CN110363797A - A kind of PET and CT method for registering images inhibited based on excessive deformation - Google Patents
A kind of PET and CT method for registering images inhibited based on excessive deformation Download PDFInfo
- Publication number
- CN110363797A CN110363797A CN201910634301.2A CN201910634301A CN110363797A CN 110363797 A CN110363797 A CN 110363797A CN 201910634301 A CN201910634301 A CN 201910634301A CN 110363797 A CN110363797 A CN 110363797A
- Authority
- CN
- China
- Prior art keywords
- pet
- image
- image block
- excessive deformation
- dimensional
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10081—Computed x-ray tomography [CT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10104—Positron emission tomography [PET]
-
- 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
Abstract
The present invention relates to Medical Image Registration fields, provide a kind of PET and CT method for registering images inhibited based on excessive deformation.The present invention acquires two-dimensional PE T/CT sequence image first, obtains PET/CT sequence chart image set and pre-processes to it, obtains PET/CT image block training set；3D U-Net convolutional neural networks building PET/CT is then based on pseudo-crystalline lattice, and combines image similarity bound term and excessive deformation that item is inhibited to construct cost function；Then neural network weight parameter is initialized, hyper parameter is set, PET/CT image block training set input PET/CT with pseudo-crystalline lattice and is iterated training to it；PET/CT image to be registered is finally matched into pseudo-crystalline lattice to the PET/CT after input training, the PET image block after generating registration.The present invention can be realized PET/CT elastic registrating, improve registration efficiency and accuracy, reduce the calculating cost inhibited to excessive deformation.
Description
Technical field
The present invention relates to Medical Image Registration field, more particularly to a kind of PET inhibited based on excessive deformation with
CT method for registering images.
Background technique
Medical figure registration plays an important role in many Medical Image Processing tasks.Usually image registration is formulated
It is optimization problem to seek spatial alternation, which measures (example by the corresponding substitution in space maximized between image
The image intensity correlation being such as registrated between image) it is corresponding come the pixel/voxel established between a pair of of fixation and mobile image.
Position emissron tomography (Positron Emission Computer Tomography, PET) is accelerated using convolution
Device generates radioactive isotope 18F, 13N, and the metabolism of human body is participated in after intravenous injection.The high tissue of metabolic rate or lesion,
The bright signal of specific hypermetabolism is presented on PET；The low tissue of metabolic rate or lesion are in low metabolism dark signal on PET.It calculates
Machine tomoscan (ComputedTomography, CT) be with X-ray beam to human body certain a part by certain thickness level into
Row scanning, when x-ray directive tissue, part ray is absorbed by tissue, and part ray passes through human body and is detected organ reception,
Signal is generated, accurately image can be positioned.
PET/CT can carry out the same machine image co-registration of function and anatomical structure, be an impressive progress of medical imaging.
Multi-modality image registration utilizes the characteristics of various imaging modes, provides complementary information for different images, increases amount of image information,
Facilitate the property that lesion is appreciated more fully and the relationship with surrounding anatomical structures, the positioning for clinical diagnosis and treatment provides
Effective method.The fusion figure for having both structural information and functional information is obtained by the PET/CT image of two kinds of different modalities of fusion
Picture is of great significance for medical image analysis and diagnosis.Since the similitude of image pixel intensities between PET/CT image is lower, lead
It is easy to produce excessive deformation after causing registration, so PET/CT registration is a challenging task.
In existing PET and CT method for registering images, it is all based on iteration optimization greatly and is registrated.It is wherein most-often used
Method is that registration problems are converted to optimization problem to minimize cost function.Common cost function includes mean square error
(MSE), mutual information (MI), normalized mutual information (NMI), normalized crosscorrelation (NCC) and gradient are related (GC).And these phases
Like property Measure Indexes directly in pixel scale movement images, and it can not reflect the higher level structure in image.Despite the presence of complete
Office's optimization method, such as simulated annealing and genetic algorithm, but they need to sample parameter space comprehensively, and this will lead
Cause excessively high calculating cost.The calculating that excessive deformation is inhibited so as to cause existing PET and CT method for registering images it is at high cost and
It is low to the efficiency and accuracy of image registration.
Summary of the invention
In view of the problems of the existing technology, the present invention provide it is a kind of based on excessive deformation inhibit PET and CT image match
Quasi- method can be realized PET/CT elastic registrating, improve registration efficiency and accuracy, and reduction is calculated as excessive deformation inhibition
This.
The technical solution of the present invention is as follows:
A kind of PET and CT method for registering images inhibited based on excessive deformation, which is characterized in that include the following steps:
Step 1: two-dimensional PE T-sequence image, the two-dimensional ct sequence image of m patient of acquisition obtains PET/CT sequence image
Collection；
Step 2: PET/CT sequence chart image set being pre-processed, PET/CT image block training set is obtained；The pretreatment
Including calculating SUV value and hu value and threshold limit range, adjustment image resolution ratio, generation image block and being normalized；
Step 3: pseudo-crystalline lattice is matched based on 3DU-Net convolutional neural networks building PET/CT；
Step 4: inhibiting item in conjunction with image similarity bound term and excessive deformation, construction PET/CT matches the cost letter of pseudo-crystalline lattice
Number；The similarity constraint item is normalized crosscorrelation NCC, the excessive deformation inhibit item be based on displacement vector field element with
The penalty term adduction of difference between gauss of distribution function element；
Step 5: the size N criticized in batch processing, regularization term weight λ, maximum is arranged in initialization neural network weight parameter
The number of iterations COUNT, e-learning rate, optimizer are decayed tactful using learning rate；
Step 6: matching the input of pseudo-crystalline lattice using PET/CT image block training set as PET/CT, output displacement vector field will
Displacement vector field and PET image block are input to space transformer jointly, the PET image block after obtaining registration, according to CT image block
Similarity constraint item is obtained with the PET image block after registration, excessive deformation inhibition item is obtained according to displacement vector field, to pass through
Cost function updates neural network weight parameter, and carries out backpropagation, is thus iterated training with pseudo-crystalline lattice to PET/CT,
Until maximum number of iterations COUNT, the PET/CT after being trained matches pseudo-crystalline lattice；
Step 7: to PET/CT image to be registered to pretreatment described in step 1 is carried out, obtained PET/CT being schemed
PET/CT after training as block to input generates the PET image block after being registrated, and visualized in pseudo-crystalline lattice.
The step 2 includes the following steps:
Step 2.1: the SUV value for calculating two-dimensional PE T-sequence image is SUV=PixelsPET×LBM×1000/
injecteddose
The hu value for calculating two-dimensional ct sequence image is Hu=PixelsCT×slopes+intercepts
Wherein, PixelsPETFor the pixel value of PET sequence image, LBM is lean body mass, and injected dose is tracer
Injection dosage；PixelsCTFor the pixel value of CT sequence image, slopes is slope, and intercepts is intercept；
Step 2.2: enhancing picture contrast processing being carried out to two-dimensional PE T-sequence image, two-dimensional ct sequence image, adjusts hu
Value window width and window level is [a1, b1], SUV value is limited in [a2, b2] in；Wherein, a1、b1、a2、b2It is constant；
Step 2.3: image resolution ratio processing, the two-dimensional ct sequence of adjustment 512 × 512 are adjusted to two-dimensional ct sequence image
The size of column image is to size H × W=128 × 128 of two-dimensional PE T-sequence image；
Step 2.4: three-dimensional data is generated respectively to two-dimensional PE T-sequence image, the two-dimensional ct sequence image of i-th of patient
[H, W, DPET, i], [H, W, DCT, i], three-dimensional data is transformed into five dimension volume data [N, H, W, Di, C], it is sampling with d pixel
Interval cuts five dimension volume datas in z-direction, generates multipair H × W × D size image block, returns to image block
One change processing, obtains image block collection, randomly selects l from image block concentration and forms PET/CT figure to PET image block and CT image block
As block training set；Wherein, i ∈ { 1,2 ..., m }, DPET, iFor the number of sections of the PET sequence image of i-th of patient, DCT, iIt is
The number of sections of the CT sequence image of i patient, DPET, i=DCT, i=Di；N is the size criticized in batch processing, and C is input nerve
The port number of network data, C=2.
In the step 2, [a1, b1]=[- 90,300], [a2, b2]=[0,5], d=32, D=64.
In the step 2.4, it is to the formula that image block is normalizedBecome the data of image block
The normal distribution that for mean value be 0 and standard deviation is 1；Wherein, x, x*The forward and backward pixel of normalized respectively in image block
Point, μ, σ are respectively mean value, the standard deviation of all pixels point in image block.
In the step 3, matching pseudo-crystalline lattice based on 3DU-Net convolutional neural networks building PET/CT includes coding path reconciliation
Code path, each paths have 4 resolution levels；The coding path has n1Layer, each layer of the coding path are equal
It is the convolutional layer that 3 × 3 × 3, step-length is 2 including a convolution kernel, each convolutional layer is followed by one BN layers and ReLU layers；It is described
Decoding paths have n2Layer, each layer of the decoding paths includes that a convolution kernel is the deconvolution that 3 × 3 × 3, step-length is 2
Layer, each warp lamination are followed by one BN layers and ReLU layers；By shortcut, by the layer of equal resolution in coding path
Decoding paths are transmitted to, original high-resolution features are provided for decoding paths；The PET/CT matches the last layer of pseudo-crystalline lattice
For 3 × 3 × 3 convolutional layer, last output channel number is 3.
In the step 4, inhibit item in conjunction with image similarity bound term and excessive deformation, construction PET/CT is with pseudo-crystalline lattice
Cost function is
Wherein, F, M are respectively CT image block, PET image block, DvFor displacement vector field matrix,For mean value be μ,
Standard deviation is the gauss of distribution function of θ, and λ is regularization term weight；
For similarity constraint item
Wherein, S is subgraph, and T is template image, and (s, t) is coordinated indexing, and S (s, t) is the pixel value of subgraph, T (s, t)
For the pixel value of template image, E (S), E (T) are respectively the average gray value of subgraph, template image；
Inhibit item for excessive deformation
Wherein, i is displacement vector field matrix DvIn element, j is the random number for following Gaussian Profile Han Shuo Yuan, f (i, j,
It θ) is penalty term,
The invention has the benefit that
The present invention is based on 3DU-Net convolutional neural networks building PET/CT to match pseudo-crystalline lattice, passes through the nothing based on deep learning
End-to-end 3D elastic registrating neural network prediction displacement vector field is supervised, the displacement prediction by voxel is carried out to image subject to registration,
And inhibition of the item to image deformation is inhibited as similarity constraint item, in conjunction with excessive deformation using normalized crosscorrelation, construct PET/CT
Cost function with pseudo-crystalline lattice is able to solve the lower caused excessive deformation problems of registration of PET/CT image similarity itself, energy
It enough realizes PET/CT elastic registrating, improves registration efficiency and accuracy, reduce the calculating cost inhibited to excessive deformation.
Detailed description of the invention
Fig. 1 is the flow chart of the PET inhibited based on excessive deformation and CT method for registering images of the invention；
Fig. 2 is that PET/CT matches pseudo-crystalline lattice in the PET and CT method for registering images of the invention inhibited based on excessive deformation
Structural schematic diagram.
Specific embodiment
Below in conjunction with the drawings and specific embodiments, the invention will be further described.
As shown in Figure 1, for the flow chart of the PET inhibited based on excessive deformation and CT method for registering images of the invention.This
The PET and CT method for registering images that are inhibited based on excessive deformation of invention, which is characterized in that include the following steps:
Step 1: two-dimensional PE T-sequence image, the two-dimensional ct sequence image of m patient of acquisition obtains PET/CT sequence image
Collection.
Step 2: PET/CT sequence chart image set being pre-processed, PET/CT image block training set is obtained；The pretreatment
Including calculating SUV value and hu value and threshold limit range, adjustment image resolution ratio, generation image block and being normalized.
The step 2 includes the following steps:
Step 2.1: the SUV value for calculating two-dimensional PE T-sequence image is SUV=PixelsPET×LBM×1000/
injecteddose
The hu value for calculating two-dimensional ct sequence image is Hu=PixelsCT×slopes+intercepts
Wherein, PixelsPETFor the pixel value of PET sequence image, LBM is lean body mass, and injected dose is tracer
Injection dosage；PixelsCTFor the pixel value of CT sequence image, slopes is slope, and intercepts is intercept；
Step 2.2: enhancing picture contrast processing being carried out to two-dimensional PE T-sequence image, two-dimensional ct sequence image, adjusts hu
Value window width and window level is [a1, b1], SUV value is limited in [a2, b2] in；Wherein, a1、b1、a2、b2It is constant；
Step 2.3: image resolution ratio processing, the two-dimensional ct sequence of adjustment 512 × 512 are adjusted to two-dimensional ct sequence image
The size of column image is to size H × W=128 × 128 of two-dimensional PE T-sequence image；
Step 2.4: three-dimensional data is generated respectively to two-dimensional PE T-sequence image, the two-dimensional ct sequence image of i-th of patient
[H, W, DPET, i], [H, W, DCT, i], three-dimensional data is transformed into five dimension volume data [N, H, W, Di, C], it is sampling with d pixel
Interval cuts five dimension volume datas in z-direction, generates multipair H × W × D size image block, returns to image block
One change processing, obtains image block collection, randomly selects l from image block concentration and forms PET/CT figure to PET image block and CT image block
As block training set；Wherein, i ∈ { 1,2 ..., m }, DPET, iFor the number of sections of the PET sequence image of i-th of patient, DCT, iIt is
The number of sections of the CT sequence image of i patient, DPET, i=DCT, i=Di；N is the size criticized in batch processing, and C is input nerve
The port number of network data, C=2.
In the present embodiment, m=176；SUV, Hu the value image of PET, CT after treatment are generated into three-dimensional data respectively
There are in ndarray；In the step 2, [a1, b1]=[- 90,300], [a2, b2]=[0,5], d=32, D=64.
For all 176 patients, the l=900 of volume data generation of wherein 141 patients is randomly choosed to SUV, Hu value
Image block is as PET/CT image block training set, 259 pairs of SUV, Hu value images for selecting the volume data of wherein 35 patients to generate
Block is as verifying collection.
In the step 2.4, it is to the formula that image block is normalizedBecome the data of image block
The normal distribution that for mean value be 0 and standard deviation is 1；Wherein, x, x*The forward and backward pixel of normalized respectively in image block
Point, μ, σ are respectively mean value, the standard deviation of all pixels point in image block.
Step 3: pseudo-crystalline lattice being matched based on 3DU-Net convolutional neural networks building PET/CT, as shown in Figure 2.
PET/CT of the invention is with the 3D U-Net that pseudo-crystalline lattice includes: that (1) is used to return displacement vector field；(2) it carries out empty
Between the component feature space converter (Spatial Transformer) that converts.In the present embodiment, in the step 3, it is based on 3DU-
It includes coding path and decoding paths that Net convolutional neural networks, which construct PET/CT with pseudo-crystalline lattice, and each paths have 4 resolutions
Rate rank；The coding path has n1Layer, each layer of the coding path include that a convolution kernel is 3 × 3 × 3, step-length
For 2 convolutional layer, each convolutional layer is followed by one BN layers and ReLU layers；The decoding paths have n2Layer, the decoding paths
Each layer include a convolution kernel be the warp lamination that 3 × 3 × 3, step-length is 2, each warp lamination is followed by a BN
Layer and ReLU layers；By shortcut, the layer of equal resolution in coding path is transmitted to decoding paths, is mentioned for decoding paths
For original high-resolution features；The convolutional layer that the last layer of the PET/CT with pseudo-crystalline lattice is 3 × 3 × 3, finally output is logical
Road number is 3.
Wherein, BN layers are batch normalization layer, and ReLU layers are the linear elementary layer of rectification, and shortcut is jump connection.Most
The convolutional layer that later layer is 3 × 3 × 3, reduces the port number of output, and last output channel number is 3 (i.e. expression three sides of x, y, z
To).
Step 4: inhibiting item in conjunction with image similarity bound term and excessive deformation, construction PET/CT matches the cost letter of pseudo-crystalline lattice
Number；The similarity constraint item is normalized crosscorrelation NCC, the excessive deformation inhibit item be based on displacement vector field element with
The penalty term adduction of difference between gauss of distribution function element.
Wherein, " excessive deformation inhibition " is defined based on the deformation degree size of 3D Deformation Field to estimate, and is drawn in cost function
Enter " excessive deformation inhibition item " to optimize with pseudo-crystalline lattice.
In the step 4, inhibit item in conjunction with image similarity bound term and excessive deformation, construction PET/CT is with pseudo-crystalline lattice
Cost function is
Wherein, F, M are respectively CT image block, PET image block, DvFor displacement vector field matrix,For mean value be μ,
Standard deviation is the gauss of distribution function of θ, and λ is regularization term weight；Wherein, F, M are respectively fixed image block, floating image block；
For similarity constraint item
Wherein, S is subgraph, and T is template image, and (s, t) is coordinated indexing, and S (s, t) is the pixel value of subgraph, T (s, t)
For the pixel value of template image, E (S), E (T) are respectively the average gray value of subgraph, template image；
Inhibit item for excessive deformation
Wherein, i is displacement vector field matrix DvIn element, j is the random number for following Gaussian Profile Han Shuo Yuan, f (i, j,
It θ) is penalty term,In the present embodiment, rule of thumb k is arranged to 2,
As | i-j | when > θ, penalty term is | i-j |k, i.e., penalty term is amplified by k power.
Step 5: nerve net is initialized using global variable initialization (global_variables_initializer)
The size N=16 criticized in batch processing, regularization term weight λ=0.3, maximum number of iterations COUNT=is arranged in network weight parameter
1000, e-learning rate be 0.001, optimizer Adam, using learning rate decay strategy.
Step 6: matching the input of pseudo-crystalline lattice using PET/CT image block training set as PET/CT, output displacement vector field will
Displacement vector field and PET image block are input to space transformer jointly, the PET image block after obtaining registration, according to CT image block
Similarity constraint item is obtained with the PET image block after registration, excessive deformation inhibition item is obtained according to displacement vector field, to pass through
Cost function updates neural network weight parameter, and carries out backpropagation, is thus iterated training with pseudo-crystalline lattice to PET/CT,
Until maximum number of iterations COUNT, the PET/CT after being trained matches pseudo-crystalline lattice.
Wherein, using the PET/CT image block of a pair of 128 × 128 × 64 sizes as the input of 3D U-Net network, output
The displacement vector field (128 × 128 × 64 × 3, respectively correspond the displacement in x, y, z direction) of equal resolution size swears displacement
Amount field and PET image block are input to space transformer jointly, export the PET image block after being registrated.
Step 7: to PET/CT image to be registered to pretreatment described in step 1 is carried out, obtained PET/CT being schemed
PET/CT after training as block to input generates the PET image block after being registrated, and visualized in pseudo-crystalline lattice.
In the present embodiment, the PET inhibited based on excessive deformation and CT method for registering images of the invention are operated in Intel
In the Windows10 system environments of core, medical figure registration is carried out based on Python and Tensorflow frame.The present invention uses
The regression problem that image registration is converted to displacement vector field based on the image registration algorithm of deep learning, i.e., prediction come from one
To the spatial correspondence between the pixel/voxel of image.Image registration is optimized simultaneously by 3D U-Net convolutional neural networks
Similarity constraint item and the excessive deformation of displacement vector field between fixed image and floating image pair inhibit item to realize.It is quantitative and
Qualitative results show that carrying out 3D PET/CT image registration using method for registering of the invention obtains good effect.Wherein,
For the model trained, a pair of new PET/CT volume data is given, can be matched in 10 seconds by primary positive calculate
Quasi- result.
Obviously, above-described embodiment is only a part of the embodiments of the present invention, instead of all the embodiments.Above-mentioned implementation
Example for explaining only the invention, is not intended to limit the scope of the present invention..Based on the above embodiment, those skilled in the art
Member's every other embodiment obtained namely all in spirit herein and original without making creative work
Made all modifications, equivalent replacement and improvement etc., are all fallen within the protection domain of application claims within reason.
Claims (6)
1. a kind of PET and CT method for registering images inhibited based on excessive deformation, which is characterized in that include the following steps:
Step 1: two-dimensional PE T-sequence image, the two-dimensional ct sequence image of m patient of acquisition obtains PET/CT sequence chart image set；
Step 2: PET/CT sequence chart image set being pre-processed, PET/CT image block training set is obtained；The pretreatment includes
It calculates SUV value and hu value and threshold limit range, adjustment image resolution ratio, generation image block and is normalized；
Step 3: pseudo-crystalline lattice is matched based on 3D U-Net convolutional neural networks building PET/CT；
Step 4: inhibiting item in conjunction with image similarity bound term and excessive deformation, construction PET/CT matches the cost function of pseudo-crystalline lattice；
The similarity constraint item is normalized crosscorrelation NCC, and it is based on displacement vector field element and height that the excessive deformation, which inhibits item,
The penalty term adduction of difference between this distribution function element；
Step 5: the size N criticized in batch processing, regularization term weight λ, greatest iteration is arranged in initialization neural network weight parameter
Number COUNT, e-learning rate, optimizer are decayed tactful using learning rate；
Step 6: matching the input of pseudo-crystalline lattice using PET/CT image block training set as PET/CT, output displacement vector field will be displaced
Vector field and PET image block are input to space transformer jointly, and the PET image block after obtaining registration according to CT image block and is matched
PET image block after standard obtains similarity constraint item, obtains excessive deformation inhibition item according to displacement vector field, to pass through cost
Function updates neural network weight parameter, and carries out backpropagation, is thus iterated training with pseudo-crystalline lattice to PET/CT, until
Maximum number of iterations COUNT, the PET/CT after being trained match pseudo-crystalline lattice；
Step 7: to PET/CT image to be registered to pretreatment described in progress step 1, the PET/CT image block that will be obtained
To the PET/CT after input training in pseudo-crystalline lattice, the PET image block after being registrated is generated, and visualized.
2. the PET and CT method for registering images according to claim 1 inhibited based on excessive deformation, which is characterized in that institute
Step 2 is stated to include the following steps:
Step 2.1: the SUV value for calculating two-dimensional PE T-sequence image is
SUV=PixelsPET×LBM×1000/injected dose
Calculate two-dimensional ct sequence image hu value be
Hu=PixelsCT×slopes+intercepts
Wherein, PixelsPETFor the pixel value of PET sequence image, LBM is lean body mass, and injected dose is tracer injection
Dosage；PixelsCTFor the pixel value of CT sequence image, slopes is slope, and intercepts is intercept；
Step 2.2: enhancing picture contrast processing being carried out to two-dimensional PE T-sequence image, two-dimensional ct sequence image, adjusts hu value window
Wide window position is [a1, b1], SUV value is limited in [a2, b2] in；Wherein, a1、b1、a2、b2It is constant；
Step 2.3: image resolution ratio processing, the two-dimensional ct sequence chart of adjustment 512 × 512 are adjusted to two-dimensional ct sequence image
The size of picture is to size H × W=128 × 128 of two-dimensional PE T-sequence image；
Step 2.4: to two-dimensional PE T-sequence image, the two-dimensional ct sequence image of i-th of patient generate respectively three-dimensional data [H,
W, DPET, i], [H, W, DCT, i], three-dimensional data is transformed into five dimension volume data [N, H, W, Di, C], using d pixel as the sampling interval
Five dimension volume datas are cut in a z-direction, multipair H × W × D size image block is generated, image block is normalized
Processing, obtains image block collection, randomly selects l from image block concentration and forms PET/CT image block to PET image block and CT image block
Training set；Wherein, i ∈ { 1,2 ..., m }, DPET, iFor the number of sections of the PET sequence image of i-th of patient, DCT, iIt is i-th
The number of sections of the CT sequence image of patient, DPET, i=DCT, i=Di；N is the size criticized in batch processing, and C is input neural network
The port number of data, C=2.
3. the PET and CT method for registering images according to claim 2 inhibited based on excessive deformation, which is characterized in that institute
It states in step 2, [a1, b1]=[- 90,300], [a2, b2]=[0,5], d=32, D=64.
4. the PET and CT method for registering images according to claim 2 inhibited based on excessive deformation, which is characterized in that institute
It states in step 2.4, is to the formula that image block is normalizedMake the data of image block become mean value 0 and
The normal distribution that standard deviation is 1；Wherein, x, x*Respectively the forward and backward pixel of normalized, μ, σ are respectively in image block
Mean value, the standard deviation of all pixels point in image block.
5. the PET and CT method for registering images according to any one of claim 2 to 4 inhibited based on excessive deformation,
Be characterized in that, in the step 3, based on 3D U-Net convolutional neural networks building PET/CT with pseudo-crystalline lattice include coding path and
Decoding paths, each paths have 4 resolution levels；The coding path has n1Layer, each layer of the coding path
It include a convolution kernel is convolutional layer that 3 × 3 × 3, step-length is 2, each convolutional layer is followed by one BN layers and ReLU layers；Institute
Stating decoding paths has n2Layer, each layer of the decoding paths includes that a convolution kernel is the warp that 3 × 3 × 3, step-length is 2
Lamination, each warp lamination are followed by one BN layers and ReLU layers；By shortcut, by equal resolution in coding path
Layer is transmitted to decoding paths, and original high-resolution features are provided for decoding paths；The PET/CT matches last of pseudo-crystalline lattice
The convolutional layer that layer is 3 × 3 × 3, last output channel number are 3.
6. the PET and CT method for registering images according to claim 5 inhibited based on excessive deformation, which is characterized in that institute
It states in step 4, inhibits item in conjunction with image similarity bound term and excessive deformation, cost function of the construction PET/CT with pseudo-crystalline lattice is
Wherein, F, M are respectively CT image block, PET image block, DvFor displacement vector field matrix,It is μ for mean value, standard
Difference is the gauss of distribution function of θ, and λ is regularization term weight；
For similarity constraint item
Wherein, S is subgraph, and T is template image, and (s, t) is coordinated indexing, and S (s, t) is the pixel value of subgraph, and T (s, t) is mould
The pixel value of plate image, E (S), E (T) are respectively the average gray value of subgraph, template image；
Inhibit item for excessive deformation
Wherein, i is displacement vector field matrix DvIn element, j be follow gauss of distribution functionRandom number, f (i, j, θ) is
Penalty term,
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910634301.2A CN110363797B (en) | 2019-07-15 | 2019-07-15 | PET and CT image registration method based on excessive deformation inhibition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910634301.2A CN110363797B (en) | 2019-07-15 | 2019-07-15 | PET and CT image registration method based on excessive deformation inhibition |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110363797A true CN110363797A (en) | 2019-10-22 |
| CN110363797B CN110363797B (en) | 2023-02-14 |
Family
ID=68219107
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910634301.2A Active CN110363797B (en) | 2019-07-15 | 2019-07-15 | PET and CT image registration method based on excessive deformation inhibition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110363797B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111260705A (en) * | 2020-01-13 | 2020-06-09 | 武汉大学 | Prostate MR image multi-task registration method based on deep convolutional neural network |
| CN111524170A (en) * | 2020-04-13 | 2020-08-11 | 中南大学 | Lung CT image registration method based on unsupervised deep learning |
| CN112598718A (en) * | 2020-12-31 | 2021-04-02 | 北京深睿博联科技有限责任公司 | Unsupervised multi-view multi-mode intelligent glasses image registration method and device |
| CN113706451A (en) * | 2021-07-07 | 2021-11-26 | 杭州脉流科技有限公司 | Method, device, system and computer-readable storage medium for intracranial aneurysm identification detection |
| CN114511602A (en) * | 2022-02-15 | 2022-05-17 | 河南工业大学 | Medical image registration method based on graph convolution Transformer |
| CN114820432A (en) * | 2022-03-08 | 2022-07-29 | 安徽慧软科技有限公司 | Radiotherapy effect evaluation method based on PET and CT elastic registration technology |
| CN115393527A (en) * | 2022-09-14 | 2022-11-25 | 北京富益辰医疗科技有限公司 | Anatomical navigation construction method and device based on multimode image and interactive equipment |
| CN116740218A (en) * | 2023-08-11 | 2023-09-12 | 南京安科医疗科技有限公司 | Heart CT imaging image quality optimization method, device and medium |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110019885A1 (en) * | 2009-07-24 | 2011-01-27 | Sarah Bond | Methods and apparatus for registration of medical images |
| CN107507234A (en) * | 2017-08-29 | 2017-12-22 | 北京大学 | Cone beam computed tomography image and x-ray image method for registering |
| CN108171738A (en) * | 2018-01-25 | 2018-06-15 | 北京雅森科技发展有限公司 | Multimodal medical image registration method based on brain function template |
| CN109074659A (en) * | 2016-05-04 | 2018-12-21 | 皇家飞利浦有限公司 | Medical image resources registration |
| CN109272443A (en) * | 2018-09-30 | 2019-01-25 | 东北大学 | A kind of PET based on full convolutional neural networks and CT method for registering images |
| CN109685811A (en) * | 2018-12-24 | 2019-04-26 | 北京大学第三医院 | PET/CT hypermetabolism lymph node dividing method based on dual path U-net convolutional neural networks |
| CN109872332A (en) * | 2019-01-31 | 2019-06-11 | 广州瑞多思医疗科技有限公司 | A kind of 3 d medical images method for registering based on U-NET neural network |
-
2019
- 2019-07-15 CN CN201910634301.2A patent/CN110363797B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110019885A1 (en) * | 2009-07-24 | 2011-01-27 | Sarah Bond | Methods and apparatus for registration of medical images |
| CN109074659A (en) * | 2016-05-04 | 2018-12-21 | 皇家飞利浦有限公司 | Medical image resources registration |
| CN107507234A (en) * | 2017-08-29 | 2017-12-22 | 北京大学 | Cone beam computed tomography image and x-ray image method for registering |
| CN108171738A (en) * | 2018-01-25 | 2018-06-15 | 北京雅森科技发展有限公司 | Multimodal medical image registration method based on brain function template |
| CN109272443A (en) * | 2018-09-30 | 2019-01-25 | 东北大学 | A kind of PET based on full convolutional neural networks and CT method for registering images |
| CN109685811A (en) * | 2018-12-24 | 2019-04-26 | 北京大学第三医院 | PET/CT hypermetabolism lymph node dividing method based on dual path U-net convolutional neural networks |
| CN109872332A (en) * | 2019-01-31 | 2019-06-11 | 广州瑞多思医疗科技有限公司 | A kind of 3 d medical images method for registering based on U-NET neural network |
Non-Patent Citations (2)
| Title |
|---|
| HESSAM SOKOOTIL 等: "Unsupervised learning for fast probabilistic diffeomorphic registration", 《MEDICAL LMAGE COMPUTING AND COMPUTER ASSISTED INTERVENTION -MICCAL 2017》 * |
| 陈明: "三维医学图像刚性配准新算法研究", 《中国优秀博硕士学位论文全文数据库 (博士)医药卫生科技辑》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111260705A (en) * | 2020-01-13 | 2020-06-09 | 武汉大学 | Prostate MR image multi-task registration method based on deep convolutional neural network |
| CN111260705B (en) * | 2020-01-13 | 2022-03-15 | 武汉大学 | Prostate MR image multi-task registration method based on deep convolutional neural network |
| CN111524170A (en) * | 2020-04-13 | 2020-08-11 | 中南大学 | Lung CT image registration method based on unsupervised deep learning |
| CN111524170B (en) * | 2020-04-13 | 2023-05-26 | 中南大学 | Pulmonary CT image registration method based on unsupervised deep learning |
| CN112598718A (en) * | 2020-12-31 | 2021-04-02 | 北京深睿博联科技有限责任公司 | Unsupervised multi-view multi-mode intelligent glasses image registration method and device |
| CN113706451A (en) * | 2021-07-07 | 2021-11-26 | 杭州脉流科技有限公司 | Method, device, system and computer-readable storage medium for intracranial aneurysm identification detection |
| CN114511602A (en) * | 2022-02-15 | 2022-05-17 | 河南工业大学 | Medical image registration method based on graph convolution Transformer |
| CN114820432A (en) * | 2022-03-08 | 2022-07-29 | 安徽慧软科技有限公司 | Radiotherapy effect evaluation method based on PET and CT elastic registration technology |
| CN114820432B (en) * | 2022-03-08 | 2023-04-11 | 安徽慧软科技有限公司 | Radiotherapy effect evaluation method based on PET and CT elastic registration technology |
| CN115393527A (en) * | 2022-09-14 | 2022-11-25 | 北京富益辰医疗科技有限公司 | Anatomical navigation construction method and device based on multimode image and interactive equipment |
| CN116740218A (en) * | 2023-08-11 | 2023-09-12 | 南京安科医疗科技有限公司 | Heart CT imaging image quality optimization method, device and medium |
| CN116740218B (en) * | 2023-08-11 | 2023-10-27 | 南京安科医疗科技有限公司 | Heart CT imaging image quality optimization method, device and medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110363797B (en) | 2023-02-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110363797A (en) | A kind of PET and CT method for registering images inhibited based on excessive deformation | |
| JP7203852B2 (en) | Estimation of full-dose PET images from low-dose PET imaging using deep learning | |
| CN109272443A (en) | A kind of PET based on full convolutional neural networks and CT method for registering images | |
| CN101076282B (en) | Dynamic tracking of moving targets | |
| US9364192B2 (en) | Error estimates in quantitative functional imaging | |
| CN109753866A (en) | With the medical image object detection of the dense feature pyramid network architecture in machine learning | |
| Gholamiankhah et al. | Deep learning-based synthetic CT generation from MR images: comparison of generative adversarial and residual neural networks | |
| US20200037974A1 (en) | Deep-learning based separation method of a mixture of dual-tracer single-acquisition pet signals with equal half-lives | |
| US11475535B2 (en) | PET-CT registration for medical imaging | |
| CN111436958B (en) | CT image generation method for PET image attenuation correction | |
| CN106491151A (en) | PET image acquisition methods and system | |
| CN102067176A (en) | Radiological imaging incorporating local motion monitoring, correction, and assessment | |
| Bardsley et al. | Hierarchical regularization for edge-preserving reconstruction of PET images | |
| CN110270015B (en) | sCT generation method based on multi-sequence MRI | |
| Hsiao et al. | Joint-MAP Bayesian tomographic reconstruction with a gamma-mixture prior | |
| Reader | The promise of new PET image reconstruction | |
| Cheon et al. | Deep learning in radiation oncology | |
| Zhang et al. | DR-only Carbon-ion radiotherapy treatment planning via deep learning | |
| US11786201B1 (en) | Method and system for material decomposition in dual- or multiple-energy x-ray based imaging | |
| Lu | Collimator optimization for single photon emission computed tomography using detection and localization tasks | |
| EP4292536A1 (en) | Method and system for material decomposition in dual- or multiple-energy x-ray based imaging | |
| EP3961570A1 (en) | Ai-based virtual fov extension for scatter correction in emission tomography | |
| Pan et al. | Full-dose PET synthesis from low-dose PET using 2D high efficiency denoising diffusion probabilistic model | |
| Hsieh et al. | Mathematical observer for 3D radiographic scanner design | |
| Yamauchi et al. | Feasibility study on image reconstruction for single-photon emission computed tomography with limited projections by neural networks |
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 |