An ultrasonic image lesion auxiliary screening method and system for a medical terminal

Abstract

The present application relates to the medical ultrasonic image processing technical field, propose a kind of medical terminal-oriented ultrasonic image focus auxiliary screening method and system, comprising: obtaining training sample set, pre-processing is obtained pre-processing multi-view ultrasonic image;Construct focus auxiliary screening model, input pre-processing multi-view ultrasonic image into light weight backbone network, extract multi-view differentiating feature, including longitudinal view feature and transverse view feature;Multi-view differentiating feature is input into implicit fusion module, constructs shared implicit state and is solved by fixed point, realizes the adaptive interaction and high-order fusion of multi-view differentiating feature, obtains global fusion feature;Global fusion feature is input into the view perception output head of view perception output module, obtains focus auxiliary screening result.The present application solves the problems that the modeling ability of dynamic complementary relationship between multi-views is insufficient in the prior art, the convergence is poor when depth balanced model is applied to ultrasonic focus screening, and the fusion stability and specificity discrimination are insufficient.

Description

Technical Field

[0001] This invention relates to the field of medical ultrasound image processing technology, and in particular to a method and system for assisting in the screening of lesions in ultrasound images for medical terminals. Background Technology

[0002] Early lesion identification is crucial for disease risk warning, and ultrasound imaging has become a key imaging tool for early clinical lesion screening and follow-up assessment. Currently, ultrasound imaging techniques for lesion screening mainly include single-view human experience-assisted analysis methods, deep learning methods based on explicit feature fusion, and multi-view learning methods employing attention mechanisms or multi-branch structures.

[0003] Single-view analysis methods typically rely on local texture, grayscale distribution, or empirical features of a single image for judgment, making it difficult to fully utilize complementary information between different views and prone to missed detections or misclassifications. Existing multi-view deep learning methods mostly employ explicit fusion methods such as feature stitching, static attention weighting, or fixed multi-branch aggregation. While these methods improve classification performance to some extent, their ability to model dynamic complementary relationships between different views is limited, making it difficult to fully characterize the high-order nonlinear correlations between multiple views in complex ultrasound scenes.

[0004] Implicit modeling methods such as deep equilibration models, which achieve deep feature representation through fixed-point solutions, are gaining increasing attention in the field of medical imaging. However, existing deep equilibration models are mostly applied to medical image reconstruction or inverse problem solving, and still face technical challenges in ultrasound lesion screening scenarios: ultrasound images generally suffer from strong artifacts, heavy noise, unclear boundaries, and significant heterogeneity between different views, which can easily lead to slow convergence, oscillation, or even divergence of fixed-point iterations; in addition, existing methods do not adequately consider the stability optimization of the implicit fusion process and the discrimination of view specificity, making it difficult to directly meet the comprehensive requirements of accuracy, robustness, and real-time performance for ultrasound-assisted lesion screening tasks. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides an ultrasound imaging-assisted lesion screening method and system for medical terminals, which solves the problems of insufficient modeling ability of dynamic complementary relationships between multiple views, poor convergence of depth equilibrium model when applied to ultrasound lesion screening, and insufficient fusion stability and specificity discrimination of the prior art.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for ultrasound-assisted lesion screening for medical terminals, comprising: A training sample set is obtained, which includes ultrasound images of longitudinal and transverse views, corresponding view labels, and lesion category labels. The ultrasound images are preprocessed to obtain preprocessed multi-view ultrasound images. A lesion-assisted screening model is constructed, comprising a lightweight backbone network, an implicit fusion module, and a view-aware output module connected in sequence; The preprocessed multi-view ultrasound images are input into a lightweight backbone network to extract multi-view differential features, including longitudinal and transverse view features. The multi-view differential features are input into the implicit fusion module to construct a shared hidden state and achieve adaptive interaction and high-order fusion of the multi-view differential features through fixed-point solving, thereby obtaining global fusion features; The global fusion features are input into the view perception output head of the view perception output module to obtain the lesion-assisted screening results.

[0007] As a further technical solution, the lightweight backbone network adopts a four-stage convolutional feature extraction structure, optimizes the configuration of the number of residual units, and the number of output channels in each stage is 32, 64, 128 and 64 respectively, and the output channels are compressed to 64 through 1×1 convolution; the activation function adopts LeakyReLU with a negative slope of 0.1.

[0008] As a further technical solution, during the training phase of the lesion-assisted screening model, the multi-view features extracted by the lightweight backbone network are mapped to the shared hidden state space, and the multi-view features are implicitly fused through a deep equalization model; the Anderson acceleration algorithm is introduced in the fixed-point solution process; and the model parameters are jointly optimized by combining a bi-term explicit regularization strategy and a multi-component loss function.

[0009] As a further technical solution, in the model output stage, a view perception output head containing longitudinal branches, transverse branches, and mixed branches is constructed; during training, view labels are used to jointly optimize different branches; during inference, the corresponding output head is dynamically selected to perform classification based on the view labels of the input ultrasound image; when the view attributes are unclear or there is inconsistency in the information of multiple views, the mixed branch is enabled to perform backtracking discrimination in order to obtain lesion-assisted screening results.

[0010] As a further technical solution, the implicit fusion process of the deep equalization model includes: concatenating multi-view features in the channel dimension and mapping them to a unified hidden state space through 1×1 convolution; the hidden state update mapping consists of 3×3 convolution, batch normalization, GELU activation function, view interaction attention mechanism and residual connection; the view interaction attention mechanism performs channel-level global average pooling on each view feature, concatenates the resulting channel description vectors and inputs them into a two-layer fully connected network, outputs channel weights through the Sigmoid function, and performs channel-wise weighted summation on each view feature based on the channel weights.

[0011] As a further technical solution, the bi-term explicit regularization strategy includes: a first regularization term is used to minimize the residual between the hidden state and the mapped output to smooth local behavior near the fixed point; and a second regularization term is used to constrain the magnitude of the hidden state to prevent gradient explosion or divergence of the iteration trajectory.

[0012] As a further technical solution, the multi-component loss function includes dynamically weighted binary cross-entropy loss and bi-term explicit regularization loss:

[0013] in, , , These represent the binary cross-entropy loss weights for long, short, and mixed headers, respectively. , , These represent the binary cross-entropy losses for long headers, short headers, and mixed headers, respectively. This represents the loss of explicit regularization for two terms.

[0014] A second aspect of the present invention provides an ultrasound imaging-assisted lesion screening system for medical terminals, comprising: The sample acquisition module is configured to: acquire a training sample set, which includes ultrasound images of longitudinal and transverse views, corresponding view labels and lesion category labels; and preprocess the ultrasound images to obtain preprocessed multi-view ultrasound images. The model building module is configured to build a lesion-assisted screening model, which includes a lightweight backbone network, an implicit fusion module, and a view-aware output module connected in sequence. The feature extraction module is configured to: input the preprocessed multi-view ultrasound images into a lightweight backbone network and extract multi-view differential features, including longitudinal view features and transverse view features; The feature fusion module is configured to: input the multi-view differential features into the implicit fusion module, construct a shared hidden state and achieve adaptive interaction and high-order fusion of the multi-view differential features through fixed-point solving, and obtain global fusion features; The result output module is configured to input the global fusion features into the view perception output head of the view perception output module to obtain the lesion-assisted screening result. A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of an ultrasound imaging-assisted lesion screening method for a medical terminal as described in the first aspect of the present invention.

[0015] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of an ultrasound imaging-assisted lesion screening method for a medical terminal as described in the first aspect of the present invention.

[0016] The above one or more technical solutions have the following beneficial effects: This invention constructs a lesion-assisted screening model that combines a lightweight backbone network with a deep, balanced implicit fusion module. This model achieves high-order interactions between features from multiple views without significantly increasing the number of parameters, overcoming the difficulty of existing explicit stitching or static attention fusion methods in fully characterizing the dynamic complementary relationships between multiple views and improving lesion identification capabilities in complex ultrasound scenarios. By introducing the Anderson acceleration algorithm during the DEQ fixed-point solution process and combining it with a bi-term explicit regularization strategy, the convergence efficiency and numerical stability of the implicit fusion process under high noise, strong artifacts, and complex echo backgrounds are significantly improved, thereby enhancing the model's training robustness and clinical application reliability. By constructing a view-aware output head containing longitudinal, transverse, and mixed branches, and dynamically selecting the output branch based on the view label during the inference phase, the lesion discrimination and adaptation capabilities under different view conditions are improved, and better compatibility is demonstrated for samples with blurred views or inconsistent information. Experimental results show that the proposed scheme achieves a recall rate of 96.30% for plaque samples and a macro-average F1 score of 94.29%. By adopting a lightweight network design and implicit fusion mechanism, the number of model parameters can be controlled at 3.31M, and the FLOPs are 1.15G. Under the experimental settings, the average inference time per image is 4.2ms and the FPS is 238. It has low computational complexity and high running efficiency, and is therefore more suitable for deployment and application in primary medical institutions, portable ultrasound scenarios and low computing power medical terminals.

[0017] The model of this invention has a high degree of consistency between the area of ​​focus and the core area of ​​the lesion, which can provide clinicians with a certain degree of visual auxiliary evidence, thereby improving the interpretability and clinical acceptability of the auxiliary screening results.

[0018] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0020] Figure 1 This is a flowchart of the overall process of the auxiliary screening method in the first embodiment; Figure 2 This is a schematic diagram of the lightweight backbone network structure of the first embodiment; Figure 3 This is a diagram of the DEQ implicit fusion module architecture of the first embodiment; Figure 4 A schematic diagram illustrating the stability and convergence speed of different DEQ configurations in the fixed-point solution process for the first embodiment; Figure 5 The image shows a cross-sectional view of the Grad-CAM attention heatmap on a typical positive sample, comparing the method of the first embodiment with the contrast model. (a) is a reference annotation map of the lesion area, (b) is the attention heatmap generated by the baseline model ResNet-18, and (c) is the attention heatmap generated by the method of this embodiment. Figure 6 The image shows a longitudinal section of the Grad-CAM attention heatmap on a typical positive sample, comparing the method of the first embodiment with the contrast model. (a) is a reference annotation map of the lesion area, (b) is the attention heatmap generated by the baseline model ResNet-18, and (c) is the attention heatmap generated by the method of this embodiment. Figure 7 The image shows a cross-sectional view of the Grad-CAM attention heatmap on a difficult positive sample, comparing the method of the first embodiment with the comparison model. (a) is a reference annotation map of the lesion area, (b) is the attention heatmap generated by the baseline model ResNet-18, and (c) is the attention heatmap generated by the method of this embodiment. Figure 8 The image shows a longitudinal section comparison of the Grad-CAM attention heatmaps of the first embodiment and the comparative model on difficult positive samples, where (a) is a reference annotation map of the lesion area, (b) is the attention heatmap generated by the baseline model ResNet-18, and (c) is the attention heatmap generated by the present method. Figure 9 The image shows a cross-sectional view of the Grad-CAM attention heatmap on difficult negative samples, comparing the method of the first embodiment with the comparison model. (a) is a reference annotation map of the lesion area, (b) is the attention heatmap generated by the baseline model ResNet-18, and (c) is the attention heatmap generated by the method of this embodiment. Figure 10 The image shows a longitudinal section comparison of the Grad-CAM attention heatmaps of the first embodiment and the comparative model on difficult negative samples, where (a) is a reference annotation map of the lesion area, (b) is the attention heatmap generated by the baseline model ResNet-18, and (c) is the attention heatmap generated by the present method. Detailed Implementation

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0023] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0024] Multi-view joint analysis has high clinical value. When the imaging quality of different views is inconsistent, or when a view is affected by noise, artifacts and low contrast, existing explicit fusion strategies often lack sufficient adaptability, resulting in the model's inadequate ability to identify small lesions, lesions with blurred boundaries, and difficult samples.

[0025] Implicit modeling methods such as depth equilibration models have advantages such as parameter sharing and memory friendliness. However, existing depth equilibration models still face new technical challenges in ultrasound lesion screening scenarios: on the one hand, ultrasound images generally have characteristics such as strong artifacts, heavy noise, unclear boundaries, and significant heterogeneity between different views, which can easily lead to slow convergence, oscillation, or even divergence in fixed-point iteration; on the other hand, existing related methods do not adequately consider the stability optimization of the implicit fusion process, view-specific discrimination, and clinical deployability, making it difficult to directly meet the comprehensive requirements of accuracy, robustness, and real-time performance for ultrasound-assisted lesion screening tasks.

[0026] How to fully utilize the complementary information between multi-view ultrasound images, improve the stability of feature fusion in complex noise scenarios, and enable the model to adapt to the clinical inference environment of low-computing-power medical terminals and even edge devices while maintaining the accuracy of lesion screening has become an urgent technical problem to be solved in this field.

[0027] Based on the above problems, this invention combines clinical application needs and adopts a technical approach of "fully utilizing multi-view samples for feature learning during the training phase and performing edge deployment for clinical scenarios during the inference phase". The training phase focuses on solving the problems of insufficient modeling of dynamic complementary information between multiple views, unstable DEQ fixed point solution under ultrasound artifact interference, and balancing lightweight and performance. The inference phase focuses on solving the problem of the model's practical deployment on low-computing-power medical terminals in clinical settings.

[0028] Example 1 like Figure 1 As shown, this embodiment discloses an ultrasound-assisted lesion screening method for medical terminals, preferably for plaque lesions in carotid artery ultrasound images. The method includes: Step S1: Acquire ultrasound images including longitudinal and transverse views and preprocess them to obtain preprocessed multi-view ultrasound images.

[0029] In this embodiment, the ultrasound images include longitudinal and transverse view images, each accompanied by corresponding view labels and image-level lesion category labels. Preferably, the longitudinal view image data uses the publicly available dataset Ar-PlaqSegm1, which contains 541 carotid ultrasound images, of which 493 are in the training set and 48 are in the test set; the transverse view image data uses a hospital-built dataset, which contains 500 carotid ultrasound images, of which 450 are in the training set and 50 are in the test set.

[0030] After acquiring ultrasound images, standardized preprocessing is performed on the images.

[0031] In this embodiment, all ultrasound images are adjusted to 256×256 pixels to eliminate the problem of inconsistent image size caused by different devices, different acquisition settings, and different acoustic window conditions; at the same time, Gaussian smoothing is applied to the images to suppress high-frequency artifacts and some speckle noise, and reduce the interference of noise on subsequent lesion edge, texture, and echo pattern modeling.

[0032] After preprocessing, standardized multi-view ultrasound image data, corresponding view labels, and lesion category labels are obtained, providing standardized input for subsequent multi-view differential feature extraction by the lightweight backbone network. In this embodiment, since it is geared towards clinical edge deployment scenarios, the size standardization and smoothing denoising methods with low computational overhead and deployment friendliness are preferred in the preprocessing stage to reduce the inference burden on the subsequent terminal side.

[0033] Step S2: Construct a lesion-assisted screening model, comprising a lightweight backbone network, an implicit fusion module, and a view-aware output module connected sequentially. The lightweight backbone network is responsible for extracting the differential features between longitudinal and transverse views; the DEQ (Depth Equalization) implicit fusion module is responsible for constructing shared hidden states and enabling high-order interactions between views; the view-aware output module is responsible for performing classification based on view attributes; and a training optimization module is also included to improve the stability of fixed-point solutions and the overall generalization ability of the model.

[0034] Through the collaborative work of various modules, an end-to-end processing chain is achieved from multi-view ultrasound image input to lesion screening result output. The screening process of the lesion-assisted screening model is as follows: Step S3: Input the preprocessed multi-view ultrasound images into a lightweight backbone network and extract multi-view differential features, including longitudinal and transverse view features.

[0035] In this embodiment, the lightweight backbone network is preferably a lightweight residual network based on an improved design of ResNet. Its core objective is to improve the encoding capability of ultrasound artifacts, low-contrast regions, and weak-response lesion features while controlling the number of parameters and computational complexity.

[0036] Specifically, such as Figure 2 As shown, the lightweight backbone network employs a four-stage convolutional feature extraction structure. Based on the original ResNet-18: The number of residual units was modified, and the number of residual blocks in each stage was adjusted from the original [2,2,2,2] to [1,2,2,1] to reduce redundant calculations; The traditional ReLU activation function is replaced with LeakyReLU with a negative slope of 0.1 to enhance the preservation of weak features with low response. The number of channels is controlled, with the output channels for each stage set to 32, 64, 128, and 64 respectively. The output channels are then compressed to 64 using a 1×1 convolution, thereby reducing the input dimension of the subsequent DEQ module. The first 32 channels of the output features can focus on the texture features of the longitudinal view, while the last 32 channels can focus on the morphological features of the transverse view.

[0037] In this embodiment, preferably, the input image is downsampled in the first stage to form a 128×128×32 feature map, and then progressively formed into 64×64×64 and 32×32×128 feature maps in the second and third stages, focusing on capturing the edge contours, internal textures, and local structural differences of the lesions; after the fourth stage, a compact feature map with a size of 16×16×64 is output. The 16×16×64 feature map, as a multi-view differential representation, will be directly input into the subsequent DEQ implicit fusion module.

[0038] Compared to existing highly complex backbone networks, the lightweight feature extraction structure adopted in this step can significantly reduce the number of parameters and computational load while maintaining strong discriminative ability, laying the foundation for subsequent deployment on low-computing-power medical terminals. Experiments show that the final number of parameters of the entire method can be controlled at 3.31M, FLOPs is 1.15G, the average inference time per image is 4.2ms, and the FPS is 238, demonstrating significant lightweight advantages.

[0039] Step S4: Input the multi-view differential features into the implicit fusion module, construct a shared hidden state, and achieve adaptive interaction and high-order fusion of the multi-view differential features through fixed-point solving to obtain global fusion features.

[0040] The longitudinal and transverse view features obtained from S3 are input into the implicit fusion module of the depth equalization model to construct a shared hidden state and achieve adaptive interaction of multi-view features through fixed-point solving. For example... Figure 3 As shown, the specific processing steps are as follows: First, the two features are concatenated in the channel dimension to form a hybrid tensor containing multi-scale complementary information; then, the hybrid tensor is mapped to a unified hidden state space through 1×1 convolution to compress redundant channels and reduce the differences between different view modalities.

[0041] In the hidden state update process, multi-view fusion is formalized into a fixed-point equation. The hidden state update mapping consists of 3×3 convolution, batch normalization, GELU activation function, view interaction attention mechanism, and residual connections. The view interaction attention mechanism performs channel-level global average pooling on the longitudinal and transverse view features respectively to obtain two channel description vectors; then, the two description vectors are concatenated and input into a two-layer fully connected network, and the channel weights are output through the Sigmoid function; finally, based on the channel weights, a channel-wise weighted summation is performed on the longitudinal and transverse features, thereby achieving adaptive adjustment of cross-view information.

[0042] Due to the generally low contrast, strong artifacts, and significant image heterogeneity of ultrasound images, traditional Picard iteration may experience slow convergence, oscillation, or divergence in some samples. Therefore, this embodiment employs Anderson Accelerator as the numerical solution tool for fixed-point iteration of the depth-equilibrium model. Specifically, a correction direction is constructed using historical hidden states with a memory depth of 5 and a residual cache, combined with a dynamic step size decay strategy to reduce the number of iterations required to achieve a given convergence accuracy and suppress residual oscillations caused by ultrasound noise.

[0043] After convergence in this step, the obtained fixed-point hidden states are used as global fusion features for subsequent view perception output heads; at the same time, the iterative residuals are retained as the basis for uncertainty quantification and participate in the loss design during the training phase, thereby further suppressing overfitting.

[0044] Step S5: Input the global fusion features into the view perception output head of the view perception output module to perform lesion classification and complete training optimization.

[0045] The global fusion features obtained in S3 are input into the view perception output head to obtain lesion-assisted screening results. In this embodiment, the view perception output head adopts a parallel three-branch multi-output head structure, including a long head (vertical section branch), a short head (horizontal section branch), and a mix head (mixed branch). Among them, the long head is used for vertical section view classification, and the input is a concatenated representation of the global fusion features and the vertical section single view features; the short head is used for horizontal section view classification, and the input is a concatenated representation of the global fusion features and the horizontal section single view features; the mix head only takes the global fusion features as input and is used as a backtracking classification branch in scenarios where view labels are missing, blurred, or the image quality is low.

[0046] During the training phase, to ensure collaborative optimization across different branches, a multi-supervised loss function is employed: the binary cross-entropy losses of the long, short, and mixed heads are weighted and summed with weights of 0.4:0.4:0.2 to improve overall generalization ability. Simultaneously, each classification branch uses a dynamically weighted binary cross-entropy loss to adapt to the differences in data distribution across different view branches; the output layer bias is initialized using the logit value based on the global proportion of positive classes in the data to reduce output bias in the early stages of training.

[0047] Specifically, to optimize model training stability and prediction accuracy, and to adapt to the data distribution characteristics of different view branches, each classification branch head adopts a dynamically weighted binary cross-entropy loss:

[0048] in, , y represents the branch prediction probability; y represents the image-level lesion category label / plaque binary classification label; The weights are adaptively updated based on the global statistical characteristics of the corresponding view branch data, such as the weights of vertically slicing branches. :

[0049] in, This represents the total number of positive samples in the longitudinal section view. This represents the total number of negative samples in the longitudinal section view.

[0050] To stabilize DEQ iterations and suppress ultrasonic noise interference, further improving training stability and prediction accuracy, a bi-term explicit regularization is introduced: the first term minimizes the residual between the hidden state and the mapped output to smooth local behavior near fixed points; the second term constrains the hidden state magnitude to prevent gradient explosion or iteration trajectory divergence. The optimizer chosen is AdamW, with an initial learning rate of 10. -4 The weight decay factor is set to 10. -5 During training, a gradient-based solution mechanism based on the implicit function theorem is employed, and the GMRES method is used to solve the Jacobian vector product to improve the stability of backpropagation in the deep equalization model module.

[0051] Specifically, the bi-term explicit regularization loss:

[0052] in, , The regularization coefficient is . It is in a hidden state. This serves as the fusion input after the backbone network features are mapped to the hidden state space. This is an iterative mapping function.

[0053] The first term minimizes the residual between the hidden state and its mapping output, causing the mapping function to have smoother local behavior near the fixed point, thus providing a more reliable residual estimate for Anderson acceleration and reducing extrapolation bias caused by drastic changes in the mapping; the second term constrains the magnitude of the hidden state to prevent gradient explosion or divergence of the iteration trajectory due to excessively large values.

[0054] Total loss function:

[0055] in, , , These represent the binary cross-entropy loss weights for long, short, and mixed headers, respectively. , , These represent the binary cross-entropy losses for long headers, short headers, and mixed headers, respectively. This represents the loss of explicit regularization for two terms.

[0056] In this embodiment, take , , , .

[0057] Step S6: Perform inference deployment on a low-computing-power medical terminal and output auxiliary screening results.

[0058] The trained lesion-assisted screening model is deployed on a low-computing-power medical terminal. In clinical scenarios, it performs real-time inference on the ultrasound images of the subjects to be screened and outputs the lesion-assisted screening results. In this embodiment, the low-computing-power medical terminal uses an NVIDIA Jetson AGX Orin edge device.

[0059] Before deployment, the trained model is converted into an inference model format suitable for terminal operation, and a preprocessing flow consistent with the input size, normalization method, and view label reading logic is configured on the terminal side. During inference, ultrasound images of the object to be tested are first acquired and standardized in the same way as during the training phase. Then, the processed images are input into the lightweight backbone network and the implicit fusion module of the deep equalization model. Subsequently, the output head is dynamically called according to the view label or view determination result.

[0060] Specifically, when the input image is determined to be a longitudinal view, the system calls the long header to output the main classification result and uses the mix header for consistency auxiliary verification; when the input image is determined to be a transverse view, the system calls the short header to output the main classification result and uses the mix header for consistency auxiliary verification; when the view label is missing, the view determination result is unclear, or the image quality is lower than a preset threshold, the system calls the mix header as the dominant output branch to improve the robustness of the model in uncertain scenarios.

[0061] The final output of the lesion-assisted screening results includes lesion category, classification confidence score, and a heatmap of the optional model interest region. This embodiment maintains high screening performance while keeping the number of parameters around 3.31M, FLOPs at 1.15G, average inference time per image at 4.2ms, and a frame rate of 238 FPS. It is well-suited for low-computing-power applications such as primary healthcare institutions, portable ultrasound equipment, and clinical bypass-assisted screening.

[0062] Through the coordinated work of the above modules, this embodiment can form a complete data processing closed loop, realizing ultrasound image lesion assisted screening for medical terminals with low computing power.

[0063] Based on the method proposed in this embodiment, the following experimental design is performed: (I) Experimental Design Experimental validation was conducted using both publicly available datasets and self-built clinical datasets.

[0064] 1. Longitudinal view dataset: The publicly available dataset Ar-PlaqSegm1 was selected as the longitudinal view dataset. This dataset contains 541 carotid ultrasound images, of which 493 are in the training set (341 images with atherosclerotic plaques and 152 images without visible plaques) and 48 are in the test set (approximately 26 images with atherosclerotic plaques and approximately 22 images without visible plaques).

[0065] 2. Cross-sectional view dataset: The cross-sectional view dataset was collected by Jinan Central Hospital and contains 500 carotid ultrasound images, including 450 images in the training set (297 images with atherosclerotic plaques and 153 images without visible plaques) and 50 images in the test set (27 images with atherosclerotic plaques and 23 images without visible plaques).

[0066] 3. Experimental Platform: The experimental platform consisted of one Intel® Core™ i5-14600KF CPU @3.50GHz, one NVIDIA GeForce RTX™ 5070Ti GPU with 16GB of video memory, and ran Ubuntu 20.04 LTS. Python was used as the programming language, and PyTorch-2.8.0 was employed as the deep learning framework. The AdamW optimizer was used during model training, with an initial learning rate set to 10. -4 The weight decay factor is set to 10. -5 .

[0067] 4. Data preprocessing: To eliminate differences in image size and grayscale distribution caused by different ultrasound equipment, operators, and acoustic window conditions, the original ultrasound images were first uniformly adjusted to 256×256 pixels, and a Gaussian smoothing algorithm was used to suppress high-frequency artifacts in the images to reduce the interference of noise on the characterization of lesion areas.

[0068] 5. Unified Evaluation Method: To ensure experimental comparability, a unified training strategy was adopted across all comparison models and ablation configurations, and an early stopping strategy was used to suppress overfitting. Precision (P), recall (R), balance score (F1), and accuracy (A) were used to evaluate overall classification performance. Simultaneously, to assess the ability to identify positive samples, positive sample precision (PM), positive sample recall (RM), and positive sample balance score (F1M) were further calculated.

[0069] 6. Experiment Content: The experiment includes the following parts: (1) Comparative experiments were conducted with various mainstream convolutional networks, Transformer-type methods and medical image classification methods to verify the advantages of the method in this embodiment in terms of overall classification performance and positive sample recall.

[0070] To comprehensively evaluate the performance of the method in this embodiment, the following representative models were selected as comparison methods: ResNet-18, DenseNet-121, ViT, Swin-Transformer, ConvNeXt, VisionMamba, MedKAN, VET-DINO, Multi-view Evidential Learning, and MedNet.

[0071] Table 1 shows the comparison results of the various methods on the test set in terms of classification performance.

[0072] Table 1 Comparative Experimental Results

[0073] As shown in Table 1, the method in this embodiment is highly competitive in terms of overall classification performance and positive sample detection capability. Specifically, the macro-average F1 score of this method reaches 0.9429, the positive sample recall rate RM reaches 0.9630, and the positive sample balance score F1M reaches 0.9369, which further improves the ability to identify positive plaque samples while ensuring overall accuracy.

[0074] To further quantify the efficiency advantages of the method in this embodiment, the number of parameters, number of floating-point operations, average inference time per image, and frame rate of different models were compared, and the results are shown in Table 2. All measurements were performed under the conditions of 256×256 single-channel grayscale image input, FP32 precision, batch=1, and the model in evaluation mode.

[0075] Table 2 Comparison of computational efficiency of different models

[0076] As shown in Table 2, the method in this embodiment exhibits significant advantages in terms of parameter quantity, computational complexity, and inference efficiency. Its parameter quantity is 3.31M, FLOPs is 1.15G, the average inference time per image is 4.2ms, and the FPS is 238. This embodiment maintains high screening performance while having low computational complexity, making it more suitable for deployment on medical terminals with low computing power.

[0077] (2) Conduct model ablation experiments, starting from the baseline model and gradually adding the DEQ implicit fusion module, DEQ optimization strategy, training initialization scheme and view-aware multi-head supervision to verify the effectiveness of each component.

[0078] To verify the effectiveness of each core component in the method of this embodiment, the following ablation experiment was designed.

[0079] Experiment #1 is the baseline model, which uses the same lightweight backbone network as the method in this embodiment, but only uses simple feature concatenation for view fusion and uses standard binary cross-entropy loss for single-head classification. Experiment #2 introduces the DEQ implicit fusion module based on #1; Experiment #3 further incorporates DEQ optimization strategies based on #2, including Anderson acceleration and view interaction attention; Experiment #4 adds a training initialization scheme to #3, including dynamic weighted loss and bias initialization; Experiment #5, based on #4, incorporates view-aware multi-head supervision to form a complete method. The experimental results for each ablation configuration are shown in Table 3.

[0080] Table 3 Ablation Experiment Results

[0081] As shown in Table 3, the model performance continuously improves with the gradual addition of each key module. After introducing the DEQ implicit fusion module, the macro-average F1 score increased from 0.7883 to 0.8530; after further adding Anderson acceleration and view interaction attention, the positive sample recall rate increased from 0.8519 to 0.8889; after adding the training initialization scheme, the macro-average F1 score increased to 0.9020 and the positive sample recall rate increased to 0.9074; when view-aware multi-head supervision was added, all indicators reached their optimal levels. All of the above modules made a real contribution to the final performance improvement.

[0082] (3) Conduct supplementary experiments with different output heads to analyze the performance differences of short head, long head, mix head and their combinations.

[0083] To further analyze the role of the view-aware multi-head module, supplementary experiments were conducted on short head, long head, mix head and their combinations, and the results are shown in Table 4.

[0084] Table 4 Performance Comparison of Different Output Heads

[0085] Table 4 shows that each individual output head possesses strong independent classification capabilities. The short head alone achieves a positive sample recall of 0.9523, indicating that the cross-sectional view itself has high discriminative value. Furthermore, multi-head collaboration can further improve overall performance. For example, when the short head is combined with the mix head, the positive sample recall increases from 0.9523 to 0.9582. When the short, long, and mix output heads work together, the model achieves the optimal overall result.

[0086] (4) Perform convergence stability analysis for different DEQ configurations.

[0087] To evaluate the numerical stability and convergence speed of different DEQ configurations in the fixed-point solution process, convergence analyses were performed on the original DEQ, the DEQ with only Anderson acceleration, and the complete method with Anderson acceleration and bi-term regularization. Figure 4 The diagram illustrates the stability and convergence speed of different DEQ configurations during the fixed-point solution process. The residual changes of each method during the iteration process were recorded in the experiment to measure their convergence efficiency and stability. The analysis results show that, compared with the DEQ method without Anderson acceleration, introducing Anderson acceleration and combining it with a stabilization strategy significantly accelerates the decrease in fixed-point residuals and makes the entire iteration process smoother, demonstrating higher numerical robustness.

[0088] (5) Interpretability analysis was performed using Grad-CAM and feature visualization.

[0089] To visually verify the lesion localization capability of the method in this embodiment and its consistency with clinical diagnostic logic, Grad-CAM technology was used to generate a model attention heatmap, and the attention patterns of ResNet-18 and the method in this embodiment were compared in typical cases. Figures 5-10 The diagram shows a comparison of Grad-CAM attention heatmaps between the method of this embodiment and the comparative model on typical and difficult samples. Figure 5 The images show cross-sectional views of typical positive samples, where (a) is a reference annotation map of the lesion area, (b) is an attention heatmap generated by the baseline model ResNet-18, and (c) is an attention heatmap generated by this method. Figure 6 The images show longitudinal cross-sectional views of typical positive samples, where (a) is a reference annotation map of the lesion area, (b) is an attention heatmap generated by the baseline model ResNet-18, and (c) is an attention heatmap generated by this method. Figure 7 The images show cross-sectional views of difficult positive samples, where (a) is a reference annotation map of the lesion area, (b) is an attention heatmap generated by the baseline model ResNet-18, and (c) is an attention heatmap generated by this method. Figure 8 The images show longitudinal cross-sectional views of difficult positive samples, where (a) is a reference annotation map of the lesion area, (b) is an attention heatmap generated by the baseline model ResNet-18, and (c) is an attention heatmap generated by this method. Figure 9 Cross-sectional views of difficult negative samples, where (a) is a reference annotation map of the lesion area, (b) is an attention heatmap generated by the baseline model ResNet-18, and (c) is an attention heatmap generated by this method; Figure 10The images show longitudinal views of difficult-to-diagnose negative samples, where (a) is a reference annotation map of the lesion area, (b) is an attention heatmap generated by the baseline model ResNet-18, and (c) is an attention heatmap generated by the proposed method. The visualization results cover three clinical scenarios: typical positive, difficult-to-diagnose, and difficult-to-diagnose, with each scenario including both transverse and longitudinal views. Analysis results show that the proposed method can focus more intently on the core plaque area in typical positive samples; in difficult-to-diagnose positive samples, it can still reliably focus on key lesion areas even in scenarios with blurred boundaries, artifact interference, or low contrast; and in difficult-to-diagnose negative samples, the degree of mis-focusing on artifact areas is lower than that of the baseline model, indicating that the proposed method has higher robustness and better clinical consistency in complex ultrasound scenarios.

[0090] Furthermore, the distribution of fused features was analyzed using feature visualization. The results show that the method in this embodiment can obtain clearer inter-class separation boundaries, which is beneficial to improving the ability to distinguish between patch and non-patch samples.

[0091] 7. Experimental Conclusions (1) Significant advantages in classification performance: The macro-average F1 value of the method in this embodiment reaches 0.9429, the accuracy reaches 0.9435, the positive sample recall reaches 0.9630, and the positive sample balance score reaches 0.9369. It performs outstandingly among a variety of mainstream comparison methods, indicating that the method in this embodiment can significantly improve the detection ability of plaque positive samples while ensuring overall performance.

[0092] (2) Lightweight and high efficiency: The method in this embodiment has 3.31M parameters, 1.15G FLOPs, an average inference time of 4.2ms per image, and 238 FPS, which is significantly better than most comparative models. This shows that the method in this embodiment can meet the accuracy requirements of clinical auxiliary screening while being more suitable for deployment in medical terminals with low computing power.

[0093] (3) The design of each module is necessary: ​​Ablation experiments show that the DEQ implicit fusion module, Anderson acceleration and view interaction attention, training initialization scheme and view perception multi-head supervision can all bring stable performance gains, indicating that the overall architecture design of the method in this embodiment has clear effectiveness.

[0094] (4) The multi-head collaborative mechanism is effective: Supplementary experiments with different output heads show that the short head, long head and mix head have high classification ability when used alone, and can achieve better comprehensive performance when the three are used in combination. This indicates that the view-aware multi-head output structure can simultaneously take into account the targeted modeling and overall generalization ability under different views.

[0095] (5) Good convergence stability and interpretability: Convergence analysis shows that the method in this embodiment has higher numerical stability in the fixed-point solution process; Visualization analysis shows that the region of interest of the method in this embodiment has high consistency with the core region of the lesion. The method in this embodiment not only has high screening performance, but also has good interpretability and clinical application value.

[0096] Example 2 This embodiment discloses an ultrasound imaging-assisted lesion screening system for medical terminals, including: The sample acquisition module is configured to: acquire a training sample set, which includes ultrasound images of longitudinal and transverse views, corresponding view labels and lesion category labels; preprocess the ultrasound images to obtain preprocessed multi-view ultrasound images. The model building module is configured to build a lesion-assisted screening model, which includes a lightweight backbone network, an implicit fusion module, and a view-aware output module connected in sequence. The feature extraction module is configured to: input preprocessed multi-view ultrasound images into a lightweight backbone network and extract multi-view differential features, including longitudinal view features and transverse view features; The feature fusion module is configured to: input multi-view differential features into the implicit fusion module, construct a shared hidden state and achieve adaptive interaction and high-order fusion of multi-view differential features through fixed-point solving, and obtain global fusion features; The result output module is configured to input global fusion features into the view perception output head of the view perception output module to obtain lesion-assisted screening results.

[0097] In this embodiment, the system may optionally include an interpretability display module, which generates a heat map based on the region of interest output by the model and overlays the heat map with the original ultrasound image to assist clinicians in interpreting lesions.

[0098] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0099] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of an ultrasound imaging-assisted lesion screening method for a medical terminal as described in Embodiment 1 of this disclosure.

[0100] Example 4 The purpose of this embodiment is to provide an electronic device.

[0101] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of an ultrasound imaging-assisted lesion screening method for a medical terminal as described in Embodiment 1 of this disclosure.

[0102] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0103] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0104] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for ultrasound-assisted lesion screening for medical terminals, characterized in that, include: A training sample set is obtained, which includes ultrasound images of longitudinal and transverse views, corresponding view labels, and lesion category labels. The ultrasound images are preprocessed to obtain preprocessed multi-view ultrasound images. A lesion-assisted screening model is constructed, comprising a lightweight backbone network, an implicit fusion module, and a view-aware output module connected in sequence; During the training phase of the lesion-assisted screening model, the multi-view features extracted by the lightweight backbone network are mapped to the shared hidden state space, and the multi-view features are implicitly fused through a deep equalization model; the Anderson acceleration algorithm is introduced in the fixed-point solution process; and the model parameters are jointly optimized by combining a bi-term explicit regularization strategy and a multi-component loss function. In the model output stage, a view perception output head containing longitudinal, transverse, and hybrid branches is constructed. During training, view labels are used to jointly optimize different branches. During inference, the corresponding output head is dynamically selected to perform classification based on the view labels of the input ultrasound image. When the view attributes are unclear or there is inconsistency in the information of multiple views, the hybrid branch is enabled to perform backtracking discrimination in order to obtain lesion-assisted screening results. The preprocessed multi-view ultrasound images are input into a lightweight backbone network to extract multi-view differential features, including longitudinal and transverse view features. The multi-view differential features are input into the implicit fusion module to construct a shared hidden state and achieve adaptive interaction and high-order fusion of the multi-view differential features through fixed-point solving, thereby obtaining global fusion features; The implicit fusion process of the deep equalization model includes: stitching together multi-view features in the channel dimension and mapping them to a unified hidden state space through 1×1 convolution; The hidden state update mapping consists of 3×3 convolution, batch normalization, GELU activation function, view interaction attention mechanism and residual connection; the view interaction attention mechanism performs channel-level global average pooling on each view feature, concatenates the resulting channel description vectors and inputs them into a two-layer fully connected network, outputs channel weights through the Sigmoid function, and performs channel-wise weighted summation on each view feature based on the channel weights; The global fusion features are input into the view perception output head of the view perception output module to obtain the lesion-assisted screening results.

2. The ultrasound imaging-assisted lesion screening method for medical terminals as described in claim 1, characterized in that, The lightweight backbone network adopts a four-stage convolutional feature extraction structure, optimizes the configuration of the number of residual units, and the number of output channels in each stage is 32, 64, 128 and 64 respectively. The output channels are compressed to 64 through 1×1 convolution. The activation function is LeakyReLU with a negative slope of 0.

1.

3. The ultrasound imaging-assisted lesion screening method for medical terminals as described in claim 1, characterized in that, The bi-term explicit regularization strategy includes: a first regularization term used to minimize the residual between the hidden state and the mapped output to smooth local behavior near the fixed point; and a second regularization term used to constrain the magnitude of the hidden state to prevent gradient explosion or divergence of the iteration trajectory.

4. The ultrasound imaging-assisted lesion screening method for medical terminals as described in claim 1, characterized in that, The multi-component loss function includes dynamically weighted binary cross-entropy loss and bivariate explicit regularization loss: in, , , These represent the binary cross-entropy loss weights for long, short, and mixed headers, respectively. , , These represent the binary cross-entropy losses for long headers, short headers, and mixed headers, respectively. This represents the loss of explicit regularization for both terms.

5. A lesion-assisted ultrasound imaging system for medical terminals, characterized in that, include: The sample acquisition module is configured to: acquire a training sample set, which includes ultrasound images of longitudinal and transverse views, corresponding view labels and lesion category labels; and preprocess the ultrasound images to obtain preprocessed multi-view ultrasound images. The model building module is configured to build a lesion-assisted screening model, which includes a lightweight backbone network, an implicit fusion module, and a view-aware output module connected in sequence. During the training phase of the lesion-assisted screening model, the multi-view features extracted by the lightweight backbone network are mapped to the shared hidden state space, and the multi-view features are implicitly fused through a deep equalization model; the Anderson acceleration algorithm is introduced in the fixed-point solution process; and the model parameters are jointly optimized by combining a bi-term explicit regularization strategy and a multi-component loss function. In the model output stage, a view perception output head containing longitudinal, transverse, and hybrid branches is constructed. During training, view labels are used to jointly optimize different branches. During inference, the corresponding output head is dynamically selected to perform classification based on the view labels of the input ultrasound image. When the view attributes are unclear or there is inconsistency in the information of multiple views, the hybrid branch is enabled to perform backtracking discrimination in order to obtain lesion-assisted screening results. The feature extraction module is configured to: input the preprocessed multi-view ultrasound images into a lightweight backbone network and extract multi-view differential features, including longitudinal view features and transverse view features; The feature fusion module is configured to: input the multi-view differential features into the implicit fusion module, construct a shared hidden state and achieve adaptive interaction and high-order fusion of the multi-view differential features through fixed-point solving, and obtain global fusion features; The implicit fusion process of the deep equalization model includes: concatenating multi-view features in the channel dimension and mapping them to a unified hidden state space through 1×1 convolution; the hidden state update mapping consists of 3×3 convolution, batch normalization, GELU activation function, view interaction attention mechanism and residual connection; the view interaction attention mechanism performs channel-level global average pooling on each view feature, concatenates the resulting channel description vectors and inputs them into a two-layer fully connected network, outputs channel weights through the Sigmoid function, and performs channel-wise weighted summation on each view feature based on the channel weights; The result output module is configured to input the global fusion features into the view perception output head of the view perception output module to obtain the lesion-assisted screening result.

6. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps of the ultrasound imaging-assisted lesion screening method for medical terminals as described in any one of claims 1-4.

7. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the ultrasound imaging-assisted lesion screening method for medical terminals as described in any one of claims 1-4.

Images