WO2022042348A1 - 医学影像标注方法和装置、设备及存储介质 - Google Patents

医学影像标注方法和装置、设备及存储介质 Download PDF

Info

Publication number
WO2022042348A1
WO2022042348A1 PCT/CN2021/112677 CN2021112677W WO2022042348A1 WO 2022042348 A1 WO2022042348 A1 WO 2022042348A1 CN 2021112677 W CN2021112677 W CN 2021112677W WO 2022042348 A1 WO2022042348 A1 WO 2022042348A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
images
similarity
labeling
annotation
Prior art date
Application number
PCT/CN2021/112677
Other languages
English (en)
French (fr)
Inventor
代黎明
姜泓羊
张冬冬
Original Assignee
北京至真互联网技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京至真互联网技术有限公司 filed Critical 北京至真互联网技术有限公司
Publication of WO2022042348A1 publication Critical patent/WO2022042348A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/40—ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/53—Querying
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
    • G06F17/147—Discrete orthonormal transforms, e.g. discrete cosine transform, discrete sine transform, and variations therefrom, e.g. modified discrete cosine transform, integer transforms approximating the discrete cosine transform
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Definitions

  • the present disclosure relates to the field of computer-aided medical technology, and in particular, to a medical image labeling method and device, equipment, and storage medium.
  • the accumulation of medical image data is the basis for deep learning technology to play a role in medical image artificial intelligence scenarios, which not only requires a sufficient amount of medical image data, but also requires medical images to have high-quality annotation information.
  • deep learning technology is divided into supervised learning and unsupervised learning, and supervised learning is very dependent on data annotation information.
  • the labeling of medical image data needs to be performed by professional doctors.
  • the quality of labeling is limited by the level of professional doctors, and the quantity of labeling is limited by the number of doctors and their energy.
  • some semi-automatic image annotation methods have been proposed. These methods take mature deep learning models as the core, and use trained deep learning models for image annotation of specific tasks.
  • the labeled images need to be reviewed or corrected by experts. Finally complete the labeling work.
  • the standardization and labeling of medical image data is the basis for the development of current medical artificial intelligence technology. How to improve the efficiency and quality of medical image labeling, improve automation, and reduce human intervention is a technical problem that urgently needs to be considered and solved.
  • a medical image labeling method including:
  • Image annotation is performed on the image to be annotated according to the annotation category.
  • the set of multi-class benchmark images obtained includes:
  • the remaining images are classified as corresponding reference images.
  • performing similarity matching between the reference image and the remaining images in the image database to obtain the color similarity includes:
  • the H component, the S component and the V component in each described HSV color space are respectively divided according to the interval of the preset number and the color feature is calculated;
  • the color similarity is obtained by using Euclidean distance according to the color histogram vector of the reference image and the remaining image.
  • performing similarity matching between the reference image and the remaining images in the image database to obtain the similarity of structural information includes:
  • the structural information similarity is obtained by using the Hamming distance according to the integer of the reference image and the residual image.
  • performing image annotation on the image to be annotated according to the annotation information includes:
  • Image annotation is performed according to the specific structure of the specific structure image
  • Image annotation is performed according to the image feature of the image feature image.
  • performing image annotation according to the specific structure of the specific structure image includes:
  • the correlation coefficient method to perform template matching on the specific structure image according to the self-built image template library to obtain the similarity of each area in the specific structure image; wherein, the images in the self-built image template library include the specific structure. ;
  • performing image annotation according to the image feature of the image feature image includes:
  • Image annotation is performed on the image feature image using the trained deep learning model.
  • a medical image labeling device including a similarity scoring module, an image acquisition module to be labelled, and an image labeling module;
  • the similarity scoring module is configured to select multiple types of reference images from the image database, and perform similarity matching between the remaining images in the image database and the selected reference images, and classify the selected reference images according to the similarity matching results.
  • the remaining images are classified into corresponding reference images, and a multi-class reference image set is constructed and obtained;
  • the to-be-labeled image acquisition module is configured to determine the to-be-labeled image from the reference image set, and to determine the to-be-labeled image's labeling category based on labeling information of the to-be-labeled image;
  • the image labeling module is configured to perform image labeling on the image to be labelled according to the labeling category.
  • a medical image annotation device including:
  • memory for storing processor-executable instructions
  • the processor is configured to implement any of the foregoing methods when executing the executable instructions.
  • non-volatile computer-readable storage medium having computer program instructions stored thereon that, when executed by a processor, implement any of the foregoing methods.
  • the remaining images are classified into the corresponding reference images according to the similarity matching results.
  • construct a multi-type reference image set determine the image to be labeled from the reference image set, determine the labeling category of the to-be-labeled image based on the labeling information of the to-be-labeled image, and perform image labeling on the to-be-labeled image according to the labeling category.
  • the present disclosure uses color similarity and content structure similarity to standardize image content, solves the problem of selecting candidate images for automatic labeling, saves the labor time cost of previous image cleaning, and improves the quality and speed of image labeling.
  • FIG. 1 shows a flowchart of a medical image labeling method according to an embodiment of the present disclosure
  • FIG. 2 shows a schematic diagram of an image standardization method of a medical image labeling method according to an embodiment of the present disclosure
  • FIG. 3 shows a schematic diagram of an automatic labeling method of a medical image labeling method according to an embodiment of the present disclosure
  • FIG. 4 shows a block diagram of a medical image labeling apparatus according to an embodiment of the present disclosure
  • FIG. 5 shows a block diagram of a medical image labeling device according to an embodiment of the present disclosure.
  • FIG. 1 shows a flowchart of a medical image labeling method according to an embodiment of the present disclosure.
  • the medical image annotation method includes:
  • Step S100 select multiple types of reference images from the image database, and perform similarity matching between the remaining images in the image database and the selected reference images, and classify the remaining images into the corresponding reference images according to the similarity matching result.
  • construct a multi-type reference image set step S200, determine the image to be labeled from the reference image set, determine the labeling category of the to-be-labeled image based on the labeling information of the to-be-labeled image, and step S300, perform image labeling on the to-be-labeled image according to the labeling category.
  • the remaining images are classified into the corresponding reference images according to the similarity matching results.
  • Obtaining a multi-class reference image set determining an image to be labeled from the reference image set, determining the labeling category of the to-be-labeled image based on the labeling information of the to-be-labeled image, and performing image labeling on the to-be-labeled image according to the labeling class.
  • the present disclosure uses color similarity and content structure similarity to standardize image content, solves the problem of selecting candidate images for automatic labeling, saves the labor time cost of previous image cleaning, and improves the quality and speed of image labeling.
  • step S100 is performed, multiple types of reference images are selected from the image database, and the remaining images in the image database are matched with the selected reference images for similarity, and the remaining images are matched according to the similarity matching result. Classify them into the corresponding benchmark images, and construct a multi-class benchmark image set.
  • similarity matching is performed between the remaining images in the image database and the selected reference images, and the remaining images are classified into the corresponding reference images according to the similarity matching results, and a multi-class reference image is constructed to obtain
  • the image collection includes: performing color similarity matching and/or structural information similarity matching on the reference image and the remaining images in the image database to obtain the corresponding color similarity and/or structural information similarity, according to the color similarity and/or structural information similarity.
  • the relationship between the similarity and the corresponding preset threshold value is used to classify the remaining images as the corresponding reference images.
  • performing similarity matching between the reference image and the remaining images in the image database to obtain the color similarity includes: converting the reference image and the remaining images into the HSV color space, and converting the H component, S component and V component in each HSV color space Divide and calculate the color features according to the preset number of intervals respectively, obtain the color histogram vector according to the color features, and obtain the color similarity according to the color histogram vector of the reference image and the remaining image using the Euclidean distance. For example, select multiple categories of reference images from an existing image database, each category of reference images includes multiple reference images, and perform color similarity matching between the reference images and the remaining images in the image database.
  • Both the reference image and the remaining image are converted into HSV color space, wherein the reference image and the remaining image are both RGB images, and further, the H component (hue), S component (saturation), and V component (light and dark) components are evenly divided , where the H component is divided into 8 intervals, that is, H ⁇ [0,1,...,7], the S component is divided into 3 intervals, that is, S ⁇ [0,1,2], the V component is divided into 3 intervals, That is, V ⁇ [0,1,2], and finally calculate the color feature according to formula 1.
  • Formula 1 is as follows:
  • the value range of F is [0,71], and each image can obtain a 72-dimensional color histogram vector, and then use the Euclidean distance to calculate the color similarity of the two images according to the color histogram vectors of the two images.
  • performing similarity matching between the reference image and the remaining image in the image database to obtain the similarity of the structural information includes: performing size adjustment and grayscale processing on the reference image and the remaining image to obtain a grayscale image. , calculate the mean value of the frequency information in the low-frequency region after the discrete cosine transform of the grayscale image, if the frequency information of any pixel in the low-frequency region is greater than the mean value, set the frequency information corresponding to the pixel to 1, if any one of the low-frequency region The frequency information of the pixel points is less than the average value, the frequency information corresponding to the pixel points is set to 0, and the frequency information of each pixel point is arranged according to the preset order to generate integers, and the Hamming distance is used to obtain similar structural information according to the integers of the reference image and the remaining images.
  • f(i,j) is the original signal
  • F(u,v) is the transformed signal
  • c(u) and c(v) are the compensation coefficients
  • N is the maximum sequence point in the time domain
  • the value of N is 32.
  • each grayscale image F(u, v) exemplarily, that is, an 8 ⁇ 8 region, that is, u ⁇ [0,7], v ⁇ [0,7], and calculate The average value of the extracted frequency information in the low-frequency area, and the frequency information of each pixel in the 8 ⁇ 8 area is compared with the average value. If the frequency information of the pixel point is greater than the average value of the frequency information in the low-frequency area, the frequency The information is set to 1. If the frequency information of the pixel is less than the average value of the frequency information of the low frequency area, the frequency information of the point is set to 0, and the obtained frequency information is arranged in a predetermined order into a 64-bit integer, that is, the image is obtained. Then, according to the 64-bit integer of the two grayscale images, the Hamming distance can be used to calculate the similarity of the structural information of the two images.
  • the similarity score is greater than or equal to the preset threshold, the remaining images corresponding to the similarity score are classified into the group corresponding to the reference image.
  • a preset similarity threshold may be used, and the preset similarity threshold includes color Similarity threshold (T color ) and content structure similarity threshold (T content ), find images similar to the reference image from the remaining images in the image database, and if it is greater than or equal to the similarity threshold, classify the remaining images as preset In the corresponding group of the multiple reference images, that is, the picture set corresponding to the reference image, otherwise, it is returned to the image database.
  • T color color
  • T content structure similarity threshold T content
  • a picture is input from the remaining images in the image database, and a similarity score is performed between this picture and each category of benchmark images, wherein the benchmark image includes K categories, and the color of the picture and the first category benchmark image is If the similarity is less than the color similarity threshold and the structural information similarity, then the picture is returned to the image database, and the similarity score is continued.
  • the color similarity between the picture and the second category reference image is greater than the color similarity threshold and the structural information similarity. Then the picture is classified into the second category of reference pictures.
  • the images in each group can be checked and verified manually, and the reference images can be adjusted and updated according to the results of the grouping.
  • Image normalization After the remaining images are divided into multiple groups, the images in each group can be checked and verified manually, and the reference images can be adjusted and updated according to the results of the grouping. Image normalization.
  • step S200 is performed, an image to be annotated is determined from the reference image set, and an annotation category of the image to be annotated is determined based on the annotation information of the image to be annotated.
  • the images to be annotated are manually selected, that is, selected from a set of benchmark images corresponding to the benchmark images, and the images to be annotated are added with the identifiers of the images to be marked, and then the images to be marked can be obtained according to the identifiers.
  • Labeling images in addition, artificially adding labeling information to each image, and then directly obtaining the labeling information of each image, so that the labeling category of the image to be labelled can be determined according to the labeling information.
  • two types of annotation information can be added to the image to be annotated according to human experience, namely specific structure annotation information and image feature annotation information.
  • the fundus image in medical images as an example, the fundus image
  • the optic disc and macula can be considered as typical specific structural information, some typical lesions, exemplary: hemorrhage, exudation, proliferative membrane, etc., will reflect different patients, different onset times, and different imaging conditions. different image features.
  • the specific structure annotation information is artificially set to "A”
  • the image feature annotation information is set to "B”. After acquiring the image to be annotated, if the annotation information corresponding to the image to be annotated is "A”, It is determined that the marked image is an image of a specific structure, and if the marked information corresponding to the image to be marked is "B", the marked image is determined to be an image feature image.
  • step S300 is performed, and image annotation is performed on the image to be annotated according to the annotation category.
  • performing image annotation on the image to be annotated according to the annotation information includes: dividing the image to be annotated into a specific structure image and an image feature image according to the annotation information, performing image annotation according to the specific structure of the specific structure image, and performing image annotation according to the specific structure of the image.
  • the image features of the feature images are used for image annotation.
  • performing image labeling according to the specific structure of the specific structure image includes: performing template matching on the specific structure image using the correlation coefficient method according to the self-built image template library to obtain the similarity of each region in the specific structure image, wherein the self-built image template library
  • the images in all include specific structures. If the similarity is greater than the preset similarity threshold, the area corresponding to the similarity will be marked on the image.
  • the template image in the image template library is used to perform template matching on a specific structure image using the correlation coefficient matching method, wherein the template matching using the correlation coefficient matching method is calculated using formula 3;
  • T(x,y) represents the template image
  • I(x,y) represents the specific structure image
  • w and h are the width and height of the template image, respectively
  • R(x,y) represents the coordinate (x,y) Correlation coefficient
  • (x 0 , y 0 ) represents the absolute coordinates on the template T
  • (x+x 0 , y+y 0 ) represents the absolute coordinates on the image I.
  • the similarity between the specific structure image and the template image is calculated by formula 3. If the similarity of a certain area in the specific structure image is greater than the preset threshold, the area will be labeled, that is, the image labeling is completed, and the label can be edited. , so as to facilitate human modification and adjustment.
  • performing image labeling according to the image features of the image feature image includes: using a deep learning model to learn the images in the labeled self-built image database to obtain a trained deep learning model, wherein the self-built deep learning model is The images in the image database all contain image features, and the trained deep learning model is used to annotate the image feature images.
  • a labelled image database that is, a labelled self-built image database, in which the images in the labelled image database are images that have been labelled by professionals.
  • a deep learning model is used to learn the labeled self-built image database.
  • the classification model ResNet or the target detection model FasterRcnn can be used to learn the labeled self-built image database, and transfer learning can be used. Enhance the learning effect of the deep learning model, and then use the trained deep learning model to perform image annotation on the image feature image, that is, generate a label, and associate the label with the corresponding area.
  • the format of the tag may be a csv file or an xml file.
  • the embodiments of the present disclosure do not limit the type of the deep learning model, as long as the purpose can be achieved, and similarly, the format of the label is not limited, and the purpose can be achieved.
  • the remaining images are classified into the corresponding reference images according to the similarity matching results.
  • construct a multi-type reference image set determine the image to be labeled from the reference image set, determine the labeling category of the to-be-labeled image based on the labeling information of the to-be-labeled image, and perform image labeling on the to-be-labeled image according to the labeling category.
  • the present disclosure uses color similarity and content structure similarity to standardize image content, and solves the problem of candidate image selection for automatic labeling, so that labor time costs for image cleaning in the early stage are saved, and the quality and speed of image labeling are improved.
  • a medical image labeling apparatus 100 is also provided. Since the working principle of the medical image labeling apparatus 100 according to the embodiment of the present disclosure is the same as or similar to that of the medical image labeling method according to the embodiment of the present disclosure, the repeated description will not be repeated.
  • the medical image labeling apparatus 100 according to the embodiment of the present disclosure includes: a similarity scoring module 110 , an image acquisition module 120 to be labelled, and an image labeling module 130 ;
  • the similarity scoring module 110 is configured to perform similarity matching between the remaining images in the image database and the reference image, and classify the remaining images into the corresponding reference image set according to the similarity matching;
  • the to-be-labeled image obtaining module 120 is configured to obtain the to-be-labeled image from the reference image set, and to determine the labeling category of the to-be-labeled image based on labeling information of the to-be-labeled image;
  • the image labeling module 130 is configured to perform image labeling on the image to be labelled according to the labeling category.
  • a medical image labeling device 200 is also provided.
  • a medical image labeling apparatus 200 includes a processor 210 and a memory 220 for storing instructions executable by the processor 210 .
  • the processor 210 is configured to implement any of the medical image labeling methods described above when executing the executable instructions.
  • the number of processors 210 may be one or more.
  • the medical image labeling apparatus 200 in the embodiment of the present disclosure may further include an input device 230 and an output device 240 .
  • the processor 210, the memory 220, the input device 230, and the output device 240 may be connected through a bus, or may be connected in other ways, which are not specifically limited here.
  • the memory 220 can be used to store software programs, computer-executable programs, and various modules, such as programs or modules corresponding to the medical image labeling method in the embodiments of the present disclosure.
  • the processor 210 executes various functional applications and data processing of the medical image labeling device 200 by running the software programs or modules stored in the memory 220 .
  • the input device 230 may be used to receive input numbers or signals. Wherein, the signal may be the generation of a key signal related to user setting and function control of the device/terminal/server.
  • the output device 240 may include a display device such as a display screen.
  • a non-volatile computer-readable storage medium on which computer program instructions are stored, and when the computer program instructions are executed by the processor 210, any one of the medical image annotation described above is implemented method.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Analysis (AREA)
  • Pure & Applied Mathematics (AREA)
  • Library & Information Science (AREA)
  • Mathematical Optimization (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Mathematics (AREA)
  • Primary Health Care (AREA)
  • Public Health (AREA)
  • Discrete Mathematics (AREA)
  • Medical Informatics (AREA)
  • General Health & Medical Sciences (AREA)
  • Epidemiology (AREA)
  • Radiology & Medical Imaging (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Algebra (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

一种医学影像标注方法和装置、设备及存储介质，该方法包括：由图像数据库中选取出多类基准图像，并将图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将剩余图像归类到对应的基准图像中，构建得到多类基准图像集合(S100)，从基准图像集合中确定待标注图像，基于待标注图像的标注信息确定待标注图像的标注类别(S200)，根据标注类别对待标注图像进行图像标注(S300)。上述方法采用色彩相似度和内容结构相似度对图像内容标准化，解决了自动标注的候选图像选取问题，节省了前期图像清洗的人工时间成本，而且提高了图像标注的质量和速度。

Description

医学影像标注方法和装置、设备及存储介质
本申请要求于2020年08月26日提交中国专利局、申请号为202010871983.1、发明名称为“医学影像标注方法和装置、设备及存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本公开涉及计算机辅助医疗技术领域，尤其涉及一种医学影像标注方法和装置、设备及存储介质。
背景技术
医学影像数据的积累是深度学习技术在医学影像人工智能场景发挥作用的基础，这不仅要求医学影像数据具有足够的数量，而且要求医学影像具备优质的标注信息。在医学人工智能技术的落地发展中，深度学习技术分为监督式学习和非监督式学习，而监督式学习十分依赖于数据的标注信息。医学图像数据的标注工作需要专业的医生来执行，标注的质量受到专业医生水平的限制，标注的数量受到医生的数量及其精力的限制，最终导致医学图像的标注工作效率难以得到提高。目前，已存在一些半自动图像标注方法被提出，这些方法以成熟的深度学习模型为核心，使用训练好的深度学习模型用于特定任务的图像标注，标注后的图像则需要专家进行复核或修正，最终完成标注的工作。医学影像数据标准化和标注是当前医学人工智能技术得以发展的基础，如何提高医学影像的标注效率和质量，提升自动化、降低人为干预，是当前急需思考和解决的技术难题。
发明内容
有鉴于此，本公开提出了一种医学影像标注方法，包括：
由图像数据库中选取出多类基准图像，并将所述图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将所述剩余图像归类到对应的基准图像中，构建得到多类基准图像集合；
从所述基准图像集合中确定待标注图像，基于待标注图像的标注信息 确定所述待标注图像的标注类别；
根据所述标注类别对所述待标注图像进行图像标注。
在一种可能的实现方式中，将所述图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将所述剩余图像归类到对应的基准图像中，构建得到多类基准图像集合包括：
将所述图像数据库中的基准图像和剩余图像进行色彩相似度匹配和/或结构信息相似度匹配，得到对应的色彩相似度和/或结构信息相似度；
根据所述色彩相似度和/或所述结构相似度与所对应的预设阈值的大小关系，将所述剩余图像归类为对应的基准图像。
在一种可能的实现方式中，将所述图像数据库中的基准图像和剩余图像进行相似度匹配得到色彩相似度包括：
将所述基准图像和所述剩余图像转换成HSV色彩空间；
将各所述HSV色彩空间中的H分量、S分量和V分量分别按预设数量的区间进行划分并计算颜色特征；
依据所述颜色特征得到颜色直方图向量；
依据所述基准图像和所述剩余图像的颜色直方图向量使用欧氏距离得到所述色彩相似度。
在一种可能的实现方式中，将所述图像数据库中的基准图像和剩余图像进行相似度匹配得到结构信息相似度包括：
将所述基准图像和所述剩余图像进行尺寸调整和灰度处理得到灰度图像；
将所述灰度图像的离散余弦变换后计算低频区的频率信息的均值；
若所述低频区中任一像素点的频率信息大于所述均值，将所述像素点对应的频率信息设置为1；
若所述低频区中任一像素点的频率信息小于所述均值，将所述像素点对应的频率信息设置为0；
将各所述像素点的频率信息依据预设顺序排列生成整数；
依据所述基准图像和所述剩余图像的所述整数使用汉明距离得到所 述结构信息相似度。
在一种可能的实现方式中，根据所述标注信息对所述待标注图像进行图像标注包括：
将所述待标注图像依据所述标注信息分为特定结构图像和影像特征图像；
依据所述特定结构图像的特定结构进行图像标注；
依据所述影像特征图像的影像特征进行图像标注。
在一种可能的实现方式中，依据所述特定结构图像的特定结构进行图像标注包括：
依据自建图像模板库对所述特定结构图像使用相关系数法进行模板匹配得到所述特定结构图像中各区域的相似度；其中，所述自建图像模板库中的图像均包括所述特定结构；
若所述相似度大于预设相似度阈值，则将所述相似度对应的区域进行图像标注。
在一种可能的实现方式中，依据所述影像特征图像的影像特征进行图像标注包括：
使用深度学习模型对已标注的自建图像数据库中的图像进行学习得到训练好的所述深度学习模型；其中，所述自建图像数据库中的图像均包含所述影像特征；
使用训练好的所述深度学习模型对所述影像特征图像进行图像标注。
根据本公开的另一方面，提供了一种医学影像标注装置，包括相似度评分模块、待标注图像获取模块和图像标注模块；
所述相似度评分模块，被配置为由图像数据库中选取出多类基准图像，并将所述图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将所述剩余图像归类到对应的基准图像中，构建得到多类基准图像集合；
所述待标注图像获取模块，被配置为从所述基准图像集合中确定待标注图像，基于待标注图像的标注信息确定所述待标注图像的标注类别；
所述图像标注模块，被配置为根据所述标注类别对所述待标注图像进行图像标注。
根据本公开的另一方面，提供了一种医学影像标注设备，包括：
处理器；
用于存储处理器可执行指令的存储器；
其中，所述处理器被配置为执行所述可执行指令时实现前面任一所述的方法。
根据本公开的另一方面，提供了一种非易失性计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现前面任一所述的方法。
这样，通过由图像数据库中选取出多类基准图像，并将图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将剩余图像归类到对应的基准图像中，构建得到多类基准图像集合，从基准图像集合中确定待标注图像，基于待标注图像的标注信息确定待标注图像的标注类别，根据标注类别对待标注图像进行图像标注。本公开采用色彩相似度和内容结构相似度对图像内容标准化，解决了自动标注的候选图像选取问题，节省了前期图像清洗的人工时间成本，而且提高了图像标注的质量和速度。
根据下面参考附图对示例性实施例的详细说明，本公开的其它特征及方面将变得清楚。
说明书附图
包含在说明书中并且构成说明书的一部分的附图与说明书一起示出了本公开的示例性实施例、特征和方面，并且用于解释本公开的原理。
图1示出本公开实施例的医学影像标注方法的流程图；
图2示出本公开实施例的医学影像标注方法的图像标准化原理图；
图3示出本公开实施例的医学影像标注方法的自动标注原理图；
图4示出本公开实施例的医学影像标注装置的框图；
图5示出本公开实施例的医学影像标注设备的框图。
具体实施方式
以下将参考附图详细说明本公开的各种示例性实施例、特征和方面。附图中相同的附图标记表示功能相同或相似的元件。尽管在附图中示出了实施例的各种方面，但是除非特别指出，不必按比例绘制附图。
在这里专用的词“示例性”意为“用作例子、实施例或说明性”。这里作为“示例性”所说明的任何实施例不必解释为优于或好于其它实施例。
另外，为了更好的说明本公开，在下文的具体实施方式中给出了众多的具体细节。本领域技术人员应当理解，没有某些具体细节，本公开同样可以实施。在一些实例中，对于本领域技术人员熟知的方法、手段、元件和电路未作详细描述，以便于凸显本公开的主旨。
图1示出根据本公开一实施例的医学影像标注方法的流程图。如图1所示，该医学影像标注方法包括：
步骤S100，由图像数据库中选取出多类基准图像，并将图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将剩余图像归类到对应的基准图像中，构建得到多类基准图像集合，步骤S200，从基准图像集合中确定待标注图像，基于待标注图像的标注信息确定待标注图像的标注类别，步骤S300，根据标注类别对待标注图像进行图像标注。
通过由图像数据库中选取出多类基准图像，并将图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将剩余图像归类到对应的基准图像中，构建得到多类基准图像集合，从基准图像集合中确定待标注图像，基于待标注图像的标注信息确定待标注图像的标注类别，根据标注类别对待标注图像进行图像标注。本公开采用色彩相似度和内容结构相似度对图像内容标准化，解决了自动标注的候选图像选取问题，节省了前期图像清洗的人工时间成本，而且提高了图像标注的质量和速度。
具体的，参见图1，执行步骤S100，由图像数据库中选取出多类基准图像，并将图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将剩余图像归类到对应的基准图像中，构建得到多类基准图像集合。
在一种可能的实现方式中，将图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将剩余图像归类到对应的基准图像中，构建得到多类基准图像集合包括：将图像数据库中的基准图像和剩余图像进行色彩相似度匹配和/或结构信息相似度匹配，得到对应的色彩相似度和/或结构信息相似度，根据色彩相似度和/或结构相似度与所对应的预设阈值的大小关系，将剩余图像归类为对应的基准图像。具体的，将图像数据库中的基准图像和剩余图像进行相似度匹配得到色彩相似度包括：将基准图像和剩余图像转换成HSV色彩空间，将各HSV色彩空间中的H分量、S分量和V分量分别按预设数量的区间进行划分并计算颜色特征，依据颜色特征得到颜色直方图向量，依据基准图像和剩余图像的颜色直方图向量使用欧氏距离得到色彩相似度。举例来说，从已有的图像数据库中选择多个类别的基准图像，每一类别基准图像包括多张基准图像，分别将基准图像与图像数据库中的剩余图像进行色彩相似度匹配，首先，将基准图像和剩余图像均转换成HSV色彩空间，其中，基准图像和剩余图像均为RGB图像，进一步的，将H分量(色调)、S分量(饱和度)、V分量(明暗)分量进行均匀划分，其中，H分量分成8个区间，即，H∈[0,1,…,7]，S分量分成3个区间，即，S∈[0,1,2]，V分量分成3个区间，即，V∈[0,1,2]，最后依据公式一计算颜色特征，公式一如下所示：
公式一：
Figure PCTCN2021112677-appb-000001
其中，F的取值范围为[0,71]，每一张图像可得到一个72维的颜色直方图向量，接着依据两个图像的颜色直方图向量使用欧式距离可以计算两个图像的色彩相似度。
进一步的，在一种可能的实现方式中，将图像数据库中的基准图像和剩余图像进行相似度匹配得到结构信息相似度包括：将基准图像和剩余图像进行尺寸调整和灰度处理得到灰度图像，将灰度图像的离散余弦变换后计算低频区的频率信息的均值，若低频区中任一像素点的频率信息大于均值，将像素点对应的频率信息设置为1，若低频区中任一像素点的频率信息小于均值，将像素点对应的频率信息设置为0，将各像素点的频率信息依据预设顺序排列生成整数，依据基准图像和剩余图像的整数使用汉明距离得到结构信息相似度。举例来说：首先，将基准图像和剩余图像的图像尺寸均缩小至32×32像素的大小，接着将基准图像和剩余图像转化成灰度图像，并分别计算基准图像和剩余图像的离散余弦变换(DCT)，二维DCT的变换公式如公式二所示：
公式二：
Figure PCTCN2021112677-appb-000002
Figure PCTCN2021112677-appb-000003
其中，f(i,j)是原始信号，F(u,v)是变换后的信号，c(u)和c(v)是补偿系数，N为时间域的最大序点，且N的值为32。
进一步的，提取每个灰度图像F(u,v)中的低频区信息，示例性的，即8×8区域，即u∈[0,7]，v∈[0,7]，并计算提取到的低频区的频率信息的均值，将8×8区域的每一个像素点的频率信息分别与均值进行比较，若像素点的频率信息大于低频区的频率信息的均值则将该点的频率信息设置为1，若像素点的频率信息小于低频区的频率信息的均值则将该点的频率信息设置为0，按照预定的顺序将得到的频率信息排列成一个64位整数，即得到该图像的内容结构特征，接着依据两个灰度图像的64位整数使用汉明距离可以计算两个图像的结构信息相似度。
进一步的，参见图1，若相似度评分大于或等于预设阈值，则将相似度评分对应的剩余图像归入基准图像对应的组。
在一种可能的实现方式中，在得到色彩相似度和结构信息相似度之后，基于色彩相似度和内容结构相似度的测量，可以依据预设的相似度阈值，预设的相似度阈值包括色彩相似度阈值(T color)和内容结构相似度阈值(T content)，从图像数据库中的剩余图像找到与基准图像相似的图像，若大于或等于相似度阈值，则将剩余图像归类为预设的多个基准图像的对应的组中，即基准图像对应的图片集合，否则返还至图像数据库。示例性的，从图像数据库的剩余图像中输入一张图片，将这张图片与各类别的基准图像做相似度评分，其中，基准图像包括K个类别，该图片与第一类别基准图像的色彩相似度小于色彩相似度阈值和结构信息相似度，则将该图片返回图像数据库，继续进行相似度评分，该图片与第二类别基准图像的色彩相似度大于色彩相似度阈值和结构信息相似度，则将该图片归入第二类别基准图片中。
另外的，参见图2，在将剩余图像分入多个组后，还可以经过人为的抽检和校验每个组中的图像，可以依据分组的结果对基准图像进行调整和更新，即完成了图像标准化。
进一步的，参见图1，执行步骤S200，从基准图像集合中确定待标注图像，基于待标注图像的标注信息确定待标注图像的标注类别。
在一种可能的实现方式中，人为的选择需要的标注的图像，即从基准图像对应的基准图像集合中选取，并将这些需要标注的图像添加待标注图像的标识，然后可以根据标识获取待标注图像，另外的，人为的还对每张图像添加标注信息，接着可以直接获取每张图片的标注信息，以使可以根据标注信息确定待标注图像的标注类别。举例来说，在确定待标注图像之后，可以根据人为经验将待标注图像添加两类标注信息，分别为特定结构标注信息和影像特征标注信息，其中以医学图像中的眼底影像为例，眼底图像中的视盘和黄斑可以被认为是典型的特定结构信息，一些典型的病灶，示例性的：出血、渗出、增殖膜等，会根据不同的患者、不同的发病时间、不同的成像条件，反映出不同的影像特征。示例性的，人为的将特定结构标注信息设置为“A”，将影像特征标注信息设置为“B”，在获取到待标注图像之后，若待标注图像对应的标注信息为“A”，则确定该标注图像为特 定结构图像，若待标注图像对应的标注信息为“B”，则确定该标注图像为影像特征图像。
进一步的，在获取到待标注图像和标注信息时，参见图1，执行步骤S300，根据标注类别对待标注图像进行图像标注。
在一种可能的实现方式中，根据标注信息对待标注图像进行图像标注包括：将待标注图像依据标注信息分为特定结构图像和影像特征图像，依据特定结构图像的特定结构进行图像标注，依据影像特征图像的影像特征进行图像标注。具体的，依据特定结构图像的特定结构进行图像标注包括：依据自建图像模板库对特定结构图像使用相关系数法进行模板匹配得到特定结构图像中各区域的相似度，其中，自建图像模板库中的图像均包括特定结构，若相似度大于预设相似度阈值，则将相似度对应的区域进行图像标注。在依据特定结构图像的特定结构进行图像标注前，还需要使用现有的图像模板库或建立图像模板库，其中，在建立图像模板库时，可以由人为的从待标注图像中选取图像模板，图像模板可以选取多个，然后建立图像模板库。举例来说，使用图像模板库中的模板图像对特定结构图像使用相关系数匹配法进行模板匹配，其中相关系数匹配法进行模板匹配使用公式三进行计算；
公式三：
Figure PCTCN2021112677-appb-000004
其中，T(x,y)表示模板图像，I(x,y)表示特定结构图像，w和h分别为模板图像的宽和高，R(x,y)表示坐标(x,y)处的相关系数，(x 0,y 0)表示模板T上的绝对坐标，(x+x 0,y+y 0)表示图像I上的绝对坐标。
通过公式三计算特定结构图像与模板图像的相似度，若特定结构图像中某区域的相似度大于预设阈值，则将该区域进行标签化处理，即完成了图像标注，且该标签可进行编辑，以使方便人为的修改调整。
在一种可能的实现方式中，依据影像特征图像的影像特征进行图像标注包括：使用深度学习模型对已标注的自建图像数据库中的图像进行学习 得到训练好的深度学习模型，其中，自建图像数据库中的图像均包含影像特征，使用训练好的深度学习模型对影像特征图像进行图像标注。在进行依据影像特征图像的影像特征进行图像标注前，需要先构建一个已标注图像数据库，即已标注的自建图像数据库，其中，已标注图像数据库中的图像为经过专业人士标注的图像，进一步的，举例来说，采用深度学习模型对已标注的自建图像数据库进行学习，具体的，可以使用分类模型ResNet或目标检测模型FasterRcnn对已标注的自建图像数据库进行学习，并且可以使用迁移学习增强深度学习模型的学习效果，接着使用训练后的深度学习模型对影像特征图像进行图像标注，即生成标签，并将标签与对应区域进行关联。其中，标签的格式可以为csv文件或者xml文件。
需要说明的是，本公开的实施例不对深度学习模型的类型进行限定，可以达到目的即可，同样的，不对标签的格式进行限定，可以达到目的即可。
另外的，参见图3，在图像标注完成后，可以进行人为的抽检及校验标注结果，若标注的结果满足要求，则将已经进行图像标注的图像和标签存储于金标图像数据库中，如果不满足要求，则对已经进行图像标注的图像进行纠正或舍弃并返还原图像数据库中，即已完成了自动标注。此外，根据自动标注的结果，调整并更新图像模板库中的模板图像，同时将金标数据添加至深度学习模型的训练数据集中去，增强深度学习模型的标注能力。
需要说明的是，尽管以上述各个步骤作为示例介绍了医学影像标注方法如上，但本领域技术人员能够理解，本公开应不限于此。事实上，用户完全可根据个人喜好和/或实际应用场景灵活设定医学影像标注方法，只要达到所需功能即可。
这样，通过由图像数据库中选取出多类基准图像，并将图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将剩余图像归类到对应的基准图像中，构建得到多类基准图像集合，从基准图像集合中确定待标注图像，基于待标注图像的标注信息确定待标注图像的标注类别，根据标注类别对待标注图像进行图像标注。本公开采用色 彩相似度和内容结构相似度对图像内容标准化，解决了自动标注的候选图像选取问题，以使节省了前期图像清洗的人工时间成本，而且提高了图像标注的质量和速度。
进一步的，根据本公开的另一方面，还提供了一种医学影像标注装置100。由于本公开实施例的医学影像标注装置100的工作原理与本公开实施例的医学影像标注方法的原理相同或相似，因此重复之处不再赘述。参见图4，本公开实施例的医学影像标注装置100包括：相似度评分模块110、待标注图像获取模块120和图像标注模块130；
相似度评分模块110，被配置为将图像数据库中的剩余图像与基准图像进行相似度匹配，依据相似度匹配将剩余图像归入对应的基准图像集合；
待标注图像获取模块120，被配置为从基准图像集合中获取待标注图像，基于待标注图像的标注信息确定待标注图像的标注类别；
图像标注模块130，被配置为根据标注类别对待标注图像进行图像标注。
更进一步地，根据本公开的另一方面，还提供了一种医学影像标注设备200。参阅图5，本公开实施例医学影像标注设备200包括处理器210以及用于存储处理器210可执行指令的存储器220。其中，处理器210被配置为执行可执行指令时实现前面任一所述的医学影像标注方法。
此处，应当指出的是，处理器210的个数可以为一个或多个。同时，在本公开实施例的医学影像标注设备200中，还可以包括输入装置230和输出装置240。其中，处理器210、存储器220、输入装置230和输出装置240之间可以通过总线连接，也可以通过其他方式连接，此处不进行具体限定。
存储器220作为一种计算机可读存储介质，可用于存储软件程序、计算机可执行程序和各种模块，如：本公开实施例的医学影像标注方法所对应的程序或模块。处理器210通过运行存储在存储器220中的软件程序或模块，从而执行医学影像标注设备200的各种功能应用及数据处理。
输入装置230可用于接收输入的数字或信号。其中，信号可以为产生与设备/终端/服务器的用户设置以及功能控制有关的键信号。输出装置240可以包括显示屏等显示设备。
根据本公开的另一方面，还提供了一种非易失性计算机可读存储介质，其上存储有计算机程序指令，计算机程序指令被处理器210执行时实现前面任一所述的医学影像标注方法。
以上已经描述了本公开的各实施例，上述说明是示例性的，并非穷尽性的，并且也不限于所披露的各实施例。在不偏离所说明的各实施例的范围和精神的情况下，对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。本文中所用术语的选择，旨在最好地解释各实施例的原理、实际应用或对市场中的技术改进，或者使本技术领域的其它普通技术人员能理解本文披露的各实施例。

Claims (10)

  1. 一种医学影像标注方法，其特征在于，包括：
    由图像数据库中选取出多类基准图像，并将所述图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将所述剩余图像归类到对应的基准图像中，构建得到多类基准图像集合；
    从所述基准图像集合中确定待标注图像，基于待标注图像的标注信息确定所述待标注图像的标注类别；
    根据所述标注类别对所述待标注图像进行图像标注。
  2. 根据权利要求1所述的方法，其特征在于，将所述图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将所述剩余图像归类到对应的基准图像中，构建得到多类基准图像集合包括：
    将所述图像数据库中的基准图像和剩余图像进行色彩相似度匹配和/或结构信息相似度匹配，得到对应的色彩相似度和/或结构信息相似度；
    根据所述色彩相似度和/或所述结构信息相似度与所对应的预设阈值的大小关系，将所述剩余图像归类为对应的基准图像。
  3. 根据权利要求2所述的方法，其特征在于，将所述图像数据库中的基准图像和剩余图像进行相似度匹配得到色彩相似度包括：
    将所述基准图像和所述剩余图像转换成HSV色彩空间；
    将各所述HSV色彩空间中的H分量、S分量和V分量分别按预设数量的区间进行划分并计算颜色特征；
    依据所述颜色特征得到颜色直方图向量；
    依据所述基准图像和所述剩余图像的颜色直方图向量使用欧氏距离得到所述色彩相似度。
  4. 根据权利要求2所述的方法，其特征在于，将所述图像数据库中的基准图像和剩余图像进行相似度匹配得到结构信息相似度包括：
    将所述基准图像和所述剩余图像进行尺寸调整和灰度处理得到灰度图像；
    将所述灰度图像的离散余弦变换后计算低频区的频率信息的均值；
    若所述低频区中任一像素点的频率信息大于所述均值，将所述像素点对应的频率信息设置为1；
    若所述低频区中任一像素点的频率信息小于所述均值，将所述像素点对应的频率信息设置为0；
    将各所述像素点的频率信息依据预设顺序排列生成整数；
    依据所述基准图像和所述剩余图像的所述整数使用汉明距离得到所述结构信息相似度。
  5. 根据权利要求1所述的方法，其特征在于，根据所述标注信息对所述待标注图像进行图像标注包括：
    将所述待标注图像依据所述标注信息分为特定结构图像和影像特征图像；
    依据所述特定结构图像的特定结构进行图像标注；
    依据所述影像特征图像的影像特征进行图像标注。
  6. 根据权利要求5所述的方法，其特征在于，依据所述特定结构图像的特定结构进行图像标注包括：
    依据自建图像模板库对所述特定结构图像使用相关系数法进行模板匹配得到所述特定结构图像中各区域的相似度；其中，所述自建图像模板库中的图像均包括所述特定结构；
    若所述相似度大于预设相似度阈值，则将所述相似度对应的区域进行图像标注。
  7. 根据权利要求5所述的方法，其特征在于，依据所述影像特征图像的影像特征进行图像标注包括：
    使用深度学习模型对已标注的自建图像数据库中的图像进行学习得到训练好的所述深度学习模型；其中，所述自建图像数据库中的图像均包含所述影像特征；
    使用训练好的所述深度学习模型对所述影像特征图像进行图像标注。
  8. 一种医学影像标注装置，其特征在于，包括相似度评分模块、待标注图像获取模块和图像标注模块；
    所述相似度评分模块，被配置为由图像数据库中选取出多类基准图像，并将所述图像数据库中的剩余图像与所选取出的基准图像进行相似度匹配，依据相似度匹配结果将所述剩余图像归类到对应的基准图像中，构建得到多类基准图像集合；
    所述待标注图像获取模块，被配置为从所述基准图像集合中确定待标注图像，基于待标注图像的标注信息确定所述待标注图像的标注类别；
    所述图像标注模块，被配置为根据所述标注类别对所述待标注图像进行图像标注。
  9. 一种医学影像标注设备，其特征在于，包括：
    处理器；
    用于存储处理器可执行指令的存储器；
    其中，所述处理器被配置为执行所述可执行指令时实现权利要求1至7中任意一项所述的方法。
  10. 一种非易失性计算机可读存储介质，其上存储有计算机程序指令，其特征在于，所述计算机程序指令被处理器执行时实现权利要求1至7中任意一项所述的方法。
PCT/CN2021/112677 2020-08-26 2021-08-16 医学影像标注方法和装置、设备及存储介质 WO2022042348A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010871983.1 2020-08-26
CN202010871983.1A CN111986785B (zh) 2020-08-26 2020-08-26 医学影像标注方法和装置、设备及存储介质

Publications (1)

Publication Number Publication Date
WO2022042348A1 true WO2022042348A1 (zh) 2022-03-03

Family

ID=73439931

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/112677 WO2022042348A1 (zh) 2020-08-26 2021-08-16 医学影像标注方法和装置、设备及存储介质

Country Status (2)

Country Link
CN (1) CN111986785B (zh)
WO (1) WO2022042348A1 (zh)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111986785B (zh) * 2020-08-26 2023-09-12 北京至真互联网技术有限公司 医学影像标注方法和装置、设备及存储介质
CN112669090A (zh) * 2021-01-06 2021-04-16 广东创意热店互联网科技有限公司 一种智能生成广告海报的方法、电子设备、计算机可读存储介质
CN112906817A (zh) * 2021-03-16 2021-06-04 中科海拓(无锡)科技有限公司 一种智能图像标注方法
CN113178248A (zh) * 2021-04-28 2021-07-27 联仁健康医疗大数据科技股份有限公司 医学图像数据库建立方法、装置、设备及存储介质
CN113159234B (zh) * 2021-05-24 2021-12-28 广州市吉华勘测股份有限公司 巡检图片的类别标注方法、装置、电子设备和存储介质
CN115713664B (zh) * 2022-12-06 2023-06-09 浙江中测新图地理信息技术有限公司 一种消防验收的智能标注方法及装置

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130195361A1 (en) * 2012-01-17 2013-08-01 Alibaba Group Holding Limited Image index generation based on similarities of image features
CN107563123A (zh) * 2017-09-27 2018-01-09 百度在线网络技术（北京）有限公司 用于标注医学图像的方法和装置
CN110264443A (zh) * 2019-05-20 2019-09-20 平安科技（深圳）有限公司 基于特征可视化的眼底图像病变标注方法、装置及介质
CN110322436A (zh) * 2019-06-19 2019-10-11 广州金域医学检验中心有限公司 医学图像处理方法、装置、存储介质及设备
CN110459299A (zh) * 2019-07-10 2019-11-15 中山大学 一种视网膜眼底彩色照片图像筛选方法
CN111986785A (zh) * 2020-08-26 2020-11-24 北京至真互联网技术有限公司 医学影像标注方法和装置、设备及存储介质

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10679345B2 (en) * 2017-12-20 2020-06-09 International Business Machines Corporation Automatic contour annotation of medical images based on correlations with medical reports
CN109003269B (zh) * 2018-07-19 2021-10-08 哈尔滨工业大学 一种能提高医生效率的医学影像病灶的标注提取方法
CN109359517A (zh) * 2018-08-31 2019-02-19 深圳市商汤科技有限公司 图像识别方法和装置、电子设备、存储介质、程序产品
CN109902672A (zh) * 2019-01-17 2019-06-18 平安科技（深圳）有限公司 图像标注方法及装置、存储介质、计算机设备
CN110993064B (zh) * 2019-11-05 2023-03-21 北京邮电大学 面向深度学习的医学影像标注方法及装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130195361A1 (en) * 2012-01-17 2013-08-01 Alibaba Group Holding Limited Image index generation based on similarities of image features
CN107563123A (zh) * 2017-09-27 2018-01-09 百度在线网络技术（北京）有限公司 用于标注医学图像的方法和装置
CN110264443A (zh) * 2019-05-20 2019-09-20 平安科技（深圳）有限公司 基于特征可视化的眼底图像病变标注方法、装置及介质
CN110322436A (zh) * 2019-06-19 2019-10-11 广州金域医学检验中心有限公司 医学图像处理方法、装置、存储介质及设备
CN110459299A (zh) * 2019-07-10 2019-11-15 中山大学 一种视网膜眼底彩色照片图像筛选方法
CN111986785A (zh) * 2020-08-26 2020-11-24 北京至真互联网技术有限公司 医学影像标注方法和装置、设备及存储介质

Also Published As

Publication number Publication date
CN111986785A (zh) 2020-11-24
CN111986785B (zh) 2023-09-12

Similar Documents

Publication Publication Date Title
WO2022042348A1 (zh) 医学影像标注方法和装置、设备及存储介质
WO2021008328A1 (zh) 图像处理方法、装置、终端及存储介质
CN110232383B (zh) 一种基于深度学习模型的病灶图像识别方法及病灶图像识别系统
CN108229509B (zh) 用于识别物体类别的方法及装置、电子设备
CN108335303B (zh) 一种应用于手掌x光片的多尺度手掌骨骼分割方法
CN110059697B (zh) 一种基于深度学习的肺结节自动分割方法
US9020250B2 (en) Methods and systems for building a universal dress style learner
CN108921057B (zh) 基于卷积神经网络的对虾形态测量方法、介质、终端设备及装置
TWI709085B (zh) 用於對車輛損傷影像進行損傷分割的方法、裝置、電腦可讀儲存媒體和計算設備
CN111931751B (zh) 深度学习训练方法、目标物识别方法、系统和存储介质
WO2022089257A1 (zh) 医学图像处理方法、装置、设备、存储介质及产品
CN110765882B (zh) 一种视频标签确定方法、装置、服务器及存储介质
KR102245218B1 (ko) 기계학습을 위한 정규화 방법 및 그 장치
CN110807775A (zh) 基于人工智能的中医舌像分割装置、方法及存储介质
US20240005494A1 (en) Methods and systems for image quality assessment
Wang et al. Medical matting: a new perspective on medical segmentation with uncertainty
CN115641323A (zh) 医学图像自动标注的方法及装置
JP2019180749A (ja) 画像処理プログラム、画像処理装置、および画像処理方法
CN111507288A (zh) 图像检测方法、装置、计算机设备和存储介质
CN113822323A (zh) 脑部扫描图像的识别处理方法、装置、设备及存储介质
CN112070181A (zh) 一种基于图像流的协同检测方法及装置、存储介质
JP2011141866A (ja) データ分類装置、データ分類システムおよびプログラム
CN107341189A (zh) 一种辅助人工对图像进行筛查、分类和存储的方法及系统
JP2015041293A (ja) 画像認識装置および画像認識方法
CN116071635A (zh) 基于结构性知识传播的图像识别方法与装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21860186

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21860186

Country of ref document: EP

Kind code of ref document: A1