CN108346154A - The method for building up of Lung neoplasm segmenting device based on Mask-RCNN neural networks - Google Patents
The method for building up of Lung neoplasm segmenting device based on Mask-RCNN neural networks Download PDFInfo
- Publication number
- CN108346154A CN108346154A CN201810091696.1A CN201810091696A CN108346154A CN 108346154 A CN108346154 A CN 108346154A CN 201810091696 A CN201810091696 A CN 201810091696A CN 108346154 A CN108346154 A CN 108346154A
- Authority
- CN
- China
- Prior art keywords
- lung neoplasm
- lung
- layer
- feature extraction
- network
- 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.)
- Granted
Links
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
- G06T2207/10012—Stereo images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10081—Computed x-ray tomography [CT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20021—Dividing image into blocks, subimages or windows
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30061—Lung
- G06T2207/30064—Lung nodule
Abstract
The invention discloses a kind of method for building up of the Lung neoplasm segmenting device based on Mask RCNN neural networks, including:Establish training sample:The three-dimensional lung CT image of acquisition is cut successively, data enhancing and difficulty divide negative sample excavation to handle, acquisition training sample set;Establish Lung neoplasm segmentation network:Network includes dense piece of sequentially connected input layer, the first maximum pond layer, 64*64*64 convolutional layers, the second maximum pond layer, dense piece of layer of 32*32*32, third maximum pond layer, 16*16*16 layer, characteristic pattern of the characteristic pattern of dense piece of layer outputs of 16*16*16 by the dense piece of layer output of up-sampling and 32*32*32 carries out Fusion Features, characteristic pattern after Fusion Features is input to RPN networks after the layer of the ponds POL;Training Lung neoplasm divides network:Lung neoplasm segmentation network is trained using training sample, obtains Lung neoplasm segmenting device.
Description
Technical field
The invention belongs to image processing fields, and in particular to a kind of Lung neoplasm segmentation based on Mask-RCNN neural networks
The method for building up of device.
Background technology
There are many existing method that Lung neoplasm in lung CT figure is detected using deep learning algorithm, but accuracy of detection is not high.
The main reason for causing precision not high be:
(1) the relatively low Lung neoplasm in certain specific types of the recall rate of detection-phase, the case where causing missing inspection so that detection
Precision is not high.
(2) size of Lung neoplasm is unbalanced, and smaller Lung neoplasm is easy ignored.
Based on above-mentioned two reason so that using the detection of deep learning algorithm and the Lung neoplasm typicalness split and generation
Table is insufficient.
Therefore, the accuracy and training network for improving Lung neoplasm detection are partitioned into more representative tubercle and become urgent need
It solves the problems, such as.
Invention content
The object of the present invention is to provide a kind of foundation sides of the Lung neoplasm segmenting device based on Mask-RCNN neural networks
Method.The 3-D view of the Lung neoplasm in lung CT can be more accurately and rapidly detected and determined in the device that this method is established.
For achieving the above object, the present invention has the beneficial effect that:
A kind of method for building up of the Lung neoplasm segmenting device based on Mask-RCNN neural networks, the method for building up include:
Establish training sample:First, the three-dimensional lung CT image of acquisition is cut into a large amount of cube fritter, then,
Cube fritter is handled using data enhancement methods, finally, dividing negative sample method for digging using difficulty, treated cube to enhancing
Body fritter is handled, and obtains most indistinguishable 2n negative sample and n positive sample forms training sample set;
Establish Lung neoplasm segmentation network:The Lung neoplasm segmentation network includes sequentially connected input layer, the first maximum pond
It is thick to change layer, 64*64*64 convolutional layers, the second maximum pond layer, dense piece of layer of 32*32*32, third maximum pond layer, 16*16*16
The 16*16*16 characteristic patterns of close piece of layer, dense piece of layer outputs of 16*16*16 are exported by dense piece of layer of up-sampling and 32*32*32
32*32*32 characteristic patterns carry out Fusion Features, the 32*32*32 characteristic patterns after Fusion Features are input to after the layer of the ponds POL
RPN networks are realized and the Lung neoplasm of 32*32*32 characteristic patterns are predicted and divided;
Training Lung neoplasm divides network:It is more than the fritter of 30mm to the Lung neoplasm that training sample is concentrated with 2 sampling frequencies
It is sampled, fritter of the Lung neoplasm more than 40mm in training sample collection sheet is sampled with 6 sampling frequencies, other sizes
Lung neoplasm with normal sample frequency sampling, the training sample after sampling is input to Lung neoplasm and divides network, with Lung neoplasm point
The error convergence for cutting the prediction output of network and really exporting is target, is trained to Lung neoplasm segmentation network, obtains lung knot
Save segmenting device.
The present invention is improved on the basis of Mask-RCNN neural networks, establishes Lung neoplasm segmentation network, the grid energy
Enough Lung neoplasm features galore extracted in training sample, and then the prediction accuracy to Lung neoplasm can be promoted.
Wherein, described the three-dimensional lung CT image of acquisition is cut into a large amount of cube fritter to include:
Three-dimensional lung CT image is cut according to the following conditions:
Condition one:A Lung neoplasm target is included at least in 70% cube fritter;
Condition two:It is randomly selected from entire lung in 30% cube fritter;
If the region that cube fritter includes has been more than lung, non-lung region is filled with meaningless values 170 in CT images
Domain;
Using the pixel in Lung neoplasm region as positive sample, the pixel in other regions is as negative sample.
Wherein, divide negative sample method for digging using difficulty treated that cube fritter is handled to enhancing, be most difficult to
2n negative sample of differentiation and n positive sample composition training sample set include:
First, network is divided using Lung neoplasm treated that cube fritter calculates to enhancing, export each pixel
Classification confidence, classification confidence indicates that the probability containing Lung neoplasm is higher in cube fritter, confidence of classifying closer to 1
Degree indicates that the probability without containing Lung neoplasm is higher in cube fritter closer to 0;
Then, the absolute value of the difference of classification confidence and true value label is calculated, the absolute value is bigger, indicates that the negative sample is got over
Difficulty is distinguished by network;
Finally, selection is most difficult to 2n negative sample being distinguished and n positive sample composition training sample set.
Wherein, dense piece of layer of the 32*32*32 includes sequentially connected 4 feature extraction groups, each feature extraction group packet
Include BN layers sequentially connected, RELU functions, CONV layers, BN layers, RELU functions, CONV layers, the output of each feature extraction group removes
Outside input as feature extraction group adjacent thereto, it is also connected to the output of all feature extraction groups thereafter, makes last
All outputs of all feature extraction groups have been merged in the output of one feature extraction group.
Dense piece of layer of the 16*16*16 includes sequentially connected 4 feature extraction groups, each feature extraction group include according to
The BN layers of secondary connection, RELU functions, CONV layers, BN layers, RELU functions, CONV layers, the output of each feature extraction group is in addition to making
Outside input for feature extraction group adjacent thereto, it is also connected to the output of all feature extraction groups thereafter, makes the last one
All outputs of all feature extraction groups have been merged in the output of feature extraction group.
Wherein, the loss function of the RPN networks is cross entropy loss function.The condition of convergence of cross entropy loss function
For:The average value of the cross entropy loss function C of continuous 3 epoch is below the loss function value of previous epoch.
Compared with prior art, the device have the advantages that being:
The device that this method is established can more accurately capture the brief summary locking nub that classical fast-RCNN is easy to miss,
Improve the precise degrees of detection;The introducing of completely new frame makes network speed on the particular problem faster, can more meet doctor
Raw actual demand;The three-dimensional Lung neoplasm image being partitioned into is more typical.
Description of the drawings
Fig. 1 is the flow chart of the method for building up for the Lung neoplasm segmenting device that embodiment provides;
Fig. 2 is the structural schematic diagram for the Lung neoplasm segmentation network that embodiment provides;
Fig. 3 is the structural schematic diagram for the dense piece of layer of 32*32*32 that embodiment provides.
Fig. 4 is the structural schematic diagram of each feature extraction group in Fig. 3.
Specific implementation mode
To make the objectives, technical solutions, and advantages of the present invention more comprehensible, with reference to the accompanying drawings and embodiments to this
Invention is described in further detail.It should be appreciated that the specific embodiments described herein are only used to explain the present invention,
Do not limit protection scope of the present invention.
Fig. 1 is the flow chart of the method for building up for the Lung neoplasm segmenting device that embodiment provides.As shown in Figure 1, the present embodiment
The method for building up of the Lung neoplasm segmenting device of offer includes the following steps:
S101 establishes training sample.
Under normal circumstances, using whole image as the input of object detection model.But for 3 dimension CT images, due to
CT images are excessively huge, and existing GPU video memorys capacity cannot be satisfied this demand, cannot be by CT images directly as object detection
The input of model.To ensure that good resolution ratio, the present embodiment can not optionally compress CT, in order to avoid lose many weights
Information is wanted, testing result is unfavorable for.Therefore, the three-dimensional lung CT image of acquisition is cut into a large amount of cube fritter.It will stand
Input of the cube fritter as model.Specifically, 128 × 128 × 128 are cut to three-dimensional lung CT image according to the following conditions
The cube fritter of (pixel):
Condition one:A Lung neoplasm target is included at least in 70% cube fritter;
Condition two:It is randomly selected from entire lung in 30% cube fritter;
If the region that cube fritter includes has been more than lung, non-lung region is filled with meaningless values 170 in CT images
Domain;
Using the pixel in Lung neoplasm region as positive sample, the pixel in other regions is as negative sample.
It should be noted that Lung neoplasm might not be in the center of cube fritter, simply by the presence of in cube fritter
In, the pixel in the Lung neoplasm region is positive sample.
After cutting, cube fritter is handled using data enhancement methods, it is gentle with the robustness for increasing training sample
Solve overfitting problem.
Since the number of negative sample is far more than positive sample, and distribution height is uneven, so there is most numerical example very
It is easy to be distinguished, but some of which possesses the negative sample of the doubtful tubercle of similar appearance.The present embodiment uses difficult point
Sample method for digging solves the problems, such as this.Detailed process is:
First, network is divided using Lung neoplasm proposed by the present invention treated that cube fritter calculates to enhancing,
The classification confidence of each pixel is exported, classification confidence indicates that the probability containing Lung neoplasm is got in cube fritter closer to 1
Height, classification confidence indicate that the probability without containing Lung neoplasm is higher in cube fritter closer to 0;
Then, the absolute value of the difference of classification confidence and true value label is calculated, the absolute value is bigger, indicates that the negative sample is got over
Difficulty is distinguished by network;
Finally, selection is most difficult to 2n negative sample being distinguished and n positive sample composition training sample set.
S102 establishes Lung neoplasm segmentation network.
As shown in Fig. 2, Lung neoplasm segmentation network includes the maximum pond layer 202 of sequentially connected input layer 201, first, 64*
The maximum pond layer 204 of 64*64 convolutional layers 203, second, dense piece of layer 205 of 32*32*32, third maximum pond layer 206,16*16*
16 dense pieces of layers 207, the 16*16*16 characteristic patterns 210 that dense piece of layer 207 of 16*16*16 exports are by up-sampling and 32*32*32
The 32*32*32 characteristic patterns 211 that dense piece of layer 205 exports carry out Fusion Features, the 32*32*32 characteristic patterns warp after Fusion Features
After the ponds POL layer 208, RPN networks 209 are input to, realizes and the Lung neoplasm of 32*32*32 characteristic patterns is predicted and divided.
First the 202, second maximum pond of maximum pond layer layer 204 and third maximum pond layer 206 are used to input
Figure carries out dimension-reduction treatment.Dense piece of dense piece of layer 205 of 64*64*64 convolutional layers 203,32*32*32,16*16*16 layer 207 are used
In to input figure progress feature extraction.The ponds POL layer 208 is after the Lung neoplasm region of 201 input figure of input layer corresponds to fusion
In characteristic pattern at corresponding characteristic point, the space characteristics and high-rise semantic feature that processing can be in conjunction with low layer in this way make final
Obtained result is more accurate reliable.
As shown in figure 3, the concrete structure of dense piece of layer 205 of 32*32*32 include sequentially connected feature extraction group A, B, C,
D, each feature extraction group as shown in figure 4, include sequentially connected BN layers 401, RELU functions 402, CONV layers 403, BN layers
404, RELU functions 405, CONV layers 406, the output of each feature extraction group is in addition to as feature extraction group adjacent thereto
Input is outer, is also connected to the output of all feature extraction groups thereafter, the output of the last one feature extraction group is made to have merged institute
There are all outputs of feature extraction group.
For feature extraction group A, the output of feature extraction group A is also connected to other than the input as feature extraction group B
The output of feature extraction group B, C, D, as shown in solid in Fig. 3.For feature extraction group B, the output of feature extraction group B in addition to
Outside input as feature extraction group C, it is also connected to the output of feature extraction group C, D, as shown by the dotted line in fig. 3.For spy
Extraction group C is levied, the output of feature extraction group C is also connected to feature extraction group D's other than the input as feature extraction group D
Output, as shown in the chain-dotted line in Fig. 3.
The structure of dense piece of layer 205 of the structure of dense piece of layer 207 of 16*16*16 and 32*32*32 is identical.
In the present embodiment, RELU functions are as activation primitive, specially f (x)=max (0, x).
Select cross entropy loss function C as the loss function of RPN networks, cross entropy loss function C is specially:
Wherein, y is desired output, that is, true value label, and a is the reality output of network, a=σ (z), z=∑s Wj×Xj
+ b, WjIt is network parameter with b.
S103, training Lung neoplasm divide network, obtain Lung neoplasm segmenting device.
Training sample is concentrated, although eliminating very small tubercle mesh in the training sample that the training sample is concentrated
Mark, but still there are nodule size and its unbalanced problem, the number of lesser tubercle is far more than major tubercle.If using uniformly adopting
Sample, the indoctrination session of network are more inclined to the detection of lesser tubercle, at the same time sacrifice the accuracy of major tubercle detection, this is and this hair
Bright goal of the invention is runed counter to.Therefore in order to solve this problem, in the present embodiment, major tubercle is increased to training sample concentration and is adopted
Sample frequency.It is specific as follows:
Fritter of the Lung neoplasm more than 30mm that training sample is concentrated is sampled with 2 sampling frequencies, with 6 samplings frequency
Rate samples fritter of the Lung neoplasm more than 40mm in training sample collection sheet, and the Lung neoplasm of other sizes is with normal sample frequency
Training sample after sampling is input to Lung neoplasm and divides network by sampling, is exported with the prediction of Lung neoplasm segmentation network and true
The error convergence of output is target, is trained to Lung neoplasm segmentation network, obtains Lung neoplasm segmenting device.
During training, the gradient of network parameter is sought using chain rule.Obtain derivative of the error to each parameter
To get to current Grad when value.According to gradient descent algorithm, parameter vector subtracts multiplying for gradient vector and learning rate
Product is primary parameter iterative process.It is below with the average value of the cross entropy loss function C of continuous 3 epoch previous
The loss function value of epoch is convergence target, by multiple parameter iteration, obtains final parameter vector to get to Lung neoplasm
Segmenting device.
After obtaining Lung neoplasm segmenting device, to sample to be tested (lung CT figure to be measured) according to the content described in S101 into
After row pretreatment, be input to Lung neoplasm segmenting device, be computed obtain the Lung neoplasm probability of sample to be tested, coordinate, diameter and
The three-dimensional segmentation result of sample to be tested.
Table 1 is the influence comparison that dense piece of layer of different numbers divides entire Lung neoplasm on network.
Table 1
It can be obtained from this table 1, dense piece of layer is more, and finally obtained Dice coefficients are bigger (result is better), and convergence needs
Number of iterations it is also fewer.But if dense piece of layer is excessive, over-fitting can be caused so that result is declined, and convergence needs
Iteration number become it is more.
Technical scheme of the present invention and advantageous effect is described in detail in above-described specific implementation mode, Ying Li
Solution is not intended to restrict the invention the foregoing is merely presently most preferred embodiment of the invention, all principle models in the present invention
Interior done any modification, supplementary, and equivalent replacement etc. are enclosed, should all be included in the protection scope of the present invention.
Claims (7)
1. a kind of method for building up of the Lung neoplasm segmenting device based on Mask-RCNN neural networks, which is characterized in that the foundation
Method includes:
Establish training sample:First, the three-dimensional lung CT image of acquisition is cut into a large amount of cube fritter, then, used
Data enhancement methods handle cube fritter, and finally, dividing negative sample method for digging using difficulty, treated that cube is small to enhancing
Block is handled, and obtains most indistinguishable 2n negative sample and n positive sample forms training sample set;
Establish Lung neoplasm segmentation network:Lung neoplasm segmentation network include sequentially connected input layer, the first maximum pond layer,
64*64*64 convolutional layers, the second maximum pond layer, dense piece of layer of 32*32*32, third maximum pond layer, dense piece of 16*16*16
Layer, the 16*16*16 characteristic patterns of dense piece of layer outputs of 16*16*16 are by up-sampling the 32* with the output of dense piece of layer of 32*32*32
32*32 characteristic patterns carry out Fusion Features, and the 32*32*32 characteristic patterns after Fusion Features are input to RPN nets after the layer of the ponds POL
Network is realized and the Lung neoplasm of 32*32*32 characteristic patterns is predicted and divided;
Training Lung neoplasm divides network:Fritter of the Lung neoplasm more than 30mm that training sample is concentrated is carried out with 2 sampling frequencies
Sampling samples fritter of the Lung neoplasm more than 40mm in training sample collection sheet with 6 sampling frequencies, the lung knot of other sizes
Section is input to Lung neoplasm with normal sample frequency sampling, by the training sample after sampling and divides network, divides network with Lung neoplasm
Prediction output and the error convergence that really exports be target, Lung neoplasm segmentation network is trained, Lung neoplasm segmentation is obtained
Device.
2. the method for building up of the Lung neoplasm segmenting device based on Mask-RCNN neural networks as described in claim 1, feature
It is, it is described the three-dimensional lung CT image of acquisition is cut into a large amount of cube fritter to include:
Three-dimensional lung CT image is cut according to the following conditions:
Condition one:A Lung neoplasm target is included at least in 70% cube fritter;
Condition two:It is randomly selected from entire lung in 30% cube fritter;
If the region that cube fritter includes has been more than lung, non-lung areas is filled with meaningless values 170 in CT images;
Using the pixel in Lung neoplasm region as positive sample, the pixel in other regions is as negative sample.
3. the method for building up of the Lung neoplasm segmenting device based on Mask-RCNN neural networks as claimed in claim 2, feature
It is, dividing negative sample method for digging using difficulty, treated that cube fritter is handled to enhancing, obtains most indistinguishable 2n
A negative sample and n positive sample composition training sample set include:
First, network is divided using Lung neoplasm treated that cube fritter calculates to enhancing, export point of each pixel
Class confidence level, classification confidence indicate that the probability containing Lung neoplasm is higher in cube fritter, classification confidence is got over closer to 1
Close to 0, indicate that the probability without containing Lung neoplasm is higher in cube fritter;
Then, calculate classification confidence and true value label absolute value of the difference, the absolute value is bigger, indicate the negative sample be more difficult to by
Network is distinguished;
Finally, selection is most difficult to 2n negative sample being distinguished and n positive sample composition training sample set.
4. the method for building up of the Lung neoplasm segmenting device based on Mask-RCNN neural networks as described in claim 1, feature
It is, dense piece of layer of the 32*32*32 includes sequentially connected 4 feature extraction groups, and each feature extraction group includes connecting successively
The BN layers that connect, RELU functions, CONV layers, BN layers, RELU functions, CONV layers, the output of each feature extraction group in addition to as with
Outside the input of its adjacent feature extraction group, it is also connected to the output of all feature extraction groups thereafter, makes the last one feature
All outputs of all feature extraction groups have been merged in the output of extraction group.
5. the method for building up of the Lung neoplasm segmenting device based on Mask-RCNN neural networks as described in claim 1, feature
It is, dense piece of layer of the 16*16*16 includes sequentially connected 4 feature extraction groups, and each feature extraction group includes connecting successively
The BN layers that connect, RELU functions, CONV layers, BN layers, RELU functions, CONV layers, the output of each feature extraction group in addition to as with
Outside the input of its adjacent feature extraction group, it is also connected to the output of all feature extraction groups thereafter, makes the last one feature
All outputs of all feature extraction groups have been merged in the output of extraction group.
6. the method for building up of the Lung neoplasm segmenting device based on Mask-RCNN neural networks as described in claim 1, feature
It is, the loss function of the RPN networks is cross entropy loss function.
7. the method for building up of the Lung neoplasm segmenting device based on Mask-RCNN neural networks as claimed in claim 6, feature
It is, the condition of convergence of cross entropy loss function is:Before the average value of the cross entropy loss function C of continuous 3 epoch is below
The loss function value of one epoch.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810091696.1A CN108346154B (en) | 2018-01-30 | 2018-01-30 | Method for establishing lung nodule segmentation device based on Mask-RCNN neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810091696.1A CN108346154B (en) | 2018-01-30 | 2018-01-30 | Method for establishing lung nodule segmentation device based on Mask-RCNN neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108346154A true CN108346154A (en) | 2018-07-31 |
CN108346154B CN108346154B (en) | 2021-09-07 |
Family
ID=62961021
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810091696.1A Active CN108346154B (en) | 2018-01-30 | 2018-01-30 | Method for establishing lung nodule segmentation device based on Mask-RCNN neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108346154B (en) |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109145769A (en) * | 2018-08-01 | 2019-01-04 | 辽宁工业大学 | The target detection network design method of blending image segmentation feature |
CN109191446A (en) * | 2018-08-30 | 2019-01-11 | 北京深睿博联科技有限责任公司 | Image processing method and device for Lung neoplasm segmentation |
CN109215091A (en) * | 2018-08-02 | 2019-01-15 | 浙江理工大学 | A kind of Fashionable Colors of Garment extraction method indicated based on figure |
CN109427060A (en) * | 2018-10-30 | 2019-03-05 | 腾讯科技(深圳)有限公司 | A kind of method, apparatus, terminal device and the medical system of image identification |
CN109446961A (en) * | 2018-10-19 | 2019-03-08 | 北京达佳互联信息技术有限公司 | Pose detection method, device, equipment and storage medium |
CN109493330A (en) * | 2018-11-06 | 2019-03-19 | 电子科技大学 | A kind of nucleus example dividing method based on multi-task learning |
CN109584248A (en) * | 2018-11-20 | 2019-04-05 | 西安电子科技大学 | Infrared surface object instance dividing method based on Fusion Features and dense connection network |
CN110310281A (en) * | 2019-07-10 | 2019-10-08 | 重庆邮电大学 | Lung neoplasm detection and dividing method in a kind of Virtual Medical based on Mask-RCNN deep learning |
CN110853041A (en) * | 2019-11-12 | 2020-02-28 | 东南大学 | Underwater pier component segmentation method based on deep learning and sonar imaging |
CN110956255A (en) * | 2019-11-26 | 2020-04-03 | 中国医学科学院肿瘤医院 | Difficult sample mining method and device, electronic equipment and computer readable storage medium |
CN111127431A (en) * | 2019-12-24 | 2020-05-08 | 杭州求是创新健康科技有限公司 | Dry eye disease grading evaluation system based on regional self-adaptive multitask neural network |
CN111341438A (en) * | 2020-02-25 | 2020-06-26 | 中国科学技术大学 | Image processing apparatus, electronic device, and medium |
WO2020168647A1 (en) * | 2019-02-21 | 2020-08-27 | 平安科技(深圳)有限公司 | Image recognition method and related device |
CN114037709B (en) * | 2021-11-05 | 2023-06-16 | 复旦大学附属肿瘤医院 | Method and device for segmenting ground glass lung nodules |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2009150004A1 (en) * | 2008-05-27 | 2009-12-17 | Telefonaktiebolaget L M Ericsson (Publ) | Index-based pixel block processing |
CN104408767A (en) * | 2014-11-20 | 2015-03-11 | 浙江大学 | Method for building sparse consistent three-dimensional human face mesh deformation model |
CN106875425A (en) * | 2017-01-22 | 2017-06-20 | 北京飞搜科技有限公司 | A kind of multi-target tracking system and implementation method based on deep learning |
CN107590797A (en) * | 2017-07-26 | 2018-01-16 | 浙江工业大学 | A kind of CT images pulmonary nodule detection method based on three-dimensional residual error neutral net |
-
2018
- 2018-01-30 CN CN201810091696.1A patent/CN108346154B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2009150004A1 (en) * | 2008-05-27 | 2009-12-17 | Telefonaktiebolaget L M Ericsson (Publ) | Index-based pixel block processing |
CN104408767A (en) * | 2014-11-20 | 2015-03-11 | 浙江大学 | Method for building sparse consistent three-dimensional human face mesh deformation model |
CN106875425A (en) * | 2017-01-22 | 2017-06-20 | 北京飞搜科技有限公司 | A kind of multi-target tracking system and implementation method based on deep learning |
CN107590797A (en) * | 2017-07-26 | 2018-01-16 | 浙江工业大学 | A kind of CT images pulmonary nodule detection method based on three-dimensional residual error neutral net |
Non-Patent Citations (1)
Title |
---|
LI, HX 等: "Learning Deep Appearance Feature for Multi-target Tracking", 《2017 INTERNATIONAL CONFERENCE ON VIRTUAL REALITY AND VISUALIZATION (ICVRV 2017)》 * |
Cited By (22)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109145769A (en) * | 2018-08-01 | 2019-01-04 | 辽宁工业大学 | The target detection network design method of blending image segmentation feature |
CN109215091A (en) * | 2018-08-02 | 2019-01-15 | 浙江理工大学 | A kind of Fashionable Colors of Garment extraction method indicated based on figure |
CN109215091B (en) * | 2018-08-02 | 2021-09-14 | 浙江理工大学 | Clothing fashion color automatic extraction method based on graph representation |
CN109191446B (en) * | 2018-08-30 | 2020-12-29 | 杭州深睿博联科技有限公司 | Image processing method and device for lung nodule segmentation |
CN109191446A (en) * | 2018-08-30 | 2019-01-11 | 北京深睿博联科技有限责任公司 | Image processing method and device for Lung neoplasm segmentation |
CN109446961A (en) * | 2018-10-19 | 2019-03-08 | 北京达佳互联信息技术有限公司 | Pose detection method, device, equipment and storage medium |
US11138422B2 (en) | 2018-10-19 | 2021-10-05 | Beijing Dajia Internet Information Technology Co., Ltd. | Posture detection method, apparatus and device, and storage medium |
CN109427060A (en) * | 2018-10-30 | 2019-03-05 | 腾讯科技(深圳)有限公司 | A kind of method, apparatus, terminal device and the medical system of image identification |
US11610310B2 (en) | 2018-10-30 | 2023-03-21 | Tencent Technology (Shenzhen) Company Limited | Method, apparatus, system, and storage medium for recognizing medical image |
US11410306B2 (en) | 2018-10-30 | 2022-08-09 | Tencent Technology (Shenzhen) Company Limited | Method, apparatus, system, and storage medium for recognizing medical image |
CN109493330A (en) * | 2018-11-06 | 2019-03-19 | 电子科技大学 | A kind of nucleus example dividing method based on multi-task learning |
CN109584248A (en) * | 2018-11-20 | 2019-04-05 | 西安电子科技大学 | Infrared surface object instance dividing method based on Fusion Features and dense connection network |
CN109584248B (en) * | 2018-11-20 | 2023-09-08 | 西安电子科技大学 | Infrared target instance segmentation method based on feature fusion and dense connection network |
WO2020168647A1 (en) * | 2019-02-21 | 2020-08-27 | 平安科技(深圳)有限公司 | Image recognition method and related device |
CN110310281A (en) * | 2019-07-10 | 2019-10-08 | 重庆邮电大学 | Lung neoplasm detection and dividing method in a kind of Virtual Medical based on Mask-RCNN deep learning |
CN110310281B (en) * | 2019-07-10 | 2023-03-03 | 重庆邮电大学 | Mask-RCNN deep learning-based pulmonary nodule detection and segmentation method in virtual medical treatment |
CN110853041A (en) * | 2019-11-12 | 2020-02-28 | 东南大学 | Underwater pier component segmentation method based on deep learning and sonar imaging |
CN110956255A (en) * | 2019-11-26 | 2020-04-03 | 中国医学科学院肿瘤医院 | Difficult sample mining method and device, electronic equipment and computer readable storage medium |
CN110956255B (en) * | 2019-11-26 | 2023-04-07 | 中国医学科学院肿瘤医院 | Difficult sample mining method and device, electronic equipment and computer readable storage medium |
CN111127431A (en) * | 2019-12-24 | 2020-05-08 | 杭州求是创新健康科技有限公司 | Dry eye disease grading evaluation system based on regional self-adaptive multitask neural network |
CN111341438A (en) * | 2020-02-25 | 2020-06-26 | 中国科学技术大学 | Image processing apparatus, electronic device, and medium |
CN114037709B (en) * | 2021-11-05 | 2023-06-16 | 复旦大学附属肿瘤医院 | Method and device for segmenting ground glass lung nodules |
Also Published As
Publication number | Publication date |
---|---|
CN108346154B (en) | 2021-09-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108346154A (en) | The method for building up of Lung neoplasm segmenting device based on Mask-RCNN neural networks | |
CN108257128A (en) | A kind of method for building up of the Lung neoplasm detection device based on 3D convolutional neural networks | |
CN109447065A (en) | A kind of method and device of breast image identification | |
CN109934200A (en) | A kind of RGB color remote sensing images cloud detection method of optic and system based on improvement M-Net | |
CN108109160A (en) | It is a kind of that interactive GrabCut tongue bodies dividing method is exempted from based on deep learning | |
CN108846826A (en) | Object detecting method, device, image processing equipment and storage medium | |
CN108717568A (en) | A kind of image characteristics extraction and training method based on Three dimensional convolution neural network | |
CN107563381A (en) | The object detection method of multiple features fusion based on full convolutional network | |
CN109363698A (en) | A kind of method and device of breast image sign identification | |
CN108830209A (en) | Based on the remote sensing images method for extracting roads for generating confrontation network | |
CN108446616B (en) | Road extraction method based on full convolution neural network ensemble learning | |
CN110390673A (en) | Cigarette automatic testing method based on deep learning under a kind of monitoring scene | |
CN108416775A (en) | A kind of ore grain size detection method based on deep learning | |
CN109363697A (en) | A kind of method and device of breast image lesion identification | |
CN109145971A (en) | Based on the single sample learning method for improving matching network model | |
CN110490188A (en) | A kind of target object rapid detection method based on SSD network improvement type | |
CN110084181A (en) | A kind of remote sensing images Ship Target Detection method based on sparse MobileNetV2 network | |
CN111583179B (en) | Lung nodule deep learning classification method based on floating cutting | |
CN108665546A (en) | A kind of multiple spot geological statistics three-dimensional modeling method of combination deep learning | |
CN105913451B (en) | A kind of natural image superpixel segmentation method based on graph model | |
Fan et al. | Rockfill material segmentation and gradation calculation based on deep learning | |
Zu et al. | Detection of common foreign objects on power grid lines based on Faster R-CNN algorithm and data augmentation method | |
Li et al. | Underwater Target Detection Based on Improved YOLOv4 | |
CN116467159A (en) | Image convolutional neural network model safety assessment method based on main modal neuron coverage | |
CN110276358A (en) | High similarity wooden unit cross section detection method under intensive stacking |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |