CN110517253A - The method of the good pernicious classification of Lung neoplasm based on 3D multiple target feature learning - Google Patents
The method of the good pernicious classification of Lung neoplasm based on 3D multiple target feature learning Download PDFInfo
- Publication number
- CN110517253A CN110517253A CN201910807278.2A CN201910807278A CN110517253A CN 110517253 A CN110517253 A CN 110517253A CN 201910807278 A CN201910807278 A CN 201910807278A CN 110517253 A CN110517253 A CN 110517253A
- Authority
- CN
- China
- Prior art keywords
- feature
- lung neoplasm
- detection
- good pernicious
- classification
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- 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
- G06T2207/20016—Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
-
- 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
- G06T2207/30061—Lung
- G06T2207/30064—Lung nodule
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of method of good pernicious classification of the Lung neoplasm based on 3D multiple target feature learning, it is low with nicety of grading mainly to solve the problems, such as that the prior art detects medical image Lung neoplasm.First with the novel feature extraction network model of SE-ResNeXt and feature pyramid system integrating.Next establishes nodule detection subnet and pernicious classification subnet and by the Shuangzi network in conjunction with Feature Selection Model, obtains the good pernicious detection of novel Lung neoplasm and disaggregated model.Finally using training set and verifying collection, according to loss function, training pattern parameter obtains the final form of novel Lung neoplasm good pernicious detection and disaggregated model, so that the test data set of no label is detected and be classified.The new feature extraction and detection network that the present invention constructs have more simplified network structure, faster training speed, higher nodule detection and nicety of grading, can be used for computer-aided medical diagnosis system compared with traditional CNN network.
Description
Technical field
The invention belongs to technical field of medical image processing, relate generally to 3D feature pyramid network and SE-ResNeXt net
The integrated new feature extraction of network and detection sorter network, are a kind of good pernicious point of the Lung neoplasm based on 3D multiple target feature learning
The method of class, can be used for computer-aided diagnosis system.
Background technique
Show that it is lethal that lung cancer has become the whole world according to the investigation of major Cancer Research Center and health organization in the world
The big cancer of rate highest first.Since lung cancer often exists in the form of Lung neoplasm in early days, detecting Lung neoplasm early can be
It largely helps to find the early stage of lung cancer, and thoracic cavity CT scan image provides possibility for the realization of this target.Radiation doctor
Teacher and clinician can intuitively be found by the visual observation to thoracic cavity CT image and diagnosing.CT technology is to calculate
Tens lung scannings can be obtained by carrying out laterally closely-spaced full lung tomoscan to tester in machine layer scanning technology
CT image.However as the rapid development of CT scan technology, imaging resolution is higher and higher, can be found in image after reconstruction
Tubercle volume it is smaller and smaller, the data volume of image increases sharply, and is human visual observation to find that the mode of Lung neoplasm mentions
Challenge is gone out.With gradualling mature for image processing and artificial intelligence technology, the area of computer aided based on convolutional neural networks is examined
Disconnected technology is come into being, it is desirable to provide Lung neoplasm region, feature extraction, classification and discriminant information in CT images, auxiliary
Doctor makes accurate diagnosis to Pulmonary Disease patients in time.
Existing lung cancer computer-aided diagnosis main flow includes: that the feature extraction of Lung neoplasm and the classification of Lung neoplasm are known
Not.In terms of feature extraction, manual features are carried out to the pathological characters of the morphological feature of tubercle, local features etc.
It extracts.Manual features are extracted time-consuming and are required the professional technique for extracting feature personnel high.In terms of the Classification and Identification of Lung neoplasm,
Convolutional neural networks technology tends to be mature, is gradually applied to field of medical image processing.But based on traditional convolutional neural networks
Structure realizes the pulmonary nodule detection of various dimensions and classification is all by the different dimensions after reducing or expanding all the time
Tubercle picture generates the combination of the feature of reflection different dimensions information as input.Although this method can effectively indicate tubercle
The various dimensional characteristics of picture, but need that input processing is multiple after changing dimension to same tubercle picture, hardware is calculated
Ability and memory size have higher requirements, it will usually as the performance bottleneck of entire algorithm, therefore can only be in limited field
Portion uses.
Summary of the invention
As shown in Figure 1, being proposed a kind of based on 3D multiple target it is an object of the invention to be directed to the deficiency of above-mentioned existing method
The method of the good pernicious classification of the Lung neoplasm of feature learning, this method is using existing popular SE-ResNeXt as core network, In
Under the premise of not increasing complicated network structure degree, solve traditional convolutional neural networks information transmitting when information lose, loss with
And the problems such as gradient disappearance or gradient explosion.And introduced feature pyramid structure, by the high level of low resolution, high semantic information
The low-level feature of feature and high-resolution, low semantic information is merged, and forms the tubercle feature of richer characterization power, solves biography
System CNN middle-shallow layer network but cannot be used for classifying more focused on detailed information, high-rise network more focused on semantic information but
The problem of accurate location information cannot be provided, optimizes performance for nodule detection and the multitask system for disliking benign classification.
As shown in Fig. 2, implementer's case design of the invention is as follows:
One, technical principle
1.SE-ResNeXt core network includes ResNeXt and SE two parts.ResNeXt is changing for ResNet network
Progressive die type, while using the Split-Transform-Merge thought of the stacking thought of VGG and Inception, it can not increase
Model accuracy rate is improved under the premise of parameter complexity.ResNeXt model uses the polymeric block of the identical topological structure of stacked in parallel
Instead of ResNet three-layer coil product structure, solve traditional CNN or fully-connected network information transmitting when information lose, loss with
And the problems such as gradient disappearance or gradient explosion.
2.SE module uses completely new feature recalibration strategy, by explicitly modeling the correlation between convolutional layer feature channel
Property carrys out the characterization ability of lift scheme, and the significance level in each feature channel is obtained in a manner of adaptive learning, with this
The weight for enhancing feature that can be information-based of selectivity, reduces the weight of useless feature, finally constructs more strong
E-learning characteristics of image.
3. feature pyramid model improves the feature extraction mode of CNN network for the height of low resolution, high semantic information
Layer feature carries out top-down side with the low-level feature of high-resolution, low semantic information and connects, so that under all scales
Feature has semantic information abundant, can preferably indicate the information of the input each dimension of picture.Using in conventional CNN model
Portion can give birth to from the supreme feature representation structure to generate same size picture different dimensions in bottom effectively under single picture view
At the method for various dimensions feature representation.To generate the stronger characteristic pattern of ability to express as next stage target detection or classification
The input of analysis task.To reinforce the feature representation of core network CNN.
Two, based on the above principles, technical solution of the present invention includes the following:
S1 obtains the known Lung neoplasm for having label according to the raw data set of Lung neoplasm public data collection LIDC-IDRI
The Lung neoplasm image in region, pre-processes such image, obtains the training set, verifying collection and test set of nodule image；
S2, foundation are integrated with the feature extraction network model of SE-ResNeXt and feature pyramid network；
S3 establishes nodule detection subnet and pernicious classification subnet；
S4 obtains novel lung nodule detection and classification mould by Feature Selection Model in conjunction with nodule detection classification Shuangzi net
Type；
Training set and verifying collection are inputted the novel good pernicious detection of Lung neoplasm and disaggregated model by S5, according to loss function,
Training pattern parameter obtains the final form of novel Lung neoplasm good pernicious detection and disaggregated model；
S6, using novel Lung neoplasm detection and good pernicious disaggregated model to the test data set of no label carry out detection with
Classification, obtain the Lung neoplasm in every medical image detection and good pernicious classification results.
The present invention has the advantage that
(1) the ResNeXt structure for using core network, can improve the entirety of model under conditions of not increasing network depth
Precision, while the loss of traditional CNN or the transmitting of fully-connected network information are avoided, loss and gradient disappearance or gradient are exploded
The problems such as.
(2) the SE module in core network, by the correlation between modeling convolution feature channel come the table of lift scheme
Sign ability obtains the rewards and punishments strategy of the significance level construction feature in each feature channel, finally in a manner of adaptive learning
Form more strong e-learning characteristics of image.
(3) feature pyramid model improves the feature extraction mode of CNN network, by low resolution, high semantic information
High-level characteristic carries out top-down side with the low-level feature of high-resolution, low semantic information and connects, so that under all scales
Feature have semantic information abundant, can preferably indicate input each dimension of picture information.To reinforce backbone network
The feature representation of network CNN.
(4) multitask feature learning method is imitated using nodule detection subnet and pernicious classification subnet, is reduced in false positive
(FPR) and in the task dependencies of pernicious classification useful information is utilized, the Generalization Capability of two tasks is improved.
Detailed description of the invention
Fig. 1 is the flow chart of the good pernicious classification method of Lung neoplasm according to the present invention based on 3D multiple target feature learning.
Fig. 2 is the grid frame of the good pernicious classification method of Lung neoplasm according to the present invention based on 3D multiple target feature learning
Frame figure.
Specific embodiment
The present invention is further explained in the light of specific embodiments.
S1 obtains the known Lung neoplasm for having label according to the raw data set of Lung neoplasm public data collection LIDC-IDRI
The Lung neoplasm image in region, pre-processes such image, obtains the training set, verifying collection and test set of nodule image,
Its feature the step S1 in include following three step by step:
S11 obtains the Lung neoplasm image with label of known Lung neoplasm ROI, to have the Lung neoplasm image of label into
Row tubercle ROI segmentation, cuts, the Lung neoplasm image after being cut；
S12 is divided the Lung neoplasm image after cutting to obtain training dataset, validation data set and test data
Collection；
S13 rotates the Lung neoplasm image in each data set tentatively obtained, obtains the four of each Lung neoplasm image
The vector image in a direction finally obtains pretreated training dataset, validation data set and test data set.
S2, foundation are integrated with the feature extraction network model of SE-ResNeXt and feature pyramid network, it is characterised in that
The step S2 include following two step by step:
S21 constructs the access SE-ResNeXt generated from the supreme different dimensions feature in bottom；
ResNeXt core network is realized from the supreme layer-by-layer concentration expression characteristic in bottom, and hyper parameter cardinality=8 is arranged,
Convolution kernel is pressed into channel packet, forms 8 parallel branch, each branch is made of alternate CONV-BN-RELU block；
The output connection SE module input of ResNeXt block, while the integrality in information exchanging process is kept, it will
The output end of the input terminal jump connection SE module of ResNeXt, the more strong e-learning characteristics of image of final output, SE mould
Block is by full articulamentum (the FC)-Sigmoid activation primitive of global average pond layer (GAP)-full articulamentum (FC)-RELU-
Alternating structure is constituted；
Each tensor X ∈ R of outputW'×H'×C'Pass through convolution operation FtrNew tensor U ∈ R is obtained afterwardsW×H×C, such as formula
(1) shown in, wherein vcRefer to c-th of filter parameter in learn one group of filter kernel；
The fixed channel dimension of initial characteristics U is constant, for the characteristic pattern in each channel, carried out by formula (2) primary complete
The average pond of office, i.e., generate statistical information z ∈ R by shrinking U on Spatial Dimension W × HC, wherein c-th of element of z is by public affairs
Formula (2) is calculated, and 1 × 1 × C descriptor based on channel is obtained；
1 × 1 × C descriptor in each characteristic pattern channel that upper step obtains is passed through into FexIt is set as the weight in each channel,
Wherein δ refers to ReLU function,
S=Fex(z, W)=σ (g (z, W))=σ (W2δ(W1z)) (3)
It is regenerated based on UFormula such as (4), whereinAnd Fscale(uc,sc) respectively refer to feature
Map uc∈RW×HWith scalar scBetween corresponding channel product；
So far, the bottom of from, supreme SE-ResNeXt network struction is completed, by SE-ResNeXt core network each stage
The feature of the last one residual error structure output is madeFor high-resolution speak in a low voice justice output feature, as next stage step S22
The input of feature pyramid network.
S22, the supplement of the feature from top to bottom enhancing access and lateral connection of construction feature pyramid network；
Firstly, stronger high-level characteristic figure semantic in feature pyramid network is done 2 times of closest up-samplings, and by SE-
Each layer of feature of ResNeXt model does 1 × 1 convolution dimension-reduction treatment, guarantees two layers of feature of lateral connection in bulk
It is upper identical.Secondly, the feature after step process is done the addition between pixel.The above process that changes is repeated, until generating most fine
Characteristic pattern.Finally, the convolution operation that 3 × 3 are carried out to fused feature, eliminates the aliasing effect of up-sampling, generates final special
The mapping ensemblen of sign；
S3 establishes nodule detection subnet and pernicious classification subnet, it is characterised in that the step S3 includes following three point
Step:
S31 takes structure similar with trunk SE-ResNeXt, enables cardinality=4, by convolution kernel by channel point
For 4 parallel branch；
S32 replaces layer FC that be fully connected in parallel organization using global average pond layer GAP, and end connects the close of 256d
Collect layer as final subnet layer；
S33 merges the output of the last one structure of among nodule detection and end layer, and being fed to has
The close layer of final 2d of sigmoid activation, for predicting the probability of the true and false positive of tubercle；
S4 obtains novel lung nodule detection and classification mould by Feature Selection Model in conjunction with nodule detection classification Shuangzi net
Type；
Training set and verifying collection are inputted the novel good pernicious detection of Lung neoplasm and disaggregated model by S5, according to loss function,
Training pattern parameter obtains the final form of novel Lung neoplasm good pernicious detection and disaggregated model；
Training set and verifying collection are inputted into the novel good pernicious detection of Lung neoplasm and disaggregated model, it is such as public according to loss function
Shown in formula (5), the Lung neoplasm image training parameter of training based on this model, thus obtain the good pernicious detection of novel Lung neoplasm with
The final form of disaggregated model.
Lglobe=α Lfp(W+wfp)+βLmalig(W+wmal) (5)
Wherein Lfp(W+wfp) and Lmalig(W+wmal) it is traditional single task loss function, definition such as formula (6) is shown,
S6, using novel Lung neoplasm detection and good pernicious disaggregated model to the test data set of no label carry out detection with
Classification, obtain the Lung neoplasm in every medical image detection and good pernicious classification results.
Application examples
With the present invention with existing Multi-Scale CNN, Nodule-Level 2D CNN, Vanilla 3D CNN,
Muti-Crop CNN and Deep 3D DPN is utilized respectively trained detection model and concentrates every medicine figure to test data
As being tested, by every kind of detector all repetitive exercise 60000 times, it is flat to the detection of test data set to obtain detection disaggregated model
Equal precision, as shown in table 1.
Table 1 is based on LIDC-IDRI data set nodule detection compared with the precision of disaggregated model
Can be seen that from the experimental result of table 1 it is proposed that the Lung neoplasm based on 3D multiple target feature learning it is good pernicious
The precision of the method for classification, detection and classification is higher than existing certain methods comprehensively, can be used for computer-aided medical diagnosis
System.
Specific embodiments of the present invention are described in detail above, those skilled in the art can be public according to the present invention
The technical disclosures opened make various various other specific variations and combinations for not departing from essence of the invention, right of the invention
It is required that be intended to cover these modifications, therefore all shapes according to the present invention, change made by principle, should all cover in protection of the invention
In range.
Claims (7)
1. the method for the good pernicious classification of Lung neoplasm based on 3D multiple target feature learning, it is characterised in that the following steps are included:
S1 obtains the known Lung neoplasm region for having label according to the raw data set of Lung neoplasm public data collection LIDC-IDRI
Lung neoplasm image, such image is pre-processed, obtain nodule image training set, verifying collection and test set；
S2, foundation are integrated with the feature extraction network model of SE-ResNeXt and feature pyramid network；
S3 establishes nodule detection subnet and pernicious classification subnet；
S4 obtains novel lung nodule detection and disaggregated model by Feature Selection Model in conjunction with nodule detection classification Shuangzi net；
Training set and verifying collection are inputted the novel good pernicious detection of Lung neoplasm and disaggregated model, according to loss function, training by S5
Model parameter obtains the final form of novel Lung neoplasm good pernicious detection and disaggregated model；
S6 is detected and is divided to the test data set of no label with good pernicious disaggregated model using novel Lung neoplasm detection
Class, obtain the Lung neoplasm in every medical image detection and good pernicious classification results.
2. the method for the good pernicious classification of Lung neoplasm according to claim 1 based on 3D multiple target feature learning, feature exist
In the step S1 include following three step by step:
S11 obtains the Lung neoplasm image with label of known Lung neoplasm ROI, ties to the Lung neoplasm image with label
ROI segmentation is saved, is cut, the Lung neoplasm image after being cut；
S12 is divided the Lung neoplasm image after cutting to obtain training dataset, validation data set and test data set；
S13 rotates the Lung neoplasm image in each data set tentatively obtained, obtains four sides of each Lung neoplasm image
To vector image, finally obtain pretreated training dataset, validation data set and test data set.
3. the method for the good pernicious classification of Lung neoplasm according to claim 1 based on 3D multiple target feature learning, feature exist
In the step S2 include it is following step by step:
S21, foundation are integrated with the feature extraction network model of SE-ResNeXt and feature pyramid network, it is characterised in that described
Step S2 include following two step by step:
Firstly, the access SE-ResNeXt that building is generated from the supreme different dimensions feature in bottom；
ResNeXt core network is realized from the supreme layer-by-layer concentration expression characteristic in bottom, and hyper parameter cardinality=8 is arranged, i.e., will
Convolution kernel presses channel packet, forms 8 parallel branch, each branch is made of alternate CONV-BN-RELU block；
The output connection SE module input of ResNeXt block, while the integrality in information exchanging process is kept, by ResNeXt's
The output end of input terminal jump connection SE module, the more strong e-learning characteristics of image of final output, SE module is by the overall situation
The alternating structure of full articulamentum (the FC)-Sigmoid activation primitive of average pond layer (GAP)-full articulamentum (FC)-RELU-
It constitutes；
Each tensor X ∈ R of outputW'×H'×C'Pass through convolution operation FtrNew tensor U ∈ R is obtained afterwardsW×H×C, such as formula (1)
It is shown, wherein vcRefer to c-th of filter parameter in learn one group of filter kernel；
The fixed channel dimension of initial characteristics U is constant, for the characteristic pattern in each channel, carried out by formula (2) primary global flat
Equal pond generates statistical information z ∈ R by shrinking U on Spatial Dimension W × HC, wherein c-th of element of z presses formula (2)
It is calculated, obtains 1 × 1 × C descriptor based on channel；
1 × 1 × C descriptor in each characteristic pattern channel that upper step obtains is passed through into FexIt is set as the weight in each channel, wherein δ
Refer to ReLU function,
S=Fex(z, W)=σ (g (z, W))=σ (W2δ(W1z)) (3)
It is regenerated based on UFormula such as (4), whereinAnd Fscale(uc,sc) respectively refer to Feature Mapping
uc∈RW×HWith scalar scBetween corresponding channel product；
So far, it is completed from the supreme SE-ResNeXt network struction in bottom, it is last by SE-ResNeXt core network each stage
The feature of one residual error structure output is madeFor high-resolution speak in a low voice justice output feature, as next stage step S22 feature
The input of pyramid network.
S22, the supplement of the feature from top to bottom enhancing access and lateral connection of construction feature pyramid network；
Firstly, stronger high-level characteristic figure semantic in feature pyramid network is done 2 times of closest up-samplings, and by SE-
Each layer of feature of ResNeXt model does 1 × 1 convolution dimension-reduction treatment, guarantees two layers of feature of lateral connection in bulk
It is upper identical；
Secondly, the feature after step process is done the addition between pixel.The above process that changes is repeated, until generating most fine feature
Figure.
Finally, the convolution operation that 3 × 3 are carried out to fused feature, eliminates the aliasing effect of up-sampling, generates final feature
Mapping ensemblen.
4. the method for the good pernicious classification of Lung neoplasm according to claim 1 based on 3D multiple target feature learning, feature exist
In the step S3 include following three step by step:
S31 takes structure similar with trunk SE-ResNeXt, enables cardinality=4, and convolution kernel is divided into 4 by channel
Parallel branch；
S32 replaces layer FC that be fully connected in parallel organization using global average pond layer GAP, and end connects the dense layer of 256d
As final subnet layer；
The output of the last one structure of among nodule detection and end layer is merged, is fed to Sigmoid by S33
The close layer of final 2d of activation, for predicting the probability of the true and false positive of tubercle；
So far, nodule detection subnet is extended from backbone and each FPN (C3-C5) feature for being inferred to of up-sampling layer, and pernicious point
Class subnet and nodule detection subnet network structure having the same.
5. the method for the good pernicious classification of Lung neoplasm according to claim 1 based on 3D multiple target feature learning, feature exist
It is combined in by Feature Selection Model and nodule detection classification Shuangzi net, obtains the good pernicious detection of novel Lung neoplasm and mould of classifying
Type.
6. the method for the good pernicious classification of Lung neoplasm according to claim 1 based on 3D multiple target feature learning, feature exist
In training set and verifying collection are inputted the novel good pernicious detection of Lung neoplasm and disaggregated model, according to loss function such as formula (5)
It is shown, the Lung neoplasm image training parameter of training based on this model, to obtain the good pernicious detection of novel Lung neoplasm and classification
The final form of model；
Lglobe=α Lfp(W+wfp)+βLmalig(W+wmal) (5)
Wherein Lfp(W+wfp) and Lmalig(W+wmal) it is traditional single task loss function, definition is as shown in formula (6).
7. the method for the good pernicious classification of Lung neoplasm according to claim 1 based on 3D multiple target feature learning, feature exist
The test data set of no label is detected and classified with good pernicious disaggregated model in being detected using novel Lung neoplasm, is obtained
The detection of Lung neoplasm in every medical image and good pernicious classification results.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910807278.2A CN110517253B (en) | 2019-08-29 | 2019-08-29 | Method for classifying benign and malignant pulmonary nodules based on 3D multi-target feature learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910807278.2A CN110517253B (en) | 2019-08-29 | 2019-08-29 | Method for classifying benign and malignant pulmonary nodules based on 3D multi-target feature learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110517253A true CN110517253A (en) | 2019-11-29 |
| CN110517253B CN110517253B (en) | 2023-04-07 |
Family
ID=68629033
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910807278.2A Active CN110517253B (en) | 2019-08-29 | 2019-08-29 | Method for classifying benign and malignant pulmonary nodules based on 3D multi-target feature learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110517253B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110931112A (en) * | 2019-12-05 | 2020-03-27 | 电子科技大学中山学院 | Brain medical image analysis method based on multi-dimensional information fusion and deep learning |
| CN111339882A (en) * | 2020-02-19 | 2020-06-26 | 山东大学 | Power transmission line hidden danger detection method based on example segmentation |
| CN111815592A (en) * | 2020-06-29 | 2020-10-23 | 郑州大学 | Training method of pulmonary nodule detection model |
| CN112232433A (en) * | 2020-10-27 | 2021-01-15 | 河北工业大学 | Lung nodule benign and malignant classification method based on dual-channel network |
| CN112241948A (en) * | 2020-09-23 | 2021-01-19 | 深圳视见医疗科技有限公司 | Lung nodule detection and analysis method and system capable of adapting to thickness of layer |
| CN112308150A (en) * | 2020-11-02 | 2021-02-02 | 平安科技（深圳）有限公司 | Target detection model training method and device, computer equipment and storage medium |
| CN112614118A (en) * | 2020-12-29 | 2021-04-06 | 浙江明峰智能医疗科技有限公司 | CT image prediction method based on deep learning and computer readable storage medium |
| CN115358976A (en) * | 2022-08-10 | 2022-11-18 | 北京医准智能科技有限公司 | Image identification method, device, equipment and storage medium |
| CN116703867A (en) * | 2023-06-09 | 2023-09-05 | 太原理工大学 | Gene mutation prediction method under cooperative driving of residual network and channel attention |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109242839A (en) * | 2018-08-29 | 2019-01-18 | 上海市肺科医院 | A kind of good pernicious classification method of CT images Lung neoplasm based on new neural network model |
| CN109993733A (en) * | 2019-03-27 | 2019-07-09 | 上海宽带技术及应用工程研究中心 | Detection method, system, storage medium, terminal and the display system of pulmonary lesions |
| CN110084813A (en) * | 2019-05-13 | 2019-08-02 | 山东大学 | A kind of good pernicious prediction technique of Lung neoplasm based on three dimensional depth learning network |
-
2019
- 2019-08-29 CN CN201910807278.2A patent/CN110517253B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109242839A (en) * | 2018-08-29 | 2019-01-18 | 上海市肺科医院 | A kind of good pernicious classification method of CT images Lung neoplasm based on new neural network model |
| CN109993733A (en) * | 2019-03-27 | 2019-07-09 | 上海宽带技术及应用工程研究中心 | Detection method, system, storage medium, terminal and the display system of pulmonary lesions |
| CN110084813A (en) * | 2019-05-13 | 2019-08-02 | 山东大学 | A kind of good pernicious prediction technique of Lung neoplasm based on three dimensional depth learning network |
Non-Patent Citations (4)
| Title |
|---|
| NASRULLAH NASRULLAH等: "Automated Lung Nodule Detection and Classification Using Deep Learning Combined with Multiple Strategies", 《SENSORS》 * |
| RAUNAK DEY等: "Diagnostic classification of lung nodules using 3D neural networks", 《2018 IEEE 15TH INTERNATIONAL SYMPOSIUM ON BIOMEDICAL IMAGING (ISBI 2018)》 * |
| 刘咏江等: "基于改进的卷积神经网络的肺结节良恶性分类", 《计算机工程与设计》 * |
| 寇大磊等: "基于深度学习的目标检测框架进展研究", 《计算机工程与应用 》 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110931112A (en) * | 2019-12-05 | 2020-03-27 | 电子科技大学中山学院 | Brain medical image analysis method based on multi-dimensional information fusion and deep learning |
| CN110931112B (en) * | 2019-12-05 | 2023-04-18 | 电子科技大学中山学院 | Brain medical image analysis method based on multi-dimensional information fusion and deep learning |
| CN111339882A (en) * | 2020-02-19 | 2020-06-26 | 山东大学 | Power transmission line hidden danger detection method based on example segmentation |
| CN111339882B (en) * | 2020-02-19 | 2022-05-31 | 山东大学 | Power transmission line hidden danger detection method based on example segmentation |
| CN111815592A (en) * | 2020-06-29 | 2020-10-23 | 郑州大学 | Training method of pulmonary nodule detection model |
| CN112241948A (en) * | 2020-09-23 | 2021-01-19 | 深圳视见医疗科技有限公司 | Lung nodule detection and analysis method and system capable of adapting to thickness of layer |
| CN112232433B (en) * | 2020-10-27 | 2022-08-26 | 河北工业大学 | Lung nodule benign and malignant classification method based on dual-channel network |
| CN112232433A (en) * | 2020-10-27 | 2021-01-15 | 河北工业大学 | Lung nodule benign and malignant classification method based on dual-channel network |
| CN112308150A (en) * | 2020-11-02 | 2021-02-02 | 平安科技（深圳）有限公司 | Target detection model training method and device, computer equipment and storage medium |
| CN112308150B (en) * | 2020-11-02 | 2022-04-15 | 平安科技（深圳）有限公司 | Target detection model training method and device, computer equipment and storage medium |
| CN112614118A (en) * | 2020-12-29 | 2021-04-06 | 浙江明峰智能医疗科技有限公司 | CT image prediction method based on deep learning and computer readable storage medium |
| CN112614118B (en) * | 2020-12-29 | 2022-06-21 | 浙江明峰智能医疗科技有限公司 | CT image prediction method based on deep learning and computer readable storage medium |
| CN115358976A (en) * | 2022-08-10 | 2022-11-18 | 北京医准智能科技有限公司 | Image identification method, device, equipment and storage medium |
| CN116703867A (en) * | 2023-06-09 | 2023-09-05 | 太原理工大学 | Gene mutation prediction method under cooperative driving of residual network and channel attention |
| CN116703867B (en) * | 2023-06-09 | 2023-12-08 | 太原理工大学 | Gene mutation prediction method under cooperative driving of residual network and channel attention |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110517253B (en) | 2023-04-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110517253A (en) | The method of the good pernicious classification of Lung neoplasm based on 3D multiple target feature learning | |
| CN108446730B (en) | CT pulmonary nodule detection device based on deep learning | |
| Tseng et al. | DNetUnet: a semi-supervised CNN of medical image segmentation for super-computing AI service | |
| JP2022518446A (en) | Medical image detection methods and devices based on deep learning, electronic devices and computer programs | |
| CN108257135A (en) | The assistant diagnosis system of medical image features is understood based on deep learning method | |
| CN110120033A (en) | Based on improved U-Net neural network three-dimensional brain tumor image partition method | |
| WO2020133636A1 (en) | Method and system for intelligent envelope detection and warning in prostate surgery | |
| CN110458249A (en) | A kind of lesion categorizing system based on deep learning Yu probability image group | |
| CN112101451B (en) | Breast cancer tissue pathological type classification method based on generation of antagonism network screening image block | |
| CN107492071A (en) | Medical image processing method and equipment | |
| CN109522966A (en) | A kind of object detection method based on intensive connection convolutional neural networks | |
| CN110310281A (en) | Lung neoplasm detection and dividing method in a kind of Virtual Medical based on Mask-RCNN deep learning | |
| CN108520535A (en) | Object classification method based on depth recovery information | |
| CN107506797A (en) | One kind is based on deep neural network and multi-modal image alzheimer disease sorting technique | |
| CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
| CN110084318A (en) | A kind of image-recognizing method of combination convolutional neural networks and gradient boosted tree | |
| CN109711426A (en) | A kind of pathological picture sorter and method based on GAN and transfer learning | |
| CN114220035A (en) | Rapid pest detection method based on improved YOLO V4 | |
| CN111932529B (en) | Image classification and segmentation method, device and system | |
| CN109902736A (en) | A kind of Lung neoplasm image classification method indicated based on autocoder construction feature | |
| CN109558912A (en) | A kind of Alzheimer's disease classification method separating convolution based on depth | |
| CN109948740A (en) | A kind of classification method based on tranquillization state brain image | |
| CN114693933A (en) | Medical image segmentation device based on generation of confrontation network and multi-scale feature fusion | |
| CN115731178A (en) | Cross-modal unsupervised domain self-adaptive medical image segmentation method | |
| CN110288028A (en) | ECG detecting method, system, equipment and computer readable storage medium |
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 |