CN110084237A - Detection model construction method, detection method and the device of Lung neoplasm - Google Patents
Detection model construction method, detection method and the device of Lung neoplasm Download PDFInfo
- Publication number
- CN110084237A CN110084237A CN201910384871.0A CN201910384871A CN110084237A CN 110084237 A CN110084237 A CN 110084237A CN 201910384871 A CN201910384871 A CN 201910384871A CN 110084237 A CN110084237 A CN 110084237A
- Authority
- CN
- China
- Prior art keywords
- image
- lung
- lung neoplasm
- sample set
- target
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/03—Recognition of patterns in medical or anatomical images
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Apparatus For Radiation Diagnosis (AREA)
Abstract
The present invention relates to detection model construction method, detection method and the devices of a kind of Lung neoplasm to determine region of interest area image according to the corresponding description information of lung images and lung images obtained in advance；Region of interest area image is handled, multiscale target sample set is obtained；Using multiscale target sample set, the target integrated rolled product neural network being pre-designed is trained, Lung neoplasm detection model is obtained.Using technical solution of the present invention, by being trained to obtain Lung neoplasm detection model to integrated convolutional neural networks using multiple dimensioned sample set, using this Lung neoplasm detection model, the accuracy of testing result can be enhanced, improve the accuracy rate of diagnostic result, and the integrated convolution neural network model that the model that neural network obtains is multiple dimensioned input is accumulated by the trained integrated rolled of multiple dimensioned sample set, single input single channel can be evaded and extract the incomprehensive caused influence to recognition result of feature, misjudgement misjudgment phenomenon is avoided, recognition accuracy and detection efficiency are improved.
Description
Technical field
The present invention relates to artificial intelligence and Medical Imaging Technology field, and in particular to a kind of detection model building of Lung neoplasm
Method, detection method and device.
Background technique
Lung cancer is the most common malignant tumour, and the Death Causes of Tumor for position of ranking the first in the world.Early detection lung at present
Cancer presence is more highly difficult, and it is abnormal to go to a doctor in time that patient can not discover own bodies situations in time, bright until there is hemoptysis etc.
Aobvious symptom.Early discovery, early intervention, early treatment have very obvious action to the survival rate for improving patients with lung cancer.Therefore, CT quilt
It is considered to detect one of most effective means of lung cancer early, lung CT image is detected, determines that the position of Lung neoplasm is
The necessary means of lung cancer " early discovery, early treatment ".
Computer-aided diagnosis system can carry out a series of processing to lung CT image by related algorithm, final right
The classification of lung CT image is predicted.The work load that doctor can not only be greatly reduced, thereby reduces because of fatigue etc.
Mistaken diagnosis caused by subjective factor, a possibility that failing to pinpoint a disease in diagnosis, more doctor provide the effective suggestion for differentiating Lung neoplasm, are the early stage of lung cancer
Prevention and treatment provides powerful guarantee with monitoring, has great significance for the diagnosis of lung cancer.
But detection model used in the method for current computer aided detection and diagnosis is conventional machines mostly
The method of the method for habit or the deep learning using the single pass convolutional neural networks structure of single input, conventional machines study
The means that method uses manual features to extract are not suitable for the big problem of processing data volume, however the form of patient and lesion is shape
Shape and color color and ever-changing, with the addition of new patient data, old feature it is possible that not applicable situation, thus
Serious error is brought, and the single pass convolutional neural networks structure of single input is only capable of carrying out for the Lung neoplasm of a certain scale
Feature extraction, but Lung neoplasm is not of uniform size, influences Detection accuracy.
Therefore, the detection model and detection method detected in the prior art to Lung neoplasm will lead to testing result not
Accurately, the accuracy rate of diagnostic result is reduced.
Summary of the invention
In view of this, the purpose of the present invention is to provide a kind of detection model construction method of Lung neoplasm, detection method and
Device, big to solve Lung neoplasm detection model error in the prior art, testing result inaccuracy reduces the accurate of diagnostic result
The problem of rate.
In order to achieve the above object, the present invention adopts the following technical scheme:
A kind of detection model construction method of Lung neoplasm, comprising:
According to the lung images and the corresponding description information of the lung images obtained in advance, area-of-interest figure is determined
Picture；
The region of interest area image is handled, multiscale target sample set is obtained；
Using the multiscale target sample set, the target integrated rolled product neural network being pre-designed is trained, is obtained
To Lung neoplasm detection model.
Further, in method described above, lung images and the lung images pair that the basis obtains in advance
The description information answered determines region of interest area image, comprising:
Image enhancement processing is carried out to the lung images obtained in advance, obtains target image；
The corresponding description information of the lung images obtained in advance is analyzed, coordinate information is obtained；
According to the target image and the coordinate information, the region of interest area image is determined.
Further, in method described above, the coordinate information includes that Lung neoplasm coordinate information and non-Lung neoplasm are sat
Mark information；
It is described according to the target image and the coordinate information, determine the region of interest area image, comprising:
The corresponding image of the Lung neoplasm coordinate information is intercepted on the target image, obtains positive sample image；Institute
It states and intercepts the corresponding image of the non-Lung neoplasm coordinate information on target image, obtain negative sample image；
The positive sample image and the negative sample image are combined, the region of interest area image is obtained.
Further, described that the region of interest area image is handled in method described above, it obtains multiple dimensioned
Target sample collection, comprising:
Based on preset n different scale, the region of interest area image is normalized, is obtained with n
The primary data sample collection of the scale, the n are the positive integer more than or equal to 2；
Data amplification processing is carried out to the primary data sample collection, obtains amplification data sample set；
The primary data sample collection and the amplification data sample set are merged, the multiscale target sample is obtained
This collection.
Further, method described above, it is described to utilize the multiscale target sample set, to the target being pre-designed
Integrated convolutional neural networks are trained, before obtaining Lung neoplasm detection model, further includes:
Obtain convolutional neural networks corresponding with each scale；
The n convolutional neural networks that will acquire are merged, and the target integrated rolled product neural network is obtained.
Further, described to utilize the multiscale target sample set in method described above, to the mesh being pre-designed
It marks integrated convolutional neural networks to be trained, obtains Lung neoplasm detection model, comprising:
The multiscale target sample set is classified, training sample set, verifying sample set and test sample collection are obtained；
Target integrated rolled product neural network is trained using the training sample set and the verifying sample set
And verifying, obtain integrated convolution neural network model；
The test sample collection is input to the integrated convolution neural network model, obtains test result；
According to the test result, test accuracy rate is determined；
Detect whether the test accuracy rate is greater than default accuracy rate；
If so, using the integrated convolution neural network model as Lung neoplasm detection model.
The present invention also provides a kind of detection methods of Lung neoplasm, comprising:
Obtain image to be detected；
Based on Lung neoplasm detection model, described image to be detected is inputted, testing result is obtained；
The Lung neoplasm detection model is constructed by the detection model construction method of above-mentioned Lung neoplasm.
The present invention also provides a kind of detection model construction devices of Lung neoplasm, comprising:
Determining module, for determining according to the lung images and the corresponding description information of the lung images obtained in advance
Region of interest area image；
Processing module obtains multiscale target sample set for handling the region of interest area image；
Training module, for utilizing the multiscale target sample set, to the target integrated rolled product nerve net being pre-designed
Network is trained, and obtains Lung neoplasm detection model.
Further, in device described above, the determining module includes: first processing units, analytical unit and figure
As determination unit；
The first processing units obtain mesh for carrying out image enhancement processing to the lung images obtained in advance
Logo image；
The analytical unit, for analyzing the corresponding description information of the lung images obtained in advance,
Obtain coordinate information；
Described image determination unit, for determining the region of interest according to the target image and the coordinate information
Area image.
The present invention also provides a kind of detection devices of Lung neoplasm, comprising: image collection module and detection module；
Described image obtains module, for obtaining image to be detected；
The detection module inputs described image to be detected, obtains testing result for being based on Lung neoplasm detection model；
The Lung neoplasm detection model is constructed by the detection model construction method of above-mentioned Lung neoplasm.
Detection model construction method, detection method and the device of Lung neoplasm of the invention are schemed according to the lung obtained in advance
Picture and the corresponding description information of lung images, determine region of interest area image；Region of interest area image is handled, is obtained more
Scaled target sample set；Using multiscale target sample set, the target integrated rolled product neural network being pre-designed is trained,
Obtain Lung neoplasm detection model.Using technical solution of the present invention, by utilizing multiscale target sample set to target integrated rolled
Product neural network is trained to obtain Lung neoplasm detection model, and it is special that convolutional neural networks distinguish over the artificial extraction of conventional machines study
Sign is to carry out deep learning to characteristics of image by convolution operation to realize the effect for automatically extracting feature, deep learning volume
The domestic demand of product neural network model is exactly big data volume, is continuously added and more news suitable for data, to subtract
Small error, the Lung neoplasm detection model constructed using the present invention, can enhance the accuracy of testing result, improve diagnostic result
Accuracy rate, and be integrated based on multiple dimensioned input by the model that multiple dimensioned sample set training integrated rolled product neural network obtains
Convolutional neural networks model can evade single input single channel and extract the incomprehensive caused influence to recognition result of feature, keep away
Fault-avoidance sentences misjudgment phenomenon, improves recognition accuracy and detection efficiency.
It should be understood that above general description and following detailed description be only it is exemplary and explanatory, not
It can the limitation present invention.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.
Fig. 1 is the flow chart of the detection model construction method embodiment one of Lung neoplasm of the invention；
Fig. 2 is the flow chart of the detection model construction method embodiment two of Lung neoplasm of the invention；
Fig. 3 is the precision line chart of training sample set in the detection model construction method of Lung neoplasm of the invention；
Fig. 4 is the precision line chart that sample set is verified in the detection model construction method of Lung neoplasm of the invention；
Fig. 5 is the structural schematic diagram of the detection method embodiment of Lung neoplasm of the invention；
Fig. 6 is the structural schematic diagram of the detection model construction device embodiment one of Lung neoplasm of the invention；
Fig. 7 is the structural schematic diagram of the detection model construction device embodiment two of Lung neoplasm of the invention；
Fig. 8 is the structural schematic diagram of the detection device embodiment of Lung neoplasm of the invention.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, technical solution of the present invention will be carried out below
Detailed description.Obviously, described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.Base
Embodiment in the present invention, those of ordinary skill in the art are obtained all without making creative work
Other embodiment belongs to the range that the present invention is protected.
Fig. 1 is the flow chart of the detection model construction method embodiment one of Lung neoplasm of the invention.As shown in Figure 1, this reality
The detection model construction method for applying the Lung neoplasm of example can specifically include following steps:
The lung images and the corresponding description information of lung images that S101, basis obtain in advance, determine area-of-interest figure
Picture；
The detection model construction method of the Lung neoplasm of the present embodiment is corresponding firstly the need of acquisition lung images and lung images
Description information, wherein lung images are preferably lung CT image, and the corresponding description information of lung images is preferably XML format
Markup information, then the present embodiment can obtain lung CT image and lung CT from LIDC-IDRI Lung neoplasm public database
The corresponding XML markup information of image, the amount of images obtained in the present embodiment are the lung CT image of 300 patients, XML mark
Information includes the corresponding XML markup information of lung CT image of 300 patients, and the present embodiment is not intended to limit the lung of acquisition
The quantity of image and description information.After getting lung CT image and the corresponding XML markup information of lung CT image, according to the figure
Picture and information determine region of interest area image.
S102, region of interest area image is handled, obtains multiscale target sample set；
Through the above steps, after determining region of interest area image, which is handled, will be handled
Image afterwards is as multiscale target sample set.
S103, using multiscale target sample set, the target integrated rolled product neural network being pre-designed is trained, is obtained
To Lung neoplasm detection model.
Through the above steps, after obtaining multiscale target sample set, using the multiscale target sample set to being pre-designed
Target integrated rolled product neural network be trained, will be most afterwards through the obtained integrated convolutional Neural of multiscale target sample set training
Network model is as Lung neoplasm detection model.
The detection model construction method of the Lung neoplasm of the present embodiment, according to the lung images and lung images pair obtained in advance
The description information answered determines region of interest area image；Region of interest area image is handled, target sample collection is obtained；It utilizes
Multiscale target sample set is trained the target integrated rolled product neural network being pre-designed, obtains Lung neoplasm detection model.
In this way by being trained to obtain Lung neoplasm detection model to target integrated rolled product neural network using multiscale target sample set,
Convolutional neural networks distinguish over the artificial extraction feature of conventional machines study, are to carry out depth to characteristics of image by convolution operation
It practises to realize the effect for automatically extracting feature, the domestic demand of deep learning convolutional neural networks model is exactly big data
Amount is continuously added and more news suitable for data, thus reduce error, then, the Lung neoplasm constructed through this embodiment
Detection model just can enhance the accuracy of testing result, improve the accuracy rate of diagnostic result, and assemble for training by multiple dimensioned sample
Practice the integrated convolution neural network model that the model that integrated convolutional neural networks obtain is multiple dimensioned input, single input list can be evaded
Channel extract feature it is incomprehensive caused by influence to recognition result, avoid misjudgement misjudgment phenomenon, improve recognition accuracy and
Detection efficiency.
Fig. 2 is the flow chart of the detection model construction method embodiment two of Lung neoplasm of the invention.As shown in Fig. 2, this reality
The detection model construction method for applying the Lung neoplasm of example can specifically include following steps:
S201, image enhancement processing is carried out to the lung images obtained in advance, obtains target image；
The detection model construction method of the Lung neoplasm of the present embodiment obtains lung images first, and wherein lung images are preferably
Lung CT image can be obtained from LIDC-IDRI Lung neoplasm public database, after getting lung CT image, be carried out to it
Image enhancement processing, to obtain target image.
The method of image enhancement processing includes histogram equalization and median filtering.Histogram equalization is enhancing image pair
Than a kind of method of degree, its basic thought is that the grey level histogram of piece image flattens, and is made each in transformed image
The distribution probability of gray value is all identical.The histogram of original image is obtained by Cumulative Distribution Function first, then histogram is repaired
It is changed to equally distributed histogram, the image after being equalized；The basic principle of median filtering is will be each in image
The gray value of a pixel is set as the intermediate value of all gray values in its neighborhood, and then eliminates isolated noise spot, to reach figure
The purpose of image intensifying, the present embodiment can choose 3 × 3 median filters and be filtered operation to original image to be enhanced
Image afterwards.
S202, the corresponding description information of the lung images obtained in advance is analyzed, obtains coordinate information；
Wherein, coordinate information includes Lung neoplasm coordinate information and non-Lung neoplasm coordinate information；
First while obtaining lung images, the corresponding description information of the lung images, each lung images are also obtained
All correspond to a set of description information.If lung images are lung CT image, description information can be corresponding for the lung CT image
XML markup information, XML markup information can be obtained from LIDC-IDRI Lung neoplasm public database, and XML markup information is by passing through
Test iconography expert mark abundant.It after getting description information, analyzes it, obtains coordinate information, wherein coordinate
Information includes Lung neoplasm coordinate information and non-Lung neoplasm coordinate information, and non-Lung neoplasm herein is that similar Lung neoplasm is not but lung knot
The part of section.
If description information is XML markup information, the basic procedure for obtaining coordinate information is as follows:
(1) sufferer number is obtained, is located at<seriesInstanceUid></SeriesInstanceUid>between character
String is the number information of patient.
(2) circulation extracts markup information.XML file includes<readingSession></readingSession>, respectively
Correspond to the markup information of each expert.Each<readingSession></readingSession>between carry out it is following
Operation:
1. searching for<unblindedReadNodule></unblindedReadNodule>.This tag memory stores up Lung neoplasm
Information.If the label includes<characteristics></characteristics>label, then it is straight for representing this Lung neoplasm
Lung neoplasm of the diameter between 3mm-30mm, each<roi></roi>in in store Lung neoplasm all coordinate informations.Its
In,<imageZposition></imageZposition>corresponding CT image corresponds to frame number；Every a pair<edgeMap></
EdgeMap>storage<xCoord>and<yCoord>information are the profile coordinates of Lung neoplasm.If do not include <
Characteristics></characteristics>label then illustrates that the tubercle is lesser tubercle, only needs to extract<roi></
Roi>in coordinate<imageZposition>,<xCoord>and<yCoord>information, which represents brief summary
The centre coordinate of section.
2. searching for<nonNodule></nonNodule>.What the label saved is the non-Lung neoplasm information of mark, the label
It only needs to extract down<imageZposition>information and<locus></locus>in a coordinate<xCoord>with<yCoord>
Information, the coordinate information represent the centre coordinate of non-Lung neoplasm tissue.
By the above-mentioned analyzing step to XML file, available Lung neoplasm coordinate information and non-Lung neoplasm coordinate information.
S203, according to target image and coordinate information, determine area-of-interest；
Through the above steps, after obtaining target image and coordinate information, lung in coordinate information is intercepted on target image
The corresponding image of tubercle coordinate information, using the image as positive sample image；Non- lung in coordinate information is intercepted on target image
The corresponding image of tubercle coordinate information, using the image as negative sample image.
After obtaining positive sample image and negative sample image, combine positive sample image and negative sample image, after combination
All images include positive sample image and negative sample image as region of interest area image, i.e. region of interest area image.
S204, it is based on preset n different scale, region of interest area image is normalized, obtained with n
The primary data sample collection of scale, n are the positive integer more than or equal to 2；
Through the above steps, after obtaining region of interest area image, region of interest area image is normalized,
In, normalized can realize that it is in two-dimensional space that bicubic interpolation, which is called bi-cubic interpolation, using bicubic interpolation algorithm
Most common interpolation method.In this approach, function f can pass through in rectangular mesh nearest 16 in the value of point (x, y)
The weighted average of a sampled point obtains, and is needed herein using two polynomial interopolation cubic functions, each direction uses one.
Data are normalized, all pictures are normalized to n scale respectively and obtain primary data sample collection.Wherein, n is
Positive integer more than or equal to 2.In the present embodiment, n is preferably 3, and data are normalized, and all pictures are distinguished normalizing
Change to 32 × 32,64 × 64 and 128 × 128 3 scales and obtains primary data sample collection.
S205, data amplification processing is carried out to primary data sample collection, obtains amplification data sample set；
Through the above steps, it after obtaining primary data sample collection, needs to carry out data amplification to the primary data sample collection
Processing, obtains amplification data sample set.Wherein, carrying out data amplification processing to primary data sample collection includes to initial data sample
All images of this concentration carry out left and right overturning and spin upside down, so that many transformed images are obtained, by transformed institute
There is image as amplification data sample set.
S206, primary data sample collection and amplification data sample set are merged, obtains multiscale target sample set；
Through the above steps, after obtaining primary data sample collection and amplification data sample set, by primary data sample collection
It is merged with amplification data sample set, so that multiple dimensioned sample set is obtained, using multiple dimensioned sample set as multiscale target sample
This collection.It include all images in primary data sample collection and amplification data sample set i.e. in multiscale target sample set.
S207, convolutional neural networks corresponding with each scale are obtained；
It learns through the above steps, by area-of-interest image normalization to n scale, therefore, it is necessary to obtain n convolution
Neural network, each convolutional neural networks are corresponding with each scale.
Convolutional neural networks are generally made of convolutional layer, pond layer, full articulamentum etc..Convolutional layer is responsible for obtaining the office of image
Portion's feature, and local feature is transmitted backward from network；Pond layer carries out down-sampling on Spatial Dimension, is responsible for reducing data
Amount；Full articulamentum will calculate classification scoring, and loss function can carry out network weight according to classification scoring and the gap of target
Adjustment.
S208, the n convolutional neural networks that will acquire are merged, and target integrated rolled product neural network is obtained；
Through the above steps, after obtaining n convolutional neural networks, n independent convolutional neural networks are merged, from
And integrated convolutional neural networks are obtained, using the integrated convolutional neural networks as target integrated rolled product neural network.This implementation
N is preferably 3 in example, and table 1 is 3 independent convolutional neural networks CNN structures, as shown in table 1:
Table 1
Wherein, convolutional neural networks CNN1 shares 1 input layer, 2 convolutional layers, 2 pond layers, 2 full articulamentums, and 1
A output layer, the convolution nuclear volume of two of them convolutional layer are respectively 8,16, and pondization is using maximum pondization strategy；Convolutional Neural net
Network CNN2 shares 1 input layer, 3 convolutional layers, 3 pond layers, 2 full articulamentums, 1 output layer, wherein three convolutional layers
Convolution nuclear volume be respectively 8,16,32, pondization is tactful using maximum pondization；Convolutional neural networks CNN3 shares 1 input layer,
4 convolutional layers, 4 pond layers, 2 full articulamentums, 1 output layer, wherein the convolution nuclear volume of four convolutional layers be respectively 8,
16,32,64；Pondization is using maximum pondization strategy.
Three separate network convolution nuclear volumes are incremented by successively, because as network depth increases, more convolution kernels can be with
Extract more profound characteristics of image.After all convolution operations access batch normalization Batch Normalization and with
Machine inactivates Dropout, the generalization ability of further lift scheme.Using nonlinear activation function ReLU as activation letter in network
Number can prevent data the phenomenon that both ends generate saturation, to avoid weight that from can not carrying out compared to traditional Sigmoid function
The situation of update.Wherein, Batch Normalization is the skill of neural metwork training, it can not only accelerate model
Convergence rate, and more importantly in alleviating deep layer network to a certain degree the problem of " gradient disperse " so that
Training deep layer network model is more easier and stablizes.Dropout just refers in each trained batch, can significantly reduce
Fitting phenomenon keeps model generalization stronger.
S209, multiscale target sample set is classified, obtains training sample set, verifying sample set and test sample
Collection；
After obtaining multiscale target sample set by step S206, classifies to multiscale target sample set, be divided into
Training sample set, verifying sample set and three kinds of test sample collection, and training sample set, verifying sample set and test sample collection are
Multiple dimensioned sample set.
S210, target integrated rolled product neural network is trained and is verified using training sample set and verifying sample set,
Obtain integrated convolution neural network model；
Through the above steps, it after obtaining target integrated rolled product neural network, training sample set and verifying sample set, utilizes
Training sample set is trained target integrated rolled product neural network, recycles verifying sample set to the target after training after training
Integrated convolutional neural networks are verified, so that integrated convolution neural network model is obtained, due to training sample set and verifying sample
This collection is multiple dimensioned sample set, so the integrated rolled obtained after training sample set training and verifying sample set verifying
Product neural network model is the integrated convolution neural network model based on multiple dimensioned input.
Lung neoplasm picture is subjected to multiple dimensioned input, difference can be excavated using the detection identification of integrated convolutional neural networks
The characteristics of image of Lung neoplasm and regular information, can evade single input single channel most possibly in this way and extract feature under scale
Influence caused by incomprehensive to recognition result, to achieve the purpose that promote Detection accuracy.
Wherein, training sample set is the data sample for carrying out models fitting, directly takes part in the process of model tune ginseng；Verifying
Sample set is the sample set that model training individually reserves in the process, it can be used for adjusting the hyper parameter of model and the energy to model
Power carries out entry evaluation.In neural network, optimal network depth is looked for validation data set, or determine backpropagation
The halt of algorithm or the quantity that hidden layer neuron is selected in neural network.
It is by the way of train in batches that target integrated rolled product neural network, which is trained and is verified, and a round is completed
The penalty values of the round training process can be returned afterwards, and by loss late, backpropagation carries out network weight to loss function from network again
The adjustment of parameter makes it possible to obtain lower loss late.Verifying and deconditioning after training loss convergence, and preservation model
For .h5 file, as final training result, i.e. the integrated convolution neural network model based on multiple dimensioned input, herein integrated
Convolutional neural networks model is integrated study model.
Integrated study understands the machine learning method for exactly gathering multiple learners from letter, sometimes also by
For multi-classifier system.Compared with single learning model, can get for integrated study model is significantly more excellent than single learning model
Generalization Capability more.Integrated convolutional neural networks by the sample data of three kinds of scales be respectively fed in CNN1, CNN2 and CNN3 into
Row detection, finally carries out integrating final result to the testing result of three networks.In integrated study theory CNN1, CNN2 and
CNN3 is Weak Classifier, and the method that Weak Classifier integrates strong classifier has averaging, ballot and weighted sum etc..The present embodiment
It is middle select ballot device mode integrated, i.e., only there are three types of classification results be all 1 when final output be just 1, otherwise tie
Fruit is 0.
S211, test sample collection is input to integrated convolution neural network model, obtains test result；
Through the above steps, after obtaining integrated convolution neural network model and test sample collection, test sample collection is defeated
Enter into integrated convolution neural network model, obtain test result, i.e., whether is Lung neoplasm.Due to multiscale target sample set packet
Image containing n scale, integrating convolution neural network model is the integrated convolution neural network model based on multiple dimensioned input, institute
With test sample concentrates image also comprising n scale, integrates in convolution neural network model also comprising n and multiple dimensioned mesh
The corresponding convolutional neural networks of standard specimen this concentration graphical rule are input to integrated convolution neural network model in test sample collection
In the process, the image that test sample is concentrated is input in corresponding convolutional neural networks according to scale correspondence, convolutional Neural
Each convolutional neural networks that network model includes can export a recognition result, also wrap in integrated convolution neural network model
Containing a voting mechanism, i.e., n convolutional neural networks for including in integrated convolution neural network model can all export an identification
As a result, can then obtain n recognition result, when only n recognition result is expressed as Lung neoplasm, the test result be just be lung knot
Section, as long as having an expression in n recognition result is not Lung neoplasm, test result be just be not Lung neoplasm.The original done so
Confirm three times because being to do true tubercle, guarantees the accuracy of strong disaggregated model.
Test sample collection is used to assess the generalization ability of mould final mask, but cannot function as adjusting ginseng, selection feature scheduling algorithm
The foundation of relevant selection.Test sample collection is that network training completes final inspection data, and network is accurate from test set
Degree can be seen that the quality of this model.
S212, according to test result, determine test accuracy rate；
Through the above steps, after obtaining test result, test result and correct result are compared, so that it is determined that test
Accuracy rate.Wherein, the calculation of test accuracy rate is that test result sample size identical with correct result is divided by test specimens
The total quantity of this concentration sample.
Whether S213, detection test accuracy rate are greater than default accuracy rate；
Through the above steps, after obtaining test accuracy rate, whether detection test accuracy rate is greater than default accuracy rate, wherein in advance
If accuracy rate is pre-set numerical value, if test accuracy rate can be more than default accuracy rate, in expression above-mentioned steps
Obtained integrated convolution neural network model is the model of accuracy rate qualification.
If S214, test accuracy rate are greater than default accuracy rate, detected integrated convolution neural network model as Lung neoplasm
Model.
Through the above steps, if detecting that test accuracy rate is greater than default accuracy rate, by integrated convolutional neural networks
Model is as Lung neoplasm detection model.If detecting that test accuracy rate is not more than default accuracy rate, then need to reacquire lung
Portion's image and description information, and integrated convolution Artificial Neural Network Structures are adjusted, to continue to integrated convolutional neural networks
Model is trained and verifies, after test accuracy rate is greater than default accuracy rate, then deconditioning.
Fig. 3 is the precision line chart of training sample set in the detection model construction method of Lung neoplasm of the invention；
Fig. 4 is the precision line chart that sample set is verified in the detection model construction method of Lung neoplasm of the invention.This implementation
Example in, to target integrated rolled product neural network be trained and verification process in, training sample set and verifying sample set training and
The precision (i.e. accuracy rate) of verifying is as shown in Figure 3 and Figure 4.Wherein, epoch when abscissa is model training, an epoch refer to
The process of completing a forward calculation and backpropagation is sent into network for all data.Since an epoch is usually too big,
Computer can not load, we can divide it into several lesser batches.Batch is exactly that each be sent into network is trained
A part of data, and batch Size is exactly the quantity of training sample in each batch.Iterations is to complete once
Batch number needed for epoch.For example, there are 2000 data, it is divided into 4 batch, then batch size is exactly 500.Fortune
All data of row are trained, and are completed 1 epoch, are needed to carry out 4 iterations.It, can before convolutional neural networks training
With the number of the size of self-setting batch size and epoch；Ordinate is accuracy rate.
In the present embodiment, it is not intended to limit the sequence that executes of step S207-S208, step S207-S208 can be in step
It is executed between either step before S209.
The detection model construction method of the Lung neoplasm of the present embodiment carries out at image enhancement the lung images obtained in advance
Reason, obtains target image；The corresponding description information of the lung images obtained in advance is analyzed, coordinate information is obtained；In mesh
The corresponding image of Lung neoplasm coordinate information is intercepted in logo image, obtains positive sample image；Upper non-Lung neoplasm is intercepted in target image
The corresponding image of coordinate information, obtains negative sample image；Positive sample image and negative sample image are combined, area-of-interest is obtained
Image；Based on preset n different scale, region of interest area image is normalized, obtains the original with n scale
Beginning set of data samples；Data amplification processing is carried out to primary data sample collection, obtains amplification data sample set；It will be multiple dimensioned original
Set of data samples and amplification data sample set merge, and obtain multiscale target sample set；It obtains corresponding with each scale
Convolutional neural networks；The n convolutional neural networks that will acquire are merged, and target integrated rolled product neural network is obtained；It will be more
Scaled target sample set is classified, and training sample set, verifying sample set and test sample collection are obtained；Using training sample set and
Verifying sample set is trained and verifies to target integrated rolled product neural network, obtains integrated convolution neural network model；It will survey
Examination sample set is input to integrated convolution neural network model, obtains test result；According to test result, test accuracy rate is determined；
Whether detection test accuracy rate is greater than default accuracy rate；If so, detecting mould for integrated convolution neural network model as Lung neoplasm
Type.In this way by the way that target integrated rolled product neural network is trained, verifies and is tested using multiscale target sample set, obtain
Lung neoplasm detection model, convolutional neural networks distinguish over the artificial extraction feature of conventional machines study, are by convolution operation to figure
As feature progress deep learning to realize the effect for automatically extracting feature, the inherent of deep learning convolutional neural networks model is needed
Asking is exactly big data volume, is continuously added and more news suitable for data, thus reduce error, then, by this implementation
The Lung neoplasm detection model of example building, just can enhance the accuracy of testing result, improve the accuracy rate of diagnostic result, and this
Target sample collection used in the examples is multiple dimensioned sample set, and the convolutional neural networks used are integrated convolutional neural networks,
The obtained Lung neoplasm detection model of training is the integrated convolution neural network model based on multiple dimensioned input, by Lung neoplasm picture into
The multiple dimensioned input of row can excavate the characteristics of image of Lung neoplasm under different scale using the detection identification of integrated convolutional neural networks
With regular information, single input single channel can be evaded most possibly in this way and extract the incomprehensive caused to identification knot of feature
The influence of fruit, network it is last determine when be integrated with the judgement of three networks as a result, can be to avoiding showing for misjudgement erroneous judgement as far as possible
As to keep recognition accuracy higher.And concurrent operation may be implemented in three networks, is promoted while saving computing resource
Detection efficiency.
Fig. 5 is the structural schematic diagram of the detection method embodiment of Lung neoplasm of the invention.As shown in figure 5, the present embodiment
The detection method of Lung neoplasm can specifically include following steps:
S301, image to be detected is obtained；
In the present embodiment, if patient wants detection lung and whether there is Lung neoplasm, the detection method of Lung neoplasm firstly the need of
Obtain image to be detected of user.Wherein, morphological segment is carried out by the lung images to patient, will owned in lung images
The image of similar Lung neoplasm is split, as image to be detected.In addition, image to be detected can also be directed to lung by doctor
Portion's image is manually marked, and by the image labeling of similar Lung neoplasms all in lung images and is split.In the present embodiment,
Lung images are preferably lung CT image, are shot by CT machine.
S302, it is based on Lung neoplasm detection model, inputs image to be detected, obtains testing result.
Through the above steps, after obtaining image to be detected, image to be detected is input to and is obtained through the foregoing embodiment
In Lung neoplasm detection model, thus the testing result exported, obtained in testing result include indicate be Lung neoplasm and
Expression is not Lung neoplasm.
The detection method of the Lung neoplasm of the present embodiment obtains image to be detected；Based on Lung neoplasm detection model, input to be checked
Altimetric image obtains testing result.Lung neoplasm detection model be target convolutional neural networks are trained using target sample collection,
What verifying and test obtained, wherein convolutional neural networks distinguish over the artificial extraction feature of conventional machines study, are grasped by convolution
Make to carry out deep learning to characteristics of image to realize the effect for automatically extracting feature, deep learning convolutional neural networks model
Domestic demand is exactly big data volume, is continuously added and more news suitable for data, thus reduce error, then, this reality
The Lung neoplasm detection model that the detection method of the Lung neoplasm of example constructs through the foregoing embodiment is applied to detect image to be detected,
The accuracy that testing result can be enhanced, improves the accuracy rate of diagnostic result, and the Lung neoplasm detection used in the present embodiment
Model is the integrated convolution neural network model based on multiple dimensioned input, and Lung neoplasm picture is carried out multiple dimensioned input, utilizes collection
The characteristics of image of Lung neoplasm and regular information can be excavated under different scale at convolutional neural networks detection identification, it in this way can be with
Evade the incomprehensive caused influence to recognition result that single input single channel extracts feature most possibly, network is last
The judgement of three networks is integrated with when judgement as a result, can be to the phenomenon that misjudging erroneous judgement is avoided as far as possible, to make recognition accuracy
It is higher.And concurrent operation may be implemented in three networks, improves detection efficiency while saving computing resource.
In order to which more comprehensively, corresponding to the detection model construction method of Lung neoplasm provided in an embodiment of the present invention, the application is also
Provide the detection model construction device of Lung neoplasm.
Fig. 6 is the structural schematic diagram of the detection model construction device embodiment one of Lung neoplasm of the invention.As shown in fig. 6,
The detection model construction device of the Lung neoplasm of the present embodiment comprises determining that module 11, processing module 12 and training module 13.
Determining module 11, for determining sense according to the lung images and the corresponding description information of lung images obtained in advance
Interest area image；
Processing module 12 obtains multiscale target sample set for handling region of interest area image；
Training module 13, for utilizing multiscale target sample set, to the target integrated rolled product neural network being pre-designed
It is trained, obtains Lung neoplasm detection model.
The detection model of the Lung neoplasm of the present embodiment constructs model, determining module 11 according to the lung images obtained in advance and
The corresponding description information of lung images, determines region of interest area image；Processing module 12 handles region of interest area image,
Obtain multiscale target sample set；Training module 13 utilizes multiscale target sample set, to the target integrated rolled product being pre-designed
Neural network is trained, and obtains Lung neoplasm detection model.In this way by utilizing multiscale target sample set to target integrated rolled
Product neural network is trained to obtain Lung neoplasm detection model, and it is special that convolutional neural networks distinguish over the artificial extraction of conventional machines study
Sign is to carry out deep learning to characteristics of image by convolution operation to realize the effect for automatically extracting feature, deep learning volume
The domestic demand of product neural network model is exactly big data volume, is continuously added and more news suitable for data, to subtract
Small error, then, the Lung neoplasm detection model constructed through this embodiment just can enhance the accuracy of testing result, improve
The accuracy rate of diagnostic result, and be multiple dimensioned input by the model that multiple dimensioned sample set training integrated rolled product neural network obtains
Integrated convolution neural network model, can evade single input single channel extract feature it is incomprehensive caused by the shadow of recognition result
It rings, avoids misjudgement misjudgment phenomenon, improve recognition accuracy and detection efficiency.
About the device in above-described embodiment, wherein modules execute the concrete mode of operation in related this method
Embodiment in be described in detail, no detailed explanation will be given here.
Fig. 7 is the structural schematic diagram of the detection model construction device embodiment two of Lung neoplasm of the invention.As shown in fig. 7,
Determining module 11 in the detection model construction device of the Lung neoplasm of the present embodiment specifically includes: first processing units 111, analysis
Unit 112 and image determination unit 113, image determination unit 113 include: interception subelement 1131 and combination subelement 1132.
First processing units 111 obtain target figure for carrying out image enhancement processing to the lung images obtained in advance
Picture；
Analytical unit 112 obtains coordinate letter for analyzing the corresponding description information of the lung images obtained in advance
Breath；
Image determination unit 113, for determining the region of interest area image according to target image and coordinate information.
Wherein, coordinate information includes Lung neoplasm coordinate information and non-Lung neoplasm coordinate information；Image determination unit 113 is first
First, the corresponding image of Lung neoplasm coordinate information is intercepted on target image, obtains positive sample image；It is intercepted on target image non-
The corresponding image of Lung neoplasm coordinate information, obtains negative sample image；Secondly, combining positive sample image and negative sample image, obtain
To region of interest area image.
Further, the processing module 12 of the present embodiment specifically includes: the second processing unit 121, third processing unit 122
With combining unit 123.
The second processing unit 121 is normalized region of interest area image for being based on preset n different scale
Processing, obtains the primary data sample collection with n scale, and n is the positive integer more than or equal to 2；
Third processing unit 122 obtains amplification data sample for carrying out data amplification processing to primary data sample collection
Collection；
Combining unit 123 obtains multiple dimensioned mesh for merging primary data sample collection and amplification data sample set
Mark sample set.
Further, the detection model construction device of the Lung neoplasm of the present embodiment further includes that neural network obtains 14 He of module
Fusion Module 15.
Neural network obtains module 14, for obtaining convolutional neural networks corresponding with each scale；
Fusion Module 15, the n convolutional neural networks for will acquire are merged, and target integrated rolled product nerve net is obtained
Network.
Further, the training module 13 of the present embodiment specifically includes: sample classification unit 131, model treatment unit
132, model measurement unit 133, the first determination unit 134, detection unit 135 and the second determination unit 136.
Sample classification unit 131 obtains training sample set, verifying sample for multiscale target sample set to be classified
This collection and test sample collection；
Model treatment unit 132, for utilizing training sample set and verifying sample set to target integrated rolled product neural network
It is trained and verifies, obtain integrated convolution neural network model；
Model measurement unit 133 obtains test knot for test sample collection to be input to integrated convolution neural network model
Fruit；
First determination unit 134, for determining test accuracy rate according to test result；
Detection unit 135, for detecting whether test accuracy rate is greater than default accuracy rate；
Second determination unit 136, if being greater than default accuracy rate for test accuracy rate, by integrated convolution neural network model
As Lung neoplasm detection model.
The detection model construction device of the Lung neoplasm of the present embodiment, by first processing units 111 to the lung obtained in advance
Portion's image carries out image enhancement processing, obtains target image；It is corresponding to the lung images obtained in advance by analytical unit 112
Description information is analyzed, and coordinate information is obtained；By image determination unit 113 according to target image and coordinate information, obtain
Region of interest area image；It is based on preset n different scale by the second processing unit 121, region of interest area image is carried out
Normalized obtains the primary data sample collection with n scale；By third processing unit 122 to primary data sample
Collection carries out data amplification processing, obtains amplification data sample set；By primary data sample collection and number is expanded by combining unit 123
It is merged according to sample set, obtains multiscale target sample set；Module 14 is obtained by neural network to obtain and each scale phase
Corresponding convolutional neural networks；It is merged by the n convolutional neural networks that Fusion Module 15 will acquire, it is integrated to obtain target
Convolutional neural networks；Multiscale target sample set is classified by sample classification unit 131, training sample set is obtained, tests
Demonstrate,prove sample set and test sample collection；It is integrated to target using training sample set and verifying sample set by model treatment unit 132
Convolutional neural networks are trained and verify, and obtain integrated convolution neural network model；It will be tested by model measurement unit 133
Sample set is input to integrated convolution neural network model, obtains test result；It is tied by the first determination unit 134 according to test
Fruit determines test accuracy rate；Detect whether test accuracy rate is greater than default accuracy rate by detection unit 135；If so, passing through the
Three determination units 136 are using integrated convolution neural network model as Lung neoplasm detection model.In this way by utilizing multiscale target
Sample set is trained, verifies and tests to target integrated rolled product neural network, obtains Lung neoplasm detection model, convolutional Neural net
Network distinguishes over the artificial extraction feature of conventional machines study, is to carry out deep learning to characteristics of image by convolution operation to realize
The effect of feature is automatically extracted, the domestic demand of deep learning convolutional neural networks model is exactly big data volume, is suitable for number
According to being continuously added and more news, thus reduce error, then, the Lung neoplasm detection model constructed through this embodiment, just
The accuracy that testing result can be enhanced, improves the accuracy rate of diagnostic result, and the target sample collection used in the present embodiment
For multiple dimensioned sample set, the convolutional neural networks used train obtained Lung neoplasm detection model to integrate convolutional neural networks
For the integrated convolution neural network model based on multiple dimensioned input, Lung neoplasm picture is subjected to multiple dimensioned input, utilizes integrated rolled
Product neural network detection identification can excavate the characteristics of image of Lung neoplasm and regular information under different scale, in this way can be maximum
Possibly evade the incomprehensive caused influence to recognition result that single input single channel extracts feature, network determines finally
When be integrated with three networks judgement as a result, can to avoid as far as possible misjudgement judge by accident the phenomenon that, to keep recognition accuracy higher.
And concurrent operation may be implemented in three networks, improves detection efficiency while saving computing resource.
About the device in above-described embodiment, wherein modules execute the concrete mode of operation in related this method
Embodiment in be described in detail, no detailed explanation will be given here.
Fig. 8 is the structural schematic diagram of the detection device embodiment of Lung neoplasm of the invention.As shown in figure 8, the present embodiment
The detection device of Lung neoplasm includes image collection module 21 and detection module 22.
Image collection module 21, for obtaining image to be detected；
Detection module 22 inputs image to be detected, obtains testing result for being based on Lung neoplasm detection model；Lung neoplasm
Detection model through the foregoing embodiment in Lung neoplasm detection model construction method building.
The detection device of the Lung neoplasm of the present embodiment obtains image to be detected by image collection module 21；Pass through detection
Module 22 is based on Lung neoplasm detection model, inputs image to be detected, obtains testing result.Lung neoplasm detection model is to utilize more rulers
Degree target sample collection is trained, verifies and tests to target integrated rolled product neural network, wherein convolutional neural networks
It distinguishes over conventional machines study and manually extracts feature, be that deep learning is carried out to characteristics of image to realize certainly by convolution operation
The dynamic effect for extracting feature, the domestic demand of deep learning convolutional neural networks model is exactly big data volume, is suitable for data
It is continuously added and more news, thus reduce error, then, the detection method of the Lung neoplasm of the present embodiment passes through above-mentioned implementation
The Lung neoplasm detection model of example building detects image to be detected, can enhance the accuracy of testing result, improves diagnosis
As a result accuracy rate, and the Lung neoplasm detection model used in the present embodiment is the integrated convolutional Neural based on multiple dimensioned input
Lung neoplasm picture is carried out multiple dimensioned input by network model, can be excavated not using the detection identification of integrated convolutional neural networks
Characteristics of image and regular information with Lung neoplasm under scale, can evade single input single channel most possibly in this way and extract feature
It is incomprehensive caused by influence to recognition result, network it is last determine when be integrated with the judgement of three networks as a result, energy
The phenomenon that avoiding misjudgement to judge by accident as far as possible, to keep recognition accuracy higher.And concurrent operation may be implemented in three networks,
Detection efficiency is improved while saving computing resource.
About the device in above-described embodiment, wherein modules execute the concrete mode of operation in related this method
Embodiment in be described in detail, no detailed explanation will be given here.
It is understood that same or similar part can mutually refer in the various embodiments described above, in some embodiments
Unspecified content may refer to the same or similar content in other embodiments.
It should be noted that in the description of the present invention, term " first ", " second " etc. are used for description purposes only, without
It can be interpreted as indication or suggestion relative importance.In addition, in the description of the present invention, unless otherwise indicated, the meaning of " multiple "
Refer at least two.
Any process described otherwise above or method description are construed as in flow chart or herein, and expression includes
It is one or more for realizing specific logical function or process the step of executable instruction code module, segment or portion
Point, and the range of the preferred embodiment of the present invention includes other realization, wherein can not press shown or discussed suitable
Sequence, including according to related function by it is basic simultaneously in the way of or in the opposite order, Lai Zhihang function, this should be of the invention
Embodiment person of ordinary skill in the field understood.
It should be appreciated that each section of the invention can be realized with hardware, software, firmware or their combination.Above-mentioned
In embodiment, software that multiple steps or method can be executed in memory and by suitable instruction execution system with storage
Or firmware is realized.It, and in another embodiment, can be under well known in the art for example, if realized with hardware
Any one of column technology or their combination are realized: having a logic gates for realizing logic function to data-signal
Discrete logic, with suitable combinational logic gate circuit specific integrated circuit, programmable gate array (PGA), scene
Programmable gate array (FPGA) etc..
Those skilled in the art are understood that realize all or part of step that above-described embodiment method carries
It suddenly is that relevant hardware can be instructed to complete by program, described program can store in a kind of computer readable storage medium
In, which when being executed, includes the steps that one or a combination set of embodiment of the method.
It, can also be in addition, each functional unit in each embodiment of the present invention can integrate in a processing module
It is that each unit physically exists alone, can also be integrated in two or more units in a module.Above-mentioned integrated mould
Block both can take the form of hardware realization, can also be realized in the form of software function module.The integrated module is such as
Fruit is realized and when sold or used as an independent product in the form of software function module, also can store in a computer
In read/write memory medium.
Storage medium mentioned above can be read-only memory, disk or CD etc..
In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show
The description of example " or " some examples " etc. means specific features, structure, material or spy described in conjunction with this embodiment or example
Point is included at least one embodiment or example of the invention.In the present specification, schematic expression of the above terms are not
Centainly refer to identical embodiment or example.Moreover, particular features, structures, materials, or characteristics described can be any
One or more embodiment or examples in can be combined in any suitable manner.
Although the embodiments of the present invention has been shown and described above, it is to be understood that above-described embodiment is example
Property, it is not considered as limiting the invention, those skilled in the art within the scope of the invention can be to above-mentioned
Embodiment is changed, modifies, replacement and variant.
Claims (10)
1. a kind of detection model construction method of Lung neoplasm characterized by comprising
According to the lung images and the corresponding description information of the lung images obtained in advance, region of interest area image is determined；
The region of interest area image is handled, multiscale target sample set is obtained；
Using the multiscale target sample set, the target integrated rolled product neural network being pre-designed is trained, lung is obtained
Nodule detection model.
2. the method according to claim 1, wherein lung images that the basis obtains in advance and the lung
The corresponding description information of image, determines region of interest area image, comprising:
Image enhancement processing is carried out to the lung images obtained in advance, obtains target image；
The corresponding description information of the lung images obtained in advance is analyzed, coordinate information is obtained；
According to the target image and the coordinate information, the region of interest area image is determined.
3. according to the method described in claim 2, it is characterized in that, the coordinate information includes Lung neoplasm coordinate information and non-lung
Tubercle coordinate information；
It is described according to the target image and the coordinate information, determine the region of interest area image, comprising:
The corresponding image of the Lung neoplasm coordinate information is intercepted on the target image, obtains positive sample image；In the mesh
The corresponding image of the non-Lung neoplasm coordinate information is intercepted in logo image, obtains negative sample image；
The positive sample image and the negative sample image are combined, the region of interest area image is obtained.
4. being obtained the method according to claim 1, wherein described handle the region of interest area image
To multiscale target sample set, comprising:
Based on preset n different scale, the region of interest area image is normalized, is obtained described with n
The primary data sample collection of scale, the n are the positive integer more than or equal to 2；
Data amplification processing is carried out to the primary data sample collection, obtains amplification data sample set；
The primary data sample collection and the amplification data sample set are merged, the multiscale target sample is obtained
Collection.
5. according to the method described in claim 4, it is characterized in that, described utilize the multiscale target sample set, to preparatory
The target integrated rolled product neural network of design is trained, before obtaining Lung neoplasm detection model, further includes:
Obtain convolutional neural networks corresponding with each scale；
The n convolutional neural networks that will acquire are merged, and the target integrated rolled product neural network is obtained.
6. the method according to claim 1, wherein described utilize the multiscale target sample set, to preparatory
The target integrated rolled product neural network of design is trained, and obtains Lung neoplasm detection model, comprising:
The multiscale target sample set is classified, training sample set, verifying sample set and test sample collection are obtained；
Target integrated rolled product neural network is trained and is tested using the training sample set and the verifying sample set
Card obtains integrated convolution neural network model；
The test sample collection is input to the integrated convolution neural network model, obtains test result；
According to the test result, test accuracy rate is determined；
Detect whether the test accuracy rate is greater than default accuracy rate；
If so, using the integrated convolution neural network model as Lung neoplasm detection model.
7. a kind of detection method of Lung neoplasm characterized by comprising
Obtain image to be detected；
Based on Lung neoplasm detection model, described image to be detected is inputted, testing result is obtained；
The detection model construction method building that the Lung neoplasm detection model passes through any one of claim 1-6 Lung neoplasm.
8. a kind of detection model construction device of Lung neoplasm characterized by comprising
Determining module, for determining that sense is emerging according to the lung images and the corresponding description information of the lung images obtained in advance
Interesting area image；
Processing module obtains multiscale target sample set for handling the region of interest area image；
Training module, for utilize the multiscale target sample set, to be pre-designed target integrated rolled product neural network into
Row training, obtains Lung neoplasm detection model.
9. device according to claim 8, which is characterized in that the determining module includes: first processing units, analysis list
Member and image determination unit；
The first processing units obtain target figure for carrying out image enhancement processing to the lung images obtained in advance
Picture；
The analytical unit is obtained for analyzing the corresponding description information of the lung images obtained in advance
Coordinate information；
Described image determination unit, for determining the area-of-interest figure according to the target image and the coordinate information
Picture.
10. a kind of detection device of Lung neoplasm characterized by comprising image collection module and detection module；
Described image obtains module, for obtaining image to be detected；
The detection module inputs described image to be detected, obtains testing result for being based on Lung neoplasm detection model；It is described
The detection model construction method building that Lung neoplasm detection model passes through any one of claim 1-6 Lung neoplasm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910384871.0A CN110084237A (en) | 2019-05-09 | 2019-05-09 | Detection model construction method, detection method and the device of Lung neoplasm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910384871.0A CN110084237A (en) | 2019-05-09 | 2019-05-09 | Detection model construction method, detection method and the device of Lung neoplasm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110084237A true CN110084237A (en) | 2019-08-02 |
Family
ID=67419543
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910384871.0A Pending CN110084237A (en) | 2019-05-09 | 2019-05-09 | Detection model construction method, detection method and the device of Lung neoplasm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110084237A (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110706185A (en) * | 2019-09-30 | 2020-01-17 | 上海数禾信息科技有限公司 | Image processing method and device, equipment and storage medium |
| CN110717518A (en) * | 2019-09-10 | 2020-01-21 | 北京深睿博联科技有限责任公司 | Persistent lung nodule identification method and device based on 3D convolutional neural network |
| CN110909615A (en) * | 2019-10-28 | 2020-03-24 | 西安交通大学 | Target detection method based on multi-scale input mixed perception neural network |
| CN111126274A (en) * | 2019-12-24 | 2020-05-08 | 深圳市检验检疫科学研究院 | Method, device, equipment and medium for detecting inbound target population |
| CN111292396A (en) * | 2020-01-16 | 2020-06-16 | 武汉轻工大学 | Image sample set generation method, device and apparatus and storage medium |
| CN111369536A (en) * | 2020-03-05 | 2020-07-03 | 上海市肺科医院（上海市职业病防治院） | Method and system for generating lung mediastinal lymph node detection model |
| CN111754453A (en) * | 2020-05-11 | 2020-10-09 | 佛山市第四人民医院(佛山市结核病防治所) | Pulmonary tuberculosis detection method and system based on chest radiography image and storage medium |
| CN111862075A (en) * | 2020-07-30 | 2020-10-30 | 西南医科大学 | Lung image analysis system and method based on deep learning |
| CN112633336A (en) * | 2020-12-10 | 2021-04-09 | 重庆大学 | Pulmonary nodule identification and classification method based on multi-scale feature fusion |
| CN112802012A (en) * | 2021-03-09 | 2021-05-14 | 中南大学湘雅医院 | Pathological image detection method, pathological image detection device, computer equipment and storage medium |
| CN112884695A (en) * | 2019-11-29 | 2021-06-01 | 天津大学 | Tuberculosis chest radiograph machine interpretation method based on deep learning model fusion framework |
| CN113139928A (en) * | 2020-01-16 | 2021-07-20 | 中移(上海)信息通信科技有限公司 | Training method of pulmonary nodule detection model and pulmonary nodule detection method |
| CN113792773A (en) * | 2021-09-01 | 2021-12-14 | 北京大学口腔医学院 | Dental root canal treatment history detection model training method, image processing method and device |
| CN116883372A (en) * | 2023-07-19 | 2023-10-13 | 重庆大学 | Method and system for adaptively identifying tumor based on blood vessel region image |
| CN116935286A (en) * | 2023-08-03 | 2023-10-24 | 广州城市职业学院 | Short video identification system |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180033144A1 (en) * | 2016-09-21 | 2018-02-01 | Realize, Inc. | Anomaly detection in volumetric images |
| CN107909572A (en) * | 2017-11-17 | 2018-04-13 | 合肥工业大学 | Pulmonary nodule detection method and system based on image enhancement |
| CN108389201A (en) * | 2018-03-16 | 2018-08-10 | 北京推想科技有限公司 | The good pernicious sorting technique of Lung neoplasm based on 3D convolutional neural networks and deep learning |
| CN109146857A (en) * | 2018-08-02 | 2019-01-04 | 常州市第二人民医院 | The Lung neoplasm self-sensing method of chest CT images |
-
2019
- 2019-05-09 CN CN201910384871.0A patent/CN110084237A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180033144A1 (en) * | 2016-09-21 | 2018-02-01 | Realize, Inc. | Anomaly detection in volumetric images |
| CN107909572A (en) * | 2017-11-17 | 2018-04-13 | 合肥工业大学 | Pulmonary nodule detection method and system based on image enhancement |
| CN108389201A (en) * | 2018-03-16 | 2018-08-10 | 北京推想科技有限公司 | The good pernicious sorting technique of Lung neoplasm based on 3D convolutional neural networks and deep learning |
| CN109146857A (en) * | 2018-08-02 | 2019-01-04 | 常州市第二人民医院 | The Lung neoplasm self-sensing method of chest CT images |
Non-Patent Citations (3)
| Title |
|---|
| WEI SHEN，等: "Multi-scale Convolutional Neural Networks for Lung Nodule Classification", 《INFORMATION PROCESSING IN MEDICAL IMAGING》 * |
| 朱国策: "基于深度卷积神经网络的医学图像肺结节检测方法研究", 《中国优秀硕士学位论文全文数据库 医药卫生科技辑》 * |
| 秦品乐，等: "基于多尺度密集网络的肺结节图像检索算法", 《计算机应用》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110717518A (en) * | 2019-09-10 | 2020-01-21 | 北京深睿博联科技有限责任公司 | Persistent lung nodule identification method and device based on 3D convolutional neural network |
| CN110717518B (en) * | 2019-09-10 | 2023-06-30 | 北京深睿博联科技有限责任公司 | Continuous lung nodule recognition method and device based on 3D convolutional neural network |
| CN110706185A (en) * | 2019-09-30 | 2020-01-17 | 上海数禾信息科技有限公司 | Image processing method and device, equipment and storage medium |
| CN110909615A (en) * | 2019-10-28 | 2020-03-24 | 西安交通大学 | Target detection method based on multi-scale input mixed perception neural network |
| CN112884695A (en) * | 2019-11-29 | 2021-06-01 | 天津大学 | Tuberculosis chest radiograph machine interpretation method based on deep learning model fusion framework |
| CN111126274A (en) * | 2019-12-24 | 2020-05-08 | 深圳市检验检疫科学研究院 | Method, device, equipment and medium for detecting inbound target population |
| CN111292396A (en) * | 2020-01-16 | 2020-06-16 | 武汉轻工大学 | Image sample set generation method, device and apparatus and storage medium |
| CN113139928B (en) * | 2020-01-16 | 2024-02-23 | 中移(上海)信息通信科技有限公司 | Training method of lung nodule detection model and lung nodule detection method |
| CN111292396B (en) * | 2020-01-16 | 2023-08-29 | 武汉轻工大学 | Image sample set generation method, device, apparatus and storage medium |
| CN113139928A (en) * | 2020-01-16 | 2021-07-20 | 中移(上海)信息通信科技有限公司 | Training method of pulmonary nodule detection model and pulmonary nodule detection method |
| CN111369536A (en) * | 2020-03-05 | 2020-07-03 | 上海市肺科医院（上海市职业病防治院） | Method and system for generating lung mediastinal lymph node detection model |
| CN111754453A (en) * | 2020-05-11 | 2020-10-09 | 佛山市第四人民医院(佛山市结核病防治所) | Pulmonary tuberculosis detection method and system based on chest radiography image and storage medium |
| CN111862075A (en) * | 2020-07-30 | 2020-10-30 | 西南医科大学 | Lung image analysis system and method based on deep learning |
| CN112633336A (en) * | 2020-12-10 | 2021-04-09 | 重庆大学 | Pulmonary nodule identification and classification method based on multi-scale feature fusion |
| CN112802012A (en) * | 2021-03-09 | 2021-05-14 | 中南大学湘雅医院 | Pathological image detection method, pathological image detection device, computer equipment and storage medium |
| CN113792773A (en) * | 2021-09-01 | 2021-12-14 | 北京大学口腔医学院 | Dental root canal treatment history detection model training method, image processing method and device |
| CN113792773B (en) * | 2021-09-01 | 2024-05-14 | 北京大学口腔医学院 | Dental root canal treatment history detection model training method, image processing method and device |
| CN116883372A (en) * | 2023-07-19 | 2023-10-13 | 重庆大学 | Method and system for adaptively identifying tumor based on blood vessel region image |
| CN116935286A (en) * | 2023-08-03 | 2023-10-24 | 广州城市职业学院 | Short video identification system |
| CN116935286B (en) * | 2023-08-03 | 2024-01-09 | 广州城市职业学院 | Short video identification system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110084237A (en) | Detection model construction method, detection method and the device of Lung neoplasm | |
| CN110060774B (en) | Thyroid nodule identification method based on generative confrontation network | |
| CN108464840B (en) | Automatic detection method and system for breast lumps | |
| CN109003672A (en) | A kind of early stage of lung cancer detection classification integration apparatus and system based on deep learning | |
| Jaszcz et al. | Lung x‐ray image segmentation using heuristic red fox optimization algorithm | |
| Yao et al. | Pneumonia Detection Using an Improved Algorithm Based on Faster R‐CNN | |
| Hu et al. | A state-of-the-art survey of artificial neural networks for whole-slide image analysis: from popular convolutional neural networks to potential visual transformers | |
| Alom et al. | MitosisNet: end-to-end mitotic cell detection by multi-task learning | |
| JP2023544466A (en) | Training method and device for diagnostic model of lung adenocarcinoma and squamous cell carcinoma based on PET/CT | |
| CN114600155A (en) | Weakly supervised multitask learning for cell detection and segmentation | |
| CN112703531A (en) | Generating annotation data for tissue images | |
| Razavi et al. | MiNuGAN: Dual segmentation of mitoses and nuclei using conditional GANs on multi-center breast H&E images | |
| CN115546605A (en) | Training method and device based on image labeling and segmentation model | |
| Yuan et al. | Neural stain normalization and unsupervised classification of cell nuclei in histopathological breast cancer images | |
| CN112036298A (en) | Cell detection method based on double-segment block convolutional neural network | |
| Sarhan et al. | Utilizing transfer learning and a customized loss function for optic disc segmentation from retinal images | |
| Barrera et al. | Automatic normalized digital color staining in the recognition of abnormal blood cells using generative adversarial networks | |
| CN108447047A (en) | Acid-fast bacilli detection method and device | |
| Bittencourt et al. | Accuracy of computer-aided image analysis in the diagnosis of odontogenic cysts: A systematic review | |
| Avanija et al. | Skin Cancer Detection using Ensemble Learning | |
| Cao et al. | 3D convolutional neural networks fusion model for lung nodule detection onclinical CT scans | |
| Sari et al. | Cholesterol level detection through eye image using fractal and decision tree | |
| Sun et al. | Detection of breast tumour tissue regions in histopathological images using convolutional neural networks | |
| CN115527204A (en) | Frame-assisted tumor microenvironment analysis method for liver cancer tissue complete slides | |
| Inamdar et al. | A Novel Attention based model for Semantic Segmentation of Prostate Glands using Histopathological Images |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190802 |