CN108324244A - The construction method and system of automatic augmentation training sample for the diagnosis of AI+MRI Image-aideds - Google Patents

The construction method and system of automatic augmentation training sample for the diagnosis of AI+MRI Image-aideds Download PDF

Info

Publication number
CN108324244A
CN108324244A CN201810005912.6A CN201810005912A CN108324244A CN 108324244 A CN108324244 A CN 108324244A CN 201810005912 A CN201810005912 A CN 201810005912A CN 108324244 A CN108324244 A CN 108324244A
Authority
CN
China
Prior art keywords
image
information
voxel
mri
obtaining
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
CN201810005912.6A
Other languages
Chinese (zh)
Inventor
汪红志
杜小霞
杨丽琴
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.)
East China Normal University
Original Assignee
East China Normal University
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 East China Normal University filed Critical East China Normal University
Priority to CN201810005912.6A priority Critical patent/CN108324244A/en
Publication of CN108324244A publication Critical patent/CN108324244A/en
Pending legal-status Critical Current

Links

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/05—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves 
    • A61B5/055—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves  involving electronic [EMR] or nuclear [NMR] magnetic resonance, e.g. magnetic resonance imaging

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Public Health (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Veterinary Medicine (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Pathology (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • Molecular Biology (AREA)
  • Surgery (AREA)
  • Psychiatry (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Physiology (AREA)
  • High Energy & Nuclear Physics (AREA)
  • Signal Processing (AREA)
  • Radiology & Medical Imaging (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)

Abstract

The invention discloses a kind of construction methods and its system of the automatic augmentation training sample for the diagnosis of AI+MRI Image-aideds, including：By the quantitative imaging technique of clinical MRI equipment, the site tissue of the typical disease of diseased individuals and the physical parameter of normal individual corresponding site tissue voxel, the quantitative information image library of chemical parameters and physiological parameter are obtained；Using quantitative information image library as input, by the mathematical model of imaging method, the MRI standard pictures and typical disease image for including the various weight features of cross sections is calculated；The MRI standard pictures and typical disease image feeding AI algorithms including the various weight features of cross sections that the step S01 quantitative information image libraries obtained or step S02 are obtained, as training sample.The present invention can provide a variety of training samples, solve the problems, such as that the inefficient and sample size of the artificial mark sample training in MRI image artificial intelligence auxiliary diagnosis is few, and can getting parms according to diagnostic image to be analyzed, targetedly sample training is provided.

Description

For AI+MRI Image-aideds diagnosis automatic augmentation training sample construction method and System
Technical field
The invention belongs to AI medical image aided diagnosis techniques fields, and AI+MRI Image-aideds are used for more particularly to one kind The construction method and system of the automatic augmentation training sample of diagnosis.
Background technology
Modern medicine has been increasingly dependent on various medical inspection technologies, especially Medical Imaging Technology.Nuclear magnetic resonance Imaging is that clinical advantage is most apparent in five big medical image patterns, the maximum pattern of potentiality, and no ionising radiation, soft tissue are differentiated Height can carry out the advantages that functional inspection and quantitative imaging so that it is more and more extensive in clinical application, is new technology newly side The key areas of method and new opplication research and development.Clinical installation amount is also quickly increasing, and has the tendency that past second-grade hospital is universal.Largely Nuclear-magnetism installation amount, output by be magnanimity nuclear magnetic resonance image.Different from the one-parameter imaging of X-ray (containing CT), nuclear-magnetism is The more weight images of multi-parameter, therefore the routine inspection at a position, the image type of generation are at least 4 times of CT images.For Difficult and complicated illness inspection, image output are then more.The image of magnanimity proposes sternness for the quality and quantity of nuclear-magnetism diagnostician Challenge.The culture of outstanding nuclear-magnetism diagnostician is a very long process, this is auxiliary to introducing artificial intelligence technology development magnetic resonance imaging It helps diagnosis to have huge potential demand, entirely autonomous diagnostic imaging is realized even if cannot settle at one go, for the nuclear-magnetism of magnanimity Image can carry out artificial intelligence anomalous identification, realize primary dcreening operation and lesion prompt, for alleviating the operating pressure of diagnostician, It is also to have important value that multi-angle, which provides diagnostic comments,.
Mathematical model introducing clinical medicine is proposed computer-aided diagnosis by nineteen fifty-nine, American scholar Ledley etc. for the first time Mathematical model, and diagnosed one group of cases of lung cancer, started the beginning of computer-aided diagnosis；Start since then, a large amount of science Family has carried out exploration in the field, but until being still within computer aided detection (computer-aided at present Detection, CAD) stage, it acts on and is to aid in radiologist's raising diagnostic accuracy and image, disease are explained Consistency.In other words, the output result of computer may only be used as a kind of supplementary means, and entirely autonomous cannot be examined by its It is disconnected.Why CAD computer aided detections can improve the diagnostic accuracy of doctor, the reason is that, in traditional diagnosis method, The diagnosis of radiologist is entirely subjective judgement process；It can thus be limited by diagnostician's experience and know-how and shadow It rings；Secondly, it is easy to omit certain subtle changes when diagnosis；Again, the diagosis difference between different physicians and between same doctor Influence.And computer objectively judges there is big advantage for correcting these mistakes and deficiency.
In recent years, the artificial intelligence based on neural network (AI) is fast-developing, it is to imitate National People's Congress's brain neuron work original A kind of mathematical processing methods of reason.Since it has the abilities such as ability of self-teaching, memory capability, predicted events development, have The feature of auxiliary diagnosis has more superior performance compared to traditional method (probabilistic method, mathematical model etc.).Therefore artificial intelligence It can be recombined with the demand of medical imaging diagnosis, expedite the emergence of the new hot spot of nearest AI+ medical images.
But either AI diagnostic imagings burning hot in the recent period still start the computer for Medical Imaging twenty years ago Auxiliary diagnosis (computer-aided diagnosis, CAD), mainly all concentrates on X ray image, CT images or ultrasonic image Scope, specific research is confined to mammary gland mostly and chest pleurotome section venereal disease becomes, in CT virtual colons scope (CTC), liver diseases CT is diagnosed.Thus, the CAD of mammary gland and Lung neoplasm lesion research can substantially represent current CAD in Medical Imaging most High level and present situation.
CAD the or AI+ images in MRI image field only have the research paper of pancreas MRI image auxiliary diagnosis at present, for The research of the MRI auxiliary diagnosis of inside the cranium position disease especially tumour etc. is not seen in document.The country organizes extensively in the recent period Artificial intelligence PK diagnosticians diagnostic imaging match, be also all to concentrate on DR images and CT image domains.
Why AI+ medical images, but rarely have research in the combination of magnetic resonance imaging imageThe reason is that MRI What the diversity of image determined.
AI application fields are divided from the source abundance of sample size, can be divided into large sample size field (as known based on face Not, character recognition, the automatic Pilot of speech recognition, security protection, athletic game etc.) and sample this field (such as needs profession mark The Medical Imaging of sample).Medical image+AI be one need profession mark small sample amount application field, i.e., there is The limited problem of data set sample number.And inside Medical Imaging, compare X-ray and CT images, a variety of figures of MR images As type becomes the application aspect of smaller sample size.For the AI application fields of small sample amount, in order to enhance training result Accuracy, often use data augmentation method carry out sample size artificial expansion.
Belong to one-parameter imaging such as DR images and CT images based on the image that X-ray generates, it is main to reflect the close of tissue Spend difference；Therefore all images belong to a kind of image, that is to say, that its standard picture only has one kind；One of AI images is important Be characterized in needing largely inputting the normal pictures and abnormal image by mark so that AI carry out after self-teaching specifically classification and The ability of judgement, to provide judging result.The image type for inputting study is more, and the adaptability and accuracy of AI is higher. But nuclear-magnetism image is multiparameter imaging, other than conventional multiple weight image, also the images such as perfusion, disperse, blood flow and Various quantitative images (relaxation time is quantitative, and chemical parameters are quantitative, and physiological parameter is quantitative etc., CEST, magnetic susceptibility sensitivity etc.), even if It is conventional imaging, the sequential parameter of different manufacturers is different, and image is also discrepant.That is, the same position, light are normal Standard picture just have too much, it is just more more for the performance of disease.For the angle of artificial intelligence, it is difficult to by have through The doctor tested completes the input of the standard and abnormal image of huge number by mark.The trained image pattern of AI systems is not Enough, adaptability and accuracy are naturally just very low.This is also the less basic reasons for being able to study of AI+MRI.
It is well known that the most key problem of AI artificial intelligence is the defeated of the original image of a large amount of standard or mark Enter.Specific bottom AI hardware (GPU) is at present mainly by the absolute monopoly of Nvida companies.The convolutional network training algorithm of middle layer And system, also there are many companies to be proposed software (such as Nvida companies release visualization training software digits), also there are many Company and research unit develop convolutional Neural algorithm (such as LeNet, AlexNet, GoogleLeNet etc.).Therefore AI Using application is concentrated mainly on, i.e., by the combination of AI and specific research field.Specific research field and AI combination, just It is embodied in the sample data set of big good standard or mark.Data set type and the sample number for inputting study are more, and AI is by instruction Experienced level is more, just more acurrate to the analytical judgment result of certain specific image.The characteristics of MRI image, determines its standard Image type is more, and the performance of same disease is also just more, therefore the bottleneck problem of AI+MRI diagnostic imagings is how reality The now input of standard picture rapidly and efficiently and abnormal mark image.
AI is directed to the problem of similar small sample number at present, is by image simply rotate, scale, color turn It changes equal image processing means and obtains new image as more samples to realize data augmentation.This treating method can solve Certainly the problem of sample number, but these sample numbers substantially still fall within the same sample, can not solve some sample diversity The problem of (such as nuclear-magnetism image).
The project is intended passing through quantitative imaging technique+virtual image technology from the learning data entrance for solving AI+ images Come the image for realizing standard and the abnormal rapidly and efficiently automatic augmentation for marking image, established for the specific study analysis judgement of follow-up AI Fixed solid data basis.
Invention content
The present invention for AI technologies apply the training sample that can be encountered with nuclear magnetic resonance image auxiliary diagnosis can not completely by Doctor marks the bottleneck problem that the method for completion is realized manually, it is proposed that one kind is based on structure standard quantitative message sample and weight The automatic augmentation construction method of image pattern.This method can solve the poor efficiency of mark realization great amount of samples manually and can not The deficiency of whole sample databases is provided, while the subjectivity error and minimal disease in the manual annotation process of doctor can also be overcome Spill tag problem.
The present invention proposes a kind of construction method of the automatic augmentation training sample for the diagnosis of AI+MRI Image-aideds, packet Include following steps：
S01：It is obtained respectively by the quantitative imaging technique of clinical MRI equipment for normal human or exemplary disease site The physics of the site tissue voxel, chemistry, physiological parameter information quantitative image library；
S02：It is input with the quantitative information image library that the first step obtains, by the mathematical model of imaging method, calculates To the MRI standard pictures and typical disease image of the various weight features of each section；
S03：The a variety of MRI standard pictures and typical disease image that second step is generated are sent into AI algorithms, as a large amount of instructions Practice sample.
It is described in the construction method of automatic augmentation training sample proposed by the present invention for the diagnosis of AI+MRI Image-aideds In step S01, the various quantitative imaging techniques of applying clinical MRI machine quantify normal human and typical disease human body Imaging obtains the physical parameters of all voxels in the position, chemical parameters, physiological parameter value set, i.e. each parameter database of human body.
In the construction method of automatic augmentation training sample proposed by the present invention for the diagnosis of AI+MRI Image-aideds, according to Quantitative imaging technique described in step S01, including and it is not limited to following technology：
1) T1 mapping technologies obtain the T1 information of each voxel；
2) T2 mapping technologies obtain the T2 information of each voxel；
3) T2* mapping technologies obtain the T2* information of each voxel；
4) QSM technologies obtain the macroscopic magnetization rate information of each voxel；
5) PDFF mapping technologies obtain the proton density fat content information of each voxel；
6) MT (magnetization transfer) technology obtains each voxel macromolecular content information
7) CEST (chemical exchange saturation transfer) technology obtains protein, content of peptides Information；
8) MRS technologies obtain each voxel chemical shift and frequency offset information；
9) FQ (Flow Quantification) obtains vascular flow rate information；
10) QPI (Quantification perfusion imaging) technology, obtains each voxel perfusion information；
11) QDI (Quantification Dif fusion imaging) technology obtains each voxel diffusion information；
12) Dynamic constrasted enhancement imaging (Dynamic Contrast Enhanced, DCE) imaging technique is obtained per individual Blood transfer constant (Ktrans) in the capillary of element, the volume fraction of Plasma volumes score (VP) and the histocyte external space (VE) etc.；
13) and dynamic magnetic susceptibility compares (Dynamic Susceptibility Contrast, DSC) imaging technique, obtains Obtain the blood volume in each voxel, blood flow, the information such as mean transit time.
It is described in the construction method of automatic augmentation training sample proposed by the present invention for the diagnosis of AI+MRI Image-aideds The physical parameter of tissue includes at least weight T1, weight T2, weight T2*, the information such as proton density；Chemical parameters at least wrap It includes：Chemical shift, macroscopic magnetization rate (QSM), proton density fat content (PDFF), the information such as chemical shift and frequency shift (FS)； Physiological parameter includes at least：Perfusion, diffusion, blood flow velocity, blood volume, Plasma volumes score, blood flow, permeability of cell membrane etc. Information.
It is described in the construction method of automatic augmentation training sample proposed by the present invention for the diagnosis of AI+MRI Image-aideds Step S02 Plays or typical cytopathic MRI image sample calculate, including：Various physics, chemistry and the physiology of tissue voxel Parameter, according to the difference of required image weights and type, using different mathematical models, directly one by one voxel be calculated it is every The half-tone information of a pixel forms the image sample data library of different weights and feature；The present invention need not be to the letter of acquisition It ceases database and carries out profile data acquisition reconstruction image again, but calculating based on mathematical model directly is carried out to each voxel To half-tone information, calculating speed is faster.
It is described in the construction method of automatic augmentation training sample proposed by the present invention for the diagnosis of AI+MRI Image-aideds The mathematical model that each voxel signal amplitude in step S02 is followed include at least it is following in it is one or more：
(1) spin echo (SE) sequence, M0For (being indicated with S)：
S∝Aρ(H)[1-exp(-TR/T1)]exp(-TE/T2)；
In formula：A indicates signal amplification effect；TR, it is repetition time, TEFor the echo time；P (H) is proton density；T1 is The spin-lattice relaxation time (also known as longitudinal relaxation time T1) of tissue；T2 is that the spin spin relaxation time of tissue (is also known as Lateral relaxation time T2)；S is acquisition signal；
(2) GRE sequences, M0For (being indicated with S)：
Remove remanent magnetization (FLASH)：
In formula, T2* it is the T considered after the uneven effect of main field2Value；A is Flip angle；T1 is spin-crystalline substance of tissue The lattice relaxation time (also known as longitudinal relaxation time T1)；T2 is spin spin relaxation time (the also known as lateral relaxation time of tissue T2)；S is acquisition signal；
It utilizes remanent magnetization (bFFSP)：
(3) IR sequences, M0For (being indicated with S)：
S∝Aρ(H)[1-2exp(-TI/T1)]{1-exp[-(TR-TI)/T1]}exp(-TE/T2)；
T in formulaIFor reversing time；T1 is the spin-lattice relaxation time (also known as longitudinal relaxation time T1) of tissue；T2 is The spin spin relaxation time (also known as lateral relaxation time T2) of tissue；S is acquisition signal；
(4) EPI sequences；
(5) DWI sequences：
S∝Aρ(H)[1-2exp(-TI/T1)]{1-exp[-(TR-TI)/T1]}exp(-TE/T2)*exp(-bD)；
In formula (5), D is diffusion coefficient parameter；B is the diffusion factor, there is b=r2*G2*delta2*(Delta- delta/3).Wherein G is diffusion gradient amplitude；Delta is gradient application time；Intervals of the Delta between symmetric gradient；It is logical It crosses and sets different b values, different degrees of DWI images can be obtained.It is proposed by the present invention to be diagnosed for AI+MRI Image-aideds Automatic augmentation training sample construction method in, can be by quantitative information figure image that step S01 is generated in the step S03 Library enters directly as AI training samples, or using the weight image of the various features generated in step S02 as training sample AI systems.
In the construction method of automatic augmentation training sample proposed by the present invention for the diagnosis of AI+MRI Image-aideds, S01 The quantitative information image library of generation includes the image library of normal portions and the quantitative information image library of typical disease.
In the construction method of automatic augmentation training sample proposed by the present invention for the diagnosis of AI+MRI Image-aideds, S02 The image of the various weights generated includes the weight image library of normal portions and the weight image library of typical disease.
Based on the above construction method, the invention also provides a kind of automatic augmentation instructions for the diagnosis of AI+MRI Image-aideds Practice the structure system of sample, including：
One normal or quantitative information comprising typical disease lesions position builds module, for obtaining each picture of MRI image Physics, chemistry and the physiologic parameter value of multiple MRI signals of element build multidimensional data matrix；
More than one kinds of arbitrary cross-section weight MRI image generation module is used for using quantitative information module as sample, according to different sequences The mathematical model of row generates the MRI image of different weights；
One automatic training sample input module, quantitative information image library and MRI image library are sent into AI systems as training sample System carries out large sample training.
The present invention can solve the inefficient and sample kind of the artificial mark sample training in MRI image artificial intelligence auxiliary diagnosis Class is difficult to the problem of all enumerating, and automatic form provides a variety of (theoretically arbitrary a variety of) training samples, i.e. automaticdata augmentation； Simultaneously this method can getting parms according to diagnostic image to be analyzed, targetedly sample training is provided.
Compared with prior art, advantages of the present invention includes：
The training sample of AI medical imaging diagnosis technology is all the image that is marked by hand with doctor to carry out at present；Exactly Due to the deficiency of training sample so that application field ripe in terms of medical imaging diagnosis AI is few at present, is concentrated mainly on Lung neoplasm detects；For many magnetic resonance imagings of image type, sample training is its bottleneck.This method is using automatic structure Thought, solve the deficiency that generates sample inefficient manually and can not all cover；
AI medical images aided diagnosis technique is after enough sample trainings at present, for any input picture, according to The analytical judgment path of system, provides final result.Automatic structure sample based on this method, due to the number of its sample structure The method for learning model and imaging has a direct correlation, thus can by selection and the identical imaging sequence of image to be analyzed and Training sample under parameter is analyzed and is judged, more targeted carry out artificial intelligence analysis helps to improve result Accuracy and adaptability.
At present compared to for human body nuclear-magnetism technology, the relevant nuclear-magnetism technology of pet, disease collection of illustrative plates and the basis of diagnosis are more Add weakness, technical staff and doctor are fewer, and check price higher, and sample size is also more rare.It is disclosed by the invention to be used for AI+ The construction method of the automatic augmentation training sample of MRI image auxiliary diagnosis, can carry out human body diseases, can also meet and dote on The AI diagnostic applications of object (such as pet dog, horse racing, the preciousness such as police dog animal).
Description of the drawings
Fig. 1 is flow chart of the present invention for the automatic training sample construction method of AI+MRI Image-aideds diagnosis；
Fig. 2 is the present invention through quantitative imaging technique：T1, T2 and matter that T1mapping technologies and T2mapping technologies obtain Sub- density quantitative information figure effect；
Fig. 3 is to use the T1 weight picture results obtained after the present invention；
Fig. 4 is to use the T2 weight picture results obtained after the present invention；
Fig. 5 is to use the T1-FLAIR picture results obtained after the present invention；
Fig. 6 is to use the T2-FLAIR image results obtained after the present invention；
Fig. 7 is to use the STIR picture results obtained after the present invention；
Fig. 8 is to use the three-dimensional ectocinerea effect obtained after the present invention.
Fig. 9 is to weigh ghosting effect using the intracranial meningeoma T1 obtained after the present invention；
Figure 10 is to weigh ghosting effect using the intracranial meningeoma T2 obtained after the present invention；
Figure 11 is using the intracranial meningeoma T1-FLAIR obtained after the present invention as effect；
Figure 12 is using the intracranial meningeoma T2-FLAIR obtained after the present invention as effect；
Figure 13 is to use the intracranial meningeoma STIR fat suppression image effects obtained after the present invention；
Figure 14 be using the standard human brain SE sequences obtained after the present invention with Parameters variation (TE=10ms, TR=20ms~ T1 weight effect series of drawing 8000ms)；
Figure 15 is using the standard human brain SE sequences obtained after the present invention with Parameters variation (TE=145ms, TR=300ms ~10500ms) T2 weight effect series of drawing；
Figure 16 is using the standard human brain SE sequences obtained after the present invention with Parameters variation (TR=100ms and 200ms, TE =10ms~145ms) effect series of drawing；
Figure 17 be using the standard human brain IR sequences obtained after the present invention with Parameters variation (TR=500ms, TE=30ms, TI=20ms~3000ms) weight variation effect series of drawing；
Figure 18 is using the standard human brain SE sequences obtained after the present invention with Parameters variation (TR=10500ms, TE=10ms ~900ms) water imaging effect series of drawing；
Figure 19 is flow chart of the present invention for the automatic training sample construction method of AI+MRI Image-aideds diagnosis；
Figure 20 is the structural schematic diagram that the present invention builds system for the automatic training sample of AI+MRI Image-aideds diagnosis.
Specific implementation mode
In conjunction with following specific examples and attached drawing, the invention will be described in further detail.Implement process, the item of the present invention Part, experimental method etc. are among the general principles and common general knowledge in the art in addition to the following content specially referred to, the present invention Content is not particularly limited.
The present embodiment is illustrated by taking the automatic training sample construction method that AI+MRI Image-aideds diagnose as an example.
As shown in Figure 1, a kind of automatic training sample construction method of AI+MRI Image-aideds diagnosis, includes the following steps：
The first step obtains the more of normal human or position comprising typical disease by various quantitative MR imaging technologies A quantitative information database (image library)；
Each power is calculated in quantitative information database by second step, the mathematical model based on different sequences and sequential parameter Weight image library；
Third walks, and above-mentioned quantitative information library and weight image library are sent into AI systems as sample respectively is trained.
The quantitative information data of multiple information can be obtained by various quantitative imaging techniques in the first step.MRI signal Quantifiable signal include three categories：Physical message, chemical information and physiologic information, design parameter may include weight T1, weight T2, proton density, chemical shift, diffusion coefficient, perfusion coefficient, elasticity, vascular flow rate, direction etc. are all to have shadow to MRI signal The parameter of the relationship of sound, to realize Perfusion Imaging, functional imaging, elastogram etc..The present embodiment is with weight T1, weight T2, proton It is illustrated for three kinds of parameters of density.
It is obtained using T1 mapping, T2 mapping and PD the mapping technologies of clinical instrumentation more in single layer or volume T1, T2 and the PD value of each pixel of tomographic image.
Use T1Mapping technologies are equivalent to the T for seeking each pixel of image1Value, and should using this T1 value as image The gray value of point.IR sequences may be used, then by constantly changing TI(reversing time) collects different values, passes through longitudinal flux Change the value that the formula fitting changed over time calculates T1.The shortcomings that this sequence is to be repeated several times, TRWhat is set is longer (3T1~5T1) causes acquisition time very long.A kind of fast method is to use DESPOT1 methods, the stable state destroyed using radio frequency into Dynamic sequence fits T by changing Flip angle with steady-state signal strength formula1Value.
Use T2Mapping technologies are equivalent to the T for seeking each pixel of image2Value, and this T2Value should as image The gray value of point.Using the SE sequences of long TR times, by the grey scale pixel value under the different TE times, at any time by cross magnetization Between the formula fitting that changes calculate T2Value.
Using PD mapping technologies, the PD values for seeking each pixel of image are equivalent to, and using this PD value as image The gray value of the point.Using long TRThe gradation of image of (3~5T1) time, the SE sequences of most short TE, acquisition are exactly the distribution of PD；
T after tissue segmentation and assignment1, T2, PD Parameter Maps are as shown in Figure 2.
Quantitative MR imaging technology has been commercialized, and is integrated in the clinic or scientific research NMR imaging equipment of each company On, which is not described herein again.
The mathematical model based on different sequences and sequential parameter in second step, quantitative information database is calculated respectively Weight image library, directly according to the mathematical model of different sequences and sequential parameter, the signal amplitude for carrying out voxel one by one calculates：
It is with the pixel amplitudes mathematical model under three kinds of basic sequences such as SE, GRE, IR：
(1) spin echo (SE) sequence, M0For (being indicated with S)：
S∝Aρ(H)[1-exp(-TR/T1)]exp(-TE/T2)；
In formula：A indicates signal amplification effect；TR, it is repetition time, TEFor the echo time；P (H) is proton density；
(2) GRE sequences, M0For (being indicated with S)：
Remove remanent magnetization (FLASH)：
In formula, T2* it is the T considered after the uneven effect of main field2Value；A is Flip angle；
It utilizes remanent magnetization (bFFSP)：
(3) IR sequences, M0For (being indicated with S)：
S∝Aρ(H)[1-2exp(-TI/T1)]{1-exp[-(TR-TI)/T1]}exp(-TE/T2)；
T in formulaIFor reversing time；
(4) DWI sequences：
S∝Aρ(H)[1-2exp(-TI/T1)]{1-exp[-(TR-TI)/T1]}exp(-TE/T2)*exp(-bD)；
Wherein, D is diffusion coefficient parameter；B is the diffusion factor, there is b=r2*G2*delta2*(Delta-delta/ 3).G is diffusion gradient amplitude in formula；Delta is gradient application time；Intervals of the Delta between symmetric gradient；Pass through setting Different b values can obtain different degrees of DWI images.Can also EPI sequences etc. be obtained as needed；
The three-dimensional cross-sections image effect obtained by second step is as shown in Figure 5.
It can be used as AI training samples using the quantitative information image and weight image of this method structure.It can also obtain simultaneously To virtual MRI numbers collection of illustrative plates, experimental teaching can be used to, with reference to comparison, technician's training, anomalous identification etc.；Virtual MRI scan technology It can also be integrated into clinical MRI scanner, by the image information of single pass, virtual scan obtains the figure of other a variety of sequences Picture, to improve scanning efficiency.
The construction method can obtain in real time arbitrary cross-section, arbitrary weight partes corporis humani position normal portions MRI collection of illustrative plates, make Ghosting effect is weighed with the T1 obtained after the present invention, as shown in Figure 3；Ghosting effect is weighed using the T2 obtained after the present invention, such as Fig. 4 institutes Show；Using the T1-FLAIR image effects obtained after the present invention, as shown in Figure 5；Schemed using the T2-FLAIR obtained after the present invention As effect, as shown in Figure 6；Using the STIR obtained after the present invention as effect, as shown in Figure 7；After the present invention, to grey matter into The effect that row three-dimensionalreconstruction obtains, as shown in Figure 8.
The construction method can also obtain in real time arbitrary cross-section, arbitrary weight partes corporis humani position typical disease MRI collection of illustrative plates, Ghosting effect is weighed using the intracranial meningeoma T1 obtained after the present invention, as shown in Figure 9；Use the encephalic meninx obtained after the present invention Tumor T2 weighs ghosting effect, as shown in Figure 10；Use the intracranial meningeoma T1-FLAIR image effects obtained after the present invention, such as Figure 11 It is shown；Using the intracranial meningeoma T2-FLAIR image effects obtained after the present invention, as shown in figure 12；Using being obtained after the present invention Intracranial meningeoma STIR fat suppression image effects, as shown in figure 13.
The protection content of the present invention is not limited to above example.Without departing from the spirit and scope of the invention, originally Field technology personnel it is conceivable that variation and advantage be all included in the present invention, and with appended claims be protect Protect range.

Claims (6)

1. a kind of construction method of automatic augmentation training sample for the diagnosis of AI+MRI Image-aideds, which is characterized in that including Following steps：
S01：By the quantitative imaging technique of clinical MRI equipment, the site tissue and just of the typical disease of diseased individuals is obtained Often physical parameter, the quantitative information image library of chemical parameters and physiological parameter of individual corresponding site tissue voxel；
S02：Using the quantitative information image library as input, by the mathematical model of imaging method, it is calculated including each The MRI standard pictures and typical disease image of the various weight features of section；
S03：The step S01 quantitative information image libraries obtained or the step S02 are obtained various including cross sections The MRI standard pictures and typical disease image of weight feature are sent into AI algorithms, as training sample.
2. the construction method of the automatic augmentation training sample according to claim 1 for the diagnosis of AI+MRI Image-aideds, It is characterized in that, in the step S01, the quantitative imaging technique includes one or more of：
T1mapping technologies, the T1 information for obtaining each voxel；
T2mapping technologies, the T2 information for obtaining each voxel；
T2*mapping technologies, the T2* information for obtaining each voxel；
QSM technologies, the macroscopic magnetization rate information for obtaining each voxel；
DFF mapping technologies, the proton density fat content information for obtaining each voxel；
MT technologies, for obtaining each voxel macromolecular content information
CEST technologies, for obtaining protein, content of peptides information；
MRS technologies, for obtaining each voxel chemical shift and frequency offset information；
FQ technologies, for obtaining vascular flow rate information；
QPI technologies, for obtaining each voxel perfusion information；
QDI technologies spread information for obtaining each voxel；
Dynamic constrasted enhancement imaging techniques, blood transfer constant in the capillary for obtaining each voxel, Plasma volumes The volume fraction of score and the histocyte external space；With
Dynamic magnetic susceptibility compares imaging technique, for obtaining the blood volume in each voxel, blood flow, mean transit time letter Breath.
3. the construction method of the automatic augmentation training sample according to claim 1 for the diagnosis of AI+MRI Image-aideds, It is characterized in that, the physical parameter includes weight T1, weight T2, weight T2*, proton density information；The chemical parameters packet It includes：Chemical shift, macroscopic magnetization rate, proton density fat content, chemical shift and frequency offset information；The physiological parameter packet It includes：Perfusion, diffusion, blood flow velocity, blood volume, Plasma volumes score, blood flow, permeability of cell membrane information.
4. the construction method of the automatic augmentation training sample according to claim 1 for the diagnosis of AI+MRI Image-aideds, It is characterized in that, the step S02 includes：To various physical parameters, chemical parameters and the physiological parameter of voxel, according to required The difference of image weights and type calculates voxel one by one using different mathematical models, obtains the ash of each pixel Information is spent, the MRI standard pictures and typical disease image of different weight features are formed.
5. the construction method of the automatic augmentation training sample according to claim 4 for the diagnosis of AI+MRI Image-aideds, It is characterized in that, the mathematical model in the step S02, each weight image that quantitative information database is calculated, according to The mathematical model of different sequences and sequential parameter, the signal amplitude for carrying out voxel one by one calculate；The signal amplitude institute of each voxel The mathematical model followed include at least it is following in it is one or more：
(1) spin-echo sequence, M0For：
S∝Aρ(H)[1-exp(-TR/T1)]exp(-TE/T2)；
In formula (1)：A indicates signal amplification effect；TRFor repetition time, TEFor the echo time；P (H) is proton density；T1 is group The spin-lattice relaxation time knitted；T2 is the spin spin relaxation time of tissue；S is acquisition signal；
(2) GRE sequences, M0For：
Remove remanent magnetization：
In formula (2), T2* it is the T considered after the uneven effect of main field2Value；A is Flip angle；T1 is spin-crystalline substance of tissue The lattice relaxation time；T2 is the spin spin relaxation time of tissue；S is acquisition signal；
Utilize remanent magnetization：
(3) IR sequences, M0For：
S∝Aρ(H)[1-2exp(-TI/T1)]{1-exp[-(TR-TI)/T1]}exp(-TE/T2)；
T in formula (3)IFor reversing time；T1 is the spin-lattice relaxation time of tissue；When T2 is the spin-spin relaxation of tissue Between；S is acquisition signal；
(4) EPI sequences；
(5) DWI sequences：
S∝Aρ(H)[1-2exp(-TI/T1)]{1-exp[-(TR-TI)/T1]}exp(-TE/T2)*exp(-bD)；
In formula (5), D is diffusion coefficient parameter；B is the diffusion factor, there is b=r2*G2*delta2* (Delta-delta/3), Wherein G is diffusion gradient amplitude, and delta is gradient application time, intervals of the Delta between symmetric gradient, by setting not Same b values, can obtain different degrees of DWI images.
6. a kind of structure system of automatic augmentation training sample for the diagnosis of AI+MRI Image-aideds, which is characterized in that including：
Quantitative information builds module, and various physical parameters, chemical parameters and physiologic parameter value for obtaining voxel build multidimensional Data matrix；
A variety of arbitrary cross-section weight MRI image generation modules are used for using quantitative information module as sample, according to not homotactic number Model is learned, the MRI image of various weights is generated；
Automatic training sample input module, is sent into the MRI image of quantitative information image library and various weights as training sample AI systems carry out large sample training.
CN201810005912.6A 2018-01-03 2018-01-03 The construction method and system of automatic augmentation training sample for the diagnosis of AI+MRI Image-aideds Pending CN108324244A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810005912.6A CN108324244A (en) 2018-01-03 2018-01-03 The construction method and system of automatic augmentation training sample for the diagnosis of AI+MRI Image-aideds

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810005912.6A CN108324244A (en) 2018-01-03 2018-01-03 The construction method and system of automatic augmentation training sample for the diagnosis of AI+MRI Image-aideds

Publications (1)

Publication Number Publication Date
CN108324244A true CN108324244A (en) 2018-07-27

Family

ID=62924616

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810005912.6A Pending CN108324244A (en) 2018-01-03 2018-01-03 The construction method and system of automatic augmentation training sample for the diagnosis of AI+MRI Image-aideds

Country Status (1)

Country Link
CN (1) CN108324244A (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108992022A (en) * 2018-08-06 2018-12-14 广州高通影像技术有限公司 A kind of AI application system when the operation applied to intelligent endoscope
CN110246109A (en) * 2019-05-15 2019-09-17 清华大学 Merge analysis system, method, apparatus and the medium of CT images and customized information
CN110322408A (en) * 2019-06-11 2019-10-11 浙江大学 Multicenter magnetic resonance image automated quality control method based on cloud platform
CN113330483A (en) * 2019-09-18 2021-08-31 拜耳公司 Predicting magnetic resonance imaging images by predictive models trained by supervised learning
CN113329688A (en) * 2019-09-18 2021-08-31 拜耳公司 Systems, methods, and computer program products for predicting, anticipating, and/or evaluating tissue characteristics
US12002203B2 (en) 2019-03-12 2024-06-04 Bayer Healthcare Llc Systems and methods for assessing a likelihood of CTEPH and identifying characteristics indicative thereof

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103699904A (en) * 2013-12-25 2014-04-02 大连理工大学 Image computer-aided diagnosis method for multi-sequence nuclear magnetic resonance images
CN104424386A (en) * 2013-08-23 2015-03-18 北京大学 Multi-parameter magnetic resonance image based prostate cancer computer auxiliary identification system
CN107133638A (en) * 2017-04-01 2017-09-05 中南民族大学 Multi-parameter MRI prostate cancer CAD method and system based on two graders
CN107240102A (en) * 2017-04-20 2017-10-10 合肥工业大学 Malignant tumour area of computer aided method of early diagnosis based on deep learning algorithm

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104424386A (en) * 2013-08-23 2015-03-18 北京大学 Multi-parameter magnetic resonance image based prostate cancer computer auxiliary identification system
CN103699904A (en) * 2013-12-25 2014-04-02 大连理工大学 Image computer-aided diagnosis method for multi-sequence nuclear magnetic resonance images
CN107133638A (en) * 2017-04-01 2017-09-05 中南民族大学 Multi-parameter MRI prostate cancer CAD method and system based on two graders
CN107240102A (en) * 2017-04-20 2017-10-10 合肥工业大学 Malignant tumour area of computer aided method of early diagnosis based on deep learning algorithm

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
李月卿主编: "《医学影像成像原理（第2版）》", 30 September 2002 *
韩鸿宾等: "《临床磁共振成像序列设计与应用 第2版》", 31 October 2007 *

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108992022A (en) * 2018-08-06 2018-12-14 广州高通影像技术有限公司 A kind of AI application system when the operation applied to intelligent endoscope
US12002203B2 (en) 2019-03-12 2024-06-04 Bayer Healthcare Llc Systems and methods for assessing a likelihood of CTEPH and identifying characteristics indicative thereof
CN110246109A (en) * 2019-05-15 2019-09-17 清华大学 Merge analysis system, method, apparatus and the medium of CT images and customized information
CN110246109B (en) * 2019-05-15 2022-03-18 清华大学 Analysis system, method, device and medium fusing CT image and personalized information
CN110322408A (en) * 2019-06-11 2019-10-11 浙江大学 Multicenter magnetic resonance image automated quality control method based on cloud platform
CN113330483A (en) * 2019-09-18 2021-08-31 拜耳公司 Predicting magnetic resonance imaging images by predictive models trained by supervised learning
CN113329688A (en) * 2019-09-18 2021-08-31 拜耳公司 Systems, methods, and computer program products for predicting, anticipating, and/or evaluating tissue characteristics
CN113329688B (en) * 2019-09-18 2023-11-28 拜耳公司 Method, system and computer readable medium for determining characteristics of tissue voxels
US11915361B2 (en) 2019-09-18 2024-02-27 Bayer Aktiengesellschaft System, method, and computer program product for predicting, anticipating, and/or assessing tissue characteristics

Similar Documents

Publication Publication Date Title
CN108324244A (en) The construction method and system of automatic augmentation training sample for the diagnosis of AI+MRI Image-aideds
CN105188528B (en) Cerebration training device and cerebration training system
Jannin et al. Validation in medical image processing.
Kamiya et al. Automated segmentation of psoas major muscle in X-ray CT images by use of a shape model: preliminary study
US20170238879A1 (en) Method of Analyzing the Brain Activity of a Subject
Rajasekaran et al. Advanced brain tumour segmentation from mri images
Kamiya Deep learning technique for musculoskeletal analysis
Gooijers et al. White matter organization in relation to upper limb motor control in healthy subjects: exploring the added value of diffusion kurtosis imaging
Gilbert et al. Mapping complex myoarchitecture in the bovine tongue with diffusion-spectrum magnetic resonance imaging
Kawamoto et al. Simultaneous Learning of Erector Spinae Muscles for Automatic Segmentation of Site-Specific Skeletal Muscles in Body CT Images
Koike et al. Accurate preoperative identification of motor speech area as termination of arcuate Fasciculus depicted by Q-ball imaging tractography
Mishra et al. Medical image processing: A challenging analysis
Yao et al. Trimodal Fusion Network Combined Global-Local Feature Extraction Strategy and Spatial-Frequency Fusion Strategy
CN117078703B (en) CT image segmentation method and system based on MRI guidance
TW201439571A (en) Method of automatically analyzing brain fiber tracts information
US11918374B2 (en) Apparatus for monitoring treatment side effects
Sille Real time segmentation of medical images
Kemper Incorporation of diffusion tensor mri in non-rigid registration for image-guided neurosurgery
Mori Two and three-dimensional analyses of brain white matter architecture using diffusion imaging
Grova et al. Validation of MRI/SPECT similarity-based registration methods using realistic simulations of normal and pathological SPECT data
Lashkari et al. Evaluation of functional connectivity in control and ADHD subjects by generalized Canonical Correlation and Probabilistic Independent Component Analysis
Tainter et al. Evaluation of 3D multimodality image registration using ROC analysis
Viar-Hernandez et al. ADAPTING MATERIAL DECOMPOSITION ACQUISITIONS FOR DECT CALIBRATIONS IN PROTON THERAPY PLANNING
Albattal Multi-Stage and Multi-Target Data-Centric Approaches to Object Detection, Localization, and Segmentation in Medical Imaging
Parvathaneni Gray Matter Surface-based Spatial Statistics in Neuroimaging Studies

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20180727