Intraoperative positioning system based on ultrasonic breast tumor rotary atherectomy

Abstract

The invention provides an intraoperative positioning system based on ultrasonic breast tumor rotary atherectomy, and relates to the technical field of medical image intelligent assistance. The system comprises a training data acquisition module, a target detection model construction module, a model training module and a target detection module instruction starting module. The training data acquisition module is used for acquiring effective training data based on an ultrasonic breast tumor rotary cutting operation video to construct a basic data set; the target detection model construction module constructs a target detection model based on an improved YOLOV11 network, and detects and tracks the positions of a lump and a cutter groove in an ultrasonic video stream in real time; the model training module is used for training a target detection model; the target detection module is used for detecting an ultrasonic breast tumor rotary cutting operation video frame and detecting and tracking the positions of a lump and a knife groove in an ultrasonic video stream in real time; the instruction starting module judges whether a rotary cutter starting instruction is sent out or not based on the positions of a lump and a cutter groove in the ultrasonic video stream.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging intelligent assistance technology, and in particular to an intraoperative positioning system based on ultrasonic breast tumor excision. Background Art

[0002] Breast lumps excision surgery, also known as the ultrasound-guided vacuum-assisted biopsy system (VABB), was introduced in 1994, clinically used in the United States in 1995, and introduced to China in 2010. After more than 20 years of development, this technology has become the preferred method for removing benign breast lumps or performing biopsies on potentially malignant lumps. Because the surgery is performed under ultrasound guidance, clinicians must be able to identify the lump and the excision blade on the ultrasound image, which requires extremely high skills from the physician. Therefore, this procedure is difficult to perform in primary hospitals at the county, township, and village levels. Using artificial intelligence methods to process surgical images in real time can help clinicians quickly locate the position of the excision blade in the ultrasound image during surgery, helping them to better advance the excision blade to a location near the lump, reduce damage to unrelated tissues, and improve the accuracy and success rate of the surgery. Therefore, this technology has important clinical significance.

[0003] Since there have been no attempts to use artificial intelligence ultrasound guidance in the field of minimally invasive excision of breast lumps, it is impossible to describe the most similar solution. The most relevant existing research results are based on the identification of benign and malignant breast lumps based on two-dimensional ultrasound images. In the domestic field, Song Pengjie et al. published an article in the Chinese Journal of Ultrasound Medicine in 2023 titled "The Value of Breast Ultrasound Artificial Intelligence-Assisted Diagnosis Combined with Breast X-ray Photography in the Diagnosis of Breast Cancer." They proposed a method to improve the diagnostic accuracy of breast cancer ultrasound diagnosis by combining ultrasound artificial intelligence and breast X-ray results and evaluated its clinical value. In 2024, Li Xiaojie et al. published an article in the Chinese Journal of Clinical Research titled "Application of Dual-Modal Ultrasound Deep Learning Prediction Model in the Diagnosis of Breast Cancer." The article proposed a prediction model constructed using deep learning technology based on breast ultrasound grayscale images and elasticity images, and compared its image reading ability with that of ultrasound doctors. It was concluded that the model can significantly improve the doctor's diagnostic efficiency in differentiating benign and malignant breast lesions. In 2024, Chen Rui et al. published an article on the application value of artificial intelligence-assisted diagnosis system in the diagnosis of BI-RADS category 4 breast nodules with a maximum diameter of ≤2cm. The article tested the performance of the artificial intelligence system in ultrasound diagnosis and showed good results.

[0004] Furthermore, in related fields abroad, Li Y et al. proposed a multimodal AI-based method for determining malignant breast tumors based on the tumor's blood oxygen metabolism level in their paper "Intelligent scoring system based on dynamic optical breast imaging for early detection of breast cancer." Vigil, Nicolle et al. also proposed a dual-modal AI-based method for determining whether a breast tumor is benign or malignant in their paper "Dual-Intended Deep Learning Model for Breast Cancer Diagnosis in Ultrasound Imaging." Current research indicates that there is no AI-assisted positioning method for excision of breast tumors.

[0005] The above existing technologies have the following two problems: the first is the discontinuity of assistance. The existing technology can only process two-dimensional ultrasound images and give relevant suggestions. This method cannot meet the requirements of real-time recognition of surgical video streams. At the same time, the lag of discontinuous auxiliary information may affect the doctor's surgical judgment. The second aspect is the singleness of the assistance. Most existing technologies only focus on the location of target resection objects such as tumors, without real-time recognition of the location of the rotary cutter in the ultrasound image. Real-time recognition of the location of the rotary cutter is crucial in this operation and has important guiding significance for doctors with little experience. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned existing technologies and provide an intraoperative positioning system based on ultrasonic breast tumor atherectomy to achieve detection and positioning of tumors and atherectomy blades under ultrasonic images.

[0007] To solve the above technical problems, the technical solution adopted by the present invention is: an intraoperative positioning system based on ultrasonic breast tumor resection, including a training data acquisition module, a target detection model construction module, a model training module, a target detection module and an instruction activation module;

[0008] The training data acquisition module acquires effective training data based on the ultrasonic breast tumor excision surgery video to construct a basic data set, and divides it into a training set, a validation set and a test set;

[0009] The target detection model construction module builds a target detection model based on an improved YOLOv11 network, which is used to detect and track the positions of lumps and grooves in ultrasound video streams in real time. The improved YOLOv11 network introduces a MobileNetv4 inverted bottleneck block and a universal bottleneck block after the first convolution-batch normalization module of the backbone network, and introduces a lightweight C3K2 module after the convolution module of the neck network. The MobileNetv4 inverted bottleneck block and the universal bottleneck block are used to extract low-level features and mid- and high-level features, respectively. The lightweight C3K2 module combines local information clipping with cross-stage partial connection technology to achieve small target detection.

[0010] The model training module trains the target detection model based on the training data;

[0011] The target detection module uses a trained target detection model to detect the video frames of ultrasonic breast tumor resection surgery, and detects and tracks the positions of the tumor and the knife groove in the ultrasound video stream in real time;

[0012] The instruction start module determines whether to issue a rotary cutter start instruction based on the positions of the mass and the knife groove in the ultrasound video stream detected in real time by the target detection module and according to the rotary cutter start determination method.

[0013] Furthermore, the system also includes a real-time visualization module; the real-time visualization module includes a display screen and a visualization interface for displaying the ultrasound video stream in real time; each frame of the ultrasound video image is detected by the target detection model, and different labels for the mass and the rotary cutter are given, and the labeling results of the mass and the rotary cutter are superimposed on the original video screen for display.

[0014] Furthermore, the specific method for the training data acquisition module to acquire effective training data and construct a basic data set based on ultrasonic breast tumor excision surgery videos is as follows:

[0015] Screen and capture video clips of ultrasonic breast tumor excision surgery; capture the original ultrasonic breast tumor excision surgery video, and capture the complete operation process from the moment the excision knife penetrates the surface of the breast tissue to the moment the knife groove opens to perform tumor excision;

[0016] A uniform sampling strategy is used for key frame extraction;

[0017] The tumor location and the operative knife location in each key frame image are annotated; the annotated key frame images are enhanced using data enhancement technology to construct a basic data set.

[0018] Furthermore, in the backbone network design of the improved YOLOV11 network, a convolutional layer combined with batch normalization is used as the basic feature extraction unit; a strategy of gradually increasing the number of channels layer by layer is adopted to achieve multi-scale feature capture of images from small scales to large scales; at the same time, two types of convolutional kernels are used for local features, and through the downsampling operation with an initial stride of 2, the size of the feature map is compressed; the MobileNetV4 inverted bottleneck block is based on the principle of depthwise separable convolution and is used in the low-level feature extraction stage. Through the design of small channel numbers and small strides, while ensuring the feature extraction effect, the computational amount is greatly reduced; the general bottleneck block combines depth convolution and expansion convolution and is used for middle and high-level feature extraction.

[0019] Furthermore, the method for judging the opening of the rotary cutter constructs an intelligent algorithm for prompting that the cutter groove of the rotary cutter reaches a specified position. This algorithm will continuously monitor the spatial position relationship between the cutter groove and the mass. When the cutter groove and the mass appear on the image at the same time, it is also necessary to meet the condition that the horizontal cutter groove is on the left side of the mass and there is a spatial intersection longitudinally to determine whether the cutter groove has reached a position suitable for opening the rotary cutter.

[0020] Furthermore, the method for judging whether the cutter groove has reached a position suitable for opening the rotary cutter is as follows:

[0021] Let the upper left corner coordinates of the cutter groove be (x1, y1), the upper left corner coordinates of the mass be (x2, y2), the length of the cutter groove be L1, the width be H1, the length of the mass be L2, and the width be H2. Based on the geometric dimensions and coordinate relationship between the cutter groove and the mass, the following three cases are judged:

[0022] First, the size of the cutter groove is larger than that of the mass: When the length difference L1 - L2 between the cutter groove and the mass is greater than n pixels, where n is a set pixel value, the horizontal judgment condition is x1 - L1 * 0.2 < x2, and the vertical judgment condition is y2 - H2 * 1.1 < y1, so as to ensure that the cutter groove covers the left side of the mass in the horizontal direction and there is an intersection between the two after opening the cutter groove vertically;

[0023] Second, the size of the cutter groove is smaller than that of the mass: When the length difference L2 - L1 is greater than n pixels, the horizontal judgment condition is adjusted to x2 - L2 * 0.2 < x1, and the vertical condition remains y2 - H2 * 1.1 < y1;

[0024] Third, the sizes of the cutter groove and the mass are smaller than the set value: If the length difference between the two is less than n pixels, on the premise of keeping the upper left corner coordinates of the mass unchanged, the length of the mass is extended to 1.2 times the original size, and then the judgment conditions of the second case are followed to achieve compatibility detection under different sizes.

[0025] The beneficial effects of the above-mentioned technical solution are as follows: the intraoperative positioning system for ultrasonic breast tumor atherectomy provided by the present invention can calculate and mark the location of the atherectomy blade groove by identifying targets with distinct morphological features in ultrasound images. The system can also identify the location of tumors in the images in real time. This system can assist the surgeon in more accurately activating the atherectomy blade during surgery, thereby improving the efficiency and success rate of the procedure and reducing damage to unrelated tissues. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A structural block diagram of an intraoperative positioning system based on ultrasonic breast tumor atherectomy provided by an embodiment of the present invention;

[0027] Figure 2 Detection results of the improved YOLOV11 network model provided by an embodiment of the present invention, where (a) is a precision-recall (PR) curve and (b) is a normalized confusion matrix.

[0028] Figure 3 An ultrasonic breast tumor image with lump and pedicle markings provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0030] In this embodiment, an intraoperative positioning system based on ultrasonic breast tumor resection is provided. Figure 1 As shown, it includes: training data acquisition module, target detection model construction module, model training module, target detection module, real-time visualization module and instruction start module;

[0031] The training data acquisition module acquires effective training data based on the ultrasonic breast tumor excision surgery video to construct a basic data set, and divides it into a training set, a validation set and a test set including:

[0032] First, video clips of ultrasonic breast tumor excision surgeries were screened and captured. Under the guidance of professional surgeons, the original videos were systematically sorted and captured. The capture range covered the entire surgical process from the moment the exciter penetrated the breast tissue surface to the moment the blade slot opened to perform tumor excision. The video data from this period accurately covered the surgical positioning process, fully preserving the dynamic changes in the exciter's trajectory and tumor position, as well as important operational details such as the surgeon's adjustment of the instrument angle and depth. This fully met the research needs of this experiment for real-time positioning of the exciter and tumor, and determination of the optimal excision site.

[0033] After valid video clips are screened, keyframes are extracted. To strike a balance between data integrity and computational efficiency, this example uses a uniform sampling strategy of one frame every 10 seconds. This sampling strategy systematically records key surgical dynamics, such as the operative trajectory of the operative blade and changes in tumor morphology, while also avoiding data redundancy caused by high-frequency sampling. This significantly reduces the computational costs of data storage and model training, providing a timely and representative image dataset for subsequent precise localization of the operative blade and tumor.

[0034] After keyframe extraction, the image annotation phase begins. This embodiment utilizes a dual review mechanism to ensure the accuracy of the annotated data: First, a specialist with at least five years of clinical experience annotates the tumor and pedicle blade positions in each keyframe image based on their specialized medical knowledge. Subsequently, an expert with at least twenty years of clinical experience reviews and confirms the annotated results. This tiered review process effectively reduces annotation errors, providing highly accurate and reliable real-world labeled data for deep learning model training and laying a solid foundation for subsequent work.

[0035] At the same time, we use a combination of data enhancement technologies such as color enhancement, mosaic enhancement, left-right flipping, and scaling (consider whether to write in detail and need to supplement pictures) to perform image enhancement, simulate diverse surgical environments and changes in ultrasound image quality, and effectively improve the robustness and detection accuracy of the model in complex clinical scenarios.

[0036] After data labeling is completed, the dataset partitioning phase begins. In this example, a total of 11,610 ultrasound images from 175 cases of data were obtained to construct a basic dataset. To ensure the effectiveness of model training and avoid overfitting, the basic dataset was hierarchically partitioned using a 7:2:1 ratio, with 122 cases in the training set, 35 cases in the validation set, and 18 cases in the test set. This partitioning strategy, by covering diverse data covering different tumor morphologies, locations, and surgical operation scenarios, can fully verify the robustness and reliability of the model in practical applications, significantly improving the model's generalization ability.

[0037] The target detection model building module builds a target detection model based on the improved YOLOV11 network, which is used to detect and track the positions of masses and grooves in the ultrasound video stream in real time;

[0038] In the backbone network design of the improved YOLOV11 network, convolutional layers (ConvBN) combined with batch normalization (Batch Normalization) are used as basic feature extraction units to build a stable feature extraction foundation. A strategy of increasing the number of channels layer by layer from 40-80-128-480-512 is adopted to capture multi-scale features of images from small to large scales. At the same time, 3×3 and 5×5 convolution kernels are flexibly used to give full play to the advantages of the two convolution kernels in local feature extraction. The downsampling operation with an initial step size of 2 is used to quickly compress the feature map size, thereby reducing the computational pressure for subsequent processing.

[0039] To overcome the computational bottlenecks of the traditional YOLOV11 network architecture and further improve model efficiency, the MobileNetV4 inverted bottleneck block and the universal bottleneck block are innovatively introduced after the first convolution-batch normalization module (ConvBN) of the backbone network. The MobileNetV4 inverted bottleneck block, based on the principle of depthwise separable convolution, is suitable for low-level feature extraction. Its design with a small number of channels and a small step size significantly reduces computation while ensuring effective feature extraction. The universal bottleneck block, which integrates depthwise convolution with dilated convolution, is primarily used for mid- and high-level feature extraction. By optimizing the computational process, it significantly reduces the number of model parameters while maintaining detection accuracy, improving overall computational efficiency.

[0040] The model's neck network integrates low- and high-resolution features through multi-level upsampling and feature fusion, enhancing the model's detection capabilities for objects of varying scales. To further improve inference efficiency, a lightweight C3K2 module is introduced after the neck network's convolutional (Conv) module. This module employs a local information clipping strategy to effectively reduce redundant computations, accelerating inference while ensuring detection accuracy remains unchanged, achieving a good balance between speed and accuracy.

[0041] In terms of modular design, the improved YOLOV11 network of the present invention has clear division of labor and efficient collaboration among modules: the ConvBN convolution layer serves as the primary link in image preprocessing, preparing for subsequent deep feature extraction; the MobileNetV4 inverted bottleneck block and the general bottleneck block are responsible for the efficient extraction of low-level and mid- and high-level features, respectively; the SPPF module uses spatial pyramid pooling technology to enhance the model's perception of multi-scale features and improve the recognition accuracy of targets of different sizes, thereby further reducing the number of parameters and avoiding parameter redundancy; the lightweight C3K2 module combines local information clipping with cross-stage partial connection (CSP) technology to demonstrate excellent performance in small target detection tasks.

[0042] Through the ConvBN convolutional layer, MobileNetV4 inverted bottleneck block and universal bottleneck block, SPPF module, C3K2 module and cross-stage part (CSP) connection technology, the modules have clear division of labor and efficient collaboration, showing excellent performance in small target detection tasks.

[0043] The optimized YOLOv11n is denoted as YOLOv11n+, and a comparison of model parameters is shown in Table 1. This example rigorously calculates and evaluates key indicators such as computational load (FLOPS), inference time, and parameter count, fully verifying the effectiveness and superiority of the proposed network architecture in terms of performance improvement.

[0044] Table 1 Parameter settings before and after improvement

[0045]

[0046] The model training module trains the target detection model based on the training data. This includes: This embodiment, based on the NVIDIA RTX 4090 graphics card and the PyTorch 2.6.0 framework, employs multiple strategies to ensure effective training. Overfitting is prevented by setting a training cycle of 333 rounds and incorporating an early stopping mechanism (training is terminated after 55 consecutive rounds of validation set performance without improvement). A stochastic gradient descent (SGD) optimizer with momentum is used, with an initial learning rate tuned to 0.01 to ensure smooth model convergence. A batch size of 64 is set to balance computational efficiency and training stability.

[0047] This example systematically compares different YOLO architectures and improved versions of the present invention in order to quantitatively evaluate the impact of each proposed improvement. This example evaluates key performance indicators, including the number of parameters, computational complexity, and inference speed under different network configurations, through controlled variable experiments. This rigorous analysis provides an empirical basis for architecture optimization and demonstrates how each improvement affects the final performance of the model. The method proposed in this example, which introduces the MobileNetV4 inverted bottleneck block and the universal bottleneck block into the Backbone part of YOLOv11n, has the best performance for real-time detection of groves and bumps. Table 2 lists the experimental results of each parameter in detail.

[0048] Table 2 Comparison of parameters of different models

[0049]

[0050] As shown in Table 2, the Yolo11n+ model has 2,140,390 parameters, the smallest of all models. This demonstrates its compact design, which helps reduce storage requirements and improve deployment flexibility. Furthermore, the model's FLOPS (Floating Points Per Second) is 4.6 GB, significantly lower than other models. This means that fewer computing resources are required to process images, reducing energy consumption and hardware costs. The Yolo11n+ model also performs well in inference speed. Its CPU inference time is only 12.2 ms, and its GPU inference time is 0.7 ms, both of which are the lowest among all models. This demonstrates that the Yolo11n+ model can provide faster response times in real-world applications, which is particularly important for applications requiring real-time processing.

[0051] The target detection module uses a trained target detection model to detect the video frames of ultrasonic breast tumor resection surgery, and detects and tracks the positions of the tumor and the knife groove in the ultrasound video stream in real time;

[0052] In this embodiment, based on the above optimized network model, the detection accuracy, recall rate and map50 value of different YOLO architectures for grooving and mass as well as the overall detection of the two are systematically compared, as shown in Table 3. Figure 2 (a) shows the precision-recall (PR) curve. Figure 2 (b) shows the normalized confusion matrix, which together show the classification performance and detection confidence characteristics of the model.

[0053] Table 3 Detection results of different models

[0054] Network Architecture Yolov11n Yolov11n+ Precision (lump) 0.830 0.862 Recall rate (lump) 0.760 0.725 map50 (lump) 0.835 0.827 Accuracy (rotary cutter) 0.783 0.850 Recall rate (rotary cutter) 0.709 0.746 map50 (rotary cutter) 0.742 0.799 Accuracy (both) 0.807 0.856 Recall (both) 0.734 0.736 map50 (both) 0.789 0.813

[0055] As can be seen from Table 3, YOLOv11n outperforms the YOLO11n+ proposed in this paper in terms of mass recognition performance. However, the optimized YOLO11n+ model proposed in this paper performs better in terms of the combined recognition performance of both the rotary cutter groove and the mass. Combined with the model training parameter performance evaluation indicators given in Table 2, the YOLO11n+ model proposed in this paper not only significantly reduces the number of model parameters, thereby achieving real-time inference, but also improves the recognition performance of both the groove and the mass.

[0056] from Figure 2The performance of the groove of the rotary cutter, the mass, and the combination of the two can be seen from the precision-recall curve in (a). When the recall rate of the groove reaches about 0.7, the precision is about 0.8. In contrast, when the recall rate of the mass reaches about 0.8, the precision still remains at about 0.9, that is, the model has higher precision and robustness in mass recognition. Overall, the average performance of the two recognitions still shows that the optimized network model proposed in this invention has good recognition ability.

[0057] The real-time visualization module developed a visualization interface based on OpenCV 4.9.0 for displaying the real-time ultrasound video stream; each frame of the ultrasound video image is quickly inferred through the object detection model, giving significantly different annotations for the mass and the rotary cutter, and accurately superimposing the annotation results of the mass and the rotary cutter on the original video screen, as Figure 3 shown.

[0058] The instruction activation module determines whether to issue a rotary cutter activation instruction based on the positions of the mass and the groove in the ultrasound video stream detected in real time by the object detection module according to the rotary cutter activation determination method;

[0059] The rotary cutter activation determination method constructs an intelligent algorithm for accurately prompting the groove of the rotary cutter to reach the specified position. Considering the significant differences in the operation methods of clinicians when operating the rotary cutter and the ultrasound probe, this embodiment designs two different execution schemes, namely surgical method 1 and surgical method 2, to adapt to diverse surgical operation habits.

[0060] Surgical method 1 is the positioning mode under the groove; when the user selects surgical method Ⅰ, the rotary cutter needs to be placed under the mass to be excised, ensuring that both the mass and the groove are presented within the ultrasound field of view. In this mode, the ultrasound probe needs to be stably placed above the skin corresponding to the mass to avoid large movements to maintain image stability. After the algorithm is started, it will continuously monitor the spatial position relationship between the groove and the mass. When both the groove and the mass appear in the image, it is also necessary to satisfy that the horizontal groove is on the left side of the mass and there is a spatial intersection vertically to determine that the groove has reached the position suitable for activating the rotary cutter; let the upper left corner coordinates of the groove be (x1, y1), the upper left corner coordinates of the mass be (x2, y2), the length of the groove be L1, the width be H1, the length of the mass be L2, and the width be H2. The specific determination logic is based on the geometric dimensions and coordinate relationships of the two, and is processed in the following three cases:

[0061] First, the size of the groove is larger than the mass: when the length difference L1 - L2 > 20 pixels is satisfied, horizontally, it is necessary to satisfy x1 - L1 * 0.2 < x2, and vertically, it is necessary to satisfy y2 - H2 * 1.1 < y1, so as to ensure that the groove covers the left side of the mass in the horizontal direction and there is an intersection between the two after the groove is opened vertically;

[0062] Second, the size of the knife groove is smaller than the mass: When the length difference L2 - L1 > 20 pixels, the horizontal determination condition is adjusted to x2 - L2 * 0.2 < x1, and the vertical direction still uses y2 - H2 * 1.1 < y1. By comparing the reverse coordinates, it is ensured that the mass can be excised after the knife groove is opened;

[0063] Third, the size of the knife groove is close to that of the mass: If the length difference between the two is less than 20 pixels, to unify the calculation logic, on the premise of keeping the upper left corner coordinates of the mass unchanged, the length of the mass is extended to 1.2 times the original size, and then the determination rules of the second case above are followed to achieve compatibility detection under different sizes.

[0064] Surgical method two is the side positioning mode of the knife groove. For surgical method two, the user needs to place the rotary cutter on the side of the mass. This operation requires the ultrasonic probe to move alternately between the rotary cutter and the skin above the mass, resulting in their alternating appearance in the ultrasonic field of view. First, through the analysis of multiple frames of images, the spatial position where the knife groove appears is accurately recorded, and the real-time position of the mass is continuously tracked and detected. The system enables the same spatial position determination algorithm as in surgical method one to dynamically calculate the coordinate and size relationship between the knife groove and the mass. When the cumulative number of frames where the positions of the two coincide reaches 500 frames, the system will trigger a prompt mechanism to clearly inform the user that the rotary cutter is in the ideal excision position, providing a reliable decision-making basis for subsequent surgical operations.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the claims of the present invention.

Claims

1. An intraoperative positioning system for ultrasonic breast tumor atherectomy, characterized by: It includes training data acquisition module, target detection model construction module, model training module, target detection module and instruction activation module; The training data acquisition module acquires effective training data based on the ultrasonic breast tumor excision surgery video to construct a basic data set, and divides it into a training set, a validation set and a test set; The target detection model construction module builds a target detection model based on an improved YOLOv11 network, which is used to detect and track the positions of lumps and grooves in ultrasound video streams in real time. The improved YOLOv11 network introduces a MobileNetv4 inverted bottleneck block and a universal bottleneck block after the first convolution-batch normalization module of the backbone network, and introduces a lightweight C3K2 module after the convolution module of the neck network. The MobileNetv4 inverted bottleneck block and the universal bottleneck block are used to extract low-level features and mid- and high-level features, respectively. The lightweight C3K2 module combines local information clipping with cross-stage partial connection technology to achieve small target detection. The model training module trains the target detection model based on the training data; The target detection module uses a trained target detection model to detect the video frames of ultrasonic breast tumor resection surgery, and detects and tracks the positions of the tumor and the knife groove in the ultrasound video stream in real time; The instruction start module determines whether to issue a rotary cutter start instruction based on the positions of the mass and the knife groove in the ultrasound video stream detected in real time by the target detection module and according to the rotary cutter start determination method.

2. The intraoperative positioning system for ultrasonic breast tumor atherectomy according to claim 1, characterized in that: The system also includes a real-time visualization module; the real-time visualization module includes a display screen and a visualization interface for displaying the ultrasound video stream in real time; each frame of the ultrasound video image is detected by the target detection model, and different labels are given for the mass and the veneer, and the labeled results of the mass and the veneer are superimposed on the original video screen for display.

3. The intraoperative positioning system for ultrasonic breast tumor atherectomy according to claim 1, characterized in that: The specific method for the training data acquisition module to acquire effective training data and construct a basic data set based on ultrasonic breast tumor excision surgery videos is as follows: Screen and capture video clips of ultrasonic breast tumor excision surgery; capture the original ultrasonic breast tumor excision surgery video, and capture the complete operation process from the moment the excision knife penetrates the surface of the breast tissue to the moment the knife groove opens to perform tumor excision; A uniform sampling strategy is used for key frame extraction; The tumor location and the operative knife location in each key frame image are annotated; the annotated key frame images are enhanced using data enhancement technology to construct a basic data set.

4. The intraoperative positioning system for ultrasonic breast tumor atherectomy according to claim 1, characterized in that: In the backbone network design of the improved YOLOV11 network, convolutional layers combined with batch normalization are used as the basic feature extraction units. A layer-by-layer channel number increasing strategy is adopted to capture multi-scale features of images from small to large scales. At the same time, two convolution kernels are used for local features, and the feature map size is compressed through a downsampling operation with an initial step size of 2. The MobileNetV4 inverted bottleneck block is based on the principle of depthwise separable convolution and is used in the low-level feature extraction stage. Through the design of small channel number and small step size, the amount of calculation is greatly reduced while ensuring the feature extraction effect. The general bottleneck block integrates depthwise convolution and dilated convolution for mid- and high-level feature extraction.

5. The intraoperative positioning system based on ultrasonic breast tumor atherectomy according to claim 1, characterized in that: The above-mentioned method for judging the opening of the rotary cutting tool constructs an intelligent algorithm for prompting that the tool groove of the rotary cutting tool reaches a specified position. This algorithm will continuously monitor the spatial position relationship between the tool groove and the mass. When the tool groove and the mass appear on the image at the same time, it is also necessary to meet the condition that the horizontal tool groove is located on the left side of the mass and there is a spatial intersection in the vertical direction to judge whether the tool groove has reached a position suitable for opening the rotary cutting tool.

6. The intraoperative positioning system for ultrasonic breast tumor atherectomy according to claim 1, characterized in that: The method for judging whether the tool groove has reached a position suitable for opening the rotary cutting tool is as follows: Let the upper left corner coordinates of the tool groove be (x1, y1), the upper left corner coordinates of the mass be (x2, y2), the length of the tool groove be L1, the width be H1, the length of the mass be L2, and the width be H2. Based on the geometric dimensions and coordinate relationship between the tool groove and the mass, the following three cases are judged: First, the size of the tool groove is larger than that of the mass: When the length difference between the tool groove and the mass L1 - L2 > n pixels, where n is a set pixel value, the horizontal judgment condition is x1 - L1 * 0.2 < x2, and the vertical judgment condition is y2 - H2 * 1.1 < y1, so as to ensure that the tool groove covers the left side of the mass in the horizontal direction and there is an intersection between the two after the tool groove is opened in the vertical direction; Second, the size of the tool groove is smaller than that of the mass: When the length difference L2 - L1 > n pixels, the horizontal judgment condition is adjusted to x2 - L2 * 0.2 < x1, and the vertical condition remains y2 - H2 * 1.1 < y1; Third, the sizes of the tool groove and the mass are smaller than the set value: If the length difference between the two is less than n pixels, on the premise of keeping the upper left corner coordinates of the mass unchanged, the length of the mass is extended to 1.2 times of the original size, and then the judgment conditions in the second case are followed to achieve compatibility detection under different sizes.

Images