CN110060235A - A kind of thyroid nodule ultrasonic image division method based on deep learning - Google Patents
A kind of thyroid nodule ultrasonic image division method based on deep learning Download PDFInfo
- Publication number
- CN110060235A CN110060235A CN201910236399.6A CN201910236399A CN110060235A CN 110060235 A CN110060235 A CN 110060235A CN 201910236399 A CN201910236399 A CN 201910236399A CN 110060235 A CN110060235 A CN 110060235A
- Authority
- CN
- China
- Prior art keywords
- image
- thyroid
- tubercle
- ultrasound image
- interest
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 33
- 208000009453 Thyroid Nodule Diseases 0.000 title claims abstract description 29
- 208000024770 Thyroid neoplasm Diseases 0.000 title claims abstract description 29
- 238000013135 deep learning Methods 0.000 title claims abstract description 13
- 238000002604 ultrasonography Methods 0.000 claims abstract description 66
- 230000011218 segmentation Effects 0.000 claims abstract description 57
- 210000001685 thyroid gland Anatomy 0.000 claims abstract description 49
- 239000000284 extract Substances 0.000 claims abstract description 5
- 238000005070 sampling Methods 0.000 claims description 30
- 238000013527 convolutional neural network Methods 0.000 claims description 11
- 238000003475 lamination Methods 0.000 claims description 3
- 238000003709 image segmentation Methods 0.000 description 12
- 238000004422 calculation algorithm Methods 0.000 description 10
- 238000012549 training Methods 0.000 description 8
- 238000003745 diagnosis Methods 0.000 description 6
- 230000006870 function Effects 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000003066 decision tree Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 210000004907 gland Anatomy 0.000 description 3
- 238000013507 mapping Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000003759 clinical diagnosis Methods 0.000 description 2
- 238000007796 conventional method Methods 0.000 description 2
- 238000012937 correction Methods 0.000 description 2
- 230000007423 decrease Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000002068 genetic effect Effects 0.000 description 2
- 230000036039 immunity Effects 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 230000000877 morphologic effect Effects 0.000 description 2
- 230000000505 pernicious effect Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000000638 solvent extraction Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 206010004412 Benign neoplasm of thyroid gland Diseases 0.000 description 1
- 206010028980 Neoplasm Diseases 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 210000000683 abdominal cavity Anatomy 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 239000003086 colorant Substances 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 238000003708 edge detection Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 230000007773 growth pattern Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 239000003550 marker Substances 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000035772 mutation Effects 0.000 description 1
- 230000001575 pathological effect Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000007637 random forest analysis Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-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
- G06T7/12—Edge-based segmentation
-
- 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
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20092—Interactive image processing based on input by user
- G06T2207/20104—Interactive definition of region of interest [ROI]
-
- 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
- G06T2207/30096—Tumor; Lesion
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Medical Informatics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Quality & Reliability (AREA)
- Radiology & Medical Imaging (AREA)
- Ultra Sonic Daignosis Equipment (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of thyroid nodule ultrasonic image division method based on deep learning, it the described method comprises the following steps: inputting original Thyroid ultrasound image, area-of-interest prognostic chart is obtained by area-of-interest semantic segmentation model, and prognostic chart is corresponded in original Thyroid ultrasound image and extracts area-of-interest；The Thyroid ultrasound interesting image regions extracted are inputted in handmarking's identification model, and coarse positioning is carried out to the tubercle in area-of-interest；Coarse positioning image is inputted into Thyroid ultrasound image tubercle edge parted pattern, tubercle prognostic chart is obtained, draws tubercle edge.The present invention effectively overcomes the problems such as thyroid nodule ultrasound image resolution ratio is low, the more bring segmentations of interference information are difficult, accuracy rate is low, finally greatly improves the accuracy of nodule segmentation.
Description
Technical field
The present invention relates to Ultrasound Image Segmentation field more particularly to a kind of thyroid nodule ultrasound figures based on deep learning
As dividing method.
Background technique
Image, semantic segmentation be image is divided into different regions according to different semantemes, and by different colors come
Visualization marks out the classification of object representated by these regions.Medical image segmentation is that the hot spot in image, semantic segmentation field is answered
One of with.The purpose of medical image segmentation is that different parts is divided the image into according to the semantic features of image itself, to face
Bed diagnosis and pathological research provide reliable basis, and the precision of segmentation not only influences other next processing steps, also
The analysis of image may be influenced as a result, having an impact in turn to clinical diagnosis and research work.
The conventional method of image segmentation is broadly divided into based on region, based on edge and based on three kinds of partitioning algorithms of classification.
Algorithm based on region has watershed segmentation method, adaptive region growing method etc..Watershed segmentation method is to be based on
Geodesic curve watershed principle of compositionality and the image segmentation algorithm proposed, can obtain preferable effect, but deposit in segmentation task
To noise-sensitive, easy over-segmentation, profile is easy to be lost the problems such as.Adaptive region growing method is then the spy using different zones
The automatic growth pattern for learning corresponding different zones of sign, achieves preferable effect on abdominal cavity CT image and brain MRI image
Fruit.
Dividing method based on edge is to solve segmentation problem by detecting the edge comprising different zones, this method
Difficult point is contradiction when edge detection between noise immunity and detection accuracy.To improve noise immunity, can generate profile missing inspection and
Position deviation, to improve detection accuracy, the pseudo-edge that noise generates will lead to unreasonable profile.
Algorithm based on classifier has random number forest method, and random forest is made of multiple decision trees, by piece image
After multiple decision tree classifications, each decision tree has a classification results, is generated by ballot mode to the final of the image
Classification results.This method is the machine in normal service learning method of image, semantic segmentation, but is not applied to medical image segmentation.
In addition there are some other algorithms can be used for image segmentation.Such as based on the dividing method of statistics, this method is utilized
Statistical theory carries out modeling to digital picture to be split to image object, and effectively noise can be inhibited to segmentation result
Image, but be affected by initial parameter.It is sought based on the dividing method of genetic algorithm using the overall situation of genetic algorithm
Excellent ability and its characteristic insensitive to initial position, can improve the performance of image segmentation, and the difficult point of this method is to adapt to
Spend the selection of function and the determination of crossover probability and mutation probability.
Summary of the invention
The present invention provides a kind of thyroid nodule ultrasonic image division method based on deep learning, effective gram of the present invention
The problems such as thyroid nodule ultrasound image resolution ratio is low, the more bring segmentations of interference information are difficult, accuracy rate is low has been taken, it is final big
Width improves the accuracy of nodule segmentation, described below:
A kind of thyroid nodule ultrasonic image division method based on deep learning, the described method comprises the following steps:
Original Thyroid ultrasound image is inputted, area-of-interest prediction is obtained by area-of-interest semantic segmentation model
Figure, and prognostic chart is corresponded in original Thyroid ultrasound image and extracts area-of-interest；
The Thyroid ultrasound interesting image regions extracted are inputted in handmarking's identification model, to area-of-interest
In tubercle carry out coarse positioning；
Coarse positioning image is inputted into Thyroid ultrasound image tubercle edge parted pattern, tubercle prognostic chart is obtained, draws knot
Save edge.
Wherein, the area-of-interest semantic segmentation model specifically: down-sampling with up-sample completely corresponding full convolution
Neural network structure,
Web vector graphic 10 convolution sums 5 Chi Hualai guarantee that the receptive field of last pixel classifications result includes whole figure
Picture；
Down-sampling carried out to characteristic pattern using a pond per convolution twice, continuous 5 times of Web vector graphic and pond layer are one by one
Corresponding warp lamination up-samples characteristic pattern；
The result up-sampled every time is summed up with corresponding pond layer, is come as the input up-sampled next time gradually extensive
The resolution ratio and details of complex pattern.
Further, Thyroid ultrasound image tubercle edge parted pattern specifically: full convolutional neural networks are used,
It, and successively will up-sampling result and pond using 2 times twice up-samplings when being up-sampled to down-sampling characteristic pattern
Change layer pool4And pool3It sums up, using restoring image detail and resolution ratio and by result as the input up-sampled next time；
Last time up-sampling step-length be 8, be directly returned to original image resolution ratio, obtain segmentation tubercle and normal tissue it is pre-
Mapping, and prognostic chart is corresponded in thyroid nodule coarse positioning image, draw tubercle edge.
The beneficial effect of the technical scheme provided by the present invention is that:
1, the present invention has preferable segmentation to hand over Thyroid ultrasound image nodule segmentation and compares, and segmentation result can be doctor
Diagnosis tubercle it is good it is pernicious submit necessary information, can be improved the accuracy, objectivity and reliability of clinical diagnosis result, improve
The working efficiency of doctor；
2, low for ultrasound image resolution ratio, the more brings of interference information divide the problems such as difficult, accuracy rate is low, the present invention
The interference information in ultrasound image is gradually reduced by thtee-stage shiplock layer, finally increases substantially nodule segmentation accuracy.
Detailed description of the invention
Fig. 1 is a kind of flow chart of thyroid nodule ultrasonic image division method based on deep learning；
Fig. 2 is the friendship of three kinds of nodule segmentation modes and the curve graph than changing with training；
Fig. 3 is Thyroid ultrasound image tubercle edge proposed by the present invention parted pattern segmentation result and artificial segmentation result
Comparison diagram.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, embodiment of the present invention is made below further
Ground detailed description.
In recent years, with the rapid development of deep learning, Medical Image Processing is better solved using deep learning mechanism
The problem research and discovery direction important as scholar and researchers.Dividing method neural network based can be solved preferably
The certainly noise in image and non-uniform problem, convolutional neural networks are far super in the precision and efficiency in natural image segmentation field
Conventional method.
Embodiment 1
In order to solve the problems, such as that background technique, the embodiment of the present invention propose a kind of first shape based on deep learning
Gland tubercle ultrasonic image division method, referring to Fig. 1, method includes the following steps:
101: extracting area-of-interest (Region ofInterest, ROI) from original Thyroid ultrasound image；
Wherein, ultrasound image can be divided into ROI and background area, and wherein ROI is to include thyroid gland external morphology and structure
Echo information region, background area is other parts in addition to ROI, comprising: ultrasound scanner model, various sweep parameters
And the information that scan position etc. is unrelated with thyroid gland itself.
102: inputting the Thyroid ultrasound image ROI extracted, coarse positioning is carried out to the tubercle in ROI；
103: coarse positioning image being inputted into Thyroid ultrasound image tubercle edge parted pattern, obtains tubercle prognostic chart, is drawn
Tubercle edge out.
In one embodiment, step 101 extracts ROI from original Thyroid ultrasound image, the specific steps are as follows:
Original Thyroid ultrasound image is inputted, obtains ROI prognostic chart by ROI semantic segmentation model, and by prognostic chart pair
ROI should be extracted into original Thyroid ultrasound image.
Wherein, ROI semantic segmentation model uses down-sampling and up-sampling corresponding full convolutional neural networks structure completely, warp
It crosses training and prediction is split to Thyroid ultrasound image.In order to learn the global information to image, Web vector graphic as far as possible
10 convolution sums, 5 Chi Hualai ensure that the receptive field of last pixel classifications result includes whole image as far as possible.Wherein, every two
Secondary convolution carries out down-sampling to characteristic pattern using a pond.In order to enable neural network to acquire the edge ROI of more details
Information, with the one-to-one warp lamination of pond layer up-samples characteristic pattern for Web vector graphic continuous 5 times, every time on adopt
The result of sample is summed up with corresponding pond layer, gradually restore as the input up-sampled next time the resolution ratio of image with
Details.
In one embodiment, step 102 carries out coarse positioning to the tubercle in ROI on the basis of step 101, specific to walk
It is rapid as follows:
The Thyroid ultrasound image ROI that input step 101 extracts, using handmarking's identification model to tubercle in ROI
Carry out coarse positioning.The artificial marker recognition model is using existing convolutional neural networks, and convolutional layer inputs 19 × 19 × 3 figure
Picture, by connect after down-sampling twice with full articulamentum, each down-sampling is by two 3 × 3 convolution and a maximum value
Pondization composition, characteristic pattern depth are 16, and full articulamentum is connect with the output layer encoded by one-hot, and output vector is [1,0] table
Show that input picture is not handmarking, output vector is that [0,1] indicates that input picture belongs to handmarking.Convolutional layer with connect entirely
Layer uses ReLU as activation primitive.After the completion of handmarking's identification, corresponding square is drawn according to the position of label in the roi
Shape frame to carry out coarse segmentation to thyroid nodule.
Wherein, handmarking's identification model is known to those skilled in the art, and the embodiment of the present invention does not repeat them here this.
In one embodiment, step 103 divides tubercle on the basis of step 101 and step 102, draws the side of tubercle
Edge, the specific steps are as follows:
Divide mould for the thyroid nodule coarse positioning image of 224 × 224 sizes as Thyroid ultrasound image tubercle edge
The original image of type input carries out edge mark to original image under tumour hospital, Medical University Of Tianjin radiologist guidance
Image label image as input after note using binaryzation.Thyroid ultrasound image tubercle edge parted pattern is using complete
Convolutional neural networks model, but it is different from ROI parted pattern, and down-sampling layer has used for reference FCN (Fully Convolutional
Networks) algorithm has used the deeper VGG19 of depth (Visual Geometry Group 19) model.Based on VGG19 points
The down-sampling process of class model shares 5 down-samplings, preceding down-sampling twice using two 3 × 3 convolution kernel to input picture into
The characteristic pattern of generation is used the input behind maximum value pond as down-sampling next time, twice down-sampling point by row sliding window convolution
Not Wei 64 peacekeepings 128 dimension, after three times down-sampling used four 3 × 3 convolution kernels to input picture carry out sliding window convolution exist
Maximum value pond is carried out, dimension is respectively 256,512,512, and saving down-sampling result is respectively pool3、pool4、pool5.?
In VGG19 sorter network, characteristic pattern is next mapped as feature vector using three layers of full articulamentum will be special using softmax
DUAL PROBLEMS OF VECTOR MAPPING is levied to sample labeling space, obtains classification results.In full convolutional neural networks, use and pool5Characteristic pattern is big
Small identical, the increased convolution kernel of dimension is by pool5All features both map to each pixel of next layer of characteristic pattern in characteristic pattern
In vector, then characteristic pattern is mapped using two 1 × 1 convolution kernels, obtains the down-sampling comprising pixel classifications information
Characteristic pattern.When being up-sampled to down-sampling characteristic pattern, first using twice 2 times up-sampling and successively will up-sampling result and
pool4And pool3It sums up using restoring image detail and resolution ratio and by result as the input up-sampled next time, finally
Primary up-sampling step-length is 8, is directly returned to original image resolution ratio, obtains the prognostic chart of segmentation tubercle and normal tissue, and will be pre-
Mapping corresponds in thyroid nodule coarse positioning image, draws tubercle edge.
In conclusion the embodiment of the present invention effectively overcomes thyroid nodule ultrasound image resolution ratio through the above steps
Low, the more brings of interference information divide the problems such as difficult, accuracy rate is low, finally greatly improve the accuracy of nodule segmentation.
Embodiment 2
The scheme in embodiment 1 is further introduced below with reference to Fig. 1, specific calculation formula, it is as detailed below
Description:
201: during being split to thyroid nodule ultrasound image, first having to from original thyroid nodule ultrasound figure
ROI is extracted as in；
The embodiment of the present invention propose a kind of full convolutional neural networks model to carry out ROI to Thyroid ultrasound image it is semantic
Segmentation.In order to learn the global information to image as far as possible, full convolutional neural networks structure has used 10 convolution sums, 5 ponds
To ensure that the receptive field of last pixel classifications result includes whole image as far as possible.
202:ROI parted pattern selects cross entropy as loss function, as shown in formula (1).
Wherein, yiIt is the true tag of classification i, piIt is the probability value of the calculated classification i of softmax, k is classification number, N
It is total sample number.
203:ROI parted pattern selects Adam optimizer, and definition is as shown in formula (2).
Wherein, θt+1Indicate the t+1 moment i.e. parameter of t+1 iterative model, θtIndicate the t moment i.e. ginseng of t iterative model
Number, η is learning rate,For the correction value of single order momentum term,For the correction value of second order momentum term, ε is a value very little
Number (generally 1e-8) is 0 to avoid denominator.
204: Thyroid ultrasound image ROI being pre-processed with Image differentiation method；
205: the handmarking that Ultrasonography doctor is added in ultrasound image tubercle part using convolutional neural networks model
It is identified, because the essence identified to handmarking is two classification problems of image, each node only has negative class 0 and just
1 two kinds of possibilities of class simplify above-mentioned formula (1) so choosing target loss function of the binary cross entropy as training,
Simplify shown in result such as formula (3)；
Wherein, yi indicates the true tag of classification i,The probability for being 1 for current sample label.
206: after the completion of handmarking's identification, corresponding rectangle frame being drawn in the roi come to first shape according to the position of label
Gland tubercle carries out coarse segmentation；
207: thyroid nodule coarse positioning image being used to input as Thyroid ultrasound image tubercle edge parted pattern
Original image, to the image label image as input after original image progress edge labelling using binaryzation；
208: dividing tubercle in coarse segmentation image, choose the target loss function that cross entropy is training；It is excellent to choose Adam
Change optimizer of the device as Thyroid ultrasound image tubercle edge parted pattern；
209: training after the completion of, using Thyroid ultrasound image tubercle edge parted pattern to thyroid gland coarse positioning image into
The segmentation of row edge.
In conclusion 201- step 209 avoids ultrasound image background area to the embodiment of the present invention through the above steps
Noise jamming, and reduce the information interference of the analogous tissue of thyroid gland areas in ROI, not only increase algorithm characteristics extraction
Quality, and improve the convergence of algorithm model.
Embodiment 3
Feasibility verifying is carried out to the scheme in Examples 1 and 2 below with reference to specific calculation formula, example, is detailed in down
Text description:
In order to compare influence of the different segmentation levels to segmentation result, 1000 that tumour hospital, Medical University Of Tianjin is provided
Raw ultrasound image data set original image as input is opened, dept. of radiology's physician guidance under Thyroid ultrasound image original image is used
Under tubercle edge mark be used as label image, 80% be used as training set, 20% be used as test set, use Thyroid ultrasound image
Tubercle edge parted pattern is split.Meanwhile thyroid gland raw ultrasound image is subjected to ROI segmentation, and in radiologist
Label image behind guidance lower mark tubercle edge synchronizes segmentation, obtains the nodule segmentation data set under ROI image, and 80%
As training set, 20% is used as test set, is equally split using Thyroid ultrasound image tubercle edge parted pattern.
Compared with carrying out nodule segmentation directly on the original image of Thyroid ultrasound image, identical parted pattern is in first
The test set that thyroid nodule is split is handed over and is compared on the tubercle coarse positioning image of shape gland ultrasound image and improves 12%；With
It carries out nodule segmentation on Thyroid ultrasound ROI image to compare, test set, which is handed over and compared, improves 7.8%.And coarse positioning image point
21.6% friendship and ratio, significantly larger than other two kinds of partitioning schemes can be reached when training starts by cutting tubercle, and be restrained more
Rapidly.
The experimental results showed that successively carrying out nodule segmentation to Thyroid ultrasound image can effectively improve gradient decline convergence
Speed, and improve the precision of final nodule segmentation, as shown in Figure 2.
Comparison thyroid nodule coarse positioning ultrasound image is drawn under tumour hospital, Medical University Of Tianjin radiologist guidance
The segmentation result of the thyroid nodule edge ultrasound image and Thyroid ultrasound image tubercle edge parted pattern produced, such as Fig. 3
It is shown, it can be seen that Thyroid ultrasound image tubercle edge segmentation result in dept. of radiology, tumour hospital, Medical University Of Tianjin
The result manually drawn under physician guidance still has certain gap in fine degree and details, but can provide to diagnosis must
The information wanted.
In Fig. 2, original image divides the initial friendship of tubercle and than only 3.7%, and ROI image divides the initial friendship of tubercle
And than be 8.2%, coarse positioning image segmentation tubercle it is initial friendship and ratio reach 21.6%, facilitate gradient decline convergence
Matter.With trained progress, it is 20 times that original image segmentation tubercle, which is handed over and compares the number of iterations when starting convergence, finally hands over and compares
Stablize 75% or so.It is 15 times that ROI segmentation tubercle, which is handed over and compares the number of iterations when starting convergence, final to hand over and than stablizing
80% or so.It is 5 times that coarse positioning image segmentation tubercle, which is handed over and compares the number of iterations when starting convergence, final to hand over and than stablizing
87% or so.As shown in Figure 2, Thyroid ultrasound image tubercle edge parted pattern exists to the nodule segmentation of Thyroid ultrasound image
Significant effect is promoted in precision and convergence rate.
In Fig. 3, first is classified as thyroid nodule coarse positioning ultrasound image, and second is classified as in Medical University Of Tianjin's tumour doctor
The thyroid nodule edge ultrasound image drawn out under institute radiologist guidance, third are classified as Thyroid ultrasound image tubercle side
The segmentation result of edge parted pattern.
From the figure 3, it may be seen that the edge of manual drawing is quicker to the variation at tubercle edge under professional radiologist guidance
Sense, Fig. 3 the first row secondary series top-right part edge and the tertial model segmentation result of the first row have larger difference, model point
It is more smooth to cut result.It is similar, in the second row thyroid model segmentation result in edge details with the side manually drawn
Edge has a small amount of difference.But the segmentation result of Thyroid ultrasound image tubercle edge parted pattern can be with accurate description knot
The morphological features such as the shape and aspect ratio of section, these are characterized in the pernicious important evidence of diagnosis Benign Thyroid Nodules, because
This segmentation result can provide help for the diagnosis of doctor.
It is instructed by comparison thyroid nodule coarse positioning ultrasound image, in tumour hospital, Medical University Of Tianjin radiologist
Under the thyroid nodule edge ultrasound image drawn out and thyroid nodule Accurate Segmentation model segmentation result, as shown in figure 3,
As can be seen that model segmentation result can be with morphological features such as the shape of accurate description tubercle and aspect ratios, it can be for doctor's
Diagnosis extends efficient help.
It will be appreciated by those skilled in the art that attached drawing is the schematic diagram of a preferred embodiment, the embodiments of the present invention
Serial number is for illustration only, does not represent the advantages or disadvantages of the embodiments.
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all in spirit of the invention and
Within principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (3)
1. a kind of thyroid nodule ultrasonic image division method based on deep learning, which is characterized in that the method includes with
Lower step:
Original Thyroid ultrasound image is inputted, area-of-interest prognostic chart is obtained by area-of-interest semantic segmentation model, and
Prognostic chart is corresponded in original Thyroid ultrasound image and extracts area-of-interest；
The Thyroid ultrasound interesting image regions extracted are inputted in handmarking's identification model, in area-of-interest
Tubercle carries out coarse positioning；
Coarse positioning image is inputted into Thyroid ultrasound image tubercle edge parted pattern, tubercle prognostic chart is obtained, draws tubercle side
Edge.
2. a kind of thyroid nodule ultrasonic image division method based on deep learning according to claim 1, feature
Be, the area-of-interest semantic segmentation model specifically: down-sampling with up-sample completely corresponding full convolutional neural networks
Structure,
Web vector graphic 10 convolution sums 5 Chi Hualai guarantee that the receptive field of last pixel classifications result includes whole image；
Down-sampling carried out to characteristic pattern using a pond per convolution twice, continuous 5 times of Web vector graphic and pond layer one-to-one correspondence
Warp lamination characteristic pattern is up-sampled；
The result up-sampled every time is summed up with corresponding pond layer, gradually restores to scheme as the input up-sampled next time
The resolution ratio and details of picture.
3. a kind of thyroid nodule ultrasonic image division method based on deep learning according to claim 1, feature
It is, Thyroid ultrasound image tubercle edge parted pattern specifically: full convolutional neural networks are used,
It, and successively will up-sampling result and pond layer using 2 times twice up-samplings when being up-sampled to down-sampling characteristic pattern
pool4And pool3It sums up, using restoring image detail and resolution ratio and by result as the input up-sampled next time；
Last time up-sampling step-length is 8, is directly returned to original image resolution ratio, obtains the prediction of segmentation tubercle and normal tissue
Figure, and prognostic chart is corresponded in thyroid nodule coarse positioning image, draw tubercle edge.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910236399.6A CN110060235A (en) | 2019-03-27 | 2019-03-27 | A kind of thyroid nodule ultrasonic image division method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910236399.6A CN110060235A (en) | 2019-03-27 | 2019-03-27 | A kind of thyroid nodule ultrasonic image division method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110060235A true CN110060235A (en) | 2019-07-26 |
Family
ID=67317388
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910236399.6A Pending CN110060235A (en) | 2019-03-27 | 2019-03-27 | A kind of thyroid nodule ultrasonic image division method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110060235A (en) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110706793A (en) * | 2019-09-25 | 2020-01-17 | 天津大学 | Attention mechanism-based thyroid nodule semi-supervised segmentation method |
| CN111161273A (en) * | 2019-12-31 | 2020-05-15 | 电子科技大学 | Medical ultrasonic image segmentation method based on deep learning |
| CN111242956A (en) * | 2020-01-09 | 2020-06-05 | 西北工业大学 | U-Net-based ultrasonic fetal heart and fetal lung deep learning joint segmentation method |
| CN111292324A (en) * | 2020-03-20 | 2020-06-16 | 电子科技大学 | Multi-target identification method and system for brachial plexus ultrasonic image |
| CN111652837A (en) * | 2020-04-16 | 2020-09-11 | 上海长征医院 | AI-based thyroid nodule left and right lobe positioning and ultrasonic report error correction method |
| CN112036395A (en) * | 2020-09-04 | 2020-12-04 | 联想(北京)有限公司 | Text classification identification method and device based on target detection |
| CN112070089A (en) * | 2020-09-23 | 2020-12-11 | 西安交通大学医学院第二附属医院 | Ultrasonic image-based intelligent diagnosis method and system for diffuse thyroid diseases |
| CN112085743A (en) * | 2020-09-04 | 2020-12-15 | 厦门大学 | Image segmentation method for renal tumor |
| CN112308065A (en) * | 2020-07-09 | 2021-02-02 | 什维新智医疗科技(上海)有限公司 | Method, system and electronic equipment for identifying features in ultrasonic image |
| CN112529894A (en) * | 2020-12-22 | 2021-03-19 | 徐州医科大学 | Thyroid nodule diagnosis method based on deep learning network |
| CN112634240A (en) * | 2020-12-25 | 2021-04-09 | 南京邮电大学 | Thyroid ultrasound image interference interface automatic identification and removal method based on threshold segmentation |
| CN113393512A (en) * | 2021-05-19 | 2021-09-14 | 哈尔滨工业大学 | Thyroid volume automatic estimation method and system based on two-dimensional ultrasonic image |
| CN113436154A (en) * | 2021-06-11 | 2021-09-24 | 北京小白世纪网络科技有限公司 | Thyroid nodule edge symptom classification method, device and system |
| CN113450326A (en) * | 2021-06-28 | 2021-09-28 | 什维新智医疗科技(上海)有限公司 | Blood vessel information extraction element of thyroid gland tubercle |
| WO2021208677A1 (en) * | 2020-04-14 | 2021-10-21 | 华为技术有限公司 | Eye bag detection method and device |
| CN113592797A (en) * | 2021-07-21 | 2021-11-02 | 山东大学 | Mammary nodule risk grade prediction system based on multi-data fusion and deep learning |
| CN113842164A (en) * | 2021-09-14 | 2021-12-28 | 上海市第六人民医院 | System and method for detecting BRAF-V600E mutation |
| CN114782676A (en) * | 2022-04-02 | 2022-07-22 | 北京广播电视台 | Method and system for extracting region of interest of video |
| CN115375854A (en) * | 2022-10-25 | 2022-11-22 | 天津市肿瘤医院（天津医科大学肿瘤医院） | Ultrasonic imaging equipment image processing method fused with liquid crystal device and related device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018015414A1 (en) * | 2016-07-21 | 2018-01-25 | Siemens Healthcare Gmbh | Method and system for artificial intelligence based medical image segmentation |
| CN107680678A (en) * | 2017-10-18 | 2018-02-09 | 北京航空航天大学 | Based on multiple dimensioned convolutional neural networks Thyroid ultrasound image tubercle auto-check system |
| US20180165808A1 (en) * | 2016-06-27 | 2018-06-14 | University Of Central Florida Research Foundation, Inc. | System and method for image-based quantification of white and brown adipose tissue at the whole-body, organ and body-region levels |
| CN108734694A (en) * | 2018-04-09 | 2018-11-02 | 华南农业大学 | Thyroid tumors ultrasonoscopy automatic identifying method based on faster r-cnn |
| CN109087327A (en) * | 2018-07-13 | 2018-12-25 | 天津大学 | A kind of thyroid nodule ultrasonic image division method cascading full convolutional neural networks |
-
2019
- 2019-03-27 CN CN201910236399.6A patent/CN110060235A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180165808A1 (en) * | 2016-06-27 | 2018-06-14 | University Of Central Florida Research Foundation, Inc. | System and method for image-based quantification of white and brown adipose tissue at the whole-body, organ and body-region levels |
| WO2018015414A1 (en) * | 2016-07-21 | 2018-01-25 | Siemens Healthcare Gmbh | Method and system for artificial intelligence based medical image segmentation |
| CN107680678A (en) * | 2017-10-18 | 2018-02-09 | 北京航空航天大学 | Based on multiple dimensioned convolutional neural networks Thyroid ultrasound image tubercle auto-check system |
| CN108734694A (en) * | 2018-04-09 | 2018-11-02 | 华南农业大学 | Thyroid tumors ultrasonoscopy automatic identifying method based on faster r-cnn |
| CN109087327A (en) * | 2018-07-13 | 2018-12-25 | 天津大学 | A kind of thyroid nodule ultrasonic image division method cascading full convolutional neural networks |
Non-Patent Citations (3)
| Title |
|---|
| RUIGUO YU,ET AL: "Localization of Thyroid Nodules in Ultrasonic Images" * |
| XIANG YING, ET AL: "Thyroid Nodule Segmentation in Ultrasound Images Based on Cascaded Convolutional Neural Network" * |
| XIANG YING, ET AL: "Thyroid Nodule Segmentation in Ultrasound Images Based on Cascaded Convolutional Neural Network", 《INTERNATIONAL CONFERENCE ON NEURAL INFORMATION PROCESSING ICONIP 2018: NEURAL INFORMATION PROCESSING》 * |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110706793A (en) * | 2019-09-25 | 2020-01-17 | 天津大学 | Attention mechanism-based thyroid nodule semi-supervised segmentation method |
| CN111161273A (en) * | 2019-12-31 | 2020-05-15 | 电子科技大学 | Medical ultrasonic image segmentation method based on deep learning |
| CN111161273B (en) * | 2019-12-31 | 2023-03-21 | 电子科技大学 | Medical ultrasonic image segmentation method based on deep learning |
| CN111242956A (en) * | 2020-01-09 | 2020-06-05 | 西北工业大学 | U-Net-based ultrasonic fetal heart and fetal lung deep learning joint segmentation method |
| CN111292324A (en) * | 2020-03-20 | 2020-06-16 | 电子科技大学 | Multi-target identification method and system for brachial plexus ultrasonic image |
| CN111292324B (en) * | 2020-03-20 | 2022-03-01 | 电子科技大学 | Multi-target identification method and system for brachial plexus ultrasonic image |
| WO2021208677A1 (en) * | 2020-04-14 | 2021-10-21 | 华为技术有限公司 | Eye bag detection method and device |
| CN113536834A (en) * | 2020-04-14 | 2021-10-22 | 华为技术有限公司 | Pouch detection method and device |
| CN111652837A (en) * | 2020-04-16 | 2020-09-11 | 上海长征医院 | AI-based thyroid nodule left and right lobe positioning and ultrasonic report error correction method |
| CN111652837B (en) * | 2020-04-16 | 2022-09-09 | 上海长征医院 | AI-based thyroid nodule left and right lobe positioning and ultrasonic report error correction method |
| CN112308065B (en) * | 2020-07-09 | 2024-05-24 | 什维新智医疗科技(上海)有限公司 | Method, system and electronic equipment for identifying characteristics in ultrasonic image |
| CN112308065A (en) * | 2020-07-09 | 2021-02-02 | 什维新智医疗科技(上海)有限公司 | Method, system and electronic equipment for identifying features in ultrasonic image |
| CN112085743A (en) * | 2020-09-04 | 2020-12-15 | 厦门大学 | Image segmentation method for renal tumor |
| CN112036395B (en) * | 2020-09-04 | 2024-05-28 | 联想(北京)有限公司 | Text classification recognition method and device based on target detection |
| CN112036395A (en) * | 2020-09-04 | 2020-12-04 | 联想(北京)有限公司 | Text classification identification method and device based on target detection |
| CN112070089A (en) * | 2020-09-23 | 2020-12-11 | 西安交通大学医学院第二附属医院 | Ultrasonic image-based intelligent diagnosis method and system for diffuse thyroid diseases |
| CN112529894A (en) * | 2020-12-22 | 2021-03-19 | 徐州医科大学 | Thyroid nodule diagnosis method based on deep learning network |
| CN112634240B (en) * | 2020-12-25 | 2022-09-23 | 南京邮电大学 | Thyroid ultrasound image interference interface automatic identification and removal method based on threshold segmentation |
| CN112634240A (en) * | 2020-12-25 | 2021-04-09 | 南京邮电大学 | Thyroid ultrasound image interference interface automatic identification and removal method based on threshold segmentation |
| CN113393512A (en) * | 2021-05-19 | 2021-09-14 | 哈尔滨工业大学 | Thyroid volume automatic estimation method and system based on two-dimensional ultrasonic image |
| CN113436154A (en) * | 2021-06-11 | 2021-09-24 | 北京小白世纪网络科技有限公司 | Thyroid nodule edge symptom classification method, device and system |
| CN113450326A (en) * | 2021-06-28 | 2021-09-28 | 什维新智医疗科技(上海)有限公司 | Blood vessel information extraction element of thyroid gland tubercle |
| CN113592797A (en) * | 2021-07-21 | 2021-11-02 | 山东大学 | Mammary nodule risk grade prediction system based on multi-data fusion and deep learning |
| CN113842164A (en) * | 2021-09-14 | 2021-12-28 | 上海市第六人民医院 | System and method for detecting BRAF-V600E mutation |
| CN113842164B (en) * | 2021-09-14 | 2024-02-20 | 上海市第六人民医院 | Systems and methods for detecting BRAF-V600E mutations |
| CN114782676A (en) * | 2022-04-02 | 2022-07-22 | 北京广播电视台 | Method and system for extracting region of interest of video |
| CN114782676B (en) * | 2022-04-02 | 2023-01-06 | 北京广播电视台 | Method and system for extracting region of interest of video |
| CN115375854A (en) * | 2022-10-25 | 2022-11-22 | 天津市肿瘤医院（天津医科大学肿瘤医院） | Ultrasonic imaging equipment image processing method fused with liquid crystal device and related device |
| CN115375854B (en) * | 2022-10-25 | 2022-12-20 | 天津市肿瘤医院（天津医科大学肿瘤医院） | Ultrasonic image equipment image processing method fused with liquid crystal device and related device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110060235A (en) | A kind of thyroid nodule ultrasonic image division method based on deep learning | |
| CN111192245B (en) | Brain tumor segmentation network and method based on U-Net network | |
| EP3432263B1 (en) | Semantic segmentation for cancer detection in digital breast tomosynthesis | |
| Wang et al. | CT male pelvic organ segmentation using fully convolutional networks with boundary sensitive representation | |
| CN105957066B (en) | CT image liver segmentation method and system based on automatic context model | |
| CN102324109B (en) | Method for three-dimensionally segmenting insubstantial pulmonary nodule based on fuzzy membership model | |
| Badura et al. | Soft computing approach to 3D lung nodule segmentation in CT | |
| CN104933709B (en) | Random walk CT lung tissue image automatic segmentation methods based on prior information | |
| CN109215033A (en) | The method and system of image segmentation | |
| CN109493317A (en) | The more vertebra dividing methods of 3D based on concatenated convolutional neural network | |
| CN102385751B (en) | Liver tumor region segmentation method based on watershed transform and classification through support vector machine | |
| CN103473767B (en) | The method and system that a kind of soft tissues of abdomen nuclear-magnetism image is cut apart | |
| CN103824295A (en) | Segmentation method of adhesion vascular pulmonary nodules in lung CT (computed tomography) image | |
| CN107977952A (en) | Medical image cutting method and device | |
| CN109447998A (en) | Based on the automatic division method under PCANet deep learning model | |
| CN104899851A (en) | Lung nodule image segmentation method | |
| CN106204600A (en) | Cerebral tumor image partition method based on multisequencing MR image related information | |
| Tan et al. | Analysis of segmentation of lung parenchyma based on deep learning methods | |
| CN112734755A (en) | Lung lobe segmentation method based on 3D full convolution neural network and multitask learning | |
| CN112767411B (en) | Synchronous segmentation method for multiple organs of lung | |
| Liu et al. | Automatic segmentation algorithm of ultrasound heart image based on convolutional neural network and image saliency | |
| CN112862783A (en) | Thyroid CT image nodule automatic diagnosis system based on neural network | |
| CN111784713A (en) | Attention mechanism-introduced U-shaped heart segmentation method | |
| Kitrungrotsakul et al. | Interactive deep refinement network for medical image segmentation | |
| CN114581474A (en) | Automatic clinical target area delineation method based on cervical cancer CT image |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190726 |
|
| RJ01 | Rejection of invention patent application after publication |