CN110490892A - A kind of Thyroid ultrasound image tubercle automatic positioning recognition methods based on USFaster R-CNN - Google Patents
A kind of Thyroid ultrasound image tubercle automatic positioning recognition methods based on USFaster R-CNN Download PDFInfo
- Publication number
- CN110490892A CN110490892A CN201910594057.1A CN201910594057A CN110490892A CN 110490892 A CN110490892 A CN 110490892A CN 201910594057 A CN201910594057 A CN 201910594057A CN 110490892 A CN110490892 A CN 110490892A
- Authority
- CN
- China
- Prior art keywords
- network
- ultrasound image
- usfaster
- thyroid
- image
- 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
- 238000002604 ultrasonography Methods 0.000 title claims abstract description 61
- 210000001685 thyroid gland Anatomy 0.000 title claims abstract description 46
- 238000000034 method Methods 0.000 title claims abstract description 27
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 28
- 238000000605 extraction Methods 0.000 claims abstract description 15
- 208000009453 Thyroid Nodule Diseases 0.000 claims abstract description 14
- 208000024770 Thyroid neoplasm Diseases 0.000 claims abstract description 14
- 230000000306 recurrent effect Effects 0.000 claims abstract description 9
- 238000003745 diagnosis Methods 0.000 claims abstract description 6
- 230000000694 effects Effects 0.000 claims description 16
- 238000012360 testing method Methods 0.000 claims description 8
- 238000013480 data collection Methods 0.000 claims description 4
- 238000013461 design Methods 0.000 claims description 4
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 230000011218 segmentation Effects 0.000 claims description 3
- 238000002372 labelling Methods 0.000 claims description 2
- 210000004218 nerve net Anatomy 0.000 claims 1
- 238000013135 deep learning Methods 0.000 abstract description 11
- 238000013473 artificial intelligence Methods 0.000 abstract description 3
- 238000012937 correction Methods 0.000 abstract description 2
- 238000012216 screening Methods 0.000 abstract description 2
- 238000003062 neural network model Methods 0.000 abstract 2
- 238000002059 diagnostic imaging Methods 0.000 abstract 1
- 239000000284 extract Substances 0.000 abstract 1
- 230000006870 function Effects 0.000 abstract 1
- 238000005457 optimization Methods 0.000 abstract 1
- 238000001514 detection method Methods 0.000 description 8
- 230000007812 deficiency Effects 0.000 description 4
- 238000007689 inspection Methods 0.000 description 4
- 208000003464 asthenopia Diseases 0.000 description 2
- 230000003211 malignant effect Effects 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 241000208340 Araliaceae Species 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
- 208000035977 Rare disease Diseases 0.000 description 1
- RTAQQCXQSZGOHL-UHFFFAOYSA-N Titanium Chemical compound [Ti] RTAQQCXQSZGOHL-UHFFFAOYSA-N 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000004195 computer-aided diagnosis Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 239000003814 drug Substances 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 210000004907 gland Anatomy 0.000 description 1
- 230000003902 lesion Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 208000038009 orphan disease Diseases 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 241000894007 species Species 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30096—Tumor; Lesion
Abstract
The invention discloses a kind of, and the Thyroid ultrasound image tubercle based on USFaster R-CNN is automatically positioned recognition methods, belongs to artificial intelligence and deep learning field.This method includes the pretreatment of Thyroid ultrasound image, deep neural network model is built, and network model training and optimization, wherein deep neural network model includes bottom convolution feature extraction network, candidate frame generates network, characteristic pattern pond layer, classification and candidate frame Recurrent networks.Using deep learning method, realize that Thyroid ultrasound characteristics of image automatically extracts, candidate frame automatically generates, screening, position correction.Realize the automatic positioning identification function of thyroid nodule.The present invention can effectively assist doctor to carry out Thyroid ultrasound diagnostic imaging, improve the objectivity and accuracy rate of diagnosis, the workload of doctor and the omission factor of Small object tubercle is effectively reduced.
Description
Technical field
The invention belongs to artificial intelligence and deep learning field, in particular to a kind of (to be directed to based on USFaster R-CNN
The more rapidly region convolutional neural networks of ultrasound image) Thyroid ultrasound image tubercle automatic positioning and recognition methods.
Background technique
In recent years, as the development of image technology, the recall rate of thyroid nodule significantly increase, thyroid nodule is very general
Time, there are about the adults of one third, and thyroid nodule is found in inspection.Although wherein most be it is benign, also there are about 4-
There are canceration risks for 5% thyroid nodule.Ultrasonic (US) is widely used in diagnosis of thyroid illness, but diagnostician
With very big subjectivity when carrying out assessment tumour by ultrasound image, this is heavily dependent on the clinical warp of different doctors
It tests.It the use of the thyroid gland computer-aided diagnosis system (CAD) of artificial intelligence is a kind of selection for solving the problems, such as this, CAD can
To provide objective reliable diagnostic comments for doctor, reference is provided for clinic.This not only can effectively solve the master of different doctors
Disagreement is seen, the accuracy rate of diagnosis is improved, the workload and omission factor of doctors can also be reduced.
Target detection is the research hotspot of computer vision and machine learning field, it can outline the mesh in image
Cursor position simultaneously identifies targeted species, is had in terms of improving detection speed and accuracy rate based on the object detection method of deep learning
Very big potentiality, the application in terms of medical field especially medical image, can save a large amount of manpower and material resources.But at this
Preceding people must allow neural network " to acquire " feature of image, so that it is worked, this generally requires a sufficiently large data
Collection carrys out supervised training network parameter.For medical image, there are significant limitation, the marks of medical image in terms of training data
Note needs a large amount of professional knowledge and energy, the quantity of the medical image of some orphan diseases that trained standard is much not achieved.
Meanwhile ultrasound image is less compared to daily characteristics of image (lacking color characteristic), contrast is low, the feature of good Malignant Nodules
Difference is unobvious, belongs to fine grit classification problem.The small volume of some tubercles, diagnostician when working long hours,
Due to visual fatigue or experience deficiency etc., it is more likely that miss lesser tubercle.These are all Thyroid ultrasound image tubercles
Positioning with identification there are the problem of.
Algorithm of target detection based on deep learning is roughly divided into two classes: first is that the two based on Region Proposal
Stage algorithm of target detection, second is that the one stage algorithm of target detection based on regression problem.The first kind needs to generate candidate
Frame classifies to target by convolutional neural networks, finally carries out the amendment of bounding box position.Second class does not generate candidate then
Frame directly converts regression problem processing for the problem of target frame positioning, directly returns to the target of prediction.The two
Can on have differences, the former is dominant on Detection accuracy and positioning accuracy, and the latter is more dominant in detection speed.
Summary of the invention
Present invention aim to address the deficiencies of existing Thyroid ultrasound image detecting method, provide one kind and are based on
The Thyroid ultrasound image tubercle of USFasterR-CNN (for the more rapidly region convolutional neural networks of ultrasound image) is automatically fixed
Method for distinguishing is known in position, and realization outlines the tubercle in Thyroid ultrasound image and judge the method for tubercle type automatically, especially
The recall rate of Small object thyroid nodule is obviously improved.
In order to achieve the above object, a kind of Thyroid ultrasound image tubercle based on USFaster R-CNN provided by the invention is certainly
Dynamic positioning identifying method, comprising the following steps:
(1) Thyroid ultrasound image is pre-processed first, intercepts out pending area, remove in ultrasound image with first
Shape gland image irrelevant information is outlined super with image labeling tool labelImg according to the label of diagnostician and diagnosis report
Thyroid nodule in acoustic image, and the mark of tubercle type is carried out, it is fabricated to ground of the xml document as training set
True label, then removes the artificial mark of the diagnostician in image, in order to avoid influence finally surpass the training effect of network
Acoustic image and corresponding xml document are divided into training set, verifying collection and test set；
(2) the USFaster R-CNN depth mind put up with the training in advance of large data collection ImageNet data set
Through network, trained network parameter is obtained, then will be trained in advance when with Thyroid ultrasound training set of images training network
Good network parameter moves in USFaster R-CNN network, so that network convergence obtains faster and has stronger general
Change ability；
(3) trained hyper parameter is finely adjusted and is optimized using verifying collection after tentatively training network model, make mould
Type reaches best and identifies and positions effect；
(4) by the Thyroid ultrasound image to be identified with positioning, i.e. test set, USFaster R-CNN network is read,
Realize the automatic positioning and identification of thyroid nodule.
The handmarking of diagnostician in removal image described in step (1) of the present invention, using Threshold segmentation by image
Binaryzation calculates all four connected regions of bianry image and sorts, and it is to need that wherein pixel, which is greater than the connected region of threshold value,
Then connected region is carried out interpolation using bilinear interpolation algorithm by the handmarking region to be looked for, reach removal handmarking
Purpose.The threshold value is 10~20, takes 20 effects best.
USFaster R-CNN deep neural network described in step (2) of the present invention includes bottom convolution feature extraction net
Network, candidate frame generate network, characteristic pattern pond layer, classification and candidate frame Recurrent networks.
The structure of bottom convolution feature extraction network of the present invention are as follows: use 4 convolutional layers, after each convolutional layer with
With 1 layer of maximum pond layer, first three convolutional layer respectively has level 2 volume to accumulate, and the 4th convolutional layer haves three layers convolution, while one spy of design
Sign splicing layer, first stacks up two layers of convolution of third layer, then the output and the 4th of dimension is risen by the convolutional layer of a 1*1
Total output of the output as feature extraction network after the result splicing of layer convolution, obtains the spy with Thyroid ultrasound image information
Sign figure.Wherein the effect of bottom convolution feature extraction network is that the mode of the image of input convolution is automatically extracted feature, is mentioned
It supplies candidate frame and generates network.The characteristics of this method combination ultrasound image: feature will less and medicine compared with natural image
The limitation of image data set quantitatively.The feature extraction network for possessing four layers of convolutional layer is devised, and at last
The non-thread sexuality of network is further increased in layer using 1*1 convolution.Feature is added between third layer and the 4th layer simultaneously to spell
Layer is connect, is specifically superimposed two convolution in third layer convolution, then rises dimension using the convolutional layer of 1*1, it is last and the 4th
Output of the results added of layer convolution as characteristic layer, can further increase third layer convolution similar to the structure of residual error in this way
The influence of feature in the entire network also can further utilize more shallow-layer features, in this way than being used alone the 4th
The characteristic effect of layer convolution is more preferable.
It is (W/8*H/8) * 9 that candidate frame of the present invention, which generates the candidate frame quantity that network generates on the ultrasound image,
Middle W, H are respectively the width and height of ultrasound image, and the side length of the size of candidate frame is 64,128 and 256, and three kinds of ratios are respectively 1:
1,1:2 and 2:1.Thyroid ultrasound image passes through bottom convolution feature extraction network, in pond under the action of layer, 8 times of sampling
To characteristic pattern, i.e., if the size of original image is W*H, the size of characteristic pattern is W/8*H/8.One namely on characteristic pattern
A pixel corresponds to 8*8 pixel of original image.Then characteristic pattern is used as the input of candidate frame generation network, candidate frame again
It generates network and uses a 3*3 size on characteristic pattern, step-length is 1 convolution sliding window, and generates on each pixel
Anchor point, the corresponding candidate frame that 9 different proportions and size are generated in original image of each anchor point.According to thyroid nodule size
Feature, three kinds of design are 1:2,1:1 and 2:1 having a size of 64,128 and 256, three kinds of ratios, are more advantageous to outline ultrasound in this way
In image doctors be easy missing inspection lesser tubercle, the network can also calculate in these frames there are the probability of target simultaneously just successive step
The size and location of frame keeps it more accurate, filters out the candidate frame that maximum probability includes target.
The convolution of a 1*1*128 is added between characteristic pattern pond of the present invention layer, classification and candidate frame Recurrent networks
Layer can increase the non-linear of network while reducing parameter, improve fitting effect.Ultrasound image containing a small amount of candidate frame
It will enter into characteristic pattern pond layer with characteristic pattern and be unified to onesize, finally enter classification and frame Recurrent networks are carried out into one
Walk the position of precision target and the classification work of target.
Hyper parameter described in step (3) of the present invention is: learning rate is usually arranged as between 0.0001~0.01, the design
Learning rate is relatively fixed as 0.0001 through many experiments, and the number of iterations is traditionally arranged to be 2000~50000 times, depends on data volume
Whether size and model reach stable convergence state, and present network architecture the number of iterations is designed as 50000, obtain final network mould
Type.
In step (4) of the present invention, when IOU > 0.5, while classification results are consistent with ground true label, then it is assumed that real
Now positioning and identification are correct.
Compared with the prior art, the present invention has the advantage that
First, the Thyroid ultrasound image tubercle provided by the invention based on USFaster R-CNN is automatically positioned identification side
Method has objectivity, effectively avoids diagnosing for assisting doctor to carry out the positioning and classification of focal zone in Thyroid ultrasound image
Doctor leads to lesser tubercle missing inspection due to visual fatigue or experience deficiency.It is a kind of full automatic positioning and recognition methods, is not necessarily to people
Work participates in, only need to be by Thyroid ultrasound image reading into deep learning network, and system will outline lesion automatically and take and mark knot
Save type.
Second, using network parameter moving method, shallower low-level image feature network and addition merging features layer.Can have
Effect makes up the few deficiency of medical image amount, while increasing the influence of shallow-layer textural characteristics in the entire network, this is more suitable
The ultrasound image less for feature.1*1 convolution operation is added, a liter peacekeeping dimensionality reduction is carried out to characteristic pattern, it is possible to reduce training ginseng
Number, accelerates the training speed of network, increases the non-thread sexuality of network.
Third, shallower feature extraction network can obtain biggish characteristic pattern, improve the resolution ratio of characteristic pattern, more great Cheng
Degree ground utilizes the textural characteristics of bottom convolution.This can allow candidate frame to generate network and generate more candidate frames in ultrasound image,
Frame is sized to [64,128,256], meets the size of thyroid nodule in ultrasound image, therefore for lesser tubercle
Recall rate has very big promotion.
Detailed description of the invention
Fig. 1 is the Thyroid ultrasound image preprocessing figure in the embodiment of the present invention；
Fig. 2 is to utilize USFaster R-CNN deep learning network training process figure in the embodiment of the present invention；
Fig. 3 is the bottom convolution feature extraction network structure in the embodiment of the present invention；
Fig. 4 is that the candidate frame in the embodiment of the present invention generates network details figure；
Fig. 5 is the Thyroid ultrasound image tubercle automatic positioning recognition effect figure in the embodiment of the present invention.
Specific embodiment
Following embodiment should not be limited the scope of the invention for invention is further described in detail.
A kind of Thyroid ultrasound image tubercle automatic positioning based on USFaster R-CNN provided in an embodiment of the present invention
Recognition methods, the specific steps are as follows:
Step 1: image procossing.Thyroid ultrasound image is obtained, collects the Thyroid ultrasound image of 300 people altogether from hospital,
Wherein 250 people are diagnosed as with benign or Malignant Nodules, and 50 people are normal.Everyone includes 5-15 images, shares 2232
Open image.Ultrasound image is pre-processed: only intercepting out using thyroid gland part in image as area-of-interest first, go
Except in image with thyroid gland irrelevant information, as shown in figure 1 shown in b.According to diagnosis report in description and ultrasound image in
Focal zone mark position ground true mark carried out to focal area make corresponding xml document, which will not scheme
It is shown as in, position and the information of focal zone has only been indicated in xml document.The corresponding xml document of every image.
Then the label that removal image teacher does focal zone when reading ultrasonic picture.Handmarking not only can image focal zone image
Texture, also will affect focal zone positioning confidence level.Removing handmarking, steps are as follows:
(1) feature more considerably higher than image peripheral regional luminance according to label area, using Threshold segmentation by image two
Value.
(2) all four connected regions of bianry image are calculated and are sorted, wherein the gray value of pixel is greater than some threshold value
Connected region seek to the label area looked for, by observation count, effect is best when threshold value takes 20.
(3) region found out in step (2) is carried out by interpolation using bilinear interpolation algorithm, reaches the mesh of removal label
's.Effect is as shown in figure 1 shown in c.
The ultrasound image of 50 people is taken out at random from 300 people as test set, in order to guarantee the extensive of test model
Ability, the image of this 50 people is not involved in any training related with building model, then by remaining 250 people according to 4:1's
Ratio classification based training collection and verifying collection.Model parameter of the training set for training building, verifying collection are used to adjust super in model
Parameter is optimal network model.
Step 2: network model is built.For deep learning network structure by bottom convolution feature extraction network, candidate frame is raw
At network, characteristic pattern pond layer, classification and candidate frame Recurrent networks composition.Wherein bottom convolution feature extraction network uses 4
Convolutional layer follows 1 layer of maximum pond layer after each convolutional layer, first three convolutional layer respectively has level 2 volume to accumulate, and the 4th convolutional layer has
3 layers of convolution, while designing a layer and two layers of convolution of third layer stacks up, then dimension is risen by the convolutional layer of a 1*1
Total output of the results added of output and the 4th layer of convolution as feature extraction network.Network structure is as shown in Figure 3.Feature extraction
Image is mainly carried out down-sampling by network, while being extracted automatically to the feature in image by convolution, is obtained with original image
As the characteristic pattern of characteristic information.Characteristic pattern is re-used as input and enters candidate frame generation network, generates the detailed process of candidate frame such as
Shown in Fig. 4.Using a 3*3 size, step-length is 1 convolution sliding window, and anchor point, Mei Gemao are generated on each pixel
The corresponding candidate frame that 9 different proportions and size are generated in original image of point.Then characteristic pattern will by two 1*1 convolution, one
Group obtains two scores, in the candidate frame generated whether include object probability, IOU is greater than the conduct prospect of threshold value, i.e.,
Containing target, the judgement less than threshold value is background, that is, does not include target.Another group obtains 4 coordinates, respectively generation candidate frame
Abscissa, ordinate is wide and high.Automatic screening is then carried out, a small amount of candidate frame containing target is left behind.Subsequently into
Characteristic pattern pond floor is unified by candidate frame area size, thus can be as the input classified below with frame Recurrent networks.Classification and
The effect of frame Recurrent networks is that further the target in candidate frame is classified and the position correction of candidate frame, is reached more
Accurate target positioning and identification.
Step 3: network training.Since Thyroid ultrasound image data set is too small, being not enough to training has very strong extensive energy
The deep learning network of power, it is easy to over-fitting occur.Network is trained with large data collection ImageNet data set first,
Network parameter is obtained, then parameter migration comes when using Thyroid ultrasound image training network, network receipts can be accelerated
It holds back, also makes network with more generalization ability, as shown in Figure 2.Ubuntu16.04 64 when the operating system environment of training network,
Hardware GPU is accelerated by Nvidia TITAN X (Pascal) video card.After obtaining preliminary network parameter, surveyed using verifying collection
Network output effect is tried, the hyper parameter in network is debugged and optimized, depth network is made to be optimal state.Final study
Rate is designed as fixing 0.0001, and the number of iterations is designed as 50000, and the loss value of network model tends towards stability, and represents deep learning
Network acquires the feature in image well.
Step 4: thyroid nodule automatic positioning and identification.The Thyroid ultrasound image reading of test set is arrived
In USFaster R-CNN deep learning network, it will realize and outline thyroid nodule automatically and judge tubercle type.Wherein wrap
The lesser tubercle of missing inspection is easy containing doctor.When IOU > 0.5, while classification results are consistent with ground true label, then it is assumed that realize
Positioning is correct with identification.Recognition effect is as shown in Figure 5.
Claims (9)
1. a kind of Thyroid ultrasound image tubercle based on USFaster R-CNN is automatically positioned recognition methods, which is characterized in that
The following steps are included:
(1) Thyroid ultrasound image is pre-processed first, intercepts out pending area, remove in ultrasound image with thyroid gland
Image irrelevant information outlines ultrasonic figure according to the label of diagnostician and diagnosis report with image labeling tool labelImg
Thyroid nodule as in, and the mark of tubercle type is carried out, the ground true for being fabricated to xml document as training set is marked
Label, then remove the artificial mark of the diagnostician in image, in order to avoid the training effect to network is influenced, finally by ultrasound image
And corresponding xml document is divided into training set, verifying collection and test set；
(2) with the large data collection ImageNet data set USFaster R-CNN depth nerve net that training has been put up in advance
Network obtains trained network parameter, then will be trained in advance when with Thyroid ultrasound training set of images training network
Network parameter moves in USFaster R-CNN network, so that network convergence obtains faster and has stronger extensive energy
Power；
(3) trained hyper parameter is finely adjusted and is optimized using verifying collection after tentatively training network model, reach model
Effect is identified and positioned to best；
(4) by the Thyroid ultrasound image to be identified with positioning, i.e. test set, USFaster R-CNN network is read, is realized
The automatic positioning and identification of thyroid nodule.
Know 2. a kind of Thyroid ultrasound image tubercle based on USFaster R-CNN according to claim 1 is automatically positioned
Other method, which is characterized in that the handmarking of the diagnostician in removal image described in step (1), it will using Threshold segmentation
Image binaryzation calculates all four connected regions of bianry image and sorts, and wherein pixel is greater than the connected region of threshold value
It is the handmarking region for needing to look for, connected region is then carried out by interpolation using bilinear interpolation algorithm, it is artificial reaches removal
The purpose of label.
Know 3. a kind of Thyroid ultrasound image tubercle based on USFaster R-CNN according to claim 2 is automatically positioned
Other method, which is characterized in that the threshold value is 10~20.
Know 4. a kind of Thyroid ultrasound image tubercle based on USFaster R-CNN according to claim 1 is automatically positioned
Other method, which is characterized in that USFaster R-CNN deep neural network described in step (2), including bottom convolution feature mention
Network is taken, candidate frame generates network, characteristic pattern pond layer, classification and candidate frame Recurrent networks.
Know 5. a kind of Thyroid ultrasound image tubercle based on USFaster R-CNN according to claim 4 is automatically positioned
Other method, which is characterized in that the structure of the bottom convolution feature extraction network are as follows: 4 convolutional layers are used, after each convolutional layer
1 layer of maximum pond layer is all followed, first three convolutional layer respectively has level 2 volume to accumulate, and the 4th convolutional layer haves three layers convolution, while design one
A merging features layer first stacks up two layers of convolution of third layer, then by the convolutional layer of 1*1 rise the output of dimension with
Total output of the output as feature extraction network after the result splicing of 4th layer of convolution, obtains with Thyroid ultrasound image information
Characteristic pattern.
Know 6. a kind of Thyroid ultrasound image tubercle based on USFaster R-CNN according to claim 4 is automatically positioned
Other method, which is characterized in that it is (W/8*H/8) * that the candidate frame, which generates the candidate frame quantity that network generates on the ultrasound image,
9, wherein W, H are respectively the width and height of ultrasound image, and the side length of the size of candidate frame is 64,128 and 256, three kinds of ratio difference
For 1:1,1:2 and 2:1.
Know 7. a kind of Thyroid ultrasound image tubercle based on USFaster R-CNN according to claim 4 is automatically positioned
Other method, which is characterized in that be added a 1*1*128's between characteristic pattern pond layer, classification and candidate frame Recurrent networks
Convolutional layer can increase the non-linear of network while reducing parameter, improve fitting effect.
Know 8. a kind of Thyroid ultrasound image tubercle based on USFaster R-CNN according to claim 1 is automatically positioned
Other method, which is characterized in that hyper parameter described in step (3) is: learning rate is 0.0001~0.01, the number of iterations 2000
~50000 times.
Know 9. a kind of Thyroid ultrasound image tubercle based on USFaster R-CNN according to claim 1 is automatically positioned
Other method, which is characterized in that in step (4), when IOU > 0.5, while classification results are consistent with ground true label, then recognize
To realize that positioning is correct with identification.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910594057.1A CN110490892A (en) | 2019-07-03 | 2019-07-03 | A kind of Thyroid ultrasound image tubercle automatic positioning recognition methods based on USFaster R-CNN |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910594057.1A CN110490892A (en) | 2019-07-03 | 2019-07-03 | A kind of Thyroid ultrasound image tubercle automatic positioning recognition methods based on USFaster R-CNN |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110490892A true CN110490892A (en) | 2019-11-22 |
Family
ID=68546544
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910594057.1A Pending CN110490892A (en) | 2019-07-03 | 2019-07-03 | A kind of Thyroid ultrasound image tubercle automatic positioning recognition methods based on USFaster R-CNN |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110490892A (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111160441A (en) * | 2019-12-24 | 2020-05-15 | 上海联影智能医疗科技有限公司 | Classification method, computer device, and storage medium |
| CN111461202A (en) * | 2020-03-30 | 2020-07-28 | 上海尽星生物科技有限责任公司 | Real-time thyroid nodule ultrasonic image identification method and device |
| CN111612752A (en) * | 2020-05-15 | 2020-09-01 | 江苏省人民医院（南京医科大学第一附属医院） | Ultrasonic image thyroid nodule intelligent detection system based on fast-RCNN |
| CN111652838A (en) * | 2020-04-16 | 2020-09-11 | 上海长征医院 | Thyroid nodule positioning and ultrasonic report error correction method based on target detection network |
| CN111820950A (en) * | 2020-06-23 | 2020-10-27 | 无锡祥生医疗科技股份有限公司 | Personalized information determination device and ultrasonic training method |
| CN111862044A (en) * | 2020-07-21 | 2020-10-30 | 长沙大端信息科技有限公司 | Ultrasonic image processing method and device, computer equipment and storage medium |
| CN112001877A (en) * | 2020-05-18 | 2020-11-27 | 天津大学 | Thyroid malignant nodule detection method based on deep learning |
| CN112036400A (en) * | 2020-07-09 | 2020-12-04 | 北京航空航天大学 | Method for constructing network for target detection and target detection method and system |
| CN112419396A (en) * | 2020-12-03 | 2021-02-26 | 前线智能科技(南京)有限公司 | Thyroid ultrasonic video automatic analysis method and system |
| CN112614108A (en) * | 2020-12-24 | 2021-04-06 | 中国人民解放军总医院第一医学中心 | Method and device for detecting nodules in thyroid ultrasound image based on deep learning |
| CN112754530A (en) * | 2021-01-18 | 2021-05-07 | 居天智慧(深圳)有限公司 | Thyroid gland discernment wireless ultrasound appearance |
| CN112927217A (en) * | 2021-03-23 | 2021-06-08 | 内蒙古大学 | Thyroid nodule invasiveness prediction method based on target detection |
| CN113555087A (en) * | 2021-07-19 | 2021-10-26 | 吉林大学第一医院 | Artificial intelligence film reading method based on convolutional neural network algorithm |
| CN113610859A (en) * | 2021-06-07 | 2021-11-05 | 东北大学 | Automatic thyroid nodule segmentation method based on ultrasonic image |
| CN115294401A (en) * | 2022-08-23 | 2022-11-04 | 北京医准智能科技有限公司 | CT image-based classification method, device, equipment and storage medium |
Citations (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106780448A (en) * | 2016-12-05 | 2017-05-31 | 清华大学 | A kind of pernicious sorting technique of ultrasonic Benign Thyroid Nodules based on transfer learning Yu Fusion Features |
| CN107239802A (en) * | 2017-06-28 | 2017-10-10 | 广东工业大学 | A kind of image classification method and device |
| CN107451615A (en) * | 2017-08-01 | 2017-12-08 | 广东工业大学 | Thyroid papillary carcinoma Ultrasound Image Recognition Method and system based on Faster RCNN |
| CN107527065A (en) * | 2017-07-25 | 2017-12-29 | 北京联合大学 | A kind of flower variety identification model method for building up based on convolutional neural networks |
| CN108108677A (en) * | 2017-12-12 | 2018-06-01 | 重庆邮电大学 | One kind is based on improved CNN facial expression recognizing methods |
| CN108509839A (en) * | 2018-02-02 | 2018-09-07 | 东华大学 | One kind being based on the efficient gestures detection recognition methods of region convolutional neural networks |
| CN108520518A (en) * | 2018-04-10 | 2018-09-11 | 复旦大学附属肿瘤医院 | A kind of thyroid tumors Ultrasound Image Recognition Method and its device |
| CN108564123A (en) * | 2018-04-10 | 2018-09-21 | 复旦大学附属肿瘤医院 | A kind of Thyroid Neoplasms smear image classification method and its device |
| CN108710919A (en) * | 2018-05-25 | 2018-10-26 | 东南大学 | A kind of crack automation delineation method based on multi-scale feature fusion deep learning |
| CN108875592A (en) * | 2018-04-13 | 2018-11-23 | 哈尔滨工程大学 | A kind of convolutional neural networks optimization method based on attention |
| CN109036467A (en) * | 2018-10-26 | 2018-12-18 | 南京邮电大学 | CFFD extracting method, speech-emotion recognition method and system based on TF-LSTM |
| CN109086678A (en) * | 2018-07-09 | 2018-12-25 | 天津大学 | A kind of pedestrian detection method extracting image multi-stage characteristics based on depth supervised learning |
| CN109117894A (en) * | 2018-08-29 | 2019-01-01 | 汕头大学 | A kind of large scale remote sensing images building classification method based on full convolutional neural networks |
| CN109242844A (en) * | 2018-09-04 | 2019-01-18 | 青岛大学附属医院 | Pancreatic tumour automatic recognition system based on deep learning, computer equipment, storage medium |
| CN109241967A (en) * | 2018-09-04 | 2019-01-18 | 青岛大学附属医院 | Thyroid ultrasound automatic image recognition system, computer equipment, storage medium based on deep neural network |
| CN109447940A (en) * | 2018-08-28 | 2019-03-08 | 天津医科大学肿瘤医院 | Convolutional neural networks training method, ultrasound image recognition positioning method and system |
| CN109493333A (en) * | 2018-11-08 | 2019-03-19 | 四川大学 | Ultrasonic Calcification in Thyroid Node point extraction algorithm based on convolutional neural networks |
| CN109543623A (en) * | 2018-11-26 | 2019-03-29 | 微医云(杭州)控股有限公司 | A kind of development of fetus condition predicting device based on Magnetic resonance imaging |
| CN109635882A (en) * | 2019-01-23 | 2019-04-16 | 福州大学 | Salient object detection method based on multi-scale convolution feature extraction and fusion |
| CN109872296A (en) * | 2019-01-04 | 2019-06-11 | 中山大学 | A kind of data enhancement methods that the thyroid nodule focal zone based on depth convolution production confrontation network generates |
| CN109886160A (en) * | 2019-01-30 | 2019-06-14 | 浙江工商大学 | It is a kind of it is non-limiting under the conditions of face identification method |
-
2019
- 2019-07-03 CN CN201910594057.1A patent/CN110490892A/en active Pending
Patent Citations (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106780448A (en) * | 2016-12-05 | 2017-05-31 | 清华大学 | A kind of pernicious sorting technique of ultrasonic Benign Thyroid Nodules based on transfer learning Yu Fusion Features |
| CN107239802A (en) * | 2017-06-28 | 2017-10-10 | 广东工业大学 | A kind of image classification method and device |
| CN107527065A (en) * | 2017-07-25 | 2017-12-29 | 北京联合大学 | A kind of flower variety identification model method for building up based on convolutional neural networks |
| CN107451615A (en) * | 2017-08-01 | 2017-12-08 | 广东工业大学 | Thyroid papillary carcinoma Ultrasound Image Recognition Method and system based on Faster RCNN |
| CN108108677A (en) * | 2017-12-12 | 2018-06-01 | 重庆邮电大学 | One kind is based on improved CNN facial expression recognizing methods |
| CN108509839A (en) * | 2018-02-02 | 2018-09-07 | 东华大学 | One kind being based on the efficient gestures detection recognition methods of region convolutional neural networks |
| CN108520518A (en) * | 2018-04-10 | 2018-09-11 | 复旦大学附属肿瘤医院 | A kind of thyroid tumors Ultrasound Image Recognition Method and its device |
| CN108564123A (en) * | 2018-04-10 | 2018-09-21 | 复旦大学附属肿瘤医院 | A kind of Thyroid Neoplasms smear image classification method and its device |
| CN108875592A (en) * | 2018-04-13 | 2018-11-23 | 哈尔滨工程大学 | A kind of convolutional neural networks optimization method based on attention |
| CN108710919A (en) * | 2018-05-25 | 2018-10-26 | 东南大学 | A kind of crack automation delineation method based on multi-scale feature fusion deep learning |
| CN109086678A (en) * | 2018-07-09 | 2018-12-25 | 天津大学 | A kind of pedestrian detection method extracting image multi-stage characteristics based on depth supervised learning |
| CN109447940A (en) * | 2018-08-28 | 2019-03-08 | 天津医科大学肿瘤医院 | Convolutional neural networks training method, ultrasound image recognition positioning method and system |
| CN109117894A (en) * | 2018-08-29 | 2019-01-01 | 汕头大学 | A kind of large scale remote sensing images building classification method based on full convolutional neural networks |
| CN109242844A (en) * | 2018-09-04 | 2019-01-18 | 青岛大学附属医院 | Pancreatic tumour automatic recognition system based on deep learning, computer equipment, storage medium |
| CN109241967A (en) * | 2018-09-04 | 2019-01-18 | 青岛大学附属医院 | Thyroid ultrasound automatic image recognition system, computer equipment, storage medium based on deep neural network |
| CN109036467A (en) * | 2018-10-26 | 2018-12-18 | 南京邮电大学 | CFFD extracting method, speech-emotion recognition method and system based on TF-LSTM |
| CN109493333A (en) * | 2018-11-08 | 2019-03-19 | 四川大学 | Ultrasonic Calcification in Thyroid Node point extraction algorithm based on convolutional neural networks |
| CN109543623A (en) * | 2018-11-26 | 2019-03-29 | 微医云(杭州)控股有限公司 | A kind of development of fetus condition predicting device based on Magnetic resonance imaging |
| CN109872296A (en) * | 2019-01-04 | 2019-06-11 | 中山大学 | A kind of data enhancement methods that the thyroid nodule focal zone based on depth convolution production confrontation network generates |
| CN109635882A (en) * | 2019-01-23 | 2019-04-16 | 福州大学 | Salient object detection method based on multi-scale convolution feature extraction and fusion |
| CN109886160A (en) * | 2019-01-30 | 2019-06-14 | 浙江工商大学 | It is a kind of it is non-limiting under the conditions of face identification method |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111160441B (en) * | 2019-12-24 | 2024-03-26 | 上海联影智能医疗科技有限公司 | Classification method, computer device, and storage medium |
| CN111160441A (en) * | 2019-12-24 | 2020-05-15 | 上海联影智能医疗科技有限公司 | Classification method, computer device, and storage medium |
| CN111461202A (en) * | 2020-03-30 | 2020-07-28 | 上海尽星生物科技有限责任公司 | Real-time thyroid nodule ultrasonic image identification method and device |
| CN111461202B (en) * | 2020-03-30 | 2023-12-05 | 上海深至信息科技有限公司 | Thyroid nodule ultrasonic image real-time identification method and device |
| CN111652838A (en) * | 2020-04-16 | 2020-09-11 | 上海长征医院 | Thyroid nodule positioning and ultrasonic report error correction method based on target detection network |
| CN111652838B (en) * | 2020-04-16 | 2022-09-09 | 上海长征医院 | Thyroid nodule positioning and ultrasonic report error correction method based on target detection network |
| CN111612752A (en) * | 2020-05-15 | 2020-09-01 | 江苏省人民医院（南京医科大学第一附属医院） | Ultrasonic image thyroid nodule intelligent detection system based on fast-RCNN |
| CN112001877A (en) * | 2020-05-18 | 2020-11-27 | 天津大学 | Thyroid malignant nodule detection method based on deep learning |
| CN111820950A (en) * | 2020-06-23 | 2020-10-27 | 无锡祥生医疗科技股份有限公司 | Personalized information determination device and ultrasonic training method |
| CN112036400A (en) * | 2020-07-09 | 2020-12-04 | 北京航空航天大学 | Method for constructing network for target detection and target detection method and system |
| CN112036400B (en) * | 2020-07-09 | 2022-04-05 | 北京航空航天大学 | Method for constructing network for target detection and target detection method and system |
| CN111862044A (en) * | 2020-07-21 | 2020-10-30 | 长沙大端信息科技有限公司 | Ultrasonic image processing method and device, computer equipment and storage medium |
| CN112419396A (en) * | 2020-12-03 | 2021-02-26 | 前线智能科技(南京)有限公司 | Thyroid ultrasonic video automatic analysis method and system |
| CN112419396B (en) * | 2020-12-03 | 2024-04-26 | 前线智能科技(南京)有限公司 | Automatic thyroid ultrasonic video analysis method and system |
| CN112614108A (en) * | 2020-12-24 | 2021-04-06 | 中国人民解放军总医院第一医学中心 | Method and device for detecting nodules in thyroid ultrasound image based on deep learning |
| CN112614108B (en) * | 2020-12-24 | 2024-04-19 | 中国人民解放军总医院第一医学中心 | Method and device for detecting nodules in thyroid ultrasound image based on deep learning |
| CN112754530A (en) * | 2021-01-18 | 2021-05-07 | 居天智慧(深圳)有限公司 | Thyroid gland discernment wireless ultrasound appearance |
| CN112927217B (en) * | 2021-03-23 | 2022-05-03 | 内蒙古大学 | Thyroid nodule invasiveness prediction method based on target detection |
| CN112927217A (en) * | 2021-03-23 | 2021-06-08 | 内蒙古大学 | Thyroid nodule invasiveness prediction method based on target detection |
| CN113610859A (en) * | 2021-06-07 | 2021-11-05 | 东北大学 | Automatic thyroid nodule segmentation method based on ultrasonic image |
| CN113610859B (en) * | 2021-06-07 | 2023-10-31 | 东北大学 | Automatic thyroid nodule segmentation method based on ultrasonic image |
| CN113555087A (en) * | 2021-07-19 | 2021-10-26 | 吉林大学第一医院 | Artificial intelligence film reading method based on convolutional neural network algorithm |
| CN115294401B (en) * | 2022-08-23 | 2023-04-14 | 北京医准智能科技有限公司 | CT image-based classification method, device, equipment and storage medium |
| CN115294401A (en) * | 2022-08-23 | 2022-11-04 | 北京医准智能科技有限公司 | CT image-based classification method, device, equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110490892A (en) | A kind of Thyroid ultrasound image tubercle automatic positioning recognition methods based on USFaster R-CNN | |
| CN108464840B (en) | Automatic detection method and system for breast lumps | |
| CN110992382A (en) | Fundus image optic cup optic disc segmentation method and system for assisting glaucoma screening | |
| CN111223553A (en) | Two-stage deep migration learning traditional Chinese medicine tongue diagnosis model | |
| CN106056595A (en) | Method for automatically identifying whether thyroid nodule is benign or malignant based on deep convolutional neural network | |
| Pan et al. | Mitosis detection techniques in H&E stained breast cancer pathological images: A comprehensive review | |
| CN109063712A (en) | A kind of multi-model Hepatic diffused lesion intelligent diagnosing method and system based on ultrasound image | |
| CN111951221A (en) | Glomerular cell image identification method based on deep neural network | |
| CN101103924A (en) | Galactophore cancer computer auxiliary diagnosis method based on galactophore X-ray radiography and system thereof | |
| CN108549912A (en) | A kind of medical image pulmonary nodule detection method based on machine learning | |
| CN112085113B (en) | Severe tumor image recognition system and method | |
| Wang et al. | A method of ultrasonic image recognition for thyroid papillary carcinoma based on deep convolution neural network | |
| CN113744183A (en) | Pulmonary nodule detection method and system | |
| CN115115841B (en) | Shadow spot image processing and analyzing method and system | |
| Wang et al. | Cataract detection based on ocular B-ultrasound images by collaborative monitoring deep learning | |
| CN109087310A (en) | Dividing method, system, storage medium and the intelligent terminal of Meibomian gland texture region | |
| Teng | Brief Review of Medical Image Segmentation Based on Deep Learning | |
| Peng et al. | Cov-TransNet: Dual branch fusion network with transformer for COVID-19 infection segmentation | |
| CN110288577A (en) | A kind of 3D realizes the control method and control device of multiple organ lesion integration | |
| CN112419246B (en) | Depth detection network for quantifying esophageal mucosa IPCLs blood vessel morphological distribution | |
| CN115880266B (en) | Intestinal polyp detection system and method based on deep learning | |
| CN117036288A (en) | Tumor subtype diagnosis method for full-slice pathological image | |
| Wang et al. | Optic disc detection based on fully convolutional neural network and structured matrix decomposition | |
| CN109816665A (en) | A kind of fast partition method and device of optical coherence tomographic image | |
| Liu et al. | One-stage attention-based network for image classification and segmentation on optical coherence tomography 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: 20191122 |