CN109685787A - Output method, device in the lobe of the lung section segmentation of CT images - Google Patents
Output method, device in the lobe of the lung section segmentation of CT images Download PDFInfo
- Publication number
- CN109685787A CN109685787A CN201811572365.6A CN201811572365A CN109685787A CN 109685787 A CN109685787 A CN 109685787A CN 201811572365 A CN201811572365 A CN 201811572365A CN 109685787 A CN109685787 A CN 109685787A
- Authority
- CN
- China
- Prior art keywords
- lung
- segmentation
- lobe
- section
- images
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/12—Edge-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/187—Segmentation; Edge detection involving region growing; involving region merging; involving connected component labelling
-
- 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
- G06T2207/20—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
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- 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
- G06T2207/30061—Lung
-
- 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
- G06T2207/30101—Blood vessel; Artery; Vein; Vascular
Abstract
The present invention relates to lobe of the lung section dividing method, device, system, storage medium and the equipment of a kind of CT images, this method comprises: detecting step, in CT images, detection exports lung outlines comprising intrapulmonary region, lung exterior domain；Screening step selects the mode of machine segmentation to filter out intrapulmonary region, and as candidate region in lung outlines；Segmentation step in the 3D level of candidate region, while carrying out blood vessel segmentation to lung section and the lobe of the lung and splitting segmentation with lung；Constitution step obtains the three-dimensional vascular distribution of lung by constructing vascular tree according to vessel segmentation；Vascular tree and lung are split segmentation result and combined, and carry out lobe of the lung section segmentation by integration step, obtain the final segmentation result of candidate region.Lobe of the lung section dividing method, device, system, storage medium and the equipment of CT images provided by the invention based on deep learning, effectively reduce error, improve diagnosis and accuracy rate, and are not limited by individual lung's morphological differences.
Description
Technical field
The present invention relates to the lobe of the lung section dividing method of image segmentation field more particularly to a kind of CT images, device, system,
Storage medium and equipment.
Background technique
Currently, lung cancer is the highest cancer of the death rate in all cancers.Lung neoplasm is the Findings form of lung cancer, in CT
The shade that increase in density is shown as in iconography, it is right by the detection and segmentation to Lung neoplasm in CT images middle lobe and lung section
The early screening of lung cancer and assessment are of great significance.In existing detection, 3D CT (Emmenlauer, M., etc.: " free,
fast and reliable stitching of large 3d datasets.”J Microscopy,233,no.1,
Pp.42-60,2009.) image data amount is big, and there are biggish individual differences, bring to the lobe of the lung and the cutting techniques of lung section
Certain difficulty.
It is different from the application of the computer vision technique of broad sense, in order to reduce answering for depth convolutional neural networks (DCNN) calculating
It is miscellaneous degree, efficiently utilize limited training data, the prior art using by medical ground information in conjunction with deep learning by the way of
It is split, for example, 1) being now with successively progressive partitioning scheme for the lobe of the lung and the best technology of lung section segmentation effect
(K.Hayashi,etc.:“Radiographic and CT appearances of the major fissures,”
Radiographics, vol.21, no.4, pp.861-874,2001.), detection positioning first is carried out to lung outlines, is recycled deep
The method of degree study is split (S.Hu, etc.: " Automatic lung segmentation for the lobe of the lung
accurate quantitation of volumetric X-ray CT images.”IEEE Trans.Med.Imag.,
Vol.20, no.6, pp.490-498, Jun.2001.), the point coordinate according to lung section is characterized to divide lung section (E.M.van
Rikxoort,etc.:“Supervised enhancement filters:Application to fissure
detection in chest CT scans,”IEEE Trans.Med.Imag.,vol.27,no.1,pp.1–10,
Jan.2008.).This mode has ignored the lung mechanics difference between different patients, and testing result has uncertainty
(I.C.Sluimer,etc.:“Towards automated segmentation of the pathological lung in
CT,"IEEE Trans.Med.Imag.,vol.24,no.8,pp.1025–1038,Aug.2005.).2) by tracheae and arteries and veins
Washing deformation is established in the analysis of pipe, and it is preferable that this mode splits segmentation effect for visible lung, but for partially not
Poor (E.van Rikxoort, etc.: " the Automatic segmentation of of visible point of column split effect
pulmonary lobes robust against incomplete fissures.”IEEE Transactions on
Medical Imaging,vol.29,no.6,pp.1286–1296,2010.).Above method carries out in mass data
Test, segmentation effect are bad.
Chinese patent literature CN103035009A discloses a kind of reconstruction based on CT images Lung neoplasm edge and segmentation side
Method, the patent first with to Gradient Features there is the mapping mode of rarefaction representation ability to carry out spatial alternation to image；Then
CT Evaluation of Image Segmentation system is established, mainly marginal information is amplified, rebuilds Lung neoplasm edge.
Although above-mentioned patent improves the accuracy rate of Lung neoplasm edge detection and segmentation in CT images, starting to figure
Before the edge of picture is detected, need to be compared with threshold value initially set, it only could in the case where meeting condition
Start to detect, conversely, needing to handle image again, this method use scope is limited by individual difference, is needed not
It is disconnected to be adjusted.
As the state-of-the-art technology LDCT (low-dose spiral CT) of screening lung cancer, LDCT dose of radiation is lower, more conventional CT agent
Amount reduces by 75%~90%, does not influence lung's imaging, the same sensibility having with CT, from American National lung in 2011 but
The early stage of lung cancer can be improved in the Notes of Key Datas of cancer Screening tests (National Lung Screening Trial, NLST) LDCT
Diagnosis and after lung cancer case fatality rate is effectively reduced, every clinical research about screening lung cancer successively provides soul-stirring
As a result, effectively having promoted application of the LDCT in screening lung cancer.Therefore, how sufficiently to activate and utilize intimate silent storage
Chest LDCT big data, be one of the most urgent problems to be solved by those skilled in the art.
In conclusion how to improve the accuracy and precision of detection and the segmentation to Lung neoplasm in CT images, Yi Jiru
What makes full use of and activates the big data of CT images, is one of those skilled in the art's urgent problem to be solved.
Summary of the invention
The CT images lobe of the lung section dividing method that solve the above-mentioned problems, the invention proposes a kind of based on deep learning,
Device, system, storage medium and equipment.
To achieve the goals above, first aspect present invention provides a kind of lobe of the lung section dividing method of CT images, the party
Method the following steps are included:
Detecting step, in CT images, detection output lung outlines, lung outlines include intrapulmonary region, lung exterior domain；
Screening step selects the mode of machine segmentation to filter out intrapulmonary region, and intrapulmonary region is made in lung outlines
For candidate region；
Segmentation step in the 3D level of candidate region, while carrying out blood vessel segmentation to lung section and the lobe of the lung and splitting segmentation with lung；
Constitution step obtains the three-dimensional vascular distribution of lung by constructing vascular tree according to vessel segmentation；
Vascular tree and lung are split segmentation result using lobe of the lung section partitioning algorithm and are combined by integration step, and carry out the lobe of the lung
Section segmentation, obtains the final segmentation result of candidate region.
Further, the FCN network of the structure based on multilayer convolutional layer is used to be detected and exported in detecting step.
Specifically, in segmentation step, blood vessel is carried out respectively to lung section and the lobe of the lung using 3D U-net and along two branch lines
Segmentation splits segmentation with lung.
Further, in constitution step, tracheae distribution situation is inferred according to three-dimensional vascular distribution, obtains bronchial tree.
Specifically, in integration step, lobe of the lung section partitioning algorithm is by input lobe of the lung section disaggregated model, by vascular tree and lung
Segmentation result integration is split, and exports final segmentation result.
Second aspect of the present invention provides a kind of lobe of the lung section segmenting device based on CT images, which includes:
Detection module, in CT images, detection output lung outlines, lung outlines to include intrapulmonary region, lung outskirt
Domain；
Screening module selects the mode of machine segmentation to filter out intrapulmonary region in lung outlines, and by intrapulmonary area
Domain is as candidate region；
Divide module, for splitting in the 3D level of candidate region, while to lung section and lobe of the lung progress blood vessel segmentation with lung
Segmentation；
Constructing module, for obtaining the three-dimensional blood vessel point of lung by constructing vascular tree according to vessel segmentation
Cloth；
Module is integrated, is combined for vascular tree and lung to be split segmentation result using lobe of the lung section partitioning algorithm, and carry out
The segmentation of lobe of the lung section, obtains the final segmentation result of candidate region.
Further, used in detection module the FCN network of the structure based on multilayer convolutional layer carry out detection and it is defeated
Out.
Specifically, in segmentation module, blood vessel is carried out respectively to lung section and the lobe of the lung using 3D U-net and along two branch lines
Segmentation splits segmentation with lung.
Further, in constructing module, tracheae distribution situation is inferred according to three-dimensional vascular distribution, obtains bronchial tree.
Specifically, in integrating module, lobe of the lung section partitioning algorithm is by input lobe of the lung section disaggregated model, by vascular tree and lung
Segmentation result integration is split, and exports final segmentation result.
Third aspect present invention provides a kind of system of lobe of the lung section segmentation realized based on CT images, including second aspect
Any one of the lobe of the lung section segmenting device based on CT images.
Fourth aspect present invention provides a kind of non-volatile memory medium, which, which has, stores it
In instruction, when executed, so that processor executes the lobe of the lung section dividing method based on CT images, instruction includes:
Detection instruction, in CT images, detection output lung outlines, lung outlines include intrapulmonary region, lung exterior domain；
Screening instruction selects the mode of machine segmentation to filter out intrapulmonary region, and intrapulmonary region is made in lung outlines
For candidate region；
Split order in the 3D level of candidate region, while carrying out blood vessel segmentation to lung section and the lobe of the lung and splitting segmentation with lung；
Construction instruction, obtains the three-dimensional vascular distribution of lung by constructing vascular tree according to vessel segmentation；
Integration instruction, splits segmentation result for vascular tree and lung using lobe of the lung section partitioning algorithm and is combined, and carry out the lobe of the lung
Section segmentation, obtains the final segmentation result of candidate region.
Fifth aspect present invention provides a kind of equipment, including memory and processor, and memory is stored with computer can
It executes instruction, processor is configured as executing instruction to implement the process of the lobe of the lung section segmentation based on CT images, and process includes:
Detecting step, in CT images, detection output lung outlines, lung outlines include intrapulmonary region, lung exterior domain；
Screening step selects the mode of machine segmentation to filter out intrapulmonary region, and intrapulmonary region is made in lung outlines
For candidate region；
Segmentation step in the 3D level of candidate region, while carrying out blood vessel segmentation to lung section and the lobe of the lung and splitting segmentation with lung；
Constitution step obtains the three-dimensional vascular distribution of lung by constructing vascular tree according to vessel segmentation；
Vascular tree and lung are split segmentation result using lobe of the lung section partitioning algorithm and are combined by integration step, and carry out the lobe of the lung
Section segmentation, obtains the final segmentation result of candidate region.
A kind of CT images lobe of the lung section dividing method, device, system, storage medium and equipment provided by the invention, are adopted first
Multilayer convolutional layer is used to accurately detect and export the lung outlines in CT images as the FCN network of main structure；Secondly it screens
Candidate region simultaneously selects 3D U-net expansion to split segmentation to lung section and lobe of the lung progress blood vessel segmentation and lung, according to vessel segmentation
Vascular tree is constructed, and tracheae distribution situation is inferred by three-dimensional vascular distribution；Recombine vascular tree and lung split segmentation result into
The segmentation of row lobe of the lung section, finally obtains the segmentation result of candidate region, this method can efficiently control cutting procedure accuracy and
Speed, finally obtained segmentation result accuracy is high, greatly reduces easily existing error in conventional segmentation mode, improves and examine
Disconnected rate, and do not limited by individual lung's morphological differences, the present invention is put forward for the first time in international and national by the big number of LDCT screening
According in conjunction with artificial intelligence technology, for solving the problems, such as that lung section is divided automatically, marked and Lung neoplasm precise positioning.
Detailed description of the invention
Fig. 1 is a kind of lobe of the lung section dividing method flow chart based on CT images provided by the invention；
Fig. 2 is a kind of lobe of the lung section segmenting device module map based on CT images provided by the invention；
Fig. 3 is that a kind of lobe of the lung section based on CT images provided by the invention divides flow diagram.
Specific embodiment
Embodiments of the present invention are illustrated by particular specific embodiment below, those skilled in the art can be by this specification
Revealed content is understood other advantages and efficacy of the present invention easily.Although description of the invention will combine preferred embodiment
It introduces together, but this feature for not representing the invention is only limitted to the embodiment.On the contrary, being invented in conjunction with embodiment
The purpose of introduction is to be possible to the other selections extended or transformation to cover claim of the invention.In order to provide pair
Depth of the invention understands, be described below in will include many concrete details.The present invention can also be real without using these details
It applies.In addition, in order to avoid confusion or obscuring emphasis of the invention, some details will be omitted in the de-scription.
Deep learning is a new field in machine learning research, and motivation is that foundation, simulation human brain are divided
The neural network of study is analysed, it imitates the mechanism of human brain to explain data, such as (" deep learning is so for image, sound and text
Fire, what it can do actually? " selected from " the deep learning world ", date 2016-05-03. is quoted).The present invention provides one kind
CT images lobe of the lung section dividing method, device, system, storage medium and equipment based on deep learning.
The embodiment of the invention provides a kind of CT images lobe of the lung section dividing methods, as shown in Figure 1, this method includes detection step
Rapid 1, screening step 2, segmentation step 3, constitution step 4, integration step 5.Specifically, in detecting step 1, CT images are inputted
Data, and the lung outlines in CT images are detected and exported, the lung outlines of output include intrapulmonary region, lung outskirt
Domain；In screening step 2, lung outlines are inputted, select the mode of machine segmentation to filter out intrapulmonary region, and intrapulmonary region is made
For candidate region；In segmentation step 3, blood vessel segmentation carried out to the lung section of candidate region and the lobe of the lung in 3D level and lung is split point
It cuts, obtains blood vessel segmentation and lung splits segmentation result, specific format is a three-dimensional lattice, has corresponded to the three-dimensional position in CT images
It sets, whether each element value is represented in dot matrix splits algorithm by blood vessel/lung and be determined as that blood vessel/lung is split；In constitution step 4, root
The three-dimensional vascular distribution of lung is obtained by constructing vascular tree according to vessel segmentation；In integration step 5, using the lobe of the lung
Vascular tree and lung are split segmentation result and are combined by section partitioning algorithm, and carry out lobe of the lung section segmentation, obtain the final of candidate region
Segmentation result.The present invention efficiently controls the accuracy and speed of segmentation using dividing layer by layer, improves the efficiency of segmentation and not
It is limited to the limitation of individual lung's morphological differences.
In recent years, based on deep learning image Segmentation Technology (Hariharan, B., Arbel_aez, P.,
Girshick,R.,Malik,J.:Hypercolumns for object segmentation and_ne-grained
Localization.In:Proc.CVPR.pp.447-456 (2015)) it yields unusually brilliant results in major international memory, general deep
It spends in convolutional neural networks (DCNN), several full articulamentums is usually added after convolutional network, so that the feature of convolutional network
Vector (Io_e, S., Szegedy, the C.:Batch normalization:Accelerating of regular length can be mapped to
deep network training by reducing internal covariate shift.CoRR abs/
1502.03167 (2015)), and being finally used.And full convolutional network such as FCN (Long, J., Shelhamer, E.,
Darrell,T.:Fully convolutional networks for semantic segmentation.In:
Proc.CVPR.pp.3431-3440 (2015)) or U-net (Ronneberger, O., Fischer, P., Brox, T.:U-
net:Convolutional networks for biomedical image segmentation.In:MICCAI.LNCS,
Vol.9351, pp.234-241.Springer (2015)), then can to image realize pixel scale classification (Milletari,
F.,Ahmadi,S.,etc.:Hough-cnn:Deep learning for segmentation of deep brain
Regions in MRI and ultrasound.CoRR abs/1601.07014 (2016)) inputs the image of arbitrary size
In FCN network, by convolutional layer feature extraction and return the up-sampling of convolutional layer, final classification results can revert to original image
Size, formed one it is end-to-end (Tran, D., Bourdev, L.D., Fergus, R., Torresani, L., Paluri, M.:
Deep end2end voxel2voxel prediction.CoRR abs/1511.06681 (2015)) network structure
(Seyedhosseini,M.,Sajjadi,M.,Tasdizen,T.:Image segmentation with cascaded
hierarchical models and logistic disjunctive normal networks.In:
Proc.ICCV.pp.2168-2175 (2013)), and image to be detected not only can be divided into prospect background by this classification
Two classes, and the classification of object can be provided automatically.U-net is also more typical one of image segmentation network, and in medicine
The image especially lobe of the lung, lung section segmentation have very big advantage: U-net first can be converted to 3-D (Fedorov, A.,
Beichel,R.,et al.:3D slicer as an image computing platform for the
Quantitative imaging network.J.Magn Reson Imaging 30 (9), 1323-1341 (2012)) level
On image segmentation (Kleesiek, J., Urban, G., Hubert, A., Schwarz, D., Maier-Hein, K.,
Bendszus,M.,Biller,A.:Deep mri brain extraction:A 3d convolutional neural
Network for skull stripping.NeuroImage (2016)), at the same U-net presented in medical domain it is huge
Big achievement, if kidney portion lesion segmentation detects, retinal vessel segmentation etc..It is calculated using the image segmentation in deep learning field
Method can greatly reduce error, improve diagnosis efficiency and accuracy.
Specifically, it will be understood by those skilled in the art that conventional segmentation mode is by Hu Shiying et al. (Hu
Shiying,Hoffman E A,Reinhardt J M.Automatic Lung segmentation for accurate
quantitation of volumetric X-ray CT images[J].IEEE Trans on Medical Imaging,
2001,20 (6): 490-498)) start one classical lung segmentation process, i.e. gray level threshold segmentation, Connected degree analysis, left
Right lung separation, morphology are closed intrapulmonary high density structures.The process uses a kind of iteration threshold optimization method, then looks for most
Big two connected components, when adhesion of lung, are isolated by the method for continuous morphological erosion, and such methods are referred to as classical
Lung dividing method.And the present invention is in detecting step 1, using the full convolutional neural networks for being different from classical dividing method
(Fully Convolutional Networks, FCN), with traditional convolutional neural networks (Convolutional Neural
Network, CNN) method of image segmentation compares, and FCN has apparent advantage: (1) can receive the input figure of arbitrary size
Picture, and do not have to require all training image and test image that there is same size；(2) more efficiently because avoid by
In use block of pixels, bring repeats the problem of storing and calculating convolution.Present invention employs mature FCN networks to be divided
It cuts, wherein main structure of the multireel lamination as FCN network, and traditional full linking layer of neural network is replaced with into convolutional layer,
Guarantee whole network is full convolutional network structure.
In order to guarantee that the resolution ratio for exporting result is consistent with input picture, by last several layers of turnup lamination by network
Output characteristic pattern is extended, and is expanded to consistent with input picture.And the present invention passes through the following aspects, in lung CT shadow
It as upper progress re -training or finely tunes to promote the accuracy of segmentation: (1) being cut using the direction lung's CT images XY (as shown in Figure 3)
Face marks out intrapulmonary region as training mark as training image；(2) it amount of training data: is adopted at random from 10K magnitude CT
Sample generates 100K magnitude CT tangent plane picture；(3) it marks accuracy: needing to guarantee training mark error within 3mm；(4) it needs
Specific aim guarantees the data volume in the specific region CT (apex pulmonis, diaphragm, outside lung)；(5) needing to guarantee specific CT type distribution, (patient is not
Same age bracket, illness, CT dosage etc.), so that lung outlines are efficiently exported, since human body distributed architecture similitude is higher, CT
Lung outlines in image are relatively clear, so this process of detecting step 1 can very quickly.
Then in screening step 2, the mode of machine segmentation is selected to filter out intrapulmonary region, machine segmentation is based on machine
The automatic lung segmentation of study, and using intrapulmonary region as candidate region, for the subsequent emphasis screening further divided, effectively
Reduce noise, accelerates detection speed.Wherein machine learning is a multi-field cross discipline, be related to probability theory, statistics,
The multiple subjects such as Approximation Theory, convextiry analysis, algorithm complexity theory.Specialize in the study that the mankind were simulated or realized to computer how
Behavior reorganizes the existing structure of knowledge and is allowed to constantly improve the performance of itself to obtain new knowledge or skills.
Further, in segmentation step 3, blood is carried out respectively to lung section and the lobe of the lung using 3D U-net and along two branch lines
Pipe segmentation splits segmentation with lung.Two of them branch line is obtained by screening candidate region (intrapulmonary region), that is, candidate regions
The edge in domain.It will be understood by those skilled in the art that encoder gradually decreases the Spatial Dimension of pond layer, decoder is gradually repaired
The details and Spatial Dimension of object.Usually there is quick connection between encoder and decoder, therefore decoder can be helped more preferable
The details of ground reparation target.U-Net is most common structure in this method.3D U-net is full convolution deep neural network
One kind is a kind of realization for traditional FCN network, can transmit top-down network and bottom-up network by great-jump-forward
(Skip connection) connection is believed with guaranteeing that final output layer can obtain high-order semanteme and low order local grain simultaneously simultaneously
Breath.Blood vessel segmentation and lung, which split segmentation and be all made of 3D U-net, to be split, and the two is using identical algorithm frame and network
Structure, difference are the mark semantic differential inputted when training, the former needs to mark out blood vessel, and the latter needs to mark out lung
It splits.Lung section and the lobe of the lung are split by 3D U-net, obtain the distribution of lung section medium vessels and 5 independent space (i.e. lungs
Leaf) and diaphragm.
Specifically, in constitution step 4, on the basis of vessel segmentation, vascular tree is constructed, obtains lung intersegmental part three
Vascular distribution is tieed up, since 3D U-net is a 3D parted pattern (as shown in Figure 3), input is the segmentation knot of a 3d space
Fruit, i.e., in 3d space each point whether be blood vessel confidence level.When constructing vascular tree, first according to a fixed threshold
Segmentation result is carried out binaryzation by value, and is smoothed, and then finds out angiosomes most by flood fill algorithm
Big connected component, as the vascular tree constructed, the i.e. final vessel segmentation of output, the confidence level based on 3d space point
Binary conversion treatment is carried out, where concrete decision be blood vessel.According to the anatomical configurations of intrapulmonary, in intrapulmonary Pulmonary Vascular and tracheae
Trend has strong consistency, and often lung's bronchuses at different levels nearby have blood vessel adjoint.However in lung CT imaging, Pulmonary Vascular
It is visually more significant, therefore algorithm is and to obtain pulmonary branches tracheorrhaphy according to the position approximation of blood vessel by detecting vascular tree
Cloth, to extract the characteristic informations such as bronchial tree distribution and shape, the information for subsequent step is integrated.Wherein, flood
Fill algorithm is to extract the points of several connections from a region to distinguish from other adjacent areas and (or dye different face respectively
Color) classic algorithm because its thinking similar to flood from a regional diffusion to it is all can reach regions and gain the name.
Further, in integration step 5, lobe of the lung section partitioning algorithm is disaggregated model neural network based.Algorithm is defeated
Entering feature is the relative position of intrapulmonary specified point (as shown in figure 3, calculating the region in the phase in the direction XYZ based on lung segmentation result
Comparative example), the blood vessel and lung of the areas adjacent split result (indicating using 64 × 64 × 64 cubes), split apart from nearest lung
With vector (position+distance) information of blood vessel.Algorithm trains lobe of the lung section disaggregated model by learning above-mentioned input, thus by blood
Guan Shu and lung split segmentation result and are combined.Algorithm output is the classification results of the position lobe of the lung and lung section.The present invention can be accurate
Ground controls the accuracy and speed of cutting procedure, for example, it is desired to guarantee amount of training data and mark accuracy, amount of training data is needed
Guarantee 100K magnitude intrapulmonary point (lobe of the lung belonging to marking and lung section), and guarantees to come from 10K magnitude CT data, and guarantee CT class
Type is distributed (patient's different age group, illness, CT dosage etc.)；Need to guarantee to mark precision 95% or more.It can be by using
The network structure (reducing convolutional layer depth and width) suitably simplified and shared adjacent domain feature are operated together to reduce and calculate
Amount, to achieve the purpose that acceleration.
Wherein, it is distributed about CT type, it, first should be according to patient age before carrying out the lobe of the lung section segmentation of CT images
Section, illness, CT dosage (such as routine CT, LDCT) are classified；Secondly it is split, can effectively subtract for same CT type
Less because of different CT type bring errors, while data statistics can be targetedly carried out, be diagnosis related disease (as shifted
Tumor, celiothelioma, Lung neoplasm etc.) valuable reference frame is provided.
As shown in figure 3, the preferred embodiment of the present invention inputs CT images data first, using FCN network in CT images
Lung outlines are detected and are exported, and the lung outlines of output include intrapulmonary region, lung exterior domain；Secondly machine segmentation is selected
Mode filters out intrapulmonary region as candidate region；Again, in 3D level, using 3D U-net to the lung section and lung of candidate region
Leaf carries out blood vessel segmentation respectively and lung splits segmentation, respectively obtains blood in 5 independent spaces (i.e. the lobe of the lung) and diaphragm and lung section
Pipe distribution, and vascular tree is constructed according to the result of lung section medium vessels distribution；Finally blood vessel segmentation and lung is combined to split segmentation result simultaneously
Lobe of the lung section partitioning algorithm carries out the segmentation of lobe of the lung section, and lobe of the lung section partitioning algorithm is disaggregated model neural network based, and algorithm passes through
Learn above-mentioned input and train lobe of the lung section disaggregated model, is combined so that vascular tree and lung are split segmentation result.Algorithm output
For the classification results of the position lobe of the lung and lung section, precisely and rapidly obtain having multiple cut zone (region 1 in such as Fig. 3,3,
4 etc.) result.
Therefore, the lobe of the lung section dividing method of a kind of CT images proposed by the present invention, using the form divided layer by layer, and introduces
FCN and 3D U-net is split respectively, and finally obtained segmentation result accuracy is high, is greatly reduced in conventional segmentation mode
Easily existing error, improves diagnosis.
The embodiment of the invention provides a kind of lobe of the lung section segmenting devices of CT images based on deep learning, as shown in Fig. 2,
The device includes detection module 10, screening module 20, segmentation module 30, constructing module 40, integrates module 50.
Detection module 10, in CT images, detection output lung outlines, lung outlines include intrapulmonary region, outside lung
Region；
Screening module 20 selects the mode of machine segmentation to filter out intrapulmonary region in lung outlines, and by intrapulmonary
Region is as candidate region；
Divide module 30, for carrying out blood vessel segmentation and lung in the 3D level of candidate region, while to the lobe of the lung and lung section
Split segmentation；
Constructing module 40, for obtaining the three-dimensional blood vessel point of lung by constructing vascular tree according to vessel segmentation
Cloth；
Module 50 is integrated, is combined, goes forward side by side for vascular tree and lung to be split segmentation result using lobe of the lung section partitioning algorithm
The segmentation of row lobe of the lung section, obtains the final segmentation result of candidate region.
Further, it in detection module 10, is split using mature FCN network, wherein multireel lamination conduct
The main structure of FCN network, and traditional full linking layer of neural network is replaced with into convolutional layer, guarantee whole network is full convolution
Network structure efficiently and fast exports the lung outlines in CT images.
Then in screening step 2, the mode of machine segmentation is selected to filter out intrapulmonary region, and using intrapulmonary region as time
Favored area, for the subsequent emphasis screening further divided.
Further, in segmentation module 30, using 3D U-net and along two branch lines to the lung section and the lobe of the lung
The blood vessel segmentation is carried out respectively and the lung splits segmentation.That is in the 3D level of candidate region, pass through 3D U-net
Lung section and the lobe of the lung are split, the distribution of lung section medium vessels and 5 independent spaces (i.e. the lobe of the lung) and diaphragm are obtained.
Further, in constructing module 40, on the basis of vessel segmentation, vascular tree is constructed, obtains lung section
Interior three-dimensional vascular distribution.According to the anatomical configurations of intrapulmonary, there is strong consistency in the trend of intrapulmonary Pulmonary Vascular and tracheae, it is past
Nearby there is blood vessel adjoint toward lung's bronchuses at different levels.However in lung CT imaging, Pulmonary Vascular is visually more significant, because
This algorithm is to obtain the distribution of pulmonary branches tracheae by detecting vascular tree, and according to the position approximation of blood vessel.
Further, in integrating module 50, lobe of the lung section partitioning algorithm is disaggregated model neural network based.Algorithm is defeated
Entering feature is the relative position of intrapulmonary specified point (as shown in figure 3, calculating the region in the phase in the direction XYZ based on lung segmentation result
Comparative example), the blood vessel and lung of the areas adjacent split result (indicating using 64 × 64 × 64 cubes), split apart from nearest lung
With vector (position+distance) information of blood vessel.Algorithm trains lobe of the lung section disaggregated model by learning above-mentioned input, thus by blood
Guan Shu and lung split segmentation result and are combined.Algorithm output is the classification results of the position lobe of the lung and lung section.
Therefore, a kind of lobe of the lung section segmenting device based on CT images of the embodiment of the present invention, using the form divided layer by layer,
And introduce FCN and 3D U-net and be split respectively, finally obtained segmentation result accuracy is high, greatly reduces conventional segmentation
Easily existing error, improves diagnosis in mode.
The embodiment of the invention also provides a kind of systems of lobe of the lung section segmentation realized based on CT images, including aforementioned implementation
The lobe of the lung section segmenting device based on CT images in mode.
The embodiment of the present invention provides a kind of non-volatile memory medium again, which, which has, stores it
In instruction, when the instruction is performed, so that processor executes the lobe of the lung section dividing method based on CT images, the instruction packet
It includes:
Detection instruction, in CT images, detection output lung outlines, lung outlines include intrapulmonary region, lung exterior domain；
Screening instruction selects the mode of machine segmentation to filter out intrapulmonary region, and intrapulmonary region is made in lung outlines
For candidate region；
Split order in the 3D level of candidate region, while carrying out blood vessel segmentation to the lobe of the lung and lung section and splitting segmentation with lung；
Construction instruction, obtains the three-dimensional vascular distribution of lung by constructing vascular tree according to vessel segmentation；
Integration instruction, splits segmentation result for vascular tree and lung using lobe of the lung section partitioning algorithm and is combined, and carry out the lobe of the lung
Section segmentation, obtains the final segmentation result of candidate region.
The embodiment of the present invention also provides a kind of equipment, including memory and processor, and memory is stored with computer can
It executes instruction, processor is configured as executing process of the instruction to implement the lobe of the lung section segmentation based on CT images, the process packet
It includes:
Detecting step, in CT images, detection output lung outlines, lung outlines include intrapulmonary region, lung exterior domain；
Screening step selects the mode of machine segmentation to filter out intrapulmonary region, and intrapulmonary region is made in lung outlines
For candidate region；
Segmentation step in the 3D level of candidate region, while carrying out blood vessel segmentation to the lobe of the lung and lung section and splitting segmentation with lung；
Constitution step obtains the three-dimensional vascular distribution of lung by constructing vascular tree according to vessel segmentation；
Vascular tree and lung are split segmentation result using lobe of the lung section partitioning algorithm and are combined by integration step, and carry out the lobe of the lung
Section segmentation, obtains the final segmentation result of candidate region.
In conclusion a kind of lobe of the lung section dividing method of CT images based on deep learning provided by the invention, device, being
System, storage medium and equipment use FCN network of the multilayer convolutional layer as main structure first, accurately detect and export CT
Lung outlines in image；Secondly screening candidate region and select 3D U-net expansion to lung section and the lobe of the lung carry out blood vessel segmentation and
Lung splits segmentation, constructs vascular tree according to vessel segmentation, and has vascular distribution to infer tracheae distribution situation；Recombine blood
Guan Shu and lung split segmentation result and carry out the segmentation of lobe of the lung section, finally obtain the segmentation result of candidate region, this method can be controlled efficiently
The accuracy and speed of cutting procedure processed, finally obtained segmentation result accuracy is high, greatly reduces in conventional segmentation mode
Easily existing error, improves diagnosis, and do not limited by individual lung's morphological differences, and the present invention is in international and national head
LDCT screening big data in conjunction with artificial intelligence technology, is divided automatically for solving lung section, is marked and Lung neoplasm essence by secondary proposition
Quasi- orientation problem.
The above-described embodiments merely illustrate the principles and effects of the present invention, and is not intended to limit the present invention.It is any ripe
The personage for knowing this technology all without departing from the spirit and scope of the present invention, carries out modifications and changes to above-described embodiment.Cause
This, institute is complete without departing from the spirit and technical ideas disclosed in the present invention by those of ordinary skill in the art such as
At all equivalent modifications or change, should be covered by the claims of the present invention.
Claims (10)
1. the output method in a kind of lobe of the lung section segmentation of CT images, which is characterized in that method includes the following steps:
Detecting step, in CT images, detection output lung outlines, the lung outlines include intrapulmonary region, lung exterior domain；
Screening step selects the mode of machine segmentation to filter out the intrapulmonary region in the lung outlines, and by the lung
Inner region is as candidate region；
Segmentation step in the 3D level of the candidate region, while carrying out blood vessel segmentation to lung section and the lobe of the lung and splitting segmentation with lung；
Constitution step obtains the three-dimensional vascular distribution of lung by constructing vascular tree according to the vessel segmentation；
The vascular tree and the lung are split segmentation result using lobe of the lung section partitioning algorithm and are combined, and carried out by integration step
The segmentation of lobe of the lung section, obtains the final segmentation result of the candidate region,
Wherein, in the detecting step, using the direction lung CT images XY section as training image, and intrapulmonary area is marked out
Domain is as training mark；Amount of training data；Mark accuracy.
2. the output method in the lobe of the lung section segmentation of CT images according to claim 1, which is characterized in that in the detection
The FCN network of the structure based on multilayer convolutional layer is used to be detected and exported in step.
3. the output method in the lobe of the lung section segmentation of CT images according to claim 1, which is characterized in that in the segmentation
In step, using 3D U-net and along two branch lines to the lung section and the lobe of the lung carry out respectively the blood vessel segmentation with it is described
Lung splits segmentation.
4. the output method in the lobe of the lung section segmentation of CT images according to claim 1, which is characterized in that in the construction
In step, tracheae distribution situation is inferred according to the three-dimensional vascular distribution, obtains bronchial tree.
5. the output method in the lobe of the lung section segmentation of CT images according to claim 1, which is characterized in that in the integration
In step, the lobe of the lung section partitioning algorithm splits segmentation knot by input lobe of the lung section disaggregated model, by the vascular tree and the lung
Fruit integration, and export final segmentation result.
6. the output device in a kind of lobe of the lung section segmentation of CT images, which is characterized in that the device includes:
Detection module, in CT images, detection output lung outlines, the lung outlines to include intrapulmonary region, lung outskirt
Domain；
Screening module selects the mode of machine segmentation to filter out the intrapulmonary region in the lung outlines, and by institute
Intrapulmonary region is stated as candidate region；
Divide module, for splitting in the 3D level of the candidate region, while to lung section and lobe of the lung progress blood vessel segmentation with lung
Segmentation；
Constructing module, for obtaining the three-dimensional blood vessel point of lung by constructing vascular tree according to the vessel segmentation
Cloth；
Module is integrated, is combined for the vascular tree and the lung to be split segmentation result using lobe of the lung section partitioning algorithm, and
The segmentation of lobe of the lung section is carried out, the final segmentation result of the candidate region is obtained,
Wherein, in the detecting step, using the direction lung CT images XY section as training image, and intrapulmonary area is marked out
Domain is as training mark；Amount of training data；Mark accuracy.
7. the output device in the lobe of the lung section segmentation of CT images according to claim 6, which is characterized in that in the detection
The FCN network of the structure based on multilayer convolutional layer is used to be detected and exported in module.
8. the output device in the lobe of the lung section segmentation of CT images according to claim 6, which is characterized in that in the segmentation
In module, using 3D U-net and along two branch lines to the lung section and the lobe of the lung carry out respectively the blood vessel segmentation with it is described
Lung splits segmentation.
9. the output device in the lobe of the lung section segmentation of CT images according to claim 6, which is characterized in that in the construction
In module, tracheae distribution situation is inferred according to the three-dimensional vascular distribution, obtains bronchial tree.
10. the output device in the lobe of the lung section segmentation of CT images according to claim 6, which is characterized in that described whole
It molds in block, the vascular tree and the lung are split segmentation by input lobe of the lung section disaggregated model by the lobe of the lung section partitioning algorithm
As a result it integrates, and exports final segmentation result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811572365.6A CN109685787A (en) | 2018-12-21 | 2018-12-21 | Output method, device in the lobe of the lung section segmentation of CT images |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811572365.6A CN109685787A (en) | 2018-12-21 | 2018-12-21 | Output method, device in the lobe of the lung section segmentation of CT images |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109685787A true CN109685787A (en) | 2019-04-26 |
Family
ID=66188185
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811572365.6A Withdrawn CN109685787A (en) | 2018-12-21 | 2018-12-21 | Output method, device in the lobe of the lung section segmentation of CT images |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109685787A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110956635A (en) * | 2019-11-15 | 2020-04-03 | 上海联影智能医疗科技有限公司 | Lung segment segmentation method, device, equipment and storage medium |
| CN111161270A (en) * | 2019-12-24 | 2020-05-15 | 上海联影智能医疗科技有限公司 | Blood vessel segmentation method for medical image, computer device and readable storage medium |
| CN111275722A (en) * | 2020-02-18 | 2020-06-12 | 广州柏视医疗科技有限公司 | Lung segment and liver segment segmentation method and system |
| CN111415742A (en) * | 2020-03-17 | 2020-07-14 | 北京青燕祥云科技有限公司 | Calculation method for predicting position of lung segment of focus through lung lobe |
| CN111681247A (en) * | 2020-04-29 | 2020-09-18 | 杭州深睿博联科技有限公司 | Lung lobe and lung segment segmentation model training method and device |
| WO2020238043A1 (en) * | 2019-05-24 | 2020-12-03 | 深圳市前海安测信息技术有限公司 | Unet network-based lung lobe segmentation method and apparatus, computer readable storage medium |
| CN113344881A (en) * | 2021-06-09 | 2021-09-03 | 西安智诊智能科技有限公司 | Lung lobe image segmentation method based on deep learning |
| CN113516669A (en) * | 2021-06-23 | 2021-10-19 | 湖北英库科技有限公司 | Trachea extraction method, device and equipment based on CT image and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2372643A3 (en) * | 2010-02-15 | 2017-01-18 | General Electric Company | System and method of pulmonary emboli detection for computed tomography |
| CN106485714A (en) * | 2016-09-07 | 2017-03-08 | 中国科学院深圳先进技术研究院 | 3-dimensional image segmenting system and its dividing method |
| CN107909581A (en) * | 2017-11-03 | 2018-04-13 | 杭州依图医疗技术有限公司 | Lobe of the lung section dividing method, device, system, storage medium and the equipment of CT images |
-
2018
- 2018-12-21 CN CN201811572365.6A patent/CN109685787A/en not_active Withdrawn
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2372643A3 (en) * | 2010-02-15 | 2017-01-18 | General Electric Company | System and method of pulmonary emboli detection for computed tomography |
| CN106485714A (en) * | 2016-09-07 | 2017-03-08 | 中国科学院深圳先进技术研究院 | 3-dimensional image segmenting system and its dividing method |
| CN107909581A (en) * | 2017-11-03 | 2018-04-13 | 杭州依图医疗技术有限公司 | Lobe of the lung section dividing method, device, system, storage medium and the equipment of CT images |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020238043A1 (en) * | 2019-05-24 | 2020-12-03 | 深圳市前海安测信息技术有限公司 | Unet network-based lung lobe segmentation method and apparatus, computer readable storage medium |
| CN110956635A (en) * | 2019-11-15 | 2020-04-03 | 上海联影智能医疗科技有限公司 | Lung segment segmentation method, device, equipment and storage medium |
| CN110956635B (en) * | 2019-11-15 | 2023-12-01 | 上海联影智能医疗科技有限公司 | Lung segment segmentation method, device, equipment and storage medium |
| CN111161270A (en) * | 2019-12-24 | 2020-05-15 | 上海联影智能医疗科技有限公司 | Blood vessel segmentation method for medical image, computer device and readable storage medium |
| CN111161270B (en) * | 2019-12-24 | 2023-10-27 | 上海联影智能医疗科技有限公司 | Vascular segmentation method for medical image, computer device and readable storage medium |
| CN111275722A (en) * | 2020-02-18 | 2020-06-12 | 广州柏视医疗科技有限公司 | Lung segment and liver segment segmentation method and system |
| CN111415742A (en) * | 2020-03-17 | 2020-07-14 | 北京青燕祥云科技有限公司 | Calculation method for predicting position of lung segment of focus through lung lobe |
| CN111681247A (en) * | 2020-04-29 | 2020-09-18 | 杭州深睿博联科技有限公司 | Lung lobe and lung segment segmentation model training method and device |
| CN111681247B (en) * | 2020-04-29 | 2023-08-25 | 杭州深睿博联科技有限公司 | Lung lobe lung segment segmentation model training method and device |
| CN113344881A (en) * | 2021-06-09 | 2021-09-03 | 西安智诊智能科技有限公司 | Lung lobe image segmentation method based on deep learning |
| CN113516669A (en) * | 2021-06-23 | 2021-10-19 | 湖北英库科技有限公司 | Trachea extraction method, device and equipment based on CT image and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107909581B (en) | Lobe of the lung section dividing method, device, system, storage medium and the equipment of CT images | |
| CN109685787A (en) | Output method, device in the lobe of the lung section segmentation of CT images | |
| Su et al. | Lung nodule detection based on faster R-CNN framework | |
| CN109035255B (en) | Method for segmenting aorta with interlayer in CT image based on convolutional neural network | |
| CN111192245B (en) | Brain tumor segmentation network and method based on U-Net network | |
| CN105574859B (en) | A kind of liver neoplasm dividing method and device based on CT images | |
| Wen et al. | A novel statistical cerebrovascular segmentation algorithm with particle swarm optimization | |
| Mu et al. | Segmentation of kidney tumor by multi-resolution VB-nets | |
| CN113744272B (en) | Automatic cerebral artery delineation method based on deep neural network | |
| CN108010013A (en) | A kind of lung CT image pulmonary nodule detection methods | |
| CN107330953A (en) | A kind of Dynamic MRI method for reconstructing based on non-convex low-rank | |
| Wang et al. | CLCU-Net: cross-level connected U-shaped network with selective feature aggregation attention module for brain tumor segmentation | |
| Ye et al. | Medical image diagnosis of prostate tumor based on PSP-Net+ VGG16 deep learning network | |
| Liu et al. | Automatic segmentation algorithm of ultrasound heart image based on convolutional neural network and image saliency | |
| Wang et al. | Accurate tumor segmentation via octave convolution neural network | |
| Xiaojie et al. | Segmentation of the aortic dissection from CT images based on spatial continuity prior model | |
| Zhao et al. | Automated coronary tree segmentation for x-ray angiography sequences using fully-convolutional neural networks | |
| Affane et al. | Robust deep 3-d architectures based on vascular patterns for liver vessel segmentation | |
| Wang et al. | A myocardial segmentation method based on adversarial learning | |
| Guo et al. | ELTS-Net: An enhanced liver tumor segmentation network with augmented receptive field and global contextual information | |
| Qiu et al. | Deep multi-scale dilated convolution network for coronary artery segmentation | |
| Wang et al. | Research on Computer Aided Detection and Segmentation of Pulmonary Nodules Based on Deep Learning and Generative Adversarial Networks | |
| CN112233058A (en) | Method for detecting lymph nodes in head and neck CT image | |
| Lewis et al. | The role of anatomical context in soft‐tissue multi‐organ segmentation of cadaveric non‐contrast‐enhanced whole body CT | |
| Zrira et al. | Automatic and Fast Whole Heart Segmentation for 3D Reconstruction |
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 | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20190426 |