CN110288609A - A kind of multi-modal whole-heartedly dirty image partition method of attention mechanism guidance - Google Patents
A kind of multi-modal whole-heartedly dirty image partition method of attention mechanism guidance Download PDFInfo
- Publication number
- CN110288609A CN110288609A CN201910461477.2A CN201910461477A CN110288609A CN 110288609 A CN110288609 A CN 110288609A CN 201910461477 A CN201910461477 A CN 201910461477A CN 110288609 A CN110288609 A CN 110288609A
- Authority
- CN
- China
- Prior art keywords
- image
- map
- mode
- attention mechanism
- segmentation
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10088—Magnetic resonance imaging [MRI]
-
- 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
- G06T2207/30048—Heart; Cardiac
Abstract
The invention discloses a kind of multi-modal whole-heartedly dirty image partition methods of attention mechanism guidance.Current mode is generated by Cycle-GAN and corresponds to the image of other mode to expand training set, and then original image and corresponding generation image pass through one and half twin networks progress image segmentations simultaneously.There are two independent encoder and a shared decoders for the half twin network, encoder is for learning the privately owned feature of mode, these features first pass through an attention mechanism module and carry out Fusion Features, and the shared feature of mode is then extracted by decoder and carries out last segmentation.The present invention makes full use of mode shared information and private information, improves segmentation precision.
Description
Technical field
The invention belongs to the field of medical imaging, in particular to a kind of multi-modal whole-heartedly dirty image partition method.
Background technique
According to American Heart Association (AHA) heart disease and stroke statistical report in 2019, in the U.S., 2019 about 1,
055,000 people's suffering from coronary heart disease, including 720,000 new and 335,000 recurrent coronary artery diseases
Example.In this sense, early diagnosis and therapy plays in terms of the death rate and disease incidence for reducing cardiovascular disease
Important function.During early diagnosis, it is complete to carry out that doctor usually collects image-forming information from different modes (for example, MR and CT)
Face investigation, one of them important prerequisite is heart minor structure of the accurate Ground Split from different modalities image.But it passes
The artificial segmentation of system is very time-consuming and laborious.Therefore, develop one whole-heartedly the dirty method divided automatically it is extremely urgent.
Although the method based on depth convolutional neural networks has been widely used for dividing other organs, these methods are answered
It is still limited for multi-modal full cardiac segmentation task, reason is: 1) mode inconsistency: the image from different modalities has
Apparent difference in appearance；2) complicated result: different heart minor structures is connection, can be even overlapped sometimes；3) each disease
Human heart in appearance also can be variant.
In recent years, there are some trials about multi-modal full cardiac segmentation.Such as Dou Qi et al. propose it is a kind of with pair
The unsupervised cross-domain generation frame of anti-study cross-module state medical image segmentation.Zheng Yefeng et al. proposes a synchronous study and turns
Change and the method for segmenting medical 3D rendering, this method can learn unpaired data set and keep their anatomical structure
It is constant.It being generated about image, the style of unpaired image can be generated to another domain from a domain migration by CycleGAN,
But it has that lacking label carries out deformation constraint.Ronneberger et al. is opened by " full convolutional neural networks FCN "
Hair, proposes " U-net " structure, and documentation & info up and down and a symmetrical expansion road are captured comprising a constricted path
Diameter is used to obtain accurate local message, is usually used in medical image segmentation.However the above method cannot all make full use of two moulds
The information or correlation that can be shared between state can not effectively overcome limitation mentioned above.
Summary of the invention
In order to solve the technical issues of above-mentioned background technique is mentioned, the invention proposes a kind of the more of attention mechanism guidance
Mode whole-heartedly dirty image partition method makes full use of mode shared information and private information, improves segmentation precision.
In order to achieve the above technical purposes, the technical solution of the present invention is as follows:
A kind of multi-modal whole-heartedly dirty image partition method of attention mechanism guidance, comprising the following steps:
(1) it generates across modality images:
It introduces and generates confrontation network, which includes 2 generators and 2 arbiters, respectively corresponds CT image and MRI figure
Picture, original CT image and MRI image are inputted respectively in corresponding generator, generate the figure for corresponding to another mode respectively
Picture；
(2) cross-module state feature learning and image segmentation:
Half twin network is constructed, which includes 2 independent encoders and 1 shared decoder, and in encoder
Attention mechanism module is set between decoder；Original CT image is inputted to wherein one together with the MRI image being generated by it
Original MRI image is inputted together with the CT image being generated by it another encoder, includes multilayer in encoder by a encoder
Down-sampling layer, the privately owned feature spectrogram of 2 encoder outputs, 2 mode, attention mechanism module merge the privately owned spy of 2 mode
Shared decoder is levied and is sent into, decoder exports the segmentation result of image.
Further, in generating confrontation network, it is defined as follows circulation consistency loss function Lcyc(GA,GB):
In above formula, xA、xBRespectively original CT image, MRI image sample,To obey pd(xA) distribution xA's
It is expected thatTo obey pd(xB) distribution xBExpectation, GA、GBRespectively correspond to the generator of CT image and MRI image；
In generating confrontation network, it is defined as follows segmentation loss function Lseg(SA/B,GA,GB):
In above formula, S is mappedA/B: A → Y, B → Y, A indicate CT mode, and B indicates MRI mode, and Y indicates that segmentation tag, i represent
One training sample, N are training sample sum, yA、yBRespectively in the true segmentation result of A mode and B mode.
Further, comprehensive circulation consistency loss function and segmentation loss function, define total loss function L (GA,GB,
DA,DB,SA/B):
L(GA,GB,DA,DB,SA/B)=LGAN(GA,DA)+LGAN(GB,DB)+λLcyc(GA,GB)+γLseg(SA/B,GA,GB)
In above formula, LGAN(GA,DA) and LGAN(GB,DB) pairs of anti-loss function of making a living, DA、DBRespectively correspond to CT image and
The arbiter of MRI image, λ, γ are the weight coefficient of loss function.
Further, in half twin network, the high-resolution feature of encoder localization, and capture more accurate
Information；The information of context is traveled to higher layers of resolution by decoder, and learns advanced semantic information.
Further, the process of the attention mechanism module is as follows:
The characteristic spectrum of 2 encoder outputs is connected first to obtain preliminary fusion map through channel layer, will tentatively be merged
Map recombinates to obtain map 1 through matrix, preliminary fusion map is successively obtained map 2 through matrix recombination and transposition, by 1 He of map
Map 2 makees the result after vector product and obtains attention map through softmax function, by attention map with tentatively merge map work
Element sums it up the result of vector product one by one with map progress is tentatively merged again, obtains final Fusion Features map.
By adopting the above technical scheme bring the utility model has the advantages that
The present invention carries out across modality images generations using improved CycleGAN to expand training set, and reduces mode
The image of layer is inconsistent；The invention proposes a kind of half twin networks of new cross-module state attention mechanism guidance, to learn
Mode can share the feature privately owned with mode, carry out multi-modal full segmenting cardiac images.The present invention efficiently solves marked 3D
The problems such as whole-heartedly dirty CT and MRI image quantity are few, the prior art underuses the relevant information of cross-module state improves segmentation essence
Degree, therefore application value with higher.
Detailed description of the invention
Fig. 1 is flow chart of the method for the present invention；
Fig. 2 is network structure of the invention；
Fig. 3 is the flow chart of attention mechanism module in the present invention；
Fig. 4 is embodiment segmentation result figure, is included (a), (b) two width subgraph.
Specific embodiment
Below with reference to attached drawing, technical solution of the present invention is described in detail.
The present invention devises a kind of multi-modal whole-heartedly dirty image partition method of attention mechanism guidance, as shown in Figs. 1-2,
The following steps are included:
1, it generates across modality images:
It introduces and generates confrontation network, which includes 2 generators and 2 arbiters, respectively corresponds CT image and MRI figure
Picture, they pass through mutual game, available extraordinary output during study.Original CT image and MRI image
It is inputted in corresponding generator respectively, generates the image for corresponding to another mode respectively.
To solve the problems, such as that generator will fight study from unpaired image, the present invention is generated using mandatory requirement
Sample GA(GB(xA)) and GB(GA(xB)) constraint that is consistent with original image, therefore define circulation consistency and lose letter
Number:
In above formula, xA、xBRespectively original CT image, MRI image sample,To obey pd(xA) distribution xA's
It is expected thatTo obey pd(xB) distribution xBExpectation, GA、GBRespectively correspond to the generator of CT image and MRI image.
If the transformation that two generators learn is reversible each other, the two generators will be consistent by recycling
Inspection without by any punishment.The generation of the problem of to prevent this data deformation, so defining another auxiliary
Map SA/B: A → Y, B → Y, wherein A represents CT mode, and B represents MR mode, and Y represents segmentation tag, and in segmentation result and
Take cross entropy as the segmentation loss function of constraint between true segmentation result:
In above formula, i represents a training sample, and N is training sample sum, yA、yBRespectively in the true of A mode and B mode
Real segmentation result.
Comprehensive circulation consistency loss function and segmentation loss function, define total loss function:
L(GA,GB,DA,DB,SA/B)=LGAN(GA,DA)+LGAN(GB,DB)+λLcyc(GA,GB)+γLseg(SA/B,GA,GB)
In above formula, LGAN(GA,DA) and LGAN(GB,DB) pairs of anti-loss function of making a living, DA、DBRespectively correspond to CT image and
The arbiter of MRI image, λ, γ are the weight coefficient of loss function.
2, cross-module state feature learning and image segmentation:
Half twin network is constructed, which includes 2 independent encoders and 1 shared decoder, and in encoder
Attention mechanism module is set between decoder.Original CT image is inputted to wherein one together with the MRI image being generated by it
Original MRI image is inputted together with the CT image being generated by it another encoder, includes multilayer in encoder by a encoder
Down-sampling layer, the privately owned feature spectrogram of 2 encoder outputs, 2 mode, attention mechanism module merge the privately owned spy of 2 mode
Shared decoder is levied and is sent into, decoder exports the segmentation result of image.
The effect of encoder is the high-resolution feature of localization, and captures more accurate information；The work of decoder
With being the semantic information that the information of context is traveled to higher layers of resolution, and learns advanced.
The present invention devises the attention mechanism module of a channel-wise between encoder and decoder.It will
Characteristic spectrum in two mode with each heart minor structure boundary information exports a new characteristic spectrum as input.
The process of attention mechanism module is as shown in Figure 3.First by the characteristic spectrum of 2 encoder outputs (assuming that having a size of C1×H×
W) connect to obtain preliminary fusion map (having a size of (C through channel layer1+C2) × H × W), it will tentatively merge map and be recombinated through matrix
To map 1, preliminary fusion map is successively obtained into map 2 through matrix recombination and transposition, after map 1 and map 2 are made vector product
Result obtain attention map (having a size of (C through softmax function1+C2)×(C1+C2)), by attention map with tentatively melt
The result that conjunction map makees vector product carries out element adduction one by one with the preliminary map that merges again, obtains final Fusion Features map
(having a size of (C1+C2)×H×W)。
Because two mode are all used to describe the same heart organ, it is assumed that the two mode may include many correlations
The feature of connection, this non-diagonal block C in attention map1×C2And C2×C1The available embodiment in part.Can also include in mode
The exclusive feature of some mode, this is in diagonal blocks C1×C1And C2×C2Part also available embodiment.
Effectiveness of the invention is hereafter verified by emulation experiment.
Stochastic gradient descent is carried out using Adam method, learning rate is 2 × 10-4, other settings are entirely by reference to CycleGAN
Setting go to train generator and arbiter.In order to accelerate training process, select to separate pre-training GA/BAnd DA/B, then complete again
Network is divided in ground training end to end.
Apply the inventive method to the Challenge contest of the field of medical imaging flagship meeting MICCAI 2017
On the public data collection that Multi-Modality Whole Heart Segmentation is provided.It contains in data set and does not match
Pair 20 MRI and 20 CT 3D rendering.The minor structure of data set has all carried out label by radiologist.Segmentation
Target is that be partitioned into 7 minor structures of heart include: left ventricle, atrium sinistrum, right ventricle, atrium dextrum, aorta, pulmonary artery and the heart
Flesh.In the training process, data set is divided into training set (10 samples) and test set (10 samples), and carries out eighty percent discount intersection
Verifying.CT mode is denoted as A, MRI mode is denoted as B.
Because the direction of data acquisition is different, all sample orientations have all been changed to by hat by software I TK-SNAP
Shape position.All be all cut there are the part of criterion label comes out, be cut into later the slice of 2D totally 2534 CT and
2208 MRI.By the size of resized to 128 × 128 after the different slice of these shapes.
In order to measure the gap of segmentation result and actual result, using Dice coefficient as evaluation index.Dice is used for
Measure the ratio being overlapped between authentic signature and segmentation result.Dice is higher, and proof segmentation precision is higher.
Cardiac segmentation visualizes comparing result as shown in figure 4, (a) in Fig. 4 is the cardiac segmentation visualization from CT to MRI
As a result, (b) being cardiac segmentation visualization result from MRI to CT, the Ours in figure is to be obtained using dividing method of the present invention
Segmentation result.As can be seen that the image generated is closely similar with original image, and missed without any apparent deformation and segmentation
Difference.
The Comparative result being split using distinct methods is as shown in table 1.As can be seen from the table, the present invention is successfully
Feature shared between mode is extracted to promote the segmentation precision of MRI image.The present invention has firstly evaluated full convolutional Neural net
Network (FCN) is respectively applied to the result of two mode segmentations as reference line.Then it has evaluated U-net and is respectively applied to two moulds
State.For the validity for further verifying the method for the present invention, CycleGAN method is used only on the data set and carries out across modal graph
As generating.It is not it is obvious that still MRI is divided that comparative experiments, which embodies the method for the present invention and acts on the segmentation precision for improving CT,
It is effective that the significantly improving of precision, which shows attention mechanism,.Also, attention mechanism efficiently avoids " bad " MRI figure
As leading astray risk caused by " good " CT image.
Table 1
| Method | Aorta | Atrium sinistrum | Left ventricle | Cardiac muscle | Right ventricle | Atrium dextrum | Pulmonary artery | It is average |
| FCN_CT | 0.8863 | 0.8163 | 0.8838 | 0.8541 | 0.7885 | 0.7940 | 0.7758 | 0.8284 |
| FCN_MRI | 0.6931 | 0.6882 | 0.8006 | 0.7161 | 0.6759 | 0.7593 | 0.6693 | 0.7146 |
| Unet_CT | 0.8992 | 0.7704 | 0.8381 | 0.8162 | 0.7643 | 0.8003 | 0.7947 | 0.8119 |
| Unet_MRI | 0.7719 | 0.6942 | 0.7751 | 0.6961 | 0.6983 | 0.7829 | 0.7171 | 0.7336 |
| CycleGAN_CT | 0.9407 | 0.8277 | 0.8362 | 0.7942 | 0.8064 | 0.8134 | 0.8103 | 0.8327 |
| CycleGAN_MRI | 0.7686 | 0.6555 | 0.7612 | 0.7038 | 0.6637 | 0.7658 | 0.6973 | 0.7165 |
| Ours_CT | 0.9282 | 0.8131 | 0.8497 | 0.7869 | 0.8066 | 0.8255 | 0.8391 | 0.8356 |
| Ours_MRI | 0.7875 | 0.6940 | 0.8031 | 0.7189 | 0.6733 | 0.7967 | 0.7147 | 0.7412 |
Embodiment is merely illustrative of the invention's technical idea, and this does not limit the scope of protection of the present invention, it is all according to
Technical idea proposed by the present invention, any changes made on the basis of the technical scheme are fallen within the scope of the present invention.
Claims (5)
1. a kind of multi-modal whole-heartedly dirty image partition method of attention mechanism guidance, which comprises the following steps:
(1) it generates across modality images:
It introduces and generates confrontation network, which includes 2 generators and 2 arbiters, respectively corresponds CT image and MRI image；
Original CT image and MRI image are inputted respectively in corresponding generator, generate the image for corresponding to another mode respectively；
(2) cross-module state feature learning and image segmentation:
Half twin network is constructed, which includes 2 independent encoders and 1 shared decoder, and in encoder and solution
Attention mechanism module is set between code device；Original CT image is inputted to one of volume together with the MRI image being generated by it
Original MRI image is inputted together with the CT image being generated by it another encoder, includes to adopt under multilayer in encoder by code device
Sample layer, the privately owned feature spectrogram of 2 encoder outputs, 2 mode, attention mechanism module merge the privately owned feature of 2 mode simultaneously
It is sent into shared decoder, decoder exports the segmentation result of image.
2. the multi-modal whole-heartedly dirty image partition method of attention mechanism guidance according to claim 1, which is characterized in that
It generates in confrontation network, is defined as follows circulation consistency loss function Lcyc(GA,GB):
In above formula, xA、xBRespectively original CT image, MRI image sample,To obey pd(xA) distribution xAExpectation,To obey pd(xB) distribution xBExpectation, GA、GBRespectively correspond to the generator of CT image and MRI image；
In generating confrontation network, it is defined as follows segmentation loss function Lseg(SA/B,GA,GB):
In above formula, S is mappedA/B: A → Y, B → Y, A indicate CT mode, and B indicates MRI mode, and Y indicates that segmentation tag, i represent one
Training sample, N are training sample sum, yA、yBRespectively in the true segmentation result of A mode and B mode.
3. the multi-modal whole-heartedly dirty image partition method of attention mechanism guidance according to claim 2, which is characterized in that comprehensive
Circulation consistency loss function and segmentation loss function are closed, total loss function L (G is definedA,GB,DA,DB,SA/B):
L(GA,GB,DA,DB,SA/B)=LGAN(GA,DA)+LGAN(GB,DB)+λLcyc(GA,GB)+γLseg(SA/B,GA,GB)
In above formula, LGAN(GA,DA) and LGAN(GB,DB) pairs of anti-loss function of making a living, DA、DBRespectively correspond to CT image and MRI
The arbiter of image, λ, γ are the weight coefficient of loss function.
4. the multi-modal whole-heartedly dirty image partition method of attention mechanism guidance according to claim 1, which is characterized in that
In half twin network, the high-resolution feature of encoder localization, and capture more accurate information；Decoder is by context
Information travel to higher layers of resolution, and learn advanced semantic information.
5. the multi-modal whole-heartedly dirty image partition method of attention mechanism guidance according to claim 1, which is characterized in that institute
The process for stating attention mechanism module is as follows:
The characteristic spectrum of 2 encoder outputs is connected first to obtain preliminary fusion map through channel layer, will tentatively merge map
It recombinates to obtain map 1 through matrix, preliminary fusion map is successively obtained into map 2 through matrix recombination and transposition, by map 1 and map
2, which make the result after vector product, obtains attention map through softmax function, by attention map with tentatively merging map makees vector
Element sums it up long-pending result one by one with map progress is tentatively merged again, obtains final Fusion Features map.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910461477.2A CN110288609B (en) | 2019-05-30 | 2019-05-30 | Multi-modal whole-heart image segmentation method guided by attention mechanism |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910461477.2A CN110288609B (en) | 2019-05-30 | 2019-05-30 | Multi-modal whole-heart image segmentation method guided by attention mechanism |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110288609A true CN110288609A (en) | 2019-09-27 |
| CN110288609B CN110288609B (en) | 2021-06-08 |
Family
ID=68002969
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910461477.2A Active CN110288609B (en) | 2019-05-30 | 2019-05-30 | Multi-modal whole-heart image segmentation method guided by attention mechanism |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110288609B (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111179207A (en) * | 2019-12-05 | 2020-05-19 | 浙江工业大学 | Cross-modal medical image synthesis method based on parallel generation network |
| CN111353499A (en) * | 2020-02-24 | 2020-06-30 | 上海交通大学 | Multi-modal medical image segmentation method, system, storage medium and electronic device |
| CN111696027A (en) * | 2020-05-20 | 2020-09-22 | 电子科技大学 | Multi-modal image style migration method based on adaptive attention mechanism |
| CN112308833A (en) * | 2020-10-29 | 2021-02-02 | 厦门大学 | One-shot brain image segmentation method based on circular consistent correlation |
| CN113177943A (en) * | 2021-06-29 | 2021-07-27 | 中南大学 | Cerebral apoplexy CT image segmentation method |
| CN113312530A (en) * | 2021-06-09 | 2021-08-27 | 哈尔滨工业大学 | Multi-mode emotion classification method taking text as core |
| CN113537057A (en) * | 2021-07-14 | 2021-10-22 | 山西中医药大学 | Facial acupuncture point automatic positioning detection system and method based on improved cycleGAN |
| CN113779298A (en) * | 2021-09-16 | 2021-12-10 | 哈尔滨工程大学 | Medical vision question-answering method based on composite loss |
| CN114842312A (en) * | 2022-05-09 | 2022-08-02 | 深圳市大数据研究院 | Generation and segmentation method and device for unpaired cross-modal image segmentation model |
| CN116883247A (en) * | 2023-09-06 | 2023-10-13 | 感跃医疗科技(成都)有限公司 | Unpaired CBCT image super-resolution generation algorithm based on Cycle-GAN |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108564119A (en) * | 2018-04-04 | 2018-09-21 | 华中科技大学 | A kind of any attitude pedestrian Picture Generation Method |
| CN108897740A (en) * | 2018-05-07 | 2018-11-27 | 内蒙古工业大学 | A kind of illiteracy Chinese machine translation method based on confrontation neural network |
| CN109325951A (en) * | 2018-08-13 | 2019-02-12 | 深圳市唯特视科技有限公司 | A method of based on the conversion and segmenting medical volume for generating confrontation network |
| CN109325931A (en) * | 2018-08-22 | 2019-02-12 | 中北大学 | Based on the multi-modality images fusion method for generating confrontation network and super-resolution network |
| CN109598745A (en) * | 2018-12-25 | 2019-04-09 | 上海联影智能医疗科技有限公司 | Method for registering images, device and computer equipment |
| CN109637634A (en) * | 2018-12-11 | 2019-04-16 | 厦门大学 | A kind of medical image synthetic method based on generation confrontation network |
| CN109685813A (en) * | 2018-12-27 | 2019-04-26 | 江西理工大学 | A kind of U-shaped Segmentation Method of Retinal Blood Vessels of adaptive scale information |
| US20190147582A1 (en) * | 2017-11-15 | 2019-05-16 | Toyota Research Institute, Inc. | Adversarial learning of photorealistic post-processing of simulation with privileged information |
| CN109801294A (en) * | 2018-12-14 | 2019-05-24 | 深圳先进技术研究院 | Three-dimensional atrium sinistrum dividing method, device, terminal device and storage medium |
-
2019
- 2019-05-30 CN CN201910461477.2A patent/CN110288609B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190147582A1 (en) * | 2017-11-15 | 2019-05-16 | Toyota Research Institute, Inc. | Adversarial learning of photorealistic post-processing of simulation with privileged information |
| CN108564119A (en) * | 2018-04-04 | 2018-09-21 | 华中科技大学 | A kind of any attitude pedestrian Picture Generation Method |
| CN108897740A (en) * | 2018-05-07 | 2018-11-27 | 内蒙古工业大学 | A kind of illiteracy Chinese machine translation method based on confrontation neural network |
| CN109325951A (en) * | 2018-08-13 | 2019-02-12 | 深圳市唯特视科技有限公司 | A method of based on the conversion and segmenting medical volume for generating confrontation network |
| CN109325931A (en) * | 2018-08-22 | 2019-02-12 | 中北大学 | Based on the multi-modality images fusion method for generating confrontation network and super-resolution network |
| CN109637634A (en) * | 2018-12-11 | 2019-04-16 | 厦门大学 | A kind of medical image synthetic method based on generation confrontation network |
| CN109801294A (en) * | 2018-12-14 | 2019-05-24 | 深圳先进技术研究院 | Three-dimensional atrium sinistrum dividing method, device, terminal device and storage medium |
| CN109598745A (en) * | 2018-12-25 | 2019-04-09 | 上海联影智能医疗科技有限公司 | Method for registering images, device and computer equipment |
| CN109685813A (en) * | 2018-12-27 | 2019-04-26 | 江西理工大学 | A kind of U-shaped Segmentation Method of Retinal Blood Vessels of adaptive scale information |
Non-Patent Citations (4)
| Title |
|---|
| QINGSONG YANG等: "Low-Dose CT Image Denoising Using a Generative Adversarial Network With Wasserstein Distance and Perceptual Loss", 《IEEE TRANSACTIONS ON MEDICAL IMAGING》 * |
| 屈宗艳: "基于注意机制的图像分割研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 李璐 等: "基于视觉显著性的目标分割算法", 《计算机科学与探索》 * |
| 郑顾平 等: "基于注意力机制的多尺度融合航拍影像语义分割", 《图学学报》 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111179207A (en) * | 2019-12-05 | 2020-05-19 | 浙江工业大学 | Cross-modal medical image synthesis method based on parallel generation network |
| CN111179207B (en) * | 2019-12-05 | 2022-04-08 | 浙江工业大学 | Cross-modal medical image synthesis method based on parallel generation network |
| CN111353499A (en) * | 2020-02-24 | 2020-06-30 | 上海交通大学 | Multi-modal medical image segmentation method, system, storage medium and electronic device |
| CN111353499B (en) * | 2020-02-24 | 2022-08-19 | 上海交通大学 | Multi-modal medical image segmentation method, system, storage medium and electronic device |
| CN111696027A (en) * | 2020-05-20 | 2020-09-22 | 电子科技大学 | Multi-modal image style migration method based on adaptive attention mechanism |
| CN112308833A (en) * | 2020-10-29 | 2021-02-02 | 厦门大学 | One-shot brain image segmentation method based on circular consistent correlation |
| CN112308833B (en) * | 2020-10-29 | 2022-09-13 | 厦门大学 | One-shot brain image segmentation method based on circular consistent correlation |
| CN113312530B (en) * | 2021-06-09 | 2022-02-15 | 哈尔滨工业大学 | Multi-mode emotion classification method taking text as core |
| CN113312530A (en) * | 2021-06-09 | 2021-08-27 | 哈尔滨工业大学 | Multi-mode emotion classification method taking text as core |
| CN113177943A (en) * | 2021-06-29 | 2021-07-27 | 中南大学 | Cerebral apoplexy CT image segmentation method |
| CN113177943B (en) * | 2021-06-29 | 2021-09-07 | 中南大学 | Cerebral apoplexy CT image segmentation method |
| CN113537057A (en) * | 2021-07-14 | 2021-10-22 | 山西中医药大学 | Facial acupuncture point automatic positioning detection system and method based on improved cycleGAN |
| CN113779298A (en) * | 2021-09-16 | 2021-12-10 | 哈尔滨工程大学 | Medical vision question-answering method based on composite loss |
| CN113779298B (en) * | 2021-09-16 | 2023-10-31 | 哈尔滨工程大学 | Medical vision question-answering method based on composite loss |
| CN114842312A (en) * | 2022-05-09 | 2022-08-02 | 深圳市大数据研究院 | Generation and segmentation method and device for unpaired cross-modal image segmentation model |
| CN114842312B (en) * | 2022-05-09 | 2023-02-10 | 深圳市大数据研究院 | Generation and segmentation method and device for unpaired cross-modal image segmentation model |
| CN116883247A (en) * | 2023-09-06 | 2023-10-13 | 感跃医疗科技(成都)有限公司 | Unpaired CBCT image super-resolution generation algorithm based on Cycle-GAN |
| CN116883247B (en) * | 2023-09-06 | 2023-11-21 | 感跃医疗科技(成都)有限公司 | Unpaired CBCT image super-resolution generation algorithm based on Cycle-GAN |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110288609B (en) | 2021-06-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110288609A (en) | A kind of multi-modal whole-heartedly dirty image partition method of attention mechanism guidance | |
| CN111709953B (en) | Output method and device in lung lobe segment segmentation of CT (computed tomography) image | |
| CN111833359B (en) | Brain tumor segmentation data enhancement method based on generation of confrontation network | |
| Liu et al. | CT synthesis from MRI using multi-cycle GAN for head-and-neck radiation therapy | |
| Hu et al. | Brain MR to PET synthesis via bidirectional generative adversarial network | |
| CN109598722B (en) | Image analysis method based on recurrent neural network | |
| CN108268870A (en) | Multi-scale feature fusion ultrasonoscopy semantic segmentation method based on confrontation study | |
| CN105760874B (en) | CT image processing system and its CT image processing method towards pneumoconiosis | |
| CN110544264A (en) | Temporal bone key anatomical structure small target segmentation method based on 3D deep supervision mechanism | |
| CN109242860A (en) | Based on the brain tumor image partition method that deep learning and weight space are integrated | |
| CN107563434A (en) | A kind of brain MRI image sorting technique based on Three dimensional convolution neutral net, device | |
| Pluim et al. | The truth is hard to make: Validation of medical image registration | |
| Mercan et al. | Virtual staining for mitosis detection in breast histopathology | |
| CN109685787A (en) | Output method, device in the lobe of the lung section segmentation of CT images | |
| CN107330953A (en) | A kind of Dynamic MRI method for reconstructing based on non-convex low-rank | |
| Tobon‐Gomez et al. | Realistic simulation of cardiac magnetic resonance studies modeling anatomical variability, trabeculae, and papillary muscles | |
| CN115578404A (en) | Liver tumor image enhancement and segmentation method based on deep learning | |
| Lindner et al. | Using synthetic training data for deep learning-based GBM segmentation | |
| CN109727197A (en) | A kind of medical image super resolution ratio reconstruction method | |
| CN114881848A (en) | Method for converting multi-sequence MR into CT | |
| Tobon-Gomez et al. | Automatic construction of 3D-ASM intensity models by simulating image acquisition: Application to myocardial gated SPECT studies | |
| Fan et al. | TR-Gan: multi-session future MRI prediction with temporal recurrent generative adversarial Network | |
| JP2022077991A (en) | Medical image processing apparatus, medical image processing method, medical image processing program, model training apparatus, and training method | |
| CN105913388A (en) | Priority constraint colorful image sparse expression restoration method | |
| CN108596900A (en) | Thyroid-related Ophthalmopathy medical image data processing unit, method, computer readable storage medium and terminal device |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |