CN109493333A - Ultrasonic Calcification in Thyroid Node point extraction algorithm based on convolutional neural networks - Google Patents
Ultrasonic Calcification in Thyroid Node point extraction algorithm based on convolutional neural networks Download PDFInfo
- Publication number
- CN109493333A CN109493333A CN201811327217.8A CN201811327217A CN109493333A CN 109493333 A CN109493333 A CN 109493333A CN 201811327217 A CN201811327217 A CN 201811327217A CN 109493333 A CN109493333 A CN 109493333A
- Authority
- CN
- China
- Prior art keywords
- image
- calcification
- convolutional neural
- neural networks
- thyroid
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06T7/10—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
- G06T2207/20084—Artificial neural networks [ANN]
Abstract
The invention discloses the ultrasonic Calcification in Thyroid Node point extraction algorithms based on convolutional neural networks, comprising the following steps: A, prepares training set image;B, final network is obtained using the training set image training convolutional neural networks;C, calcification point in thyroid nodule ultrasound image is differentiated using the final network.It is based on convolutional neural networks and extracts to calcification feature, get up to judge by the brightness of calcification point, form, with perienchyma's relationship and some Fusion Features for being not easy quantization, avoid it is traditional only differentiate by single brightness lead to the problem of judging by accident;Output label is set as the binary segmentation image for coming out calcification Image Segmentation Methods Based on Features, so that the parameter of whole network is intended to be mapped to original image on the binary segmentation image, calcification point is finally made to differentiate that accuracy rate significantly improves.
Description
Technical field
The invention belongs to computer image processing technology fields, and in particular to the ultrasonic thyroid gland based on convolutional neural networks
Tubercle calcification point extraction algorithm.
Background technique
Conventional method, which has based on local maxima Ostu method (Otsu), judges calcification feature, and method is using to knot
Section area dividing simultaneously carries out the thought of Otsu Threshold segmentation to extract calcification point.The tubercle manual segmentation of every picture is gone out first
Come, then constructs one centered on tubercle and include tubercle and tangent rectangle, then put down the rectangular area
It is divided into the rectangular area similar with former rectangle of 5 × 5 25 same sizes, and Otsu threshold value Ti is calculated to each region,
And be compared Ti with predetermined calcification segmentation threshold T, Otsu segmentation is carried out to this region if Ti is greater than T, instead
Then in this area without segmentation.Segmentation segmentation threshold being greater than respectively again after sectioning in the region of calibration threshold value
Region carries out region growing and obtains final calcified regions, and is differentiated according to result to calcification feature.
Traditional calcification point judging method is substantially the brightness based on calcification point, with the ratio of single or multiple threshold value
Compared with to determine whether being calcification, these methods have very big drawback.Firstly, calcification point is not necessarily tied in ultrasonic picture
The highest region of brightness in section and its perienchyma, has the brightness of some calcification points not high, but the comparison of it and surrounding tissue
It spends very high.Secondly, there are many thyroid nodule type, many tubercles inside and out it regional organization it is complicated, have much highlighted groups
Knit, such as cyst wall, if only using brightness as the discrimination standard of calcification point, then these highlight non-calcified tissue be easy to by
It is determined as calcification point.Third, it is the resolution ratio of different ultrasound images, clarity, whole since to obtain picture quality different for ultrasound machine
The many indexs such as body brightness, contrast all can be variant, thus only with simple brightness be extract feature, without with perinodal
It is good at correctly extracting so that the integrated environment of whole image compares.4th, when differentiating calcification point, only make
Use the brightness of calcification point as discrimination standard, and also none is comprehensive to other features such as its position, form
It uses, discrimination standard is excessively single.
Method more common at present is directly to be classified using convolutional neural networks to calcification feature, is with Alexnet
Example, Alexnet network structure as shown in Figure 1, wherein input be original image, export to classify where the image, label is by 1 He
The bivector of 0 composition is trained determining network weight to network, keeps weight constant, input picture obtains image again
Classification.
Convolutional neural networks are typically applied to classification task, it is assumed that it needs to be divided into N class, then label when it is trained is one
A N-dimensional vector being made of 1 and 0 represents the probability distribution that the sample is divided into each class.Convolutional neural networks use BP algorithm
Carry out Training, two stages of backpropagation of forward-propagating and error by information, therefore, export constantly reduce with
Error between label vector is to achieve the purpose that trained network parameter.But the parameter of whole network is made to be intended to make in this way
Original image is mapped in label vector, and label vector only has 1 and 0 two value, can not accurate representation feature required for us,
Therefore recognition accuracy is not high.
Summary of the invention
Based on this, in view of the above-mentioned problems, the present invention proposes the ultrasonic Calcification in Thyroid Node point based on convolutional neural networks
Extraction algorithm.
The technical scheme is that the ultrasonic Calcification in Thyroid Node point extraction algorithm based on convolutional neural networks, packet
Include following steps:
A, prepare training set image;
B, the training set image training convolutional neural networks are utilized;
C, calcification point in thyroid nodule ultrasound image is differentiated using trained convolutional neural networks.
Optionally, step A includes following procedure:
A1, the ultrasound image comprising thyroid nodule is obtained, selects the minimum value in tri- channels entire image RGB as figure
The gray value removal colorful blood interference of picture is cut then by every picture centered on tubercle, using tubercle extreme length as side length
Take the square area comprising a part of surrounding enviroment;
A2, the above square area is adjusted to unified size, such as 224*224, RGB triple channel is taken, as final
Training set image.
Optionally, step B includes following procedure:
B1, using training set image as input, to be partitioned into the binary image of calcification feature as label, training is divided
Network;
B2, the front portion structure for dividing network and weight remain unchanged, and modify the structure of the rear part, are changed as classifying
Network;
B3, using training set image as input, using whether there is or not calcifications to train sorter network, as final net as label
Network.
Optionally, step C is comprised the steps of:
Step C1: ultrasonic thyroid nodule image is obtained, image is handled according to method in step A1-A2, is obtained
Images to be recognized;
Step C2: using images to be recognized as the input of sorter network, predicting it, and obtains a bivector
To indicate the prediction result to the image.
The beneficial effects of the present invention are:
(1) compared to conventional method, calcification feature is extracted based on convolutional neural networks, passes through the side of deep learning
The feature of method global learning calcification point, by the brightness of calcification point, form, with perienchyma's relationship and it is some be not easy quantization
Fusion Features get up to be judged, avoid and traditional only carry out differentiating that cause erroneous judgement to be misjudged asks by single brightness
Topic.
(2) directly classify relative to traditional convolutional neural networks to calcification feature, output label is set as one by us
A binary segmentation image for coming out calcification Image Segmentation Methods Based on Features, such training will make the parameter of whole network be intended to make original image
It is mapped on the binary segmentation image, the feature in characteristic pattern extracted in convolution process layer by layer would tend to two-value
The feature that segmented image is split, and such feature is exactly what is needed the feature with it for foundation classification wanted, and makes
Calcification point differentiates that accuracy rate significantly improves.
Detailed description of the invention
Fig. 1 is the schematic diagram of traditional Alexnet network structure.
Fig. 2 is the schematic diagram of Alexnet-2 network structure.
Fig. 3 is the schematic diagram of AlexFCN_seg1 network structure.
Fig. 4 is algorithm flow chart.
Specific embodiment
The embodiment of the present invention is described in detail with reference to the accompanying drawing.
Embodiment
As shown in figure 4, the ultrasonic Calcification in Thyroid Node point extraction algorithm based on convolutional neural networks, including following step
It is rapid:
A, prepare training set image;
B, the training set image training convolutional neural networks are utilized;
C, calcification point in thyroid nodule ultrasound image is differentiated using trained convolutional neural networks.
Specifically, step A includes following procedure:
A1, the ultrasound image comprising thyroid nodule is obtained, selects the minimum value in tri- channels entire image RGB as figure
The gray value removal colorful blood interference of picture is cut then by every picture centered on tubercle, using tubercle extreme length as side length
Take the square area comprising a part of surrounding enviroment;
A2, the above square area is adjusted to unified size, such as is uniformly adjusted to 224*224, take RGB threeway
Road, as final training set image.
Step B includes following procedure:
B1, it is adjusted using traditional Alexnet network structure as template, forms segmentation network structure, be named as
AlexFCN_seg1, and train it, wherein traditional Alexnet network structure is as shown in Figure 1;
B2, the front portion structure for dividing network A lexFCN_seg1 and weight remain unchanged, and the structure of the rear part are modified, by it
It is changed as sorter network Alexnet-2;
B3, using the training set image as input, the Alexnet-2 is trained, obtained result be most
Whole network.
Step B1 includes following procedure:
1) L3, L4 layers of convolution mask quantity in Alexnet network are adjusted to 390, remove L6, L7 and L8 layers, and divide
Be not configured to convolution kernel species number, width, it is high be respectively (1024,7,7), (1024,1,1), (2,1,1) three-layer coil lamination L9,
L10, L11, and adding two layers of L12, L13 below at L11 layers makes image size revert back 224 × 224, the full convolutional neural networks
It is named as AlexFCN_seg1, structure is as shown in figure 3, the specific structure of AlexFCN_seg1 is as shown in table 1 below.
Table 1
2) AlexFCN_seg1 is trained as label using the binary image for being partitioned into calcification feature.
Step B2 includes following procedure: first five layer of L1-L5 layers of weight for retaining AlexFCN_seg1 no longer change, and remove
The full articulamentum of L6-L8 is added in the L9-L13 layer of AlexFCN_seg1 network, and size is respectively 1024,1024,2, is ordered
Entitled Alexnet-2, structure chart are shown in Fig. 2.
Step B3 includes following procedure: training Alexnet-2, using training set image as input, with whether there is or not calcification points two
Category information has as 1 as corresponding label, without as 0, this time trains the weight of three layers of full articulamentum of L6-L8 after only training,
Front L1-L5 layers of weight is no longer changed, i.e., only trains the weight of sorter network;Retain all 8 layers of power after training
Weight, the network weight no longer change, and obtained Alexnet-2 is final network.
Step C is comprised the steps of:
Step C1: ultrasonic thyroid nodule image is obtained, image is handled according to method in step A1-A2, is obtained
Images to be recognized;
Step C2: using images to be recognized as the input of Alexnet-2, predicting it, and obtain a two dimension to
Amount indicates the prediction result to the image.
The invention proposes the feature extracting method classified afterwards is first divided, first sentenced using segmentation network to required in image
Another characteristic is split training, is trained weight towards original image to be mapped to the binary image for being partitioned into feature, and
The weight of keeping characteristics thermal map and its preceding layers is constant after training, removes each layer below and adds sorter network, then
Secondary carry out classification based training, each layer sorter network that classification based training adds after only training, the trained whole weights of preservation that finish are constant, obtain
To final network.
It should be noted that the above citing is only specific embodiments of the present invention, core of the present invention is by first dividing
The way of thinking of classification based training achievees the purpose that improve image classification accuracy rate after training, and any convolutional neural networks pass through
Above-mentioned thought, which is tested, can be achieved.
A specific embodiment of the invention above described embodiment only expresses, the description thereof is more specific and detailed, but simultaneously
Limitations on the scope of the patent of the present invention therefore cannot be interpreted as.It should be pointed out that for those of ordinary skill in the art
For, without departing from the inventive concept of the premise, various modifications and improvements can be made, these belong to guarantor of the invention
Protect range.
Claims (4)
1. the ultrasonic Calcification in Thyroid Node point extraction algorithm based on convolutional neural networks, it is characterised in that: the following steps are included:
A, prepare training set image;
B, using the training set image training convolutional neural networks, final network is obtained;
C, calcification point in thyroid nodule ultrasound image is differentiated using the final network.
2. the ultrasonic Calcification in Thyroid Node point extraction algorithm according to claim 1 based on convolutional neural networks, special
Sign is that step A includes following procedure:
A1, the ultrasound image comprising thyroid nodule is obtained, selects the minimum value in tri- channels entire image RGB as image
Gray value removes colorful blood interference, then by every picture centered on tubercle, using tubercle extreme length as side length, interception packet
Square area containing a part of surrounding enviroment;
A2, the above square area is adjusted to unified size, takes RGB triple channel, as final training set image.
3. the ultrasonic Calcification in Thyroid Node point extraction algorithm according to claim 1 or 2 based on convolutional neural networks,
Be characterized in that: step B includes following procedure:
B1, using the training set image as input, to be partitioned into the binary image of calcification feature as label, training is divided
Network;
B2, the front portion structure for dividing network and weight remain unchanged, and modify the structure of the rear part, are changed as classification net
Network;
B3, using training set image as input, using whether there is or not calcifications to train sorter network, as final network as label.
4. the ultrasonic Calcification in Thyroid Node point extraction algorithm according to claim 3 based on convolutional neural networks, special
Sign is that step C is comprised the steps of:
Step C1: ultrasonic thyroid nodule image is obtained, image is handled according to method in step A1-A2, is obtained wait know
Other image;
Step C2: using images to be recognized as the input of the final network, predicting it, and obtains a bivector
To indicate the prediction result to the image.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811327217.8A CN109493333A (en) | 2018-11-08 | 2018-11-08 | Ultrasonic Calcification in Thyroid Node point extraction algorithm based on convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811327217.8A CN109493333A (en) | 2018-11-08 | 2018-11-08 | Ultrasonic Calcification in Thyroid Node point extraction algorithm based on convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109493333A true CN109493333A (en) | 2019-03-19 |
Family
ID=65695591
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811327217.8A Pending CN109493333A (en) | 2018-11-08 | 2018-11-08 | Ultrasonic Calcification in Thyroid Node point extraction algorithm based on convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109493333A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110264461A (en) * | 2019-06-25 | 2019-09-20 | 南京工程学院 | Microcalciffcation point automatic testing method based on ultrasonic tumor of breast image |
CN110288574A (en) * | 2019-06-13 | 2019-09-27 | 南通市传染病防治院(南通市第三人民医院) | A kind of adjuvant Ultrasonographic Diagnosis hepatoncus system and method |
CN110490892A (en) * | 2019-07-03 | 2019-11-22 | 中山大学 | A kind of Thyroid ultrasound image tubercle automatic positioning recognition methods based on USFaster R-CNN |
CN111160413A (en) * | 2019-12-12 | 2020-05-15 | 天津大学 | Thyroid nodule classification method based on multi-scale feature fusion |
CN112001895A (en) * | 2020-08-03 | 2020-11-27 | 什维新智医疗科技(上海)有限公司 | Thyroid calcification detection device |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106056595A (en) * | 2015-11-30 | 2016-10-26 | 浙江德尚韵兴图像科技有限公司 | Method for automatically identifying whether thyroid nodule is benign or malignant based on deep convolutional neural network |
-
2018
- 2018-11-08 CN CN201811327217.8A patent/CN109493333A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106056595A (en) * | 2015-11-30 | 2016-10-26 | 浙江德尚韵兴图像科技有限公司 | Method for automatically identifying whether thyroid nodule is benign or malignant based on deep convolutional neural network |
Non-Patent Citations (1)
Title |
---|
左东奇等: "基于卷积神经网络提取超声图像甲状腺结节钙化点的研究", 《生物医学工程学杂志》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110288574A (en) * | 2019-06-13 | 2019-09-27 | 南通市传染病防治院(南通市第三人民医院) | A kind of adjuvant Ultrasonographic Diagnosis hepatoncus system and method |
CN110264461A (en) * | 2019-06-25 | 2019-09-20 | 南京工程学院 | Microcalciffcation point automatic testing method based on ultrasonic tumor of breast image |
CN110490892A (en) * | 2019-07-03 | 2019-11-22 | 中山大学 | A kind of Thyroid ultrasound image tubercle automatic positioning recognition methods based on USFaster R-CNN |
CN111160413A (en) * | 2019-12-12 | 2020-05-15 | 天津大学 | Thyroid nodule classification method based on multi-scale feature fusion |
CN111160413B (en) * | 2019-12-12 | 2023-11-17 | 天津大学 | Thyroid nodule classification method based on multi-scale feature fusion |
CN112001895A (en) * | 2020-08-03 | 2020-11-27 | 什维新智医疗科技(上海)有限公司 | Thyroid calcification detection device |
CN112001895B (en) * | 2020-08-03 | 2021-04-02 | 什维新智医疗科技(上海)有限公司 | Thyroid calcification detection device |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109493333A (en) | Ultrasonic Calcification in Thyroid Node point extraction algorithm based on convolutional neural networks | |
CN108427912B (en) | Optical remote sensing image target detection method based on dense target feature learning | |
CN107134144B (en) | A kind of vehicle checking method for traffic monitoring | |
CN108764164B (en) | Face detection method and system based on deformable convolution network | |
CN107330453B (en) | Pornographic image identification method based on step-by-step identification and fusion key part detection | |
CN103810503B (en) | Depth study based method for detecting salient regions in natural image | |
WO2019140767A1 (en) | Recognition system for security check and control method thereof | |
CN110210475B (en) | License plate character image segmentation method based on non-binarization and edge detection | |
CN109635694B (en) | Pedestrian detection method, device and equipment and computer readable storage medium | |
CN103049763B (en) | Context-constraint-based target identification method | |
CN105809121A (en) | Multi-characteristic synergic traffic sign detection and identification method | |
CN109934200A (en) | A kind of RGB color remote sensing images cloud detection method of optic and system based on improvement M-Net | |
CN106778835A (en) | The airport target by using remote sensing image recognition methods of fusion scene information and depth characteristic | |
CN107346420A (en) | Text detection localization method under a kind of natural scene based on deep learning | |
CN108305260B (en) | Method, device and equipment for detecting angular points in image | |
CN106815604A (en) | Method for viewing points detecting based on fusion of multi-layer information | |
CN106557740B (en) | The recognition methods of oil depot target in a kind of remote sensing images | |
CN109657551B (en) | Face detection method based on context information enhancement | |
CN108960404B (en) | Image-based crowd counting method and device | |
CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
CN110263731B (en) | Single step human face detection system | |
CN111950525B (en) | Fine-grained image classification method based on destructive reconstruction learning and GoogLeNet | |
CN107944403A (en) | Pedestrian's attribute detection method and device in a kind of image | |
CN106203496A (en) | Hydrographic curve extracting method based on machine learning | |
CN105931241A (en) | Automatic marking method for natural scene 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: 20190319 |
|
RJ01 | Rejection of invention patent application after publication |