Focus identification method and device for B-mode ultrasound image, medium and equipment

Abstract

The invention discloses a focus identification method and device for a B-mode ultrasound image, a medium and equipment. The method comprises the following steps: firstly, acquiring a continuous multi-frame B-mode ultrasound image sequence through a real-time acquisition technology, and then performing feature extraction on the images by using a preset lightweight convolutional neural network so as to obtain a single-frame feature map of a plurality of continuous frames; next, the single-frame feature maps are subjected to feature fusion based on the time sequence continuity of the image, a preset fusion window and a set of preset weights to generate a plurality of fused feature representations. Finally, a lightweight inference network obtained through quantization and pruning optimization is used for conducting focus recognition on the fusion feature representations, and therefore focus recognition results of the continuous multi-frame B-mode ultrasound images are obtained. The process not only improves the accuracy and the real-time performance of identification, but also optimizes the energy consumption and the calculation efficiency of the low-end mobile edge device.

Description

Technical Field

[0001] The present invention relates to the field of lesion recognition in B-ultrasound images, and in particular to a method, device, medium and equipment for lesion recognition in B-ultrasound images. Background Art

[0002] With the development of medical imaging technology, lesion detection has become an indispensable part of modern medical diagnosis. Ultrasonic medical images (B-ultrasound) have been widely used in clinical diagnosis due to their non-invasive, convenient, and cost-effective characteristics. However, B-ultrasound images require extremely high real-time performance. Traditional lesion recognition methods rely on high-performance servers or GPU-accelerated devices to perform complex deep neural network (DNN) inference to ensure high-precision detection on a single frame. This approach is difficult to implement on low-end mobile edge devices with limited resources, mainly due to the following problems:

[0003] Single-frame inference requires a lot of computation: Existing technologies typically extract features and perform predictions on each B-ultrasound image frame based on a convolutional neural network (CNN). This results in a computational load that is proportional to the frame rate, making it difficult to achieve real-time inference on low-end devices. This is especially true when the image frame rate is high, as the cumulative computational load increases significantly, easily leading to delays and frame drops.

[0004] Lack of data redundancy: B-ultrasound image sequences have temporal continuity, and adjacent frames have similar overall structures except for minor motion. Existing technologies fail to fully utilize this data redundancy, instead performing independent calculations on areas of overlapping information, resulting in a waste of resources.

[0005] Low energy consumption and hardware resource utilization: The energy and computing power of mobile edge devices are very limited. Complex frame-by-frame reasoning inevitably causes a sharp increase in energy consumption, reduces device battery life, and when hardware resources are limited, the reasoning speed is difficult to meet clinical real-time detection requirements.

[0006] Insufficient model adaptability: To solve single-frame inference problems with high-precision networks, existing technologies often use large-scale parameter models, which are difficult to deploy on low-end devices. It is also difficult to achieve model quantization, pruning, and compression while balancing accuracy and inference speed.

[0007] These deficiencies result in the inability of existing technologies to accurately and quickly identify lesions in B-ultrasound images on low-end equipment. Summary of the Invention

[0008] The present invention provides a method for identifying lesions in B-ultrasound images to solve the problem in the prior art that the prior art cannot accurately and quickly identify lesions in B-ultrasound images on low-end equipment.

[0009] In a first aspect, the present application provides a method for identifying lesions in B-ultrasound images, comprising:

[0010] Acquire multiple continuous frames of B-ultrasound image sequences in real time;

[0011] Performing feature extraction on the B-ultrasound image sequence according to a preset lightweight convolutional neural network to obtain a plurality of single-frame feature maps of consecutive multiple frames;

[0012] Performing feature fusion on the single-frame feature maps of the continuous multiple frames based on temporal continuity, a preset fusion window, and preset weights to generate a plurality of fused feature representations;

[0013] Based on the plurality of fused feature representations, lesion recognition is performed through a preset lightweight inference network to obtain lesion recognition results of multiple consecutive frames of B-ultrasound images;

[0014] The preset lightweight inference network is obtained by quantizing and pruning a preset initial lightweight inference network according to a preset optimization formula.

[0015] The present application obtains a series of continuous multi-frame B-ultrasound images in real time and uses a preset lightweight convolutional neural network for feature extraction, effectively capturing the key information in the image sequence. Subsequently, a fusion strategy based on temporal continuity is adopted, combined with a dynamically adjusted fusion window and preset weights, to fuse the single-frame feature maps to generate a fused feature representation. This process not only makes full use of the temporal redundant information of the image sequence and reduces the computational complexity, but also enhances the robustness of the feature representation through weight adjustment. Finally, the lightweight inference network obtained by quantization and pruning optimization is used to identify lesions on the fused features, which not only improves the accuracy and real-time performance of the recognition, but also significantly reduces energy consumption, making this method particularly suitable for resource-constrained low-end mobile edge devices. Therefore, the beneficial effect of the present application is that it achieves a significant reduction in computational complexity and energy consumption while ensuring detection accuracy, and improves the adaptability and robustness of the system in dynamically changing scenarios. The present application effectively solves the problem that the existing technology cannot accurately and quickly identify lesions in B-ultrasound images on low-end devices.

[0016] Furthermore, after acquiring a continuous multi-frame B-ultrasound image sequence in real time, the method further includes:

[0017] According to the preset noise reduction filtering and pixel value normalization processing methods, the continuous multi-frame B-ultrasound image sequence is preprocessed and adjusted to a preset resolution.

[0018] The present application first obtains a continuous multi-frame B-ultrasound image sequence in real time, ensures the real-time and continuity of the data, which is crucial for the accuracy of lesion detection. Then, a preset noise reduction filter is used to effectively reduce the noise in the image and improve image quality, while the pixel value normalization process ensures that the image data is within a unified numerical range, thereby enhancing the stability and reliability of subsequent feature extraction. In addition, adjusting the image sequence to a preset resolution not only optimizes the display effect of the image, but also provides standardized input for the feature extraction of lightweight convolutional neural networks, which helps to improve the efficiency and accuracy of feature extraction. Combining these pre-processing steps, the beneficial effect is to enhance the quality of the image data, lay a solid foundation for subsequent feature extraction and lesion identification, thereby improving the accuracy and robustness of the entire lesion detection system.

[0019] Furthermore, the feature fusion of the single-frame feature maps of the continuous multiple frames is performed based on temporal continuity, a preset fusion window and preset weights, specifically:

[0020] According to the preset local contrast, the preset texture complexity and the preset motion amplitude, combined with the preset weight distribution formula, the preset weights are calculated;

[0021] According to the preset weights and the preset weighted fusion formula, weighted feature fusion is performed on the single-frame feature maps of the continuous multiple frames in a preset fusion window to obtain a plurality of fused feature representations.

[0022] The preset weight distribution formula and the preset weighted fusion formula are specifically:

[0023] The preset weight distribution formula is:

[0024]

[0025] Where C i is the local contrast score of the i-th frame, T i is the texture complexity, M i is the motion amplitude, α, β, γ are preset adjustable hyper parameters, w i For each preset weight;

[0026] The preset weighted fusion formula is:

[0027]

[0028] Where N is a preset positive integer greater than 1, f i is the feature map extracted from the i-th frame, w i is the weight of the i-th frame, F fuse It is the fusion feature representation.

[0029] This application first considers the temporal continuity of the B-ultrasound image sequence, uses the three key indicators of preset local contrast, texture complexity and motion amplitude, and combines the weight distribution formula to calculate the weight of each frame feature map. This step reasoning process ensures that the fusion process can be differentiated according to the actual importance of the image content. Subsequently, the single-frame feature maps of multiple consecutive frames are weightedly fused within the fusion window using a preset weighted fusion formula to generate a fused feature representation. This fusion strategy not only improves the richness and accuracy of the feature representation, but also adapts to changes in image content by dynamically adjusting the fusion window size, thereby enhancing the robustness of the system and the accuracy of lesion recognition while ensuring real-time performance. Therefore, the beneficial effect is that the quality and efficiency of lesion detection are improved through precise feature fusion, making it possible to achieve efficient and accurate lesion recognition on low-end mobile edge devices.

[0030] Furthermore, the step of performing feature fusion on the single-frame feature maps of the consecutive multiple frames based on temporal continuity, a preset fusion window, and preset weights to generate a plurality of fused feature representations further includes:

[0031] monitoring the difference between consecutive frames of the acquired multiple-frame B-ultrasound image sequence;

[0032] If the difference is higher than a preset threshold, dynamically adjusting the size of the preset fusion window according to a preset fusion window adjustment formula;

[0033] The preset fusion window adjustment formula is:

[0034] N t =round(N o ·e -λ·ΔI )

[0035] Where ΔI is the difference between the current frame and the previous frame, λ is the attenuation coefficient, and N t is the adjusted fusion window size, N o The size of the preset fusion window.

[0036] This application achieves adaptability to changes in image sequences by monitoring the differences between consecutive multi-frame B-ultrasound image sequences and dynamically adjusting the size of the fusion window using a fusion window adjustment formula when the difference exceeds a preset threshold. First, the single-frame feature maps of consecutive frames are weighted fused using preset weights to generate a fused feature representation, which helps to comprehensively utilize temporal information and improve the expressiveness of features. Second, by monitoring the difference between frames and dynamically adjusting the fusion window based on the difference, it can flexibly respond to changes in image content, such as rapid motion or noise, thereby maintaining the accuracy and real-time performance of feature fusion in different scenarios.

[0037] Furthermore, the preset initial lightweight inference network is quantized and pruned according to a preset optimization formula, specifically:

[0038] The preset optimization formula is:

[0039] FLOPs_reduction=original_FLOPs·a·b

[0040] Where FLOPs_reduction is the preset lightweight inference network, original_FLOPs is the preset initial lightweight inference network, a is the preset pruning ratio coefficient, and b is the preset quantization efficiency.

[0041] The initial lightweight inference network of this application is designed as an efficient structure to reduce the amount of computation (FLOPs). Then, by applying the optimization formula, the network is further pruned and quantized, where the pruning ratio coefficient (a) and quantization efficiency (b) are pre-set parameters used to guide network optimization. The pruning process removes unnecessary connections in the network, reducing the complexity and computational requirements of the model, while the quantization process converts the network parameters from high-precision representation to low-precision, further reducing the amount of computation. The beneficial effect is that this method not only significantly reduces the computational load of the model, but also improves the inference speed while maintaining the accuracy of the model, making the optimized lightweight inference network particularly suitable for running on low-end mobile edge devices with limited computing resources, thereby achieving efficient lesion identification without sacrificing too much accuracy.

[0042] Furthermore, the step of performing feature fusion on the single-frame feature maps of the continuous multiple frames based on temporal continuity, a preset fusion window, and preset weights further includes:

[0043] According to a preset average fusion formula, the single-frame feature maps of the continuous multiple frames are fused;

[0044] The preset average fusion formula is:

[0045]

[0046] Where, f i is the single-frame feature map extracted from the i-th frame, N is the number of frames continuously involved in fusion, and F fuse is the fused feature map.

[0047] This application extracts single-frame feature maps of each frame from multiple frames of B-ultrasound images acquired continuously. These feature maps contain the key information of the image. Then, based on temporal continuity, that is, the continuity and similarity between image frames, a preset average fusion formula is used to fuse these feature maps. This process reduces the differences between feature maps through averaging and enhances the stability and representativeness of the features. Ultimately, the fused feature maps can more comprehensively represent the temporal information of continuous frames, providing a richer and more reliable feature basis for subsequent lesion identification. This method effectively reduces the noise impact of single-frame features, improves the robustness of features, and reduces the amount of computation by reducing redundant information, thereby improving the accuracy and efficiency of lesion identification. It is particularly suitable for medical image analysis scenarios with limited real-time and computing resources.

[0048] In a second aspect, the present application provides a device for identifying lesions using B-ultrasound images, the device comprising:

[0049] An acquisition module is used to acquire a continuous multi-frame B-ultrasound image sequence in real time;

[0050] A feature extraction module is used to extract features from the B-ultrasound image sequence according to a preset lightweight convolutional neural network to obtain a plurality of single-frame feature maps of consecutive multiple frames;

[0051] A feature fusion module is used to fuse the single-frame feature maps of the continuous multiple frames based on temporal continuity, a preset fusion window and preset weights to generate a plurality of fused feature representations;

[0052] A lesion recognition module is used to perform lesion recognition based on the plurality of fused feature representations through a preset lightweight inference network to obtain lesion recognition results for multiple consecutive frames of B-ultrasound images;

[0053] The preset lightweight inference network is obtained by quantizing and pruning a preset initial lightweight inference network according to a preset optimization formula.

[0054] This application achieves efficient processing and accurate recognition of B-ultrasound image sequences through an integrated device design. First, the acquisition module is responsible for capturing multiple consecutive frames of B-ultrasound images in real time, ensuring the real-time and continuity of the data; then, the feature extraction module uses a lightweight convolutional neural network to extract features from these images, effectively capturing key information and reducing the computational load; then, the feature fusion module uses temporal continuity, fusion windows and preset weights to fuse single-frame feature maps, which not only enhances the robustness of the features, but also reduces the amount of computation by reducing redundant information; finally, the lesion recognition module uses a lightweight inference network optimized by quantization and pruning to identify the fused features, thereby improving the speed and accuracy of recognition. Overall, the device significantly improves the processing speed and reduces resource consumption while ensuring recognition accuracy through the collaborative work of various modules. It is particularly suitable for low-end mobile edge devices, thus showing important application value in the field of medical image analysis, especially in scenarios where fast and accurate diagnosis is required.

[0055] In a third aspect, the present application provides a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, the device containing the computer-readable storage medium is controlled to execute the method for identifying lesions in B-ultrasound images as described above. The beneficial effects thereof are the same as those of the method for identifying lesions in B-ultrasound images provided in the first aspect of the present application.

[0056] In a fourth aspect, the present application provides a terminal device comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements any one of the B-ultrasound image lesion recognition methods described in the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 : A schematic flow chart of an embodiment of a method for identifying lesions using B-ultrasound images provided in this application;

[0058] Figure 2 : A structural schematic diagram of an embodiment of a device for identifying lesions in B-ultrasound images provided in this application. DETAILED DESCRIPTION

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0060] Example 1

[0061] Please refer to Figure 1 In order to solve the problem that the existing technology cannot accurately and quickly identify lesions in B-ultrasound images on low-end devices, an embodiment of the present invention provides a method for identifying lesions in B-ultrasound images.

[0062] In this embodiment, the process of the method for identifying lesions in B-ultrasound images in this application is described in detail through steps S01-S04.

[0063] S01: Acquire a continuous multi-frame B-ultrasound image sequence in real time.

[0064] As an embodiment, the real-time acquisition of a continuous multi-frame B-ultrasound image sequence is specifically as follows:

[0065] In a primary medical institution, a mobile ultrasound diagnostic device based on ARM architecture, with 256MB of memory and no independent GPU was selected to achieve lesion detection.

[0066] The dynamic image data is acquired from the B-ultrasound probe at a frame sampling rate of F frames / second. The continuously acquired image sequence is denoted as I = {I1, I2, ..., I M}, M is the number of image frames cached simultaneously (it is recommended to take M ≥ 5);

[0067] Furthermore, the image sequence is preprocessed, and each frame of image I i , perform the following preprocessing steps:

[0068] (1) Noise reduction filtering: Use median filtering or adaptive filtering algorithm to achieve image noise reduction. The formula is as follows:

[0069] I′ i =MedianFilter(I i )

[0070] Where, Ι i is the input original B-ultrasound image of frame i, I′ i is the i-th frame image after median filtering, and MedianFilter(·) is a median filtering operation that replaces each pixel in the image with the median of all pixels in its neighborhood.

[0071] (2) Normalization: Normalize the image pixel values to the interval [0,1] to ensure the numerical stability of subsequent feature extraction. The normalization formula is described as:

[0072]

[0073] Where, Ι' i is the i-th frame image after denoising, Ι″ i (x, y) is the normalized image of the i-th frame, where (x, y) represents the pixel coordinates, Max(I′i ) is image 1' i The maximum pixel value in .

[0074] (3) Fixed image size: Each frame image is adjusted to a uniform size (such as 224×224 or other input size suitable for the network) to facilitate batch input into the network for calculation.

[0075] (4) Pseudo-color processing: When necessary (for certain specific lesion recognition algorithms), perform grayscale conversion on the pseudo-color image to reduce the processing workload.

[0076] Finally, the pre-processed image sequence I”={I1”,I2”,…,I M ”}.

[0077] The present application first obtains a continuous multi-frame B-ultrasound image sequence in real time, ensures the real-time and continuity of the data, which is crucial for the accuracy of lesion detection. Then, a preset noise reduction filter is used to effectively reduce the noise in the image and improve image quality, while the pixel value normalization process ensures that the image data is within a unified numerical range, thereby enhancing the stability and reliability of subsequent feature extraction. In addition, adjusting the image sequence to a preset resolution not only optimizes the display effect of the image, but also provides standardized input for the feature extraction of lightweight convolutional neural networks, which helps to improve the efficiency and accuracy of feature extraction. Combining these pre-processing steps, the beneficial effect is to enhance the quality of the image data, lay a solid foundation for subsequent feature extraction and lesion identification, thereby improving the accuracy and robustness of the entire lesion detection system.

[0078] S02: performing feature extraction on the B-ultrasound image sequence according to a preset lightweight convolutional neural network to obtain a plurality of consecutive single-frame feature maps.

[0079] As a preferred embodiment, the preset lightweight convolutional neural network is specifically:

[0080] In order to adapt to the computing resources of low-end devices, a lightweight convolutional neural network is used, and efficient modules such as Depthwise Separable Convolution and Shuffle Unit are used in the structure. The single-frame feature extraction model is set to CNN. Φ , whose input is I i ”, the output is the feature map The formula is described as follows:

[0081] f i =CNN φ (I i )

[0082] Where, I″ iis the i-th frame image after preprocessing, f i ∈R H×W×C is the feature map extracted from the i-th frame, where H is the height of the feature map, W is the width of the feature map, and C is the number of feature channels. φ (·) is a lightweight convolutional neural network model, and the parameter φ represents the network structure and weight.

[0083] During the design process, the following key parameters should be considered:

[0084] Table 1 Multi-frame feature extraction network parameters

[0085] Parameter name illustrate Reference value Convolution kernel size The filter size used in the convolution operation 3×3 or 5×5 Number of channels The number of channels in each layer’s output feature map 16 / 32 / 64, increasing according to the number of layers step length Convolution stride 1 or 2 Activation Function Common activation functions ReLU, ReLU6 Quantization bit width Model quantization parameters Int8

[0086] For low-end devices, it is recommended to control the network depth between 10 and 15 layers, and introduce model pruning and quantization algorithms to reduce memory usage and computing overhead.

[0087] S03: performing feature fusion on the single-frame feature maps of the continuous multiple frames based on temporal continuity, a preset fusion window, and preset weights to generate a plurality of fusion feature representations.

[0088] As a preferred embodiment, consecutive frames in a B-ultrasound image sequence reflect motion and structural changes over time, and single frames are subject to significant noise and local artifacts. Fusion of features extracted from multiple consecutive frames not only improves the robustness of lesion identification but also reduces the overall computational effort during the fusion process. This solution performs feature fusion processing on N adjacent frames (e.g., N = 3 or 5).

[0089] This application provides two main feature fusion methods:

[0090] A. Average Fusion Method:

[0091] The feature vectors of N consecutive frames are simply averaged, and the formula is:

[0092]

[0093] f i : Feature map extracted from the i-th frame.

[0094] N: The number of frames continuously involved in fusion.

[0095] F fuse : The fused feature map.

[0096] This method is simple to operate, but does not consider the problem of uneven weight of each frame information.

[0097] B. Weighted fusion method:

[0098] Determine the weight w based on the image quality, motion information or other indicators of each framei (w i ≥0, all w i The sum is 1), the formula is:

[0099]

[0100] Where, f i is the feature map extracted from the i-th frame, w i is the weight of the i-th frame, satisfying F fuse is the fused feature map.

[0101] Weight w i The calculation combines three indicators: local contrast, texture complexity, and motion amplitude, and dynamically assigns weights using the following formula:

[0102]

[0103] Among them, C i is the local contrast score of the i-th frame, T i is the texture complexity, M i is the motion amplitude, and α, β, γ are adjustable hyperparameters.

[0104] Traditional methods only rely on a single indicator (such as motion amplitude), while this method uses a comprehensive evaluation of multiple indicators to more accurately reflect the information value of each frame.

[0105] To address the computing power limitations of low-end devices, this application proposes a pre-calculated image quality score table. In actual deployment, the weight values can be directly obtained by looking up the table, avoiding the need to calculate complex indicators in real time and significantly reducing computing overhead.

[0106] The following is the detailed process of the multi-frame feature fusion module, which is divided into 5 sub-steps:

[0107] Sub-step 1: Feature normalization:

[0108] The feature f extracted for each frame i L2 normalization is performed to ensure that the features of each frame are consistent in numerical scale. The formula is:

[0109]

[0110] Where, f i is the feature map extracted from the i-th frame, f′ i is the feature map after L2 normalization, and ||·||2 is the L2 norm (Euclidean distance).

[0111] Sub-step 2: Weight calculation:

[0112] Calculate the weight w of each frame according to the local contrast, texture complexity or motion amplitude of each frame image i .

[0113] Specific algorithms available:

[0114]

[0115] Among them, ΔI i Difference operators can be used, for example:

[0116] ΔI i =I″ i -I″ i-1

[0117] (When i>1, otherwise w1 is set to a predetermined value)

[0118] Where, I″ i is the pre-processed image of the i-th frame, I″ i-1 is the pre-processed image of the i-1th frame, ΔI i is the difference image between the i-th frame and the previous frame.

[0119] Sub-step 3: Feature fusion:

[0120] Using the calculated weights of each frame, the normalized features are weighted fused to obtain the fused feature F fuse :

[0121]

[0122] Where, f i is the feature map extracted from the i-th frame, w i is the weight of the i-th frame, satisfying F fuse is the fused feature map.

[0123] Or take a simple average:

[0124]

[0125] Where, f i is the feature map extracted from the i-th frame, N is the number of frames continuously involved in fusion, F fuse is the fused feature map.

[0126] Sub-step 4: Feature Dimensionality Reduction (optional):

[0127] Since the fused features may have a high dimension, in order to adapt to the subsequent reasoning module, PCA or 1×1 convolution can be used for dimensionality reduction.

[0128] F output =Conv 1×1(F fuse )

[0129] Where, F fuse is the fused feature map, Conv 1×1 (·) is a 1x1 convolution operation used to reduce the feature dimension.

[0130] Sub-step 5: Fusion feature cache and loop update:

[0131] In order to keep the system in real-time continuous reasoning state, the fusion feature F output The data is cached for reference in the next fusion calculation to form a sliding fusion window.

[0132] To achieve real-time processing, a sliding window mechanism is employed. That is, at each step, a fixed-size window is selected from the continuous frame sequence for fusion, then the window is shifted back one frame, and the process is repeated. To cope with rapidly changing image content (such as rapid motion or sudden noise), a mechanism for dynamically adjusting the fusion window is introduced. The basic idea is to reduce the fusion window size when a significant change in the image scene is detected; conversely, increase the fusion window size. The following formula is used to calculate the fusion window size for the current frame:

[0133] N t =round(N0·e -λ·ΔI )

[0134] Where ΔI is the difference between the current frame and the previous frame, and λ is the attenuation coefficient. When the scene changes dramatically (ΔI is large), the window size is automatically reduced (for example, from 5 frames to 3 frames) to improve real-time performance; otherwise, the window size is increased to enhance robustness.

[0135] The traditional method uses a fixed window and cannot balance the accuracy and speed requirements in dynamic scenes. This application first considers the temporal continuity of the B-ultrasound image sequence, uses the three key indicators of preset local contrast, texture complexity and motion amplitude, and combines the weight distribution formula to calculate the weight of the feature map of each frame. This step reasoning process ensures that the fusion process can be differentiated according to the actual importance of the image content. Subsequently, the single-frame feature maps of multiple consecutive frames are weightedly fused within the fusion window using a preset weighted fusion formula to generate a fused feature representation. This fusion strategy not only improves the richness and accuracy of the feature representation, but also adapts to changes in image content by dynamically adjusting the fusion window size, thereby enhancing the robustness of the system and the accuracy of lesion recognition while ensuring real-time performance. Therefore, the beneficial effect is that the quality and efficiency of lesion detection are improved through precise feature fusion, making it possible to achieve efficient and accurate lesion recognition on low-end mobile edge devices.

[0136] S04: Based on the multiple fusion feature representations, lesion identification is performed through a preset lightweight inference network to obtain lesion identification results of multiple consecutive frames of B-ultrasound images; wherein, the preset lightweight inference network is obtained by quantizing and pruning a preset initial lightweight inference network according to a preset optimization formula.

[0137] As a preferred embodiment, the lesion identification is performed by using a preset lightweight inference network, specifically:

[0138] The fused feature F output As input, it enters the lightweight inference network for lesion classification or segmentation. The network design can be customized based on existing MobileNet, SqueezeNet, ShuffleNet, etc. The network structure example is as follows:

[0139] (1) Input layer: F output The dimensions are [H', W', C'];

[0140] (2) Middle layer: Several lightweight convolution blocks, each of which includes depthwise separable convolution, BatchNormalization, and ReLU activation function;

[0141] (3) Global average pooling layer: converts the feature map into a one-dimensional vector;

[0142] (4) Fully connected classification layer: outputs the probability vector of the lesion category or the prediction map of the segmented area.

[0143] Let the inference network be NET_Ψ, the formula is:

[0144] Y=NET_Ψ(F output )

[0145] Among them, Y is the final lesion identification result.

[0146] To address the limited computing power of low-end mobile edge devices, the network NET_Ψ has been streamlined in terms of both number of layers and width. Furthermore, the network model can be quantized and pruned to reduce the number of parameters and improve inference speed.

[0147] Algorithm optimization formulas are as follows:

[0148] FLOPs_reduction = original_FLOPs × (pruning ratio) × (quantization efficiency)

[0149] To meet the real-time requirements, this solution designs a mechanism to dynamically adjust the fusion window so that when the image content changes significantly, the number of frame fusions is automatically shortened (for example, from 5 frames to 3 frames) to balance the amount of computation and accuracy.

[0150] The initial lightweight inference network of this application is designed as an efficient structure to reduce the amount of computation (FLOPs). Then, by applying the optimization formula, the network is further pruned and quantized, where the pruning ratio coefficient (a) and quantization efficiency (b) are pre-set parameters used to guide network optimization. The pruning process removes unnecessary connections in the network, reducing the complexity and computational requirements of the model, while the quantization process converts the network parameters from high-precision representation to low-precision, further reducing the amount of computation. The beneficial effect is that this method not only significantly reduces the computational load of the model, but also improves the inference speed while maintaining the accuracy of the model, making the optimized lightweight inference network particularly suitable for running on low-end mobile edge devices with limited computing resources, thereby achieving efficient lesion identification without sacrificing too much accuracy.

[0151] Furthermore, the result Y output by the inference module may be a probability map or region annotation of the lesion. The post-processing module adopts the following measures:

[0152] (1) Thresholding operation: Binarize the probability map using a fixed or adaptive threshold:

[0153] B(x,y)=1if Y(x,y)>T else 0

[0154] (2) Morphological filtering: using operations such as dilation and corrosion to eliminate noise and isolated areas;

[0155] (3) Boundary smoothing: Use Gaussian filtering or median filtering to smooth the edges to ensure the continuity of the detection area.

[0156] The post-processed image results are directly presented on the mobile edge device display and can also be wirelessly transmitted to a central server for further evaluation.

[0157] Furthermore, the lesion recognition method of the present application can reduce the amount of calculation, because the FLOPs required for single-frame feature extraction is F CNN , the total FLOPs of the traditional frame-by-frame method under N frames is about N×F CNN ; In the multi-frame fusion method of this application, only single-frame CNN extraction (or parallel calculation) plus fusion operation is required, and the fusion calculation amount F fuse Much smaller than F CNN Assume that the fusion calculation amount is approximately α×F CNN (α<<1), the total amount of calculation is about F total =N×F CNN +α×F CNN .

[0158] This application uses the fusion module to effectively eliminate data redundancy. In practice, α can be controlled below 0.1, thereby reducing the overall computational complexity by a ratio of approximately (1-0.1 / N).

[0159] As a preferred embodiment, test results show that compared with the traditional frame-by-frame reasoning method, after implementing the present application solution:

[0160] (1) The average inference latency is reduced by approximately 80%;

[0161] (2) Energy consumption is reduced by about 50%, and it can still operate stably under low power consumption conditions;

[0162] (3) The recall rate of lesion detection increased by approximately 10%, and the false alarm rate decreased by approximately 5%.

[0163] Table 2 Experimental comparison data

[0164]

[0165]

[0166] The above data proves that in typical practical applications, the method of this application can achieve efficient and accurate lesion detection on low-end mobile edge devices, while saving hardware resources and supporting real-time medical applications.

[0167] The present application obtains a series of continuous multi-frame B-ultrasound images in real time and uses a preset lightweight convolutional neural network for feature extraction, effectively capturing the key information in the image sequence. Subsequently, a fusion strategy based on temporal continuity is adopted, combined with a dynamically adjusted fusion window and preset weights, to fuse the single-frame feature maps to generate a fused feature representation. This process not only makes full use of the temporal redundant information of the image sequence and reduces the computational complexity, but also enhances the robustness of the feature representation through weight adjustment. Finally, the lightweight inference network obtained by quantization and pruning optimization is used to identify lesions on the fused features, which not only improves the accuracy and real-time performance of the recognition, but also significantly reduces energy consumption, making this method particularly suitable for resource-constrained low-end mobile edge devices. Therefore, the beneficial effect of the present application is that it achieves a significant reduction in computational complexity and energy consumption while ensuring detection accuracy, and improves the adaptability and robustness of the system in dynamically changing scenarios. The present application effectively solves the problem that the existing technology cannot accurately and quickly identify lesions in B-ultrasound images on low-end devices.

[0168] Example 2

[0169] Please refer to Figure 2 , which is a lesion recognition device for B-ultrasound images provided in an embodiment of the present application.

[0170] In this embodiment, the device for identifying lesions based on B-ultrasound images includes an acquisition module 10 , a feature extraction module 20 , a feature fusion module 30 and a lesion identification module 40 .

[0171] The acquisition module 10 is used to acquire a continuous multi-frame B-ultrasound image sequence in real time.

[0172] As an embodiment, the real-time acquisition of a continuous multi-frame B-ultrasound image sequence is specifically as follows:

[0173] In a primary medical institution, a mobile ultrasound diagnostic device based on ARM architecture, with 256MB of memory and no independent GPU was selected to achieve lesion detection.

[0174] The dynamic image data is acquired from the B-ultrasound probe at a frame sampling rate of F frames / second. The continuously acquired image sequence is denoted as I = {I1, I2, ..., I M}, M is the number of image frames cached simultaneously (it is recommended to take M ≥ 5);

[0175] Furthermore, the image sequence is preprocessed, and each frame of image I i , perform the following preprocessing steps:

[0176] (1) Noise reduction filtering: Use median filtering or adaptive filtering algorithm to achieve image noise reduction. The formula is as follows:

[0177] I′ i =MedianFilter(I i )

[0178] Where, Ι i is the input i-th frame original B-ultrasound image, I' i is the i-th frame image after median filtering, and MedianFilter(·) is a median filtering operation that replaces each pixel in the image with the median of all pixels in its neighborhood.

[0179] (2) Normalization: Normalize the image pixel values to the interval [0,1] to ensure the numerical stability of subsequent feature extraction. The normalization formula is described as:

[0180]

[0181] Where, Ι' i is the i-th frame image after denoising, Ι″ i (x, y) is the normalized image of the i-th frame, where (x, y) represents the pixel coordinates, Max(I′ i ) is image 1' i The maximum pixel value in .

[0182] (3) Fixed image size: Each frame image is adjusted to a uniform size (such as 224×224 or other input size suitable for the network) to facilitate batch input into the network for calculation.

[0183] (4) Pseudo-color processing: When necessary (for certain specific lesion recognition algorithms), perform grayscale conversion on the pseudo-color image to reduce the processing workload.

[0184] Finally, the pre-processed image sequence I”={I1”,I2”,…,I M ”}.

[0185] The present application first obtains a continuous multi-frame B-ultrasound image sequence in real time, ensures the real-time and continuity of the data, which is crucial for the accuracy of lesion detection. Then, a preset noise reduction filter is used to effectively reduce the noise in the image and improve image quality, while the pixel value normalization process ensures that the image data is within a unified numerical range, thereby enhancing the stability and reliability of subsequent feature extraction. In addition, adjusting the image sequence to a preset resolution not only optimizes the display effect of the image, but also provides standardized input for the feature extraction of lightweight convolutional neural networks, which helps to improve the efficiency and accuracy of feature extraction. Combining these pre-processing steps, the beneficial effect is to enhance the quality of the image data, lay a solid foundation for subsequent feature extraction and lesion identification, thereby improving the accuracy and robustness of the entire lesion detection system.

[0186] The feature extraction module 20 is used to extract features from the B-ultrasound image sequence according to a preset lightweight convolutional neural network to obtain a plurality of continuous single-frame feature maps.

[0187] As a preferred embodiment, the preset lightweight convolutional neural network is specifically:

[0188] In order to adapt to the computing resources of low-end devices, a lightweight convolutional neural network is used, and efficient modules such as Depthwise Separable Convolution and Shuffle Unit are used in the structure. The single-frame feature extraction model is set to CNN. Φ , whose input is I i ”, the output is the feature map The formula is described as follows:

[0189] f i =CNN φ (I i )

[0190] Where, I″ i is the i-th frame image after preprocessing, f i ∈R H×W×Cis the feature map extracted from the i-th frame, where H is the height of the feature map, W is the width of the feature map, and C is the number of feature channels. φ (·) is a lightweight convolutional neural network model, and the parameter φ represents the network structure and weight.

[0191] During the design process, the following key parameters should be considered:

[0192] Table 3 Multi-frame feature extraction network parameters

[0193] Parameter name illustrate Reference value Convolution kernel size The filter size used in the convolution operation 3×3 or 5×5 Number of channels The number of channels in each layer’s output feature map 16 / 32 / 64, increasing according to the number of layers step length Convolution stride 1 or 2 Activation Function Common activation functions ReLU, ReLU6 Quantization bit width Model quantization parameters Int8

[0194] For low-end devices, it is recommended to control the network depth between 10 and 15 layers, and introduce model pruning and quantization algorithms to reduce memory usage and computing overhead.

[0195] The feature fusion module 30 is used to perform feature fusion on the single-frame feature maps of the continuous multiple frames based on temporal continuity, a preset fusion window and preset weights to generate a plurality of fused feature representations.

[0196] As a preferred embodiment, consecutive frames in a B-ultrasound image sequence reflect motion and structural changes over time, and single frames are subject to significant noise and local artifacts. Fusion of features extracted from multiple consecutive frames not only improves the robustness of lesion identification but also reduces the overall computational effort during the fusion process. This solution performs feature fusion processing on N adjacent frames (e.g., N = 3 or 5).

[0197] This application provides two main feature fusion methods:

[0198] A. Average Fusion Method:

[0199] The feature vectors of N consecutive frames are simply averaged, and the formula is:

[0200]

[0201] f i : Feature map extracted from the i-th frame.

[0202] N: The number of frames continuously involved in fusion.

[0203] F fuse : The fused feature map.

[0204] This method is simple to operate, but does not consider the problem of uneven weight of each frame information.

[0205] B. Weighted fusion method:

[0206] Determine the weight w based on the image quality, motion information or other indicators of each frame i (w i ≥0, all wi The sum is 1), the formula is:

[0207]

[0208] Where, f i is the feature map extracted from the i-th frame, w i is the weight of the i-th frame, satisfying F fuse is the fused feature map.

[0209] Weight w i The calculation combines three indicators: local contrast, texture complexity, and motion amplitude, and dynamically assigns weights using the following formula:

[0210]

[0211] Among them, C i is the local contrast score of the i-th frame, T i is the texture complexity, M i is the motion amplitude, and α, β, γ are adjustable hyperparameters.

[0212] Traditional methods only rely on a single indicator (such as motion amplitude), while this method uses a comprehensive evaluation of multiple indicators to more accurately reflect the information value of each frame.

[0213] To address the computing power limitations of low-end devices, this application proposes a pre-calculated image quality score table. In actual deployment, the weight values can be directly obtained by looking up the table, avoiding the need to calculate complex indicators in real time and significantly reducing computing overhead.

[0214] The following is the detailed process of the multi-frame feature fusion module, which is divided into 5 sub-steps:

[0215] Sub-step 1: Feature normalization:

[0216] The feature f extracted for each frame i L2 normalization is performed to ensure that the features of each frame are consistent in numerical scale. The formula is:

[0217]

[0218] Where, f i is the feature map extracted from the i-th frame, f′ i is the feature map after L2 normalization, and ‖·‖2 is the L2 norm (Euclidean distance).

[0219] Sub-step 2: Weight calculation:

[0220] Calculate the weight w of each frame according to the local contrast, texture complexity or motion amplitude of each frame image i .

[0221] Specific algorithms available:

[0222]

[0223] Among them, ΔI i Difference operators can be used, for example:

[0224] ΔI i =I″ i -I″ i-1

[0225] (When i>1, otherwise w1 is set to a predetermined value)

[0226] Where, I″ i is the pre-processed image of the i-th frame, I″ i-1 is the pre-processed image of the i-1th frame, ΔI i is the difference image between the i-th frame and the previous frame.

[0227] Sub-step 3: Feature fusion:

[0228] Using the calculated weights of each frame, the normalized features are weighted fused to obtain the fused feature F fuse :

[0229]

[0230] Where, f i is the feature map extracted from the i-th frame, w i is the weight of the i-th frame, satisfying F fuse is the fused feature map.

[0231] Or take a simple average:

[0232]

[0233] Where, f i is the feature map extracted from the i-th frame, N is the number of frames continuously involved in fusion, F fuse is the fused feature map.

[0234] Sub-step 4: Feature Dimensionality Reduction (optional):

[0235] Since the fused features may have a high dimension, in order to adapt to the subsequent reasoning module, PCA or 1×1 convolution can be used for dimensionality reduction.

[0236] F output =Conv 1×1 (F fuse )

[0237] Where, F fuseis the fused feature map, Conv 1×1 (·) is a 1x1 convolution operation used to reduce the feature dimension.

[0238] Sub-step 5: Fusion feature cache and loop update:

[0239] In order to keep the system in real-time continuous reasoning state, the fusion feature F output The data is cached for reference in the next fusion calculation to form a sliding fusion window.

[0240] To achieve real-time processing, a sliding window mechanism is employed. That is, at each step, a fixed-size window is selected from the continuous frame sequence for fusion, then the window is shifted back one frame, and the process is repeated. To cope with rapidly changing image content (such as rapid motion or sudden noise), a mechanism for dynamically adjusting the fusion window is introduced. The basic idea is to reduce the fusion window size when a significant change in the image scene is detected; conversely, increase the fusion window size. The following formula is used to calculate the fusion window size for the current frame:

[0241] N t =round(N0·e -λ·ΔI )

[0242] Where ΔI is the difference between the current frame and the previous frame, and λ is the attenuation coefficient. When the scene changes dramatically (ΔI is large), the window size is automatically reduced (for example, from 5 frames to 3 frames) to improve real-time performance; otherwise, the window size is increased to enhance robustness.

[0243] The traditional method uses a fixed window and cannot balance the accuracy and speed requirements in dynamic scenes. This application first considers the temporal continuity of the B-ultrasound image sequence, uses the three key indicators of preset local contrast, texture complexity and motion amplitude, and combines the weight distribution formula to calculate the weight of the feature map of each frame. This step reasoning process ensures that the fusion process can be differentiated according to the actual importance of the image content. Subsequently, the single-frame feature maps of multiple consecutive frames are weightedly fused within the fusion window using a preset weighted fusion formula to generate a fused feature representation. This fusion strategy not only improves the richness and accuracy of the feature representation, but also adapts to changes in image content by dynamically adjusting the fusion window size, thereby enhancing the robustness of the system and the accuracy of lesion recognition while ensuring real-time performance. Therefore, the beneficial effect is that the quality and efficiency of lesion detection are improved through precise feature fusion, making it possible to achieve efficient and accurate lesion recognition on low-end mobile edge devices.

[0244] The lesion recognition module 40 is used to perform lesion recognition based on the multiple fusion feature representations through a preset lightweight inference network to obtain lesion recognition results of multiple consecutive frames of B-ultrasound images; wherein, the preset lightweight inference network is obtained by quantizing and pruning the preset initial lightweight inference network according to a preset optimization formula.

[0245] As a preferred embodiment, the lesion identification is performed by using a preset lightweight inference network, specifically:

[0246] The fused feature F output As input, it enters the lightweight inference network for lesion classification or segmentation. The network design can be customized based on existing MobileNet, SqueezeNet, ShuffleNet, etc. The network structure example is as follows:

[0247] (1) Input layer: F output The dimensions are [H', W', C'];

[0248] (2) Middle layer: Several lightweight convolution blocks, each of which includes depthwise separable convolution, BatchNormalization, and ReLU activation function;

[0249] (3) Global average pooling layer: converts the feature map into a one-dimensional vector;

[0250] (4) Fully connected classification layer: outputs the probability vector of the lesion category or the prediction map of the segmented area.

[0251] Let the inference network be NET_Ψ, the formula is:

[0252] Y=NET_Ψ(F output )

[0253] Among them, Y is the final lesion identification result.

[0254] To address the limited computing power of low-end mobile edge devices, the network NET_Ψ has been streamlined in terms of both number of layers and width. Furthermore, the network model can be quantized and pruned to reduce the number of parameters and improve inference speed.

[0255] Algorithm optimization formulas are as follows:

[0256] FLOPs_reduction = original_FLOPs × (pruning ratio) × (quantization efficiency)

[0257] To meet the real-time requirements, this solution designs a mechanism to dynamically adjust the fusion window so that when the image content changes significantly, the number of frame fusions is automatically shortened (for example, from 5 frames to 3 frames) to balance the amount of computation and accuracy.

[0258] The initial lightweight inference network of this application is designed as an efficient structure to reduce the amount of computation (FLOPs). Then, by applying the optimization formula, the network is further pruned and quantized, where the pruning ratio coefficient (a) and quantization efficiency (b) are pre-set parameters used to guide network optimization. The pruning process removes unnecessary connections in the network, reducing the complexity and computational requirements of the model, while the quantization process converts the network parameters from high-precision representation to low-precision, further reducing the amount of computation. The beneficial effect is that this method not only significantly reduces the computational load of the model, but also improves the inference speed while maintaining the accuracy of the model, making the optimized lightweight inference network particularly suitable for running on low-end mobile edge devices with limited computing resources, thereby achieving efficient lesion identification without sacrificing too much accuracy.

[0259] Furthermore, the result Y output by the inference module may be a probability map or region annotation of the lesion. The post-processing module adopts the following measures:

[0260] (1) Thresholding operation: Binarize the probability map using a fixed or adaptive threshold:

[0261] B(x,y)=1if Y(x,y)>T else 0

[0262] (2) Morphological filtering: using operations such as dilation and corrosion to eliminate noise and isolated areas;

[0263] (3) Boundary smoothing: Use Gaussian filtering or median filtering to smooth the edges to ensure the continuity of the detection area.

[0264] The post-processed image results are directly presented on the mobile edge device display and can also be wirelessly transmitted to a central server for further evaluation.

[0265] Furthermore, the lesion recognition method of the present application can reduce the amount of calculation, because the FLOPs required for single-frame feature extraction is F CNN , the total FLOPs of the traditional frame-by-frame method under N frames is about N×F CNN ; In the multi-frame fusion method of this application, only single-frame CNN extraction (or parallel calculation) plus fusion operation is required, and the fusion calculation amount F fuse Much smaller than F CNN Assume that the fusion calculation amount is approximately α×F CNN (α<<1), the total amount of calculation is about F total =N×F CNN +α×F CNN .

[0266] This application uses the fusion module to effectively eliminate data redundancy. In practice, α can be controlled below 0.1, thereby reducing the overall computational complexity by a ratio of approximately (1-0.1 / N).

[0267] As a preferred embodiment, test results show that compared with the traditional frame-by-frame reasoning method, after implementing the present application solution:

[0268] (1) The average inference latency is reduced by approximately 80%;

[0269] (2) Energy consumption is reduced by about 50%, and it can still operate stably under low power consumption conditions;

[0270] (3) The recall rate of lesion detection increased by approximately 10%, and the false alarm rate decreased by approximately 5%.

[0271] Table 4 Experimental comparison data

[0272] index Traditional frame-by-frame reasoning method Multi-frame fusion method of the present invention Average inference latency 200ms / frame 25ms / fusion window Energy consumption per second 3w 1.5w Detection accuracy (F1 score) 0.82 0.9

[0273] The above data proves that in typical practical applications, the method of this application can achieve efficient and accurate lesion detection on low-end mobile edge devices, while saving hardware resources and supporting real-time medical applications.

[0274] This application achieves efficient processing and accurate recognition of B-ultrasound image sequences through an integrated device design. First, the acquisition module is responsible for capturing multiple consecutive frames of B-ultrasound images in real time, ensuring the real-time and continuity of the data; then, the feature extraction module uses a lightweight convolutional neural network to extract features from these images, effectively capturing key information and reducing the computational load; then, the feature fusion module uses temporal continuity, fusion windows and preset weights to fuse single-frame feature maps, which not only enhances the robustness of the features, but also reduces the amount of computation by reducing redundant information; finally, the lesion recognition module uses a lightweight inference network optimized by quantization and pruning to identify the fused features, thereby improving the speed and accuracy of recognition. Overall, the device significantly improves the processing speed and reduces resource consumption while ensuring recognition accuracy through the collaborative work of various modules. It is particularly suitable for low-end mobile edge devices, thus showing important application value in the field of medical image analysis, especially in scenarios where fast and accurate diagnosis is required.

[0275] Example 3:

[0276] An embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the method for identifying lesions in B-ultrasound images;

[0277] Wherein, the method for identifying lesions in B-ultrasound images, if implemented in the form of a software functional unit and used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing related hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by a processor, it can implement the steps of the above-mentioned various method embodiments. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0278] Example 4

[0279] The present application provides a terminal device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements any one of the B-ultrasound image lesion recognition methods described in Example 1.

[0280] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for identifying lesions in B-ultrasound images, characterized in that: include: Acquire multiple continuous frames of B-ultrasound image sequences in real time; Performing feature extraction on the B-ultrasound image sequence according to a preset lightweight convolutional neural network to obtain a plurality of single-frame feature maps of consecutive multiple frames; Performing feature fusion on the single-frame feature maps of the continuous multiple frames based on temporal continuity, a preset fusion window, and preset weights to generate a plurality of fused feature representations; Based on the plurality of fused feature representations, lesion recognition is performed through a preset lightweight inference network to obtain lesion recognition results of multiple consecutive frames of B-ultrasound images; The preset lightweight inference network is obtained by quantizing and pruning a preset initial lightweight inference network according to a preset optimization formula.

2. The method for identifying lesions using B-ultrasound images according to claim 1, wherein: After acquiring a continuous multi-frame B-ultrasound image sequence in real time, the method further includes: According to the preset noise reduction filtering and pixel value normalization processing methods, the continuous multi-frame B-ultrasound image sequence is preprocessed and adjusted to a preset resolution.

3. The method for identifying lesions using B-ultrasound images according to claim 1, wherein: The feature fusion of the single-frame feature maps of the continuous multiple frames is performed based on temporal continuity, a preset fusion window, and preset weights, specifically: According to the preset local contrast, the preset texture complexity and the preset motion amplitude, combined with the preset weight distribution formula, the preset weights are calculated; According to the preset weights and the preset weighted fusion formula, weighted feature fusion is performed on the single-frame feature maps of the continuous multiple frames in a preset fusion window to obtain a plurality of fused feature representations.

4. The method for identifying lesions using B-ultrasound images according to claim 3, wherein: The preset weight distribution formula and the preset weighted fusion formula are specifically: The preset weight distribution formula is: Where C i is the local contrast score of the i-th frame, T i is the texture complexity, M i is the motion amplitude, α, β, γ are preset adjustable hyper parameters, w i For each preset weight; The preset weighted fusion formula is: Where N is a preset positive integer greater than 1, f i is the feature map extracted from the i-th frame, w i is the weight of the i-th frame, F fuse It is the fusion feature representation.

5. The method for identifying lesions using B-ultrasound images according to claim 1, wherein: The step of fusing the single-frame feature maps of the consecutive multiple frames based on temporal continuity, a preset fusion window, and preset weights to generate a plurality of fused feature representations further includes: monitoring the difference between consecutive frames of the acquired multiple-frame B-ultrasound image sequence; If the difference is higher than a preset threshold, dynamically adjusting the size of the preset fusion window according to a preset fusion window adjustment formula; The preset fusion window adjustment formula is: N t =round(N o ·e -λ·ΔI ) Where ΔI is the difference between the current frame and the previous frame, λ is the attenuation coefficient, and N t is the adjusted fusion window size, N o The size of the preset fusion window.

6. The method for identifying lesions using B-ultrasound images according to claim 1, wherein: The preset initial lightweight inference network is quantized and pruned according to the preset optimization formula, specifically: The preset optimization formula is: FLOPs_reduction=original_FLOPs·a·b Where FLOPs_reduction is the preset lightweight inference network, original_FLOPs is the preset initial lightweight inference network, a is the preset pruning ratio coefficient, and b is the preset quantization efficiency.

7. The method for identifying lesions using B-ultrasound images according to claim 1, wherein: The step of fusing the single-frame feature maps of the continuous multiple frames based on temporal continuity, a preset fusion window, and preset weights further includes: According to a preset average fusion formula, the single-frame feature maps of the continuous multiple frames are fused; The preset average fusion formula is: Where, f i is the single-frame feature map extracted from the i-th frame, N is the number of frames continuously involved in fusion, and F fuse is the fused feature map.

8. A device for identifying lesions using B-ultrasound images, characterized in that: include: An acquisition module is used to acquire a continuous multi-frame B-ultrasound image sequence in real time; A feature extraction module is used to extract features from the B-ultrasound image sequence according to a preset lightweight convolutional neural network to obtain a plurality of single-frame feature maps of consecutive multiple frames; A feature fusion module is used to fuse the single-frame feature maps of the continuous multiple frames based on temporal continuity, a preset fusion window and preset weights to generate a plurality of fused feature representations; A lesion recognition module is used to perform lesion recognition based on the plurality of fused feature representations through a preset lightweight inference network to obtain lesion recognition results for multiple consecutive frames of B-ultrasound images; The preset lightweight inference network is obtained by quantizing and pruning a preset initial lightweight inference network according to a preset optimization formula.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the method for identifying lesions in B-ultrasound images according to any one of claims 1 to 7.

10. A terminal device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the method for identifying lesions in B-ultrasound images according to any one of claims 1 to 7 is implemented.

Images