CN110895828B - Model and method for generating MR (magnetic resonance) image simulating heterogeneous flexible biological tissue - Google Patents
Model and method for generating MR (magnetic resonance) image simulating heterogeneous flexible biological tissue Download PDFInfo
- Publication number
- CN110895828B CN110895828B CN201911217179.5A CN201911217179A CN110895828B CN 110895828 B CN110895828 B CN 110895828B CN 201911217179 A CN201911217179 A CN 201911217179A CN 110895828 B CN110895828 B CN 110895828B
- Authority
- CN
- China
- Prior art keywords
- image
- module
- sub
- array
- generating
- 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.)
- Active
Links
Images
Classifications
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
The invention relates to a model and a method for generating MR images simulating heterogeneous flexible biological tissues, wherein the model comprises a segmentation module, an automatic coding module and an image generation module, and the segmentation module is used for segmenting an input full-size real MR image x into a plurality of array sub-image modules x a The automatic coding module comprises a coder F enc Decoder F dec And discriminator D JSD Encoder F enc For converting an array sub-image module into n codes q (x) a ) A =1,2, … n, decoder F dec For combining each code q (x) a ) Module for converting into corresponding array pseudo sub-image
Discriminator D JSD For combining the codes q (x) a ) And random noise p (x) a ) Made as a new code q' (x) a ). The invention learns a direct mapping from the original full-size real MR image x and then uses this mapping and the constrained noise vectors to generate a large number of MR images.
Description
Technical Field
The invention relates to a generation method for simulating a three-dimensional nuclear magnetic resonance image, in particular to a model and a method for generating an MR image for simulating a heterogeneous flexible biological tissue.
Background
In recent years, methods widely used for medical image synthesis can be roughly classified into three types: 1. the automatic encoder can encode the features shown by the high-dimensional data into low-dimensional vectors, and then convert the low-dimensional vectors into high-dimensional vectors represented by the original data, so as to realize the generation of images; 2. generating a countermeasure network, namely continuously gaming through a generating network G (Generator) and a discriminating network D (Discriminator), further enabling G to learn the distribution of data, and after training is finished, G can generate a vivid image from a section of random number; 3. combination of an autoencoder with a generation countermeasure network.
For the autoencoder, kommrusch S et al proposed a network named LuNG that trained an autoencoder neural network with 3 potential feature neurons in the bottleneck layer using a training image as input, using the autoencoder as a generation model, the output of the autoencoder being subjected to a clean-up algorithm to increase the likelihood of generating an image that can be used for medical research. However, the input of the LuNG network is that 51 seed images are modified into 816 images as a training set, that is, the network still needs a large number of medical images as a training set, and we are generated due to lack of data. Furthermore, the generation of images by means of an automatic encoder is one-to-one, and if a large amount of data is required, the cost of the method is relatively high.
Goodfellow et al propose a method for image synthesis using GAN, which synthesizes data directly from random noise vectors, rather than outputting one-to-one with an image as input. Since then, GAN has rapidly evolved and continues to improve. Li Z et al propose structure-enhanced GAN (SEGAN) for recovering the missing function in global and local scales to compute the correlation between the split patches, and a new generator SU-Net that includes different sized convolution filters at each layer in order to capture multi-scale structure information. Although this approach has been successful in dealing with its target task, the direct mapping from random noise to the generated image completely ignores the inherent complexity differences between the synthesis tasks of MR images, and thus the synthesis tasks encounter greater challenges.
Given the difference in task complexity, j-y.zhu et al propose a synthesis method combining an auto-encoder with GAN, which is one-to-many image generation. The steganographic z is first sampled randomly from a known distribution and the generator maps the input image and the steganographic samples z to the generated output samples. The method is trained in an unsupervised manner, wherein only the encoding of the actual data and the corresponding images are used. However, this method is difficult to capture meaningful information for MRI because medical images are more biased towards structural coherence than natural images, so that the learning process is dominated by modeling structure rather than texture distribution. Furthermore, if the amount of training data is limited, the variety of synthetic data is still limited because the method cannot generate "invisible" data from random inputs due to overfitting these codes. In summary, existing methods have difficulty synthesizing MRI data with sufficient diversity, which contains meaningful structural information.
Although both the automatic encoder and the GAN can generate pictures, the MR images have a complex structure, unlike general clothing, human face and other images, so a scheme with better generation effect needs to be adopted. The existing GAN model has strong pertinence in generating images, namely, one group of data sets corresponds to one model, and if another data set is changed, one model needs to be retrained, so that an MR image does not find an available model to generate pictures, and a model which can be used for generating the MR image is constructed according to the idea of combining GAN and an automatic encoder.
Disclosure of Invention
The technical problem to be solved by the invention is to provide a model and a method for generating MR images, namely magnetic resonance images, of simulated heterogeneous flexible biological tissues aiming at the defects.
In order to solve the technical problems, the invention adopts the following technical scheme:
generating simulated heterogeneous flexible raw materialThe model of the object tissue MR image comprises a segmentation module, an automatic coding module and an image generation module, wherein the segmentation module is used for segmenting an input full-size real MR image x into a plurality of array sub-image modules x a Said automatic coding module comprising a coder F enc Decoder F dec And discriminator D JSD Said encoder F enc For converting an array sub-image module into n codes q (x) a ) A =1,2, … n, said decoder F dec For encoding each code q (x) a ) Module for converting into corresponding array pseudo sub-image
The discriminator D JSD For combining the codes q (x) a ) And random noise p (x) a ) Made as a new code q' (x) a ) Where a =1,2, … n, the image generation module includes a generator G for generating each new code q' (x) and a helicelayer module a ) And array pseudo sub-image block->
Conversion into a corresponding synthetic patch>
The SpliceLayer module is used for combining the patch->
Spliced into a full-size image->
Said +>
The MR image is the simulated MR image of the heterogeneous flexible biological tissue.
A method of generating MR images of simulated heterogeneous flexible biological tissue, comprising the steps of:
step S1: decomposing an input full-size real MR image x into n array sub-imagesModule x a ，a＝1,2，…n；
Step S2: all array subimage modules x a Array input encoder F enc To obtain each array sub-image module x a Corresponding code q (x) a )，a＝1,2,…n；
Step S3, the code q (x) a ) Input decoder F dec Module for obtaining array pseudo sub-image
The code q (x) a ) And random noise p (x) a ) Input, a =1,2, … n input discriminator D JSD To obtain a new code q' (x) a )，a＝1,2，…n；
Step S4, array pseudo sub-image module
And a new code q' (x) a ) A =1,2, … n input generator G, resulting in a synthesized patch ÷ based on>
Wherein a =1,2, … n;
step S5, synthesizing the patch
The input SpliceLayer module, which will make n synthesized patches->
Coherent stitching into a full-size image->
Is/are>
The MR image is the simulated MR image of the heterogeneous flexible biological tissue.
Furthermore, the stitching method of the SpliceLayer module is that each composite image unit is stitched
Corresponds to each original array sub-image module x a And determines a combined image unit on the basis of the identifier of the image unit>
Locating the position of a sub-image in the matrix, the full-size image->
By n equal-sized synthetic image units->
And (4) forming.
Further, the discriminator D JSD The calculation formula of (2) is as follows:
wherein E x～p(x) Is a desire for x.
Further, the random noise p (x) a ) The vector length of a =1,2, … n is set to 128.
Further, the loss function of the generator G is
Said loss function +>
Is composed of
And &>
Sum, said->
And &>
The calculation formula of (2) is as follows:
wherein
Computing the a-th sub-image x a And its array pseudo-sub-image module>
Pixel-level euclidean distance between them, where a =1,2, … n, n is a natural number.
The beneficial effects of the invention are as follows:
the invention learns a direct mapping in an original full-size real MR image x, and then generates a large number of MR images by using the mapping and constrained noise vectors.
Drawings
FIG. 1 is a general schematic of the present invention;
FIG. 2 is a schematic diagram of the generator G;
fig. 3 is a schematic structural diagram of the discriminator D.
Detailed Description
The principles and features of this invention are described below in conjunction with the following drawings, which are set forth to illustrate, but are not to be construed to limit the scope of the invention.
As shown in figure 1, the model for generating the MR image simulating the heterogeneous flexible biological tissue comprises a segmentation module, an automatic coding module and an image generation module, wherein the segmentation module is used for segmenting an input full-size real MR image x into a plurality of array sub-image modules x a Said automatic coding module comprising a coder F enc Decoder F dec And discriminator D JSD Said encoder F enc For converting an array sub-image module into n codes q (x) a ) A =1,2, … n, said decoder F dec For combining each code q (x) a ) Module for converting into corresponding array pseudo sub-image
The discriminator D JSD For encoding the code q (x) a ) And random noise p (x) a ) Made as a new code q' (x) a ) Wherein a =1,2, … n, the image generation module includes a generator G, spliceLayer module and a discriminator D, the generator G for applying each new code q' (x) a ) And array pseudo sub-image block->
Conversion into the corresponding synthetic patch->
The SpliceLayer module is used for combining the patch->
Spliced into a full-size image->
The discriminator D is used to determine whether the MR image x is full-size or not by inputting a full-size real MR image x and a full-size image->
Discriminating full size images>
Whether true or false, the->
The MR image of the simulated inhomogeneous flexible biological tissue is obtained.
A method of generating MR images simulating inhomogeneous flexible biological tissue, comprising the steps of:
(1) Decomposition of an input full-size real MR image x into n array sub-image modules x a ，a＝1,2，…n；
(2) All array subimage modules x a Array input encoder F enc To obtain a code q (x) a )，a＝1,2,…n；
(3) The code q (x) a ) A =1,2, … n, input decoder F dec Module for obtaining array pseudo sub-image
The code q (x) a ) A =1,2, … n, and random noise p (x) a ) A =1,2, … n input discriminator D JSD To obtain a new code q' (x) a )，a＝1,2,…n；
The random noise p (x) a ) The vector length of a =1,2, … n is set to 128;
to be able to generate reasonable MR images from noise vectors instead of MRI encoding, we encode q (x) a ) Is reconstructed into a predefined distribution p (x) a ) A =1,2, … n. The MRI data distribution is located on a 128-dimensional manifold, represented as a latent space, and the reconstructed distribution q' (x) is calculated by minimizing Jensen Shannon divergence (i.e., JSD) a ) Predefined distribution p (x) of a =1,2, … n a ) A =1,2, … n and actual MRI encoding distribution q (x) a ) Loss between a =1,2, … n minimizes JSD values. Thus, q (x) a ) And p (x) a ) The JSD value calculation formula between is:
wherein E x～p(x) Is a desire for x.
(4) Pseudo sub-image module of array
And a new code q' (x) a ) A =1,2, … n input generator G, resulting in a resultant patch ≥ r>
Wherein a =1,2, · · n;
as shown in fig. 2, the generator G introduces a new code q' (x) in a conventional manner, using a common encoder-decoder strategy a ) A =1,2, … n, the encoder portion of generator G acts as a feature extractor, with the multi-layer structure therein capturing dramatically the local and more global data representation. Each layer in the generator consists of one residual block and the size of the residual block is marked on each layer in fig. 2. 128-dimensional noise code q' (x) a ) All connected to the first layer. Unless otherwise noted, all layers of the generator G and arbiter D use a kernel size of 4 and a stride of 2. Generator G and discriminator D do not involve pooling layers;
the loss function of the generator G is
The loss function->
Is composed of
And &>
Sum, said->
And &>
Comprises the following steps:
wherein
Calculate the ith sub-image x a And its pseudo sub-image module>
Pixel-level euclidean distance between them, where a =1,2, … n, n is a natural number.
(5) Will synthesize the patch
Wherein a =1,2 · n is input to the SpliceLayer module, which will synthesize a patch &>
The consecutive stitching is a full size image->
The generator G will create a number n of equally sized patches
The SpliceLayer module holds each synthesized image unit>
Corresponds to each original image unit x a And determines a combined image unit based on the identifier of the image unit>
Positioning sub-images in the matrix such that the patches are placed in the matrix and spliced into a complete image, said full-size image ÷ or>
By a number of equally sized combined image units->
And (4) forming. />
(6) Will full size image
And inputting the full-size real MR image into a discriminator D to determine the authenticity of the synthetic image, and verifying the authenticity of the MR image of the simulated heterogeneous flexible biological tissue to conclude that the MR image of the simulated heterogeneous flexible biological tissue can be falsified.
As shown in FIG. 3, discriminator D uses the Convolition-BatchNorm-LeakyRelu component as a building block to determine whether the newly created full image is true or false;
the loss function formula of the discriminator D is as follows:
an auto-encoder (auto encoder) is an unsupervised neural network model that can be trained on implicit features of the input data, called encoding (coding), which in the present invention is the encoder F enc Through an encoder F enc Acquiring the code of the image; at the same time, the original input data can be reconstructed using the learned new features, called decoding, which in the present invention is decoder F dec By means of a decoder F dec The encoding is decoded into an image.
The invention learns a direct mapping from the original full-size real MR image x and then uses this mapping and the constrained noise vectorA large number of MR images are generated. More specifically, the original image is represented by x, q (x) a ) Representing constrained noise vectors, by mimicking the image formation process, let G:
representing the original image x and the noise code p (x) a ) An image generation process as input and generates an output image->
The length of the noise code is set to 128, and the generator is usually a rather complex function in view of the complexity of the image forming process, but nevertheless, an end-to-end (image-to-image) image compositor is accomplished by employing a powerful GAN deep learning framework and the flexibility of an auto-encoder.
The foregoing is illustrative of the best mode of the invention and details not described herein are within the common general knowledge of a person of ordinary skill in the art. The scope of the present invention is defined by the appended claims, and any equivalent modifications based on the technical teaching of the present invention are also within the scope of the present invention.
Claims (5)
1. The model for generating the MR image simulating the heterogeneous flexible biological tissue is characterized by comprising a segmentation module, an automatic coding module and an image generation module, wherein the segmentation module is used for segmenting an input full-size real MR image x into a plurality of array sub-image modules x a Said automatic coding module comprising a coder F enc A decoder F dec And discriminator D JSD Said encoder F enc For converting an array sub-image module into n codes q (x) a ) A =1,2, … n, said decoder F dec For encoding each code q (x) a ) Module for converting into corresponding array pseudo sub-image
The discriminator D JSD For combining the codes q (x) a ) And random noise p (x) a ) Made as a new code q' (x) a ) Where a =1,2, … n, the image generation module includes a generator G for generating each new code q' (x) and a helicelayer module a ) And array pseudo sub-image block->
Conversion into the corresponding synthetic patch->
The SpliceLayer module is used for combining the patch->
Stitched into full size images
Is/are>
The MR image of the simulated heterogeneous flexible biological tissue is obtained; the stitching method of the SpliceLayer module is that each combined image unit is matched and matched>
Corresponds to each original array sub-image module x a And determines a combined image unit based on the identifier of the image unit>
Locating a position of a sub-image in a matrix, the full-size image->
By n equal-sized synthetic image units->
And (4) forming.
2. The method for generating the MR image of the simulated heterogeneous flexible biological tissue is characterized by comprising the following steps of:
step S1: decomposition of an input full-size real MR image x into n array sub-image modules x a ，a＝1,2，…n；
Step S2: all array subimage modules x a Array input encoder F of enc To obtain each array sub-image module x a Corresponding code q (x) a )，a＝1,2,…n；
Step S3, the code q (x) a ) Input decoder F dec Module for obtaining array pseudo sub-image
The code q (x) a ) And random noise p (x) a ) A =1,2, … n input discriminator D JSD To obtain a new code q' (x) a )，a＝1,2，…n；
S4, array pseudo sub-image module
And a new code q' (x) a ) A =1,2, … n input generator G, resulting in a resultant patch ≥ r>
Wherein a =1,2, … n;
step S5, synthesizing the patch
Inputting the n composite patches into a SpliceLayer module
The consecutive stitching is a full size image->
Said +>
The MR image of the simulated heterogeneous flexible biological tissue is obtained; the stitching method of the SpliceLayer module is that each combined image unit is matched and matched>
Corresponds to each original array sub-image module x a And determines a combined image unit based on the identifier of the image unit>
Locating a position of a sub-image in a matrix, the full size image->
By n equal-sized synthetic image units->
And (4) forming.
3. Method for generating MR images of simulated inhomogeneous flexible biological tissue according to claim 2, characterized in that the discriminator D is used JSD The calculation formula of (2) is as follows:
wherein E x～p(x) Is a desire for x.
4. Method for generating MR images of simulated inhomogeneous flexible biological tissue according to claim 2, characterized in that the random noise p (x) a ) The vector length of a =1,2, … n is set to 128.
5. Method for generating an MR image of simulated inhomogeneous flexible biological tissue according to claim 2, characterized in that the loss function of the generator G is
The loss function->
Is->
And &>
Sum, said->
And &>
The calculation formula of (2) is as follows:
wherein
Computing the a-th sub-image x a And its array pseudo-sub-image module>
Pixel-level euclidean distance between, where a =1,2, … n, n being a natural number. />
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911217179.5A CN110895828B (en) | 2019-12-03 | 2019-12-03 | Model and method for generating MR (magnetic resonance) image simulating heterogeneous flexible biological tissue |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911217179.5A CN110895828B (en) | 2019-12-03 | 2019-12-03 | Model and method for generating MR (magnetic resonance) image simulating heterogeneous flexible biological tissue |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110895828A CN110895828A (en) | 2020-03-20 |
| CN110895828B true CN110895828B (en) | 2023-04-18 |
Family
ID=69787574
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911217179.5A Active CN110895828B (en) | 2019-12-03 | 2019-12-03 | Model and method for generating MR (magnetic resonance) image simulating heterogeneous flexible biological tissue |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110895828B (en) |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107784676A (en) * | 2017-09-20 | 2018-03-09 | 中国科学院计算技术研究所 | Compressed sensing calculation matrix optimization method and system based on autocoder network |
| CN107909621A (en) * | 2017-11-16 | 2018-04-13 | 深圳市唯特视科技有限公司 | It is a kind of based on it is twin into confrontation network medical image synthetic method |
| CN108038821A (en) * | 2017-11-20 | 2018-05-15 | 河海大学 | A kind of image Style Transfer method based on production confrontation network |
| CN108564611A (en) * | 2018-03-09 | 2018-09-21 | 天津大学 | A kind of monocular image depth estimation method generating confrontation network based on condition |
| CN109637634A (en) * | 2018-12-11 | 2019-04-16 | 厦门大学 | A kind of medical image synthetic method based on generation confrontation network |
| CN110097059A (en) * | 2019-03-22 | 2019-08-06 | 中国科学院自动化研究所 | Based on file and picture binary coding method, system, the device for generating confrontation network |
| CN110265117A (en) * | 2019-06-05 | 2019-09-20 | 深圳大学 | Medical image generation method and device |
| CN110503187A (en) * | 2019-07-26 | 2019-11-26 | 江苏大学 | A kind of implementation method of the generation confrontation network model generated for functional magnetic resonance imaging data |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10624558B2 (en) * | 2017-08-10 | 2020-04-21 | Siemens Healthcare Gmbh | Protocol independent image processing with adversarial networks |
| US10592779B2 (en) * | 2017-12-21 | 2020-03-17 | International Business Machines Corporation | Generative adversarial network medical image generation for training of a classifier |
-
2019
- 2019-12-03 CN CN201911217179.5A patent/CN110895828B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107784676A (en) * | 2017-09-20 | 2018-03-09 | 中国科学院计算技术研究所 | Compressed sensing calculation matrix optimization method and system based on autocoder network |
| CN107909621A (en) * | 2017-11-16 | 2018-04-13 | 深圳市唯特视科技有限公司 | It is a kind of based on it is twin into confrontation network medical image synthetic method |
| CN108038821A (en) * | 2017-11-20 | 2018-05-15 | 河海大学 | A kind of image Style Transfer method based on production confrontation network |
| CN108564611A (en) * | 2018-03-09 | 2018-09-21 | 天津大学 | A kind of monocular image depth estimation method generating confrontation network based on condition |
| CN109637634A (en) * | 2018-12-11 | 2019-04-16 | 厦门大学 | A kind of medical image synthetic method based on generation confrontation network |
| CN110097059A (en) * | 2019-03-22 | 2019-08-06 | 中国科学院自动化研究所 | Based on file and picture binary coding method, system, the device for generating confrontation network |
| CN110265117A (en) * | 2019-06-05 | 2019-09-20 | 深圳大学 | Medical image generation method and device |
| CN110503187A (en) * | 2019-07-26 | 2019-11-26 | 江苏大学 | A kind of implementation method of the generation confrontation network model generated for functional magnetic resonance imaging data |
Non-Patent Citations (4)
| Title |
|---|
| Changhee HAN.Combining Noise-to-Image and Image-to-Image GANs: Brain MR Image Augmentation for Tumor Detection.IEEE.2019,全文. * |
| Jun-Yan Zhu.Toward Multimodal Image-to-Image Translation.31st Conference on Neural Information Processing Systems (NIPS 2017).2017,全文. * |
| Yong Guo.Auto-Embedding Generative Adversarial Networks For High Resolution Image Synthesis.IEEE TRANSACTIONS ON MULTIMEDIA.2019,第21卷(第11期),全文. * |
| 陈锟 ; 乔沁 ; 宋志坚 ; .生成对抗网络在医学图像处理中的应用.生命科学仪器.2018,(Z1),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110895828A (en) | 2020-03-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2022267641A1 (en) | Image defogging method and system based on cyclic generative adversarial network | |
| CN113099208B (en) | Method and device for generating dynamic human body free viewpoint video based on nerve radiation field | |
| CN113240613B (en) | Image restoration method based on edge information reconstruction | |
| CN110659727B (en) | Sketch-based image generation method | |
| CN110599395B (en) | Target image generation method, device, server and storage medium | |
| CN111597946B (en) | Processing method of image generator, image generation method and device | |
| CN111260741B (en) | Three-dimensional ultrasonic simulation method and device by utilizing generated countermeasure network | |
| CN116309650B (en) | Medical image segmentation method and system based on double-branch embedded attention mechanism | |
| CN110363068B (en) | High-resolution pedestrian image generation method based on multiscale circulation generation type countermeasure network | |
| CN113379601B (en) | Real world image super-resolution method and system based on degradation variable-component self-encoder | |
| CN112347850A (en) | Infrared image conversion method, living body detection method, device and readable storage medium | |
| CN113344777B (en) | Face changing and replaying method and device based on three-dimensional face decomposition | |
| CN111797891A (en) | Unpaired heterogeneous face image generation method and device based on generation countermeasure network | |
| CN114581613A (en) | Trajectory constraint-based human body model posture and shape optimization method and system | |
| CN110827232A (en) | Cross-modal MRI (magnetic resonance imaging) synthesis method based on morphological feature GAN (gain) | |
| CN114926343A (en) | Image super-resolution method based on pyramid fusion attention network | |
| CN116402865A (en) | Multi-mode image registration method, device and medium using diffusion model | |
| WO2023109702A1 (en) | Substance decomposition method based on non-enhanced ct, intelligent terminal, and storage medium | |
| CN113160032A (en) | Unsupervised multi-mode image conversion method based on generation countermeasure network | |
| CN117437420A (en) | Cross-modal medical image segmentation method and system | |
| CN115049739A (en) | Binocular vision stereo matching method based on edge detection | |
| CN117094895B (en) | Image panorama stitching method and system | |
| CN110895828B (en) | Model and method for generating MR (magnetic resonance) image simulating heterogeneous flexible biological tissue | |
| CN113379606A (en) | Face super-resolution method based on pre-training generation model | |
| CN112509129A (en) | Spatial view field image generation method based on improved GAN 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |