An echocardiogram standard section classification method and device, electronic equipment and storage medium

Abstract

The application relates to the technical field of medical image data processing, and discloses an echocardiogram standard section classification method and device, electronic equipment and a storage medium, which comprises the following steps: acquiring echocardiogram medical image data to be classified; inputting the image data into a self-adaptive structure enhancement module, performing residual structure enhancement by using a channel re-labeling branch, a convolution mixed branch, a multi-scale branch and a self-adaptive router, and obtaining enhanced input features; inputting the features into a multi-level pyramid feature extraction network to extract second and third pyramid level features and a network end main feature; inputting the network end main feature into a main classification branch to output a main classification score, inputting the second and third pyramid level features into a double auxiliary fusion mapping module to output a second auxiliary classification score and a third auxiliary classification score; generating normalized fusion weights through learnable fusion parameters, and performing weighted summation on the corresponding scores to output a final fusion score and determine a standard section category.

Description

Technical Field

[0001] This invention relates to a method, apparatus, electronic device, and storage medium for classifying standard sections of echocardiography, belonging to the field of medical image data processing technology. Background Technology

[0002] Currently, using network architecture to extract features in a hierarchical manner to assist in cross-section recognition is the mainstream technical approach. By expanding the receptive field step by step through multi-layer convolution and pooling, the original pixels are converted into deep semantic features. Spatial downsampling is used to filter local clutter in order to focus on the anatomical tissue contour. This transformation process essentially relies on semantic abstraction to replace shallow spatial details, sacrificing local geometric cues in exchange for category translation invariance.

[0003] However, ultrasound images generally suffer from speckle noise and blurred boundaries due to the acoustic imaging principle. In practical applications, different standard sections of echocardiograms have high structural similarity, and key discrimination relies on weak local spatial geometric cues. Existing pathways are rigidly designed to rely solely on semantic features at the network's end to make decisions, implicitly accepting the implicit cost of the gradual dilution of intermediate layer representation information and its complete discarding during the inference stage. This results in insufficient network perception of weak boundary differences, making it prone to classification confusion due to the loss of structural information in complex situations. Moreover, there are also shortcomings in temporal control methods and dynamic data processing. For example, publication number C Chinese invention patent application N121962763A discloses an echocardiogram section classification method based on dynamic video. It attempts to model the temporal correlation of multi-frame spatial feature sequences through spatiotemporal data augmentation and the ViewFormer module. However, before feature aggregation, this scheme relies solely on the high-level semantic output at the end of the final stage of the spatial feature encoder. When processing high dynamic video streams, the local fine-grained geometric cues and anatomical boundary features contained in the intermediate layers have been diluted or discarded in the spatial domain encoding stage. This results in a lack of underlying spatial details to support the temporal control decision. When faced with similar sections with strong speckle noise interference, it is still easy to cause classification confusion.

[0004] Therefore, the technical problem to be solved by this invention is how to maintain the feature generalization ability while constructing a nonlinear spatial clue completion and hierarchical representation collaborative discrimination mechanism in the front-end topology and decision layer, so as to achieve adaptive fusion of deep semantics and intermediate local fine-grained features, thereby solving the problem of confusion in the classification of similar sections caused by speckle noise interference and loss of shallow structural information in ultrasound images. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A method for classifying standard sections of echocardiography, comprising the following steps:

[0006] Step S1: Obtain the echocardiographic medical image data to be classified;

[0007] Step S2: Input the echocardiogram medical image data into the adaptive structure enhancement module. The adaptive structure enhancement module includes a channel recalibration branch, a convolutional hybrid branch, a multi-scale branch, and an adaptive router. Each branch generates a corresponding enhancement response. The adaptive router generates branch fusion weights, which are then written back to the original input after residual scaling to obtain the enhanced input features. The enhanced input features are then input into a multi-level pyramid feature extraction network. The resolution is reduced and the number of channels is increased through four pyramid levels. The features of the second pyramid level with a dimension of 128, the features of the third pyramid level with a dimension of 320, and the main features of the network terminal with a dimension of 512 are explicitly extracted.

[0008] Step S3: Input the main features at the end of the network into the main classification branch to output the high-level semantic classification score; input the second pyramid level features and the third pyramid level features into the dual-auxiliary fusion mapping module, project them onto the category classification score space through global average pooling and fully connected layer linear mapping, and output the second auxiliary classification score and the third auxiliary classification score.

[0009] Step S4: Normalized primary fusion weight, normalized secondary fusion weight, and normalized tertiary fusion weight are generated through learnable fusion parameters. These weights are then multiplied by the high-level semantic classification score, the secondary fusion score, and the tertiary fusion score, respectively, and summed. The final fusion score is calculated and output within the category classification score space to determine the standard section category of the echocardiogram medical image data.

[0010] Preferably, when training the multi-level pyramid feature extraction network, the network parameters of the multi-level pyramid feature extraction network are optimized by performing multi-classification supervision constraint calculation on the high-level semantic classification score, the second auxiliary classification score, the third auxiliary classification score, and the final fusion score through the joint loss function; the joint loss function is composed of the sum of the main classification cross-entropy loss term, the second auxiliary cross-entropy loss term, the third auxiliary cross-entropy loss term, and the fusion branch cross-entropy loss term.

[0011] Preferably, step S2, which reduces resolution and increases the number of channels through four pyramid levels, includes: Step S21, the first pyramid level receives echocardiogram medical image data, performs downsampling processing and feature channel adjustment, and outputs first-level features; Step S22, the second pyramid level receives the first-level features, performs downsampling processing and feature channel adjustment, and outputs second-level features; Step S23, the third pyramid level receives the second-level features, performs downsampling processing and feature channel adjustment, and outputs third-level features; Step S24, the fourth pyramid level receives the third-level features, performs downsampling processing and feature channel adjustment, and outputs the main features at the network end.

[0012] Preferably, step S3, which inputs the second pyramid-level features and the third pyramid-level features into the dual-aided fusion mapping module, and projects them onto the category classification score space through global average pooling and fully connected layer linear mapping to output the second and third auxiliary classification scores, includes: Step S31, performing global average pooling on the second pyramid-level features through the first auxiliary mapping unit in the dual-aided fusion mapping module, and linearly mapping them to the category classification score space through the first fully connected layer to output the second auxiliary classification score; Step S32, performing global average pooling on the third pyramid-level features through the second auxiliary mapping unit in the dual-aided fusion mapping module. The local average pooling process is linearly mapped to the category classification score space through the second fully connected layer, and the third auxiliary classification score is output. The process of obtaining the echocardiographic medical image data to be classified in step S1 includes: receiving the original ultrasound image and performing structural enhancement processing on the original ultrasound image through an adaptive structural enhancement module. The adaptive structural enhancement module includes a channel recalibration branch, a convolutional hybrid branch, a multi-scale branch, and an adaptive router. Each branch generates a corresponding enhancement response, the adaptive router generates branch fusion weights, and writes them back to the original input after residual scaling, generating structurally enhanced image features as the echocardiographic medical image data to be classified.

[0013] Preferably, in steps S21, S22, S23 and S24, the downsampling process in each pyramid level is configured with a fixed spatial downsampling ratio of 2, so that the spatial resolution of the feature decreases by 1 / 2 along each pyramid level, and the number of feature channels increases along each pyramid level according to the preset level configuration.

[0014] Preferably, when calculating the joint loss function, corresponding weight coefficients are configured for the main classification cross-entropy loss term, the second auxiliary cross-entropy loss term, the third auxiliary cross-entropy loss term, and the fusion branch cross-entropy loss term, respectively. The weight coefficient of the fusion branch cross-entropy loss term is greater than the weight coefficient of the second auxiliary cross-entropy loss term and is also greater than the weight coefficient of the third auxiliary cross-entropy loss term.

[0015] Preferably, the echocardiographic medical image data to be classified is imported from the ex vivo ultrasound storage medium into the data calculation module and transformed in the background. The final output of the fusion score is a standard cross-sectional category numerical index as a neutral label. In the inference processing stage and the network training stage, the second auxiliary classification score and the third auxiliary classification score participate in the weighted summation calculation together with the high-level semantic classification score in the category classification score space, so as to maintain the consistency of the feature interaction topology between inference processing and network training.

[0016] An echocardiographic standard section classification device is provided to implement an echocardiographic standard section classification method.

[0017] An electronic device for standard echocardiographic section classification includes: a memory for storing computer program instructions; and a processor for executing the computer program instructions to implement the steps of the method.

[0018] A storage medium for standard echocardiographic plane classification, the storage medium storing a computer program that, when executed by a processor, implements the steps of the method described.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. In standard section classification of echocardiography, the input image is fed into a parallel enhancement topology consisting of a channel recalibration branch, a convolutional hybrid branch, and a multi-scale branch. An adaptive router is used to extract a global description of the input features to dynamically generate fusion weights. This allows channel responses, spatial neighborhood interactions, and multi-scale structural cues to be fused together based on the current image quality fluctuations. The fusion residuals are written back to the original input using a learnable residual scaling factor. This dynamically completes blurred boundaries and local textures while maintaining the stability of the main structure of the echocardiogram, effectively avoiding the damage to shallow representation information caused by the sudden drop in spatial resolution during the hierarchical encoding stage of traditional feature extraction networks.

[0021] 2. The feature extraction stage adopts a four-stage pyramid architecture. While gradually reducing the spatial resolution of the feature map, the number of feature channels is increased to construct hierarchical visual representations. The second-stage features and the third-stage features are explicitly retained as the input basis for the back-end dual-auxiliary fusion head. This allows the rich local structure and spatial details contained in the second-stage features and the mid-level discriminative features in the third-stage features that balance spatial compression and semantic expression to be fully inherited. This overcomes the defect of traditional feature classification networks that lose local contour structure and spatial neighborhood geometric clues due to the blurring and stripping of anatomical boundaries when making decisions by relying solely on the high-level semantics at the end of the backbone.

[0022] 3. The backend dual-auxiliary fusion head uses global average pooling and linear mapping to independently map the second-stage features and third-stage features to the category space, generating auxiliary classification outputs in the category classification score space. These outputs, along with the main classification branch outputs, are weighted and recombined through a set of normalized learnable fusion parameters. This allows mid-level discriminative evidence rich in spatial organization to directly participate in the final classification decision during the inference stage. This adaptive interaction method of the score space can organically integrate deep high-level semantics with mid-level local fine-grained features, eliminating classification confusion caused by the lack of obvious global appearance differences between similar cross-sections in multi-level collaborative voting. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the classification of echocardiographic sections according to the present invention.

[0024] Figure 2 This is a structural diagram of the echocardiography classification device of the present invention;

[0025] Figure 3 This is a schematic diagram of the overall algorithm topology framework of the multi-level classification system in this invention;

[0026] Figure 4 This is a schematic diagram of the microstructure topology of the adaptive structure enhancement module in this invention;

[0027] Figure 5 This is a schematic diagram of the microstructure topology of the dual auxiliary fusion head at the back end in this invention.

[0028] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0030] A standard echocardiographic section classification method establishes an overall processing path at the macroscopic topology level, encompassing input enhancement, hierarchical encoding, and fusion discrimination. Let the input echocardiographic image be denoted as X, and the front-end adaptive structural enhancement process be denoted as f. pre (X), the feature encoding process based on the pyramid backbone network is denoted as f. back (X), the backend dual-aid fusion discrimination process is denoted as f. rear (X), then the overall classification mapping process is explicitly expressed by the following mathematical relation: y=f rear (f back (f pre(X))), the method includes the following steps:

[0031] Step S1: Obtain the echocardiographic medical image data to be classified;

[0032] Step S2: Input the echocardiogram medical image data into the adaptive structure enhancement module. The adaptive structure enhancement module includes a channel recalibration branch, a convolutional hybrid branch, a multi-scale branch, and an adaptive router. Each branch generates a corresponding enhancement response. The adaptive router generates branch fusion weights, which are then written back to the original input after residual scaling to obtain the enhanced input features. The enhanced input features are then input into a multi-level pyramid feature extraction network. The resolution is reduced and the number of channels is increased through four pyramid levels. The features of the second pyramid level with a dimension of 128, the features of the third pyramid level with a dimension of 320, and the main features of the network terminal with a dimension of 512 are explicitly extracted.

[0033] Step S3: Input the main features at the end of the network into the main classification branch to output the high-level semantic classification score; input the second pyramid level features and the third pyramid level features into the dual-auxiliary fusion mapping module, project them onto the category classification score space through global average pooling and fully connected layer linear mapping, and output the second auxiliary classification score and the third auxiliary classification score.

[0034] Step S4: Normalized primary fusion weight, normalized secondary fusion weight, and normalized tertiary fusion weight are generated through learnable fusion parameters. These weights are then multiplied by the high-level semantic classification score, the secondary fusion score, and the tertiary fusion score, respectively, and summed. The final fusion score is calculated and output within the category classification score space to determine the standard section category of the echocardiogram medical image data.

[0035] Preferably, when training the multi-level pyramid feature extraction network, the network parameters of the multi-level pyramid feature extraction network are optimized by performing multi-classification supervision constraint calculation on the high-level semantic classification score, the second auxiliary classification score, the third auxiliary classification score, and the final fusion score through the joint loss function; the joint loss function is composed of the sum of the main classification cross-entropy loss term, the second auxiliary cross-entropy loss term, the third auxiliary cross-entropy loss term, and the fusion branch cross-entropy loss term.

[0036] Preferably, step S2, which reduces resolution and increases the number of channels through four pyramid levels, includes: Step S21, the first pyramid level receives echocardiogram medical image data, performs downsampling processing and feature channel adjustment, and outputs first-level features; Step S22, the second pyramid level receives the first-level features, performs downsampling processing and feature channel adjustment, and outputs second-level features; Step S23, the third pyramid level receives the second-level features, performs downsampling processing and feature channel adjustment, and outputs third-level features; Step S24, the fourth pyramid level receives the third-level features, performs downsampling processing and feature channel adjustment, and outputs the main features at the network end.

[0037] Preferably, step S3, which inputs the second pyramid-level features and the third pyramid-level features into the dual-aided fusion mapping module, and projects them onto the category classification score space through global average pooling and fully connected layer linear mapping to output the second and third auxiliary classification scores, includes: Step S31, performing global average pooling on the second pyramid-level features through the first auxiliary mapping unit in the dual-aided fusion mapping module, and linearly mapping them to the category classification score space through the first fully connected layer to output the second auxiliary classification score; Step S32, performing global average pooling on the third pyramid-level features through the second auxiliary mapping unit in the dual-aided fusion mapping module. The local average pooling process is linearly mapped to the category classification score space through the second fully connected layer, and the third auxiliary classification score is output. The process of obtaining the echocardiographic medical image data to be classified in step S1 includes: receiving the original ultrasound image and performing structural enhancement processing on the original ultrasound image through an adaptive structural enhancement module. The adaptive structural enhancement module includes a channel recalibration branch, a convolutional hybrid branch, a multi-scale branch, and an adaptive router. Each branch generates a corresponding enhancement response, the adaptive router generates branch fusion weights, and writes them back to the original input after residual scaling, generating structurally enhanced image features as the echocardiographic medical image data to be classified.

[0038] Preferably, in steps S21, S22, S23 and S24, the downsampling process in each pyramid level is configured with a fixed spatial downsampling ratio of 2, so that the spatial resolution of the feature decreases by 1 / 2 along each pyramid level, and the number of feature channels increases along each pyramid level according to the preset level configuration.

[0039] Preferably, when calculating the joint loss function, corresponding weight coefficients are configured for the main classification cross-entropy loss term, the second auxiliary cross-entropy loss term, the third auxiliary cross-entropy loss term, and the fusion branch cross-entropy loss term, respectively. The weight coefficient of the fusion branch cross-entropy loss term is greater than the weight coefficient of the second auxiliary cross-entropy loss term and is also greater than the weight coefficient of the third auxiliary cross-entropy loss term.

[0040] Preferably, the echocardiographic medical image data to be classified is imported from the ex vivo ultrasound storage medium into the data calculation module and transformed in the background. The final output of the fusion score is a standard cross-sectional category numerical index as a neutral label. In the inference processing stage and the network training stage, the second auxiliary classification score and the third auxiliary classification score participate in the weighted summation calculation together with the high-level semantic classification score in the category classification score space, so as to maintain the consistency of the feature interaction topology between inference processing and network training.

[0041] An echocardiographic standard section classification device is provided to implement an echocardiographic standard section classification method.

[0042] An electronic device for standard echocardiographic section classification includes: a memory for storing computer program instructions; and a processor for executing the computer program instructions to implement the steps of any one of methods 1 to 7.

[0043] A storage medium for the classification of standard echocardiographic sections, the storage medium storing a computer program, which, when executed by a processor, implements the steps of methods 1-7.

[0044] Example 1: A current standard echocardiographic section classification method is applied to an echocardiographic imaging workstation. It improves section recognition accuracy through multi-level feature fusion. The specific process is as follows: After image data acquisition, it first enters the adaptive structure enhancement module. This module consists of a channel recalibration branch, a convolutional hybrid branch, a multi-scale branch, and an adaptive router. The adaptive router dynamically generates fusion weights for each branch based on the global description of the input features. The outputs of the three branches are weighted and superimposed according to the proportions generated by the adaptive router. Finally, the enhanced residual features are written back to the original input image using a residual scaling factor. Its mathematical flow and residual write-back mechanism are explicitly represented by the following formula: and ,in, Indicates input features, The set parameters represent the enhanced input features. The feature indices of each independent branch in the parallel enhancement topology are explicitly defined, where Indicates a channel recalibration branch. Indicates the convolutional hybrid branch, This represents multi-scale branches. Through the precise constraint of this set, a one-to-one mapping and traversal relationship is established between mathematical weighted summation terms and specific physical branch structures. Indicates the first Nonlinear transformation function with enhanced branches, The fusion weights are dynamically generated by the adaptive router based on the input, and the sum of the weights of each branch is equal to 1. Indicates global average pooling. This represents a lightweight routing mapping function that uniformly adopts a mapping matrix topology consistent with the formula, used to project the global description vector obtained by global average pooling onto a low-dimensional weight activation space corresponding to the number of branches. The learnable residual scaling factor is then used. Subsequently, the enhanced input features are fed into a multi-level pyramid feature extraction network as a highly preferred instantiated skeleton. The PVTv2-B2 network model is adopted, which adopts a four-stage pyramid structure. The channel configuration of the four stages is explicitly specified as [64, 128, 320, 512]. The second pyramid level outputs the second pyramid level feature F2 with a dimension of 128; the third pyramid level outputs the third pyramid level feature F3 with a dimension of 320; and the fourth pyramid level outputs the network terminal main feature with a dimension of 512. Feature F2 retains relatively more local structure and spatial details, while feature F3 has a higher level of discriminative information between spatial compression and semantic representation. This invention explicitly retains and extracts F2 and F3 as the input basis for the back-end dual-auxiliary fusion head.

[0045] In the classification and auxiliary fusion stage, the main features at the end of the network are input into the main classification branch, and after passing through a fully connected layer, a high-level semantic classification score is output. Let z be the high-level semantic classification score output by the main classification branch. main ∈R K Where K is the number of categories, in this embodiment K=4. The standard cross-section categories include the apical five-chamber view, the parasternal aortic short-axis view, the parasternal long-axis view, and the subxiphoid right ventricular outflow tract view. Simultaneously, the second pyramid level features and the third pyramid level features are input into the dual-auxiliary fusion mapping module. The first auxiliary mapping unit in the dual-auxiliary fusion mapping module performs global average pooling on the second pyramid level features, linearly maps them to the category classification score space through the first fully connected layer, and outputs the second auxiliary classification score. The second auxiliary mapping unit performs global average pooling on the features of the third pyramid level, and then linearly maps them to the category classification score space through the second fully connected layer, outputting the third auxiliary classification score. .

[0046] In the final prediction stage, let the three sets of learnable fusion parameters be as follows: , and Then the normalized fusion weights It is generated by the following formula: , satisfying ρ main +ρ aux2 +ρ aux3 =1, and based on this, the scores are weighted and summed within the category classification score space to calculate and output the final fusion score z.f The process satisfies the following formula: This fusion occurs in the category classification score space, rather than the feature layer. Therefore, the intermediate auxiliary branch is not only used for intermediate layer supervision during the training phase, but also directly participates in the final classification decision during the inference phase. The features of the second and third pyramid levels are directly received through the dual auxiliary fusion mapping module, so that these local anatomical contours that have not yet been deeply diluted can be directly fed back to the final prediction layer in the form of classification scores, thereby achieving the collaborative discrimination of semantic generalization and structural details at the logical level.

[0047] Example 2: This experiment verifies the effectiveness of the sectional classification scheme disclosed in this invention by constructing an echocardiographic image analysis verification platform. The experimental platform is built based on a general-purpose high-performance computing unit, configured with a graphics processor with 64GB of video memory and a memory redundancy environment that meets the requirements of parallel computing. To simulate clinical image acquisition conditions, the experiment uses 5000 cases of two-dimensional echocardiographic image data collected from different clinical databases as test samples, including standard sectional images as well as low-quality images caused by differences in patient position, respiratory motion interference, and improper gain settings. The experimental group uses the multi-level pyramid-shaped feature extraction network and dual-auxiliary fusion mapping module of this invention, while the control group uses a baseline network model containing only single-core feature extraction and no auxiliary fusion mechanism. The network input image size is standardized to 224×224 pixels, and the number of feature channels at the pyramid level is configured to be 64, 128, 320, and 512, ensuring that the network maintains spatial resolution in the shallow layers while achieving effective compression of semantic information in the deep layers. During the network training phase, the main classification branch, the two auxiliary classification branches, and the fusion branch jointly receive classification supervision. The refined mathematical expression of the total joint loss function L is shown below: ,in, Represents cross-entropy loss, For real category labels, , and Configure the weight coefficients for the second auxiliary cross-entropy loss term, the third auxiliary cross-entropy loss term, and the cross-entropy loss term of the fusion branch, in addition to the primary classification cross-entropy loss term. Also configure the weight coefficients for the fusion branch. Greater than the second auxiliary weight coefficient And greater than the third auxiliary weight coefficient This joint supervision method enables the main branch, intermediate auxiliary branches, and fused output to simultaneously optimize towards the correct classification direction. During the inference phase, the final fused score is directly used. As a classification criterion, the weight generation network acts as a dynamic regulator. Based on the adaptive update mode of the cross-entropy loss function feedback, it dynamically allocates fusion weights according to the scores of each branch, ensuring that the output contribution of the main classification branch and the two auxiliary branches is automatically adjusted under different signal-to-noise ratio conditions.

[0048] The fusion weights are dynamically allocated based on the scores of each branch to ensure automatic adjustment of the output contribution of the main classification branch and the two auxiliary branches under different signal-to-noise ratio conditions. Specifically, the contribution refers to the high-level semantic classification score output by the main classification branch in the category classification score space. The second auxiliary classification score output by the first auxiliary mapping unit and the third auxiliary classification score output by the second auxiliary mapping unit. The final fusion score is calculated by weighted summation using an adaptive update mode. At that time, the normalized fusion weight ratio of each is... , and The dynamic adjustment mechanism of the decision influence on the final standard section category digital index output is as follows: when the input ultrasound image presents a low signal-to-noise ratio due to differences in patient positioning or improper gain settings, the high-level semantic features of the main classification branch are easily blurred by speckle noise due to spatial resolution compression, leading to a decrease in its classification score. The confidence level decreases; at this point, the system passes the total joint loss function. The adaptive gradient backpropagation feedback triggers the weight generation network to act as a dynamic regulator, adaptively increasing the second auxiliary fusion weights rich in fine-grained local anatomical contours and spatial geometric cues. And the third auxiliary fusion weight This involves explicitly increasing the output contribution of the two auxiliary branches at the scoring decision level to compensate for the damage or dilution of deep semantic information, ensuring that the system can still output a stable and accurate final fused score under complex operating conditions. .

[0049] Experimental observations revealed that when speckle noise or artifacts interfered with ultrasound images, the classification accuracy of the control group decreased from 92.4% under the baseline condition to 81.2%, showing significant performance degradation. The experimental group, by using a structural enhancement module to repair and complete low-quality features, maintained the classification accuracy at 89.8% under this condition. Data comparison showed that when the proportion of the third auxiliary classification score to the final fusion weight was between 0.2 and 0.4, the system exhibited the least fluctuation in classification scores for difficult samples. The determination of this weight proportion range was based on gradient stability testing during network training. In the optimization phase of the clinical trial dataset, experiments using a weight scan with a step size of 0.05 revealed that if the weight was below 0.2, the contribution of the intermediate-scale geometric information carried by the third-level features to the loss function was insufficient. Insufficient weighting makes it difficult for the network to distinguish geometrically similar sections such as the long and short axes of the left ventricle when processing low signal-to-noise ratio images. If the weight is higher than 0.4, the third-level features still contain a lot of unfiltered speckle background noise, and excessive weighting will cause the classification results to be overly sensitive to image noise, leading to result jumps. Experimental data shows that the mutual information between classification branches reaches the optimal balance in the range of 0.2 to 0.4, which enables the joint loss function to maintain a fast and stable convergence state under image streams from different devices. If the weight of this branch is lower than 0.1, the system's ability to distinguish standard section edge information decreases due to the lack of intermediate layer feature supplementation, and the test accuracy decreases by about 3.5%. If the weight exceeds 0.6, the high-level semantic expression is interfered with by local noise, resulting in nonlinear degradation of the overall classification stability.

[0050] Experimental results confirm that by dynamically adjusting the weights in the classification score space through the dual-auxiliary fusion mapping module, the system effectively balances the structural details of shallow features with the semantic generalization of deep features. This data proves that the network architecture improves the stability of echocardiogram section recognition without increasing the additional inference computing hardware overhead by organically combining high-level classification responses and intermediate-level classification responses. Furthermore, the disclosed weight allocation interval has verifiable engineering implementation boundaries.

[0051] Example 3: The core technical solution of this invention revolves around the adaptive structure enhancement module, the multi-level pyramid-shaped feature extraction network, and the back-end dual-auxiliary fusion head; the content related to abnormal data injection early warning, network transmission monitoring, and back-end control strategies is not part of the core technical content of the original classification network solution of this invention, and therefore will not be elaborated in this example.

[0052] Example 4: The content related to operational status calibration, temperature drift compensation, and hardware load correction is not part of the core technology of the original classification network scheme of this invention. The focus of this invention is on the adaptive structure enhancement module, the hierarchical feature preservation based on the multi-level pyramid feature extraction network, and the back-end dual-auxiliary fusion mechanism. Therefore, this example will not elaborate on the device operational status calibration and offset compensation methods.

[0053] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for classifying standard sections of echocardiography, characterized in that, Includes the following steps: Step S1: Obtain the echocardiographic medical image data to be classified; Step S2: Input the echocardiogram medical image data into the adaptive structure enhancement module. The adaptive structure enhancement module includes a channel recalibration branch, a convolutional hybrid branch, a multi-scale branch, and an adaptive router. Each branch generates a corresponding enhancement response. The adaptive router generates branch fusion weights and writes them back to the original input after residual scaling to obtain the enhanced input features. The enhanced input features are then fed into a multi-level pyramid feature extraction network. The resolution is reduced and the number of channels is increased through four pyramid levels. The features of the second pyramid level with a dimension of 128, the features of the third pyramid level with a dimension of 320, and the main features at the end of the network with a dimension of 512 are explicitly extracted. Step S3: Input the main features at the end of the network into the main classification branch to output the high-level semantic classification score; The second and third pyramid level features are input into the dual-aid fusion mapping module, and then projected onto the category classification score space through global average pooling and fully connected layer linear mapping, and output the second and third auxiliary classification scores. Step S4: Normalized primary fusion weight, normalized secondary fusion weight, and normalized tertiary fusion weight are generated through learnable fusion parameters. These weights are then multiplied by the high-level semantic classification score, the secondary fusion score, and the tertiary fusion score, respectively, and summed. The final fusion score is calculated and output within the category classification score space to determine the standard section category of the echocardiogram medical image data.

2. The method for classifying standard sections of echocardiography according to claim 1, characterized in that, When training the multi-level pyramid feature extraction network, the network parameters of the multi-level pyramid feature extraction network are optimized by performing multi-class supervision constraint calculation on the high-level semantic classification score, the second auxiliary classification score, the third auxiliary classification score and the final fusion score through the joint loss function. The joint loss function is composed of the sum of the main classification cross-entropy loss term, the second auxiliary cross-entropy loss term, the third auxiliary cross-entropy loss term and the fusion branch cross-entropy loss term.

3. The method for classifying standard sections of echocardiography according to claim 1, characterized in that, Step S2 involves reducing resolution and increasing the number of channels through four pyramid levels: Step S21, the first pyramid level receives echocardiogram medical image data, performs downsampling processing and feature channel adjustment, and outputs first-level features; Step S22, the second pyramid level receives the first-level features, performs downsampling processing and feature channel adjustment, and outputs second-level features; Step S23, the third pyramid level receives the second-level features, performs downsampling processing and feature channel adjustment, and outputs third-level features; Step S24, the fourth pyramid level receives the third-level features, performs downsampling processing and feature channel adjustment, and outputs the main features at the network end.

4. The method for classifying standard sections of echocardiography according to claim 1, characterized in that, Step S3 involves inputting the second and third pyramid-level features into the dual-aided fusion mapping module. The features are then projected onto the category classification score space through global average pooling and linear mapping via a fully connected layer, outputting the second and third auxiliary classification scores. This includes: Step S31, performing global average pooling on the second pyramid-level features using the first auxiliary mapping unit in the dual-aided fusion mapping module, and linearly mapping them to the category classification score space via a first fully connected layer to output the second auxiliary classification score; Step S32, performing global average pooling on the third pyramid-level features using the second auxiliary mapping unit in the dual-aided fusion mapping module. The process involves equal pooling, linear mapping to the category classification score space via a second fully connected layer, and outputting a third auxiliary classification score. Step S1 involves obtaining the echocardiographic medical image data to be classified, including: receiving the original ultrasound image and performing structural enhancement processing on the original ultrasound image through an adaptive structural enhancement module. The adaptive structural enhancement module includes a channel recalibration branch, a convolutional hybrid branch, a multi-scale branch, and an adaptive router. Each branch generates a corresponding enhancement response, the adaptive router generates branch fusion weights, and writes them back to the original input after residual scaling, generating structurally enhanced image features as the echocardiographic medical image data to be classified.

5. The method for classifying standard sections of echocardiography according to claim 3, characterized in that, In steps S21, S22, S23, and S24, the downsampling process in each pyramid level is configured with a fixed spatial downsampling ratio of 2, so that the spatial resolution of the feature decreases by 1 / 2 layer by layer along each pyramid level, and the number of feature channels increases layer by layer along each pyramid level according to the preset level configuration.

6. The method for classifying standard sections of echocardiography according to claim 2, characterized in that, When calculating the joint loss function, corresponding weight coefficients are configured for the main classification cross-entropy loss term, the second auxiliary cross-entropy loss term, the third auxiliary cross-entropy loss term, and the fusion branch cross-entropy loss term. The weight coefficient of the fusion branch cross-entropy loss term is greater than the weight coefficient of the second auxiliary cross-entropy loss term, and is also greater than the weight coefficient of the third auxiliary cross-entropy loss term.

7. The method for classifying standard sections of echocardiography according to claim 1, characterized in that, The echocardiographic medical image data to be classified is imported from the ex vivo ultrasound storage medium into the data calculation module and transformed in the background. The final output of the fusion score is a standard cross-sectional category numerical index as a neutral label. In the inference processing stage and the network training stage, the second auxiliary classification score and the third auxiliary classification score participate in the weighted summation calculation together with the high-level semantic classification score in the category classification score space to maintain the consistency of the feature interaction topology between inference processing and network training.

8. A standard section classification device for echocardiography, characterized in that, The echocardiography standard section classification device is used to implement the echocardiography standard section classification method as described in claim 1.

9. An electronic device for classifying standard sections of echocardiography, characterized in that, include: Memory is used to store computer program instructions; A processor for executing computer program instructions to implement the steps of the method as described in claim 1.

10. A storage medium for the classification of standard echocardiographic sections, characterized in that, A storage medium for standard echocardiographic section classification stores a computer program that, when executed by a processor, implements the steps of the method described in claim 1.

Images