CN106875409B - A kind of light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method - Google Patents
A kind of light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method Download PDFInfo
- Publication number
- CN106875409B CN106875409B CN201710181583.6A CN201710181583A CN106875409B CN 106875409 B CN106875409 B CN 106875409B CN 201710181583 A CN201710181583 A CN 201710181583A CN 106875409 B CN106875409 B CN 106875409B
- Authority
- CN
- China
- Prior art keywords
- region
- sorted
- image
- feature
- sticking patch
- 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.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 39
- 206010021620 Incisional hernias Diseases 0.000 title claims abstract description 28
- 238000002604 ultrasonography Methods 0.000 title claims abstract description 27
- 210000003195 fascia Anatomy 0.000 claims abstract description 24
- 238000000605 extraction Methods 0.000 claims abstract description 18
- 238000012545 processing Methods 0.000 claims abstract description 10
- 238000001514 detection method Methods 0.000 claims description 12
- 239000000523 sample Substances 0.000 claims description 10
- 230000007613 environmental effect Effects 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 9
- 238000001228 spectrum Methods 0.000 claims description 7
- 230000009466 transformation Effects 0.000 claims description 6
- 238000001914 filtration Methods 0.000 claims description 5
- 239000011159 matrix material Substances 0.000 claims description 5
- 230000003595 spectral effect Effects 0.000 claims description 5
- 238000012217 deletion Methods 0.000 claims description 4
- 230000037430 deletion Effects 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 230000004807 localization Effects 0.000 claims description 3
- 238000011946 reduction process Methods 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims description 2
- 230000009467 reduction Effects 0.000 claims description 2
- 101150052147 ALLC gene Proteins 0.000 claims 1
- 235000013399 edible fruits Nutrition 0.000 claims 1
- 230000008602 contraction Effects 0.000 abstract description 4
- 230000000694 effects Effects 0.000 abstract description 4
- 230000002980 postoperative effect Effects 0.000 abstract description 3
- 210000000481 breast Anatomy 0.000 abstract description 2
- 230000004069 differentiation Effects 0.000 abstract description 2
- 238000012360 testing method Methods 0.000 description 9
- 210000001519 tissue Anatomy 0.000 description 9
- 239000000284 extract Substances 0.000 description 6
- 238000012549 training Methods 0.000 description 6
- 238000002474 experimental method Methods 0.000 description 5
- 210000002784 stomach Anatomy 0.000 description 5
- 238000002513 implantation Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 206010019909 Hernia Diseases 0.000 description 2
- 210000001015 abdomen Anatomy 0.000 description 2
- 210000003484 anatomy Anatomy 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 239000002775 capsule Substances 0.000 description 2
- 230000007812 deficiency Effects 0.000 description 2
- 238000009499 grossing Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 230000003902 lesion Effects 0.000 description 2
- 230000008439 repair process Effects 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 238000010408 sweeping Methods 0.000 description 2
- 206010060954 Abdominal Hernia Diseases 0.000 description 1
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 208000035091 Ventral Hernia Diseases 0.000 description 1
- 210000003815 abdominal wall Anatomy 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 239000006071 cream Substances 0.000 description 1
- 238000003748 differential diagnosis Methods 0.000 description 1
- 238000009792 diffusion process Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 210000004907 gland Anatomy 0.000 description 1
- 238000007912 intraperitoneal administration Methods 0.000 description 1
- 210000003205 muscle Anatomy 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000001356 surgical procedure Methods 0.000 description 1
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
-
- G06T5/70—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- 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
- G06T2207/10132—Ultrasound image
- G06T2207/10136—3D ultrasound image
-
- 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
- G06T2207/20036—Morphological image processing
-
- 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
- G06T2207/20092—Interactive image processing based on input by user
- G06T2207/20104—Interactive definition of region of interest [ROI]
Abstract
The invention belongs to technical field of image processing, specially a kind of light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method.The present invention quantify to extract from the associated texture characteristic parameter in region to be sorted in the three-dimensional volume of interest (VOI) of dynamicization three-dimensional breast ultrasound (ABUS) image automatically using textural characteristics extraction algorithm first, for use in the differentiation to sticking patch and fascia;Then for spatial alternations more sensitive issues such as the curling postoperative to incisional hernia sticking patch of 2 d texture parameter, contractions, three-D grain parameter and three-dimensional position parameter are introduced to improve the robustness of light-type sticking patch classification and identification algorithm;Finally feature selecting is carried out using class spacing algorithm and sequential advancement search method.The method of the present invention feature selecting effect is good, high-efficient, can effectively improve the nicety of grading of light-type incisional hernia sticking patch three-dimensional ultrasound pattern, is convenient for automatic Classification and Identification.
Description
Technical field
The invention belongs to technical field of image processing, and in particular to a kind of light-type incisional hernia sticking patch three-dimensional ultrasound pattern is special
Levy extracting method.
Background technique
Light-type sticking patch thickness usually less than 0.5mm, weaves aperture in 3mm or more, foreign matter residual quantity very little leads to its time
Acoustical signal is weaker, and the high echo picture of the linear in imaging results is not obvious enough.Meanwhile in ventral hernia repair art, light-type
Sticking patch can be placed on gap (Sublay) and abdomen before stomach wall fascia between (Onlay), flesh between (Inlay), fascia and peritonaeum
Intracavitary peritonaeum inner wall (IPOM) four position (as shown in Figure 2) relevant to stomach wall anatomical structure level, and all implantation positions
It sets and is all closer to fascia tissue.However, in the cross section view of HHUS (hand-held ultrasound, two-dimentional hand-held ultrasound)
In (cross section and/or sagittal plane), since light-type sticking patch and fascia tissue are usually all shown as a linear high echogenic area
Domain [9].Therefore, what the linear marking feature of HHUS cross section view was reflected is that light-type sticking patch and fascia tissue are overlapped
Hybrid texture attribute.Fascia can be regarded as anatomy noise when light-type sticking patch identifies, and not only be harmful to artificial detection, more
The linear marking is greatly reduced in computer picture recognition to the differential diagnosis value of fascia and light-type sticking patch.
Since light-type sticking patch and fascia tissue are overlapped in HHUS image, it is difficult light-type sticking patch from its week
It is distinguished in the fascia tissue enclosed.However, ABUS (automated 3-D breast ultrasound, automatized three-dimensional cream
Gland is ultrasonic) additional diagnostic message is provided for the identification of light-type sticking patch by its coronal-plane (surgical plane), new is coronal
Face view can isolate light-type sticking patch from fascia tissue, so that presenting significant light-type sticking patch mesh texture can
Depending on changing.Therefore, ABUS provides a possibility that selectively analyzing light-type sticking patch coronal-plane textural characteristics, has and provides more
Add accurate characteristic parameter to describe the potential of mesh texture, can inherently be provided for the identification of light-type sticking patch more acurrate
Characteristic parameter.
Summary of the invention
It is an object of the invention to propose a kind of light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method, so as to
Light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature can be extracted, accurately convenient for classification is known automatically in subsequent processes
Not.
Light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method proposed by the present invention uses textural characteristics first
Extraction algorithm quantifies to extract region to be sorted in the VOI (volume of interest, volume of interest) of ABUS image automatically
Associated texture characteristic parameter, for use in the differentiation to sticking patch and fascia;Then incisional hernia is mended for 2 d texture parameter
The spatial alternations more sensitive issues such as the postoperative curling of piece, contraction, introduce three-D grain parameter and three-dimensional position parameter, to mention
The robustness of high light-type sticking patch classification and identification algorithm;Finally feature is carried out using class spacing algorithm and sequential advancement search method
Selection.
The specific steps of the present invention are as follows:
S1: 2-D (two-dimensional, two dimension) foreground mask of one ABUS coronal-plane of building;
S2: VOI is extracted;
S3: pretreatment operation is carried out to VOI image;
S4: after pretreatment, giving the region to be sorted of one group of VOI, carries out one by one to each region to be sorted special
Sign is extracted, and extracts 40 features in total to each region to be sorted;
S5: carrying out feature selecting to 40 features of extraction, final to choose 11 classification for making light-type sticking patch and fascia
Error reaches the smallest feature and combines as feature.
The relevant technical details being related to below with regard to step in the method for the present invention are further described specifically:
1. in step S1, the 2-D foreground mask of one coronal-plane of the building the steps include:
S1.1: all ABUS coronal-planes for being located at 0.5 to 0.9 times of scanning total depth are sliced C1-CnIt takes out, to all C1-
CnThe pixel of same position does average value processing in image, obtains a coronal-plane mean value image Cmean;
S1.2: using Otsu algorithm to image CmeanThreshold process is carried out, bianry image C is obtainedbinary;
S1.3: using morphology opening operation to CbinaryConnected region in image carries out edge smoothing processing, and most to it
The black cavity as caused by lesion or shade is filled in big white connected region, obtains coronal-plane 2-D foreground mask image
Cmask。
2. in step S2, the extraction VOI the steps include:
S2.1: from ABUS coronal-plane mask images CmaskThe upper left corner starts, and image is divided into the image of 50 × 50 pixels
Block;Encountered at boundary on the right and lower boundary pixel deficiency can not piecemeal when, then add appropriate black region;
S2.2: by the coronal-plane position feature of all black picture blocks (the sum of binaryzation value of all pixels point is 0)
0 is set to, by all image blocks with the intersection of foreground mask boundary (the sum of binaryzation value of all pixels point is between 1 to 49)
Coronal-plane position feature be set to 1, by the coronal of all full white image blocks (the sum of binaryzation value of all pixels point be 50)
Face position feature is set to 2;
S2.3: the image block that all coronal-plane position features are not 0 is chosen for current ROI (region of one by one
Interest, current interest region);Using all cross sections ABUS relevant to ROI and sagittal plane image-region, by 2-D
ROI be extended to the VOI of 3-D;Current VOI is sent into subsequent characteristics extraction module one by one, until completing time to all VOI
It goes through.
3. it is described that pretreatment operation is carried out to VOI image in step S3, it the steps include:
S3.1: 3-D ISRAD (intelligent speckle reducing anisotropic is used
Diffusion, intelligent spot noise reduction anisotropy parameter) VOI progress three-dimensional filtering processing of the algorithm [1] to automatically extracting out,
It is thin such as light-type sticking patch reticular texture without destroying to filter out speckle noise contained in ABUS image homogeneous region as far as possible
Save feature;
S3.2: binary conversion treatment is done to the single frames cross-sectional image for being located at the center VOI using Otsu algorithm [8], is obtained standby
Select white connected region;
S3.3: the white connected region using all areas in opening operation deletion bianry image less than 15 pixels;
S3.4: extracting the circumscribed bounding box of minimum of each white connected region, calculate the width of bounding box, height and
Apex coordinate;
S3.5: the white connected region of width < 15 pixels of bounding box is deleted;
S3.6: being numbered mark to all remaining white connected regions, using as region to be sorted.
4. it is described that feature extraction is carried out to each region to be sorted one by one in step S4, it the steps include:
S4.1: the 2-D textural characteristics in region to be sorted are extracted;
S4.2: the 3-D textural characteristics in region to be sorted are extracted;
S4.3: extracting the local feature of scan depths locating for region to be sorted, is based on sweeping locating for region to be sorted with one
Retouch the local feature f in the direction Y- of the i.e. ABUS of depthdepthTo characterize the occurrence probability of sticking patch;
S4.4: extracting the environmental characteristic in region to be sorted Yu hernical sac positional relationship, is based on region to be sorted and hernia with one
The location parameter f of capsule positional relationshipadjacencyTo characterize the occurrence probability of sticking patch.
Here it is main using by theoretical proof effectively and GLCM (the gray level co- that is used widely
Occurrence matrix, gray level co-occurrence matrixes) gray value in area image to be sorted converts texture information by [2].
Meanwhile for light-type sticking patch the characteristics of coronal-plane can show significant reticular texture, be also added into be more suited to it is netted
FD (fractal dimension, fractal dimension) [3] feature of texture analysis.It may go out after surgery for light-type sticking patch
The clinical picture of existing contraction, the curling with spatial alternation feature etc. has also been further introduced into 3-D GLCM [4] and 3-D FD
[5] to extract the three-D grain feature in region to be sorted.For ABUS scan attribute and incisional hernia sticking patch implantation position characteristic, also
Propose the position feature (i.e. local feature and environmental characteristic are corresponding) in two regions to be sorted.To sum up, in total using 40 ginsengs
Number to carry out feature statement to each region to be sorted.Wherein, including 25 2-D textural characteristics, 13 3-D textural characteristics, 1
Based on the local feature of scan depths locating for region to be sorted, 1 environment based on region to be sorted and hernical sac positional relationship is special
Sign.
5. in step S4.1, the 2-D textural characteristics for extracting region to be sorted are carried out in two steps:
The first step calculates 12 2-D GLCM textural characteristics of region cross section to be sorted single frames slice.Firstly, extracting
The single frames cross section noise-reduced image at the center VOI;Secondly, doing upper and lower 5 pixel in scan depths direction to each region to be sorted
Region extension;Then, it is partitioned into the cross section single frames slice in each region to be sorted;Finally, calculating 12 to single frames slice
Descriptor, the 2-D GLCM textural characteristics as region cross section to be sorted single frames slice;
Second step calculates the 12 2-D GLCM textural characteristics and 1 2-D FD of coronal-plane Slice Sequence in region to be sorted
Feature.Firstly, extracting all coronal-plane Slice Sequences corresponding with scan depths locating for each region to be sorted;Secondly, right
In each descriptor fi(number of plies that i=1...n, n are coronal-plane Slice Sequence), uses descriptor fiSuccessively to every coronal-plane
Slice is calculated, and obtains one group of characteristic value [f1...fn];Finally, taking [f1...fn] mean value FmeanAs this descriptor pair
In the characteristic value of coronal-plane Slice Sequence in region to be sorted.
For 2-D GLCM, such as the definition initially by scholar Haralick in document [2] to GLCM, in two dimensional image
The differences in spatial location of two pixels can use motion vectorTo describe.Distance of the d between two pixels,
For the angle of two pixels and reference axis.Set a distance d is given for one, on 4 independent directionsThere may be 8 adjacent pixels to appearance altogether, as shown in Figure 4.So, in two dimensional imageDirection is separated by a pair of of pixel pair of d, is respectively provided with the probability of gray scale i and j appearance, i.e.,It is denoted as pij.It will be by pij
The gray level co-occurrence matrixes of image can be obtained in the matrix normalization of composition.
Haralick proposes calculated 14 textural characteristics from GLCM in document [2] altogether.The present invention has selected it
In 12 features, include energy (f1- Energy), contrast (f2- Contrast), correlation (f3- Correlation), variance
(f4- Variance), homogeney (f5- Homogeneity), mean value (f6- Sum Average), entropy (f7- Entropy), from phase
Close (f8- Autocorrelation), otherness (f9- Dissimilarity), cluster shade (f10- Cluster Shade), collection
Group's protrusion (f11- Cluster Prominence) and maximum probability (f12-Maximum Probability)。
Wherein, NgFor grey level, p (i, j) is the probability that (i, j) occurs in gray scale, i.e., gray level co-occurrence matrixes normalize
Result.The average and standard deviation of matrix row and column isWith
For 2-D FD, two-dimensional fractal dimension is estimated by the power spectrum of image Fourier transformation.Use following FFT
(fast fourier transform, Fast Fourier Transform (FFT)) to carry out DFT (discrete fourier to two dimensional image
Transform, discrete Fourier transform):
Wherein, I is the two dimensional image region having a size of (M, N), and u and v are the spatial frequency (u=in the direction x and y respectively
0,1 ... M-1, v=0,1 ... N-1).Power spectral density P is estimated as follows by F (u, v):
P (u, v)=| F (u, v) |2 (14)
To calculate 2-D FD, it is averaged to P along the radial section direction across FFT frequency domain.Frequency space is by equably
24 directions are divided into, and 30 points are equably sampled to the radial component in each direction.Calculate log (Pf) to log's (f)
Least square fitting, whereinRadial frequency [10] are represented, then FD is relevant to this double-log in the form of following
Slope of a curve β:
Wherein, DTIt is topological dimension, for two dimensional image, DT=2.
6. in step S4.2, the 3-D textural characteristics for extracting region to be sorted, to the volume number in each region to be sorted
According to calculating 12 3-D GLCM textural characteristics and 1 3-D FD feature.
For 3-D GLCM, 2-D GLCM is extended to 3-D GLCM by the method for document [4].Two in 3-D image
The differences in spatial location of tissue points can use motion vectorTo describe.Distance of the d between two tissue points,For two bodies
Azimuth between vegetarian refreshments, zenith angle of the θ between two tissue points.Set a distance d is given for one, on 13 independent directions, altogether
There may be 26 adjacent voxels to appearance, as shown in Figure 5.Here, identical 12 still in selective extraction and 2-D GLCM
3-D GLCM textural characteristics.
For 3-D FD, three-dimensional fractal dimension is estimated by the power spectrum of the 3-D Fourier transformation of volume images.It uses
Following 3-D FFT to carry out 3-D DFT to entire 3-D image:
Wherein, I is the 3-D image region having a size of (M, N, K), and u, v and w are in the space of x, y and z directionss frequency respectively
Rate.Power spectral density P estimates as follows:
P (u, v, w)=| F (u, v, w) |2 (17)
To calculate 3-D FD, it is averaged to P along the radial sector direction across 3-D FFT frequency domain.Frequency space is equal
Equal part is carried out in 24 azimuth directions and 12 zenith angular direction evenly, and the radial component in each direction is equably sampled
30 points.Calculate log (Pf) to the least square fitting of log (f), whereinRepresent radial frequency
[10], then 3-D FD is relevant to the slope β of this double logarithmic curve in the form of following:
Wherein, DTIt is topological dimension, for 3-D image, DT=3.
7. in step S4.3, the local feature for extracting scan depths locating for region to be sorted, based on stomach wall along ABUS
The anatomical features of Y-direction and the sticking patch implantation position feature of four kinds of strain less repair formulas are scanned, light-type sticking patch is in abdomen
Occurrence probability in 1-3cm depth areas below skin is maximum.It is therefore proposed that one is based on scan depths locating for region to be sorted
The local feature f in (direction Y- of ABUS)depthCharacterize the occurrence probability of sticking patch, and along ABUS scan depths direction to the part
Feature as shown in Figure 6 the case where divided, i.e., carry out interval division, respectively section from probe interface to skin following depth
3~6cm of 0~1cm, 1~3cm of section and section.According to the depth parameter in region to be sorted, to local feature fdepthSetting is such as
Under: 1) section 0~1cm, light-type sticking patch occurrence probability is medium on the upper side, i.e. fdepthBe set as 3., the reason is that this regional location compared with
Shallowly, only Onlay and Inlay art formula is possible to place sticking patch in this region.But the two is answered in clinic at present because recurrence rate is higher
With not too much more.2) 1~3cm of section, light-type sticking patch occurrence probability highest, i.e. fdepthIt is set as 4., the reason is that this region is just
It is the position for placing sticking patch in current clinic using most Sublay and IPOM art formulas well;3) in the section 3~6cm, with depth
A possibility that increase of degree, placement sticking patch, is gradually reduced, therefore from top to bottom, i.e. the f in this regiondepth2. and 1. it is set as.
8. in step S4.4, the environmental characteristic for extracting region and hernical sac positional relationship to be sorted is divided to following two stepping
Row.
The first step proposes a kind of quick hernical sac detection location algorithm based on ABUS data to determine hernical sac position.It is first
First, all black objects in ABUS coronal image are detected;Secondly, calculating maximum black after filtering out pseudo- black objects
The size of target i.e. hernical sac in the direction X-, Y- and Z-;Third will to ensure that entire hernical sac is all contained among VOI volume
Hernical sac size all expands 40 pixels in three directions;4th, the VOI appearance cut out containing hernical sac is concentrated from ABUS volume data
Product;5th, spot noise reduction process is carried out to VOI using 3-D ISRAD algorithm;Finally, dividing from every frame cross-sectional image of VOI
Hernical sac profile is cut out, detection and localization process of the algorithm to hernical sac in ABUS data are completed.
Second step is adjusted coronal-plane position feature after the completion of hernical sac detection.To with hernical sac coronal-plane projected area
The coronal-plane position feature in the region to be sorted of domain overlapping or intersection adds 2;To within the distance 3cm of hernical sac coronal-plane projected area
The coronal-plane position feature in region to be sorted adds 1.There is the probability of sticking patch in the to be sorted region closer apart from hernical sac position
It is bigger, it is therefore proposed that a location parameter f based on region to be sorted Yu hernical sac positional relationshipadjacencyTo characterize sticking patch
Occurrence probability.So using the coronal-plane position feature value in region to be sorted as environmental characteristic fadjacencyValue.Wait divide
The f in class regionadjacencyValue is bigger, and corresponding sticking patch occurrence probability is bigger.
9. in step S5,40 features of described pair of extraction carry out feature selecting, the steps include:
(1) between the class for calculating separately each feature using DBC (distance between class, class spacing) method [6]
Away from.Firstly, being free of 55 of sticking patch in all 18 cases (case group) of incision tract sticking patch containing light-type and incision tract
In case (control group), in conjunction with this experiment case load, 278 and the present invention for representing all different cases are artificially chosen by doctor
The identical typical sample region of the area size to be sorted automatically obtained.The typical case of light-type sticking patch is contained only including 125
Region and 153 contain only the representative region of fascia;Secondly, extracted 40 features are to all 278 sample areas before use
Do calculation of characteristic parameters and normalization in domain;Third calculates separately the class spacing of each feature;
(2) sorting from large to small by class spacing to 40 features chooses biggish preceding 25 features of class spacing, as spy
Levy the primary dcreening operation result of selection;
(3) SFS (sequential forward selection, before sequence is used to 25 features tentatively selected
Into search) method [7] selected, to obtain making the highest feature combination of classification accuracy rate.Firstly, true by doctor from 278
In fixed sample, training set 139 and test set 139 are randomly selected.Secondly, classifier selects SVM (support vector
Machine, support vector machines), classifier is trained with training set, with test set testing classification effect.In order to reduce experiment sample
Limited caused error, carries out 100 random experiments altogether, is evenly dividing training set and test set at random again every time.Root
According to final result, when choosing 4 2-D textural characteristics (energy, homogeney, entropy, the fractal dimension of two-dimensional coronal face image), 6
3-D textural characteristics (energy, correlation, homogeney, entropy, auto-correlation, the fractal dimension of three-dimensional volume to be sorted) and 1 area to be sorted
Three dimensionality location parameter (the f in domainadjacency) this 11 features when, the error in classification of light-type sticking patch and fascia can be made to reach
Minimum combines so finally choosing 11 features as feature.
As previously mentioned, being extracted 40 features in total to each region to be sorted.Since there are one between these features
Fixed redundancy, and the classification capacity of each feature is also not quite similar.Meanwhile in order to reduce intrinsic dimensionality, it is accurate to improve classification
Property, reduce classification the working time.The present invention is by selecting DBC and SFS to carry out feature selecting to above-mentioned 40 features.Finally, it selects
It takes 11 the error in classification of light-type sticking patch and fascia is made to reach the smallest feature to combine as feature.They are 4 2-D textures
Feature (energy, homogeney, entropy, the fractal dimension of two-dimensional coronal face image), 6 3-D textural characteristics be (three-dimensional volume to be sorted
Energy, correlation, homogeney, entropy, auto-correlation, fractal dimension) and 1 region to be sorted three dimensionality location parameter (fadjacency)。
Compared with prior art, the present invention is not only to the light-type sticking patch having a extensive future, using ABUS technology to it
Effectively reliable feature extraction is carried out, and realizes the choosing of the optimization to light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature
It selects, effectively improves the accuracy of Classification and Identification, reduces the classification working time.The method of the present invention feature selecting effect is good, high-efficient,
The nicety of grading of light-type incisional hernia sticking patch three-dimensional ultrasound pattern can be effectively improved.
Detailed description of the invention
Fig. 1 is light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method flow chart of the invention.
Fig. 2 is abdominal wall structure and four kinds of sticking patch implantation position schematic diagrames.Light-type sticking patch (black dotted lines) can be placed in
Before stomach wall fascia (Onlay), it is placed in the gap (Sublay) being placed between fascia and peritonaeum between flesh (Inlay), or be placed in
Intraperitoneal peritonaeum inner wall (IPOM).
Fig. 3 is the typical ABUS image of light-type sticking patch and fascia.Wherein, (a) and (b) is respectively light-type sticking patch and muscle
Image of the film in ABUS coronal-plane;(c) and (d) is image of the two in the cross section ABUS.
Fig. 4 is pixel to the spatial relation figure in two dimensional image.The central pixel point given for one is (white
Color), it is independent at 4On direction, there may be 8 pixels at a distance of normalized cumulant d altogether to (grey).
Fig. 5 is pixel to the spatial relation figure in 3-D image.The centerbody vegetarian refreshments given for one is (white
Color), it is independent at 13On direction, there may be 26 voxels at a distance of normalized cumulant d altogether to (grey).
Fig. 6 is that the ABUS imaging schematic diagram in stomach wall region and the local feature value based on scan depths set schematic diagram.
Fig. 7 is automatically extracting for ROI.Wherein, (a) foreground mask image Cmask;(b) by CmaskIt is added at lower boundary black
Color region, and it is divided into the image block of 50 × 50 pixels;(c) ROI extract as a result, coronal-plane position feature be 1 ROI with thin white
Color frame indicates that the ROI that coronal-plane position feature is 2 is indicated with thick white box.
Fig. 8 is VOI expansion process.Wherein, three orthogonal planes (A: cross section, the S: arrow of (a) ABUS original volume data
Shape face, C: coronal-plane);(b) the three orthogonal plane views of VOI;(c) 3-D view of VOI.
Fig. 9 .VOI image preprocessing process schematic.Wherein, (a) is located at the single frames cross section original image at the center VOI;
(b) 3-D ISRAD noise-reduced image;(c) bianry image;(d) small objects of image (c) delete result;(e) image (d) is most
Small circumscribed bounding box extraction results;(f) narrow goal deletion and alternative target number annotation results.
Figure 10 is the cross section single frames slice and coronal-plane Slice Sequence in 5 regions to be sorted in Fig. 9 (f).Wherein, (a)
5 single frames in cross section are sliced;(b) 5 groups of Slice Sequences of coronal-plane.
Figure 11 is that the quick hernical sac detection location algorithm based on ABUS data automatically extracts result to the VOI containing hernical sac.(a)
Three orthogonal planes (A: cross section, S: sagittal plane, C: coronal-plane) of ABUS original volume data;(b) three orthogonal planes of VOI
View;(c) 3-D view of VOI.
Figure 12 is the adjustment result of coronal-plane position feature in region to be sorted.(a) the three-dimensional reconstruction view of hernical sac testing result
Figure;(b) projection of hernical sac coronal-plane and all regions to be sorted;(c) it needs to be adjusted coronal-plane position feature to be sorted
Region.Coronal-plane position feature adds 2 region to be sorted to indicate with thick white box, and coronal-plane position feature adds 1 area to be sorted
Domain is indicated with thin white box.
Figure 13 is the DBC calculated result of all 40 features.Wherein feature f39It is scan depths feature fdepth, feature f40
It is region coronal-plane position feature f to be sortedadjacency。
Specific embodiment
The present invention is illustrated with reference to the accompanying drawings and detailed description, it is clear that described embodiment is only
It is a part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, ordinary skill people
Member's every other embodiment obtained without making creative work, shall fall within the protection scope of the present invention.
Referring to attached drawing 1, the light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method of the technical program include with
Under step:
S1: the 2-D foreground mask of one ABUS coronal-plane of building, shown in experimental result such as Fig. 7 (a);
(1) all ABUS coronal-planes for being located at 0.5 to 0.9 times of scanning total depth are sliced C1-CnIt takes out, to all C1-Cn
The pixel of same position does average value processing in image, obtains a coronal-plane mean value image Cmean;
(2) using Otsu algorithm to image CmeanThreshold process is carried out, bianry image C is obtainedbinary;
(3) using morphology opening operation to CbinaryConnected region in image carries out edge smoothing processing, and to its maximum
The black cavity as caused by lesion or shade is filled in white connected region, obtains coronal-plane 2-D foreground mask image
Cmask。
S2: VOI is extracted;
(1) from ABUS coronal-plane mask images CmaskThe upper left corner starts, and image is divided into the image block of 50 × 50 pixels.
Encountered at boundary on the right and lower boundary pixel deficiency can not piecemeal when, then add appropriate black region, experimental result such as Fig. 7
(b) shown in;
(2) the coronal-plane position feature of all black picture blocks (the sum of binaryzation value of all pixels point is 0) is set
It is 0, by all image blocks intersected with foreground mask boundary (the sum of binaryzation value of all pixels point is between 1 to 49)
Coronal-plane position feature is set to 1, by the coronal-plane of all full white image blocks (the sum of binaryzation value of all pixels point is 50)
Position feature is set to 2, shown in experimental result such as Fig. 7 (c);
(3) image block that all coronal-plane position features are not 0 is chosen for current ROI one by one.Using all with ROI phase
The ROI of 2-D, is extended to the VOI of 3-D by the cross section ABUS of pass and sagittal plane image-region.Current VOI is sent into one by one subsequent
Characteristic extracting module, until completing the traversal to all VOI, experimental results are shown in figure 8.
S3: pretreatment operation is carried out to VOI image, experimental result is as shown in Figure 9;
(1) three-dimensional filtering processing is carried out to the VOI automatically extracted out using 3-D ISRAD algorithm;
(2) binary conversion treatment is done to the single frames cross-sectional image for being located at the center VOI using Otsu algorithm, obtains alternative white
Connected region;
(3) the white connected region using all areas in opening operation deletion bianry image less than 15 pixels;
(4) the circumscribed bounding box of minimum for extracting each white connected region, calculates width, height and the top of bounding box
Point coordinate;
(5) the white connected region of width < 15 pixels of bounding box is deleted;
(6) mark is numbered to all remaining white connected regions, using as region to be sorted.
S4: after pretreatment, giving the region to be sorted of one group of VOI, can carry out one by one to each region to be sorted
Feature extraction extracts 40 features to each region to be sorted in total;
(1) the 2-D textural characteristics in region to be sorted are extracted;
(a) 12 2-D GLCM textural characteristics of region cross section to be sorted single frames slice are calculated.Firstly, extracting VOI
The single frames cross section noise-reduced image at center;Secondly, doing the region of upper and lower 5 pixel in scan depths direction to each region to be sorted
Extension;Then, it is partitioned into the cross section single frames slice in each region to be sorted;Finally, calculating 12 descriptions to single frames slice
Symbol, the 2-D GLCM textural characteristics as region cross section to be sorted single frames slice.Figure 10 (a) gives as shown in Fig. 9
5 region cross section to be sorted single frames slice;
(b) the 12 2-D GLCM textural characteristics and 1 2-D FD feature of coronal-plane Slice Sequence in region to be sorted are calculated.
Firstly, extracting all coronal-plane Slice Sequences corresponding with scan depths locating for each region to be sorted;Secondly, for each
Descriptor fi(number of plies that i=1...n, n are coronal-plane Slice Sequence), uses descriptor fiSuccessively to every coronal-plane be sliced into
Row calculates, and obtains one group of characteristic value [f1...fn];Finally, taking [f1...fn] mean value FmeanAs this descriptor for wait divide
The characteristic value of class region coronal-plane Slice Sequence.It is cut as Figure 10 (b) gives 5 region coronal-planes to be sorted shown in Fig. 9
Piece sequence;
12 2-D GLCM textural characteristics include energy (f1- Energy), contrast (f2- Contrast), correlation (f3-
Correlation), variance (f4- Variance), homogeney (f5- Homogeneity), mean value (f6- Sum Average), entropy
(f7- Entropy), auto-correlation (f8- Autocorrelation), otherness (f9- Dissimilarity), cluster shade (f10-
Cluster Shade), the prominent (f of cluster11- Cluster Prominence) and maximum probability (f12-Maximum
Probability), calculated respectively by formula 1-12;
For 2-D FD, two-dimensional fractal dimension is estimated by the power spectrum of image Fourier transformation.Use formula 13FFT
To carry out two dimensional image DFT, spectrum density P is estimated by F (u, v) by formula 14;
To calculate 2-D FD, it is averaged to P along the radial section direction across FFT frequency domain.Frequency space is by equably
24 directions are divided into, and 30 points are equably sampled to the radial component in each direction.Calculate log (Pf) to log's (f)
Least square fitting, whereinRadial frequency is represented, then FD is relevant to this double logarithmic curve by formula 15
Slope β;
(2) the 3-D textural characteristics for extracting region to be sorted, calculate 12 3-D GLCM textural characteristics and 1 3-D FD is special
Sign.2-D GLCM is extended to 3-D GLCM by the method for document [4].For 3-D FD, three-dimensional fractal dimension is by volume images
The power spectrum of 3-D Fourier transformation estimated.3-D is carried out to entire 3-D image using the 3-D FFT of formula 16
DFT, power spectral density P are estimated by formula 17;
To calculate 3-D FD, it is averaged to P along the radial sector direction across 3-D FFT frequency domain.Frequency space is equal
Equal part is carried out in 24 azimuth directions and 12 zenith angular direction evenly, and the radial component in each direction is equably sampled
30 points.Calculate log (Pf) to the least square fitting of log (f), whereinRadial frequency is represented,
Then 3-D FD is relevant to the slope β of this double logarithmic curve by formula 18;
(3) local feature for extracting scan depths locating for region to be sorted proposes that one is based on sweeping locating for region to be sorted
Retouch the local feature f in depth (direction Y- of ABUS)depthCharacterize the occurrence probability of sticking patch, and along ABUS scan depths direction pair
The local feature as shown in Figure 6 the case where divided.According to the depth parameter in region to be sorted, to local feature fdepthSetting
It is as follows: 1) 0~1cm of section, fdepth=3., and 2) 1~3cm of section, fdepth=4., and 3) in the section 3~6cm, by this region
fdepth2. and 1. feature is set as;
(4) environmental characteristic for extracting region and hernical sac positional relationship to be sorted proposes that one is based on region to be sorted and hernia
The location parameter f of capsule positional relationshipadjacencyTo characterize the occurrence probability of sticking patch.
To determine hernical sac position, a kind of quick hernical sac detection location algorithm based on ABUS data is proposed.Firstly, detection
All black objects in ABUS coronal image;Secondly, calculating maximum black target also just after filtering out pseudo- black objects
It is size of the hernical sac in the direction X-, Y- and Z-;Third, to ensure that entire hernical sac is all contained among VOI volume, by hernical sac size
All expand 40 pixels in three directions;4th, the VOI volume containing hernical sac is cut out from ABUS volume data concentration;5th, make
Spot noise reduction process is carried out to VOI with 3-D ISRAD algorithm;Finally, being partitioned into hernical sac wheel from every frame cross-sectional image of VOI
Exterior feature completes detection and localization process of the algorithm to hernical sac in ABUS data.Algorithm automatically extracts result as schemed to the VOI's containing hernical sac
Shown in 11.
After the completion of hernical sac detection, coronal-plane position feature is adjusted.To Chong Die with hernical sac coronal-plane view field
Or the coronal-plane position feature in the region to be sorted of intersection adds 2;To with it is to be sorted within the distance 3cm of hernical sac coronal-plane projected area
The coronal-plane position feature in region adds 1.Finally using the coronal-plane position feature value in region to be sorted as environmental characteristic
fadjacencyValue.The adjustment result of coronal-plane position feature in region to be sorted is as shown in figure 12.
S5: carrying out feature selecting to 40 features of extraction, final to choose 11 classification for making light-type sticking patch and fascia
Error reaches the smallest feature and combines as feature;
(1) the class spacing of each feature is calculated separately using DBC.Firstly, in the whole of incision tract sticking patch containing light-type
18 cases (case group) and incision tract are selected 278 by doctor without in 55 cases (control group) of sticking patch manually
Typical sample region identical with the area size to be sorted that this paper automatically obtains.Light-type is contained only including 125 to mend
The representative region of piece and 153 contain only the representative region of fascia.Secondly, extracted 40 features are to all 278 before use
A sample areas does calculation of characteristic parameters and normalization.Third, calculate separately the class spacing of each feature and to 40 features by
Class spacing is ranked up from big to small.Finally, choosing primary dcreening operation result of biggish preceding 25 features of class spacing as feature selecting.
Experimental result is as shown in figure 13;
(2) sorting from large to small by class spacing to 40 features chooses biggish preceding 25 features of class spacing;
(3) 25 features tentatively selected are selected using SFS.Firstly, the sample determined by doctor from 278
In this, training set 139 and test set 139 are randomly selected.Secondly, classifier selects SVM, classifier is trained with training set,
With test set testing classification effect.In order to reduce the limited caused error of experiment sample, 100 random experiments are carried out altogether, often
It is secondary to be evenly dividing training set and test set at random again.Finally, 11 errors in classification for making light-type sticking patch and fascia are chosen
Reach the smallest feature to combine as feature.They be 4 2-D textural characteristics (energy of two-dimensional coronal face image, homogeney,
Entropy, fractal dimension), 6 3-D textural characteristics (energy, correlation, homogeney, entropys, auto-correlation, FRACTAL DIMENSION of three-dimensional volume to be sorted
Number) and 1 region to be sorted three dimensionality location parameter (fadjacency)。
Present invention selection makes the error in classification of light-type sticking patch and fascia reach the smallest feature as feature combination, realizes
To the optimum choice of light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature, effectively improves the accuracy of Classification and Identification, reduces
Classify the working time.The present invention is sensitive for spatial alternations such as the curling postoperative to incisional hernia sticking patch of 2 d texture parameter, contractions
Problem proposes using three-D grain parameter and is aided with ABUS three dimensionality location parameter and extracts light-type incisional hernia sticking patch feature
Method, the nicety of grading of light-type incisional hernia sticking patch three-dimensional ultrasound pattern can be effectively improved.
Bibliography
[1]Wu J,Wang Y,Yu J,Shi X,Zhang J,Chen Y,Pang Y."Intelligent speckle
reducing anisotropic diffusion algorithm for automated 3-D ultrasound
images."Journal of the Optical Society of America A,2015,32(2):248-257.
[2]Haralick R.M.,Shanmugam K.,Dinstein I.H."Textural features for
image classification."IEEE Transactions on Systems,Man and Cybernetics 1973,
SMC-3(6):610-621.
[3]Chaudhuri B.B.,Sarkar N."Texture segmentation using fractal
dimension."IEEE Transactions on Pattern Analysis and Machine Intelligence,
1995,17(1):72-77.
[4]Chen W.,Giger M.L.,Li H.,Bick U.,Newstead G.M."Volumetric texture
analysis of breast lesions on contrast-enhanced magnetic resonance images."
Magnetic Resonance in Medicine,2007,58(3):562-571.
[5]Tang D.,Marangoni A.G."3D fractal dimension of fat crystal
networks."Chemical Physics Letters,2006,433(1-3):248-252.
[6]Li H.,Jiang T.,Zhang K."Efficient and robust feature extraction by
maximum margin criterion."IEEE Transactions on Neural Networks,2006,17(1):
157-165.
[7]Gunal S.,Gerek O.N.,Ece D.G.,Edizkan R."The search for optimal
feature set in power quality event classification."Expert Systems with
Applications,2009,36(7):10266-10273.
[8]Otsu N.A threshold selection method from gray-level histograms[J]
.IEEE Transactions on Systems,Man,and Cybernetics,1979,9(1):62-66.
[9]Wu J.,Wang Y.,Yu J.,Chen Y.,Pang Y.,Diao X.,Qiu Z."Identification
of implanted mesh after incisional hernia repair using an automated breast
volume scanner."Journal of Ultrasound in Medicine,2015,34(6):1071-1081.
[10]Kontos D.,Bakic P.R.,Carton A.K.,Troxel A.B.,Conant E.F.,Maidment
A.D."Parenchymal texture analysis in digital breast tomosynthesis for breast
cancer risk estimation:a preliminary study."Academic Radiology,2009,16(3):
283-298.。
Claims (9)
1. a kind of light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method, which comprises the following steps:
S1: the two-dimentional 2-D foreground mask of one ABUS coronal-plane of building;
S2: volume of interest VOI is extracted;
S3: pretreatment operation is carried out to VOI image;
S4:, can be to each 40 features of extracted region to be sorted to the region to be sorted of given one group of VOI;
S5: selecting 40 features of extraction, and final choose makes the error in classification of light-type sticking patch and fascia reach minimum
Feature as feature combine;
In step S1, the specific steps of the 2-D foreground mask of one ABUS coronal-plane of the building are as follows:
S1.1: all ABUS coronal-planes for being located at 0.5 to 0.9 times of scanning total depth are slicedC 1-C nIt takes out, to allC 1-C nFigure
The pixel of same position does average value processing as in, obtains a coronal-plane mean value imageC mean;
S1.2: using Otsu algorithm to imageC meanThreshold process is carried out, bianry image is obtainedC binary;
S1.3: morphology opening operation pair is usedC binaryImage is handled, and coronal-plane 2-D foreground mask image is obtainedC mask。
2. light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method according to claim 1, which is characterized in that
In step S2, the specific steps for extracting VOI are as follows:
S2.1: from ABUS coronal-plane mask imagesC maskThe upper left corner starts, and image is divided into the image block of 50 × 50 pixels;
S2.2: being set to 0 for the coronal-plane position feature of all black picture blocks, by all images intersected with foreground mask boundary
The coronal-plane position feature of block is set to 1, and the coronal-plane position feature of all full white image blocks is set to 2;
S2.3: the image block that all coronal-plane position features are not 0 is chosen for current interest region ROI one by one;Using institute
There are the cross section ABUS relevant to ROI and sagittal plane image-region, the ROI of 2-D is extended to the VOI of 3-D;It one by one will be current
VOI is sent into subsequent characteristics extraction module, until completing the traversal to all VOI.
3. light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method according to claim 2, which is characterized in that
In step S3, the specific steps that pretreatment operation is carried out to VOI image are as follows:
S3.1: three-dimensional filter is carried out to the VOI automatically extracted out using 3-D intelligence spot noise reduction anisotropy parameter ISRAD algorithm
Wave processing;
S3.2: binary conversion treatment is done to the single frames cross-sectional image for being located at the center VOI using Otsu algorithm;
S3.3: the white connected region using all areas in opening operation deletion bianry image less than 15 pixels;
S3.4: the circumscribed bounding box of minimum of each white connected region of extraction, width, height and the vertex for calculating bounding box are sat
Mark;
S3.5: the white connected region of width < 15 pixels of bounding box is deleted;
S3.6: being numbered mark to all remaining white connected regions, using as region to be sorted.
4. light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method according to claim 3, which is characterized in that
In step S4, the specific steps of feature extraction are carried out to each region to be sorted one by one are as follows:
S4.1: the 2-D textural characteristics in region to be sorted are extracted, are carried out in two steps;It is single to calculate region cross section to be sorted for the first step
12 2-D gray level co-occurrence matrixes GLCM textural characteristics of frame slice;Second step calculates coronal-plane Slice Sequence in region to be sorted
12 2-D GLCM textural characteristics and 1 2-D fractal dimension FD feature, are total up to 25;
S4.2: the 3-D textural characteristics in region to be sorted are extracted;To the volume data in each region to be sorted, 12 3-D are calculated
GLCM textural characteristics and 1 3-D FD feature, are total up to 13;
S4.3: extracting the local feature of scan depths locating for region to be sorted, deep based on scanning locating for region to be sorted with one
Degree is the local feature f in the direction Y- of ABUSdepthTo characterize the occurrence probability of sticking patch;
S4.4: extracting the environmental characteristic in region to be sorted Yu hernical sac positional relationship, with one based on region to be sorted and hernical sac position
Set the location parameter f of relationshipadjacencyTo characterize the occurrence probability of sticking patch;
12 GLCM textural characteristics, including 2-D GLCM and 3-D GLCM, specifically: energy f1, contrast f2, correlation f3, variance
f4, homogeney f5, mean value f6, entropy f7, auto-correlation f8, otherness f9, cluster shade f10, the prominent f of cluster11With maximum probability f12。
5. light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method according to claim 4, which is characterized in that
In step S5,40 features of described pair of extraction carry out the specific steps of feature selecting are as follows: firstly, using class spacing method DBC points
The class spacing of each feature is not calculated;Secondly, being sorted from large to small to 40 features by class spacing, before selection class spacing is biggish
25 features;Finally, being selected using sequential advancement search method SFS 25 features tentatively selected, to obtain making classifying
The highest feature combination of accuracy.
6. light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method according to claim 4, which is characterized in that
In step S4.1,12 2-D gray level co-occurrence matrixes GLCM textures for calculating region cross section to be sorted single frames slice are special
Sign, specific steps are as follows: firstly, extracting the single frames cross section noise-reduced image at the center VOI;Secondly, existing to each region to be sorted
Do the region extension of upper and lower 5 pixel in scan depths direction;Then, it is partitioned into the cross section single frames slice in each region to be sorted;
12 descriptors are calculated finally, being sliced to the single frames, the 2-D GLCM texture as region cross section to be sorted single frames slice
Feature;
12 2-D GLCM textural characteristics for calculating coronal-plane Slice Sequence in region to be sorted, specific steps are as follows: for 2-D
GLCM, differences in spatial location motion vector the D(φ, d of two pixels in two dimensional image) it describes, d is two pixels
Between distance, φ be two pixels and reference axis angle;Set a distance d is given for one, on 4 independent directions: φ=
0,45,90,135, there may be 8 adjacent pixels to appearance altogether, then, the direction φ in two dimensional image is separated by d's
A pair of of pixel pair is respectively provided with the probability of gray scale i and j appearance, i.e. p(i, j/ φ, d), it is denoted as pij;It will be by pijThe matrix of composition
Normalization obtains the gray level co-occurrence matrixes of image;Then, the calculating formula of 12 2-D GLCM textural characteristics is as follows:
Wherein, NgIt is the probability that (i, j) occurs in gray scale for grey level, p(i, j), i.e. the normalized knot of gray level co-occurrence matrixes
Fruit;The average and standard deviation of matrix row and column is,, and,;
1 2-D fractal dimension FD feature for calculating coronal-plane Slice Sequence in region to be sorted, specific steps are as follows: for 2-D
FD, two-dimensional fractal dimension are estimated by the power spectrum of image Fourier transformation;Come using following Fast Fourier Transform (FFT) FFT
Discrete Fourier transform DFT is carried out to two dimensional image:
(13)
Wherein, I is the two dimensional image region having a size of (M, N), and u and v are the spatial frequency in the direction x and y respectively, u=0,
1 ... M-1, v=0,1 ... N-1;Power spectral density P is v) estimated as follows by F(u:
(14)
To calculate 2-D FD, it is averaged to P along the radial section direction across FFT frequency domain, frequency space is by equably equal part
For 24 directions, and 30 points are equably sampled to the radial component in each direction;Calculate log(Pf) to log(f) and minimum
Two multiply fitting, whereinRadial frequency is represented, then FD is relevant to this double logarithmic curve in the form of following
Slope β:
(15)
Wherein, DTIt is topological dimension, for two dimensional image, DT = 2。
7. light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method according to claim 4, which is characterized in that
In step S4.2,12 3-D GLCM textural characteristics of the calculating, specific steps are as follows: for 3-D GLCM, in 3-D image
Two tissue points differences in spatial location motion vector D(φ, θ, d) describe, distance of the d between two tissue points, φ is
Azimuth between two tissue points, zenith angle of the θ between two tissue points;Set a distance d is given for one, in 13 independent directions
On, there may be 26 adjacent voxels to appearance altogether;Here, identical 12 3-D still in selective extraction and 2-D GLCM
GLCM textural characteristics;
1 3-D FD feature of the calculating, specific steps are as follows: for 3-D FD, three-dimensional fractal dimension by volume images Fu 3-D
In the power spectrum of leaf transformation estimated;3-D DFT is carried out to entire 3-D image using following 3-D FFT:
(16)
Wherein,IBe having a size of (M,N,K) 3-D image region,u, vWithwBe respectivelyx, yWithzThe spatial frequency in direction,
Power spectral densityPEstimate as follows:
(17)
To calculate 3-D FD, it is averaged to P along the radial sector direction across 3-D FFT frequency domain;Frequency space is by equably
Equal part is carried out in 24 azimuth directions and 12 zenith angular direction, and 30 are equably sampled to the radial component in each direction
Point, calculates log(Pf) to log(f) and least square fitting, whereinRadial frequency is represented,
Then 3-D FD is relevant to the slope β of this double logarithmic curve in the form of following:
(18)
Wherein, DTIt is topological dimension, for 3-D image, DT = 3。
8. light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method according to claim 4, which is characterized in that
In step S4.3, the local feature f in the direction scan depths, that is, ABUS Y- locating for region to be sorteddepthIt is provided that edge
ABUS scan depths direction carries out interval division to skin following depth by probe interface to local feature: being respectively section 0 ~ 1
3 ~ 6 cm of cm, 1 ~ 3 cm of section and section;According to the depth parameter in region to be sorted, to local feature fdepthSetting: 1) section 0
~ 1 cm, light-type sticking patch occurrence probability is medium on the upper side, i.e. fdepthIt is set as 3.;2) 1 ~ 3 cm of section, light-type sticking patch go out
Existing probability highest, i.e. fdepthIt is set as 4.;3) in 3 ~ 6 sections cm, with the increase of depth, a possibility that placing sticking patch by
It is decrescence small, therefore from top to bottom, i.e. the f in this regiondepth2. and 1. it is set as.
9. light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method according to claim 4, which is characterized in that
In step S4.4, the environmental characteristic for extracting region and hernical sac positional relationship to be sorted is divided to following two step to carry out:
The first step detects location algorithm using the quick hernical sac based on ABUS data, to determine hernical sac position: firstly, detection
All black objects in ABUS coronal image;Secondly, calculating maximum black target also just after filtering out pseudo- black objects
It is size of the hernical sac in the direction X-, Y- and Z-;Third, to ensure that entire hernical sac is all contained among VOI volume, by hernical sac size
All expand 40 pixels in three directions;4th, the VOI volume containing hernical sac is cut out from ABUS volume data concentration;5th,
Spot noise reduction process is carried out to VOI using 3-D ISRAD algorithm;Finally, being partitioned into hernical sac from every frame cross-sectional image of VOI
Profile completes detection and localization process to hernical sac in ABUS data;
Second step is adjusted coronal-plane position feature after the completion of hernical sac detection: to heavy with hernical sac coronal-plane view field
The coronal-plane position feature in folded or intersection region to be sorted adds 2;To within 3 cm of hernical sac coronal-plane projected area distance to
The coronal-plane position feature of specification area adds 1;There is the probability of sticking patch just in the to be sorted region closer apart from hernical sac position
It is bigger, therefore, with a location parameter f based on region to be sorted Yu hernical sac positional relationshipadjacencyTo characterize the appearance of sticking patch
Probability;So using the coronal-plane position feature value in region to be sorted as environmental characteristic fadjacencyValue;Area to be sorted
The f in domainadjacencyValue is bigger, and corresponding sticking patch occurrence probability is bigger.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710181583.6A CN106875409B (en) | 2017-03-24 | 2017-03-24 | A kind of light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710181583.6A CN106875409B (en) | 2017-03-24 | 2017-03-24 | A kind of light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106875409A CN106875409A (en) | 2017-06-20 |
CN106875409B true CN106875409B (en) | 2019-06-21 |
Family
ID=59172836
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710181583.6A Expired - Fee Related CN106875409B (en) | 2017-03-24 | 2017-03-24 | A kind of light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106875409B (en) |
Families Citing this family (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107610107B (en) * | 2017-09-01 | 2019-09-13 | 华中科技大学 | A kind of three-dimensional vascular plaque features of ultrasound pattern based on dimension describes method |
CN109300170B (en) * | 2018-10-18 | 2022-10-28 | 云南大学 | Method for transmitting shadow of portrait photo |
CN110246125A (en) * | 2019-05-31 | 2019-09-17 | 天津大学 | Teat placement automatic testing method based on ABUS coronal image |
CN110991339B (en) * | 2019-12-02 | 2023-04-28 | 太原科技大学 | Three-dimensional palate wrinkle identification method adopting cyclic frequency spectrum |
CN111275617B (en) * | 2020-01-09 | 2023-04-07 | 云南大学 | Automatic splicing method and system for ABUS breast ultrasound panorama and storage medium |
CN111210427B (en) * | 2020-01-17 | 2023-03-10 | 云南大学 | Time-change-based method for analyzing post-operation shrinkage of in-vivo light-weight patch |
CN111626986B (en) * | 2020-04-29 | 2022-06-21 | 云南大学 | Three-dimensional ultrasonic abdominal wall hernia patch detection method and system based on deep learning |
CN111528836B (en) * | 2020-05-06 | 2023-04-28 | 北京工业大学 | Brain function network feature extraction method based on dynamic directional transfer function |
CN116188488B (en) * | 2023-01-10 | 2024-01-16 | 广东省第二人民医院(广东省卫生应急医院) | Gray gradient-based B-ultrasonic image focus region segmentation method and device |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104657984A (en) * | 2015-01-28 | 2015-05-27 | 复旦大学 | Automatic extraction method of three-dimensional breast full-volume image regions of interest |
-
2017
- 2017-03-24 CN CN201710181583.6A patent/CN106875409B/en not_active Expired - Fee Related
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104657984A (en) * | 2015-01-28 | 2015-05-27 | 复旦大学 | Automatic extraction method of three-dimensional breast full-volume image regions of interest |
Non-Patent Citations (3)
Title |
---|
A Fast Detection and Diagnosis Algorithm for Abdominal Incisional Hernia Lesions With Automated 3D Ultrasound Images;Jun Wu et al.;《2013 6th International Conference on Biomedical Engineering and Informatics (BMEI 2013)》;20131231;第86-90页 |
Identification of Implanted Mesh After Incisional Hernia Repair Using an Automated Breast Volume Scanner;Jun Wu et al.;《J Ultrasound Med》;20151231;第34卷;第1071-1081页 |
基于自动化三维超声冠状面图像的腹壁手术切口检测算法;吴俊 等;《仪器仪表学报》;20161231;第37卷;第23-27页 |
Also Published As
Publication number | Publication date |
---|---|
CN106875409A (en) | 2017-06-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106875409B (en) | A kind of light-type incisional hernia sticking patch three-dimensional ultrasound pattern feature extracting method | |
Lo et al. | Multi-dimensional tumor detection in automated whole breast ultrasound using topographic watershed | |
Huang et al. | Watershed segmentation for breast tumor in 2-D sonography | |
Ma et al. | Two graph theory based methods for identifying the pectoral muscle in mammograms | |
Tzikopoulos et al. | A fully automated scheme for mammographic segmentation and classification based on breast density and asymmetry | |
CN108364294A (en) | Abdominal CT images multiple organ dividing method based on super-pixel | |
Abd-Elaziz et al. | Liver tumors segmentation from abdominal CT images using region growing and morphological processing | |
US20050027188A1 (en) | Method and apparatus for automatically detecting breast lesions and tumors in images | |
Flores et al. | Breast ultrasound despeckling using anisotropic diffusion guided by texture descriptors | |
CN101727537A (en) | Computer determining method of mammary gland CR image based on double visual angle information fusion | |
WO2006086467A1 (en) | System and method for using learned discriminative models to segment three dimensional colon image data | |
US10376230B2 (en) | Obtaining breast density measurements and classifications | |
Xian et al. | Multiple-domain knowledge based MRF model for tumor segmentation in breast ultrasound images | |
CN101859441B (en) | Image-based computer-aided analytical method for performing non-invasive monitoring to degree of tumor-infiltrated surrounding tissue | |
Jiang et al. | Segmentation of prostate ultrasound images: the state of the art and the future directions of segmentation algorithms | |
Pengiran Mohamad et al. | Transition of traditional method to deep learning based computer-aided system for breast cancer using Automated Breast Ultrasound System (ABUS) images: a review | |
KR101892949B1 (en) | Method and apparatus for hierarchical organ segmentation using shape and intensity information based on multi-atlas in abdominal ct images | |
Edwin et al. | Liver and tumour segmentation from abdominal CT images using adaptive threshold method | |
Armya et al. | Medical images segmentation based on unsupervised algorithms: a review | |
Pashoutan et al. | Automatic breast tumor classification using a level set method and feature extraction in mammography | |
Onal et al. | Image based measurements for evaluation of pelvic organ prolapse | |
Zhang | Novel approaches to image segmentation based on neutrosophic logic | |
Abdellatif et al. | K2. Automatic pectoral muscle boundary detection in mammograms using eigenvectors segmentation | |
Rodrigues et al. | A new methodology based on q-entropy for breast lesion classification in 3-D ultrasound images | |
da Cruz | Automatic analysis of mammography images |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190621 Termination date: 20200324 |