CN105719303B - Magnetic resonance prostate 3D rendering dividing method based on more depth belief networks - Google Patents
Magnetic resonance prostate 3D rendering dividing method based on more depth belief networks Download PDFInfo
- Publication number
- CN105719303B CN105719303B CN201610048013.5A CN201610048013A CN105719303B CN 105719303 B CN105719303 B CN 105719303B CN 201610048013 A CN201610048013 A CN 201610048013A CN 105719303 B CN105719303 B CN 105719303B
- Authority
- CN
- China
- Prior art keywords
- network
- test set
- grader
- result
- obtains
- 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
Classifications
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30081—Prostate
Abstract
The invention discloses a kind of magnetic resonance MRI prostate 3D rendering dividing methods based on more depth belief networks.The problem of prior art is mainly solved because of artificial selection feature, causes segmentation precision low.Its implementation is:The data set of each patient is obtained according to patient image, each patient data set is divided into three parts；The identical depth belief network of three structures and three softmax graders is respectively trained with this three parts training set；Three test sets are input in three networks；Split with output of three softmax graders to three networks, and the segmentation result of three test sets is sequentially overlapped, obtain the first segmentation result of test set；Segmentation result at the beginning of test set is handled using three dimensional morphology, obtains final segmentation result.The present invention can efficiently use the information characteristics of magnetic resonance MRI sequence image levels without artificial selection, improve the accuracy rate of segmentation.
Description
Technical field
The invention belongs to a kind of technical field of image processing, more particularly to a kind of magnetic resonance three-dimensional image partition method can
For the processing to medical image.
Background technology
Prostate cancer is one of most common malignant tumour of male, is ranked in the developed country such as America and Europe prostate-cancer incidence
First, the death rate is only second to lung cancer.In recent years as China enters aged society, prostate-cancer incidence shows a rising trend.
Prostate surrounding tissue organ is numerous and complicated, and prostate is as the chestnut of a back-off, like circular base position
Above prostate, it is pointed downward., close to bladder, front is pubis for it, and rear is close to rectum.Surrounding tissue and prostate
It is adhered serious.Image partition method is the scientific basic of diagnoses and treatment prostatic disorders, and diagnosis of prostate disease treatment need to
One of key technology to be solved.
At present, 2D features, magnetic resonance MRI prostates are mainly used for the dividing method of magnetic resonance MRI prostate images
Image is sequence image, and 2D features cannot utilize the information between image levels, and the main feature using artificial selection is in 2D
On split.This artificial selection relies primarily on the experience of people and a large amount of repetition experiments obtain, and a good feature is almost
Looked for the strength of whole team, to take a substantial amount of time and grope just obtain repeatedly.Since medical image is complicated more
Become, personalized strong, the feature by artificial selection is difficult to be suitable for all complicated medical images, influences the accurate of clinical diagnosis
Property, so being badly in need of wanting a kind of more flexible robust, the accurate dividing method of science in face of complicated magnetic resonance MRI image.
The content of the invention
It is an object of the invention to the deficiency for above-mentioned prior art, proposes a kind of magnetic based on more depth belief networks
Resonate prostate 3D rendering dividing method, extracts feature from image in an adaptive way, avoids lacking for artificial selection feature
Fall into, improve the magnetic resonance MRI prostates image segmentation accuracy rate of complexity.
The present invention thinking be：By using sequence 3D rendering block as input data, more depth belief networks, energy are constructed
It is enough adaptively to extract characteristics of image, and then improve forefront MRI image segmentation effect；The three-dimensional segmentation result energy that the present invention obtains
Enough help doctor to have body of prostate more intuitively to recognize, improve the accuracy of clinical diagnosis and treatment.
According to above-mentioned thinking, technical solution of the present invention includes as follows：
(1) n × m width magnetic resonance MRI prostate images are obtained, wherein n=10 represents patient's number, and m≤50 represent each
Patient is up to 50 width images；
(2) all pixels point of each image is pre-processed, will be by 13 × 13 × 3 centered on each pixel
Cuboid be converted into the row vector of 1 × 507 dimension, obtain patient data set be made of m × l row vector, wherein l expressions are often
Pixel number in piece image, l≤7200；
(3) training set is constructed：
9 patients are chosen from 10 patients, before selected each patientA the first instruction of row vector composition
Practice collection, it is middleA row vector forms the second training set, afterA row vector forms the 3rd training set；
By remaining 1 patient, it is divided into three test sets according still further to as above method, is respectively the first test set, the second survey
Examination collection and the 3rd test set；
(4) depth belief network is built, is learnt to obtain network one respectively with the first training set, is learnt with the second training set
To network two, learn to obtain network three with the 3rd training set, these three networks have identical structure；
(5) train a softmax grader to obtain grader one with network one, a softmax is trained with network two
Grader obtains grader two, trains a softmax grader to obtain grader three with network three；
(6) test set is split：
Its result input grader one is calculated by the first test set input network one, obtains point of the first test set
Cut result；
Its result input grader two is calculated by the second test set input network two, obtains point of the second test set
Cut result；
Its result input grader three is calculated by the 3rd test set input network three, obtains point of the 3rd test set
Cut result；
The segmentation result of three test sets is successively merged, obtains the segmentation result of test set；
(7) three dimensional morphology is carried out to test set segmentation result to handle to obtain final segmentation result.
The present invention has the following advantages compared with prior art：
1. the present invention carries out depth abstractdesription using depth belief network to data, image itself can be made full use of to believe
Breath, without artificial selection feature, adaptivity is good, improves segmentation effect.
2. the present invention uses input of the 3D rendering block as depth belief network, magnetic resonance MRI sequences can be more effectively utilized
The information of image levels.
Brief description of the drawings
Fig. 1 is that the present invention realizes general flow chart；
Fig. 2 is the present invention to testing patient's magnetic resonance MRI prostate image segmentation figures.
Specific implementation method
Referring to the drawings 1, step is as follows for of the invention realizing：
Step 1. obtains patient image.
N × m width magnetic resonance MRI prostate images are obtained from hospital, wherein n=10 represents patient's number, and m≤50 represent every
A patient is up to 50 width images.
Step 2. obtains patient data set according to patient image.
2a) obtain 13 × 13 × 3 cuboid centered on each pixel per piece image；
Each cuboid 2b) is converted into the row vector of 1 × 507 dimension, obtains every patient by m × l row vector group
Into data set, wherein l represented per the pixel number in piece image, l≤7200.
Step 3. constructs training set and test set.
3a) choose 9 patients in ten patients；
Before 3b) selected patient data is concentratedA row vector forms the first training set, middleIt is a
Row vector forms the second training set, afterA row vector forms the 3rd training set；
Before 3c) remaining patient data is concentratedA row vector forms the first test set, middleA row
Vector the second test set of composition, afterA row vector forms the 3rd test set.
Step 4. trains the identical depth belief network of three parameters.
Three depth belief networks 4a) are built, i.e., are sequentially overlapped completion using three limited Boltzmann machines, three limited
The number of nodes of Boltzmann machine is respectively 500,1000,500；
The learning rate for 4b) setting three depth belief networks is 0.01, and initial dynamical learning rate is 0.5, stablizes dynamic
Learning rate is 0.9；
4c) use is limited Boltzmann machine to each layer of sdpecific dispersion Algorithm for Training；
4d) using Back Propagation Algorithm fine setting entire depth belief network；
4e) use input of first training set as first depth belief network, use 4c) arrive 4d) method to
One depth belief network is trained, and obtains network one；
4f) use input of second training set as second depth belief network, use 4c) arrive 4d) method to
Two depth belief networks are trained, and obtain network two；
4g) use input of the 3rd training set as the 3rd depth belief network, use 4c) arrive 4d) method to
Three depth belief networks are trained, and obtain network three；
Step 5. trains three softmax graders.
The error of gradient descent algorithm minimization softmax graders 5a) is used, tries to achieve the optimal of softmax graders
Parameter, completes the training to softmax graders；
The first training set 5b) is input to network one and obtains the output of network one, first softmax is used as by the use of this output
The input of grader, uses 5a) method first softmax grader is trained to obtain grader one；
The second training set 5c) is input to network two and obtains the output of network two, second softmax is used as by the use of this output
The input of grader, uses 5a) method second softmax grader is trained to obtain grader two；
The 3rd training set 5d) is input to network three and obtains the output of network three, the 3rd softmax is used as by the use of this output
The input of grader, uses 5a) method the 3rd softmax grader is trained to obtain grader three.
Step 6. splits test set.
6a) its result input grader one is calculated by the first test set input network one, obtains the first test set
Segmentation result；
6b) its result input grader two is calculated by the second test set input network two, obtains the second test set
Segmentation result；
6c) its result input grader three is calculated by the 3rd test set input network three, obtains the 3rd test set
Segmentation result；
6d) result of three test sets is sequentially overlapped to obtain the first segmentation result of test set.
Step 7. carries out three dimensional morphology processing to first segmentation result.
The first segmentation result of test set is carried out three-dimensional expansion and handles to obtain expansion process result by (7a)；
Expansion process result is carried out three-dimensional corrosion treatment and obtains corrosion treatment result by (7b)；
Corrosion treatment result is carried out three-dimensional expansion and handles to obtain final segmentation result by (7c) again.
In conclusion the magnetic resonance MRI prostate 3D renderings segmentation side proposed by the present invention based on more depth belief networks
Method, the data set of each patient is obtained according to patient image, each patient data set is divided into three parts, and instructed with this three parts
Practice collection and the identical depth belief network of three parameters and three softmax graders are respectively trained, three test sets are input to
In three networks, split with output of three softmax graders to three networks, obtain the segmentation knot of three test sets
Fruit obtains the segmentation result of test set after being sequentially overlapped, finally test set segmentation result is handled using three dimensional morphology,
Segmentation result to the end is obtained, as shown in Figure 2.Fig. 2 (a) is the segmentation result of the present invention in its figure, and Fig. 2 (b) is medical practitioner
Manual segmentation result, Fig. 2 (c) are the fusion schematic diagram of segmentation result of the present invention and doctor's craft segmentation result.
The present invention can be good at splitting magnetic resonance MRI prostate images it can be seen from figure (2), and obtain
Fabulous segmentation effect, can provide strong help for diagnoses and treatment of the doctor to prostate patient.
Claims (5)
1. the magnetic resonance MRI prostate 3D rendering dividing methods based on more depth belief networks, including：
(1) n × m width magnetic resonance MRI prostate images are obtained, wherein n=10 represents patient's number, and m≤50 represent each patient
Be up to 50 width images；
(2) all pixels point of each image is pre-processed, will be by 13 × 13 × 3 length centered on each pixel
Cube is converted into the row vector of 1 × 507 dimension, obtains the patient data set being made of m × l row vector, and wherein l represents each width
Pixel number in image, l≤7200；
(3) training set is constructed：
9 patients are chosen from 10 patients, before selected each patientA row vector forms the first training set,
It is middleA row vector forms the second training set, afterA row vector forms the 3rd training set；
By remaining 1 patient, it is divided into three test sets according still further to as above method, is respectively the first test set, the second test set
With the 3rd test set；
(4) depth belief network is built, is learnt to obtain network one respectively with the first training set, learns to obtain net with the second training set
Network two, learns to obtain network three with the 3rd training set, these three networks have identical structure；
(5) train a softmax grader to obtain grader one using network one, a softmax is trained using network two
Grader obtains grader two, trains a softmax grader to obtain grader three using network three；
(6) test set is split：
Its result input grader one is calculated by the first test set input network one, obtains the segmentation knot of the first test set
Fruit；
Its result input grader two is calculated by the second test set input network two, obtains the segmentation knot of the second test set
Fruit；
Its result input grader three is calculated by the 3rd test set input network three, obtains the segmentation knot of the 3rd test set
Fruit；
The segmentation result of three test sets is successively merged, obtains the segmentation result of test set；
(7) three dimensional morphology is carried out to test set segmentation result to handle to obtain final segmentation result.
2. the magnetic resonance MRI prostate 3D rendering dividing methods according to claim 1 based on more depth belief networks, its
Middle step (4) builds depth belief network, and the depth that one three layers of construction is sequentially overlapped using three limited Boltzmann machines is believed
Network is read, the number of nodes of three limited Boltzmann machines is respectively 500,1000,500.
3. the magnetic resonance MRI prostate 3D rendering dividing methods according to claim 1 based on more depth belief networks, its
Middle step (5) trains a softmax grader to obtain grader one using network one, carries out in accordance with the following steps：
First training set is input in network one and obtains the output of network one by (5a), using this output as softmax graders
Training set；
Training set is input in softmax graders by (5b), using gradient descent algorithm minimization error, tries to achieve grader
Optimized parameter, obtains grader one.
4. the magnetic resonance MRI prostate 3D rendering dividing methods according to claim 1 based on more depth belief networks, its
Its result input grader one is calculated the first test set input network one by middle step (6), obtains the first test set
Segmentation result, carries out in accordance with the following steps：
Its result input grader one is calculated the first test set input network one by (6a), obtains point of the first test set
Cut result.
5. the magnetic resonance MRI prostate 3D rendering dividing methods according to claim 1 based on more depth belief networks, its
Middle step (7) carries out three dimensional morphology to the first segmentation result of test set and handles to obtain final segmentation result, according to following step
It is rapid to carry out：
The first segmentation result of test set is carried out three-dimensional expansion and handles to obtain expansion process result by (7a)；
Expansion process result is carried out three-dimensional corrosion treatment and obtains corrosion treatment result by (7b)；
Corrosion treatment result is carried out three-dimensional expansion and handles to obtain final segmentation result by (7c) again.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610048013.5A CN105719303B (en) | 2016-01-25 | 2016-01-25 | Magnetic resonance prostate 3D rendering dividing method based on more depth belief networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610048013.5A CN105719303B (en) | 2016-01-25 | 2016-01-25 | Magnetic resonance prostate 3D rendering dividing method based on more depth belief networks |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105719303A CN105719303A (en) | 2016-06-29 |
| CN105719303B true CN105719303B (en) | 2018-04-17 |
Family
ID=56154887
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610048013.5A Active CN105719303B (en) | 2016-01-25 | 2016-01-25 | Magnetic resonance prostate 3D rendering dividing method based on more depth belief networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105719303B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106683104B (en) * | 2017-01-06 | 2019-07-05 | 西北工业大学 | Prostate Magnetic Resonance Image Segmentation method based on integrated depth convolutional neural networks |
| CN109102512B (en) * | 2018-08-06 | 2021-03-09 | 西安电子科技大学 | DBN neural network-based MRI brain tumor image segmentation method |
| CN109448005B (en) * | 2018-10-31 | 2019-12-27 | 数坤（北京）网络科技有限公司 | Network model segmentation method and equipment for coronary artery |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103996023A (en) * | 2014-05-09 | 2014-08-20 | 清华大学深圳研究生院 | Light field face recognition method based on depth belief network |
| US8898093B1 (en) * | 2012-06-25 | 2014-11-25 | The Boeing Company | Systems and methods for analyzing data using deep belief networks (DBN) and identifying a pattern in a graph |
| CN104408692A (en) * | 2014-11-25 | 2015-03-11 | 南京信息工程大学 | Image fuzzy model parameter analysis method based on depth learning |
| CN104657751A (en) * | 2015-03-12 | 2015-05-27 | 华北电力大学（保定） | Mainline direction feature based deep belief network image classification method |
| CN105138975A (en) * | 2015-08-13 | 2015-12-09 | 济南大学 | Human body complexion area segmentation method based on deep belief network |
-
2016
- 2016-01-25 CN CN201610048013.5A patent/CN105719303B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8898093B1 (en) * | 2012-06-25 | 2014-11-25 | The Boeing Company | Systems and methods for analyzing data using deep belief networks (DBN) and identifying a pattern in a graph |
| CN103996023A (en) * | 2014-05-09 | 2014-08-20 | 清华大学深圳研究生院 | Light field face recognition method based on depth belief network |
| CN104408692A (en) * | 2014-11-25 | 2015-03-11 | 南京信息工程大学 | Image fuzzy model parameter analysis method based on depth learning |
| CN104657751A (en) * | 2015-03-12 | 2015-05-27 | 华北电力大学（保定） | Mainline direction feature based deep belief network image classification method |
| CN105138975A (en) * | 2015-08-13 | 2015-12-09 | 济南大学 | Human body complexion area segmentation method based on deep belief network |
Non-Patent Citations (2)
| Title |
|---|
| Efficient Training of Convolutional Deep Belief Networks in the Frequency Domain for Application to High-Resolution 2D and 3D Images;Tom Brosch等;《Neural Computation》;20151231;全文 * |
| 基于辨别性深度信念网络的说话人分割;马勇等;《清华大学学报（自然科学版）》;20130630;第53卷(第6期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105719303A (en) | 2016-06-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111047594B (en) | Tumor MRI weak supervised learning analysis modeling method and model thereof | |
| CN109598722B (en) | Image analysis method based on recurrent neural network | |
| CN111445390A (en) | Wide residual attention-based three-dimensional medical image super-resolution reconstruction method | |
| CN107909585A (en) | Inner membrance dividing method in a kind of blood vessel of intravascular ultrasound image | |
| CN109636811A (en) | Integration method, the device of the lobe of the lung section segmentation of CT images | |
| Zhang et al. | Multi‐needle localization with attention U‐net in US‐guided HDR prostate brachytherapy | |
| CN108535675A (en) | A kind of magnetic resonance multichannel method for reconstructing being in harmony certainly based on deep learning and data | |
| CN105719303B (en) | Magnetic resonance prostate 3D rendering dividing method based on more depth belief networks | |
| CN109377520A (en) | Cardiac image registration arrangement and method based on semi-supervised circulation GAN | |
| CN110338844A (en) | The display processing method and 3-D supersonic imaging method and system of three-dimensional imaging data | |
| CN112488976B (en) | Multi-modal medical image fusion method based on DARTS network | |
| CN110008992B (en) | Deep learning method for prostate cancer auxiliary diagnosis | |
| CN107563434A (en) | A kind of brain MRI image sorting technique based on Three dimensional convolution neutral net, device | |
| WO2020057074A1 (en) | Model training method and device for plaque segmentation, apparatus, and storage medium | |
| CN103985109B (en) | Feature-level medical image fusion method based on 3D (three dimension) shearlet transform | |
| CN110363802A (en) | Prostate figure registration system and method based on automatic segmentation and pelvis alignment | |
| Parikh et al. | Three dimensional virtual reality model of the normal female pelvic floor | |
| CN103268605B (en) | Based on brain map and the brain image method for registering of higher order statistical deformation model | |
| CN112641471A (en) | Bladder capacity determination and three-dimensional shape assessment method and system special for radiotherapy | |
| CN112419322A (en) | Temporal bone external semicircular canal segmentation method based on 3D multi-scale multi-pooling feature fusion network | |
| CN115841457A (en) | Three-dimensional medical image segmentation method fusing multi-view information | |
| CN116721253A (en) | Abdominal CT image multi-organ segmentation method based on deep learning | |
| Wang et al. | Triplanar convolutional neural network for automatic liver and tumor image segmentation | |
| CN113205148B (en) | Medical image frame interpolation method and terminal for iterative interlayer information fusion | |
| CN115375621A (en) | Liver tumor ablation path planning method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |