WO2021143019A1 - 基于人工智能的图像搜索方法、装置、设备及介质 - Google Patents

基于人工智能的图像搜索方法、装置、设备及介质 Download PDF

Info

Publication number
WO2021143019A1
WO2021143019A1 PCT/CN2020/093333 CN2020093333W WO2021143019A1 WO 2021143019 A1 WO2021143019 A1 WO 2021143019A1 CN 2020093333 W CN2020093333 W CN 2020093333W WO 2021143019 A1 WO2021143019 A1 WO 2021143019A1
Authority
WO
WIPO (PCT)
Prior art keywords
historical
distribution map
dose distribution
target
image
Prior art date
Application number
PCT/CN2020/093333
Other languages
English (en)
French (fr)
Inventor
张晓颖
王季勇
Original Assignee
平安科技(深圳)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 平安科技(深圳)有限公司 filed Critical 平安科技(深圳)有限公司
Publication of WO2021143019A1 publication Critical patent/WO2021143019A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/58Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/583Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/30Determination of transform parameters for the alignment of images, i.e. image registration
    • G06T7/33Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10072Tomographic images
    • G06T2207/10081Computed x-ray tomography [CT]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20081Training; Learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20084Artificial neural networks [ANN]

Definitions

  • This application relates to the technical field of artificial intelligence medical image processing, and in particular to an artificial intelligence-based image search method, device, equipment and medium.
  • the embodiments of the present application provide an artificial intelligence-based image search method, device, equipment, and medium to solve the problem of low efficiency or low accuracy of current radiotherapy dose distribution maps.
  • the radiotherapy plan database is queried based on the target feature vector, and a target dose distribution map matching the target feature vector is obtained.
  • An image search request acquisition module for acquiring an image search request, the image search request including a target user identification, an original CT image corresponding to the target user identification, and an original organ-at-risk outline drawing;
  • An original dose distribution map acquisition module configured to input the original CT map and the original organ-at-risk map into a dose analysis model to generate an original dose distribution map corresponding to the target user identification;
  • the standard dose distribution map acquisition module is used to perform registration processing on the original dose distribution map to acquire the standard dose distribution map
  • a target feature vector acquiring module configured to input the standard dose distribution map into an image search model to acquire a target feature vector corresponding to the target user identifier
  • a computer device includes a memory, a processor, and computer-readable instructions that are stored in the memory and can run on the processor, and the processor implements the following steps when the processor executes the computer-readable instructions:
  • the radiotherapy plan database is queried based on the target feature vector, and a target dose distribution map matching the target feature vector is obtained.
  • One or more readable storage media storing computer readable instructions, when the computer readable instructions are executed by one or more processors, the one or more processors execute the following steps:
  • the radiotherapy plan database is queried based on the target feature vector, and a target dose distribution map matching the target feature vector is obtained.
  • FIG. 1 is a schematic diagram of an application environment of an image search method based on artificial intelligence in an embodiment of the present application
  • FIG. 2 is a flowchart of an image search method based on artificial intelligence in an embodiment of the present application
  • Fig. 3 is another flowchart of an image search method based on artificial intelligence in an embodiment of the present application
  • FIG. 5 is another flowchart of an image search method based on artificial intelligence in an embodiment of the present application.
  • Fig. 10 is a schematic diagram of a computer device in an embodiment of the present application.
  • the image search method based on artificial intelligence provided in the embodiment of the present application can be applied to the application environment shown in FIG. 1.
  • the artificial intelligence-based image search method is applied in an image search system.
  • the image search system includes a client and a server as shown in FIG.
  • Corresponding original original CT image and original organ-at-risk map are generated and processed to generate a standard dose distribution diagram, and the historical registration distribution diagram similar to the standard dose distribution diagram is quickly searched through the image search model as the target dose distribution diagram to improve the target
  • the acquisition efficiency and accuracy of the dose distribution map can be used as a reference for clinicians to develop target radiotherapy plans.
  • the target radiotherapy plan is a radiotherapy plan for target users.
  • the client is also called the client, which refers to the program that corresponds to the server and provides local services to the client.
  • the client can be installed on, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices.
  • the server can be implemented as an independent server or a server cluster composed of multiple servers.
  • an image search method based on artificial intelligence is provided.
  • the method is applied to the server in FIG. 1 as an example for description, including the following steps:
  • S201 Acquire an image search request, where the image search request includes the target user identification, the original CT image corresponding to the target user identification, and the original outline drawing of the organ at risk.
  • the original CT image is the image obtained by the target user through a CT scan.
  • CT Computer Tomography
  • the detector receives the X-rays that pass through the layer and converts it into visible light, which changes from photoelectric to photoelectric conversion.
  • the electrical signal is converted into a digital signal by an analog/digital converter.
  • the cross-sectional or three-dimensional image of the body under examination is the original CT image for use
  • the original CT image found small lesions in the inspected part.
  • the inspected part includes a diseased part and a non-diseased part.
  • the inspected part is the lung
  • the diseased part is the right lung
  • the non-diseased part is the left lung and organs around the left lung.
  • Organs at risk refer to important non-diseased tissues or organs within the radiation range of radiotherapy rays.
  • the original organ-at-risk sketch map refers to a map obtained by sketching non-diseased important tissues or organs within the radiation range of radiotherapy rays in the original CT image, and the original organ-at-risk sketch map corresponds to the target user.
  • the clinician generates an image search request with the target user ID, the original CT image corresponding to the target user ID, and the original threat sketch map by clicking the image search button on the client, and sends the image search request to The server, so that the server can obtain the image search request.
  • S202 Input the original CT image and the original outline of the organ at risk into the dose analysis model, and generate an original dose distribution diagram corresponding to the target user identification.
  • the original dose distribution map is the radiation dose predicted by the dose analysis model, that is, the radiation dose required to predict the diseased part of the target user when the target user is subjected to radiotherapy. Since the original dose distribution map is generated by the dose analysis model, the original dose distribution map may ignore the spatial location between different organs, and the accuracy is not high and cannot meet the clinical standards. Therefore, clinicians cannot directly generate the original dose distribution map. Target radiotherapy plan.
  • S203 Perform registration processing on the original dose distribution map to obtain a standard dose distribution map.
  • the registration process is used to compare or fuse images acquired by different users under different conditions, so that images can be accurately searched later. Understandably, due to differences in body types of different users, different users’ organ sizes or spatial positions are different. By looking for a spatial transformation, the dose distribution map or CT map of different users can be mapped to another image to map different users’ The same image corresponds to the points at the same position in space one by one to eliminate the differences between different individuals.
  • the image registration algorithm is used to register the original CT image with the standard CT image to obtain the standard registration parameters, and the original dose distribution diagram is converted based on the standard registration parameters to obtain the standard dose distribution diagram to ensure that the target user identification corresponds to
  • the organs of the original CT image and the organs of the standard CT image are in corresponding positions, which can eliminate the influence of the size and spatial location of the organs of different individuals on the image search, ensuring that similar images can be searched later, and improving image search Accuracy.
  • the standard CT map refers to a general CT map template.
  • S204 Input the standard dose distribution map into the image search model, and obtain the target feature vector corresponding to the target user identification.
  • the image search model refers to a pre-trained model used to identify dose distribution maps to output feature vectors.
  • the image search model is specifically a model generated by convolutional neural network training based on a three-element loss function, which can ensure that the distance between the feature vectors generated by the image search model for similar images is small, and specifically allows similar registration distribution maps to pass.
  • the distance of the corresponding feature vector generated by the image search model is small, and the distance of the feature vector generated by the image search model for dissimilar images is large. Specifically, the distance of the corresponding feature vector generated by the image search model for dissimilar registration distribution maps is large.
  • the image search model ensures that the distance of the feature vector generated by the image search model for similar image images is smaller than the distance of the feature vector generated by the image search model for dissimilar images, so that the follow-up can accurately obtain the history similar to the standard dose distribution map.
  • the standard dose distribution map is input into the image search model to obtain the target feature vector corresponding to the target user identification, which provides a basis for the subsequent search for the standard dose distribution map.
  • S205 Query the radiotherapy plan database based on the target feature vector, and obtain a target dose distribution map matching the target feature vector.
  • the radiotherapy plan database refers to a database used to store user data corresponding to historical users after radiotherapy.
  • the user data includes associated stored historical user IDs, historical CT maps, historical dose distribution maps, historical radiotherapy plans, and historical dose distributions.
  • historical users refer to users who have undergone radiotherapy.
  • the historical user identifier is an identifier used to uniquely identify historical users.
  • Historical CT images are images acquired by historical users through CT scans.
  • the historical dose distribution map is a dose distribution map formed by historical users during radiotherapy.
  • Historical radiotherapy plans are radiotherapy plans collected by historical users during radiotherapy.
  • the historical feature vector is the feature vector obtained by inputting the historical dose distribution map into the image search model.
  • the target dose distribution map is a historical registration distribution map similar to the standard dose distribution map.
  • the historical feature vector is obtained through the image search model, which can ensure that the distance of the feature vector of the similar historical dose distribution map is small, and the distance of the feature vector of the dissimilar historical dose distribution map is large, so that the follow-up can accurately obtain the similarity to the standard dose distribution map
  • the historical registration distribution map is used to determine the target dose distribution map based on the similar historical dose distribution map.
  • the similarity calculation formula is used to calculate the similarity between the target feature vector and the historical feature vector corresponding to any historical user identifier in the radiotherapy plan database, and the top M (M is a positive integer) historical features with the largest similarity are obtained Vector, the historical dose distribution map corresponding to these historical feature vectors is used as the target dose distribution map, so that the server can find the associated stored historical radiotherapy plan in the target radiotherapy plan database according to the target dose distribution map, and send it to the client for the clinician to follow Historical radiotherapy plans to develop target radiotherapy plans to improve the efficiency and accuracy of the formulation of target radiotherapy plans. Understandably, because the historical radiotherapy plan is a plan that the historical user has already performed radiotherapy, the historical radiotherapy plan has a strong reference value. It aims to shorten the time required for the clinician to formulate the target radiotherapy plan and shorten the radiotherapy cycle of the target user. The user provides timely radiotherapy.
  • the artificial intelligence-based image search method obtaineds the original CT image and the original organ-at-risk outline in the image search request, and inputs the original CT image and the original organ-at-risk outline into the dose analysis model to generate the original dose distribution map , Provide technical support for image search. Perform registration processing on the original dose distribution map and obtain the standard dose distribution map to eliminate the differences between different individuals, so as to achieve the purpose of information fusion.
  • the artificial intelligence-based image search method before step S204, that is, before inputting the standard dose distribution map into the image search model and obtaining the target feature vector corresponding to the target user identifier, the artificial intelligence-based image search method further includes:
  • the historical user image data includes historical CT images, historical organ-at-risk maps, and historical dose distribution maps.
  • the first historical user identification is an identification of a historical user in the image database.
  • the historical user image data is image data stored in the image database in association with the first historical user identification.
  • the historical user image data includes, but is not limited to, historical CT images, historical organ-at-risk maps, and historical dose distribution maps.
  • the historical organ-at-risk sketch map refers to a map obtained by sketching non-diseased important tissues or organs within the radiation range of radiotherapy rays in the historical CT map of the same historical user.
  • S302 Input the historical CT map and the historical organ-at-risk map corresponding to the first historical user identification into the dose analysis model, and obtain the analysis dose distribution map corresponding to the first historical user identification.
  • the analyzed dose distribution map is the historical CT map and the historical organ-at-risk map corresponding to the first historical user identification.
  • the dose distribution model is used to predict the predicted dose distribution map, and the analyzed dose distribution map is used as the training image Search the training data of the model.
  • the historical registration distribution map refers to a map obtained after image registration processing is performed on the historical dose distribution map.
  • the analytical registration distribution map refers to the image obtained after image registration processing is performed on the analytical dose distribution map.
  • an image registration algorithm may be used to register the historical CT image of the first historical user identification with the standard CT image, to obtain the registration parameter corresponding to the first historical user identification, and to compare the historical dose distribution map based on the registration parameter.
  • Perform image registration processing to obtain a historical registration distribution map, perform image registration processing on the analysis dose distribution map based on the registration parameters, and obtain an analysis registration distribution map.
  • the historical registration distribution map and the analysis of the registration distribution map are obtained through registration processing, which can ensure that the trained model is more accurate.
  • S304 Query the image database, and determine the comparative dose distribution map corresponding to the first historical user identifier based on the historical dose distribution map of other historical user identifiers.
  • the historical dose distribution maps corresponding to other historical user identifiers other than the first historical user identifier are acquired from the image database, and the historical dose distribution maps corresponding to other historical user identifiers are registered to generate a comparative dose distribution map to eliminate The difference between images of different users ensures the accuracy of the training image search model. Understandably, in order to obtain more training samples, the historical dose distribution maps corresponding to multiple historical user identifiers other than the first historical user identifier can be registered to obtain a comparative dose distribution map to obtain a sufficient number of Training samples.
  • the historical registration distribution map, the analysis registration distribution map, and the comparative dose distribution map corresponding to the first historical user identification are used as training samples. It is understandable that since the historical registration distribution map and the analysis registration distribution map are the same history The user’s image data has higher similarity, and the contrast dose distribution map is the dose distribution map obtained after registration processing on the historical dose distribution map except the first historical user ID, and is corresponding to the first historical user ID.
  • the historical registration distribution map and the analysis registration distribution map are dissimilar dose distribution maps to ensure that the generated image search model can make the distance between the feature vectors corresponding to dissimilar images large, and specifically make the dissimilar registration distribution maps correspond The feature vector distance of is large, which can ensure the accuracy of subsequent image search.
  • S306 Input the training samples into the convolutional neural network based on the ternary loss function for model training, and obtain an image search model.
  • the historical registration distribution map, the analysis registration distribution map, and the contrast dose distribution map corresponding to the first historical user identification are input into the convolutional neural network based on the weight sharing of the ternary loss function for training, when the loss is less than the function convergence Value, it means that the training of the image search model is completed.
  • the function convergence value is a preset value used to evaluate whether the loss function meets the convergence requirement, and it can be zero.
  • the ternary loss function is M (M is a positive integer) represents the number of training samples, i (i is a positive integer, i ⁇ M) represents the i-th group of training samples, x a represents the vector corresponding to the historical registration distribution map, and x p represents the analysis registration distribution
  • the artificial intelligence-based image search method inputs the historical CT map and the historical organ-at-risk map corresponding to the first historical user identification into the dose analysis model to quickly obtain the analysis dose distribution corresponding to the first historical user identification picture. Based on the historical dose distribution map and the analysis dose distribution map corresponding to the first historical user identification, the historical registration distribution map and the analysis registration distribution map are obtained, so as to eliminate the influence of the difference between the images of different users on the model training. Query the image database, determine the contrast dose distribution map corresponding to the first historical user ID based on the historical dose distribution map of other historical user IDs, compare the historical registration distribution map corresponding to the first historical user ID, analyze the registration distribution map and the contrast dose The distribution map is used as a training sample to ensure the accuracy of subsequent image searches.
  • the feature vector distance corresponding to the registration distribution map is large.
  • step S303 which is based on the historical dose distribution map and the analyzed dose distribution map corresponding to the first historical user identification, obtains the historical registration distribution map and analyzes the registration distribution map, including:
  • S401 Use an image registration algorithm to register the historical CT image corresponding to the first historical user identifier with the standard CT image, and obtain historical registration parameters.
  • the historical CT image is preprocessed to provide a basis for registration.
  • the preprocessing process includes: noise elimination processing on the historical CT image to eliminate interference factors; when the pixel size of the historical CT image is different from that of the standard CT image , The size of the historical CT image is adjusted to match the pixel size of the historical CT image and the standard CT image, so that the characteristics of the historical CT image and the standard CT image are corresponding, so as to ensure the accuracy of the acquired historical registration parameters.
  • the first feature is selected from the specific position of the preprocessed historical CT image
  • the second feature is selected from the position corresponding to the standard CT image and the historical CT image, so as to obtain the first feature and the second feature corresponding to the same specific position
  • the three-dimensional coordinates of the second feature of the standard CT image determine the registration function of the preprocessed historical CT image and the standard CT image based on the two three-dimensional coordinates, and obtain the historical registration parameters based on the registration function.
  • the historical registration parameters are the parameters of the registration function
  • the historical CT images are resampled to verify the accuracy of the historical registration parameters.
  • S402 Perform image registration on the historical dose distribution map and the analyzed dose distribution map based on the historical registration parameters, and obtain the historical registration distribution map and analyze the registration distribution map.
  • the historical dose distribution map and the analysis dose distribution map corresponding to the first historical user identification are registered according to the historical registration parameters, that is, the historical dose distribution map and the analysis dose distribution map corresponding to the first historical user identification according to the historical registration parameters
  • the distribution map undergoes spatial transformation.
  • the spatial transformation can be conversions such as rotation, reduction, and enlargement to obtain historical registration distribution maps and analyze registration distribution maps. Since the training samples are all registered, the difference between images of different users can be eliminated and the image generation can be ensured. The accuracy of the search model.
  • the artificial intelligence-based image search method uses an image registration algorithm to register the historical CT image corresponding to the first historical user identifier with the standard CT image to obtain historical registration parameters.
  • Image registration is performed on the historical dose distribution map and the analysis dose distribution map based on the historical registration parameters, and the historical registration distribution map and the analysis registration distribution map can be obtained, which can eliminate the difference between the images of different users and ensure the accuracy of the generated image search model sex.
  • step S304 that is, querying the image database, and determining the contrast dose distribution map corresponding to the first historical user identifier based on the historical dose distribution maps of other historical user identifiers, includes:
  • S501 Determine a target area location based on the historical registration distribution map corresponding to the first historical user identifier.
  • the target area refers to the tumor site.
  • the historical registration distribution map includes the target area and the organ at risk.
  • the target area is outlined in the historical registration distribution map corresponding to the first historical user ID by computer or manually, so that other histories of the same target area can be found later.
  • the historical dose distribution map corresponding to the user ID For example, if the target area is a lung tumor, the historical dose distribution map of the lung tumor is filtered from the image database, thereby reducing the number of image searches and improving the efficiency of subsequent acquisition of the contrast dose distribution map.
  • S502 Query the image database based on the target area, and obtain a contrast dose distribution map whose similarity to the historical registration distribution map is less than a first preset threshold.
  • the first preset threshold is a threshold used to determine whether the dose distribution maps corresponding to different historical user identifiers meet the similarity standard.
  • the image matching algorithm is used to calculate the similarity between the historical registration distribution map corresponding to the first historical user identifier and the historical registration distribution map corresponding to other historical user identifiers, and other historical user identifiers whose similarity is less than the first preset threshold are calculated.
  • the corresponding historical registration distribution map is determined as the contrast dose distribution map of the first historical user identification to obtain samples of the training image search model.
  • the historical registration distribution map corresponding to each user ID may be different, the historical registration distribution map corresponding to X (X is a positive integer) other historical user IDs with the smallest similarity can be obtained as the training sample , To ensure that the number of samples used to train the image search model is sufficient.
  • the image matching algorithm includes, but is not limited to, a grayscale-based matching algorithm and a feature-based matching algorithm.
  • the artificial intelligence-based image search method determines the target area based on the historical registration distribution map corresponding to the first historical user identification, thereby reducing the number of image searches and improving the efficiency of obtaining the contrast dose distribution map.
  • the image database is queried based on the target area, and the contrast dose distribution map whose similarity with the historical registration distribution map is less than the first preset threshold is obtained to obtain samples of the training image search model and provide technology for the training image search model.
  • step S304 that is, query the image database, and determine the historical dose distribution map of other historical user IDs as the contrast dose distribution map corresponding to the first historical user ID, including:
  • S601 Obtain a historical registration distribution map corresponding to the second historical user identifier from the image database.
  • the second historical user identifier refers to the identifier of any other historical user except the first historical user identifier.
  • the historical registration distribution map corresponding to the second historical user identifier is obtained from the image database, so as to subsequently obtain the contrast dose distribution map for training.
  • S602 Obtain the first DVH map corresponding to the first historical user identification based on the historical registration distribution map and the historical organ-at-risk outline map based on the first historical user identification, and the historical organ-at-risk outline map based on the first historical user identification and the second historical user Identify the corresponding historical registration distribution map to generate a corresponding second DVH map.
  • DVH is the abbreviation of Dose-Volume Histogram, which refers to the dose volume histogram.
  • the ordinate in the DVH diagram represents the volume of the diseased part, and the abscissa represents the dose of radiotherapy.
  • the dose-volume histogram specifically includes two curves, one curve reflects the dose-volume relationship of the target area in the radiotherapy plan, and the other curve reflects the dose-volume relationship of the organ at risk in the radiotherapy plan.
  • the target area and the organ at risk are outlined in advance on the historical CT map corresponding to the first historical user ID, and the area where the target area is located is converted into a vector 1 representation, and the area at which the organ at risk is located is converted into a vector 0 representation to generate The first vector matrix; in the same way, the value of the radiotherapy dose in the historical registration distribution map corresponding to the first historical user ID is converted into the corresponding second vector matrix; then the first vector matrix and the second vector matrix are multiplied, thereby Obtain the dose and volume curve of the target area corresponding to the first historical user ID.
  • the value of the radiotherapy dose in the historical registration distribution map corresponding to the second historical user identification is transformed into a fourth vector matrix, and the first vector matrix is multiplied by the fourth vector matrix to obtain the second historical user identification The relationship between the dose and the volume of the target area; relatively, the third vector matrix is multiplied by the fourth vector matrix to obtain the dose-volume relationship of the organ-at-risk identified by the second historical user to generate the second DVH map.
  • the target similarity is a value used to indicate the degree of similarity between the first DVH picture and the second DVH picture.
  • the dose and volume curve of the target area in the first DVH diagram take N points with equal intervals
  • the dose and volume curve of the target area in the second DVH diagram take N points with equal intervals
  • the distance difference between the N points of the dose and volume curve of the target area in the DVH diagram and the second DVH diagram forms the first coordinate difference
  • the interval between the dose and volume curve of the organ-at-risk in the first DVH diagram For equal N points, take N points with equal intervals on the dose and volume curve of the organ-at-risk in the second DVH chart, and calculate the N points in the dose-volume curve of the organ-at-risk in the first DVH chart and the second DVH chart
  • the distance difference forms the second coordinate difference; the average value of the first coordinate difference and the second coordinate difference is calculated as the target similarity, and then the contrast dose distribution map can be accurately determined according to the target similarity.
  • the first DVH map is obtained based on the historical registration distribution map and the historical organ-at-risk sketch map of the first historical user identification
  • the second DVH map is obtained according to the historical organ-at-risk sketch map and the second historical organ-at-risk map identified by the first historical user.
  • the historical registration distribution map corresponding to the historical user identifier is obtained. If the historical registration distribution map corresponding to the first historical user identifier is similar to the historical registration distribution map corresponding to the second historical user identifier, then the first DVH map and the second DVH map are similar. The diagram should also be similar. On the contrary, if the historical registration distribution map corresponding to the first historical user identifier and the historical registration distribution map corresponding to the second historical user identifier are not similar, the first DVH map and the second DVH map should also be dissimilar.
  • the second preset threshold is used to determine whether the first DVH picture and the second DVH picture meet a similar standard value.
  • the target similarity is less than the preset threshold, it means that the first DVH map and the second DVH map are not similar, and the historical dose distribution map corresponding to the second historical user identifier is used as the historical registration with the first historical user identifier
  • the contrast dose distribution map corresponding to the distribution map, the dissimilar historical dose distribution map is used as the contrast dose distribution map to ensure that the generated image search model can accurately identify the distance between similar images and dissimilar images, and improve the generated image search The accuracy of the model ensures the subsequent input of the dose distribution map into the feature vector generated by the image search model.
  • the historical registration distribution map corresponding to the second historical user identifier is obtained from the image database, so as to subsequently obtain the contrast dose distribution map for training, based on the first historical user
  • the marked historical registration distribution map and the historical organ-at-risk outline map Obtain the first DVH map corresponding to the first historical user identification, and the historical registration-at-risk map based on the first historical user identification and the historical registration distribution corresponding to any historical user identification
  • the map generates the corresponding second DVH map.
  • the similarity algorithm is used to calculate the similarity between the first DVH map and the second DVH map, and then the contrast dose distribution map can be accurately determined according to the target similarity. If the target similarity is less than the preset threshold, the historical dose distribution map identified by other users is used as the contrast dose distribution map corresponding to the historical registration distribution map identified by the first historical user.
  • step S205 querying the radiotherapy plan database based on the target feature vector to obtain a target dose distribution map matching the target feature vector includes:
  • S701 Query the radiotherapy plan database, and obtain a historical feature vector corresponding to any historical user identifier.
  • the server can query the radiotherapy plan database to quickly obtain the historical feature vectors corresponding to all historical user IDs.
  • S702 Calculate the target similarity value between the target feature vector and the historical feature vector.
  • the target similarity value is a value indicating the degree of similarity between the target feature vector and the historical feature vector.
  • the server can quickly calculate the target similarity value between the target feature vector and the historical feature vector through the similarity algorithm.
  • the similarity algorithm includes but is not limited to the cosine similarity algorithm, the Euclidean distance algorithm, and the Manhattan algorithm.
  • the third preset threshold is a value used to determine whether the historical feature vector and the target feature vector meet the similarity standard.
  • the image search model provided in this embodiment queries the radiotherapy plan database to obtain the historical feature vector corresponding to any historical user identifier. Calculate the target similarity value between the target feature vector and the historical feature vector. If the target similarity value is greater than the third preset threshold, the historical registration distribution map corresponding to the historical feature vector is determined as the target dose distribution map, so that the associated and stored historical radiotherapy plans can be searched based on the historical dose distribution map to provide clinicians Develop the historical radiotherapy plan of the target user as a reference.
  • the artificial intelligence-based image search method before step S205, before querying the radiotherapy plan database based on the feature vector corresponding to the target user identifier, the artificial intelligence-based image search method further includes:
  • the image data of all historical users who have undergone radiotherapy is stored in the image database, and the image data of each historical user is stored in association with the corresponding historical user ID and stored in the server.
  • the server can query by keyword matching, etc. Algorithm to quickly obtain the historical dose distribution map of all historical users. For example, the server can search for all historical dose distribution maps through the keyword "dose distribution map".
  • S802 Use an image registration algorithm to perform registration processing on the historical dose distribution map to obtain the historical registration distribution map.
  • an image registration algorithm is used to perform registration processing on the historical dose distribution map, and obtaining the historical registration distribution map is consistent with step S401. To avoid repetition, details are not described herein again.
  • inputting the historical registration distribution map into the image search model, and generating the corresponding historical feature vector is consistent with the process of generating the target feature vector in step S204. To avoid repetition, details are not described herein.
  • S804 Store the historical feature vector of each historical user identification and the corresponding historical registration distribution map in a radiotherapy plan database in association.
  • the image search request acquisition module 901 is configured to acquire an image search request, the image search request includes a target user identification, an original CT image corresponding to the target user identification, and an original outline of the organ at risk.
  • the original dose distribution map acquisition module 902 is used to input the original CT image and the original organ-at-risk map into the dose analysis model to generate an original dose distribution map corresponding to the target user identification.
  • the standard dose distribution map acquisition module 903 is used to perform registration processing on the original dose distribution map to obtain the standard dose distribution map.
  • the target feature vector obtaining module 904 is configured to input the standard dose distribution map into the image search model to obtain the target feature vector corresponding to the target user identification.
  • the target dose distribution map acquisition module 905 is configured to query the radiotherapy plan database based on the target feature vector to obtain a target dose distribution map matching the target feature vector.
  • the analysis dose distribution map acquisition module is used to input the historical CT map and the historical risk device outline map corresponding to the first historical user identification into the dose analysis model to obtain the analysis dose distribution map corresponding to the first historical user identification.
  • the image registration processing module is used to obtain the historical registration distribution map and analyze the registration distribution map based on the historical dose distribution map and the analysis dose distribution map corresponding to the first historical user identification.
  • the training sample determination module is configured to use the historical registration distribution map, the analysis registration distribution map, and the contrast dose distribution map corresponding to the first historical user identification as training samples.
  • the image search model acquisition module is used to input the training samples into the convolutional neural network based on the ternary loss function for model training to obtain the image search model.
  • the image registration processing module includes:
  • the historical registration parameter acquisition unit is configured to use an image registration algorithm to register the historical CT image corresponding to the first historical user identifier with the standard CT image to obtain the historical registration parameter.
  • the registration distribution map acquisition unit is used to perform image registration on the historical dose distribution map and the analyzed dose distribution map based on the historical registration parameters, and obtain the historical registration distribution map and analyze the registration distribution map.
  • the comparative dose distribution map determination module includes: a target area location determination unit and a first determination unit.
  • the target region location determining unit is configured to determine the target region location based on the historical registration distribution map corresponding to the first historical user identification.
  • the comparative dose distribution map determination module includes: a historical registration distribution map acquisition unit, a DVH map acquisition unit, a target similarity acquisition unit, and a second judgment unit.
  • the historical registration distribution map obtaining unit is configured to obtain the historical registration distribution map corresponding to the second historical user identifier from the image database.
  • the target similarity acquisition unit is used to calculate the similarity between the first DVH map and the second DVH map by using a similarity algorithm to acquire the target similarity.
  • the second judgment unit is configured to, if the target similarity is less than the second preset threshold, use the historical registration distribution map of the second historical user identification as the contrast dose distribution corresponding to the historical registration distribution map of the first historical user identification picture.
  • the target dose distribution map acquisition module 905 includes: a radiotherapy plan database query unit, a feature vector calculation unit, and a third judgment unit.
  • the radiotherapy plan database query unit is used to query the radiotherapy plan database and obtain the historical feature vector corresponding to any historical user identifier.
  • the feature vector calculation unit is used to calculate the target similarity value between the target feature vector and the historical feature vector.
  • the third judgment unit is configured to determine the historical registration distribution map corresponding to the historical feature vector as the target dose distribution map if the target similarity value is greater than the third preset threshold.
  • the registration processing unit is used to perform registration processing on the historical dose distribution map by using an image registration algorithm to obtain the historical registration distribution map.
  • the historical feature vector generating unit is used to input the historical registration distribution map into the image search model to generate the corresponding historical feature vector.
  • the radiotherapy plan database generating unit is used to associate the historical feature vector identified by each historical user with the corresponding historical registration distribution map and store them in the radiotherapy plan database.
  • Each module in the above artificial intelligence-based image search device can be implemented in whole or in part by software, hardware, and a combination thereof.
  • the above-mentioned modules may be embedded in the form of hardware or independent of the processor in the computer equipment, or may be stored in the memory of the computer equipment in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure diagram may be as shown in FIG. 10.
  • the computer equipment includes a processor, a memory, a network interface, and a database connected through a system bus.
  • the processor of the computer device is used to provide calculation and control capabilities.
  • the memory of the computer device includes a readable storage medium and an internal memory.
  • the readable storage medium stores an operating system, computer readable instructions, and a database.
  • the internal memory provides an environment for the operation of the operating system and computer readable instructions in the readable storage medium.
  • the database of the computer equipment is used to store historical registration distribution maps.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • the computer readable instructions are executed by the processor to realize an artificial intelligence-based image search method.
  • the readable storage medium provided in this embodiment includes a non-volatile readable storage medium and a volatile readable storage medium.
  • a computer device including a memory, a processor, and computer-readable instructions stored in the memory and capable of running on the processor.
  • the processor executes the computer-readable instructions to implement the The steps of the artificial intelligence image search method, such as steps S201-S205 shown in FIG. 2, or the steps shown in FIG. 3 to FIG. 8, are not repeated here to avoid repetition.
  • the functions of the modules/units in the embodiment of the artificial intelligence-based image search device are realized, for example, the image search request acquisition module 901 and the original dose distribution map acquisition module shown in FIG. 9 902.
  • the functions of the standard dose distribution map acquisition module 903, the target feature vector acquisition module 904, and the target dose distribution map acquisition module 905 are not repeated here to avoid repetition.
  • one or more readable storage media storing computer readable instructions are provided.
  • the readable storage medium stores computer readable instructions, and the computer readable instructions are executed by a processor to implement the foregoing implementations.
  • the steps of the artificial intelligence-based image search method in the example such as steps S201-S205 shown in FIG. 2, or the steps shown in FIG. 3 to FIG. 8, are not repeated here to avoid repetition.
  • the processor executes the computer-readable instructions, the functions of the modules/units in the embodiment of the artificial intelligence-based image search device are realized, for example, the image search request acquisition module 901 and the original dose distribution map acquisition module shown in FIG. 9 902.
  • the functions of the standard dose distribution map acquisition module 903, the target feature vector acquisition module 904, and the target dose distribution map acquisition module 905 are not repeated here to avoid repetition.
  • the readable storage medium provided in this embodiment includes a non-volatile readable storage medium and a volatile readable storage medium.
  • Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Channel (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Library & Information Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Radiation-Therapy Devices (AREA)
  • Apparatus For Radiation Diagnosis (AREA)

Abstract

一种基于人工智能的图像搜索方法、装置、设备及介质,涉及人工智能医学图像处理领域。该基于人工智能的图像搜索方法包括:获取图像搜索请求,图像搜索请求包括目标用户标识、与目标用户标识相对应的原始CT图和原始危及器官勾画图(S201);将原始CT图和原始危及器官勾画图输入剂量分析模型,生成与目标用户标识对应的原始剂量分布图(S202);对原始剂量分布图进行配准处理,获取标准剂量分布图(S203);将标准剂量分布图输入图像搜索模型,获取目标用户标识对应的目标特征向量(S204);基于目标特征向量查询放疗计划数据库,获取与目标特征向量相匹配的目标剂量分布图(S205)。该基于人工智能的图像搜索方法可提高目标剂量分布图的获取效率和准确度。

Description

基于人工智能的图像搜索方法、装置、设备及介质
本申请要求于 20200114日提交中国专利局、申请号为 202010038730.6,发明名称为“ 基于人工智能的图像搜索方法、装置、设备及介质”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
本申请涉及人工智能医学图像处理技术领域,尤其涉及一种基于人工智能的图像搜索方法、装置、设备及介质。
背景技术
肿瘤放射放疗(简称放疗)是利用放射线如放射性同位素产生的α、β、γ射线和各类x射线放疗机或加速器产生的x射线、电子线、质子束及其它粒子束等放疗恶性肿瘤的一种局部放疗方法。在放疗过程中,需基于剂量分布图设计放疗计划,以便基于该放疗计划进行放疗。在放疗计划设计中,剂量分布图通常由医师和物理师手工操作,由于目标器官与周围组织的相似性高,处方剂量水平多样化,并且目标器官附近有许多敏感的关键结构,往往需要很长的时间来确定,放疗计划的好坏依赖于医师和物理师的临床经验和医学知识,制作过程工作量大,效率低下,主观性较强,差异性较大。
随着计算机技术的发展,已有学者们提出了利用深度学习训练的模型自动生成可应用于患者进行放疗的剂量分布图,发明人意识到此时生成的剂量分布图往往忽略器官间空间位置关系,精确度不高,使得其在临床应用的参考价值不高。
发明内容
本申请实施例提供一种基于人工智能的图像搜索方法、装置、设备及介质,以解决当前放疗剂量分布图获取效率较低或者精确度不高的问题。
一种基于人工智能的图像搜索方法,包括:
获取图像搜索请求,所述图像搜索请求包括目标用户标识、与所述目标用户标识相对应的原始CT图和原始危及器官勾画图;
将所述原始CT图和所述原始危及器官勾画图输入剂量分析模型,生成与所述目标用户标识对应的原始剂量分布图;
对所述原始剂量分布图进行配准处理,获取标准剂量分布图;
将所述标准剂量分布图输入图像搜索模型,获取所述目标用户标识对应的目标特征向量;
基于所述目标特征向量查询放疗计划数据库,获取与所述目标特征向量相匹配的目标剂量分布图。
一种基于人工智能的图像搜索装置,包括:
图像搜索请求获取模块,用于获取图像搜索请求,所述图像搜索请求包括目标用户标识、与所述目标用户标识相对应的原始CT图和原始危及器官勾画图;
原始剂量分布图获取模块,用于将所述原始CT图和所述原始危及器官勾画图输入剂量分析模型,生成与所述目标用户标识对应的原始剂量分布图;
标准剂量分布图获取模块,用于对所述原始剂量分布图进行配准处理,获取标准剂量分布图;
目标特征向量获取模块,用于将所述标准剂量分布图输入图像搜索模型,获取所述目标用户标识对应的目标特征向量;
目标剂量分布图获取模块,用于基于所述目标特征向量查询放疗计划数据库,获取与所述目标特征向量相匹配的目标剂量分布图。
一种计算机设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令,所述处理器执行所述计算机可读指令时实现如下步骤:
获取图像搜索请求,所述图像搜索请求包括目标用户标识、与所述目标用户标识相对应的原始CT图和原始危及器官勾画图;
将所述原始CT图和所述原始危及器官勾画图输入剂量分析模型,生成与所述目标用户标识对应的原始剂量分布图;
对所述原始剂量分布图进行配准处理,获取标准剂量分布图;
将所述标准剂量分布图输入图像搜索模型,获取所述目标用户标识对应的目标特征向量;
基于所述目标特征向量查询放疗计划数据库,获取与所述目标特征向量相匹配的目标剂量分布图。
一个或多个存储有计算机可读指令的可读存储介质,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器执行如下步骤:
获取图像搜索请求,所述图像搜索请求包括目标用户标识、与所述目标用户标识相对应的原始CT图和原始危及器官勾画图;
将所述原始CT图和所述原始危及器官勾画图输入剂量分析模型,生成与所述目标用户标识对应的原始剂量分布图;
对所述原始剂量分布图进行配准处理,获取标准剂量分布图;
将所述标准剂量分布图输入图像搜索模型,获取所述目标用户标识对应的目标特征向量;
基于所述目标特征向量查询放疗计划数据库,获取与所述目标特征向量相匹配的目标剂量分布图。
本申请的一个或多个实施例的细节在下面的附图和描述中提出,本申请的其他特征和优点将从说明书、附图以及权利要求变得明显。
上述基于人工智能的图像搜索方法、装置、设备及介质,获取图像搜索请求中的原始CT图和原始危及器官勾画图,将所述原始CT图和所述原始危及器官勾画图输入剂量分析模型,生成原始剂量分布图,为图像搜索提供技术支持。对所述原始剂量分布图进行配准处理,获取标准剂量分布图,以消除不同个体间的差异,从而达到信息融合的目的,将所述标准剂量分布图输入图像搜索模型,获取所述目标用户标识对应的目标特征向量,基于所述目标特征向量查询放疗计划数据库,获取与所述目标特征向量相匹配的目标剂量分布图,以便服务器根据目标剂量分布图在目标放疗计划数据库中查找关联存储的历史放疗计划,并发送给客户端,以便临床医生根据历史放疗计划制定目标放疗计划,提高目标放疗计划的制定效率和精确度。
附图说明
为了更清楚地说明本申请实施例的技术方案,下面将对本申请实施例的描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动性的前提下,还可以根据这些附图获得其他的附图。
图1是本申请一实施例中基于人工智能的图像搜索方法的一应用环境示意图;
图2是本申请一实施例中基于人工智能的图像搜索方法的一流程图;
图3是本申请一实施例中基于人工智能的图像搜索方法的另一流程图;
图4是本申请一实施例中基于人工智能的图像搜索方法的另一流程图;
图5是本申请一实施例中基于人工智能的图像搜索方法的另一流程图;
图6是本申请一实施例中基于人工智能的图像搜索方法的另一流程图;
图7是本申请一实施例中基于人工智能的图像搜索方法的另一流程图;
图8是本申请一实施例中基于人工智能的图像搜索方法的另一流程图;
图9是本申请一实施例中基于人工智能的图像搜索装置的一示意图;
图10是本申请一实施例中计算机设备的一示意图。
具体实施方式
下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例是本申请一部分实施例,而不是全部的实施例。基于本申请中的实施例,本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例,都属于本申请保护的范围。
本申请实施例提供的基于人工智能的图像搜索方法,该基于人工智能的图像搜索方法可应用如图1所示的应用环境中。具体地,该基于人工智能的图像搜索方法应用在图像搜索系统中,该图像搜索系统包括如图1所示的客户端和服务器,客户端与服务器通过网络进行通信,用于实现对目标用户标识对应的原始原始CT图和原始危及器官勾画图生成进行处理,生成标准剂量分布图,通过图像搜索模型快速搜索到与标准剂量分布图相似的历史配准分布图作为目标剂量分布图,以提高目标剂量分布图的获取效率和准确度,从而为临床医生制定目标放疗计划做参考。该目标放疗计划是用于针对目标用户制定的放疗计划。其中,客户端又称为用户端,是指与服务器相对应,为客户提供本地服务的程序。客户端可安装在但不限于各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备上。服务器可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在一实施例中,如图2所示,提供一种基于人工智能的图像搜索方法,以该方法应用在图1中的服务器为例进行说明,包括如下步骤:
S201:获取图像搜索请求,图像搜索请求包括目标用户标识、与目标用户标识相对应的原始CT图和原始危及器官勾画图。
其中,目标用户是指进行检测确定肿瘤分布情况,以便临床医生制定目标放疗计划的用户。目标放疗计划是针对目标用户制定的放疗计划,一个用户对应一个放疗计划。目标用户标识是用于唯一识别目标用户的标识,例如,目标用户标识可以是目标用户的姓名和目标用户的身份证号等。
原始CT图是目标用户通过CT扫描获得的图像。CT(Computed Tomography,即电子计算机断层扫描),是用X射线束对人体被检查部位所在的层面进行扫描,由探测器接收透过该层面的X射线,转变为可见光后,由光电转换变为电信号,再经模拟/数字转换器(analog/digital converter)转为数字信号,计算机对这些数字信号进行处理后,得到的人体被检查部位的断面或立体的图像即为原始CT图,以便利用该原始CT图发现被检查部位的细小病变。可以理解地,被检查部位包括患病部位和非患病部位,例如,被检查部位是肺部,患病部位是右肺,非患病部位是左肺和左肺周围的器官等。
危及器官是指在放疗射线的放射范围内的非患病的重要组织或器官。原始危及器官勾画图是指在原始CT图中对在放疗射线的放射范围内的非患病的重要组织或器官进行勾画获得的图,原始危及器官勾画图与目标用户对应。
具体地,目标用户到医院进行CT扫描,以获得该目标用户的原始CT图,并根据原始CT图勾勒出危及器官,以获得原始危及器官勾画图,将目标用户的原始CT图和原始危及器官勾画图等图像数据与目标用户标识进行关联存储在图像数据库中。可以理解地,每一用户的图像数据与用户标识一一对应,以便管理,其中,图像数据包括但不限于CT图和危及器官勾画图。图像数据库是用于存储所有用户的图像数据的库。
作为一示例,临床医生通过点击客户端上的图像搜索按钮,生成带有目标用户标识、目标用户标识相对应的原始CT图和原始危及勾画图的图像搜索请求,并将该图像搜索请求发送给服务器,以使服务器获取到图像搜索请求。
S202:将原始CT图和原始危及器官勾画图输入剂量分析模型,生成与目标用户标识对应的原始剂量分布图。
其中,剂量分析模型是用于生成预测的剂量分布图的模型。该剂量分析模型是基于深度神经网络训练生成的模型,该剂量分析模型是采用深度神经网络对训练样本进行训练所形成的模型,该训练样本包括同一用户标识对应的训练CT图、训练危及器官勾画图和对应的训练剂量分布图。
原始剂量分布图是剂量分析模型预测的射线剂量,即对目标用户进行放疗时,预测目标用户患病部位所需的射线剂量。由于原始剂量分布图是通过剂量分析模型生成的,原始剂量分布图中可能忽略不同器官之间存在的空间位置,精确度不高,无法满足临床标准,因此临床医生无法直接由原始剂量分布图生成目标放疗计划。
具体地,将原始CT图和原始危及器官勾画图输入剂量分析模型,从而快速生成与目标用户标识对应的原始剂量分布图,为图像搜索提供技术支持。
S203:对原始剂量分布图进行配准处理,获取标准剂量分布图。
其中,配准处理是用于比较或融合不同用户在不同条件下获取的图像,以便后续可精准搜索图像。可以理解地,由于不同用户体型存在差异导致不同用户的器官大小或者空间位置情况不同,通过寻找一种空间变换把不同用户的剂量分布图或者CT图等映射到另一幅图像,将不同用户的相同图像对应于空间同一位置的点一一对应起来,以消除不同个体间的差异。
具体地,采用图像配准算法对原始CT图与标准CT图进行配准,获得标准配准参数,基于标准配准参数对原始剂量分布图进行转换,获取标准剂量分布图,确保目标用户标识对应的原始CT图的器官与标准CT图的器官处于相对应的位置,可以排除因为不同个体的器官的大小、空间位置等因素对图像搜索的影响,确保后续可搜索到相似的图像,提高图像搜索的准确度。标准CT图是指通用的CT图模板。
S204:将标准剂量分布图输入图像搜索模型,获取目标用户标识对应的目标特征向量。
其中,图像搜索模型是指预先训练好的用于识别剂量分布图,以输出特征向量的模型。该图像搜索模型具体是应用基于三元损失函数的卷积神经网络训练生成的模型,可确保相似的图像像通过图像搜索模型生成的特征向量的距离小,具体是使相似的配准分布图通过图像搜索模型生成对应的特征向量距离小,不相似的图像像通过图像搜索模型生成的特征向量的距离大,具体是使不相似的配准分布图通过图像搜索模型生成对应的特征向量距离大,即该图像搜索模型确保相似的图像像通过图像搜索模型生成的特征向量的距离小于不相似的图像像通过图像搜索模型生成的特征向量的距离,以便后续可准确获取与标准剂量分布图相似的历史配准分布图,从而确定目标剂量分布图。具体地,将标准剂量分布图输入图像搜索模型,从而获取目标用户标识对应的目标特征向量,为后续寻找与标准剂量分布图提供基础。
S205:基于目标特征向量查询放疗计划数据库,获取与目标特征向量相匹配的目标剂量分布图。
其中,放疗计划数据库是指用于存储经过放疗后的历史用户对应的用户数据的数据库,该用户数据包括关联存储的历史用户标识、历史CT图、历史剂量分布图、历史放疗计划和历史剂量分布图对应的历史特征向量。
其中,历史用户是指已经进行放疗的用户。历史用户标识是用于唯一识别历史用户的标识。历史CT图是历史用户通过CT扫描获取的图像。历史剂量分布图是历史用户在放疗过程中形成的剂量分布图。历史放疗计划是历史用户在放疗过程中采集的放疗计划。历史 特征向量是将历史剂量分布图输入图像搜索模型获得的特征向量。目标剂量分布图是与标准剂量分布图相似的历史配准分布图。
历史特征向量通过图像搜索模型获得的,可以确保相似的历史剂量分布图的特征向量的距离小,不相似的历史剂量分布图的特征向量的距离大,以便后续可准确获取与标准剂量分布图相似的历史配准分布图,从而基于相似的历史剂量分布图确定目标剂量分布图。
具体地,采用相似度计算公式计算目标特征向量和放疗计划数据库中的任一历史用户标识对应的历史特征向量的相似度,并获取相似度最大的前M(M为正整数)个的历史特征向量,将这些历史特征向量对应的历史剂量分布图作为目标剂量分布图,以便服务器根据目标剂量分布图在目标放疗计划数据库中查找关联存储的历史放疗计划,并发送给客户端,以便临床医生根据历史放疗计划制定目标放疗计划,提高目标放疗计划的制定效率和精确度。可以理解地,由于历史放疗计划是历史用户已经进行放疗的计划,则历史放疗计划具有较强的参考价值,以缩短临床医生制定目标放疗计划所需的时间,缩短目标用户的放疗周期,为目标用户提供及时放疗。
本实施例所提供的基于人工智能的图像搜索方法,获取图像搜索请求中的原始CT图和原始危及器官勾画图,将原始CT图和原始危及器官勾画图输入剂量分析模型,生成原始剂量分布图,为图像搜索提供技术支持。对原始剂量分布图进行配准处理,获取标准剂量分布图,以消除不同个体间的差异,从而达到信息融合的目的,将标准剂量分布图输入图像搜索模型,获取目标用户标识对应的目标特征向量,基于目标特征向量查询放疗计划数据库,获取与目标特征向量相匹配的目标剂量分布图,以便服务器根据目标剂量分布图在目标放疗计划数据库中查找关联存储的历史放疗计划,并发送给客户端,以便临床医生根据历史放疗计划制定目标放疗计划,提高目标放疗计划的制定效率和精确度。
在一实施例中,如3所示,在步骤S204之前,即在将标准剂量分布图输入图像搜索模型,获取目标用户标识对应的目标特征向量之前,基于人工智能的图像搜索方法还包括:
S301:获取第一历史用户标识的历史用户图像数据,历史用户图像数据包括历史CT图、历史危及器官勾画图和历史剂量分布图。
其中,第一历史用户标识是图像数据库中的一个历史用户的标识。历史用户图像数据是与第一历史用户标识关联存储在图像数据库中的图像数据。该历史用户图像数据包括但不限于历史CT图、历史危及器官勾画图和历史剂量分布图。其中,历史危及器官勾画图是指在同一历史用户的历史CT图中对在放疗射线的放射范围内的非患病的重要组织或器官进行勾画获得的图。
S302:将第一历史用户标识对应的历史CT图和历史危及器官勾画图输入到剂量分析模型,获取第一历史用户标识对应的分析剂量分布图。
其中,分析剂量分布图是第一历史用户标识对应的历史CT图和历史危及器官勾画图通过剂量分布模型进行预测,所获得的预测的剂量分布图,以将分析剂量分布图作为用于训练图像搜索模型的训练数据。
S303:基于第一历史用户标识对应的历史剂量分布图和分析剂量分布图,获取历史配准分布图和分析配准分布图。
其中,历史配准分布图是指对历史剂量分布图进行图像配准处理后获得的图。分析配准分布图是指对分析剂量分布图进行图像配准处理后获得的图像。
具体地,可以采用图像配准算法对第一历史用户标识的历史CT图与标准CT图进行配准,以获得第一历史用户标识对应的配准参数,基于该配准参数对历史剂量分布图进行图像配准处理,获取历史配准分布图,基于该配准参数对对分析剂量分布图进行图像配准处理,获取分析配准分布图。经过配准处理获得历史配准分布图和分析配准分布图,可以确保训练的模型更加精确。
S304:查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定第一历史用 户标识对应的对比剂量分布图。
具体地,从图像数据库中获取除第一历史用户标识以外的其他历史用户标识对应的历史剂量分布图,对其他历史用户标识对应的历史剂量分布图进行配准处理生成对比剂量分布图,以消除不同用户的图像间差异,确保训练图像搜索模型的准确性。可以理解地,为了获取更多的训练样本,可以将除第一历史用户标识以外的多个其他历史用户标识对应的历史剂量分布图进行配准处理,获取对比剂量分布图,以获取数量充足的训练样本。
S305:将第一历史用户标识对应的历史配准分布图、分析配准分布图和对比剂量分布图作为训练样本。
具体地,将第一历史用户标识对应的历史配准分布图、分析配准分布图和对比剂量分布图作为训练样本,可以理解地,由于历史配准分布图和分析配准分布图是同一历史用户的图像数据,则相似性较高,而对比剂量分布图是对除第一历史用户标识以外的历史剂量分布图进行配准处理后获取的剂量分布图,是与第一历史用户标识对应的历史配准分布图和分析配准分布图不相似的剂量分布图,以确保生成的图像搜索模型可使不相似的图像对应的特征向量的距离大,具体是使不相似的配准分布图对应的特征向量距离大,从而可保障保后续图像搜索的准确性。
S306:将训练样本输入基于三元损失函数的卷积神经网络进行模型训练,获取图像搜索模型。
具体地,将第一历史用户标识对应的历史配准分布图、分析配准分布图和对比剂量分布图输入基于三元损失函数的权重共享的卷积神经网络进行训练,当损失为小于函数收敛值时,则表示该图像搜索模型训练完成。该函数收敛值是预先设置的用于评估损失函数是否达到收敛要求的值,可以为零。其中,三元损失函数为
Figure PCTCN2020093333-appb-000001
M(M为正整数)表示训练样本的数量,i(i为正整数,i≤M)表示第i组训练样本,x a表示历史配准分布图对应的向量,x p表示分析配准分布图对应的向量,x n表示对比剂量分布图对应的向量,
Figure PCTCN2020093333-appb-000002
表示历史配准分布图和分析配准分布图之间的欧式距离度量,
Figure PCTCN2020093333-appb-000003
表示历史配准分布图和对比剂量分布图之间的欧式距离度量,α是指x a与x n之间的距离和x a与x p之间的距离之间最小的间隔,+表示[]内的值大于函数收敛值的时候,取[]内的值为损失;[]内的值小于函数收敛值的时候,则该图像搜索模型训练完成。
本实施例所提供的基于人工智能的图像搜索方法,将第一历史用户标识对应的历史CT图和历史危及器官勾画图输入到剂量分析模型,以快速获取第一历史用户标识对应的分析剂量分布图。基于第一历史用户标识对应的历史剂量分布图和分析剂量分布图,获取历史配准分布图和分析配准分布图,以消除不同用户的图像间差异对模型训练的影响。查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定第一历史用户标识对应的对比剂量分布图,将第一历史用户标识对应的历史配准分布图、分析配准分布图和对比剂量分布图作为训练样本,从而可保障保后续图像搜索的准确性。将训练样本输入基于三元损失函数的卷积神经网络进行模型训练,快速获取图像搜索模型,以确保生成的图像搜索模型可使不相似的图像对应的特征向量的距离大,具体是使不相似的配准分布图对应的特征向量距离大。
在一实施例中,如图4所示,步骤S303,即基于第一历史用户标识对应的历史剂量分 布图和分析剂量分布图,获取历史配准分布图和分析配准分布图,包括:
S401:采用图像配准算法将第一历史用户标识对应的历史CT图与标准CT图进行配准,获取历史配准参数。
具体地,对历史CT图进行预处理,为配准提供基础,该预处理过程包括:对历史CT图进行消除噪声处理,以消除干扰因素;当历史CT图与标准CT图的像素大小不同时,对历史CT图的尺寸进行调整,以使历史CT图与标准CT图的像素大小相匹配,以使历史CT图和标准CT图的特征对应,从而保证获取的历史配准参数的准确性。
从预处理后的历史CT图的特定位置上选取第一特征,并在标准CT图与历史CT图对应的位置选取第二特征,以获取基于同一特定位置对应的第一特征和第二特征,为了更好地确定历史CT图和标准CT图的对应关系,可以在历史CT图和标准CT图取多个对应的位置的特征;建立三维坐标系,确定历史CT图的第一特征的三维坐标和标准CT图的第二特征的三维坐标,依据两个三维坐标确定预处理后的历史CT图和标准CT图的配准函数,基于该配准函数获取历史配准参数。具体地,历史配准参数是配准函数的参数,对历史CT图重新采样验证历史配准参数准确性。
S402:基于历史配准参数对历史剂量分布图和分析剂量分布图进行图像配准,获取历史配准分布图和分析配准分布图。
具体地,根据历史配准参数对第一历史用户标识对应的历史剂量分布图和分析剂量分布图进行配准,即根据历史配准参数对第一历史用户标识对应的历史剂量分布图和分析剂量分布图进行空间变换。其中,空间变换可以是旋转、缩小和放大等转换,以获取历史配准分布图和分析配准分布图,由于训练样本均进行配准处理,可以消除不同用户的图像间的差异,确保生成图像搜索模型的准确性。
本实施例所提供的基于人工智能的图像搜索方法,采用图像配准算法将第一历史用户标识对应的历史CT图与标准CT图进行配准,获取历史配准参数。基于历史配准参数对历史剂量分布图和分析剂量分布图进行图像配准,获取历史配准分布图和分析配准分布图,可以消除不同用户的图像间的差异,确保生成图像搜索模型的准确性。
在一实施例中,如图5所示,步骤S304,即查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定第一历史用户标识对应的对比剂量分布图,包括:
S501:基于第一历史用户标识对应的历史配准分布图确定靶区部位。
其中,靶区部位是指肿瘤部位。历史配准分布图中包括靶区部位和危及器官部位,通过计算机或者人工在第一历史用户标识对应的历史配准分布图中勾勒出靶区部位,以便后续查找相同的靶区部位的其他历史用户标识对应的历史剂量分布图。例如,靶区部位为肺部肿瘤,则从图像数据库中筛选出肺部肿瘤的历史剂量分布图,从而减少图像查找数量,提高后续获取对比剂量分布图的效率。
S502:基于靶区部位查询图像数据库,获取与历史配准分布图相似度小于第一预设阈值的对比剂量分布图。
其中,第一预设阈值是用于判断不同历史用户标识对应的剂量分布图是否达到相似标准的阈值。
具体地,通过图像匹配算法计算第一历史用户标识对应的历史配准分布图和其他历史用户标识对应的历史配准分布图的相似度,将相似度小于第一预设阈值的其他历史用户标识对应的历史配准分布图确定为第一历史用户标识的对比剂量分布图,以获取训练图像搜索模型的样本。可以理解地,由于每一用户标识对应的历史配准分布图可能不一样,因此,可以获取相似度最小的X(X为正整数)个其他历史用户标识对应的历史配准分布图作为训练样本,确保用于训练图像搜索模型的样本数量充足。本实施例中,图像匹配算法包括但不限于基于灰度的匹配算法和基于特征的匹配算法。
本实施例所提供的基于人工智能的图像搜索方法,基于第一历史用户标识对应的历史 配准分布图确定靶区部位,从而减少图像查找数量,提高获取对比剂量分布图的效率。基于靶区部位查询图像数据库,获取与历史配准分布图相似度小于第一预设阈值的对比剂量分布图,以获取训练图像搜索模型的样本,为训练图像搜索模型提供技术。
在一实施例中,如图6所示,步骤S304,即查询图像数据库,将其他历史用户标识的历史剂量分布图,确定为第一历史用户标识对应的对比剂量分布图,包括:
S601:从图像数据库获取第二历史用户标识对应的历史配准分布图。
其中,第二历史用户标识是指除第一历史用户标识外的其他任意一个历史用户的标识。具体地,从图像数据库中获取第二历史用户标识对应的历史配准分布图,以便后续获取到用于进行训练的对比剂量分布图。
S602:基于第一历史用户标识的历史配准分布图和历史危及器官勾画图获取第一历史用户标识对应的第一DVH图,基于第一历史用户标识的历史危及器官勾画图和第二历史用户标识对应的历史配准分布图生成对应的第二DVH图。
其中,DVH是Dose-Volume Histogram的缩写,是指剂量体积直方图。DVH图中纵坐标代表患病部位的体积,横坐标代表放疗的剂量。剂量体积直方图具体包括了两条曲线,其中一条曲线体现放疗计划中靶区部位的剂量-体积的关系,另一条曲线体现了放疗计划中危及器官的剂量-体积的关系。
具体地,预先在第一历史用户标识对应的历史CT图上勾画出靶区部位和危及器官,并将靶区部位所在区域转化为向量1表示,将危及器官所在区域转化为向量0表示,生成第一向量矩阵;同理,将第一历史用户标识对应的历史配准分布图中放疗剂量的数值转化为相应的第二向量矩阵;再将第一向量矩阵与第二向量矩阵相乘,从而得到第一历史用户标识对应的靶区部位的剂量与体积曲线。可以理解地,由于靶区部位所在区域的向量为1,则保留了靶区部位的剂量与体积的关系,将第一向量矩阵与第二向量矩阵相乘,可得到第一历史用户标识对应的靶区部位的剂量与体积曲线。相对地,将靶区部位所在区域转化为向量0表示,将危及器官所在区域转化为向量1表示,生成第三向量矩阵,第二向量矩阵与第三向量矩阵相乘,以获取危及器官中剂量-体积的关系,即第一历史用户标识对应的危及器官的剂量与体积曲线,从而生成第一DVH图。
同理地,根据第二历史用户标识对应的历史配准分布图中放疗剂量的数值转化为第四向量矩阵,第一向量矩阵与第四向量矩阵相乘,以得到第二历史用户标识对应的靶区部位的剂量与体积的关系;相对地,第三向量矩阵与第四向量矩阵相乘,以获取第二历史用户标识的危及器官的剂量-体积的关系,以生成第二DVH图。
S603:采用相似度算法对第一DVH图和第二DVH图进行相似度计算,获取目标相似度。
其中,目标相似度是用于表示第一DVH图和第二DVH图之间相似程度的值。
具体地,在第一DVH图的靶区部位的剂量与体积曲线取间隔相等的N个点,在第二DVH图中靶区部位的剂量与体积曲线取间隔相等的N个点,计算第一DVH图和第二DVH图中靶区部位的剂量与体积曲线N个点的距离差值,形成第一坐标差值;同理地,在第一DVH图的危及器官的剂量与体积曲线取间隔相等的N个点,在第二DVH图的危及器官的剂量与体积曲线取间隔相等的N个点,计算第一DVH图和第二DVH图的危及器官的剂量与体积曲线中N个点的距离差值,形成第二坐标差值;求第一坐标差值和第二坐标差值的平均值,作为目标相似度,后续可以根据目标相似度,精准地确定对比剂量分布图。可以理解地,第一DVH图是根据第一历史用户标识的历史配准分布图和历史危及器官勾画图获取的,第二DVH图是根据第一历史用户标识的历史危及器官勾画图和第二历史用户标识对应的历史配准分布图获取的,若第一历史用户标识对应的历史配准分布图和第二历史用户标识对应的历史配准分布图相似,则第一DVH图和第二DVH图也应该是相似的。相反,若第一历史用户标识对应的历史配准分布图和第二历史用户标识对应的历史配准分布图不相似,则第一DVH图和第二DVH图也应该是不相似的。
S604:若目标相似度小于第二预设阈值,则将第二历史用户标识的历史配准分布图作为与第一历史用户标识的历史配准分布图相对应的对比剂量分布图。
其中,第二预设阈值用于判断第一DVH图和第二DVH图是否达到相似标准的值。
具体地,目标相似度小于预设阈值时,说明第一DVH图和第二DVH图并不相似,则将第二历史用户标识对应的历史剂量分布图作为与第一历史用户标识的历史配准分布图相对应的对比剂量分布图,将不相似的历史剂量分布图作为对比剂量分布图,以确保生成的图像搜索模型可以准确识别相似的图像与不相似的图像的距离,提高生成的图像搜索模型准确度,确保后续将剂量分布图输入图像搜索模型生成的特征向量。
本实施例所提供的基于人工智能的图像搜索方法中,从图像数据库获取第二历史用户标识对应的历史配准分布图,以便后续获取用于进行训练的对比剂量分布图,基于第一历史用户标识的历史配准分布图和历史危及器官勾画图获取第一历史用户标识对应的第一DVH图,基于第一历史用户标识的历史危及器官勾画图和任一历史用户标识对应的历史配准分布图生成对应的第二DVH图。采用相似度算法对第一DVH图和第二DVH图进行相似度计算,后续可以根据目标相似度精准地确定对比剂量分布图。若目标相似度小于预设阈值,则将其他用户标识的历史剂量分布图作为与第一历史用户标识的历史配准分布图相对应的对比剂量分布图。
在一实施例中,如图7所示,步骤S205,基于目标特征向量查询放疗计划数据库,获取与目标特征向量相匹配的目标剂量分布图,包括:
S701:查询放疗计划数据库,获取任一历史用户标识对应的历史特征向量。
具体地,在训练好图像搜索模型后,则将所有历史用户标识对应的历史剂量分布图进行配准,将配准后的历史配准分布图输入图像搜索模型,以生成每一历史用户标识对应的历史特征向量,并存储在放疗计划数据库中,因此,服务器查询放疗计划数据库,则可以快速获取到所有历史用户标识对应的历史特征向量。
S702:计算目标特征向量与历史特征向量的目标相似值。
其中,目标相似值是表示目标特征向量与历史特征向量的相似程度的值。
具体地,服务器可以通过相似度算法以快速地计算目标特征向量与历史特征向量的目标相似值。本实施例中,相似度算法包括但不限于余弦相似度算法、欧式距离算法和曼哈顿算法等。
S703:若目标相似值大于第三预设阈值,则将历史特征向量对应的历史配准分布图确定为目标剂量分布图。
其中,第三预设阈值是用于判断历史特征向量与目标特征向量是否达到相似标准的值。
具体地,计算目标特征向量与历史特征向量的目标相似值,将目标相似值按照由大到小进行排序,根据排序结果,选取目标相似值大于第三预设阈值的前M(M为正整数)个历史特征向量对应的历史配准分布图,将选取出来的历史剂量分布图作为目标剂量分布图,以便后续根据该历史剂量分布图查找关联存储的历史放疗计划,为临床医生提供制定目标用户的历史放疗计划作为参考。
本实施例所提供的图像搜索模型,查询放疗计划数据库,获取任一历史用户标识对应的历史特征向量。计算目标特征向量与历史特征向量的目标相似值。若目标相似值大于第三预设阈值,则将历史特征向量对应的历史配准分布图确定为目标剂量分布图,以便后续根据该历史剂量分布图查找关联存储的历史放疗计划,为临床医生提供制定目标用户的历史放疗计划作为参考。
在一实施例中,如图8所示,在步骤S205之前,在基于目标用户标识对应的特征向量查询放疗计划数据库之前,基于人工智能的图像搜索方法还包括:
S801:从图像数据库中获取N个历史剂量分布图。
具体地,将所有经过放疗的历史用户的图像数据存储在图像数据库中,且每一历史用户的图像数据与对应的历史用户标识关联存储,并存储在服务器中,服务器可以通过关键字匹配等查询算法,快速获取所有历史用户的历史剂量分布图,例如,服务器可以通过关键字“剂量分布图”查找所有的历史剂量分布图。
S802:采用图像配准算法对历史剂量分布图进行配准处理,获取历史配准分布图。
具体地,通过采用图像配准算法对历史剂量分布图进行配准处理,获取历史配准分布图与步骤S401一致,为避免重复,在此不再赘述。
S803:将历史配准分布图输入图像搜索模型,生成对应的历史特征向量。
具体地,将历史配准分布图输入图像搜索模型,生成对应的历史特征向量与步骤S204生成目标特征向量的过程一致,为避免重复,在此不再赘述。
S804:将每一历史用户标识的历史特征向量和对应的历史配准分布图关联存储在放疗计划数据库。
具体地,将相同历史用户标识的历史特征向量、历史配准分布图、历史剂量分布图和历史放疗计划关联存储在放疗计划数据库中,以便后续查找与目标特征向量相似的历史特征向量,从而获得与目标用户相似患病部位的历史用户的历史放疗计划,为临床医生提供参考。
本实施例所提供的图像搜索模型,从图像数据库中获取N个历史剂量分布图,采用图像配准算法对历史剂量分布图进行配准处理,获取历史配准分布图,将历史配准分布图输入图像搜索模型,生成对应的历史特征向量,以便后续计算历史特征向量与目标特征向量的目标相似值。将每一历史用户标识的历史特征向量和对应的历史配准分布图关联存储在放疗计划数据库,以便后续查找与目标特征向量相似的历史特征向量,从而获得与目标用户相似患病部位的历史用户的历史放疗计划,为临床医生提供参考。
应理解,上述实施例中各步骤的序号的大小并不意味着执行顺序的先后,各过程的执行顺序应以其功能和内在逻辑确定,而不应对本申请实施例的实施过程构成任何限定。
在一实施例中,提供一种基于人工智能的图像搜索装置,该基于人工智能的图像搜索装置与上述实施例中基于人工智能的图像搜索方法一一对应。如图9所示,该基于人工智能的图像搜索装置包括图像搜索请求获取模块901、原始剂量分布图获取模块902、标准剂量分布图获取模块903、目标特征向量获取模块904和目标剂量分布图获取模块905。各功能模块详细说明如下:
图像搜索请求获取模块901,用于获取图像搜索请求,图像搜索请求包括目标用户标识、与目标用户标识相对应的原始CT图和原始危及器官勾画图。
原始剂量分布图获取模块902,用于将原始CT图和原始危及器官勾画图输入剂量分析模型,生成与目标用户标识对应的原始剂量分布图。
标准剂量分布图获取模块903,用于对原始剂量分布图进行配准处理,获取标准剂量分布图。
目标特征向量获取模块904,用于将标准剂量分布图输入图像搜索模型,获取目标用户标识对应的目标特征向量。
目标剂量分布图获取模块905,用于基于目标特征向量查询放疗计划数据库,获取与目标特征向量相匹配的目标剂量分布图。
进一步地,在目标特征向量获取模块904之前,基于人工智能的图像搜索装置还包括:历史用户图像数据获取模块、分析剂量分布图获取模块、图像配准处理模块、对比剂量分布图确定模块、训练样本确定模块和图像搜索模型获取模块。
历史用户图像数据获取模块,用于获取第一历史用户标识的历史用户图像数据,历史用户图像数据包括历史CT图、历史危及器官勾画图和历史剂量分布图。
分析剂量分布图获取模块,用于将第一历史用户标识对应的历史CT图和历史危及器 官勾画图输入到剂量分析模型,获取第一历史用户标识对应的分析剂量分布图。
图像配准处理模块,用于基于第一历史用户标识对应的历史剂量分布图和分析剂量分布图,获取历史配准分布图和分析配准分布图。
对比剂量分布图确定模块,用于查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定第一历史用户标识对应的对比剂量分布图。
训练样本确定模块,用于将第一历史用户标识对应的历史配准分布图、分析配准分布图和对比剂量分布图作为训练样本。
图像搜索模型获取模块,用于将训练样本输入基于三元损失函数的卷积神经网络进行模型训练,获取图像搜索模型。
进一步地,图像配准处理模块,包括:
历史配准参数获取单元,用于采用图像配准算法将第一历史用户标识对应的历史CT图与标准CT图进行配准,获取历史配准参数。
配准分布图获取单元,用于基于历史配准参数对历史剂量分布图和分析剂量分布图进行图像配准,获取历史配准分布图和分析配准分布图。
进一步地,对比剂量分布图确定模块,包括:靶区部位确定单元和第一判断单元。
靶区部位确定单元,用于基于第一历史用户标识对应的历史配准分布图确定靶区部位。
第一判断单元,用于基于靶区部位查询图像数据库,获取与历史配准分布图相似度小于第一预设阈值的对比剂量分布图。
进一步地,对比剂量分布图确定模块,包括:历史配准分布图获取单元、DVH图获取单元、目标相似度获取单元和第二判断单元。
历史配准分布图获取单元,用于从图像数据库获取第二历史用户标识对应的历史配准分布图。
DVH图获取单元,用于基于第一历史用户标识的历史配准分布图和历史危及器官勾画图获取第一历史用户标识对应的第一DVH图,基于第一历史用户标识的历史危及器官勾画图和第二历史用户标识对应的历史配准分布图生成对应的第二DVH图。
目标相似度获取单元,用于采用相似度算法对第一DVH图和第二DVH图进行相似度计算,获取目标相似度。
第二判断单元,用于若目标相似度小于第二预设阈值,则将第二历史用户标识的历史配准分布图作为与第一历史用户标识的历史配准分布图相对应的对比剂量分布图。
进一步地,目标剂量分布图获取模块905,包括:放疗计划数据库查询单元、特征向量计算单元和第三判断单元。
放疗计划数据库查询单元,用于查询放疗计划数据库,获取任一历史用户标识对应的历史特征向量。
特征向量计算单元,用于计算目标特征向量与历史特征向量的目标相似值。
第三判断单元,用于若目标相似值大于第三预设阈值,则将历史特征向量对应的历史配准分布图确定为目标剂量分布图。
进一步地,在目标剂量分布图获取模块905之前,基于人工智能的图像搜索装置还包括:历史剂量分布图获取单元、配准处理单元、历史特征向量生成单元和放疗计划数据库生成单元。
历史剂量分布图获取单元,用于从图像数据库中获取N个历史剂量分布图。
配准处理单元,用于采用图像配准算法对历史剂量分布图进行配准处理,获取历史配准分布图。
历史特征向量生成单元,用于将历史配准分布图输入图像搜索模型,生成对应的历史特征向量。
放疗计划数据库生成单元,用于将每一历史用户标识的历史特征向量和对应的历史配准分布图关联存储在放疗计划数据库。
关于基于人工智能的图像搜索装置的具体限定可以参见上文中对于基于人工智能的图像搜索方法的限定,在此不再赘述。上述基于人工智能的图像搜索装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中,也可以以软件形式存储于计算机设备中的存储器中,以便于处理器调用执行以上各个模块对应的操作。
在一个实施例中,提供了一种计算机设备,该计算机设备可以是服务器,其内部结构图可以如图10所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中,该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括可读存储介质、内存储器。该可读存储介质存储有操作系统、计算机可读指令和数据库。该内存储器为可读存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储历史配准分布图。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种基于人工智能的图像搜索方法。本实施例所提供的可读存储介质包括非易失性可读存储介质和易失性可读存储介质。
在一个实施例中,提供了一种计算机设备,包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机可读指令,处理器执行计算机可读指令时实现上述实施例中基于人工智能的图像搜索方法的步骤,例如图2所示的步骤S201-S205,或者图3至图8中所示的步骤,为避免重复,这里不再赘述。或者,处理器执行计算机可读指令时实现基于人工智能的图像搜索装置这一实施例中的各模块/单元的功能,例如图9所示的图像搜索请求获取模块901、原始剂量分布图获取模块902、标准剂量分布图获取模块903、目标特征向量获取模块904和目标剂量分布图获取模块905的功能,为避免重复,这里不再赘述。
在一实施例中,提供了一个或多个存储有计算机可读指令的可读存储介质,该可读存储介质上存储有计算机可读指令,该计算机可读指令被处理器执行时实现上述实施例中基于人工智能的图像搜索方法的步骤,例如图2所示的步骤S201-S205,或者图3至图8中所示的步骤,为避免重复,这里不再赘述。或者,处理器执行计算机可读指令时实现基于人工智能的图像搜索装置这一实施例中的各模块/单元的功能,例如图9所示的图像搜索请求获取模块901、原始剂量分布图获取模块902、标准剂量分布图获取模块903、目标特征向量获取模块904和目标剂量分布图获取模块905的功能,为避免重复,这里不再赘述。本实施例所提供的可读存储介质包括非易失性可读存储介质和易失性可读存储介质。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,是可以通过计算机可读指令来指令相关的硬件来完成,所述的计算机可读指令可存储于一非易失性计算机可读取存储介质中,该计算机可读指令在执行时,可包括如上述各方法的实施例的流程。其中,本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用,均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限,RAM以多种形式可得,诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
所属领域的技术人员可以清楚地了解到,为了描述的方便和简洁,仅以上述各功能单元、模块的划分进行举例说明,实际应用中,可以根据需要而将上述功能分配由不同的功能单元、模块完成,即将所述装置的内部结构划分成不同的功能单元或模块,以完成以上 描述的全部或者部分功能。
以上所述实施例仅用以说明本申请的技术方案,而非对其限制;尽管参照前述实施例对本申请进行了详细的说明,本领域的普通技术人员应当理解:其依然可以对前述各实施例所记载的技术方案进行修改,或者对其中部分技术特征进行等同替换;而这些修改或者替换,并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围,均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种基于人工智能的图像搜索方法,其中,包括:
    获取图像搜索请求,所述图像搜索请求包括目标用户标识、与所述目标用户标识相对应的原始CT图和原始危及器官勾画图;
    将所述原始CT图和所述原始危及器官勾画图输入剂量分析模型,生成与所述目标用户标识对应的原始剂量分布图;
    对所述原始剂量分布图进行配准处理,获取标准剂量分布图;
    将所述标准剂量分布图输入图像搜索模型,获取所述目标用户标识对应的目标特征向量;
    基于所述目标特征向量查询放疗计划数据库,获取与所述目标特征向量相匹配的目标剂量分布图。
  2. 如权利要求1所述的基于人工智能的图像搜索方法,其中,在所述将所述标准剂量分布图输入图像搜索模型,获取所述目标用户标识对应的目标特征向量之前,所述基于人工智能的图像搜索方法还包括:
    获取第一历史用户标识的历史用户图像数据,所述历史用户图像数据包括历史CT图、历史危及器官勾画图和历史剂量分布图;
    将第一历史用户标识对应的历史CT图和历史危及器官勾画图输入到剂量分析模型,获取所述第一历史用户标识对应的分析剂量分布图;
    基于第一历史用户标识对应的历史剂量分布图和分析剂量分布图,获取历史配准分布图和分析配准分布图;
    查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定所述第一历史用户标识对应的对比剂量分布图;
    将所述第一历史用户标识对应的历史配准分布图、分析配准分布图和对比剂量分布图作为训练样本;
    将所述训练样本输入基于三元损失函数的卷积神经网络进行模型训练,获取图像搜索模型。
  3. 如权利要求2所述的基于人工智能的图像搜索方法,其中,所述基于第一历史用户标识对应的历史剂量分布图和分析剂量分布图,获取历史配准分布图和分析配准分布图,包括:
    采用图像配准算法将所述第一历史用户标识对应的历史CT图与标准CT图进行配准,获取历史配准参数;
    基于所述历史配准参数对历史剂量分布图和分析剂量分布图进行图像配准,获取历史配准分布图和分析配准分布图。
  4. 如权利要求2所述的基于人工智能的图像搜索方法,其中,所述查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定所述第一历史用户标识对应的对比剂量分布图,包括:
    基于所述第一历史用户标识对应的历史配准分布图确定靶区部位;
    基于所述靶区部位查询图像数据库,获取与所述历史配准分布图相似度小于第一预设阈值的对比剂量分布图。
  5. 如权利要求2所述的基于人工智能的图像搜索方法,其中,所述查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定所述第一历史用户标识对应的对比剂量分布图,包括:
    从图像数据库获取第二历史用户标识对应的历史配准分布图;
    基于所述第一历史用户标识的历史配准分布图和历史危及器官勾画图,获取所述第一 历史用户标识对应的第一DVH图,基于第一历史用户标识的历史危及器官勾画图和第二历史用户标识对应的历史配准分布图,生成第二历史用户标识对应的第二DVH图;
    采用相似度算法对所述第一DVH图和第二DVH图进行相似度计算,获取目标相似度;
    若所述目标相似度小于第二预设阈值,则将所述第二历史用户标识的历史配准分布图作为与所述第一历史用户标识的历史配准分布图相对应的对比剂量分布图。
  6. 如权利要求1所述的基于人工智能的图像搜索方法,其中,所述基于所述目标特征向量查询放疗计划数据库,获取与所述目标特征向量相匹配的目标剂量分布图,包括:
    查询放疗计划数据库,获取任一历史用户标识对应的历史特征向量;
    计算所述目标特征向量与所述历史特征向量的目标相似值;
    若所述目标相似值大于第三预设阈值,则将所述历史特征向量对应的历史配准分布图确定为目标剂量分布图。
  7. 如权利要求1所述的基于人工智能的图像搜索方法,其中,在所述基于所述目标特征向量查询放疗计划数据库之前,所述基于人工智能的图像搜索方法还包括:
    从图像数据库中获取N个历史剂量分布图;
    采用所述图像配准算法对所述历史剂量分布图进行配准处理,获取历史配准分布图;
    将所述历史配准分布图输入图像搜索模型,生成对应的历史特征向量;
    将每一历史用户标识的所述历史特征向量和对应的所述历史配准分布图关联存储在放疗计划数据库。
  8. 一种基于人工智能的图像搜索装置,其中,包括:
    图像搜索请求获取模块,用于获取图像搜索请求,所述图像搜索请求包括目标用户标识、与所述目标用户标识相对应的原始CT图和原始危及器官勾画图;
    原始剂量分布图获取模块,用于将所述原始CT图和所述原始危及器官勾画图输入剂量分析模型,生成与所述目标用户标识对应的原始剂量分布图;
    标准剂量分布图获取模块,用于对所述原始剂量分布图进行配准处理,获取标准剂量分布图;
    目标特征向量获取模块,用于将所述标准剂量分布图输入图像搜索模型,获取所述目标用户标识对应的目标特征向量;
    目标剂量分布图获取模块,用于基于所述目标特征向量查询放疗计划数据库,获取与所述目标特征向量相匹配的目标剂量分布图。
  9. 一种计算机设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令,其中,所述处理器执行所述计算机可读指令时实现如下步骤:
    获取图像搜索请求,所述图像搜索请求包括目标用户标识、与所述目标用户标识相对应的原始CT图和原始危及器官勾画图;
    将所述原始CT图和所述原始危及器官勾画图输入剂量分析模型,生成与所述目标用户标识对应的原始剂量分布图;
    对所述原始剂量分布图进行配准处理,获取标准剂量分布图;
    将所述标准剂量分布图输入图像搜索模型,获取所述目标用户标识对应的目标特征向量;
    基于所述目标特征向量查询放疗计划数据库,获取与所述目标特征向量相匹配的目标剂量分布图。
  10. 如权利要求9所述的计算机设备,其中,在所述将所述标准剂量分布图输入图像搜索模型,获取所述目标用户标识对应的目标特征向量之前,所述处理器执行所述计算机可读指令时还实现如下步骤:
    获取第一历史用户标识的历史用户图像数据,所述历史用户图像数据包括历史CT图、历史危及器官勾画图和历史剂量分布图;
    将第一历史用户标识对应的历史CT图和历史危及器官勾画图输入到剂量分析模型,获取所述第一历史用户标识对应的分析剂量分布图;
    基于第一历史用户标识对应的历史剂量分布图和分析剂量分布图,获取历史配准分布图和分析配准分布图;
    查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定所述第一历史用户标识对应的对比剂量分布图;
    将所述第一历史用户标识对应的历史配准分布图、分析配准分布图和对比剂量分布图作为训练样本;
    将所述训练样本输入基于三元损失函数的卷积神经网络进行模型训练,获取图像搜索模型。
  11. 如权利要求10所述的计算机设备,其中,所述基于第一历史用户标识对应的历史剂量分布图和分析剂量分布图,获取历史配准分布图和分析配准分布图,包括:
    采用图像配准算法将所述第一历史用户标识对应的历史CT图与标准CT图进行配准,获取历史配准参数;
    基于所述历史配准参数对历史剂量分布图和分析剂量分布图进行图像配准,获取历史配准分布图和分析配准分布图。
  12. 如权利要求10所述的计算机设备,其中,所述查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定所述第一历史用户标识对应的对比剂量分布图,包括:
    基于所述第一历史用户标识对应的历史配准分布图确定靶区部位;
    基于所述靶区部位查询图像数据库,获取与所述历史配准分布图相似度小于第一预设阈值的对比剂量分布图。
  13. 如权利要求10所述的计算机设备,其中,所述查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定所述第一历史用户标识对应的对比剂量分布图,包括:
    从图像数据库获取第二历史用户标识对应的历史配准分布图;
    基于所述第一历史用户标识的历史配准分布图和历史危及器官勾画图,获取所述第一历史用户标识对应的第一DVH图,基于第一历史用户标识的历史危及器官勾画图和第二历史用户标识对应的历史配准分布图,生成第二历史用户标识对应的第二DVH图;
    采用相似度算法对所述第一DVH图和第二DVH图进行相似度计算,获取目标相似度;
    若所述目标相似度小于第二预设阈值,则将所述第二历史用户标识的历史配准分布图作为与所述第一历史用户标识的历史配准分布图相对应的对比剂量分布图。
  14. 如权利要求9所述的计算机设备,其中,所述基于所述目标特征向量查询放疗计划数据库,获取与所述目标特征向量相匹配的目标剂量分布图,包括:
    查询放疗计划数据库,获取任一历史用户标识对应的历史特征向量;
    计算所述目标特征向量与所述历史特征向量的目标相似值;
    若所述目标相似值大于第三预设阈值,则将所述历史特征向量对应的历史配准分布图确定为目标剂量分布图。
  15. 一个或多个存储有计算机可读指令的可读存储介质,其中,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器执行如下步骤:
    获取图像搜索请求,所述图像搜索请求包括目标用户标识、与所述目标用户标识相对应的原始CT图和原始危及器官勾画图;
    将所述原始CT图和所述原始危及器官勾画图输入剂量分析模型,生成与所述目标用户标识对应的原始剂量分布图;
    对所述原始剂量分布图进行配准处理,获取标准剂量分布图;
    将所述标准剂量分布图输入图像搜索模型,获取所述目标用户标识对应的目标特征向量;
    基于所述目标特征向量查询放疗计划数据库,获取与所述目标特征向量相匹配的目标剂量分布图。
  16. 如权利要求15所述的可读存储介质,其中,在所述将所述标准剂量分布图输入图像搜索模型,获取所述目标用户标识对应的目标特征向量之前,所述基于人工智能的图像搜索方法还包括:
    获取第一历史用户标识的历史用户图像数据,所述历史用户图像数据包括历史CT图、历史危及器官勾画图和历史剂量分布图;
    将第一历史用户标识对应的历史CT图和历史危及器官勾画图输入到剂量分析模型,获取所述第一历史用户标识对应的分析剂量分布图;
    基于第一历史用户标识对应的历史剂量分布图和分析剂量分布图,获取历史配准分布图和分析配准分布图;
    查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定所述第一历史用户标识对应的对比剂量分布图;
    将所述第一历史用户标识对应的历史配准分布图、分析配准分布图和对比剂量分布图作为训练样本;
    将所述训练样本输入基于三元损失函数的卷积神经网络进行模型训练,获取图像搜索模型。
  17. 如权利要求16所述的可读存储介质,其中,所述基于第一历史用户标识对应的历史剂量分布图和分析剂量分布图,获取历史配准分布图和分析配准分布图,包括:
    采用图像配准算法将所述第一历史用户标识对应的历史CT图与标准CT图进行配准,获取历史配准参数;
    基于所述历史配准参数对历史剂量分布图和分析剂量分布图进行图像配准,获取历史配准分布图和分析配准分布图。
  18. 如权利要求16所述的可读存储介质,其中,,所述查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定所述第一历史用户标识对应的对比剂量分布图,包括:
    基于所述第一历史用户标识对应的历史配准分布图确定靶区部位;
    基于所述靶区部位查询图像数据库,获取与所述历史配准分布图相似度小于第一预设阈值的对比剂量分布图。
  19. 如权利要求16所述的可读存储介质,其中,,所述查询图像数据库,基于其他历史用户标识的历史剂量分布图,确定所述第一历史用户标识对应的对比剂量分布图,包括:
    从图像数据库获取第二历史用户标识对应的历史配准分布图;
    基于所述第一历史用户标识的历史配准分布图和历史危及器官勾画图,获取所述第一历史用户标识对应的第一DVH图,基于第一历史用户标识的历史危及器官勾画图和第二历史用户标识对应的历史配准分布图,生成第二历史用户标识对应的第二DVH图;
    采用相似度算法对所述第一DVH图和第二DVH图进行相似度计算,获取目标相似度;
    若所述目标相似度小于第二预设阈值,则将所述第二历史用户标识的历史配准分布图作为与所述第一历史用户标识的历史配准分布图相对应的对比剂量分布图。
  20. 如权利要求15所述的可读存储介质,其中,所述基于所述目标特征向量查询放疗计划数据库,获取与所述目标特征向量相匹配的目标剂量分布图,包括:
    查询放疗计划数据库,获取任一历史用户标识对应的历史特征向量;
    计算所述目标特征向量与所述历史特征向量的目标相似值;
    若所述目标相似值大于第三预设阈值,则将所述历史特征向量对应的历史配准分布图确定为目标剂量分布图。
PCT/CN2020/093333 2020-01-14 2020-05-29 基于人工智能的图像搜索方法、装置、设备及介质 WO2021143019A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010038730.6A CN111241331A (zh) 2020-01-14 2020-01-14 基于人工智能的图像搜索方法、装置、设备及介质
CN202010038730.6 2020-01-14

Publications (1)

Publication Number Publication Date
WO2021143019A1 true WO2021143019A1 (zh) 2021-07-22

Family

ID=70872501

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/093333 WO2021143019A1 (zh) 2020-01-14 2020-05-29 基于人工智能的图像搜索方法、装置、设备及介质

Country Status (2)

Country Link
CN (1) CN111241331A (zh)
WO (1) WO2021143019A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115938591B (zh) * 2023-02-23 2023-05-09 福建自贸试验区厦门片区Manteia数据科技有限公司 基于放疗的剂量分布区间确定装置及电子设备
CN116030938B (zh) * 2023-03-29 2023-06-13 福建自贸试验区厦门片区Manteia数据科技有限公司 放疗剂量分布区间的确定装置及电子设备
CN116052840B (zh) * 2023-03-31 2023-06-13 福建自贸试验区厦门片区Manteia数据科技有限公司 基于放疗的剂量分布数据处理装置、电子设备及存储介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7046762B2 (en) * 1999-11-05 2006-05-16 Georgia Tech Research Corporation Systems and methods for global optimization of treatment planning for external beam radiation therapy
CN102184334A (zh) * 2011-05-19 2011-09-14 南京航空航天大学 一种基于检索的放射治疗计划系统及检索方法
CN104036109A (zh) * 2014-03-14 2014-09-10 上海大图医疗科技有限公司 基于图像的病例检索、勾画及治疗计划系统和方法
CN104636618A (zh) * 2015-02-09 2015-05-20 上海联影医疗科技有限公司 放疗计划系统
CN108717866A (zh) * 2018-04-03 2018-10-30 陈辛元 一种预测放疗计划剂量分布的方法、装置、设备及存储介质
CN109872804A (zh) * 2019-02-02 2019-06-11 上海赛提菲克医疗器械有限公司 一种自动放疗计划系统及其使用方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7046762B2 (en) * 1999-11-05 2006-05-16 Georgia Tech Research Corporation Systems and methods for global optimization of treatment planning for external beam radiation therapy
CN102184334A (zh) * 2011-05-19 2011-09-14 南京航空航天大学 一种基于检索的放射治疗计划系统及检索方法
CN104036109A (zh) * 2014-03-14 2014-09-10 上海大图医疗科技有限公司 基于图像的病例检索、勾画及治疗计划系统和方法
CN104636618A (zh) * 2015-02-09 2015-05-20 上海联影医疗科技有限公司 放疗计划系统
CN108717866A (zh) * 2018-04-03 2018-10-30 陈辛元 一种预测放疗计划剂量分布的方法、装置、设备及存储介质
CN109872804A (zh) * 2019-02-02 2019-06-11 上海赛提菲克医疗器械有限公司 一种自动放疗计划系统及其使用方法

Also Published As

Publication number Publication date
CN111241331A (zh) 2020-06-05

Similar Documents

Publication Publication Date Title
WO2021143019A1 (zh) 基于人工智能的图像搜索方法、装置、设备及介质
Men et al. More accurate and efficient segmentation of organs‐at‐risk in radiotherapy with convolutional neural networks cascades
US11715203B2 (en) Image processing method and apparatus, server, and storage medium
Shiri et al. Decentralized distributed multi-institutional PET image segmentation using a federated deep learning framework
Zhou et al. A method of using deep learning to predict three‐dimensional dose distributions for intensity‐modulated radiotherapy of rectal cancer
US10850120B2 (en) Selecting a dose prediction model based on clinical goals
US20220076052A1 (en) Similarity determining method and device, network training method and device, search method and device, and electronic device and storage medium
WO2022178997A1 (zh) 医学影像配准方法、装置、计算机设备及存储介质
Wu et al. Development and evaluation of a deep learning algorithm for rib segmentation and fracture detection from multicenter chest CT images
CN111709485A (zh) 医学影像处理方法、装置和计算机设备
US20210204898A1 (en) Medical image analyzing system and method thereof
Chao et al. Image‐based modeling of tumor shrinkage in head and neck radiation therapy a
CN112001979B (zh) 运动伪影处理方法、系统、可读存储介质和设备
WO2021135063A1 (zh) 病理数据分析方法、装置、设备及存储介质
Guan et al. Big data analytics on lung cancer diagnosis framework with deep learning
Haq et al. Dynamic multiatlas selection‐based consensus segmentation of head and neck structures from CT images
CN116350958A (zh) 放疗计划参数的设定方法、装置、设备、介质和程序产品
Yan et al. A pseudoinverse deformation vector field generator and its applications
Javaid et al. Semantic segmentation of computed tomography for radiotherapy with deep learning: compensating insufficient annotation quality using contour augmentation
CN112635023B (zh) 鼻咽癌的剂量预测模型的生成方法、剂量预测方法及装置
WO2022217715A1 (zh) 相似患者的识别方法、装置、计算机设备和存储介质
Jian et al. DBPNDNet: dual-branch networks using 3DCNN toward pulmonary nodule detection
CN111681237B (zh) 图像配准方法、装置、计算机设备和存储介质
Vazquez et al. A deep learning-based approach for statistical robustness evaluation in proton therapy treatment planning: a feasibility study
Davey et al. Automated gross tumor volume contour generation for large‐scale analysis of early‐stage lung cancer patients planned with 4D‐CT

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: 20913617

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: 20913617

Country of ref document: EP

Kind code of ref document: A1