Ultrasound video object detection method and system based on improved object detection model

Abstract

This invention discloses a method and system for ultrasound video target detection based on an improved target detection model, belonging to the fields of intelligent medical image detection, computer vision, and video stream processing. The method includes: performing lightweight preprocessing on the ultrasound video stream to output a standardized valid frame sequence; inputting the valid frame sequence into the backbone network of the improved target detection model to extract clean depth feature maps; after multi-scale fusion via a neck network, generating target detection boxes and confidence scores using a dynamic detection head, and outputting video-level detection results. The system includes a video input module, a lightweight preprocessing module, a backbone network module, a neck network module, a dynamic detection head module, and a result output module. This invention, using the above method and system, achieves real-time, high-precision target detection in ultrasound video through dual Mona cascaded filtering, global contextual convolution, and a triple attention dynamic detection head. The model is lightweight and efficient, meeting the requirements of embedded real-time deployment.

Description

Technical Field

[0001] This invention relates to the fields of intelligent medical image detection, computer vision and video stream processing, and in particular to an ultrasound video target detection method and system based on an improved target detection model. Background Technology

[0002] Ultrasound imaging, due to its non-invasive, real-time, and portable advantages, has been widely used in clinical examinations and primary care screening. With the popularization of portable ultrasound devices (such as handheld ultrasound), ultrasound examinations are gradually evolving from traditional static image acquisition to dynamic video scanning. However, portable ultrasound devices are limited by hardware cost and size, and the acquired ultrasound videos often suffer from low resolution, significant inter-frame noise interference, blurred target boundaries, and high inter-frame redundancy, posing a serious challenge to subsequent automated target detection.

[0003] Current object detection technologies are primarily geared towards natural images or high-resolution static medical images. While mainstream detection models (such as the YOLO series and Faster R-CNN) perform well in general object detection tasks, their designs do not fully consider the dynamic characteristics and low-quality properties of ultrasound videos. Directly applying existing models to ultrasound video object detection has the following drawbacks: Poor process adaptability: Existing methods typically extract images frame by frame from the video and then detect each frame separately. The temporal correlation between frames is not effectively utilized, resulting in low detection efficiency, susceptibility to motion artifacts, and difficulty in achieving real-time processing.

[0004] Insufficient model detection performance: Due to the high overlap between background noise (such as anatomical texture) and target signal in the frequency domain and spatial distribution in ultrasound video frames, general detection models are prone to misdetecting background texture as targets. At the same time, for targets with large scale variations and irregular shapes, existing models are prone to problems such as missing small targets, incomplete target bounding, and low bounding box regression accuracy.

[0005] Real-time performance is difficult to meet: the inference speed of existing detection models cannot match the normal frame rate of ultrasound video (25fps~30fps), making it difficult to achieve end-to-end real-time detection on resource-constrained embedded devices.

[0006] Furthermore, ultrasound videos exhibit high inter-frame redundancy, with numerous invalid frames (such as blurry frames or frames with a low target area ratio) not only increasing computational burden but also introducing unnecessary noise interference. Current technologies lack lightweight preprocessing mechanisms for ultrasound video streams, failing to effectively reduce computational load while maintaining detection accuracy.

[0007] Therefore, there is an urgent need for a target detection method that adapts to the dynamic features and low-quality attributes of ultrasound video, enabling real-time preprocessing of video streams, efficient feature extraction, and inter-frame information fusion, improving detection accuracy while meeting real-time requirements, and possessing lightweight deployment capabilities. Summary of the Invention

[0008] The purpose of this invention is to provide an ultrasonic video target detection method and system based on an improved target detection model. To address the problem of low quality, high noise, and poor real-time performance of handheld ultrasonic videos, which makes patch detection difficult, this invention proposes a collaborative optimization strategy for the entire process, including lightweight video stream preprocessing, feature purification, global reconstruction, dynamic perception, and inter-frame fusion. It also constructs a YOLO11-MGD-T fast detection network to achieve end-to-end real-time automated detection.

[0009] To achieve the above objectives, this invention provides an ultrasonic video target detection method based on an improved target detection model, comprising the following steps: S1. Perform lightweight preprocessing on the frame image data in the ultrasound video stream. The preprocessing includes frame filtering, denoising and alignment, and outputs a standardized effective frame sequence. S2. Input the effective frame sequence into the backbone network of the improved object detection model to extract clean depth feature maps; S3. Input the deep feature map into the neck network of the improved target detection model, perform multi-scale feature fusion, and output the fused feature map; S4. Input the fused feature map into the dynamic detection head to generate target detection boxes and confidence scores, and output video-level detection results.

[0010] Preferably, the frame filtering in step S1 specifically includes: Get each frame of image A lightweight segmentation model based on MobileNetV3 is used to locate the target region in the frame image and calculate the image sharpness. and target area Percentage, only retain and The frames are considered valid frames; Among them, clarity The calculation formula is: ; In the formula, The width of the frame image in pixels. The height of the frame image is in pixels. and The frame images obtained by using the Sobel gradient operator are respectively in direction and Gradient of direction; Target area percentage The calculation formula is: ; In the formula, The pixel area of ​​the target region in the frame image output by the lightweight segmentation model.

[0011] Preferably, the noise reduction in step S1 specifically includes: Denoising is performed on each frame of the image using a multi-scale Gaussian bilateral filter, with the following formula: ; In the formula, The coordinates of the current pixel. The width of the filter window is half its width, and the window size is [value missing]. , and These represent the offsets of neighboring pixels relative to the center pixel in the horizontal and vertical directions, respectively. This represents the grayscale value of the original frame at the neighboring pixels. For spatial domain kernel functions, For grayscale kernel functions, The normalized weights are equal to the weights of all neighboring pixels. sum.

[0012] Preferably, the alignment in step S1 specifically includes: The denoised frames are then subjected to grayscale normalization and spatial alignment. The grayscale normalization formula is as follows: ; In the formula, For the denoised frame at the pixel grayscale value at that location This represents the average grayscale value of all pixels in the current denoised frame. The standard deviation of grayscale; Spatial alignment uses affine transformations to move the target region to the center of the image.

[0013] Preferably, in step S2, the backbone network is C2PSA_Mona, where each PSABlock in the original YOLO11 C2PSA module contains a spatial attention branch and a feedforward neural network. C2PSA_Mona embeds a first-level Mona module after the spatial attention branch and a second-level Mona module after the feedforward neural network, forming a double-Mona cascaded filter: The first-level Mona suppresses background noise through soft thresholding: the feature response of the spatial attention branch output. Perform soft threshold transformation: ; In the formula, The output feature response value after soft thresholding. It is a soft threshold function. For symbolic functions, for The absolute value, For the network to base on input features The threshold parameter learned adaptively. This indicates that the value within the parentheses is the maximum value between 0 and 0. The second-level Mona enhances the target-related channels through feature recalibration: it re-calibrates the feature maps processed by the first-level Mona and the feedforward neural network. Perform recalibration: ; In the formula, This indicates element-wise multiplication. It uses the Sigmoid activation function, with an output range of (0,1). A global context modeling function used to generate channel weight vectors. To Apply soft threshold filtering, This is the clean depth feature map that the backbone network outputs at the end.

[0014] Preferably, in step S3, the neck network is C3k2_GCConv, and the global context convolution GCConv replaces the standard convolution; GCConv captures long-range dependencies through parallel multi-scale convolution branches during the training phase and reparameterizes them into a single convolution operator during the inference phase.

[0015] Preferably, in step S4, the dynamic detection head is DyHead, which sequentially applies scale-aware attention. Spatial perception and attention and task-aware attention Transform the fused feature map: ; In the formula, To fuse feature maps, For the real number field, The number of levels in the feature pyramid. The number of spatial locations at each level, Number of channels; For scale-aware attention functions, For spatial awareness attention function, For task-aware attention functions; This is the feature representation after triple attention transformation, used to generate object detection boxes and confidence scores.

[0016] Preferably, the improved object detection model is trained using the following loss function: ; In the formula, The intersection-union ratio (IUU) based on the auxiliary bounding box is calculated as follows: ; in, It is the inner auxiliary box of the real box. The inner auxiliary box of the predicted box. Indicates the area of ​​the region. Indicates intersection, Represents the union; The auxiliary border is generated by the following formula: ; in, The center coordinates of the original bounding box. and These are the width and height of the original frame, respectively. To control the hyperparameters of the auxiliary bounding box scale, when Generate inner auxiliary box in time. Generate an outer auxiliary box at the time. These are the coordinates of the left, right, top, and bottom boundaries of the auxiliary border, respectively. and These are the Euclidean distances between the top-left corner and the bottom-right corner of the predicted bounding box and the ground truth bounding box, respectively. and These represent the width and height in pixels of the input image, respectively.

[0017] This invention also provides an ultrasonic video target detection system based on an improved target detection model, used to execute the above-described ultrasonic video target detection method based on the improved target detection model, comprising: The video input module is used to acquire frame image data from the ultrasound video stream; The lightweight preprocessing module is used to perform real-time frame filtering, denoising, and alignment on frame image data, and output a standardized and effective frame sequence. The backbone network module is used to extract clean depth feature maps from valid frame sequences; The neck network module is used to perform multi-scale feature fusion on the deep feature map and output a fused feature map. The dynamic detection head module is used to generate target detection boxes and confidence scores from the fused feature map; The results output module is used to output video-level detection results and real-time visual annotations.

[0018] Preferably, it also includes a lightweight inter-frame fusion module for performing temporal weighted fusion and non-maximum suppression on the detection results of multiple consecutive frames; the system is lightweightly deployed on an embedded device, and the lightweight preprocessing module runs synchronously with the video input module to achieve real-time processing.

[0019] Therefore, the ultrasonic video target detection method and system based on the improved target detection model described above have the following beneficial effects: (1) Achieve end-to-end real-time automated detection of ultrasound video. Through lightweight preprocessing (frame filtering, denoising, alignment) and inter-frame fusion module, no manual intervention is required, and the end-to-end detection time of a single frame is less than 35 milliseconds, which matches the real-time acquisition frame rate of 30fps of ultrasound video, significantly improving detection efficiency.

[0020] (2) Significantly improves detection accuracy and robustness. To address issues such as background noise, variable target scale, and irregular shape in low-quality ultrasound videos, mechanisms such as dual Mona cascade filtering, global contextual convolution, and dynamic detection heads are employed to effectively suppress anatomical noise, enhance global semantic association, and improve target integrity recognition capabilities. Experiments show that the model achieves 90.96% mAP@0.5, a 7.32% improvement over the benchmark YOLO11, with an accuracy of 89.80% and a recall of 89.00%.

[0021] (3) The model is lightweight and easy to deploy in embedded systems. Through channel pruning, depthwise separable convolution replacement and inference acceleration optimization, the number of parameters is reduced by 38% compared with the benchmark, and the model size is reduced to 8.5 megabytes. It can be deployed in the embedded system of portable devices such as handheld ultrasound, meeting the real-time detection needs of rapid screening scenarios at the grassroots level.

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0023] Figure 1 This is a flowchart of the ultrasonic video target detection method based on the improved target detection model of the present invention; Figure 2 This is a schematic diagram of the network structure of the improved target detection model of the present invention (i.e., YOLO11-MGD-T network). Figure 3 This is a schematic diagram of the C2PSA_Mona module and its sub-modules of the present invention; Figure 4 This is a schematic diagram of the internal structure of the C3k2_GCConv feature fusion module of the present invention (integrated GCConv). Figure 5 This is a schematic diagram of the internal structure of the dynamic detection head (DyHead) of this invention. Detailed Implementation

[0024] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0026] Example This embodiment uses carotid artery plaques as the detection target to describe in detail the ultrasound video target detection method and system based on an improved target detection model of the present invention. It should be noted that the technical solution of the present invention is not limited to carotid artery plaque detection, but can also be applied to other types of ultrasound video target detection tasks.

[0027] I. Overview of the improved target detection model.

[0028] The improved target detection model described in this invention refers to the YOLO11-MGD-T model, which is obtained by improving the baseline YOLO11 model through multiple modules to address the characteristics of handheld ultrasound video. Figure 2 As shown. Specifically includes: Backbone network: An improved C2PSA_Mona network is adopted, which suppresses anatomical noise in ultrasound images and extracts clean depth feature maps through a dual Mona cascade filtering mechanism.

[0029] Neck network: An improved C3k2_GCConv network is adopted, which achieves multi-scale feature fusion through global contextual convolution (GCConv) to solve the problem of target "semantic fragmentation".

[0030] Dynamic detection head: Employs DyHead, which adaptively adjusts the detection focus through a triple attention mechanism of scale, space, and task.

[0031] Loss function: Inner_MPDIoU is used, combined with auxiliary bounding box and corner distance constraints to improve bounding box regression accuracy.

[0032] The model has been designed with lightweight features (the number of parameters has been reduced by 38% compared to YOLO11, and the size has been reduced to 8.5M), and it can be deployed on embedded devices to achieve real-time detection.

[0033] II. Methods and steps, such as Figure 1 As shown.

[0034] S1: Lightweight Preprocessing This step performs real-time preprocessing on each frame of image data in the ultrasound video stream, including frame filtering, noise reduction, and alignment.

[0035] S11: Real-time frame filtering Get each frame of image ( (The frame number is used to locate the target region in the frame image (the carotid artery region in this embodiment) using a lightweight segmentation model based on MobileNetV3. This segmentation model has only 0.8M parameters and an inference time of less than 1ms / frame.

[0036] Calculate the sharpness of the frame image and target area Percentage, only retain and The frames are considered valid frames; Among them, clarity The calculation formula is: ; In the formula, The width of the frame image in pixels. The height of the frame image is in pixels. , , and The frame images obtained by using the Sobel gradient operator are respectively in direction and Gradient of direction; Target area percentage The calculation formula is: ; In the formula, The pixel area of ​​the target region in the frame image output by the lightweight segmentation model.

[0037] S12: Lightweight noise reduction To address anatomical noise (such as the linear textures produced by the sternocleidomastoid muscle and connective tissue of blood vessel walls) and Gaussian noise in ultrasound video frames, multi-scale Gaussian bilateral filtering is employed for denoising. ; In the formula, The coordinates of the current pixel. The width of the filter window is half its width, and the window size is [value missing]. In this embodiment, =2 Window 5x5 and These represent the offsets of neighboring pixels relative to the center pixel in the horizontal and vertical directions, respectively. This represents the grayscale value of the original frame at the neighboring pixels. For spatial domain kernel functions (dependent on Euclidean distance). It is a grayscale kernel function (dependent on grayscale difference). The normalized weights are equal to the weights of all neighboring pixels. The algorithm reduces computational cost by 60% compared to traditional denoising algorithms, takes less than 2ms to denoise a single frame, and preserves target edges.

[0038] S13: Standardized Alignment Perform grayscale normalization on the denoised frames: ; In the formula, For the denoised frame at the pixel grayscale value at that location , These are the mean and standard deviation of grayscale values ​​for all pixels in the current denoised frame, respectively, to eliminate grayscale differences caused by device and environment.

[0039] Then, based on the target region segmented in S11, the target region (carotid artery lumen) is moved to the center of the image through affine transformation to avoid acquisition angle deviation. The alignment operation has low computational cost and high speed.

[0040] After preprocessing, the valid frames are arranged into a sequence according to time order. The total preprocessing time for a single frame is less than 5ms, matching a video frame rate of 25 / 30fps, achieving real-time processing without delay.

[0041] S2: Backbone Network Feature Extraction This step inputs the preprocessed valid frame sequence into the improved backbone network C2PSA_Mona to extract clean depth feature maps.

[0042] Definition: C2PSA_Mona is an improvement on the original YOLO11 C2PSA module, such as... Figure 3 As shown. Each PSABlock in the original C2PSA module contains a spatial attention branch and a feedforward neural network (FNN). The C2PSA_Mona of this invention embeds a first-level Mona module after the attention branch and a second-level Mona module after the FNN, forming a double Mona cascaded filter.

[0043] First-level Mona (soft threshold filtering): Feature response of spatial attention branch output Perform soft threshold transformation: ; In the formula, The output feature response value after soft thresholding. It is a soft threshold function. For symbolic functions, for The absolute value, For the network to base on input features The threshold parameter learned adaptively. This indicates that the value in parentheses is the maximum value between 0 and 0; this operation sets responses with absolute values ​​less than the threshold to zero (suppressing high-frequency linear background noise) and retains responses with values ​​greater than the threshold (preserving patchy features).

[0044] Second-level Mona (feature recalibration): The feature map after processing by the first-level Mona and feedforward neural network Perform recalibration: ; In the formula, This indicates element-wise multiplication. It uses the Sigmoid activation function, with an output range of (0,1). A global context modeling function used to generate channel weight vectors. To Apply soft threshold filtering, This is the clean depth feature map output by the backbone network. Through this mechanism, the channel encoding the target semantic information is amplified, while the background clutter channel is suppressed.

[0045] S3: Neck Network Feature Fusion like Figure 4 As shown, the C3k2_GCConv module described in this invention is an improvement on the original YOLO11 C3k2 module. The original C3k2 module consists of two standard convolutional layers and multiple stacked C3k sub-modules in between, with each C3k sub-module containing multiple standard 3×3 convolutions. The improvement of this invention lies in replacing the standard convolutions in the bottleneck layer of the C3k sub-module with global context convolutions (GCConv), resulting in the C3k_GCConv sub-module, while maintaining the cross-stage partial network (CSPNet) structure of C3k2.

[0046] Specifically, the internal structure of the C3k2_GCConv module is as follows (see...). Figure 4 ): Input splitting: The input feature map first undergoes channel transformation through a convolutional layer (Conv), and then a split operation divides the feature map into two branches along the channel dimension. The left branch (denoted as Branch_A) retains the original features and is used for subsequent shortcut connections; the right branch (denoted as Branch_B) enters the stacked C3k_GCConv submodule.

[0047] The C3k_GCConv submodule contains three convolutional layers. The bottleneck layer in the middle uses GCConv, while the two convolutional layers before and after it remain standard 1×1 convolutions. The training-inference dual-mode design of GCConv is as follows: Training phase: GCConv runs multiple convolutional branches in parallel, including a 1×1 global context branch (which transforms the entire feature map after calculating global average pooling) and multiple local convolutional branches with different dilation rates, thereby capturing long-distance dependencies at different scales.

[0048] Inference phase: Through reparameterization, the weights of all branches are equivalently merged into a standard convolution, so the computational cost during inference is exactly the same as that of a standard convolution, without adding any latency.

[0049] Multi-submodule stacking: Based on the parameter n of C3k2 (n=2 or 3 in this embodiment), the above C3k_GCConv submodules are stacked multiple times to deepen the feature abstraction level. During the stacking process, the output of each submodule serves as the input of the next submodule, gradually expanding the receptive field and enhancing global semantic association.

[0050] Feature fusion: The stacked Branch_B output and the original features of Branch_A are concatenated along the channel dimension by the Concat operation, and then compressed by an output convolutional layer (Conv) to obtain the final fused feature map.

[0051] Long-distance dependency modeling mechanism: For carotid plaques containing heterogeneous components (e.g., hyperechoic calcification caps and hypoechoic lipid kernels), traditional 3×3 convolutions can only see local image regions, easily misclassifying the calcification cap and lipid kernel as two independent targets. GCConv's global context branch, however, can directly "see" the entire feature map, establishing pixel-level associations between the lipid kernel and calcification cap, thus semantically connecting them into a complete plaque. Simultaneously, multi-scale local branches ensure the preservation of detailed information. This design significantly improves the model's ability to recognize the integrity of large-scale, irregularly shaped targets with uneven internal echoes without increasing the inference burden.

[0052] This step uses an improved neck network C3k2_GCConv to fuse the multi-scale feature maps output by S2.

[0053] Definition: GCConv (Global Context Convolution) is a reparameterized convolutional architecture. During training, it captures long-range dependencies through parallel multi-scale convolutional branches (including global context branches and local feature branches); during inference, these branches are losslessly reparameterized into a single convolutional operator without increasing computational cost.

[0054] In this embodiment, C3k2_GCConv applies GCConv to the feature pyramid (FPN) and PANet feature aggregation path. For targets like carotid artery plaques with heterogeneous internal components (e.g., a strong echogenic calcification cap + a hypoechoic lipid core), traditional convolution tends to fragment them. However, GCConv reconnects fragmented features into a whole through global semantic association, significantly improving the ability to identify the integrity of large-scale and mixed echogenic targets. Output fused feature map. ( For the real number field, The number of levels in the feature pyramid. The number of spatial locations at each level, (Number of channels).

[0055] S4: Dynamic Detection Head and Output This step uses the DyHead dynamic detection head to generate target detection boxes and confidence scores, and outputs video-level results.

[0056] Definition: such as Figure 5 As shown, DyHead unifies the attention mechanisms of scale, space, and task into a single framework, applying them sequentially: ; In the formula, For scale-aware attention function (dynamically adjust the weights of different levels to solve multi-scale problems). For spatial awareness attention function (focusing on the discriminative spatial location of the target and suppressing blood flow artifacts). For task-aware attention functions (dynamically switch channels on and off according to classification / localization requirements); This is the feature representation after triple attention transformation, used to generate object detection boxes and confidence scores.

[0057] The specific implementation of each attention branch is as follows: Scale-aware attention: The input features are compressed spatially using global average pooling (Avg pooling), and then weights for each layer are generated through two fully connected layers (Conv 1×1, ReLU, Hard sigmoid) to dynamically adjust the importance of features at different scales. For carotid artery plaques, this mechanism allows small calcifications to receive high weights for high-resolution shallow features, while large diffuse plaques receive high weights for deep semantic features.

[0058] Spatial-aware attention: Employing the concept of deformable convolution, the offset of each location is learned through a standard convolutional layer (Conv 3×3), and then normalized by Sigmoid to guide the network to focus on the most discriminative spatial location of the plaque (such as the lipid core area and the plaque shoulder), effectively suppressing blood flow signal artifacts in the vascular lumen.

[0059] Task-aware attention: The features output by spatial attention are globally averaged and pooled, then activated by two fully connected layers (FC) and ReLU, and finally normalized to generate channel-dimensional gating signals. This mechanism dynamically enables or suppresses specific feature channels according to the different needs of classification tasks (which rely on texture information) and localization tasks (which rely on edge gradients), thus achieving intelligent resource allocation.

[0060] Training Loss Function: The improved object detection model is trained using the Inner_MPDIoU loss function. This loss function combines the auxiliary bounding box gradient control of Inner-IoU with the corner distance constraint of MPDIoU, making it particularly suitable for slender, wall-hugging targets (such as carotid artery plaques). Its final form is: ; In the formula, The intersection-union ratio (IUU) based on the auxiliary bounding box is calculated as follows: ; in, It is the inner auxiliary box of the real box. The inner auxiliary box of the predicted box. Indicates the area of ​​the region. Indicates intersection, Represents the union; The auxiliary border is generated by the following formula: ; in, The center coordinates of the original bounding box. and These are the width and height of the original frame, respectively. To control the hyperparameters of the auxiliary bounding box scale, when Generate inner auxiliary box in time. Generate an outer auxiliary box at the time. These are the coordinates of the left, right, top, and bottom boundaries of the auxiliary border, respectively. and These are the Euclidean distances between the top-left corner and the bottom-right corner of the predicted bounding box and the ground truth bounding box, respectively. and These represent the width and height in pixels of the input image (256×256), respectively. In this embodiment, during training, the following values ​​are used: (Inner auxiliary box) is used for fine regression.

[0061] Inference Phase: For each input frame, the dynamic detection head directly outputs the detection bounding box and confidence score. Then, inter-frame result fusion is performed: Non-maximum suppression (NMS) is used to remove duplicate boxes, with an NMS threshold set to 0.5. Simultaneously, a temporal attention mechanism is used to weight and fuse the detection results of multiple consecutive frames, amplifying the weight of clear frames and suppressing the weight of motion artifact frames. Finally, a video-level detection report is output, including the target's occurrence time, location, and confidence score, and the detection bounding boxes are annotated in real-time within the video stream.

[0062] III. Experimental Data and Results

[0063] (I) Preparation of Experimental Data From November 2024 to May 2025, handheld ultrasound carotid artery examination videos were collected from 720 subjects in the Department of Cardiology and Department of Neurology of the affiliated hospital. Among them, there were 425 males and 294 females, with a mean age of 70.38±9.74 years. The videos were acquired by a TUORen handheld color ultrasound imaging system (5-12MHz linear array probe) at a frame rate of 25fps, in MP4 format, with a single video segment duration of 10-30 seconds.

[0064] The acquired video stream was frame extracted, yielding a total of 2078 valid frames. The ratio of patches present to absent was 7:5, and the image resolution was uniformly set to 256×256 pixels. Invalid frames and invalid videos were strictly screened according to the inclusion and exclusion criteria.

[0065] A three-level quality control strategy of "single-group double physician annotation + dispute review" was adopted. The LabelMe tool was used to annotate the effective frames. The intra-group correlation coefficient (ICC) analysis showed that the ICC values ​​were all greater than 0.95 (F test, p<0.001), which ensured the objectivity and consistency of the annotation.

[0066] The labeled videos were divided into a training set (576 videos, 1662 valid frames) and a test set (144 videos, 416 valid frames) in an 8:2 ratio to provide data support for model training and performance verification.

[0067] (II) Model Construction and Training The YOLO11-MGD-T fast detection network was built based on the PyTorch framework. The lightweight backbone network C2PSA_Mona-Lite, the high-efficiency neck network C3k2_GCConv-Lite, the lightweight DyHead-Lite dynamic detection head, the Inner_MPDIoU lightweight loss function, and the lightweight inter-frame fusion module were built in sequence. The overall number of network parameters was reduced by 38% compared with the benchmark YOLO11.

[0068] Training parameter settings: batch size is set to 32, initial learning rate is 0.001, stochastic gradient descent (SGD) optimizer is used, momentum is 0.937, weight decay is 0.0005, training epochs are 300, and the learning rate is dynamically decayed using a cosine annealing strategy; mixed precision training is used during training to improve training speed while avoiding gradient vanishing.

[0069] Training environment: The deep learning training environment was built using an NVIDIA RTX 4090 graphics card, CUDA 11.8, and cuDNN 8.7. The total training time was less than 48 hours. After training, the optimal model weights were saved.

[0070] (III) Model Performance Validation and Video Detection Evaluation metrics: Precision (P), recall (R), and mAP@0.5 were used as the evaluation metrics for detection precision, and single-frame inference time, end-to-end detection time, and video detection frame rate were used as the evaluation metrics for detection speed, to comprehensively verify the model's "high precision + high speed" performance.

[0071] Performance Validation: Experiments were conducted on the test set, and the results showed that the YOLO11-MGD-T model achieved an mAP@0.5 of 90.96%, which is 7.32% higher than the baseline YOLO11, with an accuracy of 89.80% and a recall of 89.00%. The model's single-frame inference time was <20ms, the end-to-end detection time was <35ms / frame, and the video detection frame rate reached 30fps, achieving real-time detection.

[0072] Automated detection and verification of video streams: The test set of 144 handheld ultrasound videos were input into the detection system to achieve fully automated detection. The system can accurately output plaque detection results and video-level detection reports, and annotate the plaque location in real time in the video stream. There are no background false detections or missed detections of small plaques. The consistency between the detection results and the physician's manual diagnosis is over 98.5%.

[0073] (iv) Model Deployment and Practical Application Lightweight model deployment: The trained YOLO11-MGD-T model is quantized (FP32→FP16), converted using ONNX, and accelerated and optimized using TensorRT to generate an inference engine suitable for embedded devices, reducing the model size to 8.5M.

[0074] End-side deployment: The optimized model is deployed on the embedded system of the TUORen handheld color ultrasound imaging system (Rockchip RK3588, Android 12) to achieve rapid and automated local detection of the handheld ultrasound device.

[0075] Practical application: Clinical trials were conducted in primary community health service centers, and a total of 200 residents were screened for carotid artery plaques. The detection process realizes "handheld ultrasound video acquisition → real-time automated detection by the system → instant output of results", with a single test taking less than 1 minute. This is more than 10 times more efficient than traditional manual diagnosis. The consistency between the test results and the diagnoses of physicians in higher-level hospitals reached 98%, meeting the clinical needs of rapid screening at the primary level.

[0076] This embodiment also provides an ultrasonic video target detection system based on an improved target detection model, including: Video input module: Acquires frame image data from the ultrasound video stream.

[0077] Lightweight preprocessing module: performs real-time frame filtering (including lightweight segmentation model), multi-scale Gaussian bilateral filtering for noise reduction, grayscale normalization and affine transformation alignment on frame image data, and outputs a standardized effective frame sequence.

[0078] Backbone network module: The C2PSA_Mona network is used, which includes dual Mona cascaded filtering to extract clean depth feature maps.

[0079] Neck network module: The C3k2_GCConv network is used, and GCConv is used for multi-scale feature fusion.

[0080] Dynamic detection head module: Uses DyHead to generate target detection boxes and confidence scores.

[0081] Results output module: Outputs video-level detection reports and real-time visual annotations.

[0082] The system also includes a lightweight inter-frame fusion module for performing temporal weighted fusion and non-maximum suppression (NMS threshold 0.5) on detection results from multiple consecutive frames. The system is lightweight and deployed on embedded devices (such as handheld ultrasound devices based on RK3588), and the preprocessing module runs synchronously with the video input module to achieve real-time processing.

[0083] Therefore, the present invention adopts the above-mentioned ultrasound video target detection method and system based on the improved target detection model, achieving a detection accuracy of 90.96% mAP@0.5, which is 7.32% higher than the baseline; single-frame inference <20ms, end-to-end <35ms / frame, matching 30fps real-time processing; the number of parameters is reduced by 38%, and the model size is reduced to 8.5M, which can be deployed on embedded devices; the detection is fully automated without human intervention, and the consistency with physician diagnosis reaches more than 98%.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An ultrasonic video target detection method based on an improved target detection model, characterized in that, Includes the following steps: S1. Perform lightweight preprocessing on the frame image data in the ultrasound video stream. The preprocessing includes frame filtering, denoising and alignment, and outputs a standardized effective frame sequence. S2. Input the effective frame sequence into the backbone network of the improved object detection model to extract clean depth feature maps; S3. Input the deep feature map into the neck network of the improved target detection model, perform multi-scale feature fusion, and output the fused feature map; S4. Input the fused feature map into the dynamic detection head to generate target detection boxes and confidence scores, and output video-level detection results.

2. The ultrasonic video target detection method based on the improved target detection model according to claim 1, characterized in that, The frame filtering in step S1 specifically includes: Get each frame of image A lightweight segmentation model based on MobileNetV3 is used to locate the target region in the frame image and calculate the image sharpness. and target area Percentage, only retain and The frames are considered valid frames; Among them, clarity The calculation formula is: ； In the formula, The width of the frame image in pixels. The height of the frame image in pixels. and The frame images obtained by using the Sobel gradient operator are respectively in direction and Gradient of direction; Target area percentage The calculation formula is: ； In the formula, The pixel area of ​​the target region in the frame image output by the lightweight segmentation model.

3. The ultrasonic video target detection method based on the improved target detection model according to claim 2, characterized in that, The noise reduction in step S1 specifically includes: Denoising is performed on each frame of the image using a multi-scale Gaussian bilateral filter, with the following formula: ； In the formula, For the denoised frame in pixels grayscale value at that location The coordinates of the current pixel. The width of the filter window is half its width, and the window size is [value missing]. , and These represent the offsets of neighboring pixels relative to the center pixel in the horizontal and vertical directions, respectively. This represents the grayscale value of the original frame at the neighboring pixels. For spatial domain kernel functions, For grayscale kernel functions, The normalized weights are equal to the weights of all neighboring pixels. sum.

4. The ultrasonic video target detection method based on the improved target detection model according to claim 3, characterized in that, The alignment in step S1 specifically includes: The denoised frames are then subjected to grayscale normalization and spatial alignment. The grayscale normalization formula is as follows: ； In the formula, This represents the average grayscale value of all pixels in the current denoised frame. The standard deviation of grayscale; Spatial alignment uses affine transformations to move the target region to the center of the image.

5. The ultrasonic video target detection method based on the improved target detection model according to claim 1, characterized in that, In step S2, the backbone network is C2PSA_Mona, where each PSABlock in the original YOLO11 C2PSA module contains a spatial attention branch and a feedforward neural network. C2PSA_Mona embeds a first-level Mona module after the spatial attention branch and a second-level Mona module after the feedforward neural network, forming a double-Mona cascaded filter: The first-level Mona suppresses background noise through soft thresholding: the feature response of the spatial attention branch output. Perform soft threshold transformation: ； In the formula, The output feature response value after soft thresholding. It is a soft threshold function. For symbolic functions, for The absolute value, For the network to base on input features The threshold parameter learned adaptively. This indicates that the value within the parentheses is the maximum value between 0 and 0. The second-level Mona enhances the target-related channels through feature recalibration: it re-calibrates the feature maps processed by the first-level Mona and the feedforward neural network. Perform recalibration: ； In the formula, This indicates element-wise multiplication. It uses the Sigmoid activation function, with an output range of (0,1). A global context modeling function used to generate channel weight vectors. To Apply soft threshold filtering, This is the clean depth feature map that the backbone network outputs at the end.

6. The ultrasonic video target detection method based on the improved target detection model according to claim 1, characterized in that, In step S3, the neck network is C3k2_GCConv, and the standard convolution is replaced by the global context convolution GCConv; GCConv captures long-range dependencies through parallel multi-scale convolution branches during the training phase and reparameterizes them into a single convolution operator during the inference phase.

7. The ultrasonic video target detection method based on the improved target detection model according to claim 1, characterized in that, In step S4, the dynamic detection head is DyHead, and scale-aware attention is applied sequentially. Spatial perception and attention and task-aware attention Transform the fused feature map: ； In the formula, To fuse feature maps, For the real number field, The number of levels in the feature pyramid. The number of spatial locations at each level, Number of channels; For scale-aware attention functions, For spatial awareness attention function, For task-aware attention functions; This is the feature representation after triple attention transformation, used to generate object detection boxes and confidence scores.

8. The ultrasonic video target detection method based on the improved target detection model according to claim 1, characterized in that, The improved object detection model is trained using the following loss function: ； In the formula, The intersection-union ratio (IUU) based on the auxiliary bounding box is calculated as follows: ； in, It is the inner auxiliary box of the real box. The inner auxiliary box of the predicted box. Indicates the area of ​​the region. Indicates intersection, Represents the union; The auxiliary border is generated by the following formula: ； in, The center coordinates of the original bounding box. and These are the width and height of the original frame, respectively. To control the hyperparameters of the auxiliary bounding box scale, when Generate inner auxiliary box in time. Generate an outer auxiliary box at the time. These are the coordinates of the left, right, top, and bottom boundaries of the auxiliary border, respectively. and These are the Euclidean distances between the top-left corner and the bottom-right corner of the predicted bounding box and the ground truth bounding box, respectively. and These represent the width and height in pixels of the input image, respectively.

9. An ultrasonic video target detection system based on an improved target detection model, used to execute the ultrasonic video target detection method based on an improved target detection model as described in any one of claims 1-8, characterized in that, include: The video input module is used to acquire frame image data from the ultrasound video stream; The lightweight preprocessing module is used to perform real-time frame filtering, denoising, and alignment on frame image data, and output a standardized and effective frame sequence. The backbone network module is used to extract clean depth feature maps from valid frame sequences; The neck network module is used to perform multi-scale feature fusion on the deep feature map and output a fused feature map. The dynamic detection head module is used to generate target detection boxes and confidence scores from the fused feature map; The results output module is used to output video-level detection results and real-time visual annotations.

10. The ultrasonic video target detection system based on the improved target detection model according to claim 9, characterized in that, It also includes a lightweight inter-frame fusion module, which performs temporal weighted fusion and non-maximum suppression on the detection results of multiple consecutive frames; the system is lightweight and deployed on an embedded device, and the lightweight preprocessing module runs synchronously with the video input module to achieve real-time processing.

Images