CN108898140A - Brain tumor image segmentation algorithm based on improved full convolutional neural networks - Google Patents
Brain tumor image segmentation algorithm based on improved full convolutional neural networks Download PDFInfo
- Publication number
- CN108898140A CN108898140A CN201810590091.7A CN201810590091A CN108898140A CN 108898140 A CN108898140 A CN 108898140A CN 201810590091 A CN201810590091 A CN 201810590091A CN 108898140 A CN108898140 A CN 108898140A
- Authority
- CN
- China
- Prior art keywords
- fcnn
- algorithm
- brain tumor
- image
- network
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
- G06F18/2193—Validation; Performance evaluation; Active pattern learning techniques based on specific statistical tests
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Probability & Statistics with Applications (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
The present invention relates to a kind of MR brain tumor image partition methods based on improved full convolutional neural networks algorithm, and steps are as follows：The first step：Image preprocessing；Second step：Improved FCNN coarse segmentation algorithm：Based on FCNN network, batch regularization layer is added after each convolutional layer to accelerate the training speed of network, improve the precision of model, and Fusion Features three times are carried out to the brain tumor feature of Chi Huahou, to obtain finer brain tumor feature, improved FCNN network, i.e. FCNN-4s network are established；Third step：Blending algorithm is cut in the subdivision of FCNN-4s and CRF：According to the coarse segmentation result of FCNN-4s in second step, energy function obtains pixel original home label probability value in initialization CRF model, CRF model, two class probability maps of continuous iterated revision FCNN-4s prediction then are calculated using following steps, fusion results are cut in the subdivision for obtaining CRF.
Description
Technical field
The present invention is a key areas in field of medical imaging, and medical image and deep learning algorithm are combined
Come, completes the Accurate Segmentation of brain tumor nuclear magnetic resonance image.
Background technique
Malignant brain tumor is one of cancer types most fearful in the world, it will usually make decrease of cognitive function, the life of patient
Bioplasm quantitative change is poor.The most common brain tumor is primary central nervous system lymphoma and glioma in adult, wherein glioma
80% or more of malignant tumour is accounted for, therefore glioma is the key object of lesion segmentation.But since glioma may alternatively appear in greatly
Any position of brain and size is indefinite, in irregular shape, so that its segmentation is still a challenging task, therefore such as
Using modern information technologies, efficient and full-automatic Ground Split brain tumor becomes an important research direction for what.Nuclear magnetic resonance at
Picture (Magnetic Resonance Imaging, MRI) technology has Noninvasive, has to brain tumor image segmentation important
Booster action[3-4].Brain tumor MRI sequence includes T1 weighting (T1-weighted), T1C (Contrast enhanced T1-
Weighted images), T2 weighting (T2-weighted images) and FLAIR (Fluid Attenuated Inversion
The imaging sequences such as Recovery) clinically usually combine the position and size of four kinds of common diagnosing tumours of image.
Supervised learning algorithm based on training sample and label obtains a disaggregated model by study, realizes image with this
Classification or segmentation task.Be based especially on deep learning convolutional neural networks (Convolutional Neural Network,
CNN the computer vision fields such as image classification, target detection and semantic segmentation) are had been widely used for.With condition random field
(Conditional Random Fields, CRF), extreme random tree (Extremely Randomized Trees, ERT) and
The tradition supervision machine learning algorithm such as support vector machines (Support Vector Machine, SVM) is compared, and deep learning is based on
Method extract feature independent of artificial, but learn the complexity to high level and work transformation matrix from training data automatically
Feature.Image is carried out convolution with core characteristic parameter and is formed by CNN by stacking multiple convolutional layers, pond layer and full articulamentum
The feature learning model of strong robustness and adaptivity.Havaei etc.[1]Propose a kind of deep learning with two access CNN
Model, including a convolution access and a full connecting path；Pereira etc.[2]Using the deeper CNN structure of the number of plies, and mould
Network parameter is reduced in type with the convolution kernel of multiple 3 × 3 small sizes, improves arithmetic speed, strengthens mentioning for tumor boundaries information
It takes；Shi Dongli etc. combines fuzzy inference system, establishes learning rules and is judged the probability of CNN prediction tumour point to improve again
Segmentation precision.Although the partitioning algorithm based on deep learning makes substantial progress, the brain tumor image segmentation based on CNN exists
Following problems：(1) label of adjacent pixel blocks has independence, does not account for the correlation and consistency between label；(2)
The size of block of pixels limits the size of receptive field, and network can only extract local feature, causes CNN that can not be partitioned into fine
Tumor boundaries；(3) CNN is directed to each adjacent pixel blocks convolution one by one, causes to calculate upper redundancy with higher.
For the defect of CNN segmentation brain tumor, the present invention proposes a kind of based on improved full convolutional neural networks (Fully
Convolutional Neural Network, FCNN) and condition random field (Conditional Random Fields, CRF)
Full-automatic brain tumor MR image segmentation algorithm.Gray scale normalization is done to multi-modal MR brain tumor image first and gray level image melts
The pretreatment of conjunction then establishes FCNN model to pretreatment image and carries out coarse segmentation, then based between segmented image label
Probability spectrum model CRF is fused in FCNN, carries out fine boundary segmentation, improves the segmentation precision of brain tumor by correlation.
[1]Havaei M,DavyA,Wardefarley D,et al.Brain tumor segmentation with
Deep Neural Networks[J].Medical ImageAnalysis,2017,35:18-31.
[2]Pereira S,PintoA,Alves V,et al.Brain Tumor Segmentation Using
Convolutional Neural Networks in MRI Images[J].IEEE Transactions on Medical
Imaging,2016,35(5):1240-1251.
Summary of the invention
For existing machine learning and CNN the algorithm problem not high to brain tumor image segmentation precision, the present invention provides one
MR brain tumor image partition method of the kind based on improved full convolutional neural networks and condition random field, passes through improved full convolution
Neural network algorithm FCNN-4s obtains coarse segmentation as a result, then using in statistical learning probability CRF algorithm amendment coarse segmentation result
Brain tumor boundary.Technical solution is as follows：
A kind of MR brain tumor image partition method based on improved full convolutional neural networks algorithm, steps are as follows：
The first step：Image preprocessing
Gray scale normalization is carried out to tri- kinds of mode of brain tumor image FLAIR, T2 and T1C, it is then logical respectively as R, G, B
Road carry out simple gray level image fusion so as to convolution nuclear energy study to different modalities different characteristic, will be after gray scale merges
Pretreatment image as algorithm training and test data.
Second step：Improved FCNN coarse segmentation algorithm
Based on FCNN network, batch regularization layer is added after each convolutional layer to accelerate the training speed of network, improved
The precision of model, and Fusion Features three times are carried out to the brain tumor feature of Chi Huahou and are built with obtaining finer brain tumor feature
Improved FCNN network, i.e. FCNN-4s network are found, improved FCNN coarse segmentation algorithm steps are as follows：
1) behind 5 ponds, 1/2 size of original image, 1/4 ruler are respectively obtained to by the fused pretreatment image of gray scale
Very little, 1/8 size, 1/16 size and 1/32 size eigenmatrix；
2) 2 times of up-samplings are carried out to the high dimensional feature vector matrix of 1/32 size first, the feature vector square with Pool4 layers
Battle array carries out first time fusion；2 times of up-samplings then are carried out to the Fusion Features matrix obtained for the first time, the feature with Pool3 layers
Vector matrix carries out second and merges；And then this fusion feature matrix is subjected to 2 times of up-samplings, with Pool2 layers of feature to
Moment matrix carries out third time fusion；The eigenmatrix of third time fusion is then subjected to 4 times of up-samplings, is obtained and original brain tumor
The eigenmatrix of image same size, each pixel of last Prob layers of output are judged as tumour point and non-tumour point respectively
Two probability maps, obtain improved FCNN, i.e. FCNN-4s, coarse segmentation result；
Third step：Blending algorithm is cut in the subdivision of FCNN-4s and CRF
According to the coarse segmentation of FCNN-4s in second step as a result, to obtain pixel original for energy function in initialization CRF model
Belong to label probability value, then calculates CRF model, two class probability of continuous iterated revision FCNN-4s prediction using following steps
Fusion results are cut in map, the subdivision for obtaining CRF.
Preferably, the method for third step is as follows：
1)：The probability map for belonging to tumour and the probability map for being not belonging to tumour are filtered respectively with Gaussian filter
Wave obtains two class filter results；
2)：Filter result distribution weight in step 1) is subjected to multiply-add summation, then to the probability graph of each classification
Spectrum carries out conversion update according to label compatibility matrix and obtains a little to energy potential function；
3)：Unitary energy potential function is sought to the output of FCNN-4s, then integration step) point in 2 is to energy potential function
Seek entire energy function；
4)：3) probability value of pixel ownership label is sought into the entire energy function value normalization in, and presses maximum probability
Label corresponding to the pixel is taken, is moved back until probability value iteration that each pixel belongs to respective label reaches 90% or more
Algorithm recycles out, otherwise the pixel loss backpropagation currently calculated is returned in FCNN-4s algorithm to the study for continuing network parameter
With amendment, whether each pixel is tumour pixel after accurate judgement is good, and algorithm can complete the segmentation of brain tumor image.
MRI is the important supplementary means of cerebral disease clinical diagnosis, and the present invention proposes a kind of based on the improved of deep learning
The partitioning algorithm of full convolutional neural networks and condition random field, experiment show that the accurate of MR brain tumor image may be implemented in the algorithm
Point, average Dice value is up to 0.9129.With compared with algorithm, Dice, Precision and Recall of this algorithm have very big
Promotion, and algorithm stability is stronger, and the time that width brain tumor image cost is divided in prediction is also less, in average 1s just
The segmentation of achievable width brain tumor image.The present invention effectively supports the segmentation of brain tumor nuclear magnetic resonance image, is MRI
The development that advanced optimizes of computer aided measurement technology provides reference, and the mode to expert's manual measurement is to mend well
It fills.
Detailed description of the invention
Fig. 1 partitioning algorithm block diagram of the present invention；
The improved FCNN-4s coarse segmentation network structure of Fig. 2；
Extracted brain tumor feature (a) FCNN-8s (b) FCNN-4s (c) label of Fig. 3 heterogeneous networks structure；
The extracted brain tumor feature of Fig. 4 inventive algorithm and label figure (a) brain tumor characteristic pattern (b) label figure；
Fig. 5 algorithm figure, (a) FLAIR (b) T1C (c) T2 compared with the segmentation result on fusion modality images in single mode
(d) grayscale fusion image (e) single mode segmentation figure (f) merges mode segmentation figure (g) label；
The segmentation result of five kinds of algorithms of Fig. 6 compares figure, (a) FLAIR (b) T1C (c) T2 (d) grayscale fusion image
(e) Havaei algorithm (f) Pereira algorithm (g) FCNN-8s (h) FCNN-4s
(I) inventive algorithm (k) label.
Specific embodiment
The brain tumor image segmentation algorithm process of the improved full convolutional neural networks of the present invention is as shown in Figure 1.First to returning
The one multi-modal MR brain tumor image data changed carries out gray level image fusion；Then using the partial data of fusion as training set,
Partial data utilizes the blending algorithm of training set training FCNN and CRF as test set；Model is finally carried out on test set
The metrics evaluation of test and segmentation result.
1) image preprocessing
Original each pixel of MR brain tumor image is 16bit, but is all pair in digital image processing techniques
8bit image is handled, therefore the present invention first carries out gray scale normalization to MR image, and the gray value of each pixel is uniformly pressed
It is reduced between range 0-255.In view of neural network is only not enough to accurately by the feature that single modality images data are extracted
Ground Split brain tumor boundary, so according to the imaging characteristics of tetra- kinds of modality images of MR, to tri- kinds of mode difference of FLAIR, T2 and T1C
Simple gray level image fusion is carried out as R, G, channel B so as to the different characteristic of convolution nuclear energy study to different modalities.
2) improved FCNN-4s coarse segmentation algorithm
In order to obtain good tumoral character, fine boundary segmentation is carried out convenient for probability map CRF model.The present invention couple
Original FCNN algorithm improves, and proposes fine Fusion Features model, while in order to accelerate the convergence rate of network, improving
Batch regularization (Batch Normalization, BN) layer is added in network by the precision of model, is formed shown in Fig. 2
FCNN-4s algorithm.Wherein Data is data Layer, and Conv is convolutional layer, and Pool is pond layer, and Up is up-sampling layer, and Prob is defeated
Probability map layer out.Fuse is characterized mixing operation, and fusion is in the dimension of high dimensional feature matrix and low-dimensional eigenmatrix and greatly
It is small it is identical under the premise of, the data that two objective matrixs correspond to dimension are subjected to simple sum operation.Before Crop is indicated if merged
The dimension of two objective matrixs and it is not of uniform size when (Pool4 layers of output size are 28 × 28 × 512 in such as table 1, and Up_
Score1 layers of output size are 18 × 18 × 2), algorithm is to cut two objective matrix dimensions and matrix that is in the same size and carrying out
Operation.When dimension is inconsistent, algorithm is by low-dimensional eigenmatrix with 1 × 1 × Nhigh (Nhigh is high dimensional feature matrix quantity)
Convolution kernel carry out convolution so that the size of low-dimensional eigenmatrix value will not be changed while two matrix dimensions are identical again；When big
When small inconsistent, because the characteristic information of image focuses primarily upon middle section, algorithm is extracted from the intermediate of low-dimensional eigenmatrix
Size identical with high dimensional feature matrix.
Table 1 is the design parameter of network, the name that wherein Layer is each layer；Kernel indicates convolution kernel or Chi Huahe
Dimension and size；Stride indicates the step-length of sliding window during convolution algorithm；Pad indicate calculate before to input matrix into
The certain edge of row expands；Output-Size indicates the dimension and size of output eigenmatrix.Algorithm is by the spy on different dimensions
Sign matrix is merged three times, keeps the brain tumor feature of acquisition finer, while the BN layers of parameter that can make each repetitive exercise
Weight distribution does not change a lot to accelerate convergence rate.
Table 1
In order to verify the superiority for improving network, compared fusion, the FCNN-8s network of feature and fusion are special three times twice
The extracted brain tumor characteristic effect of the FCNN-4s network of sign.Fig. 3 illustrates the extracted final brain tumor feature letter of each network
Breath, yellow degree is deeper, and representative is that the probability of brain tumor is bigger, and the blue degree the deep then bigger for the probability of background.Thus may be used
To find out, the tumour general outline for merging the FCNN-4s ratio FCNN-8s of more features information is more obvious, is more advantageous to subsequent
The boundary segmentation of CRF model.
3) blending algorithm is cut in the subdivision of FCNN-4s and CRF
The brain tumor boundary that FCNN-4s is obtained not is very well, the brain tumor boundary after segmentation to be corrected with CRF with this.
Judge in MR brain tumor image whether each pixel is tumour point, indicates non-tumour pixel with 0 in label l, 1 indicates swollen
Tumor pixel belongs to two classification problems.After getting two probability maps of FCNN-4s, energy letter is initialized first
Number obtains pixel original home label probability value, then calculates CRF model, continuous iterated revision FCNN- using following steps
Two class probability maps of 4s prediction, blending algorithm are as follows.
Step 1 is respectively filtered two class probability maps of l=0 and l=1 with Gaussian filter, obtains two classes
Filter result；
Filter result distribution weight in step 1 is carried out multiply-add summation by step 2, then to the probability graph of each classification
Spectrum carries out conversion update according to label compatibility matrix and obtains a little to energy potential function；
Step 3 seeks unitary energy potential function to the output of FCNN-4s network, and then the point in integration step 2 is to energy
Potential function result seeks entire energy function；
The probability value of pixel ownership label is sought in the normalization of entire energy function value obtained in step 3 by step 4, and is pressed
Maximum probability takes label corresponding to the pixel, until the probability value iteration that each pixel belongs to respective label reaches
90% or more exits algorithm circulation, otherwise returns in FCNN-4s algorithm the pixel loss backpropagation currently calculated and continues network
The study and amendment of parameter.
After improved FCNN-4s network and CRF algorithm fusion, the present invention is extracted the characteristic pattern of brain tumor again
Picture improves the rough segmentation of FCNN-4s algorithm figure 4, it is seen that algorithm brain tumor proposed by the present invention boundary is finer
It cuts as a result, the precision divided is higher.
4) comparison and analysis of experimental result
For difference of the algorithm more proposed by the present invention on single mode image and blending image segmentation effect, brain is chosen
The obvious FLAIR single mode image of tumor boundaries is split in result with the pretreatment image merged using gray level image
Comparison.It can be seen that from experimental result Fig. 5, when dividing FLAIR single mode image with inventive algorithm, can correctly be partitioned into
The general outline of brain tumor, but there is over-segmentation phenomenon there are a large amount of scatterplot in non-tumor region；When segmentation gray scale merges mould
When state image, since the tumor information of three kinds of mode can be complementary to one another, the extraction of the features such as boundary is strengthened, so that brain tumor
Profile the phenomenon that not only being segmented correctly out, decreasing over-segmentation to a certain extent, the tumor boundaries of segmentation are also more
Add fine.
Then using the quality of three kinds of index quantification analysis segmentation results, as can be seen from Table 2, multi-modal fusion image is opposite
Single mode image has larger promotion on the segmentation precisions such as index of similarity, sensitivity and positive prediction rate, especially
Dice improves 6.33%.
Table 2
In order to verify the superiority of mentioned innovatory algorithm, the present invention and the simple division algorithm for only carrying out Fusion Features twice
FCNN-8s and without merge CRF FCNN-4s algorithm be compared, while test also with famous scholar Havaei and
Traditional CNN algorithm of the propositions such as Pereira compares, and as can be seen from Figure 6, there are apparent mistakes for the algorithm that Havaei is proposed
Segmentation, the brain tumor boundary of segmentation is unobvious and there are more isolated scatterplots；The algorithm segmentation performance that Pereira is proposed is whole
It is slightly promoted compared with Havaei, and reduces the over-segmentation of brain tumor to a certain extent；Although FCNN-8s obtains smooth brain
Lesion segmentation profile, but lead to since Fusion Features are less that brain tumor boundary is not fine enough and segmentation precision is not high； FCNN-4s
There is certain enhancing on obtaining brain tumor boundary information for opposite FCNN-8s, but the brain tumor boundary that is partitioned into or inadequate
It is fine and smooth.Especially when dividing the 3rd complicated brain tumor image, due to the too complicated segmentation result for leading to various algorithms of tumor boundaries
It is all barely satisfactory, but satisfied result can be partitioned into using inventive algorithm.Generally speaking, more using Fusion Features
Add fine FCNN-4s model, and merges the end-to-end algorithm structure that CRF is formed and similar pixel is enabled to obtain phase
Same label effectively solves the problems, such as over-segmentation and the less divided of brain tumor to refine the boundary of brain tumor.
In addition from table 3 it is also seen that inventive algorithm is with respect to other algorithms segmentation precision with higher, Dice compared with
5.47% and 3.60% has been respectively increased in the algorithm that two scholars propose, the FCNN algorithm compared with amixis CRF is even more average improves
8.52%.Again from the point of view of the average time for dividing a width brain tumor image, inventive algorithm reality with higher in prediction
Shi Xing can averagely complete the segmentation of a width brain tumor image in 1 second.
Table 3.
Claims (2)
1. a kind of MR brain tumor image partition method based on improved full convolutional neural networks algorithm, steps are as follows：
The first step：Image preprocessing
Gray scale normalization is carried out to tri- kinds of mode of brain tumor image FLAIR, T2 and T1C, is then carried out respectively as R, G, channel B
Simple gray level image fusion will pass through the fused pre- place of gray scale so that the study of convolution nuclear energy is to the different characteristic of different modalities
Manage data of the image as algorithm training and test.
Second step：Improved FCNN coarse segmentation algorithm
Based on FCNN network, batch regularization layer is added after each convolutional layer to accelerate the training speed of network, improve model
Precision, and Fusion Features three times are carried out to the brain tumor feature of Chi Huahou, to obtain finer brain tumor feature, foundation changes
Into FCNN network, i.e. FCNN-4s network, improved FCNN coarse segmentation algorithm steps are as follows：
1) to behind 5 ponds, respectively obtained by the fused pretreatment image of gray scale 1/2 size of original image, 1/4 size,
The eigenmatrix of 1/8 size, 1/16 size and 1/32 size；
2) 2 times of up-samplings are carried out to the high dimensional feature vector matrix of 1/32 size first, with Pool4 layers of eigenvectors matrix into
Row merges for the first time；2 times of up-samplings then are carried out to the Fusion Features matrix obtained for the first time, the feature vector with Pool3 layers
Matrix carries out second and merges；And then this fusion feature matrix is subjected to 2 times of up-samplings, the feature vector square with Pool2 layers
Battle array carries out third time fusion；The eigenmatrix of third time fusion is then subjected to 4 times of up-samplings, is obtained and original brain tumor image
The eigenmatrix of same size, each pixel of last Prob layers of output are judged as the two of tumour point and non-tumour point respectively
Probability map, obtains improved FCNN, i.e. FCNN-4s, coarse segmentation result；
Third step：Blending algorithm is cut in the subdivision of FCNN-4s and CRF
According to the coarse segmentation of FCNN-4s in second step as a result, energy function obtains pixel original home in initialization CRF model
Label probability value, then using following steps calculating CRF model, the two class probability maps that continuous iterated revision FCNN-4s is predicted,
Fusion results are cut in the subdivision for obtaining CRF.
2. partitioning algorithm according to claim 1, which is characterized in that the method for third step is as follows：
1)：The probability map for belonging to tumour and the probability map for being not belonging to tumour are filtered respectively with Gaussian filter, obtained
To two class filter results；
2)：Filter result distribution weight in step 1) is subjected to multiply-add summation, then to the probability map root of each classification
Conversion update is carried out according to label compatibility matrix to obtain a little to energy potential function；
3)：Unitary energy potential function is sought to the output of FCNN-4s, then integration step) point in 2 seeks energy potential function
Entire energy function；
4)：3) probability value of pixel ownership label is sought into the entire energy function value normalization in, and takes this by maximum probability
Label corresponding to pixel exits calculation until probability value iteration that each pixel belongs to respective label reaches 90% or more
Method circulation, otherwise by the pixel loss backpropagation currently calculated return FCNN-4s algorithm in continue network parameter study with repair
Just, whether each pixel is tumour pixel after accurate judgement is good, and algorithm can complete the segmentation of brain tumor image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810590091.7A CN108898140A (en) | 2018-06-08 | 2018-06-08 | Brain tumor image segmentation algorithm based on improved full convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810590091.7A CN108898140A (en) | 2018-06-08 | 2018-06-08 | Brain tumor image segmentation algorithm based on improved full convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108898140A true CN108898140A (en) | 2018-11-27 |
Family
ID=64344645
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810590091.7A Pending CN108898140A (en) | 2018-06-08 | 2018-06-08 | Brain tumor image segmentation algorithm based on improved full convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108898140A (en) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109840913A (en) * | 2019-01-21 | 2019-06-04 | 中南民族大学 | The method and system of lump segmentation in a kind of mammography X |
| CN109902748A (en) * | 2019-03-04 | 2019-06-18 | 中国计量大学 | A kind of image, semantic dividing method based on the full convolutional neural networks of fusion of multi-layer information |
| CN110084823A (en) * | 2019-04-18 | 2019-08-02 | 天津大学 | Three-dimensional brain tumor image partition method based on cascade anisotropy FCNN |
| CN110120033A (en) * | 2019-04-12 | 2019-08-13 | 天津大学 | Based on improved U-Net neural network three-dimensional brain tumor image partition method |
| CN110120048A (en) * | 2019-04-12 | 2019-08-13 | 天津大学 | In conjunction with the three-dimensional brain tumor image partition method for improving U-Net and CMF |
| CN110245747A (en) * | 2019-06-21 | 2019-09-17 | 华中师范大学 | Image processing method and device based on full convolutional neural networks |
| CN110349170A (en) * | 2019-07-13 | 2019-10-18 | 长春工业大学 | A kind of full connection CRF cascade FCN and K mean value brain tumor partitioning algorithm |
| CN110706214A (en) * | 2019-09-23 | 2020-01-17 | 东南大学 | Three-dimensional U-Net brain tumor segmentation method fusing condition randomness and residual error |
| CN110992320A (en) * | 2019-11-22 | 2020-04-10 | 电子科技大学 | Medical image segmentation network based on double interleaving |
| CN111127487A (en) * | 2019-12-27 | 2020-05-08 | 电子科技大学 | Real-time multi-tissue medical image segmentation method |
| CN111414922A (en) * | 2019-01-07 | 2020-07-14 | 阿里巴巴集团控股有限公司 | Feature extraction method, image processing method, model training method and device |
| CN111626913A (en) * | 2019-02-27 | 2020-09-04 | 顺丰科技有限公司 | Image processing method, device and storage medium |
| CN111667499A (en) * | 2020-06-05 | 2020-09-15 | 济南博观智能科技有限公司 | Image segmentation method, device and equipment for traffic signal lamp and storage medium |
| CN111914948A (en) * | 2020-08-20 | 2020-11-10 | 上海海事大学 | Ocean current machine blade attachment self-adaptive identification method based on rough and fine semantic segmentation network |
| CN111932447A (en) * | 2020-08-04 | 2020-11-13 | 中国建设银行股份有限公司 | Picture processing method, device, equipment and storage medium |
| CN112308077A (en) * | 2020-11-02 | 2021-02-02 | 中科麦迪人工智能研究院(苏州)有限公司 | Sample data acquisition method, image segmentation method, device, equipment and medium |
| CN112315451A (en) * | 2020-11-30 | 2021-02-05 | 沈阳航空航天大学 | Brain tissue segmentation method based on image clipping and convolutional neural network |
| CN112529915A (en) * | 2020-12-17 | 2021-03-19 | 山东大学 | Brain tumor image segmentation method and system |
| CN112767305A (en) * | 2020-12-15 | 2021-05-07 | 首都医科大学附属北京儿童医院 | Ultrasonic cardiogram identification method and device for congenital heart disease |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106600571A (en) * | 2016-11-07 | 2017-04-26 | 中国科学院自动化研究所 | Brain tumor automatic segmentation method through fusion of full convolutional neural network and conditional random field |
| CN106803256A (en) * | 2017-01-13 | 2017-06-06 | 深圳市唯特视科技有限公司 | A kind of 3D shape based on projection convolutional network is split and semantic marker method |
| CN107016681A (en) * | 2017-03-29 | 2017-08-04 | 浙江师范大学 | Brain MRI lesion segmentation approach based on full convolutional network |
| CN107527044A (en) * | 2017-09-18 | 2017-12-29 | 北京邮电大学 | A kind of multiple car plate clarification methods and device based on search |
| CN107729819A (en) * | 2017-09-22 | 2018-02-23 | 华中科技大学 | A kind of face mask method based on sparse full convolutional neural networks |
| CN107749061A (en) * | 2017-09-11 | 2018-03-02 | 天津大学 | Based on improved full convolutional neural networks brain tumor image partition method and device |
| CN108062756A (en) * | 2018-01-29 | 2018-05-22 | 重庆理工大学 | Image, semantic dividing method based on the full convolutional network of depth and condition random field |
-
2018
- 2018-06-08 CN CN201810590091.7A patent/CN108898140A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106600571A (en) * | 2016-11-07 | 2017-04-26 | 中国科学院自动化研究所 | Brain tumor automatic segmentation method through fusion of full convolutional neural network and conditional random field |
| CN106803256A (en) * | 2017-01-13 | 2017-06-06 | 深圳市唯特视科技有限公司 | A kind of 3D shape based on projection convolutional network is split and semantic marker method |
| CN107016681A (en) * | 2017-03-29 | 2017-08-04 | 浙江师范大学 | Brain MRI lesion segmentation approach based on full convolutional network |
| CN107749061A (en) * | 2017-09-11 | 2018-03-02 | 天津大学 | Based on improved full convolutional neural networks brain tumor image partition method and device |
| CN107527044A (en) * | 2017-09-18 | 2017-12-29 | 北京邮电大学 | A kind of multiple car plate clarification methods and device based on search |
| CN107729819A (en) * | 2017-09-22 | 2018-02-23 | 华中科技大学 | A kind of face mask method based on sparse full convolutional neural networks |
| CN108062756A (en) * | 2018-01-29 | 2018-05-22 | 重庆理工大学 | Image, semantic dividing method based on the full convolutional network of depth and condition random field |
Non-Patent Citations (1)
| Title |
|---|
| BING SHUAI 等: "Improving Fully Convolution Network for Semantic Segmentation", 《ARXIV》 * |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111414922B (en) * | 2019-01-07 | 2022-11-15 | 阿里巴巴集团控股有限公司 | Feature extraction method, image processing method, model training method and device |
| CN111414922A (en) * | 2019-01-07 | 2020-07-14 | 阿里巴巴集团控股有限公司 | Feature extraction method, image processing method, model training method and device |
| CN109840913A (en) * | 2019-01-21 | 2019-06-04 | 中南民族大学 | The method and system of lump segmentation in a kind of mammography X |
| CN109840913B (en) * | 2019-01-21 | 2020-12-29 | 中南民族大学 | Method and system for segmenting tumor in mammary X-ray image |
| CN111626913B (en) * | 2019-02-27 | 2024-04-05 | 顺丰科技有限公司 | Image processing method, device and storage medium |
| CN111626913A (en) * | 2019-02-27 | 2020-09-04 | 顺丰科技有限公司 | Image processing method, device and storage medium |
| CN109902748A (en) * | 2019-03-04 | 2019-06-18 | 中国计量大学 | A kind of image, semantic dividing method based on the full convolutional neural networks of fusion of multi-layer information |
| CN110120048B (en) * | 2019-04-12 | 2023-06-06 | 天津大学 | Three-dimensional brain tumor image segmentation method combining improved U-Net and CMF |
| CN110120048A (en) * | 2019-04-12 | 2019-08-13 | 天津大学 | In conjunction with the three-dimensional brain tumor image partition method for improving U-Net and CMF |
| CN110120033A (en) * | 2019-04-12 | 2019-08-13 | 天津大学 | Based on improved U-Net neural network three-dimensional brain tumor image partition method |
| CN110084823A (en) * | 2019-04-18 | 2019-08-02 | 天津大学 | Three-dimensional brain tumor image partition method based on cascade anisotropy FCNN |
| CN110245747A (en) * | 2019-06-21 | 2019-09-17 | 华中师范大学 | Image processing method and device based on full convolutional neural networks |
| CN110245747B (en) * | 2019-06-21 | 2021-10-19 | 华中师范大学 | Image processing method and device based on full convolution neural network |
| CN110349170A (en) * | 2019-07-13 | 2019-10-18 | 长春工业大学 | A kind of full connection CRF cascade FCN and K mean value brain tumor partitioning algorithm |
| CN110706214A (en) * | 2019-09-23 | 2020-01-17 | 东南大学 | Three-dimensional U-Net brain tumor segmentation method fusing condition randomness and residual error |
| CN110706214B (en) * | 2019-09-23 | 2022-06-17 | 东南大学 | Three-dimensional U-Net brain tumor segmentation method fusing condition randomness and residual error |
| CN110992320A (en) * | 2019-11-22 | 2020-04-10 | 电子科技大学 | Medical image segmentation network based on double interleaving |
| CN111127487B (en) * | 2019-12-27 | 2022-04-19 | 电子科技大学 | Real-time multi-tissue medical image segmentation method |
| CN111127487A (en) * | 2019-12-27 | 2020-05-08 | 电子科技大学 | Real-time multi-tissue medical image segmentation method |
| CN111667499A (en) * | 2020-06-05 | 2020-09-15 | 济南博观智能科技有限公司 | Image segmentation method, device and equipment for traffic signal lamp and storage medium |
| CN111932447A (en) * | 2020-08-04 | 2020-11-13 | 中国建设银行股份有限公司 | Picture processing method, device, equipment and storage medium |
| CN111932447B (en) * | 2020-08-04 | 2024-03-22 | 中国建设银行股份有限公司 | Picture processing method, device, equipment and storage medium |
| CN111914948A (en) * | 2020-08-20 | 2020-11-10 | 上海海事大学 | Ocean current machine blade attachment self-adaptive identification method based on rough and fine semantic segmentation network |
| CN112308077A (en) * | 2020-11-02 | 2021-02-02 | 中科麦迪人工智能研究院(苏州)有限公司 | Sample data acquisition method, image segmentation method, device, equipment and medium |
| WO2022089267A1 (en) * | 2020-11-02 | 2022-05-05 | 中科麦迪人工智能研究院(苏州)有限公司 | Sample data acquisition method, image segmentation method and apparatus, device and medium |
| CN112315451A (en) * | 2020-11-30 | 2021-02-05 | 沈阳航空航天大学 | Brain tissue segmentation method based on image clipping and convolutional neural network |
| CN112767305A (en) * | 2020-12-15 | 2021-05-07 | 首都医科大学附属北京儿童医院 | Ultrasonic cardiogram identification method and device for congenital heart disease |
| CN112767305B (en) * | 2020-12-15 | 2024-03-08 | 首都医科大学附属北京儿童医院 | Method and device for identifying echocardiography of congenital heart disease |
| CN112529915A (en) * | 2020-12-17 | 2021-03-19 | 山东大学 | Brain tumor image segmentation method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108898140A (en) | Brain tumor image segmentation algorithm based on improved full convolutional neural networks | |
| Zhang et al. | Attention gate resU-Net for automatic MRI brain tumor segmentation | |
| Mlynarski et al. | 3D convolutional neural networks for tumor segmentation using long-range 2D context | |
| Noori et al. | Attention-guided version of 2D UNet for automatic brain tumor segmentation | |
| CN111192245B (en) | Brain tumor segmentation network and method based on U-Net network | |
| Chang et al. | A mix-pooling CNN architecture with FCRF for brain tumor segmentation | |
| Liu et al. | A cascaded dual-pathway residual network for lung nodule segmentation in CT images | |
| Li et al. | Dilated-inception net: multi-scale feature aggregation for cardiac right ventricle segmentation | |
| Xie et al. | Automatic segmentation of dermoscopy images using self-generating neural networks seeded by genetic algorithm | |
| Bai et al. | Liver tumor segmentation based on multi-scale candidate generation and fractal residual network | |
| Cinar et al. | A hybrid DenseNet121-UNet model for brain tumor segmentation from MR Images | |
| Wu et al. | Automatic glioma segmentation based on adaptive superpixel | |
| CN109360208A (en) | A kind of medical image cutting method based on one way multitask convolutional neural networks | |
| CN110120048A (en) | In conjunction with the three-dimensional brain tumor image partition method for improving U-Net and CMF | |
| CN111709952A (en) | MRI brain tumor automatic segmentation method based on edge feature optimization and double-flow decoding convolutional neural network | |
| Weninger et al. | Multi-task learning for brain tumor segmentation | |
| Mohammed et al. | A survey of methods for brain tumor segmentation-based MRI images | |
| Li et al. | Brain tumor segmentation using 3D generative adversarial networks | |
| Chen et al. | Skin lesion segmentation using recurrent attentional convolutional networks | |
| Barzegar et al. | WLFS: Weighted label fusion learning framework for glioma tumor segmentation in brain MRI | |
| Zhao et al. | Al-net: Attention learning network based on multi-task learning for cervical nucleus segmentation | |
| Cetin et al. | Multiple sclerosis lesion detection in multimodal MRI using simple clustering-based segmentation and classification | |
| Banerjee et al. | A CADe system for gliomas in brain MRI using convolutional neural networks | |
| Soleymanifard et al. | Segmentation of whole tumor using localized active contour and trained neural network in boundaries | |
| Le et al. | Offset curves loss for imbalanced problem in medical segmentation |
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 |
Application publication date: 20181127 |
|
| RJ01 | Rejection of invention patent application after publication |