A medical ultrasound image dataset construction method, system, and storage medium
By using a multimodal segmentation model and a similarity-driven mask propagation mechanism, the problem of low efficiency in constructing medical ultrasound image datasets is solved, and high-quality datasets can be constructed efficiently and at low cost.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2025-06-25
- Publication Date
- 2026-05-29
Smart Images
Figure CN120766056B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ultrasound image processing technology, and in particular to a method, system and storage medium for constructing medical ultrasound image datasets. Background Technology
[0002] Acquiring and interpreting optimal ultrasound images is a fundamental skill during ultrasound-guided regional nerve blocks, which involves the identification of key ultrasound anatomical structures. However, anatomical structures can be complex and varied, and despite continuous improvements in ultrasound image quality, the ability to acquire or interpret correct ultrasound views remains a limiting factor. Therefore, clinicians need not only a solid theoretical foundation but also extensive practical clinical experience. This limits the widespread adoption of ultrasound-guided nerve blocks in hospitals at all levels and also means that training an anesthesiologist proficient in ultrasound-guided nerve blocks requires a considerable amount of time.
[0003] In recent years, with the emergence of Artificial Intelligence (AI), especially deep learning-based AI technology, it has begun to make its mark in the field of medical imaging, providing doctors with powerful auxiliary tools to accurately segment target tissues and greatly improve the efficiency and accuracy of medical image analysis. This technology can be used for teaching, training, and even clinical applications. However, deep learning models typically require large, high-quality datasets for training, while the annotation of medical images, especially ultrasound images, is complex and time-consuming, demanding a high level of expertise from annotators. This has become a significant factor limiting the widespread application of AI in the field of medical imaging.
[0004] Traditional ultrasound image datasets mostly used polygon methods for annotating different tissues. This approach was not only time-consuming but also required experienced physicians to perform the annotations manually. This method demanded highly skilled annotators, resulting in low data acquisition efficiency. With the development of semi-automatic segmentation models in recent years, such as the Segment Anything Model (SAM), a cue-based semi-automatic segmentation model, semi-automatic annotation of untrained images has become possible. This technology has been quickly applied to medical image dataset construction, effectively improving dataset construction efficiency and reducing the time cost for physician annotation. However, existing medical semi-automatic segmentation models largely rely on visual cues (bounding boxes or positive and negative sample points) and textual prompts. The desired segmentation result is obtained by selecting the bounding box or positive sample points of the segmentation region. While this improves annotation efficiency, it still requires physicians with extensive clinical experience, making data acquisition costs still very high. Therefore, improving the construction efficiency and data quality of ultrasound image datasets while reducing the professional skill requirements for annotators has become an urgent problem to be solved. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art by providing a method, system, and storage medium for constructing medical ultrasound image datasets, improving the efficiency and quality of ultrasound image dataset construction, and reducing the requirements for the professional level of annotation personnel.
[0006] The objective of this invention can be achieved through the following technical solutions:
[0007] A method for constructing a medical ultrasound image dataset includes the following steps:
[0008] S1: Acquire ultrasound images, mask and annotate the anatomical structures in the ultrasound images, and construct the initial dataset;
[0009] S2: Train the pre-built multimodal segmentation model based on the initial dataset;
[0010] S3: Obtain the ultrasound image to be labeled, calculate the similarity between the ultrasound image to be labeled and the labeled ultrasound images in the initial dataset, and select the most similar set of labeled ultrasound images as reference images.
[0011] S4: Based on the calculated similarity, the mask corresponding to the reference image is weighted and averaged, and converted into a binary mask, which is used as the fusion mask of the ultrasound image to be labeled; based on the calculated similarity, the bounding rectangle region corresponding to the mask of the reference image is weighted and fused to obtain a weighted fused bounding rectangle, which is used as a bounding box prompt for the ultrasound image to be labeled.
[0012] S5: Use the ultrasound image to be labeled, the corresponding fusion mask, text labels, and bounding box prompts as input to the trained multimodal segmentation model to generate the segmentation result of the ultrasound image to be labeled;
[0013] S6: For the segmentation results of the ultrasound images to be labeled output by the multimodal segmentation model, select a confidence index to calculate the confidence level, and set a confidence threshold for quality assessment and review.
[0014] S7: After adding the ultrasound image segmentation results approved in S6 to the initial dataset, the new initial dataset is used to train the multimodal segmentation model and to select reference images for subsequent ultrasound images to be labeled.
[0015] Furthermore, the similarity calculation process includes:
[0016] Adjust the image resolution of the ultrasound image to be labeled and the labeled ultrasound images in the initial dataset to the same size;
[0017] The feature vectors of each image in the ultrasound image to be labeled and the labeled ultrasound image are obtained using a pre-trained image encoder.
[0018] The cosine similarity method is used to calculate the similarity between the feature vector of the ultrasound image to be labeled and the feature vectors of each labeled ultrasound image. The corresponding calculation expression is:
[0019]
[0020] In the formula, s i Let A be the similarity between the ultrasound image to be labeled and the i-th labeled ultrasound image, A be the feature vector of the ultrasound image to be labeled, B be the feature vector of the labeled ultrasound image, · be the vector dot product, and ‖·‖ be the Euclidean distance between the feature vectors.
[0021] Furthermore, during the similarity calculation process, the method assigns higher weights to temporally similar frames based on temporal continuity, resulting in a weighted cosine similarity, which is used as the final similarity calculation result. The corresponding calculation expression is:
[0022] s i =ω t s i
[0023] In the formula, ω t With s i The product magnitude is given to s i ω t The time weight is assigned based on the similarity between the time sequence of the labeled ultrasound image and the time sequence of the ultrasound image to be labeled.
[0024] Furthermore, the generation process of the binary mask is specifically as follows:
[0025] The weighted average of the masks of the reference images is calculated based on the similarity scores. The corresponding calculation expression is:
[0026]
[0027] In the formula, M fused (p) is the weighted average mask at pixel position p, M i M is the mask of the i-th reference image. i (p) is the mask value of the i-th reference image at pixel position p, s i It is the similarity between the ultrasound image to be labeled and the i-th reference image;
[0028] The weighted average mask is converted into a binary mask by setting a threshold. The corresponding calculation expression is:
[0029]
[0030] In the formula, M binary (p) is the binary mask at pixel position p, and T is the threshold of the binary mask.
[0031] Furthermore, the specific process for generating the weighted fused bounding rectangle is as follows:
[0032] For each reference image mask, calculate the corresponding bounding rectangle region B. i =[x min ,y min ,x max ,y max ], where x min y min x max y max Let x and y represent the minimum and maximum values of the bounding rectangle on the x and y axes, respectively. Then, based on the calculated similarity, a weighted fused bounding rectangle is calculated. The corresponding calculation expression is:
[0033]
[0034] In the formula, B fused For the weighted fused bounding rectangle, s i Let x be the similarity between the ultrasound image to be labeled and the i-th reference image, where n is the number of reference images, and x is the similarity between the reference image and the reference image. min,i ,y min,i ,x max,i ,y max,i Let x represent the minimum and maximum values of the bounding rectangular region of the i-th reference image along the x and y axes.
[0035] Furthermore, the method uses the predicted intersection-union ratio as a confidence index, and the quality assessment and review process is as follows: if the prediction result of the multimodal segmentation model is higher than the confidence threshold, it is considered that the prediction result does not need to be manually corrected and is directly added to the initial dataset; if the prediction result is lower than the confidence threshold, the prediction result is manually corrected and then added to the initial dataset.
[0036] Furthermore, in step S1, the polygon annotation method is used to complete the annotation of the anatomical structures in the ultrasound image.
[0037] Furthermore, the multimodal segmentation model is an improved SAM model, which includes a prompt encoding module for processing masks, text labels, and box selection prompts.
[0038] This embodiment also provides a system for implementing the above-described method for constructing a medical ultrasound image dataset, including:
[0039] The image import module is used to load ultrasound images;
[0040] The manual annotation module is used to mask and annotate the anatomical structures in the loaded ultrasound images with text labels to build the initial dataset;
[0041] The multimodal segmentation model training module is used to train a pre-built multimodal segmentation model based on an initial dataset;
[0042] The image similarity calculation module is used to acquire ultrasound images to be labeled, calculate the similarity between the ultrasound images to be labeled and the labeled ultrasound images in the initial dataset, and select the most similar set of labeled ultrasound images as reference images.
[0043] The mask propagation module is used to perform weighted averaging of the masks corresponding to the reference images based on the calculated similarity, and convert them into binary masks, which are then used as the fusion mask for the ultrasound image to be labeled; it also performs weighted fusion of the bounding rectangle regions corresponding to the masks of the reference images based on the calculated similarity, resulting in a weighted fusion bounding rectangle, which serves as a bounding box selection prompt for the ultrasound image to be labeled; the ultrasound image to be labeled, the corresponding fusion mask, text labels, and bounding box selection prompts are used as input to the trained multimodal segmentation model to generate the segmentation result of the ultrasound image to be labeled, which is then added to the initial dataset;
[0044] The dataset management module is used to organize and maintain datasets.
[0045] This embodiment also provides a computer-readable storage medium storing a computer program, which is executed by a processor using the method described above.
[0046] Compared with the prior art, the present invention has the following advantages:
[0047] This invention introduces a multimodal segmentation model and a similarity-driven mask propagation mechanism. Using a manually labeled initial dataset, the multimodal segmentation model is trained. Simultaneously, reference images are selected by calculating the similarity between the ultrasound image to be labeled and the labeled dataset. A weighted average of the masks and bounding rectangles of each reference image is then calculated based on the similarity to obtain a fused mask and bounding box hints. These, along with text labels, serve as input to the multimodal segmentation model, enabling automatic generation of predictive masks. Furthermore, confidence assessment ensures the accuracy of the annotations, significantly reducing the workload of manual annotation and lowering the professional skill requirements for annotators. This greatly improves the efficiency of dataset annotation and the speed of manual correction, enabling the rapid construction of high-quality, large-scale medical ultrasound image datasets. Attached Figure Description
[0048] Figure 1 This is a flowchart illustrating a method for constructing a medical ultrasound image dataset provided in an embodiment of the present invention;
[0049] Figure 2 This is a schematic diagram illustrating an embodiment of the present invention for obtaining a reference image based on image similarity;
[0050] Figure 3 This is a schematic diagram of a confidence assessment process provided in an embodiment of the present invention. Detailed Implementation
[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0052] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0053] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0054] Example 1
[0055] like Figure 1 As shown, this embodiment provides a method for constructing a medical ultrasound image dataset, including the following steps:
[0056] S1: Acquire ultrasound images, mask and annotate the anatomical structures in the ultrasound images, and construct the initial dataset;
[0057] S2: Train the pre-built multimodal segmentation model based on the initial dataset;
[0058] S3: Obtain the ultrasound image to be labeled, calculate the similarity between the ultrasound image to be labeled and the labeled ultrasound images in the initial dataset, and select the most similar set of labeled ultrasound images as reference images.
[0059] S4: Based on the calculated similarity, the mask corresponding to the reference image is weighted and averaged, and converted into a binary mask, which is used as the fusion mask of the ultrasound image to be labeled; based on the calculated similarity, the bounding rectangle region corresponding to the mask of the reference image is weighted and fused to obtain a weighted fused bounding rectangle, which is used as a bounding box prompt for the ultrasound image to be labeled.
[0060] S5: Take the ultrasound image to be labeled, the corresponding fusion mask, text labels, and bounding box prompts as input to the trained multimodal segmentation model to generate the segmentation result of the ultrasound image to be labeled.
[0061] The fusion of masks, text labels, and bounding box cues is obtained through a weighted fusion of reference images and serves as input to the multimodal segmentation model. Using only one or two of these cues makes the model susceptible to data quality and annotation errors. For example, using only bounding box cues may lead to misclassification due to inaccurate annotations, while using only mask cues may affect the segmentation results due to noise or artifacts. By fusing multiple cues, the model can form complementary relationships among different cues, reducing its dependence on a single cue and thus improving its adaptability to diverse and complex scenes.
[0062] S6: For the prediction results of the ultrasound images to be labeled output by the multimodal segmentation model, the prediction crossover ratio is used as the confidence index, and a confidence threshold is set. If the prediction result of the multimodal segmentation model is higher than the confidence threshold, it is considered that the prediction result does not need to be manually corrected and is directly added to the initial dataset; if the prediction result is lower than the confidence threshold, the prediction result is manually corrected and then added to the initial dataset.
[0063] Essentially, the quality of the generated segmentation mask is assessed through a confidence evaluation mechanism, and the intersection-union ratio score is used to determine whether additional visual cues or manual corrections from the user are needed.
[0064] S7: After adding the ultrasound image segmentation results approved in S6 to the initial dataset, the new initial dataset is used to train the multimodal segmentation model and to select reference images for subsequent ultrasound images to be labeled.
[0065] This is equivalent to adding the reviewed and corrected labeled data to the already labeled dataset for iterative updates.
[0066] Preferably, in step S1, a polygon annotation method is used to accurately annotate the key anatomical structures in the ultrasound image.
[0067] In this embodiment, the process of constructing the initial dataset specifically includes:
[0068] Import the ultrasound images to be labeled;
[0069] It provides annotation tools such as polygons, brushes, and erasers to adapt to the complexity and variability of different anatomical structures;
[0070] Supports real-time previewing and modification of different label annotation results;
[0071] Supports exporting the true grayscale image of the annotation results;
[0072] The manual annotation was done by doctors with at least 5 years of clinical experience to ensure the high quality of the initial dataset, ultimately resulting in a mask and text label for each ultrasound image.
[0073] In step S2, the multimodal segmentation model is an improved Segment Anything Model (SAM) model. The SAM model has a cue encoding module for processing multimodal inputs such as masks, text labels and box selection prompts, and outputs a segmentation mask and a predicted intersection-over-union ratio.
[0074] Since the SAM model's prompt encoder does not provide a specific implementation method for using text prompts as input, this invention uses CLIP Text Encoder to generate text feature vectors for different ultrasound tissue labels, and projects them onto the prompt encoder as input through two layers of MLP. During model training, CLIP Text Encoder is frozen and not trained.
[0075] In step S3, the similarity calculation process includes:
[0076] Adjust the image resolution of the ultrasound images to be labeled and the labeled ultrasound images in the initial dataset to the same size M×N;
[0077] The feature vectors of each image in the ultrasound image to be labeled and the labeled ultrasound image are obtained using a pre-trained image encoder.
[0078] The cosine similarity method is used to calculate the similarity between the feature vectors of the ultrasound image to be labeled and those of the labeled ultrasound image. This method is particularly suitable for ultrasound images acquired from the same anatomical location, different patients, or the same patient at different time points. If the two images have similar anatomical structures, the cosine similarity will be closer to 1; if the content of the two images differs significantly, the cosine similarity will be closer to 0 or a negative value. The formula for calculating cosine similarity is:
[0079]
[0080] In the formula, s i Let A be the similarity between the ultrasound image to be labeled and the i-th labeled ultrasound image, where A is the feature vector of the ultrasound image to be labeled, B is the feature vector of the labeled ultrasound image (e.g., the 512-dimensional vector extracted by the CLIP text encoder), A·B represents the vector dot product, and ||A|| and ||B|| represent the Euclidean distance between the vectors.
[0081] Preferably, for consecutive frames captured from a video stream, considering that images that are close in time are more likely to be similar, temporal continuity is taken into account, and frames that are close in time are given higher weights; the weighted cosine similarity calculation formula is:
[0082] s i=ω t s i
[0083] In the formula, ω t With s i The product magnitude is given to s i ω t The time weight (e.g., 1.2) is assigned based on the similarity between the time sequence of the labeled ultrasound image and the time sequence of the ultrasound image to be labeled, with higher weights given to several frames near the current frame.
[0084] Set a similarity threshold (e.g., 0.9) to filter out images with too low similarity and retain only a set of highly similar candidate images as reference images.
[0085] In step S4, the specific process of generating the binary mask is as follows:
[0086] For each text label, a weighted average of the reference mask of the reference image is calculated based on similarity, and a binary mask is generated by thresholding. For each pixel position p, the formula for calculating the weighted average mask is:
[0087]
[0088] In the formula, M fused (p) is the weighted average mask at pixel position p, M i M is the mask of the i-th reference image. i (p) is the mask value (0 or 1) of the i-th reference image at pixel position p, s i It is the similarity between the ultrasound image to be labeled and the i-th reference image;
[0089] For the weighted average mask M fused It is a floating-point matrix between [0,1]. It is converted into a binary mask by setting a threshold. For each pixel position p, the threshold calculation formula is:
[0090]
[0091] In the formula, M binary (p) is the binary mask at pixel position p, and T is the threshold of the binary mask (e.g., 0.6), which can be dynamically adjusted to ensure that the generated binary mask is neither too large nor too small.
[0092] The process of generating the weighted fused bounding rectangle includes: extracting the bounding rectangle region of the reference mask, and generating a more stable selection tooltip through weighted fusion; specifically:
[0093] For each reference image, the mask M i Calculate the corresponding bounding rectangle region B. i =[xmin ,y min ,x max ,y max ], where x min y min x max y max Let x and y represent the minimum and maximum values of the bounding rectangle on the x and y axes, respectively. Then, based on the calculated similarity, a weighted fused bounding rectangle is calculated. The corresponding calculation expression is:
[0094]
[0095] In the formula, B fused For the weighted fused bounding rectangle, s i Let x be the similarity between the ultrasound image to be labeled and the i-th reference image, where n is the number of reference images, and x is the similarity between the reference image and the reference image. min,i ,y min,i ,x max,i ,y max,i Let x represent the minimum and maximum values of the bounding rectangular region of the i-th reference image along the x and y axes.
[0096] Align and match the fused binary mask and bounding box prompts with the image to be labeled;
[0097] In step S5, the text labels, fusion mask, and bounding box prompts are used as inputs to the multimodal segmentation model to generate the segmentation results of the image to be labeled.
[0098] In this embodiment, the confidence assessment in step S6 specifically includes:
[0099] The predicted IoU (Intersection over Union) ratio output by the multimodal segmentation model is used as a confidence index. If it is higher than a preset threshold (e.g., 0.8), the predicted mask is considered reliable and does not require manual correction; the predicted mask is directly output. If it is lower than the threshold, the user needs to obtain the corrected predicted mask by clicking on a prompt or by directly using a drawing tool. Users can efficiently correct the generated mask by clicking on positive sample points in false negative regions or negative sample points in false positive regions.
[0100] In this embodiment, the data iteration update process in step S7 specifically includes:
[0101] The reviewed and corrected labeled data is added to the labeled dataset. This serves two purposes: firstly, to iteratively optimize the multimodal segmentation model and improve segmentation and labeling accuracy; and secondly, to select reference images for the ultrasound images to be labeled.
[0102] Example 2
[0103] This embodiment provides a system for implementing a medical ultrasound image dataset construction method as described in Embodiment 1, comprising:
[0104] The image import module is used to load ultrasound images;
[0105] The manual annotation module is used to mask and annotate the anatomical structures in the loaded ultrasound images with text labels to build the initial dataset;
[0106] The multimodal segmentation model training module is used to train a pre-built multimodal segmentation model based on an initial dataset;
[0107] The image similarity calculation module is used to acquire ultrasound images to be labeled, calculate the similarity between the ultrasound images to be labeled and the labeled ultrasound images in the initial dataset, and select the most similar set of labeled ultrasound images as reference images.
[0108] The mask propagation module is used to perform weighted averaging of the masks corresponding to the reference images based on the calculated similarity, and convert them into binary masks, which are then used as the fusion mask for the ultrasound image to be labeled. It also performs weighted fusion of the bounding rectangle regions corresponding to the masks of the reference images based on the calculated similarity, resulting in a weighted fusion bounding rectangle, which serves as a bounding box selection prompt for the ultrasound image to be labeled. The ultrasound image to be labeled, the corresponding fusion mask, text labels, and bounding box selection prompts are used as input to the trained multimodal segmentation model to generate the segmentation result of the ultrasound image to be labeled, which is then added to the initial dataset.
[0109] Confidence assessment module: Calculates the quality score of the segmentation results, evaluates the quality of the generated segmentation mask, and provides manual correction function;
[0110] The dataset management module is used to organize and maintain datasets.
[0111] It should be noted that the specific content and beneficial effects of the system in this application can be found in the above method embodiments, and will not be repeated here.
[0112] The functions described above herein can be implemented as computer software programs, which are tangibly contained in a computer-readable storage medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded into and / or installed onto a device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, the functions described above can be performed.
[0113] The computer program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This computer program code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the computer program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.
[0114] This embodiment also provides a computer-readable storage medium storing a computer program, which is executed by a processor as the medical ultrasound image dataset construction method of Embodiment 1.
[0115] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium can be a machine-readable signal medium or a machine-readable storage medium. A computer-readable storage medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.
[0116] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A method for constructing a medical ultrasound image dataset, characterized in that, Includes the following steps: S1: Acquire ultrasound images, mask and annotate the anatomical structures in the ultrasound images, and construct the initial dataset; S2: Train the pre-built multimodal segmentation model based on the initial dataset; S3: Obtain the ultrasound image to be labeled, calculate the similarity between the ultrasound image to be labeled and the labeled ultrasound images in the initial dataset, and select the most similar set of labeled ultrasound images as reference images. S4: Based on the calculated similarity, the mask corresponding to the reference image is weighted and averaged, and converted into a binary mask, which is used as the fusion mask of the ultrasound image to be labeled; based on the calculated similarity, the bounding rectangle region corresponding to the mask of the reference image is weighted and fused to obtain a weighted fused bounding rectangle, which is used as a bounding box prompt for the ultrasound image to be labeled. S5: Use the ultrasound image to be labeled, the corresponding fusion mask, text labels, and bounding box prompts as input to the trained multimodal segmentation model to generate the segmentation result of the ultrasound image to be labeled; S6: For the segmentation results of the ultrasound images to be labeled output by the multimodal segmentation model, select a confidence index to calculate the confidence level, and set a confidence threshold for quality assessment and review. S7: After adding the ultrasound image segmentation results approved in S6 to the initial dataset, the new initial dataset is used to train the multimodal segmentation model and to select reference images for subsequent ultrasound images to be labeled. The specific process of generating the binary mask is as follows: The weighted average of the masks of the reference images is calculated based on the similarity scores. The corresponding calculation expression is: In the formula, pixel position The weighted average mask, It is the mask of the i-th reference image. It is the mask value of the i-th reference image at pixel position p. It is the similarity between the ultrasound image to be labeled and the i-th reference image; The weighted average mask is converted into a binary mask by setting a threshold. The corresponding calculation expression is: In the formula, pixel position The binary mask, The threshold value for the binary mask; The specific process for generating the weighted fused bounding rectangle is as follows: For each reference image mask, calculate the corresponding bounding rectangle region. ,in, Let x and y represent the minimum and maximum values of the bounding rectangle on the x and y axes, respectively. Then, based on the calculated similarity, a weighted fused bounding rectangle is calculated. The corresponding calculation expression is: In the formula, To weighted merge the bounding rectangles, Let be the similarity between the ultrasound image to be labeled and the i-th reference image. n The number of reference images, Let x represent the minimum and maximum values of the bounding rectangular region of the i-th reference image along the x and y axes.
2. The method for constructing a medical ultrasound image dataset according to claim 1, characterized in that, The similarity calculation process includes: Adjust the image resolution of the ultrasound image to be labeled and the labeled ultrasound images in the initial dataset to the same size; The feature vectors of each image in the ultrasound image to be labeled and the labeled ultrasound image are obtained using a pre-trained image encoder. The cosine similarity method is used to calculate the similarity between the feature vector of the ultrasound image to be labeled and the feature vectors of each labeled ultrasound image. The corresponding calculation expression is: In the formula, Let be the similarity between the ultrasound image to be labeled and the i-th labeled ultrasound image. Let be the feature vector of the ultrasound image to be labeled. The feature vector of the labeled ultrasound image. For vector dot product, is the Euclidean distance between the feature vectors.
3. The method for constructing a medical ultrasound image dataset according to claim 2, characterized in that, The method, during the similarity calculation process, also assigns higher weights to temporally similar frames based on temporal continuity, obtaining a weighted cosine similarity as the final similarity calculation result. The corresponding calculation expression is: In the formula, and Assign the product to , The time weight is assigned based on the similarity between the time sequence of the labeled ultrasound image and the time sequence of the ultrasound image to be labeled.
4. The method for constructing a medical ultrasound image dataset according to claim 1, characterized in that, The method uses the predicted intersection-union ratio as a confidence index. The quality assessment and review process is as follows: if the prediction result of the multimodal segmentation model is higher than the confidence threshold, it is considered that the prediction result does not need to be manually corrected and is directly added to the initial dataset; if the prediction result is lower than the confidence threshold, the prediction result is manually corrected and then added to the initial dataset.
5. The method for constructing a medical ultrasound image dataset according to claim 1, characterized in that, In step S1, the polygon annotation method is used to annotate the anatomical structures in the ultrasound image.
6. The method for constructing a medical ultrasound image dataset according to claim 1, characterized in that, The multimodal segmentation model is an improved SAM model, which includes a prompt encoding module for processing masks, text labels, and box selection prompts.
7. A system for implementing a method for constructing a medical ultrasound image dataset as described in any one of claims 1-6, characterized in that, include: The image import module is used to load ultrasound images; The manual annotation module is used to mask and annotate the anatomical structures in the loaded ultrasound images with text labels to build the initial dataset; The multimodal segmentation model training module is used to train a pre-built multimodal segmentation model based on an initial dataset; The image similarity calculation module is used to acquire ultrasound images to be labeled, calculate the similarity between the ultrasound images to be labeled and the labeled ultrasound images in the initial dataset, and select the most similar set of labeled ultrasound images as reference images. The mask propagation module is used to perform weighted averaging of the masks corresponding to the reference images based on the calculated similarity, and convert them into binary masks, which are then used as the fusion mask for the ultrasound image to be labeled; it also performs weighted fusion of the bounding rectangle regions corresponding to the masks of the reference images based on the calculated similarity, resulting in a weighted fusion bounding rectangle, which serves as a bounding box selection prompt for the ultrasound image to be labeled; the ultrasound image to be labeled, the corresponding fusion mask, text labels, and bounding box selection prompts are used as input to the trained multimodal segmentation model to generate the segmentation result of the ultrasound image to be labeled, which is then added to the initial dataset; The dataset management module is used to organize and maintain datasets.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is executed by a processor according to any one of claims 1 to 6.