Disclosure of Invention
The invention provides a non-calcified plaque detection method and a device thereof, which can obtain accurate non-calcified plaque on an image.
One aspect of the present invention provides a method for detecting non-calcified plaque, including: carrying out similarity model training on the blood vessel image sample to obtain a similarity network training model of the blood vessel image; predicting a preliminary non-calcified plaque classification result in the blood vessel image through a similarity network training model; and screening out the asymmetric similar part in the preliminary classification result of the non-calcified plaque to obtain the final classification result of the non-calcified plaque.
In an implementation manner, the similarity training of the blood vessel image sample to obtain a similarity network training model of the blood vessel image includes: inputting a label containing a non-calcified plaque pixel region in a blood vessel image sample, and giving a penalty term to the non-calcified plaque pixel region to obtain the label containing the penalty term non-calcified plaque pixel region; modifying the similarity measurement function through the label of the non-calcified plaque pixel area containing the punishment item; and training the similarity network through the modified similarity measurement function to obtain a similarity network training model.
In one embodiment, the predicting the preliminary non-calcified plaque classification result in the blood vessel image by the similarity network training model comprises: determining a blood vessel region in the blood vessel image, and obtaining a similar blood vessel region and a dissimilar blood vessel region in the blood vessel region through a similarity network training model; identifying a vessel combination region including at least the obtained one similar vessel region and one dissimilar vessel region; predicting the vessel combination region as the preliminary non-calcified plaque classification result.
In one embodiment, the screening out the non-symmetrical similar parts of the preliminary non-calcified plaque classification result to obtain the final non-calcified plaque classification result comprises: generating a symmetric similarity label on a non-calcified plaque sample, and performing similarity training on a similarity network training model through the non-calcified plaque sample containing the symmetric similarity label or performing symmetric similarity model training to obtain a symmetric similarity model of the non-calcified plaque; and screening out the preliminary classification result of the non-calcified plaque through the symmetrical similar model of the non-calcified plaque, and confirming the classification result of the non-calcified plaque.
In an implementation manner, the generating a symmetric similarity label on a non-calcified plaque sample, and performing similarity training on a similarity network training model or performing symmetric similarity model training on the non-calcified plaque sample containing the symmetric similarity label to obtain a symmetric similarity model of the non-calcified plaque includes: inputting a label containing a symmetrical similarity cut pixel region in a non-calcified plaque sample, and giving a penalty item to the symmetrical similarity cut pixel region to obtain a label containing the penalty item and the symmetrical similarity cut pixel region; modifying a similarity measurement function through modifying the label containing the symmetrical similarity block pixel area, and training a similarity network training model through the modified similarity measurement function to obtain a symmetrical similar model of the non-calcified plaque; or modifying the similarity measurement function by modifying the label containing the symmetrical similarity block pixel area, and carrying out symmetrical similarity model training by the modified similarity measurement function.
Another aspect of the present invention provides a detection apparatus of non-calcified plaque, comprising: the training module is used for carrying out similarity model training on the blood vessel image samples to obtain a similarity network training model of the blood vessel images; the prediction module is used for predicting a preliminary non-calcified plaque classification result in the blood vessel image through the similarity network training model; and the screening module is used for screening out the asymmetric similar part in the preliminary non-calcified plaque classification result to obtain a final non-calcified plaque classification result.
In one embodiment, the training module comprises: the input unit is used for inputting a label containing a non-calcified plaque pixel area in a blood vessel image sample, giving a penalty term to the non-calcified plaque pixel area and obtaining the label containing the penalty term and the non-calcified plaque pixel area; the modification unit is used for modifying the similarity measurement function through the label of the non-calcified plaque pixel area containing the penalty term; and the training unit is used for training the similarity network through the modified similarity measurement function to obtain a similarity network training model.
In one embodiment, the prediction module comprises: the determining unit is used for determining a blood vessel region in the blood vessel image and obtaining a similar blood vessel region and a dissimilar blood vessel region in the blood vessel region through a similarity network training model; an identifying unit for identifying a vessel combination region including at least one similar vessel region and one dissimilar vessel region obtained; a prediction unit for predicting the vessel combination region as the preliminary non-calcified plaque classification result.
In one embodiment, the sifting module comprises: the generation training unit is used for generating a symmetric similarity label on a non-calcified plaque sample, and performing similarity training on a similarity network training model through the non-calcified plaque sample containing the symmetric similarity label or performing symmetric similar model training to obtain a symmetric similar model of the non-calcified plaque; and the screening and confirming unit is used for screening and removing the preliminary non-calcified plaque classification result through the symmetrical similar model of the non-calcified plaque and confirming the classification result of the non-calcified plaque.
In one embodiment, the generating the training unit includes: the input subunit is used for inputting a label containing a symmetric similarity cut pixel region in a non-calcified plaque sample, and giving a penalty item to the symmetric similarity cut pixel region to obtain a label containing the penalty item and the symmetric similarity cut pixel region; the modifying subunit is used for modifying the similarity measurement function through modifying the label containing the symmetric similarity block pixel area, and training the similarity network training model through the modified similarity measurement function to obtain a symmetric similarity model of the non-calcified plaque; or modifying the similarity measurement function by modifying the label containing the symmetrical similarity block pixel area, and carrying out symmetrical similarity model training by the modified similarity measurement function.
According to the non-calcified plaque detection method and the non-calcified plaque detection device, model training is carried out by using the blood vessel image sample to obtain the similarity network training model, the preliminary non-calcified plaque classification result in the image is predicted through the similarity network training model, the preliminary non-calcified plaque classification result is screened out, and the accuracy of the non-calcified plaque classification result is further confirmed, so that the final non-calcified plaque classification result with accurate position, accurate quantity and high reliability is obtained, the position of the non-calcified plaque is not required to be manually searched and determined, the automatic searching of the non-calcified plaque classification result is realized, and the labor cost is saved.
Detailed Description
In order to make the objects, features and advantages of the present invention more obvious and understandable, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Fig. 1 shows a flow chart of a method for detecting non-calcified plaque according to an embodiment of the present invention.
Referring to fig. 1, in one aspect, the embodiment of the present invention provides a method for detecting non-calcified plaque, including the following steps: 101, performing similarity model training on a blood vessel image sample to obtain a similarity network training model of the blood vessel image; step 102, predicting a preliminary non-calcified plaque classification result in a blood vessel image through a similarity network training model; and 103, screening out asymmetrical similar parts in the preliminary classification result of the non-calcified plaque to obtain a final classification result of the non-calcified plaque.
By the detection method provided by the embodiment of the invention, the specific position of the non-calcified plaque can be automatically detected on the image containing the blood vessel image without manual operation, the accuracy of the detected non-calcified plaque is high, the exact position and the exact number of the non-calcified plaque can be obtained, and the condition of missed detection is avoided.
The blood vessel image of the embodiment of the invention is a CT image. The detection method provided by the embodiment of the invention mainly comprises three steps. Before processing a CT image, a similarity network training model needs to be obtained, and the similarity network training model is obtained by performing similarity network model training on a blood vessel image sample. It should be understood that the blood vessel image sample of the embodiment of the present invention should include the blood vessel region of the non-calcified plaque in addition to the normal blood vessel region. The normal blood vessel region is a continuous similar blood vessel, and there is a difference between the blood vessel region where the non-calcified plaque occurs and the normal blood vessel region. In the training process, the similarity is defined as 1 for continuous similar vessel regions, namely normal vessel regions, and the similarity is defined as 0 for dissimilar vessel regions with non-calcified plaques. The label is used for neural network model training, and a similarity network training model with more accurate prediction effect on the blood vessel image can be obtained.
And predicting the CT image by using the similarity network training model, so that a preliminary non-calcified plaque classification result in the CT image can be obtained. In order to avoid multiple selection, the embodiment of the invention also screens out the preliminary non-calcified plaque classification result and removes the asymmetric similar part, so that the final non-calcified plaque classification result with accurate position and accurate quantity is obtained on the CT image.
The embodiment of the present invention includes, in step 101: firstly, inputting a label containing a non-calcified plaque pixel area in a blood vessel image sample, giving a penalty term to the non-calcified plaque pixel area, and obtaining the label containing the penalty term and the non-calcified plaque pixel area. The similarity metric function is then modified by tags containing punishment terms non-calcified plaque pixel regions. And then, training the similarity network through the modified similarity measurement function to obtain a similarity network training model.
Specifically, in the process of performing similarity model training, a label containing a non-calcified plaque pixel region is required to perform model training, in the process of performing model training through the label containing the non-calcified plaque pixel region, a penalty item is given to the non-calcified plaque pixel region in the label, a similarity measurement function is modified through the label containing the penalty item, and then a similarity network is trained through the modified similarity measurement function. The difference between the non-calcified plaque pixel region and other pixel regions in the CT image, particularly the difference between the non-calcified plaque pixel region and a normal blood vessel region can be improved, so that the similarity network training model has higher accuracy in predicting the candidate region of the non-calcified plaque.
For example, the penalty term in the embodiment of the present invention may be obtained by multiplying the pixel regions of the non-calcified plaque by "x-1" or "x-2", or by multiplying by another numerical value or by using another calculation method to enlarge the difference between the pixel region of the non-calcified plaque and another pixel region in the CT image, thereby enlarging the similarity between the pixel region of the non-calcified plaque and another pixel region in the CT image, and improving the accuracy of the preliminary classification result of the non-calcified plaque predicted by the similarity network training model, so as to further obtain a more accurate location and an accurate number of classification results of the non-calcified plaque on the CT image.
The embodiment of the invention comprises the following steps in step 102: firstly, determining a blood vessel region in a blood vessel image, and training a model through a similarity network to obtain a similar blood vessel region and a dissimilar blood vessel region in the blood vessel region. Then, a vessel combination region including at least one similar vessel region and one dissimilar vessel region is identified. And then, predicting the blood vessel combination area as a preliminary non-calcified plaque classification result.
Specifically, the purpose of distinguishing similar vessel regions from dissimilar vessel regions is achieved through a similarity network training model. It should be understood that, in the CT image, the similar blood vessel region is a portion similar to the blood vessel region, i.e. a blood vessel portion without non-calcified plaque; the dissimilar blood vessel region is a portion dissimilar to the blood vessel region, i.e., a portion of the blood vessel in which the non-calcified plaque occurs. For example, in the prediction, there is a blood vessel combination region including a similar blood vessel region, below the similar blood vessel region, there is a dissimilar blood vessel region, and further below, there is a similar blood vessel region, and such a blood vessel combination region is determined as a candidate region of a non-calcified plaque.
It should be noted that this example is only one of the blood vessel combination regions capable of determining the preliminary non-calcified plaque classification result, as long as the blood vessel combination including at least one dissimilar blood vessel region and at least one similar blood vessel region is satisfied, the preliminary non-calcified plaque classification result should be determined.
The embodiment of the present invention comprises in step 103: firstly, generating a symmetric similarity label on a non-calcified plaque sample, and performing similarity training on a similarity network training model through the non-calcified plaque sample containing the symmetric similarity label or performing symmetric similarity model training to obtain a symmetric similarity model of the non-calcified plaque. Then, the candidate regions of the non-calcified plaque are screened out through the symmetrical similar model of the non-calcified plaque, and the classification result of the non-calcified plaque is confirmed.
In order to further improve the accuracy of the classification result of the non-calcified plaque, the embodiment of the invention needs to further screen after obtaining the preliminary classification result of the non-calcified plaque. By utilizing the characteristic that non-calcified plaques have vertical symmetry, when designing screened training data, vertically symmetrical blocks are cut for the non-calcified plaque area. The areas of non-calcified plaque here are the data used in the model training process. Further, the symmetric similarity model of the non-calcified plaque obtained by model training may be a new model obtained by model training for screening, or may be further trained on a similarity network training model. The embodiment of the invention does not limit the specific implementation mode of obtaining the symmetrical similar model of the non-calcified plaque, and only needs to satisfy the requirement that the obtained symmetrical similar model of the non-calcified plaque can screen out the preliminary classification result of the non-calcified plaque through the up-down symmetry of the non-calcified plaque.
In the embodiment of the present invention, a symmetric similarity label is generated on a non-calcified plaque sample, and a similarity network training model is subjected to similarity training or symmetric similarity model training through the non-calcified plaque sample containing the symmetric similarity label to obtain a symmetric similarity model of the non-calcified plaque, wherein the step specifically includes: firstly, inputting a label containing a symmetrical similarity cut pixel region in a non-calcified plaque sample, and giving a penalty term to the symmetrical similarity cut pixel region to obtain the label containing the penalty term and the symmetrical similarity cut pixel region. And then, modifying the similarity measurement function by modifying the label of the pixel area containing the symmetrical similarity cut block, and training the similarity network training model by the modified similarity measurement function to obtain the symmetrical similarity model of the non-calcified plaque. Or modifying the similarity measurement function by modifying the label containing the symmetrical similarity block pixel area, and carrying out symmetrical similarity model training by the modified similarity measurement function.
Specifically, in the embodiment of the invention, in the model training process of the symmetric similar model of the non-calcified plaque, a penalty term can be added to the training data to design the loss function. For example, the difference between the pixel region of the non-calcified plaque and the other pixel regions in the CT image can be enlarged by multiplying other values or by other calculation methods by "x 1" and "x 2" of the non-calcified plaque and by multiplying the blood vessel region, i.e., the region where the calcified plaque does not occur "x-1" and "x-2". It should be noted that, when "the non-calcified plaque pixel region x-1 or x-2" and "the non-calcified plaque region x 1 or x 2" and the blood vessel region, i.e. the region where calcified plaque does not occur x-1 or x-2 ", in step 103, the difference between the non-calcified plaque and the normal blood vessel portion in the CT image can be further improved, and the accuracy of the screening process can be improved, so as to meet the requirement of obtaining an accurate position and an accurate number of classification results of the non-calcified plaque on the CT image.
The CT image referred to in the embodiment of the present invention is original CT image volume data, and the original CT image volume data is divided and then subjected to search prediction screening by a BUFFER to obtain a calcified plaque classification result. However, the image of the present invention is not limited to CT original image volume data, and other original images may be predicted.
Fig. 2 is a schematic structural diagram of a non-calcified plaque detection apparatus according to an embodiment of the present invention.
Referring to fig. 2, in one aspect, an embodiment of the present invention provides an apparatus for detecting non-calcified plaque, including: the training module 201 is configured to perform similarity model training on the blood vessel image sample to obtain a similarity network training model of the blood vessel image. The prediction module 202 is configured to predict a preliminary non-calcified plaque classification result in the blood vessel image through the similarity network training model. And the screening module 203 is used for screening out asymmetric similar parts in the preliminary non-calcified plaque classification result to obtain a final non-calcified plaque classification result.
The training module 201 of the embodiment of the invention comprises: the input unit 2011 is configured to input a label of a non-calcified plaque pixel region in the blood vessel image sample, and give a penalty term to the non-calcified plaque pixel region to obtain a label of the non-calcified plaque pixel region containing the penalty term. A modifying unit 2012, configured to modify the similarity measure function by using the label containing the punishment term non-calcified plaque pixel region. And the training unit 2013 is used for training the similarity network through the modified similarity measurement function to obtain a similarity network training model.
The prediction module 202 of the embodiment of the present invention includes: the determining unit 2021 is configured to determine a blood vessel region in the blood vessel image, and obtain a similar blood vessel region and a dissimilar blood vessel region in the blood vessel region through a similarity network training model. An identifying unit 2022 for identifying a vessel combination region comprising at least the obtained one similar vessel region and one dissimilar vessel region. A prediction unit 2023, configured to predict the blood vessel combination region as a preliminary non-calcified plaque classification result.
The screening module 203 of the embodiment of the invention comprises: the generation training unit 2031 is configured to generate a symmetric similarity label on the non-calcified plaque sample, and perform similarity training on the similarity network training model through the non-calcified plaque sample containing the symmetric similarity label, or perform symmetric similarity model training to obtain a symmetric similarity model of the non-calcified plaque. The screening confirmation unit 2032 is configured to screen out the candidate region of the non-calcified plaque by using the symmetric similarity model of the non-calcified plaque, and confirm the classification result of the non-calcified plaque.
The generation training unit 2031 of the embodiment of the present invention includes: an input subunit 20311, configured to input a label containing a symmetric similarity diced pixel region in a non-calcified plaque sample, and give a penalty term to the symmetric similarity diced pixel region, to obtain a label containing a penalty term and a symmetric similarity diced pixel region; a modifying subunit 20322, configured to modify the similarity metric function by modifying the label containing the symmetric similarity tile pixel region, and train the similarity network training model by using the modified similarity metric function, to obtain a symmetric similarity model of the non-calcified plaque; or modifying the similarity measurement function by modifying the label containing the symmetrical similarity block pixel area, and carrying out symmetrical similarity model training by the modified similarity measurement function.
In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.
Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality" means two or more unless specifically defined otherwise.
The above description is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present invention, and all the changes or substitutions should be covered within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the appended claims.