A B-ultrasound image lesion identification method, device, medium and equipment

Abstract

The application discloses a kind of B ultrasound image's lesion identification method, device, medium and equipment.The application is first obtained by real-time acquisition technology continuous multiple frames B ultrasound image sequence, then utilize a pre-set lightweight convolutional neural network to these images feature extraction, to obtain multiple continuous frame single frame feature map.Next, these single frame feature map will be based on the time sequence continuity of image, a pre-set fusion window and a group of preset weight feature fusion, to generate multiple fusion feature representation.Finally, utilize a lightweight inference network that is optimized by quantization and pruning, to these fusion feature representation carries out lesion identification, to obtain the lesion identification result of continuous multiple frames B ultrasound image.This process not only improves the accuracy and real-time of identification, but also optimizes the energy consumption and computing efficiency of low-end mobile edge device.

Description

TECHNICAL FIELD

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

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

[0003] Large amount of single-frame inference calculation: the existing technology usually extracts features and makes predictions for each frame of B-ultrasound image based on convolutional neural network (CNN), resulting in a calculation amount proportional to the frame rate, which is difficult to realize real-time inference on low-end devices, especially when the image frame rate is high, the cumulative calculation amount increases significantly, which easily leads to delay and frame loss.

[0004] Lack of data redundancy utilization: B-ultrasound image sequences have temporal continuity, and the overall structure is similar between adjacent frames except for slight motion. The existing technology fails to fully utilize this data redundancy characteristic, but instead repeatedly calculates the information overlapping area independently, resulting in resource waste.

[0005] Low energy consumption and hardware resource utilization: mobile edge devices have limited energy and computing power, and frame-by-frame complex inference inevitably leads to a significant increase in energy consumption, reducing device battery life, and the inference speed is difficult to meet the clinical real-time detection requirements when hardware resources are limited.

[0006] Insufficient model adaptability: in order to solve the single-frame inference problem with high-precision network, the existing technology often uses large-scale parameter models, which are difficult to deploy on low-end devices, and it is difficult to achieve model quantization, pruning and compression while considering accuracy and inference speed.

[0007] These deficiencies result in the inability of the existing technology to accurately and quickly recognize lesions in B-ultrasound images on low-end devices. SUMMARY

[0008] The present application provides a lesion recognition method for B-ultrasound images to solve the problem that the existing technology cannot accurately and quickly recognize lesions in B-ultrasound images on low-end devices.

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

[0010] real-time acquisition of a continuous multi-frame B-ultrasound image sequence;

[0011] feature extraction of the B-ultrasound image sequence according to a preset lightweight convolutional neural network, to obtain a plurality of continuous multi-frame single-frame feature maps;

[0012] feature fusion of the continuous multi-frame single-frame feature maps based on time sequence continuity, a preset fusion window, and preset weights, to generate a plurality of fusion feature representations;

[0013] lesion identification of the continuous multi-frame B-ultrasound image sequence based on the plurality of fusion feature representations through a preset lightweight inference network, to obtain a lesion identification result of the continuous multi-frame B-ultrasound image sequence.

[0014] wherein the preset lightweight inference network is obtained by quantization and pruning of a preset initial lightweight inference network according to a preset optimization formula.

[0015] The present application effectively captures the key information in the image sequence by real-time acquisition of a continuous multi-frame B-ultrasound image sequence and feature extraction using a preset lightweight convolutional neural network. Subsequently, a fusion strategy based on time sequence continuity is adopted, combined with a dynamically adjusted fusion window and preset weights, to fuse single-frame feature maps and generate fusion feature representations. This process not only fully utilizes the temporal redundancy information of the image sequence, reducing the computational complexity, but also enhances the robustness of the feature representation through weight adjustment. Finally, the fusion features are identified by a lightweight inference network optimized by quantization and pruning, which not only improves the accuracy and real-time performance of the identification, but also significantly reduces the energy consumption, making the method particularly suitable for low-end mobile edge devices with limited resources. Therefore, the beneficial effects of the present application are to realize a significant reduction in computational load and energy consumption while ensuring detection accuracy, improving the adaptability and robustness of the system in dynamic 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] Further, after the real-time acquisition of the continuous multi-frame B-ultrasound image sequence, the method further comprises:

[0017] preprocessing and adjusting the continuous multi-frame B-ultrasound image sequence to a preset resolution according to a preset denoising filter processing and pixel value normalization processing method.

[0018] The application firstly acquires a continuous multi-frame B-ultrasound image sequence in real time, ensuring the real-time and continuity of the data, which is crucial for the accuracy of lesion detection. Then, a preset noise reduction filtering process is adopted to effectively reduce the noise in the image and improve the image quality. The pixel value normalization process ensures that the image data is within a unified numerical range, 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 the lightweight convolutional neural network, which helps to improve the efficiency and accuracy of feature extraction. The comprehensive pre-processing steps enhance the quality of image data, laying a solid foundation for subsequent feature extraction and lesion recognition, thereby improving the accuracy and robustness of the entire lesion detection system.

[0019] Further, the single-frame feature maps of the continuous multi-frames are fused based on time sequence continuity, a preset fusion window and preset weights, specifically:

[0020] According to the preset local contrast, preset texture complexity and preset motion amplitude, and combining a preset weight distribution formula, preset weights are calculated.

[0021] According to the preset weights and a preset weighted fusion formula, weighted feature fusion is performed on the single-frame feature maps of the continuous multi-frames in the preset fusion window, obtaining a plurality of fusion 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] In the formula, 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 a, b, g are preset adjustable hyperparameters, w i is the preset weight.

[0026] The preset weighted fusion formula is:

[0027]

[0028] In the formula, 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, and F fuse is the fusion feature representation.

[0029] The application firstly considers the time sequence continuity of the B-ultrasound image sequence, calculates the weight of each frame feature map by using three key indicators of preset local contrast, texture complexity and motion amplitude, and combining a weight distribution formula. This step ensures that the image content can be processed differently according to the actual importance during the fusion process. Subsequently, the preset weighted fusion formula is used to perform weighted fusion on the single frame feature maps of the continuous multiple frames in the fusion window to generate a fusion feature representation. This fusion strategy not only improves the richness and accuracy of the feature representation, but also adjusts the size of the fusion window dynamically to adapt to the changes in image content, thereby ensuring real-time performance while enhancing the robustness of the system and the accuracy of lesion identification. Therefore, the beneficial effect is to improve the quality and efficiency of lesion detection by precise feature fusion, making it possible to achieve efficient and accurate lesion identification on low-end mobile edge devices.

[0030] Further, the feature fusion of the single frame feature maps of the continuous multiple frames based on time sequence continuity, a preset fusion window and preset weights to generate a plurality of fusion feature representations further comprises:

[0031] monitoring the difference degree between the continuous frames of the obtained B-ultrasound image sequence of the continuous multiple frames;

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

[0033] wherein the preset fusion window adjustment formula is:

[0034] N t = round(N o ·e -λ·ΔI )

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

[0036] The application realizes the adaptability to the changes in the image sequence by monitoring the difference degree between the continuous multiple frames of the B-ultrasound image sequence and dynamically adjusting the size of the fusion window using the fusion window adjustment formula when the difference degree is higher than the preset threshold. Firstly, the single frame feature maps of the continuous multiple frames are weighted and fused by using the preset weight to generate a fusion feature representation, which helps to utilize the time sequence information comprehensively and improve the expression ability of the features. Secondly, by monitoring the frame difference degree and dynamically adjusting the fusion window according to the difference degree, the changes in the image content such as rapid motion or noise can be flexibly handled, thereby maintaining the accuracy and real-time performance of feature fusion in different scenarios.

[0037] Further, 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] In the formula, FLOPs_reduction is the preset lightweight inference network, original_FLOPs is the preset initial lightweight inference network, a is a preset pruning ratio coefficient, and b is a preset quantization efficiency.

[0041] The initial lightweight inference network of the present application is designed as an efficient structure, aiming to reduce the amount of calculation (FLOPs). Then, by applying the optimization formula, the network is further pruned and quantized, where the pruning ratio coefficient (a) and the quantization efficiency (b) are preset parameters for guiding network optimization. The pruning process removes unnecessary connections in the network, reducing the complexity and computational demand of the model, while the quantization process converts the network parameters from high-precision representation to low-precision, further reducing the amount of calculation. 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, so that the optimized lightweight inference network is 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] Further, the single-frame feature map of the continuous multiple frames is fused based on the time sequence continuity, a preset fusion window, and a preset weight for each, which further includes:

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

[0044] The preset average fusion formula is:

[0045]

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

[0047] The application extracts single-frame feature maps from each frame of the continuously acquired multiple frames of B-ultrasound images, which contain key information of the images. Then, based on the time sequence continuity, i.e. the continuity and similarity between image frames, the feature maps are fused using a preset average fusion formula. This process reduces the differences between the feature maps through averaging and enhances the stability and representativeness of the features. Finally, the fused feature maps can more comprehensively represent the time sequence information of the continuous frames, providing a more abundant and reliable feature basis for subsequent lesion identification. This method effectively reduces the noise influence of single-frame features, improves the robustness of the features, and reduces the amount of calculation by reducing redundant information, thereby improving the accuracy and efficiency of lesion identification, and is particularly suitable for real-time and limited medical image analysis scenarios.

[0048] In a second aspect, the application provides a lesion identification device for B-ultrasound images, which comprises:

[0049] An acquisition module is configured to acquire a sequence of continuous multiple frames of B-ultrasound images in real time.

[0050] A feature extraction module is configured to perform feature extraction on the sequence of B-ultrasound images according to a preset lightweight convolutional neural network, to obtain a plurality of single-frame feature maps of continuous multiple frames.

[0051] A feature fusion module is configured to perform feature fusion on the plurality of single-frame feature maps of continuous multiple frames based on time sequence continuity, a preset fusion window, and preset weights, to generate a plurality of fused feature representations.

[0052] A lesion identification module is configured to perform lesion identification based on the plurality of fused feature representations through a preset lightweight inference network, to obtain a lesion identification result for the continuous multiple frames of B-ultrasound images.

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

[0054] The application realizes efficient processing and accurate recognition of B-ultrasound image sequences through an integrated device design. First, the acquisition module is responsible for capturing continuous multiple 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 computational load. Next, the feature fusion module fuses single-frame feature maps using temporal continuity, a fusion window, and a preset weight, which not only enhances the robustness of the features but also reduces computational load by reducing redundant information. Finally, the lesion recognition module uses a lightweight inference network optimized by quantization and pruning to recognize the fused features, improving the speed and accuracy of recognition. Overall, through the collaborative work of the modules, the device significantly improves the processing speed and reduces resource consumption while ensuring recognition accuracy, making it particularly suitable for low-end mobile edge devices. Therefore, it has important application value in the field of medical image analysis, especially in scenarios that require fast and accurate diagnosis.

[0055] In a third aspect, the application provides a computer-readable storage medium comprising a stored computer program, wherein the computer program, when executed, controls a device in which the computer-readable storage medium is located to perform the lesion recognition method for B-ultrasound images as described. The beneficial effects are the same as those of the lesion recognition method for B-ultrasound images provided in the first aspect of the application.

[0056] In a fourth aspect, the 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 the processor implements any one of the lesion recognition methods for B-ultrasound images as described in the first aspect when executing the computer program. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 FIG. 1 is a flowchart of an embodiment of the lesion recognition method for B-ultrasound images provided by the application;

[0058] Figure 2 FIG. 2 is a structural diagram of an embodiment of the lesion recognition device for B-ultrasound images provided by the application. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the application.

[0060] Embodiment One

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

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

[0063] S01: Real-time acquisition of a continuous multi-frame B-ultrasound image sequence.

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

[0065] In a primary medical institution, a mobile ultrasonic diagnostic instrument based on an ARM architecture, with a memory of 256 MB and without an independent GPU, is selected to realize lesion detection.

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

[0067] Further, the image sequence is preprocessed, and each frame of image Ii is preprocessed as follows: i

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

[0069] I′ i = MedianFilter(I i )

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

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

[0072]

[0073] In the formula, Ι' i is the i-th image after noise reduction, and Ι″ i (x,y) is the i-th image after normalization processing, where (x,y) represents the pixel coordinates, and Max(I′​​i ) is the maximum pixel value in the image I i .

[0074] (3) Image size fixing: adjust each frame of image to a uniform size (such as 224x224 or other input size suitable for the network) to facilitate batch transmission into the network for calculation.

[0075] (4) Pseudo-color processing: perform gray scale transformation on the pseudo-color image when necessary (for certain specific lesion identification algorithms) to reduce the processing amount.

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

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

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

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

[0080] To adapt to the computing resources of low-end devices, a lightweight convolutional neural network is adopted, which uses efficient modules such as Depthwise Separable Convolution and Shuffle Unit in its structure. The single-frame feature extraction model is set to CNN Φ , with I i as input and feature map as output.

[0081] f i = CNN φ (I'' i )

[0082] In the formula, I'' iFor the pre-processed i-th frame image, f i ∈R H×W×C is the feature map extracted for 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 weights.

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

[0084] Table 1: Network parameter table for multi-frame feature extraction

[0085] Parameter name Explanation Reference value Convolution kernel size Filter size used in convolution operation 3x3 or 5x5 Channel number Number of channels of output feature map of each layer 16 / 32 / 64, growing according to the number of layers Step Convolution step 1 or 2 Activation function Common activation function ReLU, ReLU6 Quantization bit width Model quantization parameter Int8

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

[0087] S03: Based on the temporal continuity, the pre-set fusion window, and the pre-set weights, the single-frame feature maps of the continuous multiple frames are fused to generate a plurality of fusion feature representations.

[0088] As a preferred embodiment, the continuous frames in the B-mode image sequence reflect the motion and structural changes within a certain time, and there are more noise and local artifacts in single frames. By fusing the features extracted from continuous multiple frames, the robustness of lesion identification can be improved, and the overall computational load can be reduced during the fusion process. The present scheme performs feature fusion processing on adjacent N frames (for example, N = 3 or 5).

[0089] The present application provides two main feature fusion methods:

[0090] A. Average fusion method:

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

[0092]

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

[0094] N: the number of frames participating in fusion.

[0095] F fuse : the fused feature map.

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

[0097] B. Weighted fusion method:

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

[0099]

[0100] In the formula, f i is the feature map extracted from the i-th frame, w i is the weight of the i-th frame, and satisfies F fuse is the fused feature map.

[0101] The weight w i combines local contrast, texture complexity, and motion amplitude three indicators, and dynamically allocates the weight through the following formula:

[0102]

[0103] 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, and a, b, g are adjustable hyperparameters.

[0104] The traditional method only relies on a single indicator (such as motion amplitude), while the present method comprehensively evaluates multiple indicators to more accurately reflect the information value of each frame.

[0105] In view of the computational power limitation of low-end devices, the present application proposes a pre-computed image quality score table. In actual deployment, the weight value is directly obtained by looking up the table, avoiding real-time calculation of complex indicators, and significantly reducing the computational 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 standardization:

[0108] The feature f i extracted from each frame is subjected to L2 normalization processing to ensure that the features of each frame are consistent in numerical scale, and the formula is:

[0109]

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

[0111] Sub-step 2: Weight calculation:

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

[0113] The specific algorithm can be selected as follows:

[0114]

[0115] where ΔI i The difference operator can be used, for example:

[0116] ΔI i = I" i - I" i-1

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

[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-1)-th frame, and ΔI i is the difference image between the i-th frame and the previous frame.

[0119] Sub-step 3: Feature fusion:

[0120] The normalized features are weighted and fused using the calculated weights of each frame to obtain the fused features 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, and F is the fused feature map. fuse

[0123] Or a simple average is used:

[0124]

[0125] where f i is the feature map extracted from the i-th frame, N is the number of consecutive frames participating in fusion, and F fuse is the fused feature map.

[0126] Sub-step 4: Feature dimension reduction (optional):

[0127] Since the fused features may have high dimensions, PCA or 1x1 convolution can be used for dimension reduction to adapt to the subsequent inference module.

[0128] F output = Conv 1×1 ​(F fuse )

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

[0130] Sub-step 5: Fusing feature cache and cyclic updating:

[0131] To keep the system in real-time continuous inference state, the fused feature F output is cached for the next fusion calculation reference, forming a sliding fusion window.

[0132] To achieve real-time processing, a sliding window mechanism is adopted. That is, in each step, a fixed size window is selected from the continuous frame sequence for fusion, and then the window is moved back by one frame, and the process is repeated. To deal with the case where the image content changes quickly (such as fast motion or sudden noise), a mechanism for dynamically adjusting the fusion window is introduced. The basic idea is: when a significant change in the image scene is detected, the fusion window size is reduced; otherwise, the fusion window size is increased. The following formula is used to calculate the fusion window size of the current frame:

[0133] N t = round(N0·e -λ·ΔI )

[0134] where ΔI is the difference degree of the current frame and the previous frame, and λ is the decay coefficient. When the scene changes dramatically (ΔI is large), the window size is automatically reduced (such as 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, which cannot balance the accuracy and speed requirements in dynamic scenes. The present application first considers the temporal continuity of the B-ultrasound image sequence, and calculates the weight of each frame feature map using a pre-set local contrast, texture complexity and motion amplitude three key indicators combined with a weight distribution formula. This step ensures that the fusion process can be differentiated according to the actual importance of the image content. Subsequently, a pre-set weighted fusion formula is used to perform weighted fusion on the single frame feature maps of the continuous multiple frames within the fusion window to generate a fused feature representation. This fusion strategy not only improves the richness and accuracy of the feature representation, but also adjusts the fusion window size dynamically to adapt to changes in image content, thereby ensuring real-time performance while enhancing the robustness of the system and the accuracy of lesion identification. Therefore, the beneficial effect is to improve the quality and efficiency of lesion detection through accurate feature fusion, making it possible to achieve efficient and accurate lesion identification on low-end mobile edge devices.

[0136] S04: performing lesion identification on the plurality of fused feature representations through a preset lightweight inference network to obtain a lesion identification result of the continuous multiple 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 performing lesion identification through the preset lightweight inference network specifically comprises:

[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 according to existing MobileNet, SqueezeNet, ShuffleNet, etc. The network structure is as follows:

[0139] (1) input layer: F output The dimension is [H', W', C'];

[0140] (2) intermediate layer: a plurality of lightweight convolutional blocks, each block including depth 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 a probability vector of lesion categories or a prediction map of segmentation regions.

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

[0144] Y = NET_Ψ(F output )

[0145] wherein Y is the final lesion identification result.

[0146] In view of the limited computing power of low-end mobile edge devices, the network NET_Ψ is simplified in terms of the number of layers and width. At the same time, the network model can be quantized and pruned to reduce the parameter quantity and improve the inference speed.

[0147] The algorithm optimization formula is, for example:

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

[0149] For real-time requirements, the scheme designs a mechanism for dynamically adjusting the fusion window, so that when the image content changes greatly, the number of frame fusion is automatically shortened (such as from 5 frames to 3 frames), to balance the calculation amount and accuracy.

[0150] The initial lightweight inference network of the present application is designed as an efficient structure aiming to reduce the amount of calculation (FLOPs). Then, through the application of optimization formula, the network is further pruned and quantized, wherein the pruning proportion coefficient (a) and the quantization efficiency (b) are preset parameters for guiding the network optimization. The pruning process removes unnecessary connections in the network, reducing the complexity and calculation demand of the model, while the quantization process converts the network parameters from high-precision representation to low-precision, further reducing the amount of calculation. The beneficial effect lies in the fact that this method not only significantly reduces the calculation load of the model, but also improves the inference speed, while maintaining the accuracy of the model, so that the optimized lightweight inference network is 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 can be a probability map or a region annotation of the lesion. The post-processing module adopts the following measures:

[0152] (1) Thresholding operation: fixed or adaptive thresholding is used for binaryzation of the probability map:

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

[0154] (2) Morphological filtering: operations such as dilation and erosion are used to eliminate noise and isolated regions;

[0155] (3) Boundary smoothing: Gaussian filtering or median filtering is used to smooth the edges to ensure the continuity of the detected regions.

[0156] The post-processed image result is directly presented on the display screen of the mobile edge device, or can be wirelessly transmitted to the central server for further evaluation.

[0157] Furthermore, the lesion identification method of the present application can reduce the amount of calculation, because the FLOPs required for single-frame feature extraction is F CNN , and the total FLOPs of the traditional frame-by-frame method is approximately NxF CNN under N frames; while in the multi-frame fusion method of the present application, only single-frame CNN extraction (or parallel calculation) plus fusion operation is required, and the fusion calculation amount F fuse is much smaller than F CNN . Assuming that the fusion calculation amount is approximately axF CNN (α << 1), the total calculation amount is approximately F total = NxF CNN + axF CNN .

[0158] The application effectively eliminates data redundancy by using a fusion module. In practice, a can be controlled to be below 0.1, thereby reducing the overall calculation amount by a proportion of (1-0.1 / N).

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

[0160] (1) The average inference delay is reduced by about 80%;

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

[0162] (3) The recall rate of lesion detection is increased by about 10%, and the false positive rate is reduced by about 5%.

[0163] Table 2 Comparison of experimental data

[0164]

[0165]

[0166] The above data prove that, in typical practical applications, the method of the 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 application acquires a sequence of continuous multiple B-ultrasound images in real time, and uses a pre-set lightweight convolutional neural network to extract features, effectively capturing key information in the image sequence. Subsequently, a fusion strategy based on time sequence continuity is used, combining a dynamically adjusted fusion window and pre-set weights to fuse single-frame feature maps and generate a fused feature representation. This process not only makes full use of the temporal redundancy information of the image sequence, reducing the computational complexity, but also enhances the robustness of the feature representation through weight adjustment. Finally, a lightweight inference network optimized through quantization and pruning is used for lesion recognition of the fused features, which not only improves the accuracy and real-time performance of recognition, but also significantly reduces energy consumption, making the method particularly suitable for resource-constrained low-end mobile edge devices. Therefore, the beneficial effects of the application are to significantly reduce the calculation amount and energy consumption while ensuring detection accuracy, improve the adaptability and robustness of the system in dynamic changing scenarios, and effectively solve the problem that the prior art cannot accurately and quickly recognize lesions in B-ultrasound images on low-end devices.

[0168] Embodiment Two

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

[0170] In the embodiment, the lesion recognition device of the B-ultrasound image comprises an acquisition module 10, a feature extraction module 20, a feature fusion module 30 and a lesion recognition module 40.

[0171] The acquisition module 10 is configured to acquire a sequence of B-ultrasound images of continuous multiple frames in real time.

[0172] As an embodiment, the sequence of B-ultrasound images of continuous multiple frames is acquired in real time, specifically as follows.

[0173] In a primary medical institution, a mobile ultrasonic diagnostic instrument based on ARM architecture, with a memory of 256 MB and without independent GPU, is selected to realize lesion detection.

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

[0175] Further, the image sequence is preprocessed, and each image I i is preprocessed as follows:

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

[0177] I′ i = MedianFilter(I i )

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

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

[0180]

[0181] In the formula, Ι' i is the i-th image after noise reduction, and Ι″ i (x, y) is the i-th image after normalization, where (x, y) represents the pixel coordinates, and Max(I′ i ) is the maximum pixel value in the image Ι' i .

[0182] (3) Image size fixing: adjust each frame of image to a uniform size (such as 224x224 or other suitable input size of the network) to facilitate batch transmission into the network for calculation.

[0183] (4) Pseudo-color processing: perform gray scale transformation on the pseudo-color image when necessary (for certain specific lesion identification algorithms) to reduce the processing amount.

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

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

[0186] The feature extraction module 20 is configured to perform feature extraction on the B-ultrasound image sequence according to a pre-set lightweight convolutional neural network, to obtain a plurality of continuous single-frame feature maps.

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

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

[0189] f i = CNN φ (I i )

[0190] In the formula, I'' is the i-th pre-processed image, f i ∈R i H×W×C ​Feature map extracted for 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 weights.

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

[0192] Table 3. Network parameter table for multi-frame feature extraction

[0193] Parameter name Explanation Reference value Convolution kernel size Filter size used in convolution operation 3x3 or 5x5 Channel number Number of channels of output feature map of each layer 16 / 32 / 64, growing according to the number of layers Step Convolution step 1 or 2 Activation function Common activation function ReLU, ReLU6 Quantization bit width Model quantization parameter Int8

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

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

[0196] As a preferred embodiment, the continuous frames in the B-ultrasound image sequence reflect the motion and structural changes within a certain time, and the single-frame has more noise and local artifacts. By fusing the features extracted from the continuous multiple frames, the robustness of lesion recognition can be improved, and the overall computational load can be reduced during the fusion process. The present scheme performs feature fusion processing on adjacent N frames (for example, N = 3 or 5).

[0197] The present application provides two main feature fusion methods:

[0198] A. Average fusion method:

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

[0200]

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

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

[0203] F fuse : Feature map after fusion.

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

[0205] B. Weighted fusion method:

[0206] The weight w i (w i ≥ 0, all wi The sum is 1, and the formula is:

[0207]

[0208] In the formula, 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] The weight w i combines local contrast, texture complexity, and motion amplitude, and dynamically allocates the weight through the following formula:

[0210]

[0211] 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, and a, b, and g are adjustable hyperparameters.

[0212] The traditional method only relies on a single indicator (such as motion amplitude), while the present method comprehensively evaluates multiple indicators to more accurately reflect the information value of each frame.

[0213] In view of the computational power limitation of low-end devices, the present application proposes a pre-computed image quality score table. In actual deployment, the weight value is directly obtained by looking up the table, avoiding real-time calculation of complex indicators, and significantly reducing the computational 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 standardization:

[0216] The features f i extracted from each frame are subjected to L2 normalization processing to ensure that the features of each frame are consistent in numerical scale, and the formula is:

[0217]

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

[0219] Sub-step 2: Weight calculation:

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

[0221] The specific algorithm can be selected as follows:

[0222]

[0223] where ΔI i The difference operator can be used, for example:

[0224] ΔI i = I″ i -I″ i-1

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

[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-1)-th frame, and ΔI i is the difference image between the i-th frame and the previous frame.

[0227] Sub-step 3: Feature fusion:

[0228] The normalized features are weighted and fused using the calculated weights of each frame to obtain the fused features 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, and F is the fused feature map. fuse

[0231] Or a simple average is used:

[0232]

[0233] where f i is the feature map extracted from the i-th frame, N is the number of consecutive frames participating in fusion, and F fuse is the fused feature map.

[0234] Sub-step 4: Feature dimension reduction (optional):

[0235] Since the fused features may have a high dimension, PCA or 1x1 convolution can be used for dimension reduction to adapt to the subsequent inference module.

[0236] F output = Conv 1×1 (F fuse )

[0237] where F fuse ​Conv 1×1 (·) is a 1x1 convolution operation to reduce the feature dimension.

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

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

[0240] In order to realize real-time processing, a sliding window mechanism is adopted. That is, in each step, a fixed size window is selected from the continuous frame sequence for fusion, and then the window is moved back by one frame, and the process is repeated. In order to deal with the case where the image content changes quickly (such as fast motion or sudden noise), a mechanism for dynamically adjusting the fusion window is introduced. The basic idea is: when a significant change in the image scene is detected, the fusion window size is reduced; otherwise, the fusion window size is increased. The following formula is used to calculate the fusion window size of the current frame:

[0241] N t = round(N0·e -λ·ΔI )

[0242] Where ΔI is the difference degree of the current frame and the previous frame, λ is the attenuation coefficient. When the scene changes dramatically (ΔI is large), the window size is automatically reduced (such as 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, which cannot balance the accuracy and speed requirements in a dynamic scene. The present application first considers the temporal continuity of the B-ultrasound image sequence, and calculates the weight of each frame feature map using a pre-set local contrast, texture complexity and motion amplitude three key indicators combined with a weight distribution formula. This step ensures that the image content can be processed differently according to its actual importance during the fusion process. Subsequently, a pre-set weighted fusion formula is used to perform weighted fusion on the single frame feature maps of the continuous multiple frames within the fusion window to generate a fusion feature representation. This fusion strategy not only improves the richness and accuracy of the feature representation, but also adjusts the fusion window size dynamically to adapt to changes in image content, thereby ensuring real-time performance while enhancing the robustness of the system and the accuracy of lesion identification. Therefore, the beneficial effect is to improve the quality and efficiency of lesion detection by precise feature fusion, making it possible to achieve efficient and accurate lesion identification on low-end mobile edge devices.

[0244] The lesion recognition module 40 is configured to perform lesion recognition on the plurality of fused feature representations through a preset lightweight inference network to obtain a lesion recognition result of the continuous multiple frames of B-ultrasound images.

[0245] As a preferred embodiment, the lesion recognition through the preset lightweight inference network is specifically:

[0246] The fused feature F output As input, the lightweight inference network is entered to perform lesion classification or segmentation. The network design can be customized according to existing MobileNet, SqueezeNet, ShuffleNet, etc. For example, the network structure is as follows:

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

[0248] (2) Intermediate layer: a plurality of lightweight convolutional blocks, each block including a depth separable convolution, a BatchNormalization, and a 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 a probability vector of lesion categories or a prediction map of segmentation regions.

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

[0252] Y = NET_Ψ(F output )

[0253] Where Y is the final lesion recognition result.

[0254] In view of the limited computing power of low-end mobile edge devices, the network NET_Ψ is simplified in terms of the number of layers and width. At the same time, the network model can be quantized and pruned to reduce the parameter quantity and improve the inference speed.

[0255] For example, the algorithm optimization formula is:

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

[0257] For real-time requirements, the present scheme designs a mechanism for dynamically adjusting the fusion window, so that when the image content changes greatly, the number of frame fusions is automatically shortened (such as from 5 frames to 3 frames), to balance the calculation amount and accuracy.

[0258] The initial lightweight inference network of the present application is designed as an efficient structure aiming to reduce the amount of calculation (FLOPs). Then, through the application of optimization formula, the network is further pruned and quantized, wherein the pruning proportion coefficient (a) and the quantization efficiency (b) are preset parameters for guiding the network optimization. The pruning process removes unnecessary connections in the network, reducing the complexity and calculation demand of the model, while the quantization process converts the network parameters from high-precision representation to low-precision, further reducing the amount of calculation. The beneficial effect lies in the fact that this method not only significantly reduces the calculation load of the model, but also improves the inference speed, while maintaining the accuracy of the model, so that the optimized lightweight inference network is 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 can be a probability map or a region annotation of the lesion. The post-processing module adopts the following measures:

[0260] (1) Thresholding operation: fixed or adaptive thresholding is used for binaryzation of the probability map:

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

[0262] (2) Morphological filtering: operations such as dilation and erosion are used to eliminate noise and isolated regions;

[0263] (3) Boundary smoothing: Gaussian filtering or median filtering is used to smooth the edges to ensure the continuity of the detected regions.

[0264] The post-processed image result is directly presented on the display screen of the mobile edge device, or can be wirelessly transmitted to the central server for further evaluation.

[0265] Furthermore, the lesion identification method of the present application can reduce the amount of calculation, because the FLOPs required for single-frame feature extraction is F CNN , and the total FLOPs of the traditional frame-by-frame method is approximately NxF CNN under N frames; while in the multi-frame fusion method of the present application, only single-frame CNN extraction (or parallel calculation) plus fusion operation is required, and the fusion calculation amount F fuse is much smaller than F CNN . Assuming that the fusion calculation amount is approximately axF CNN (α << 1), the total calculation amount is approximately F total = NxF CNN + axF CNN .

[0266] The application effectively eliminates data redundancy by using a fusion module. In practice, a can be controlled below 0.1, thereby reducing the overall calculation amount by about (1-0.1 / N) percentage.

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

[0268] (1) The average inference delay is reduced by about 80%;

[0269] (2) Energy consumption is reduced by about 50%, and stable operation can still be achieved under low-power conditions;

[0270] (3) The lesion detection recall rate is increased by about 10%, and the false positive rate is reduced by about 5%.

[0271] Table 4 Experimental comparison data

[0272] Indicator Traditional frame-by-frame inference method Multi-frame fusion method of the application Average inference delay 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 the 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] The application realizes efficient processing and accurate recognition of B-ultrasound image sequences through integrated device design. First, the acquisition module is responsible for real-time capture of consecutive multiple B-ultrasound images, ensuring real-time and continuity of data; then, the feature extraction module uses a lightweight convolutional neural network to extract features from these images, effectively capturing key information and reducing computational load; next, the feature fusion module fuses single-frame feature maps using temporal continuity, fusion window and preset weight, which not only enhances the robustness of the features, but also reduces the computational load by reducing redundant information; finally, the lesion recognition module uses a lightweight inference network optimized by quantization and pruning to recognize the fused features, improving the speed and accuracy of recognition. Overall, through the cooperative work of each module, the device significantly improves the processing speed and reduces resource consumption while ensuring recognition accuracy, and is particularly suitable for low-end mobile edge devices, thereby showing important application value in the field of medical image analysis, especially in scenarios that require fast and accurate diagnosis.

[0275] Embodiment three:

[0276] The embodiment of the application provides a computer readable storage medium, which comprises a stored computer program, wherein the computer readable storage medium controls the device where the computer readable storage medium is located to execute the lesion recognition method of the B-ultrasound image when the computer program runs.

[0277] If the B-ultrasound image lesion identification method is implemented in the form of a software function unit and used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When the processor executes the computer program, the steps of the above-mentioned various method embodiments can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate form, etc. The computer-readable medium can include any entity or device that can carry the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0278] Embodiment Four

[0279] The application provides a terminal device, which comprises 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, any one of the B-ultrasound image lesion identification methods described in Embodiment One is implemented.

[0280] The above-mentioned specific embodiments further illustrate the purpose, technical solutions, and beneficial effects of the application. It should be understood that the above-mentioned embodiments are only specific embodiments of the application and are not used to limit the protection scope of the application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the application should be included in the protection scope of the application.

Claims

1. A method for lesion identification in B-ultrasound images, characterized in that, include: Real-time acquisition of multiple consecutive frames of ultrasound image sequences; The ultrasound image sequence is subjected to feature extraction using a preset lightweight convolutional neural network to obtain several consecutive multi-frame single-frame feature maps. The single-frame feature maps of the continuous multi-frames are fused based on temporal continuity, a preset fusion window, and preset weights to generate several fused feature representations. 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 specifically involves: Based on the preset local contrast, preset texture complexity, and preset motion amplitude, combined with the preset weight allocation formula, the preset weights are calculated. Based on the preset weights and preset weighted fusion formula, the single-frame feature maps of the consecutive multiple frames are subjected to weighted feature fusion in a preset fusion window to obtain several fused feature representations. 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 several fused feature representations also includes: The difference between consecutive frames of the acquired multi-frame B-mode ultrasound image sequence is monitored. If the difference is higher than a preset threshold, the size of the preset fusion window is dynamically adjusted according to a preset fusion window adjustment formula. The preset fusion window adjustment formula is as follows: In the formula, ΔI is the difference between the current frame and the previous frame, λ is the attenuation coefficient, and N t N is the adjusted fusion window size. o The preset size of the fusion window; 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 further includes: According to the preset average fusion formula, the single-frame feature maps of the consecutive multiple frames are fused. The preset average fusion formula is: In the formula, f i Let F be the single-frame feature map extracted from the i-th frame, N be the number of consecutive frames participating in the fusion, and F be the feature map extracted from the i-th frame. fuse The fused feature map; Based on the aforementioned fusion feature representations, lesion identification is performed through a preset lightweight inference network to obtain lesion identification results for multiple consecutive frames of B-mode 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 lesion identification in B-ultrasound images according to claim 1, characterized in that, After acquiring a sequence of multiple consecutive ultrasound images in real time, the method further includes: Based on the preset noise reduction filtering and pixel value normalization processing methods, the continuous multi-frame B-mode ultrasound image sequence is preprocessed and adjusted to the preset resolution.

3. The method for lesion identification in B-ultrasound images according to claim 1, characterized in that, The preset weight allocation formula and the preset weighted fusion formula are specifically as follows: The preset weight allocation formula is as follows: In the formula, Score the local contrast of the i-th frame. For texture complexity, The motion amplitude is denoted by α, β, and γ, which are preset adjustable hyperparameters. i These are the preset weights; The preset weighted fusion formula is: In the formula, N is a preset positive integer greater than 1. The feature map extracted for the i-th frame. Let i be the weight of the i-th frame. This represents the fusion feature representation.

4. The method for lesion identification in B-ultrasound images according to claim 1, characterized in that, The step of quantizing and pruning the preset initial lightweight inference network according to the preset optimization formula specifically involves: The preset optimization formula is: In the formula, 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.

5. A lesion identification device for B-ultrasound images, characterized in that, include: The acquisition module is used to acquire a sequence of multiple consecutive ultrasound images in real time. The feature extraction module is used to extract features from the ultrasound image sequence based on a preset lightweight convolutional neural network to obtain a number of consecutive multi-frame single-frame feature maps. The 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 several fused feature representations. 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 specifically involves: Based on the preset local contrast, preset texture complexity, and preset motion amplitude, combined with the preset weight allocation formula, the preset weights are calculated. Based on the preset weights and preset weighted fusion formula, the single-frame feature maps of the consecutive multiple frames are subjected to weighted feature fusion in a preset fusion window to obtain several fused feature representations. 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 several fused feature representations also includes: The difference between consecutive frames of the acquired multi-frame B-mode ultrasound image sequence is monitored. If the difference is higher than a preset threshold, the size of the preset fusion window is dynamically adjusted according to a preset fusion window adjustment formula. The preset fusion window adjustment formula is as follows: In the formula, ΔI is the difference between the current frame and the previous frame, λ is the attenuation coefficient, and N t N is the adjusted fusion window size. o The preset size of the fusion window; 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 further includes: According to the preset average fusion formula, the single-frame feature maps of the consecutive multiple frames are fused. The preset average fusion formula is: In the formula, f i Let F be the single-frame feature map extracted from the i-th frame, N be the number of consecutive frames participating in the fusion, and F be the feature map extracted from the i-th frame. fuse The fused feature map; The lesion identification module is used to identify lesions based on the aforementioned fusion feature representations through a preset lightweight inference network, thereby obtaining lesion identification results for multiple consecutive frames of B-mode 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.

6. 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, it controls the device on which the computer-readable storage medium is located to perform the lesion identification method of ultrasound images as described in any one of claims 1 to 4.

7. A terminal device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the lesion identification method for ultrasound images as described in any one of claims 1 to 4.

Images