images, do not directly determine the quality level. Existing methods typically encode these visual factors uniformly as overall features, making it difficult to distinguish between quality determinants and co-occurring visual degradation factors. This easily leads to models relying on non-deterministic visual cues, thus affecting the stability and robustness of quality assessment results.

[0004] In recent years, causal learning has been used to identify more stable quality information from relevant features. Existing correlation methods mainly reduce the influence of unstable visual factors through features-dependent conditions, causal adaptation, or domain-invariant representations, and have some role in cross-device or cross-domain quality assessment. However, these methods typically focus on obtaining stable feature representations using causal relationships, lacking further structured description of the directed relationship structure formed between multiple visual factors. In other words, simply estimating the directional effect of one visual factor on another cannot fully reflect the hierarchical relationships and topological structures formed by multiple quality-related factors, thus making it difficult to reliably distinguish between key factors that dominate quality changes and visual factors that merely accompany them.

[0005] On the other hand, geometric learning can describe complex data structures using distances or topological relationships between features, and existing methods have applied geometric relationships to tasks such as sample-level representation learning, structure preservation, and knowledge transfer. However, such geometric modeling typically focuses on the global structural relationships between different samples, without addressing the geometric organization of directed causal relationships between different quality-related visual factors within a single echocardiogram. Therefore, simply using sample-level geometric relationship modeling is still insufficient to describe the asymmetric and hierarchical causal interaction structure between quality factors.

[0006] Furthermore, echocardiographic quality depends not only on the local visual factors themselves but also on the combined influence of the spatial-semantic relationships between different visual factors. Existing methods typically compress local visual features and relational information into a unified feature space, causing variations in the visual factors themselves to mix with changes in the relational structure between them. When imaging conditions or quality degradation patterns change, the model struggles to separately determine the impact of local visual changes and relational structure changes on the quality assessment results, resulting in feature entanglement between relational representations and local quality representations.

[0007] In summary, existing methods for evaluating echocardiographic quality have the following two main shortcomings: 1. The directional relationship structure between quality determinants and co-occurring visual degradation factors is difficult to effectively distinguish. Existing methods can learn the statistical correlation between visual factors and quality outcomes, but lack joint modeling of the directional effects and topological structures among different visual factors. This can easily lead to the inclusion of visual factors that co-occur with low-quality images but do not play a decisive role in the quality representation. 2. There is representational entanglement between causal relationship structures and local quality features. Existing feature aggregation methods typically compress relational information and local visual information in a unified manner, making it difficult to separately represent the different effects of changes in relational structure and local visual changes on quality evaluation, thus limiting the model's ability to express complex quality degradation patterns.

[0008] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0009] The purpose of this invention is to propose an echocardiographic quality assessment method based on nonlinear causal geometric learning. By simultaneously and explicitly modeling the directed causal relationships and geometric topological structures between visual factors, and further differentially aggregating the relationship structure with local quality features, the quality determinants can be strengthened in the relationship space, while co-occurring but non-deterministic visual factors are suppressed. This results in a more stable quality representation with topological awareness and discriminative ability, improving the accuracy, stability, and consistency of the evaluation results with the actual quality annotation under complex imaging conditions.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: The echocardiography quality assessment method based on nonlinear causal geometric learning includes the following steps: Step 1. Acquire echocardiogram images and preprocess them, then construct a training dataset; Step 2. Build an echocardiography quality assessment model, which includes a feature extraction network, a causal geometry learning module, a Transformer encoder, a nonlinear relationship aggregation module, a view classification head, and a quality prediction head; The preprocessed echocardiogram is input into the feature extraction network to obtain visual factor features; the causal geometry learning module models the directed causal relationship and its geometric topology between different visual factors to obtain causal enhancement features; Subsequently, a task token sequence is constructed based on the causal enhancement features, view token, quality token, and location embedding, and the task token sequence is input into the Transformer encoder to obtain relation enhancement features; The relation enhancement features output by the Transformer encoder are respectively fed into the view classification auxiliary branch and the quality evaluation main branch; The view classification auxiliary branch predicts the view category of the echocardiogram using the view classification head; The main branch of quality assessment first performs nonlinear mapping, importance weighting, and residual aggregation on the relationship enhancement features through the nonlinear relationship aggregation module, and then outputs the echocardiogram quality assessment results by the quality prediction head; Step 3. Train the echocardiography quality assessment model built in Step 2 based on the training dataset constructed in Step 1, and use the trained echocardiography quality assessment model to perform echocardiography quality assessment.

[0011] The present invention has the following advantages: As described above, this invention discloses an echocardiographic quality assessment method based on nonlinear causal geometric learning. This method constructs an echocardiographic quality assessment model. In terms of network structure, the model incorporates a causal geometric learning module. By estimating the directed conditional dependencies between different visual factors, it utilizes the asymmetry of the causal interaction matrix to calculate the causal effects of each visual factor and constrains the hierarchical topology of multi-hop causal relationships in latent hyperbolic space. This strengthens the visual factors that truly determine echocardiographic quality, suppresses visual factors that co-occur with quality degradation but do not have a decisive effect, and reduces the interference of unstable visual cues on quality representation. The model also includes a nonlinear relationship aggregation module. Through dimension-wise nonlinear mapping, relationship importance weighting, and residual aggregation, it differentiates the enhanced patchtokens for different relationships. This strengthens key relationship information closely related to quality assessment, suppresses redundant relationship information, and, while preserving basic quality features, alleviates the representational entanglement between the causal relationship structure and local visual features, improving the compactness and discriminativeness of the quality representation. By collaboratively connecting the causal geometry learning module, the Transformer encoder, the nonlinear relation aggregation module, and the view classification auxiliary branch, and jointly training the model using geometric topology loss, view classification loss, and quality loss, the causal geometry learning module first performs causal enhancement and topological constraints on visual factors. The Transformer encoder further models the contextual relationships between different visual factors, the nonlinear relation aggregation module performs differential aggregation of relation features, and the view classification auxiliary branch provides anatomical view semantic supervision for quality assessment. These components work together to enable the model to obtain a more stable, topology-aware, and discriminative quality representation, thereby improving the accuracy and stability of echocardiographic quality assessment under complex imaging conditions and enhancing the consistency between the assessment results and the actual quality annotations. Attached Figure Description

[0012] Figure 1 This is a network structure diagram of the echocardiography quality assessment model in an embodiment of the present invention; Figure 2 This is a network structure diagram of the causal geometry learning module in an embodiment of the present invention; Figure 3 This is a network structure diagram of the nonlinear relationship aggregation module in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the visualization analysis of the region of interest of the proposed echocardiographic quality assessment model on three datasets: CACTUS, CLINAv-GKD, and Echo-Q. in Figure 4 (a), (b), and (c) in the figure represent three datasets: CACTUS, CLINAV-GKD, and Echo-Q, respectively. Detailed Implementation

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: To address the aforementioned shortcomings of existing echocardiography quality assessment methods, this invention describes an echocardiography quality assessment method based on nonlinear causal geometric learning, which includes the following steps: Step 1. Acquire echocardiogram images and preprocess them, then construct a training dataset.

[0014] First, extract image frames from the echocardiogram video sequence, or directly obtain a single echocardiogram image.

[0015] The original echocardiogram images were cropped to retain the effective fan-shaped imaging area, while irrelevant background outside the effective imaging area was removed. The cropped images were then resized uniformly. pixels, to obtain preprocessed echocardiogram images .

[0016] For each preprocessed echocardiogram image, a view category label is set. and quality label .

[0017] View category labels This is used to indicate the preset standard echocardiographic view category to which the image belongs. The specific view category is determined based on the standard views contained in the training dataset. Examples include the parasternal left ventricular long-axis view, apical two-chamber view, apical three-chamber view, apical four-chamber view, apical five-chamber view, and subxiphoid four-chamber view, aortic level short-axis view, papillary muscle level short-axis view, and mitral valve level short-axis view.

[0018] Quality Label It is used to annotate based on the display integrity, boundary clarity, image noise, contrast, and artifacts of key anatomical structures such as heart chambers, myocardium, and valves.

[0019] Quality Label Numerical quality scoring or discrete quality grading is used.

[0020] Numerical quality scores are used to represent the degree of continuous variation in echocardiographic quality; discrete quality grades divide echocardiography into multiple quality categories from low to high according to a preset quality classification standard.

[0021] A higher quality score or quality grade indicates better echocardiographic imaging quality.

[0022] In the datasets used in this embodiment, the CACTUS dataset uses 11 quality levels from 0 to 10, while the CLINAv-GKD dataset and the Echo-Q dataset both use 5 quality levels (Excellent, Good, Medium, Poor, Very Poor).

[0023] The first Preprocessed echocardiogram images and their corresponding view category labels and quality label Matching is performed to form training samples. The training dataset consists of multiple training samples.

[0024] View category labels and quality labels are used to supervise the training of the view classification auxiliary branch and the quality evaluation main branch, respectively.

[0025] When the quality label uses a numerical quality score, the model outputs the corresponding predicted quality score; when the quality label uses a discrete quality level, the model outputs the corresponding predicted quality level.

[0026] The model view classification auxiliary branch (i.e., the view classification head mentioned below) outputs the predicted probability of each preset view category and takes the view category with the highest predicted probability as the predicted view category.

[0027] Step 2. Construct an echocardiography quality assessment model, the network structure of which is as follows: Figure 1 As shown.

[0028] This echocardiography quality assessment model takes preprocessed echocardiogram images as input, obtains quality assessment results through causal geometric relationship modeling and nonlinear relationship aggregation, and introduces a view classification task as auxiliary supervision.

[0029] like Figure 2 As shown, the echocardiography quality assessment model includes a feature extraction network, a causal geometry learning module, a Transformer encoder, a nonlinear relationship aggregation module, a view classification head, and a quality prediction head.

[0030] The overall processing flow of the echocardiography quality assessment model is as follows: The preprocessed echocardiogram is input into the feature extraction network to obtain visual factor features; the causal geometry learning module models the directed causal relationship and its geometric topology between different visual factors to obtain causal enhancement features; Subsequently, a task token sequence is constructed based on the causal enhancement features, view token, quality token, and location embedding, and the task token sequence is input into the Transformer encoder to obtain relation enhancement features; The relation enhancement features output by the Transformer encoder are respectively fed into the view classification auxiliary branch and the quality evaluation main branch; The view classification auxiliary branch predicts the view category of the echocardiogram using the view classification head; this view classification auxiliary branch is used to provide anatomical view semantic supervision for quality assessment. The main branch of quality assessment first performs nonlinear mapping, importance weighting, and residual aggregation on the relationship enhancement features through the nonlinear relationship aggregation module, and then outputs the echocardiogram quality assessment results by the quality prediction head.

[0031] The following section provides a detailed explanation of each component of the echocardiography quality assessment model.

[0032] In this embodiment, the feature extraction network, such as the ResNet network, has the following processing flow: The feature extraction network first receives preprocessed echocardiogram images. and extract from A visual factor set composed of visual factor feature maps .

[0033] in For the first Each visual factor feature map represents the spatial response of the corresponding visual pattern in the input image. The dimensions of the visual factor feature map are its height and width.

[0034] Then, the feature maps of each visual factor are stacked along the channel direction to obtain the visual factor tensor. .

[0035] Visual Factor Tensor Input the causal geometry learning module.

[0036] This causal geometry learning module models the directed causal relationships and their geometric topology among different visual factors, and enhances the visual factor tensors based on the causal effects of each visual factor, outputting causal enhancement features. .

[0037] like Figure 2 The network structure of the causal geometry learning module is shown.

[0038] This causal geometry learning module is used to explicitly model the directed causal relationships between different visual factors and to constrain the hierarchical topology of the causal relationships using latent hyperbolic space. This strengthens visual factors that are decisive for image quality and suppresses visual factors that co-occur with quality degradation but do not play a decisive role.

[0039] This causal geometry learning module uses visual factor tensors As input, the system sequentially performs directed causal relationship estimation, causal effect calculation, causal feature enhancement, and latent hyperbolic geometric topology learning, outputting causal enhanced features. .

[0040] For visual factor set Feature maps of any two visual factors and First, calculate the feature map of visual factors. Pointing to visual factor feature map Directed conditional dependency : ; in, This represents element-wise multiplication at corresponding spatial locations. This represents the stability constant to prevent the denominator from being zero. This represents the generalized Lehmer mean nonlinear aggregation operator.

[0041] in, This represents a stability constant to prevent the denominator from being zero. In this embodiment, Pick .

[0042] This definition defines a directed dependency as the strength of the conditional dependency between two features. Represented by visual factor feature map When used as a conditional reference, by point to The directed conditional dependency strength is used to characterize visual factors. Visual factors The relative influence of the two is used as an estimate of the strength of the directed causal interaction between them.

[0043] For the Visual Factor Feature Map The generalized Lehmer mean nonlinear aggregation operator is expressed as: .

[0044] in, Indicates the first The visual factor feature map in the first The response value at each spatial location, This represents the total number of spatial locations in the feature map of each visual factor. and These are learnable parameters.

[0045] By adjusting and This generalized Lehmer mean nonlinear aggregation operator can adaptively change its sensitivity to different response intensities and response distributions, thereby improving the stability of directed conditional dependency estimation.

[0046] Calculate the directed conditional dependencies for each combination of visual factors and construct a causal interaction matrix. This enables the estimation of directed causal relationships.

[0047] Due to directed conditional dependency and These represent two conditional dependencies in opposite directions, and therefore they are usually not equal; therefore, a causal interaction matrix is ​​used. Asymmetry, calculating the causal action vector .

[0048] ; in Representing the causal interaction matrix The transpose of the matrix, This represents the Sigmoid function. The parameter representing the sensitivity to controlling asymmetric relationships (in this embodiment) Take 0.1). Represents a column vector consisting entirely of 1s. Indicates in-sample standardization; Causal action vector The first in The nth component represents the nth component. The relative dominance of each visual factor in all directed causal relationships; based on the causal action vector Visual factor tensor Channel-level modulation is performed to obtain causal enhancement features. : ; in, Represents the causal modulation intensity parameter (and In this embodiment, A value of 0.5 is used to keep the causal modulation coefficient positive, thereby enhancing visual factors with strong causal effects and suppressing non-deterministic visual factors while preserving the basic information of the original features. This represents the hyperbolic tangent function.

[0049] During calculation, Extended to the The entire spatial location of the feature map of each visual factor enhances the visual factors with strong causal effects and suppresses the influence of non-deterministic visual factors on quality characterization.

[0050] To further maintain the causal interaction matrix The described hierarchical topology, for the first Spatial aggregation and feature mapping of visual factors yield Euclidean embeddings. Then, it is projected onto the latent hyperbolic space through an exponential mapping at the origin. , in, Indicates the first Geometric embedding of visual elements; Indicates curvature parameter as The exponential mapping of the latent hyperbolic space at the origin; for the causal interaction matrix The singular value decomposition is expressed by the following formula: ; in, and Let these represent the left singular vector matrix and the right singular vector matrix, respectively. Represents the sum of the singular values The vector is composed of singular values; then, a weighted power series expansion is performed on each singular value.

[0051] Specifically, for the first A singular value its expanded structural spectral components Represented as: , in, Indicates the relation order index. These correspond to second-order, fourth-order, and even-order relations, respectively. This represents the highest order of the weighted power series expansion. This represents a scale parameter that controls the degree of contribution of different order relationships.

[0052] The structure spectrum is composed of all structure spectrum components. .

[0053] The intensity of multi-hop structural interactions between different visual factors is represented as follows: ; in Indicates the first The visual factor and the first The intensity of multi-hop structural interactions between visual factors.

[0054] Construct a geometric topological loss based on the interaction strength of multi-hop structures: ; in, Indicates the relationship with the first Visual factors that dominate causal relationships Indicates the relationship with the first The causal relationship between visual factors is suppressed.

[0055] This geometric topological loss This allows the dominant causal relationships to have a tighter geometric structure in the latent hyperbolic space, while keeping the suppressed relationships relatively separate.

[0056] The causal geometry learning module ultimately outputs causal enhancement features. and geometric topology loss .

[0057] Among them, causal enhancement features Used for token construction in subsequent tasks and input into the Transformer encoder, geometric topology loss. For joint training of echocardiography quality assessment models.

[0058] The process of constructing the task token sequence is as follows: First, the causal enhancement features obtained from the causal geometry learning module are... Partition and map to Each Patch is embedded, and the corresponding position embedding is added to each Patch embedding.

[0059] Set the initial view token for view classification before the patch embedding sequence. and the initial quality token used for quality evaluation This constitutes a task token sequence. The formula is expressed as follows: ; in, Indicates the first A patch embedding, This indicates the corresponding position embedding. .

[0060] Task token sequence The input is a Transformer encoder, which models the contextual relationships between different visual factors through a self-attention mechanism. After the Transformer encoder outputs, view relationship features corresponding to the view tokens are extracted. .

[0061] and quality token The set of relationship enhancement patch tokens is denoted as { }, .

[0062] View relationship features The input view classification header is mapped by a multilayer perceptron to obtain the predicted view category. This branch is used to provide semantic supervision of anatomical views, helping the network distinguish quality features under different standard cross-sectional conditions.

[0063] Relationship Enhancement Patch Token Collection { The input nonlinear relationship aggregation module generates quality relationship features through nonlinear mapping, relationship importance weighting, and residual aggregation. , Input the quality prediction header to obtain the quality prediction results. .

[0064] The nonlinear relation aggregation module is used to further aggregate relation enhancement patch tokens into compact and discriminative quality relation features, while alleviating feature entanglement between relation structure information and local visual information.

[0065] This module uses quality tokens. and relationship enhancement Patch token set As input, the algorithm sequentially performs basic relation aggregation, dimension-wise nonlinear mapping, relation importance weighting, and residual aggregation, and outputs the aggregated quality relation features. .

[0066] like Figure 3 The network structure of the nonlinear relationship aggregation module is shown below, and its processing flow is as follows: First, the quality token With relationship enhancement Patch token set { Perform basic relation aggregation to obtain basic quality relation features. .

[0067] This basic quality relationship feature is used to represent the overall relationship between the quality token and different patch tokens, and is the basis for subsequent nonlinear relationship enhancement and residual aggregation.

[0068] For the A relationship enhancement patch token , and its first Each channel component is denoted as ; The nonlinear relationship aggregation module performs dimension-wise function decomposition on each channel component and models the nonlinear relationships in different channels using multiple learnable piecewise spline basis functions, as expressed in the following formula: ; in For the first The nonlinear relation response is obtained by performing a dimension-wise nonlinear mapping on the relation enhancement patch token.

[0069] This represents the channel dimension of the relationship enhancement Patchtoken. This represents the number of piecewise spline basis functions corresponding to each channel. Indicates the action on the first The first channel A learnable piecewise spline basis function This represents the corresponding learnable combination weights.

[0070] Through dimension-wise nonlinear mapping, different channels and different relational patterns can be modulated using different nonlinear functions, thereby avoiding the homogenization of all relational features by using a uniform linear mapping.

[0071] This process allows for a more detailed description of the different impacts of various relational structures on echocardiographic quality assessment results.

[0072] After obtaining the nonlinear relation response corresponding to each relation enhancement patch token, the importance weight of each relation enhancement patch token is calculated by the relation importance estimation unit. .

[0073] The relation importance estimation unit consists of a first fully connected layer, a ReLU activation layer, a second fully connected layer, and a Softmax normalization layer connected sequentially. Its processing flow is as follows: First, the first A relationship enhancement Patchtoken The first fully connected layer performs feature mapping, followed by a non-linear transformation using the ReLU activation function; then, the transformed features are mapped to scalar importance scores through a second fully connected layer. : ; in, and These represent the weight parameters of the first fully connected layer and the second fully connected layer, respectively. and These represent the corresponding bias parameters.

[0074] Finally, the importance scores of all relation-enhancing Patchtokens are normalized using Softmax to obtain the [number of]th [item / value]. Each relationship enhances the importance weight of Patchtoken: .

[0075] in For the first This relationship enhances the relative contribution of Patchtoken to the quality assessment results.

[0076] Normalize all importance weights to satisfy: and .

[0077] The non-linear relational responses of each relation-enhancing patch token are then applied. According to the corresponding importance weight Weighting is performed, and the relationship between residuals and basic quality characteristics is analyzed. By fusing the data, we obtain the aggregated quality relationship characteristics. : .

[0078] in, This represents the learnable residual scaling factor, used to control the modulation strength of the nonlinear relation enhancement term on the basic quality relation features. The initial value is set to 1, and it is adaptively updated through backpropagation during model training.

[0079] Residual joins are used to enhance key relationship patterns related to quality assessment and suppress redundant or non-deterministic relationship information; residual joins are used to retain effective information in the basic quality relationship features and avoid loss of original quality semantics caused by nonlinear relationship aggregation.

[0080] The aggregated quality relationship characteristics Input a nonlinear quality prediction head. The nonlinear quality prediction head utilizes a set of learnable piecewise univariate functions to... Mapping to the quality prediction space, we obtain the quality prediction results, expressed by the following formula: ; in This indicates the number of piecewise univariate functions in the quality prediction header. Indicates the first A learnable piecewise single-variable function This represents the corresponding learnable combination weights. This represents the final predicted quality value of the echocardiogram.

[0081] Step 3. Train the echocardiography quality assessment model built in Step 2 based on the training dataset constructed in Step 1, and use the trained echocardiography quality assessment model to perform echocardiography quality assessment.

[0082] During the training of the echocardiography quality assessment model, a total training loss is constructed based on geometric topology loss, view classification loss, and quality loss. : ; in The geometric topology loss is generated by the causal geometric learning module. For view classification loss, For quality loss. Based on the predicted view category. and quality prediction results Calculate the view classification loss and quality loss separately.

[0083] , and These represent the weight coefficients for geometric topology loss, view classification loss, and quality loss, respectively, used to adjust the impact of each loss term on the model training process. In this embodiment, all three weight coefficients are set to 1.

[0084] During training, based on total training loss Backpropagation is performed to jointly update the parameters involved in training the echocardiography quality assessment model until the model reaches the preset number of training rounds or the total training loss converges.

[0085] The parameters used in the training specifically include: the parameters of the feature extraction network, and the generalized Lehmer mean parameter in the causal geometry learning module. and In addition, Euclidean embedding mapping parameters, Patch embedding mapping parameters, initial view token, initial quality token and position embedding, Transformer encoder parameters, view classification header parameters, piecewise spline basis function parameters in the nonlinear relation aggregation module, and learnable combinatorial weights. Importance estimation parameters and residual scaling factor And the piecewise univariate function parameters and learnable combined weights in the nonlinear quality prediction head. wait.

[0086] After the echocardiography input model is trained, the echocardiogram to be evaluated is input into the trained echocardiogram input model to obtain the corresponding view category and output the corresponding quality score or quality level.

[0087] During the inference phase, only the forward signal flow is retained. The input echocardiogram image is sequentially processed through feature extraction, causal geometry learning, Transformer encoding, and nonlinear relation aggregation, and finally outputs the corresponding view category, quality score, or quality level.

[0088] In addition, to verify the effectiveness of the method of the present invention (GeoCAN), comparative experiments were conducted on the public datasets CACTUS, CLINAv-GKD, and the clinical dataset Echo-Q.

[0089] The comparison methods include KonCept, IQT, TRIQ, LAR-IQA, Causal-IQA, and TreS.

[0090] Among them, KonCept is a no-reference image quality assessment method based on deep convolutional networks, which predicts quality scores by extracting high-level perceptual features of images; IQT is a perceptual image quality assessment method based on Transformer, which models the relationship between perceptual features of images through the Transformer structure; TRIQ is an image quality assessment method based on Transformer encoder, which extracts local features through convolutional neural networks and aggregates global quality information using Transformer encoder; LAR-IQA is a lightweight, accurate and robust no-reference image quality assessment method, which achieves quality prediction under different distortion conditions through a lightweight network structure; Causal-IQA is a generalized image quality assessment method based on causal inference, which reduces the impact of confounding factors on quality prediction through causal optimization; and TreS is a no-reference image quality assessment method based on Transformer, relative ranking and self-consistency, which achieves quality prediction by jointly modeling local features, global features and relative quality relationships between images.

[0091] Evaluation metrics include Quality Assurance (QA), View Classification (CA), Pearson Linear Correlation Coefficient (PLCC), Spearman Rank Correlation Coefficient (SRCC), and Kendall Rank Correlation Coefficient (KRCC).

[0092] Table 1. Comparison of the method of this invention with other methods on the CACTUS and CLINAv-GKD datasets.

[0093] As shown in Table 1, on the CACTUS and CLINAv-GKD datasets, the QA of the method of this invention reached 0.8740 and 0.9696, respectively; PLCC reached 0.9850 and 0.9897, respectively; SRCC reached 0.9833 and 0.9899, ​​respectively; and KRCC reached 0.9587 and 0.9831, respectively, which are generally better than the compared methods; the view classification accuracy is also basically consistent with the optimal results.

[0094] Table 2 shows a quantitative comparison of the method of the present invention with other methods on the clinical dataset Echo-Q.

[0095] Table 2 Comparison of the method of the present invention with other methods on the clinical dataset Echo-Q.

[0096] As shown in Table 2, on the clinical dataset Echo-Q, the QA, CA, PLCC, SRCC and KRCC of the method of the present invention reached 0.8721, 0.9923, 0.9027, 0.8983 and 0.8770 respectively, all of which achieved the best results.

[0097] The experimental results above show that the method of the present invention can effectively improve the accuracy of echocardiographic quality prediction and its consistency with the actual quality label, and has good stability under different datasets and complex clinical imaging conditions.

[0098] To further verify the effectiveness of the method of this invention, a visualization analysis was performed on the model interest regions on the CACTUS, CLINAv-GKD, and Echo-Q datasets, such as... Figure 4 As shown in (a), (b), and (c) of the diagram. Figure 4 As can be seen, the method of the present invention can focus attention on the boundaries of the heart chambers, the myocardial region and other anatomical topologies that are closely related to image quality, rather than image noise, local artifacts or visually degraded areas related to the surface of the quality result, under different views and imaging conditions.

[0099] The above results demonstrate that the causal geometry learning module of this invention can strengthen the structural factors that truly determine image quality and suppress co-occurring but non-deterministic visual factors. Furthermore, the region of interest in the model of this invention exhibits structured relationships between multiple anatomical regions, rather than merely isolated local responses. This indicates that the nonlinear relationship aggregation module can effectively fuse the relationship information between different visual factors, alleviating the representational entanglement between relational structures and local quality features.

[0100] The attention patterns on the three datasets above show good consistency, indicating that the method of the present invention can obtain a stable quality characterization with anatomical topology awareness and provide good interpretability for the quality evaluation results.

[0101] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A method for evaluating echocardiographic quality based on nonlinear causal geometric learning, characterized in that, Includes the following steps: Step 1. Acquire echocardiogram images and preprocess them, then construct a training dataset; Step 2. Build an echocardiography quality assessment model, which includes a feature extraction network, a causal geometry learning module, a Transformer encoder, a nonlinear relationship aggregation module, a view classification head, and a quality prediction head; The preprocessed echocardiogram is input into a feature extraction network to obtain visual factor features; The causal geometry learning module models the directed causal relationships and their geometric topology between different visual factors to obtain causal enhancement features; Subsequently, a task token sequence is constructed based on the causal enhancement features, view token, quality token, and location embedding, and the task token sequence is input into the Transformer encoder to obtain relation enhancement features; The relation enhancement features output by the Transformer encoder are respectively fed into the view classification auxiliary branch and the quality evaluation main branch; The view classification auxiliary branch predicts the view category of the echocardiogram using the view classification head; The main branch of quality assessment first performs nonlinear mapping, importance weighting, and residual aggregation on the relationship enhancement features through the nonlinear relationship aggregation module, and then outputs the echocardiogram quality assessment results by the quality prediction head; Step 3. Train the echocardiography quality assessment model built in Step 2 based on the training dataset constructed in Step 1, and use the trained echocardiography quality assessment model to perform echocardiography quality assessment.

2. The echocardiography quality assessment method based on nonlinear causal geometric learning according to claim 1, characterized in that, Step 1 specifically involves: Acquire raw echocardiogram images, retain the effective fan-shaped imaging region in the images, and uniformly adjust the image size. pixels, to obtain preprocessed echocardiogram images ; Based on the anatomical sections shown in the echocardiogram, this is the preprocessed echocardiogram image. Set view category labels Quality labels are assigned based on the completeness of key anatomical structures, the clarity of their boundaries, and the presence of artifacts. ; The quality label is either a quality grade or a quality score; Preprocessed echocardiogram images With the corresponding view category label and quality label Perform matching to form training samples The training dataset consists of multiple training samples.

3. The echocardiographic quality assessment method based on nonlinear causal geometric learning according to claim 1, characterized in that, In step 2, the processing flow of the feature extraction network is as follows: The feature extraction network first receives preprocessed echocardiogram images. and extract from A visual factor set composed of visual factor feature maps , For the first Visual factor feature map; in The dimensions of the visual factor feature maps are defined as their height and width. Then, the visual factor feature maps are stacked along the channel direction to obtain the visual factor tensor. .

4. The echocardiography quality assessment method based on nonlinear causal geometric learning according to claim 3, characterized in that, In step 2, the processing flow of the causal geometry learning module is as follows: For visual factor set Feature maps of any two visual factors and First, calculate the feature map of visual factors. Pointing to visual factor feature map Directed conditional dependency : ； in, This represents element-wise multiplication at corresponding spatial locations. This represents the stability constant to prevent the denominator from being zero. It is a generalized Lehmer mean nonlinear aggregation operator; Indicates the feature map of a given visual factor Visual factor feature map under the condition The strength of conditional dependence, i.e., by point to Directed conditional dependency; For the Visual Factor Feature Map The generalized Lehmer mean nonlinear aggregation operator is expressed as: ； in, Indicates the first The visual factor feature map in the first The response value at each spatial location, This represents the total number of spatial locations in the feature map of each visual factor. and These are learnable parameters; Calculate the directed conditional dependencies for each combination of visual factors and construct a causal interaction matrix. This enables the estimation of directed causal relationships; Due to directed conditional dependency and These represent two conditional dependencies in opposite directions, and therefore they are usually not equal; therefore, a causal interaction matrix is ​​used. Asymmetry, calculating the causal action vector : ； in Representing the causal interaction matrix The transpose of the matrix, This represents the Sigmoid function. A parameter representing the degree of sensitivity to asymmetric relationships. Represents a column vector consisting entirely of 1s. Indicates in-sample standardization; Causal action vector The first in The nth component represents the nth component. The relative dominance of visual factors in all directed causal relationships; Based on the causal action vector Visual factor tensor Channel-level modulation is performed to obtain causal enhancement features. : ； in, Indicates the causal modulation intensity parameter. This represents the hyperbolic tangent function.

5. The echocardiographic quality assessment method based on nonlinear causal geometric learning according to claim 4, characterized in that, In step 2, the process of constructing the task token sequence is as follows: First, the causal enhancement features obtained from the causal geometry learning module are... Partition and map to Each Patch embedding is configured, and the corresponding position embedding is added to each Patch embedding. Set the initial view token for view classification before the patch embedding sequence. and the initial quality token used for quality evaluation This constitutes a task token sequence. The formula is expressed as follows: ； in, Indicates the first A patch embedding, This indicates the corresponding position embedding. .

6. The echocardiography quality assessment method based on nonlinear causal geometric learning according to claim 5, characterized in that, In step 2, the task token sequence Input Transformer encoder; The Transformer encoder models the contextual relationships between different visual factors through a self-attention mechanism; After the Transformer encoder outputs, the view relationship features corresponding to the view token are extracted. and quality token The set of relationship enhancement patch tokens is denoted as { }, ; View relationship features The input view classification header is mapped by a multilayer perceptron to obtain the predicted view category. ; Relationship Enhancement Patch Token Collection { The input nonlinear relationship aggregation module generates quality relationship features through nonlinear mapping, relationship importance weighting, and residual aggregation. , Input the quality prediction header to obtain the quality prediction results. .

7. The echocardiography quality assessment method based on nonlinear causal geometric learning according to claim 6, characterized in that, In step 2, the processing flow of the nonlinear relationship aggregation module is as follows: First, use the quality token. With relationship enhancement Patch token set { Perform basic relation aggregation to obtain basic quality relation features. For the first A relationship enhancement patch token , and its first Each channel component is denoted as ; The nonlinear relationship aggregation module performs dimension-wise function decomposition on each channel component and models the nonlinear relationships in different channels using multiple learnable piecewise spline basis functions, as expressed in the following formula: ； in For the first The nonlinear relation response is obtained by performing a dimension-wise nonlinear mapping on a relation enhancement patch token; The channel dimension of the relation-enhancing Patchtoken; This represents the number of piecewise spline basis functions corresponding to each channel. Indicates the action on the first The first channel A learnable piecewise spline basis function This represents the corresponding learnable combination weights; After obtaining the non-linear relation response corresponding to each relation enhancement patch token, the importance weight of each relation enhancement patch token is calculated. Normalize all importance weights to satisfy: and ; Subsequently, the non-linear relational responses of each relation-enhancing patch token are... According to the corresponding importance weight Weighting is performed, and the relationship between residuals and basic quality characteristics is analyzed. By fusing the data, we obtain the aggregated quality relationship characteristics. : ;in, This represents the learnable residual scaling factor.

8. The echocardiography quality assessment method based on nonlinear causal geometric learning according to claim 7, characterized in that, In step 2, the processing flow of the nonlinear quality prediction head is as follows: The nonlinear quality prediction head utilizes a set of learnable piecewise univariate functions to express quality relationship characteristics. Mapping to the quality prediction space, we obtain the quality prediction results, expressed by the following formula: ； in This represents the final predicted echocardiogram quality value. This indicates the number of piecewise univariate functions in the quality prediction header. Indicates the first A learnable piecewise single-variable function This represents the corresponding learnable combination weights.

9. The echocardiographic quality assessment method based on nonlinear causal geometric learning according to claim 8, characterized in that, In step 3, the total training loss is constructed during the training of the echocardiography quality assessment model. ; ； in, Represents geometric topology loss. For view classification loss, For quality loss; , and These represent the weighting coefficients for geometric topology loss, view classification loss, and quality loss, respectively. During training, backpropagation is performed based on the total training loss to jointly update the parameters involved in the training of the echocardiography quality assessment model until the echocardiography quality assessment model reaches the preset number of training rounds or the loss converges. After the echocardiography input model is trained, the echocardiogram to be evaluated is input into the trained echocardiogram input model to obtain the corresponding view category and output the corresponding quality score or quality level.

10. The echocardiographic quality assessment method based on nonlinear causal geometric learning according to claim 9, characterized in that, In step 3, the construction process of the geometric topology loss is as follows: For the Spatial aggregation and feature mapping of visual factors yield Euclidean embeddings. Then, it is projected onto the latent hyperbolic space through the exponential mapping at the origin, as expressed by the following formula: ;in, Indicates the first Geometric embedding of visual elements; Indicates curvature parameter as The exponential mapping of the latent hyperbolic space at the origin; for the causal interaction matrix Perform singular value decomposition: ； in, and Let these represent the left singular vector matrix and the right singular vector matrix, respectively. Represents the sum of the singular values The vector is composed of singular values; then, a weighted power series expansion is performed on each singular value. Specifically, for the first A singular value its expanded structural spectral components Represented as: ； in, Indicates the relation order index. These correspond to second-order, fourth-order, and even-order relations, respectively. This represents the highest order of the weighted power series expansion. A scale parameter representing the degree of contribution controlled by different order relationships; The structure spectrum is composed of all structure spectrum components. ; The intensity of multi-hop structural interactions between different visual factors is represented as follows: ； in Indicates the first The visual factor and the first The intensity of multi-hop structural interactions between visual factors; Construct a geometric topological loss based on the interaction strength of multi-hop structures: ;in, Indicates the relationship with the first Visual factors that dominate causal relationships Indicates the relationship with the first The causal relationship between visual factors is suppressed.

Images