CN113781461A - Intelligent patient monitoring and sequencing method - Google Patents

Intelligent patient monitoring and sequencing method Download PDF

Info

Publication number
CN113781461A
CN113781461A CN202111089390.0A CN202111089390A CN113781461A CN 113781461 A CN113781461 A CN 113781461A CN 202111089390 A CN202111089390 A CN 202111089390A CN 113781461 A CN113781461 A CN 113781461A
Authority
CN
China
Prior art keywords
network
image
constructing
artifact
identification
Prior art date
Legal status (The legal status 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 status listed.)
Pending
Application number
CN202111089390.0A
Other languages
Chinese (zh)
Inventor
曾栋
谌高峰
王子丹
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guangdong Provincial Laboratory Of Artificial Intelligence And Digital Economy Guangzhou
Original Assignee
Guangdong Provincial Laboratory Of Artificial Intelligence And Digital Economy Guangzhou
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 Guangdong Provincial Laboratory Of Artificial Intelligence And Digital Economy Guangzhou filed Critical Guangdong Provincial Laboratory Of Artificial Intelligence And Digital Economy Guangzhou
Priority to CN202111089390.0A priority Critical patent/CN113781461A/en
Publication of CN113781461A publication Critical patent/CN113781461A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—2D [Two Dimensional] image generation
    • G06T11/003—Reconstruction from projections, e.g. tomography
    • G06T11/005—Specific pre-processing for tomographic reconstruction, e.g. calibration, source positioning, rebinning, scatter correction, retrospective gating
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—2D [Two Dimensional] image generation
    • G06T11/003—Reconstruction from projections, e.g. tomography
    • G06T11/006—Inverse problem, transformation from projection-space into object-space, e.g. transform methods, back-projection, algebraic methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—2D [Two Dimensional] image generation
    • G06T11/003—Reconstruction from projections, e.g. tomography
    • G06T11/008—Specific post-processing after tomographic reconstruction, e.g. voxelisation, metal artifact correction
    • G06T5/73—
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10081—Computed x-ray tomography [CT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30101—Blood vessel; Artery; Vein; Vascular
    • G06T2207/30104—Vascular flow; Blood flow; Perfusion
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2211/00—Image generation
    • G06T2211/40—Computed tomography
    • G06T2211/404—Angiography
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2211/00—Image generation
    • G06T2211/40—Computed tomography
    • G06T2211/421—Filtered back projection [FBP]

Abstract

The invention provides a patient intelligent monitoring and sequencing method, which comprises S1, carrying out self-adaptive optimization on a CT scanning protocol based on a depth recognition model; s2, identifying and automatically correcting the CT artifact based on a countermeasure technology; s3, CT quantitative imaging is carried out based on the incremental learning technology; s4, carrying out intelligent analysis on the CT image based on the transfer learning technology; and S5, making an auxiliary decision on the CT image. The invention constructs a strategy network for intelligent monitoring and sequencing of emergency patients, and the network architecture can be based on various deep convolutional network technologies, such as a high-dimensional convolutional neural network, and aims to screen CT projection data of cardiovascular critical and severe patients; the high-dimensional convolutional neural network finally determines whether the image contains the acute cardiovascular disease, so that the image is diagnosed according to the urgent-heavy priority sequence instead of the original acquisition sequence, limited medical resources of a hospital are optimized, and the purpose of quickly diagnosing and treating the urgent-heavy priority patient is achieved.

Description

Intelligent patient monitoring and sequencing method
Technical Field
The invention relates to the field of sequencing, in particular to an intelligent patient monitoring and sequencing method.
Background
With the continuous improvement of the medical level of China, at present that CT equipment is increasingly popularized, the traditional technical innovation of CT products is kept since the time of marketization, new generation products are continuously developed so as to develop a new field of clinical application, the original products are qualitatively changed due to the updating and upgrading of the products, new functions are provided, new requirements of customers can be met, and sufficient profits are kept for enterprises while new using function values of the customers are created. In recent five years, the AI technology has become mature in data, algorithms, computing power and other aspects, and has often made breakthrough, thus actually solving the practical problems and actually creating economic effects. As a significant representative of large data distribution, the medical industry is expected to become one of the widest industries with application prospects, so that the method has great commercialization potential.
Therefore, it is particularly necessary to solve the problem of the imbalance of traditional medical resources by fusing the artificial intelligence technology and the large-scale medical CT data. Therefore, an intelligent patient monitoring and sequencing method is provided.
Disclosure of Invention
In view of the above problems, the present invention aims to provide a method for sequencing patient intelligent monitoring, comprising:
s1, carrying out self-adaptive optimization on the CT scanning protocol based on the depth recognition model;
s2, identifying and automatically correcting the CT artifact based on a countermeasure technology;
s3, CT quantitative imaging is carried out based on the incremental learning technology;
s4, carrying out intelligent analysis on the CT image based on the transfer learning technology;
and S5, making an auxiliary decision on the CT image.
Preferably, the S1 includes:
s11, constructing an organ region anatomical feature identification network;
and S12, constructing a CT personalized scanning protocol task guide network.
Preferably, the S11 includes:
s1101, constructing a three-dimensional positioning image imaging sub-network:
the depth residual error learning sub-network comprises a noise estimation block, an artifact estimation block and an image filtering block, wherein the noise estimation block, the artifact estimation block and the image filtering block are respectively processed through a residual error learning network, a multi-scale wavelet and a convolutional neural network;
carrying out artifact estimation based on a processing flow of multi-scale wavelet transform, further improving the image quality of the microdose image by an iteration method, and finally obtaining a three-dimensional positioning image;
s1102, constructing an organ region identification subnetwork:
taking the recovered micro-dose three-dimensional positioning image output by the deep residual learning network as the input of the identification network;
wherein the organ region identification network comprises an encoding process and a decoding process;
wherein, the coding process adopts the structure of residual block, and the decoding process adopts the full convolution network; the training set is defined as X, where XiFor the ith input image, YiLabels for the ith input image, wherein I is defined as a different reconstruction target region; simultaneous definition of
Figure BDA0003266756520000021
Probability of kth pixel of ith input image;
Figure BDA0003266756520000022
the probability map for image level prediction can be obtained by all computations at pixel level, and the cost function is:
Figure BDA0003266756520000023
the three-dimensional positioning image imaging sub-network and the organ region identification sub-network are cascaded to obtain a final organ region anatomical feature identification network, L represents a cost function, and r represents a control parameter.
Preferably, the S12 includes:
s1201, constructing a depth recognition network, wherein the constructed depth recognition model is based on a bidirectional depth recursive network;
s1202, three-dimensional positioning image and target obtained from organ region anatomical feature identification network are utilizedThe target organ area is used as the input of a high-dimensional image feature extraction network to extract the morphological features of the three-dimensional positioning image and the target organ area
Figure BDA0003266756520000024
Texture features
Figure BDA0003266756520000025
S1203, personalized scanning parameter estimation, task-driven, patient-driven multi-objective optimization,
wherein, the optimization process of the personalized scanning parameter estimation in the step S1203 is solved through an objective equation
Figure BDA0003266756520000026
Wherein omegaAIs the adaptive scanning parameter to be solved; omegaRAdaptive reconstruction parameters to be solved; s represents an estimate of the local noise power spectrum, T represents an estimate of the local modulation transfer function,
Figure BDA0003266756520000027
is based on task-driven and patient-driven parameter estimation, and contains patient core information
Figure BDA0003266756520000028
And high-dimensional morphological characteristics of three-dimensional positioning image
Figure BDA0003266756520000029
Texture features
Figure BDA00032667565200000210
Namely, it is
Figure BDA00032667565200000211
Finally, optimizing the self-adaptive exposure parameter omega under each angle through an ADMM-Net networkAAnd an adaptive reconstruction parameter omegaR(ii) a The cascade network adopts a small batch random gradient descent method to carry out hierarchical training on the network，fx，fy，fzIndicating the position information of the image f at coordinate points x, y, z.
Preferably, the S2 includes:
s21, constructing an antagonistic active learning network for automatic CT artifact identification;
s22, constructing a confrontation learning network for automatically correcting the CT artifact;
the S21 includes:
performing network research design based on active learning and antagonistic learning theory technology; the active learning network can assist artifact labeling in data, wherein the active learning network A is (C, Q, S, L and U), and C is a machine learning model and is used for CT artifact identification and identification; q is a query function, and a committee-based selection algorithm is adopted as a query strategy and is used for screening data containing large information amount from the unlabeled sample pool U; s is a supervisor, selects correct labels in the sample U, and trains a classifier and the next round of query by using the obtained new knowledge;
the S22 includes:
s2201, constructing an artifact automatic correction confrontation learning network with artifact identification type indexes, and identifying artifact types:
s2202, according to the artifact identification result, marking and indexing each type of artifact;
s2203, aiming at different artifacts, adopting different countermeasure networks to automatically correct;
s2204, according to the identification result and range in the artifact identification network, a feature extraction method is adopted to perform feature analysis and extraction on the artifact part in the chord graph, and specific loss functions are constructed for different features so as to realize artifact automatic correction based on artifact structural feature difference;
s2205, for the artifacts meeting the preset conditions, performing multi-scale directional field processing on the CT projection data according to the identification result, extracting multi-scale directional data information of the artifact part in the projection data, and setting a countermeasure network cost function according to the multi-scale directional data information to obtain the projection data after artifact correction.
Preferably, the S3 includes:
s31, constructing a CT quantitative imaging semi-supervised incremental learning network; constructing a low-dose CT quantitative imaging supervised learning network:
the low-dose CT quantitative imaging supervised learning network comprises a fully-connected filter layer, a sine back-projection layer and a residual convolutional neural network, wherein the fully-connected filter layer is optimally designed based on a filter kernel of a filter back-projection algorithm, the sine back-projection layer corresponds to a back-projection operator in the filter back-projection algorithm, the residual convolutional neural network is used for further optimizing a reconstruction result, and a network cost function is designed to be 2 norm root mean square error and is recorded as:
Figure BDA0003266756520000031
wherein
Figure BDA0003266756520000032
Reconstructing the final predicted image of the network for CT, xH refThe CT image is a target CT image, N is the number of training samples, and theta is a parameter needing to be learned;
s32, constructing a semi-supervised incremental learning network: inputting noise-containing CT projection data acquired from an external source into a pre-trained FBP net after passing through various deformation fields, performing self-adaptive weighted fusion on CT image results to obtain a fused high-quality CT target image, and performing depth optimization on FBP net parameters again to obtain a semi-supervised quantitative imaging model;
s33, constructing a CT quantitative imaging unsupervised incremental learning network; designing a CT quantitative imaging unsupervised incremental learning network based on the statistical characteristics of noisy CT projection data before Log transformation and the sparsity of a high-order derivative of the CT projection data; an objective function in a deep learning network is constructed based on a maximum posterior probability framework, the objective function comprises a data fidelity item based on a likelihood function and a prior information item based on piecewise linearity, the objective function can more accurately describe CT projection data and can carry out training optimization along a gradient direction, and the expression is as follows:
Figure BDA0003266756520000041
wherein, C1Representing noisy CT projection training data set, yiThe ith noisy CT projection training data is obtained, N represents the number of data sets, G represents the number of X-ray photons satisfying the composite Poisson statistical characteristics, I represents the number of photons received by a detector, and I represents the number of the X-ray photons received by the detectoriRepresenting the number of photons received by the i-th detector, GiRepresents IiApproximately middle value of, Gj| A Denotes the multiplication value of G in the j-th projection, I0jRepresenting the number of incident photons in the j projection, ε represents the electronic noise, and obeys a mean of 0 and a variance of σ2Gaussian distribution of (D)2Representing a second order difference operator, fθRepresenting unsupervised learning network mapping, wherein theta is a parameter needing to be learned in the network mapping, a high-quality CT image is obtained through reconstruction by utilizing a filtering back-projection algorithm, and fθ(yi)jRepresenting the training data y for the ith noisy CT projectioniUnsupervised learning network mapping on pixel j.
Preferably, the S4 includes:
s41, constructing a vascular qualitative identification network based on the video depth migration learning technology, including:
s4101, constructing a dynamic depth of field area identification network: the training set is defined as s { (X)i,Yi) I ═ 1,2,3, …, n }, where X isiInputting images for a three-dimensional sequence of ith time frame, YiFor the ith time frame of tag data,
Figure BDA0003266756520000042
for the network output result, the cost function is:
Figure BDA0003266756520000043
s4102, constructing a video depth of field migration learning network;
s42, constructing a blood vessel quantitative recognition network based on medical multi-modal migration imaging, comprising the following steps:
s4201, constructing a blood vessel quantitative identification network: the training set is defined as S { (X)i，Yi) I ═ 1,2,3, …, n }, where X isiFor the ith input image, YiE {0,1} is the label for the ith input image, where Yi1 is defined as an abnormal image, Yi0 is defined as a non-anomalous image;
definition of
Figure BDA0003266756520000044
Is the probability of the kth pixel of the ith input image, where k ═ 1,2,3, …, | Xi|}，|Xi| represents XiTotal number of pixels, if
Figure BDA0003266756520000045
For a probability map of image level prediction, then
Figure BDA0003266756520000046
All through the pixel level
Figure BDA0003266756520000047
And calculating to obtain the cost function:
Figure BDA0003266756520000051
wherein I (-) is an indicator function,
Figure BDA0003266756520000052
calculated by the Soft max Function, r is a control parameter; l isMILRepresenting a cost function;
s4202, constructing the multi-mode migration learning network.
Preferably, the S5 includes:
s51, constructing a strategy network for intelligent monitoring and sequencing of emergency patients;
and S52, constructing a monitoring sequencing result obtained by a depth network generated by the auxiliary report facing the vascular diseases.
The invention has the following effective effects:
1. the invention constructs a strategy network for intelligent monitoring and sequencing of emergency patients, and the network architecture can be based on various deep convolutional network technologies, such as a high-dimensional convolutional neural network, and aims to screen CT projection data of cardiovascular critical and severe patients; the high-dimensional convolutional neural network finally determines whether the image contains the acute cardiovascular disease, so that the image is diagnosed according to the urgent-heavy priority sequence instead of the original acquisition sequence, limited medical resources of a hospital are optimized, and the purpose of quickly diagnosing and treating the urgent-heavy priority patient is achieved.
2. Aiming at three problems faced by CT imaging in clinical application, the invention combines an Artificial Intelligence (AI) technology, brings each problem into a deep learning network, sequentially progresses each sub-network simultaneously, integrates the sub-networks into a whole, has definite target, and constructs a new framework of CT imaging and intelligent analysis facing multi-task scanning protocol self-adaptive optimization, micro-radiation dose CT quantitative imaging and CT imaging intelligent auxiliary diagnosis.
3. The new framework of the CT-AI image system constructed by the invention is clinical task oriented and has more clinical practical significance, and the framework deeply analyzes the potential and the demand of each component of the CT image system based on the CT image scanning-reconstruction-auxiliary analysis integrated system, thereby being convenient for the subsequent reconstruction and upgrade of the CT image system.
4. The invention effectively solves the problems that the current CT CAD x technology is mostly based on CT images, and omits richer and more detailed medical information contained in projection data; according to the integrated framework, the prediction error of the output end can be finally transmitted to the CT projection data of the input end, which is equivalent to positioning and identifying cardiovascular abnormalities by utilizing the projection information under multiple angles from the projection directly, namely predicting from the original data.
Drawings
The invention is further illustrated by means of the attached drawings, but the embodiments in the drawings do not constitute any limitation to the invention, and for a person skilled in the art, other drawings can be obtained on the basis of the following drawings without inventive effort.
Fig. 1 is a diagram of an exemplary embodiment of a patient intelligent monitoring and sequencing method according to the present invention.
Detailed Description
Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the accompanying drawings are illustrative only for the purpose of explaining the present invention, and are not to be construed as limiting the present invention.
In one embodiment, as shown in FIG. 1, the invention provides a patient intelligent monitoring sequencing method, comprising
The technical scheme adopted by the invention is as follows:
an intelligent patient monitoring and sequencing method comprises the following steps:
s1, carrying out self-adaptive optimization on the CT scanning protocol based on the depth recognition model, including:
s11, constructing an organ region anatomical feature identification network;
s12, constructing a task guide network of the CT personalized scanning protocol;
the identification network can accurately estimate the target organ anatomical region under the micro radiation dose, and provides guidance for the optimization of a subsequent CT scanning protocol;
s2, identifying and automatically correcting the CT artifact based on a countermeasure technology;
s3, CT quantitative imaging is carried out based on the incremental learning technology;
s4, carrying out intelligent analysis on the CT image based on the transfer learning technology;
and S5, making an auxiliary decision on the CT image.
Wherein the S11 includes the following steps:
s1101, constructing a three-dimensional positioning image imaging sub-network; the depth residual error learning sub-network comprises a noise estimation block, an artifact estimation block and an image filtering block, wherein the noise estimation, the artifact estimation and the image filtering are respectively processed through a residual error learning network, a multi-scale wavelet and a convolutional neural network; carrying out artifact estimation based on a processing flow of multi-scale wavelet transform, further improving the image quality of the microdose image by an iteration method, and finally obtaining a high-quality three-dimensional positioning image;
s1102, construction of organsArea recognition subnetwork: the recovered micro-dose three-dimensional positioning image output by the deep residual learning network is used as the input of an identification network, and the segmentation and identification effects at the pixel level are achieved by utilizing the segmentation and identification network structure based on the semantics; wherein the organ region identification network comprises an encoding process and a decoding process; wherein, the coding process adopts the structure of residual block, and the decoding process adopts the full convolution network; the training set is defined as X, where XiFor the ith input image, YiLabels for the ith input image, wherein I is defined as a different reconstruction target region; simultaneous definition of
Figure BDA0003266756520000061
Probability of kth pixel of ith input image;
Figure BDA0003266756520000062
the probability map for image level prediction can be obtained by all computations at pixel level, and the cost function is:
Figure BDA0003266756520000071
the three-dimensional positioning image imaging sub-network and the organ region identification sub-network are cascaded to obtain a final organ region anatomical feature identification network, L represents a cost function, and r represents a control parameter.
S12, constructing a CT personalized scanning protocol task guide network, comprising the following steps:
s1201, constructing a deep recognition network, and quickly and accurately extracting the physical sign information of the patient, including biochemical indexes, age, height and the like. The built depth recognition model is based on a bi-directional depth recursive network, wherein the recursive network has been used with great success and wide application in a plurality of natural language processing. The core information of the patient can be efficiently extracted by utilizing a bidirectional deep recursive network structure
Figure BDA0003266756520000072
And implements two-dimensional vectorization thereofCan assist in optimizing CT adaptive scanning parameters according to the high-dimensional sign information
S1202, extracting morphological characteristics of the three-dimensional positioning image and the target organ region by using the high-quality three-dimensional positioning image and the target organ region obtained from the organ region anatomical feature identification network as input of a high-dimensional image feature extraction network
Figure BDA0003266756520000073
Texture features
Figure BDA0003266756520000074
Different from the conventional adaptive scanning protocol which only sets the adaptive scanning protocol according to the two-dimensional positioning image, the self-adaptive scanning parameter of each exposure projection can be accurately estimated by the self-adaptive scanning protocol. Wherein the high-dimensional image feature extraction network may be constituted by a full convolution network.
S1203, personalized scanning parameter estimation, task-driven and patient-driven multi-objective optimization is performed, wherein the multi-objective optimization comprises a scanning range, tube current, tube voltage and the like, and the purpose is to reduce radiation damage of a patient and ensure quality of a reconstructed image and diagnosis precision.
Wherein, the optimization process of the personalized scanning parameter estimation in the step S1203 is solved through an objective equation
Figure BDA0003266756520000075
Wherein omegaAIs the adaptive scanning parameter to be solved; omegaRAdaptive reconstruction parameters to be solved; s represents an estimate of the local noise power spectrum, T represents an estimate of the local modulation transfer function,
Figure BDA0003266756520000076
is based on task-driven and patient-driven parameter estimation, and contains patient core information
Figure BDA0003266756520000077
And high-dimensional morphological characteristics of three-dimensional positioning image
Figure BDA0003266756520000078
Texture features
Figure BDA0003266756520000079
Namely, it is
Figure BDA00032667565200000710
Finally, optimizing the self-adaptive exposure parameter omega under each angle through an ADMM-Net networkAAnd an adaptive reconstruction parameter omegaR(ii) a The cascade network adopts a small batch random gradient descent method to carry out grading training on the network, fx，fy，fzThe position information of the image f at coordinate points x, y and z, and the convergence of the network training are analyzed in a preliminary experiment.
Wherein S2 includes:
s21, constructing an antagonistic active learning network for automatic identification of CT artifacts;
s22, constructing a confrontation learning network for automatically correcting the CT artifact;
s21, performing network research design based on active learning and antagonistic learning theory technology; the active learning network can assist artifact labeling in data, wherein the active learning network A is (C, Q, S, L and U), and C is a machine learning model and is used for CT artifact identification and identification; q is a query function, and a committee-based selection algorithm is adopted as a query strategy and is used for screening data containing large information amount from the unlabeled sample pool U; s is a supervisor, selects correct labels in the sample U, and trains a classifier and the next round of query by using the obtained new knowledge; finally, high identification and fine labeling of common artifacts of high-precision CT under less identification data are realized;
s22, constructing a confrontation learning network for CT artifact automatic correction, which comprises the following steps:
s2201, constructing an artifact automatic correction and confrontation learning network with artifact identification type indexes, and identifying different artifact types, such as metal artifacts, ring artifacts, windmill artifacts and the like:
s2202, according to the artifact identification result, marking and indexing each type of artifact;
s2203, automatically correcting by adopting different countermeasure networks according to different artifacts;
s2204, under the condition that aliasing exists in various artifacts or under the condition that different artifact representations are similar, according to the identification result and range in the artifact identification network, firstly, a feature extraction method is adopted to carry out feature analysis and extraction on the artifact part in the chord graph, and specific loss functions are constructed for different features so as to realize artifact automatic correction based on artifact structural feature difference;
s2105, for the artifacts meeting preset conditions, such as strong artifacts affecting accurate identification of anatomical structures or affecting standard values of CT images, such as metal artifacts, ring artifacts and the like, multi-scale directional field processing is performed on the CT projection data according to identification results, such as a multi-scale directional transformation field and a multi-scale contour transformation field, so as to extract multi-scale directional data information of artifact parts in the projection data, and a targeted countermeasure network cost function is designed according to the multi-scale data information, so that projection data after artifact correction are finally obtained.
Wherein the S3 includes the following steps:
s31, constructing a CT quantitative imaging semi-supervised incremental learning network; constructing a low-dose CT quantitative imaging supervised learning network:
for existing sample data of CT 'noisy image-target image', estimating internal features of the sample by using a 'projection domain-image domain' depth filtering back projection network, and using the estimated features as a supervised learning imaging network; the low-dose CT quantitative imaging supervised learning network comprises a fully-connected filter layer, a sine back-projection layer and a residual convolutional neural network, wherein the fully-connected filter layer is optimally designed based on a filter kernel of a filter back-projection algorithm, the sine back-projection layer corresponds to a back-projection operator in the filter back-projection algorithm, the residual convolutional neural network is used for further optimizing a reconstruction result, and a network cost function is designed to be 2 norm root mean square error and is recorded as:
Figure BDA0003266756520000081
wherein
Figure BDA0003266756520000082
Reconstructing the final predicted image of the network for CT, xH refThe CT image is a target CT image, N is the number of training samples, and theta is a parameter needing to be learned;
s32, constructing a semi-supervised incremental learning network: according to the project, firstly, noise-containing CT projection data acquired from an external source are input into FBP net trained in advance after passing through various deformation fields, then, the network reconstruction CT image result is subjected to self-adaptive weighted fusion to obtain a fused high-quality CT target image, further, the FBP net parameters are subjected to depth optimization again by using continuously-increased noise-containing CT projection data-target CT image data, and finally, semi-supervised quantitative imaging model building is completed;
s33, constructing a CT quantitative imaging unsupervised incremental learning network; and designing a CT quantitative imaging unsupervised incremental learning network based on the statistical characteristics of noisy CT projection data before Log transformation, such as composite Poisson-Gaussian noise distribution and the sparsity of the high-order derivative of the CT projection data. Based on a maximum posterior probability framework, the project designs an objective function in a deep learning network, the objective function comprises a data fidelity item based on a likelihood function and a prior information item based on piecewise linearity, the objective function can more accurately describe CT projection data and can carry out training optimization along a gradient direction, and the expression is as follows:
Figure BDA0003266756520000091
wherein, C1Representing noisy CT projection training data set, yiThe ith noisy CT projection training data is obtained, N represents the number of data sets, G represents the number of X-ray photons satisfying the composite Poisson statistical characteristics, I represents the number of photons received by a detector, and I represents the number of the X-ray photons received by the detectoriRepresenting the number of photons received by the i-th detector, GiRepresents IiApproximately middle value of, Gj| A Denotes the multiplication value of G in the j-th projection, I0jRepresenting the number of incident photons in the j projection, ε represents the electronic noise, and obeys a mean of 0 and a variance of σ2Gaussian distribution of (D)2Representing a second order difference operator, fθRepresenting unsupervised learning network mapping, wherein theta is a parameter needing to be learned in the network mapping, a high-quality CT image is obtained through reconstruction by utilizing a filtering back-projection algorithm, and fθ(yi)jRepresents y for the ith measurementiUnsupervised learning network mapping on pixel j;
wherein the S4 includes the following steps: s41, constructing a vascular qualitative identification network based on a video depth of field migration learning technology; the cardiovascular perfusion CT imaging data comprises a plurality of image data of continuous time frames, the data has high time continuity and spatial correlation, and the characteristics also exist in the natural scene video data. Compared with the perfusion CT image which is difficult to directly train the network due to the lack of the labeling data, the natural video data has a fine label and is easy to train the network. Therefore, based on 4, the study is intended to adopt a transfer learning technique to use a video depth of field analysis network for quantitative assessment of cardiovascular state
S4101, constructing a dynamic depth of field/area identification network: the training set is defined as s { (X)i，Yi) I ═ 1,2,3, …, n }, where X isiInputting images for a three-dimensional sequence of ith time frame, YiFor ith time frame label data (such as video depth of field image/CT cardiovascular region image), i.e. two-dimensional probability hotspot map (Hot map), different regions can be quantitatively distinguished,
Figure BDA0003266756520000092
for the network output result, the cost function is:
Figure BDA0003266756520000101
s4102, constructing a video depth of field migration learning network; the method comprises the steps of firstly inputting natural video data pairs into the recognition network, performing network pre-training to obtain a depth-of-field network, then formulating a transfer learning strategy, and performing transfer learning training on the depth-of-field network by using cardiovascular perfusion CT data to obtain a blood vessel region quantitative evaluation network, wherein partial parameters of the depth-of-field network in the evaluation network are frozen, only parameters of other layers are trained, high-dimensional robust characteristics of cardiovascular and cerebrovascular CT sequence images are deeply mined, cardiovascular qualitative recognition S42 is realized, and a blood vessel quantitative recognition network based on medical multi-modal transfer imaging is constructed;
the method comprises the following steps of S4201, constructing a blood vessel quantitative identification network: the training set is defined as S { (X)i，Yi) I ═ 1,2,3, …, n }, where X isiFor the ith input image, YiE {0,1} is the label for the ith input image, where Yi1 is defined as an abnormal image, Yi0 is defined as a non-anomalous image. Simultaneous definition of
Figure BDA0003266756520000102
Is the probability of the kth pixel of the ith input image, where k ═ 1,2,3, …, | Xi|}，|Xi| represents XiTotal number of pixels. If it is not
Figure BDA0003266756520000103
For a probability map of image level prediction, then
Figure BDA0003266756520000104
Can pass through all of the pixel levels
Figure BDA0003266756520000105
And calculating to obtain the cost function:
Figure BDA0003266756520000106
wherein I (-) is an indicator function,
Figure BDA0003266756520000107
calculated by the Soft max Function, r is a control parameter; l isMILRepresenting a cost function;
s4202, constructing a multi-mode migration learning network: for cerebral perfusion, firstly inputting natural image data pairs into the recognition network, performing network pre-training, and then performing transfer learning training on the recognition network after natural image training by using cerebral perfusion MRI data pairs to obtain a blood vessel region quantitative evaluation network; finally, a transfer learning strategy is formulated, brain perfusion CT projection data are firstly reconstructed by brain perfusion CT data, then a blood vessel region quantitative evaluation network is subjected to transfer learning, brain MRI is fused, the relevance of an internal anatomical structure of a CT image and the diversity of image information of different modes are optimized, and blood vessel region quantitative evaluation network parameters are optimized, wherein the transfer learning strategy can be selected according to the size of a brain perfusion CT data set and the number of parameters: (1) freezing the parameters of the first n layers, namely not changing the values of the n layers when training the cerebral perfusion CT blood vessel region quantitative evaluation network; (2) the first n layers of parameters are not frozen, but the values of the network parameters are continuously adjusted, namely two strategies of fine adjustment are used for carrying out transfer learning; finally, quantitatively identifying a cerebral stroke region of the cerebral perfusion CT image; similarly, for myocardial perfusion data, firstly inputting the natural image data pair into the identification network, performing network pre-training, and then performing transfer learning training on the identification network trained by the natural image by using myocardial perfusion PET data to obtain a blood vessel region quantitative evaluation network; finally, a transfer learning strategy is formulated, the myocardial perfusion CT data is used for reconstructing brain perfusion CT projection data, then transfer learning is carried out on a blood vessel region quantitative evaluation network, myocardial PET is fused, the correlation of an internal anatomical structure and the diversity of different modal image information in a CT image are optimized, and the parameters of the blood vessel region quantitative evaluation network are optimized, wherein two strategies similar to the former strategies are selected for parameter training according to the size of a myocardial perfusion CT data set and the number of parameters; finally, myocardial perfusion CT image myocardial infarction area is identified quantitatively.
Wherein the S5 includes the following steps: s51, constructing a strategy network for intelligent monitoring and sequencing of emergency patients; a strategy network for intelligent monitoring and sequencing of emergency patients is provided, and a network architecture can be based on a plurality of deep convolutional network technologies, such as a high-dimensional convolutional neural network, and aims to screen CT projection data of cardiovascular critical and serious patients. The high-dimensional convolutional neural network finally determines whether the image contains the acute cardiovascular disease, so that the image is diagnosed according to the urgent and serious priority sequence instead of the original acquisition sequence, limited medical resources of a hospital are optimized, and the purpose of quickly diagnosing and treating the urgent and serious priority patient is achieved;
s52, constructing a monitoring sequencing result obtained by a depth network generated by a vascular disease auxiliary report, preferentially judging the CT projection data of patients with acute and severe cardiovascular diseases according to the optimal sequence of 'acute and severe priority', and generating a disease auxiliary report by using a recurrent neural network according to a cardiovascular disease sample library.
For CT projection data, a description image maximum probability expression form is constructed:
Figure BDA0003266756520000111
where θ is the model parameter, I is the CT projection data, and S represents the correct transcription, and there is no length constraint. Chain rules are typically applied to simulate S0，…,SNN is the length of the particular example.
Figure BDA0003266756520000112
When the model is trained, (S, I) forms a training example pair, and a random gradient descent algorithm is used for optimizing the sum log p (S | I) of logarithmic probabilities.
Wherein, the method also comprises the construction probability p of the long-term and short-term memory network LSTM for short; LSTM is designed in a web-expanded form, and a copy of the LSTM memory can be created for the image and each sentence, so that all LSTM share the same parameters in time, and all repeated joins are converted to feed-forward joins in an expanded version, then the expansion process is as follows:
x-1＝CNN(I)
xt＝WeSt，t∈{0，…，N-1}
pt+1＝LSTM(xt)，t∈{0，…,N-1}
where we represent each word as a single heat vector StThe size of which is equal to the size of the dictionary. Note that we use two special words S0And SNRepresenting the beginning and end of a sentence, respectively. Both images and words are mapped to the same space, the images being viewed by use of the viewFeel CNN and embed W by worde. Image I is input only once at t-1 to inform LSTM of the image content. Our penalty function is the sum of the negative log likelihood of the correct word at each step and can be expressed as follows:
Figure BDA0003266756520000121
by optimizing all parameters of the LSTM, the top level of the image embedder CNN and the word embedding WeTo minimize the loss function of the above equation; the network finally inputs the cardiovascular disease classification recognition result for assisting the subsequent medical diagnosis.
The invention has the following effective effects:
1. the invention constructs a strategy network for intelligent monitoring and sequencing of emergency patients, and the network architecture can be based on various deep convolutional network technologies, such as a high-dimensional convolutional neural network, and aims to screen CT projection data of cardiovascular critical and severe patients; the high-dimensional convolutional neural network finally determines whether the image contains the acute cardiovascular disease, so that the image is diagnosed according to the urgent-heavy priority sequence instead of the original acquisition sequence, limited medical resources of a hospital are optimized, and the purpose of quickly diagnosing and treating the urgent-heavy priority patient is achieved.
2. Aiming at three problems faced by CT imaging in clinical application, the invention combines an Artificial Intelligence (AI) technology, brings each problem into a deep learning network, sequentially progresses each sub-network simultaneously, integrates the sub-networks into a whole, has definite target, and constructs a new framework of CT imaging and intelligent analysis facing multi-task scanning protocol self-adaptive optimization, micro-radiation dose CT quantitative imaging and CT imaging intelligent auxiliary diagnosis.
3. The new framework of the CT-AI image system constructed by the invention is clinical task oriented and has more clinical practical significance, and the framework deeply analyzes the potential and the demand of each component of the CT image system based on the CT image scanning-reconstruction-auxiliary analysis integrated system, thereby being convenient for the subsequent reconstruction and upgrade of the CT image system.
4. The invention effectively solves the problems that the current CT CAD x technology is mostly based on CT images, and omits richer and more detailed medical information contained in projection data; according to the integrated framework, the prediction error of the output end can be finally transmitted to the CT projection data of the input end, which is equivalent to positioning and identifying cardiovascular abnormalities by utilizing the projection information under multiple angles from the projection directly, namely predicting from the original data.
While embodiments of the invention have been shown and described, it will be understood by those skilled in the art that: various changes, modifications, substitutions and alterations can be made to the embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims (8)

1. An intelligent patient monitoring and sequencing method is characterized by comprising the following steps:
s1, carrying out self-adaptive optimization on the CT scanning protocol based on the depth recognition model;
s2, identifying and automatically correcting the CT artifact based on a countermeasure technology;
s3, CT quantitative imaging is carried out based on the incremental learning technology;
s4, carrying out intelligent analysis on the CT image based on the transfer learning technology;
and S5, making an auxiliary decision on the CT image.
2. The method for sequencing intelligent patient monitoring of claim 1, wherein said S1 comprises:
s11, constructing an organ region anatomical feature identification network;
and S12, constructing a CT personalized scanning protocol task guide network.
3. The method for sequencing intelligent patient monitoring of claim 2, wherein said S11 comprises:
s1101, constructing a three-dimensional positioning image imaging sub-network:
the depth residual error learning sub-network comprises a noise estimation block, an artifact estimation block and an image filtering block, wherein the noise estimation block, the artifact estimation block and the image filtering block are respectively processed through a residual error learning network, a multi-scale wavelet and a convolutional neural network;
carrying out artifact estimation based on a processing flow of multi-scale wavelet transform, further improving the image quality of the microdose image by an iteration method, and finally obtaining a three-dimensional positioning image;
s1102, constructing an organ region identification subnetwork:
taking the recovered micro-dose three-dimensional positioning image output by the deep residual learning network as the input of the identification network;
wherein the organ region identification network comprises an encoding process and a decoding process;
wherein, the coding process adopts the structure of residual block, and the decoding process adopts the full convolution network; the training set is defined as X, where XiFor the ith input image, YiLabels for the ith input image, wherein I is defined as a different reconstruction target region; simultaneous definition of
Figure FDA0003266756510000011
Probability of kth pixel of ith input image;
Figure FDA0003266756510000012
the probability map for image level prediction can be obtained by all computations at pixel level, and the cost function is:
Figure FDA0003266756510000013
the three-dimensional positioning image imaging sub-network and the organ region identification sub-network are cascaded to obtain a final organ region anatomical feature identification network, L represents a cost function, and r represents a control parameter.
4. The method for sequencing intelligent patient monitoring of claim 3, wherein said S12 comprises:
s1201, constructing a depth recognition network, wherein the constructed depth recognition model is based on a bidirectional depth recursive network;
s1202, extracting morphological characteristics of the three-dimensional positioning image and the target organ region by using the three-dimensional positioning image and the target organ region obtained from the organ region anatomical feature identification network as input of a high-dimensional image feature extraction network
Figure FDA0003266756510000021
Texture features
Figure FDA0003266756510000022
S1203, personalized scanning parameter estimation, task-driven, patient-driven multi-objective optimization,
wherein, the optimization process of the personalized scanning parameter estimation in the step S1203 is solved through an objective equation
Figure FDA0003266756510000023
Wherein omegaAIs the adaptive scanning parameter to be solved; omegaRAdaptive reconstruction parameters to be solved; s represents an estimate of the local noise power spectrum, T represents an estimate of the local modulation transfer function,
Figure FDA0003266756510000024
is based on task-driven and patient-driven parameter estimation, and contains patient core information
Figure FDA0003266756510000025
And high-dimensional morphological characteristics of three-dimensional positioning image
Figure FDA0003266756510000026
Texture features
Figure FDA0003266756510000027
Namely, it is
Figure FDA0003266756510000028
Finally, optimizing the self-adaptive exposure parameter omega under each angle through an ADMM-Net networkAAnd an adaptive reconstruction parameter omegaR(ii) a The cascade network adopts a small batch random gradient descent method to carry out grading training on the network, fx，fy，fzIndicating the position information of the image f at coordinate points x, y, z.
5. The method for sequencing intelligent patient monitoring of claim 1, wherein said S2 comprises:
s21, constructing an antagonistic active learning network for automatic CT artifact identification;
s22, constructing a confrontation learning network for automatically correcting the CT artifact;
the S21 includes:
performing network research design based on active learning and antagonistic learning theory technology; the active learning network can assist artifact labeling in data, wherein the active learning network A is (C, Q, S, L and U), and C is a machine learning model and is used for CT artifact identification and identification; q is a query function, and a committee-based selection algorithm is adopted as a query strategy and is used for screening data containing large information amount from the unlabeled sample pool U; s is a supervisor, selects correct labels in the sample U, and trains a classifier and the next round of query by using the obtained new knowledge;
the S22 includes:
s2201, constructing an artifact automatic correction confrontation learning network with artifact identification type indexes, and identifying artifact types:
s2202, according to the artifact identification result, marking and indexing each type of artifact;
s2203, aiming at different artifacts, adopting different countermeasure networks to automatically correct;
s2204, according to the identification result and range in the artifact identification network, a feature extraction method is adopted to perform feature analysis and extraction on the artifact part in the chord graph, and specific loss functions are constructed for different features so as to realize artifact automatic correction based on artifact structural feature difference;
s2205, for the artifacts meeting the preset conditions, performing multi-scale directional field processing on the CT projection data according to the identification result, extracting multi-scale directional data information of the artifact part in the projection data, and setting a countermeasure network cost function according to the multi-scale directional data information to obtain the projection data after artifact correction.
6. The method for sequencing intelligent patient monitoring of claim 1, wherein said S3 comprises:
s31, constructing a CT quantitative imaging semi-supervised incremental learning network; constructing a low-dose CT quantitative imaging supervised learning network:
the low-dose CT quantitative imaging supervised learning network comprises a fully-connected filter layer, a sine back-projection layer and a residual convolutional neural network, wherein the fully-connected filter layer is optimally designed based on a filter kernel of a filter back-projection algorithm, the sine back-projection layer corresponds to a back-projection operator in the filter back-projection algorithm, the residual convolutional neural network is used for further optimizing a reconstruction result, and a network cost function is designed to be 2 norm root mean square error and is recorded as:
Figure FDA0003266756510000031
wherein
Figure FDA0003266756510000032
Reconstructing the final predicted image of the network for CT, xH refThe CT image is a target CT image, N is the number of training samples, and theta is a parameter needing to be learned;
s32, constructing a semi-supervised incremental learning network: inputting noise-containing CT projection data acquired from an external source into a pre-trained FBPnet after passing through various deformation fields, performing self-adaptive weighted fusion on CT image results to obtain a fused high-quality CT target image, and performing depth optimization on FBPnet parameters again to obtain a semi-supervised quantitative imaging model;
s33, constructing a CT quantitative imaging unsupervised incremental learning network; designing a CT quantitative imaging unsupervised incremental learning network based on the statistical characteristics of noisy CT projection data before Log transformation and the sparsity of a high-order derivative of the CT projection data; an objective function in a deep learning network is constructed based on a maximum posterior probability framework, the objective function comprises a data fidelity item based on a likelihood function and a prior information item based on piecewise linearity, the objective function can more accurately describe CT projection data and can carry out training optimization along a gradient direction, and the expression is as follows:
Figure FDA0003266756510000033
wherein, C1Representing noisy CT projection training data set, yiThe ith noisy CT projection training data is obtained, N represents the number of data sets, G represents the number of X-ray photons satisfying the composite Poisson statistical characteristics, I represents the number of photons received by a detector, and I represents the number of the X-ray photons received by the detectoriRepresenting the number of photons received by the i-th detector, GiRepresents IiApproximately middle value of, Gj| A Denotes the multiplication value of G in the j-th projection, I0jRepresenting the number of incident photons in the j projection, ε represents the electronic noise, and obeys a mean of 0 and a variance of σ2Gaussian distribution of (D)2Representing a second order difference operator, fθRepresenting unsupervised learning network mapping, wherein theta is a parameter needing to be learned in the network mapping, a high-quality CT image is obtained through reconstruction by utilizing a filtering back-projection algorithm, and fθ(yi)jRepresenting the training data y for the ith noisy CT projectioniUnsupervised learning network mapping on pixel j.
7. The method for sequencing intelligent patient monitoring of claim 1, wherein said S4 comprises:
s41, constructing a vascular qualitative identification network based on the video depth migration learning technology, including:
s4101, constructing a dynamic depth of field area identification network: the training set is defined as s { (X)i，Yi) 1,2,3, aiInputting images for a three-dimensional sequence of ith time frame, YiFor the ith time frame of tag data,
Figure FDA0003266756510000041
for the network output result, the cost function is:
Figure FDA0003266756510000042
s4102, constructing a video depth of field migration learning network;
s42, constructing a blood vessel quantitative recognition network based on medical multi-modal migration imaging, comprising the following steps:
s4201, constructing a blood vessel quantitative identification network: the training set is defined as S { (X)i，Yi) 1,2,3, aiFor the ith input image, YiE {0,1} is the label for the ith input image, where Yi1 is defined as an abnormal image, Yi0 is defined as a non-anomalous image;
definition of
Figure FDA0003266756510000043
Is the probability of the kth pixel of the ith input image, where k ═ 1,2,3i|}，|Xi| represents XiTotal number of pixels, if
Figure FDA0003266756510000044
For a probability map of image level prediction, then
Figure FDA0003266756510000045
All through the pixel level
Figure FDA0003266756510000046
And calculating to obtain the cost function:
Figure FDA0003266756510000047
wherein I (-) is an indicator function,
Figure FDA0003266756510000048
calculated by the Soft max Function, r is a control parameter; l isMILRepresenting a cost function;
s4202, constructing the multi-mode migration learning network.
8. The method for sequencing intelligent patient monitoring of claim 1, wherein said S5 comprises:
s51, constructing a strategy network for intelligent monitoring and sequencing of emergency patients;
and S52, constructing a monitoring sequencing result obtained by a depth network generated by the auxiliary report facing the vascular diseases.
CN202111089390.0A 2021-09-16 2021-09-16 Intelligent patient monitoring and sequencing method Pending CN113781461A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111089390.0A CN113781461A (en) 2021-09-16 2021-09-16 Intelligent patient monitoring and sequencing method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111089390.0A CN113781461A (en) 2021-09-16 2021-09-16 Intelligent patient monitoring and sequencing method

Publications (1)

Publication Number Publication Date
CN113781461A true CN113781461A (en) 2021-12-10

Family

ID=78851677

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111089390.0A Pending CN113781461A (en) 2021-09-16 2021-09-16 Intelligent patient monitoring and sequencing method

Country Status (1)

Country Link
CN (1) CN113781461A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115345803A (en) * 2022-10-19 2022-11-15 合肥视微科技有限公司 Residual error network-based annular artifact correction method in CT (computed tomography) tomography
CN115836855A (en) * 2023-02-22 2023-03-24 首都医科大学附属北京朝阳医院 Mobile magnetic resonance equipment imaging method and device, storage medium and terminal

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107481297A (en) * 2017-08-31 2017-12-15 南方医科大学 A kind of CT image rebuilding methods based on convolutional neural networks
US20170372193A1 (en) * 2016-06-23 2017-12-28 Siemens Healthcare Gmbh Image Correction Using A Deep Generative Machine-Learning Model
CN109035169A (en) * 2018-07-19 2018-12-18 西安交通大学 A kind of unsupervised/semi-supervised CT image reconstruction depth network training method
CN111968110A (en) * 2020-09-02 2020-11-20 广州海兆印丰信息科技有限公司 CT imaging method, device, storage medium and computer equipment
CN111968111A (en) * 2020-09-02 2020-11-20 广州海兆印丰信息科技有限公司 Method and device for identifying visceral organs or artifacts of CT (computed tomography) image
CN111968167A (en) * 2020-09-02 2020-11-20 广州海兆印丰信息科技有限公司 Image processing method and device for CT three-dimensional positioning image and computer equipment
CN111968108A (en) * 2020-09-02 2020-11-20 广州海兆印丰信息科技有限公司 CT intelligent imaging method, device and system based on intelligent scanning protocol
CN112037146A (en) * 2020-09-02 2020-12-04 广州海兆印丰信息科技有限公司 Medical image artifact automatic correction method and device and computer equipment
CN112348792A (en) * 2020-11-04 2021-02-09 广东工业大学 X-ray chest radiography image classification method based on small sample learning and self-supervision learning
CN112508063A (en) * 2020-11-23 2021-03-16 刘勇志 Medical image classification method based on incremental learning

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170372193A1 (en) * 2016-06-23 2017-12-28 Siemens Healthcare Gmbh Image Correction Using A Deep Generative Machine-Learning Model
CN107481297A (en) * 2017-08-31 2017-12-15 南方医科大学 A kind of CT image rebuilding methods based on convolutional neural networks
CN109035169A (en) * 2018-07-19 2018-12-18 西安交通大学 A kind of unsupervised/semi-supervised CT image reconstruction depth network training method
CN111968110A (en) * 2020-09-02 2020-11-20 广州海兆印丰信息科技有限公司 CT imaging method, device, storage medium and computer equipment
CN111968111A (en) * 2020-09-02 2020-11-20 广州海兆印丰信息科技有限公司 Method and device for identifying visceral organs or artifacts of CT (computed tomography) image
CN111968167A (en) * 2020-09-02 2020-11-20 广州海兆印丰信息科技有限公司 Image processing method and device for CT three-dimensional positioning image and computer equipment
CN111968108A (en) * 2020-09-02 2020-11-20 广州海兆印丰信息科技有限公司 CT intelligent imaging method, device and system based on intelligent scanning protocol
CN112037146A (en) * 2020-09-02 2020-12-04 广州海兆印丰信息科技有限公司 Medical image artifact automatic correction method and device and computer equipment
CN112348792A (en) * 2020-11-04 2021-02-09 广东工业大学 X-ray chest radiography image classification method based on small sample learning and self-supervision learning
CN112508063A (en) * 2020-11-23 2021-03-16 刘勇志 Medical image classification method based on incremental learning

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
PRATYUSH KUMAR.ET AL: "Example Mining for Incremental Learning in Medical Imaging", 《ARXIV》, 31 December 2018 (2018-12-31) *
施俊等: "深度学习在医学影像中的应用综述", 《中国图象图形学报》, vol. 25, no. 10, 31 December 2020 (2020-12-31) *
朱森华等: "人工智能技术在医学影像产业的应用与思考", 《人工智能》, no. 03, 10 June 2020 (2020-06-10) *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115345803A (en) * 2022-10-19 2022-11-15 合肥视微科技有限公司 Residual error network-based annular artifact correction method in CT (computed tomography) tomography
CN115836855A (en) * 2023-02-22 2023-03-24 首都医科大学附属北京朝阳医院 Mobile magnetic resonance equipment imaging method and device, storage medium and terminal

Similar Documents

Publication Publication Date Title
Liang et al. MCFNet: Multi-layer concatenation fusion network for medical images fusion
CN110503654B (en) Medical image segmentation method and system based on generation countermeasure network and electronic equipment
CN111476292B (en) Small sample element learning training method for medical image classification processing artificial intelligence
CN110992351B (en) sMRI image classification method and device based on multi-input convolution neural network
CN109544518B (en) Method and system applied to bone maturity assessment
CN110930421A (en) Segmentation method for CBCT (Cone Beam computed tomography) tooth image
CN113314205B (en) Efficient medical image labeling and learning system
CN111583285B (en) Liver image semantic segmentation method based on edge attention strategy
CN111667483B (en) Training method of segmentation model of multi-modal image, image processing method and device
CN113781461A (en) Intelligent patient monitoring and sequencing method
CN111667027B (en) Multi-modal image segmentation model training method, image processing method and device
CN113688862B (en) Brain image classification method based on semi-supervised federal learning and terminal equipment
CN111798439A (en) Medical image quality interpretation method and system for online and offline fusion and storage medium
CN114897914A (en) Semi-supervised CT image segmentation method based on confrontation training
CN112529063A (en) Depth domain adaptive classification method suitable for Parkinson voice data set
CN110477907A (en) A kind of intelligence assists in identifying the method for epilepsy outbreak
La Rosa A deep learning approach to bone segmentation in CT scans
CN111784713A (en) Attention mechanism-introduced U-shaped heart segmentation method
CN116563549A (en) Magnetic resonance image heart segmentation method based on coarse-granularity weak annotation
Zhao et al. MPSHT: multiple progressive sampling hybrid model multi-organ segmentation
CN117274599A (en) Brain magnetic resonance segmentation method and system based on combined double-task self-encoder
CN112750131A (en) Pelvis nuclear magnetic resonance image musculoskeletal segmentation method based on scale and sequence relation
CN116433679A (en) Inner ear labyrinth multi-level labeling pseudo tag generation and segmentation method based on spatial position structure priori
CN110613445A (en) DWNN framework-based electrocardiosignal identification method
CN115798709A (en) Alzheimer disease classification device and method based on multitask graph isomorphic network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination