CN110084810A - A kind of Lung neoplasm image detecting method, model training method, device and storage medium - Google Patents
A kind of Lung neoplasm image detecting method, model training method, device and storage medium Download PDFInfo
- Publication number
- CN110084810A CN110084810A CN201910374504.2A CN201910374504A CN110084810A CN 110084810 A CN110084810 A CN 110084810A CN 201910374504 A CN201910374504 A CN 201910374504A CN 110084810 A CN110084810 A CN 110084810A
- Authority
- CN
- China
- Prior art keywords
- lung neoplasm
- dimensional feature
- image
- dimensional
- data
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- 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
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30061—Lung
- G06T2207/30064—Lung nodule
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Quality & Reliability (AREA)
- Radiology & Medical Imaging (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
- Apparatus For Radiation Diagnosis (AREA)
Abstract
The invention discloses a kind of Lung neoplasm image detecting method, model training method, device and storage mediums, are related to medical image processing field.The model training method of the Lung neoplasm image detection includes: to be pre-processed by convolutional network to CT Lung neoplasm data, obtains Lung neoplasm characteristic image；The three-dimensional feature data of Lung neoplasm characteristic image are obtained by Xception network structure；The three-dimensional feature data of Lung neoplasm characteristic image are stacked, the one or four dimensional feature is obtained；Three dimensional convolution kernel processing is carried out to the one or four dimensional feature, obtains the two or four dimensional feature；According to the two or four dimensional feature, class probability is calculated；Class probability is the probability that each pixel is Lung neoplasm；When class probability meets the condition of convergence of model, the Lung neoplasm image detection model trained is obtained.Using two-dimensional convolution+three dimensional convolution kernel processing framework, computational efficiency is improved, the requirement to hardware resource is reduced.
Description
Technical field
This application involves medical image processing fields, instruct in particular to a kind of Lung neoplasm image detecting method, model
Practice method, apparatus and storage medium.
Background technique
Lung neoplasm is a kind of granulomatous diseases of multisystem multiple organ that the cause of disease is unknown, often invades lung, the leaching of bilateral hilus pulumonis
It fawns on, organs, the chest rate of being invaded such as eye, skin are up to 80%~90%.CT is a kind of effective work for Lung neoplasm detection
Tool, but the contradiction between CT images quantity and image department doctor's quantity makes the automatic detection algorithm of Lung neoplasm become a kind of urgent
Demand.A large amount of research has been carried out for Lung neoplasm automatic measurement technique at present, tradition is mainly put forth effort in these researchs
Machine learning and deep learning field.
The prior art generally uses two methods: the first, two-dimensional convolution extracts feature, using recirculating network to each number of plies
It is analyzed according to being associated；Second, use Three dimensional convolution network processes CT data.The first scheme can only obtain a classification knot
Fruit, the serial structure of recirculating network can reduce computational efficiency, meanwhile, recirculating network is not suitable for lung to the calculating between any each layer
The disease surveillance of this local association of tubercle.Second of technical solution needs the graphics processor of a large amount of sample and high configuration,
Training time is long.
Need a kind of Lung neoplasm image detection model training method that computational efficiency is high to solve the above problems at present.
Summary of the invention
The embodiment of the present invention be designed to provide a kind of Lung neoplasm image detecting method, model training method, device and
Storage medium solves the problems, such as that computational efficiency is low using two-dimensional convolution+three dimensional convolution kernel processing model framework.
To achieve the goals above, technical solution used in the embodiment of the present invention is as follows:
In a first aspect, the embodiment of the present invention proposes a kind of model training method of Lung neoplasm image detection, this method comprises:
CT Lung neoplasm data are pre-processed by convolutional network, obtain Lung neoplasm characteristic image；Pass through Xception network structure
Obtain the three-dimensional feature data of Lung neoplasm characteristic image；The three-dimensional feature data of Lung neoplasm characteristic image are stacked, obtain the one or four
Dimensional feature；Three dimensional convolution kernel processing is carried out to the one or four dimensional feature, obtains the two or four dimensional feature；According to the two or four dimensional feature, meter
Calculate class probability；Class probability is the probability that each pixel is Lung neoplasm；When class probability meets the condition of convergence of model,
Obtain the Lung neoplasm image detection model trained.
Optionally, Lung neoplasm spy is obtained by Xception network structure in the three-dimensional feature data of Lung neoplasm characteristic image
Before the three-dimensional feature data for levying image, this method further include: above-mentioned according to the two-dimensional image data pre-training of open source
Xception network structure.
Optionally, method further include: when class probability is unsatisfactory for the condition of convergence of model, then reacquire CT Lung neoplasm
Data execute transfer training to Xception network structure.
Specifically, three dimensional convolution kernel processing is carried out to the one or four dimensional feature, obtaining output channel number is corresponding classification class
Not Shuo the two or four dimensional feature, comprising:
One or four dimensional feature is handled with the three dimensional convolution kernel of 1x1x1, obtains the three or four that output channel number is 256
Dimensional feature；The output channel number of one or four dimensional feature is 2048.By the three or four dimensional feature successively use 5x1x3,5x3x1,
The three dimensional convolution kernel of 7x1x3,7x3x1 are handled, and the four or four dimensional feature that output channel number is 256 is obtained；Use modified line
Property unit is modified the four or four dimensional feature.Four or four dimensional feature after amendment is up-sampled, obtaining output channel number is
256 the May 4th dimensional feature.The attribute mapping layer that size is corresponded in the May 4th dimensional feature and Xception network structure is carried out
Matrix is added, and obtains the six or four dimensional feature that output channel number is 256.Six or four dimensional feature, two the three-dimensional of 3x3x3 are rolled up
Product core is handled, and the seven or four dimensional feature that output channel number is 256 is obtained；Using linear revision unit to the seven or four dimensional feature
It is modified.Revised seven or four dimensional feature is handled with the three dimensional convolution kernel of 1x1x1, acquisition output channel number is phase
Two or four dimensional feature of the class categories number answered.
Second aspect, the embodiment of the present invention also propose a kind of Lung neoplasm image detecting method, and this method is applied to by upper
State the Lung neoplasm image detection model trained that the model training method of Lung neoplasm image detection obtains.Lung neoplasm image inspection
Survey method includes: to be divided into N/2 according to by convolutional network successively to extract adjacent N layer CT image data to all CT picture numbers
It is completed according to extracting, and CT image data is pre-processed, obtain Lung neoplasm characteristic image.It is obtained by Xception network structure
Take the three-dimensional feature data of Lung neoplasm characteristic image；The three-dimensional feature data of Lung neoplasm characteristic image are stacked, it is four-dimensional to obtain first
Feature；Three dimensional convolution kernel processing is carried out to the one or four dimensional feature, obtains the two or four dimensional feature.According to the two or four dimensional feature, calculate
Class probability；Class probability is the probability that each pixel is Lung neoplasm.
Optionally, this method further include: according to class probability, N layers of CT image data are predicted, obtain Lung neoplasm inspection
The prediction result of survey；The average value of the prediction result of N layer CT image data adjacent twice is taken to detect as Lung neoplasm final
Prediction result.
The third aspect, the embodiment of the present invention also propose a kind of model training apparatus of Lung neoplasm image detection, comprising: first
Processing module and first obtains module；First obtains module, for obtaining CT Lung neoplasm data by convolutional network；First processing
Module, for pre-processing CT Lung neoplasm data；First, which obtains module, obtains Lung neoplasm characteristic image.First acquisition module is also used to
The three-dimensional feature data of Lung neoplasm characteristic image are obtained by Xception network structure；First processing module is also used to stack lung
The three-dimensional feature data of tubercle characteristic image, the first acquisition module are also used to obtain the one or four dimensional feature.First processing module is also
For carrying out three dimensional convolution kernel processing to the one or four dimensional feature, the first acquisition module is also used to obtain the two or four dimensional feature；First
Processing module is also used to calculate class probability according to the two or four dimensional feature；Class probability is that each pixel is the several of Lung neoplasm
Rate；When class probability meets the condition of convergence of model, the first acquisition module is also used to obtain the Lung neoplasm image inspection trained
Survey model.
Optionally, first processing module is also used to the two-dimensional image data pre-training Xception network knot according to open source
Structure.
Optionally, when class probability is unsatisfactory for the condition of convergence of model, the first acquisition module is also used to reacquire CT
Lung neoplasm data；First processing module is also used to execute transfer training to Xception network structure.
Fourth aspect, the embodiment of the present invention also propose a kind of Lung neoplasm image detection device, comprising: Second processing module and
Second obtains module.
Second obtains module, successively extracts adjacent N layer CT image data for being divided into N/2 according to by convolutional network
It extracts and completes to all CT image datas；Second processing module is used for pre-treatment CT image data；Second acquisition module is also used to
Obtain Lung neoplasm characteristic image.
Second processing module is also used to obtain Lung neoplasm feature according to Lung neoplasm characteristic image and Xception network structure
The three-dimensional feature data of image, stack the three-dimensional feature data of Lung neoplasm characteristic image, and the second acquisition module is also used to obtain the
One or four dimensional features.Second processing module is also used to carry out the one or four dimensional feature three dimensional convolution kernel processing, and second obtains module also
For obtaining the two or four dimensional feature.Second processing module is also used to calculate class probability according to the two or four dimensional feature；Class probability
It is the probability that each pixel is Lung neoplasm.
Optionally, Second processing module is also used to predict N tomographic image data according to class probability, obtain lung knot
Save the prediction result of detection；Second processing module is also used to take being averaged for the prediction result of N layer CT image data adjacent twice
It is worth the final prediction result detected as Lung neoplasm.
5th aspect, the embodiment of the present invention also propose a kind of computer readable storage medium, are stored thereon with computer journey
Sequence when computer program is read out by the processor and runs, realizes above-mentioned Lung neoplasm image detecting method and the inspection of Lung neoplasm image
The model training method of survey.
The invention discloses a kind of Lung neoplasm image detecting method, model training method, device and storage mediums, are related to curing
Learn image processing field.The model training method of the Lung neoplasm image detection includes: by convolutional network to CT Lung neoplasm data
It is pre-processed, obtains Lung neoplasm characteristic image；The three-dimensional for obtaining Lung neoplasm characteristic image by Xception network structure is special
Levy data；The three-dimensional feature data of Lung neoplasm characteristic image are stacked, the one or four dimensional feature is obtained；Three are carried out to the one or four dimensional feature
Convolution kernel processing is tieed up, the two or four dimensional feature is obtained；According to the two or four dimensional feature, class probability is calculated；Class probability is each picture
Vegetarian refreshments is the probability of Lung neoplasm；When class probability meets the condition of convergence of model, the Lung neoplasm image detection trained is obtained
Model.Using two-dimensional convolution+three dimensional convolution kernel framework, computational efficiency is improved, reduces the requirement to hardware resource.
Other features and advantages of the present invention will be illustrated in subsequent specification, also, partly be become from specification
It is clear that by implementing understanding of the embodiment of the present invention.The objectives and other advantages of the invention can be by written theory
Specifically noted structure is achieved and obtained in bright book, claims and attached drawing.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, below will be to needed in the embodiment attached
Figure is briefly described, it should be understood that the following drawings illustrates only certain embodiments of the present invention, therefore is not construed as pair
The restriction of range for those of ordinary skill in the art without creative efforts, can also be according to this
A little attached drawings obtain other relevant attached drawings.
Fig. 1 shows a kind of model training method schematic diagram of Lung neoplasm image detection provided by the embodiment of the present invention.
Fig. 2 shows a kind of schematic diagrames of pre-training Xception network structure provided by the embodiment of the present invention.
Fig. 3 shows a kind of Three dimensional convolution processing schematic provided by the embodiment of the present invention.
Fig. 4 shows a kind of pulmonary nodule detection method schematic diagram provided by the embodiment of the present invention.
Fig. 5 shows a kind of schematic diagram of Optimization Prediction result provided by the embodiment of the present invention.
Fig. 6 shows a kind of model training apparatus schematic diagram of Lung neoplasm image detection provided in an embodiment of the present invention.
Fig. 7 shows a kind of Lung neoplasm image detection device of proposition of the embodiment of the present invention.
Icon: the model training apparatus of 300- Lung neoplasm image detection, 301 first obtain module 301, the processing of 302- first
Module, 400- Lung neoplasm image detection device, 401- second obtain module, 402- Second processing module.
Specific embodiment
Below in conjunction with attached drawing in the embodiment of the present invention, technical solution in the embodiment of the present invention carries out clear, complete
Ground description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.Usually exist
The component of the embodiment of the present invention described and illustrated in attached drawing can be arranged and be designed with a variety of different configurations herein.Cause
This, is not intended to limit claimed invention to the detailed description of the embodiment of the present invention provided in the accompanying drawings below
Range, but it is merely representative of selected embodiment of the invention.Based on the embodiment of the present invention, those skilled in the art are not doing
Every other embodiment obtained under the premise of creative work out, shall fall within the protection scope of the present invention.
It should also be noted that similar label and letter indicate similar terms in following attached drawing, therefore, once a certain Xiang Yi
It is defined in a attached drawing, does not then need that it is further defined and explained in subsequent attached drawing.Meanwhile in the disclosure
In description, term " first ", " second " etc. are only used for distinguishing description, are not understood to indicate or imply relative importance.
In order to improve the model training efficiency of Lung neoplasm image detection, the present invention proposes a kind of mould of Lung neoplasm image detection
Type training method is a kind of model training method of Lung neoplasm image detection provided by the embodiment of the present invention referring to Fig. 1, Fig. 1
Schematic diagram, this method comprises:
Step 100 pre-processes CT Lung neoplasm data by convolutional network, obtains Lung neoplasm characteristic image.
It can be by the CT Lung neoplasm data that convolutional network obtains and randomly select a certain number of data, specifically extract
Data bulk can be set according to actual hardware resource: in actual use, the 32 layer image data randomly selected
A preferable training can be realized.The size of the Lung neoplasm characteristic image of acquisition can be 496x496x3.
Step 101, the three-dimensional feature data that Lung neoplasm characteristic image is obtained by Xception network structure.
Step 102, the three-dimensional feature data for stacking Lung neoplasm characteristic image obtain the one or four dimensional feature.
The change of the output channel number of data may be implemented into the treatment process of the one or four dimensional feature in Lung neoplasm characteristic image
Change.
Step 103 carries out three dimensional convolution kernel processing to the one or four dimensional feature, obtains the two or four dimensional feature.
Step 104, according to the two or four dimensional feature, calculate class probability.
Above-mentioned class probability is the probability that each pixel is Lung neoplasm.
Step 105 judges whether class probability meets the condition of convergence of model.
The condition of convergence of model can be frequency of training, or the accuracy rate of Lung neoplasm is measured according to class probability, can also
To use statistical F1 score, can also be it is average friendship and than (Mean Intersection over Union, referred to as:
MIoU)。
When class probability meets the condition of convergence of model, 106 are thened follow the steps；
The Lung neoplasm image detection model that step 106, acquisition have been trained.
When class probability is unsatisfactory for the condition of convergence of model, then CT Lung neoplasm data are reacquired, to Xception net
Network structure executes transfer training；Re-execute above-mentioned steps 100, step 101, step 102, step 103, step 104 and step
Rapid 105, until model is restrained.
Using two-dimensional convolution+three dimensional convolution kernel processing model training framework, most feature extraction work is by two dimension
Convolutional network is completed, and calculation amount is small, improves the computational efficiency of model training.Meanwhile this method is also in the hard of lower configuration
Model training can also be carried out under part environment.The image number using only 32 CT tomographies may be implemented in model training in network
According to primary training is completed, the computational efficiency of model is improved, further reduced the requirement to hardware configuration.
Optionally, in order to obtain a preferably original template in the training process, before step 101, it is added one
Pre-training process, such as Fig. 2, Fig. 2 are a kind of schematic diagrames of pre-training Xception network structure provided by the embodiment of the present invention.
The model training method of above-mentioned Lung neoplasm image detection can also include:
Step 107, the two-dimensional image data pre-training Xception network structure according to open source.
Specifically, pre-training Xception network structure can be its argument section.
Optionally, in order to obtain a preferable three dimensional convolution kernel as a result, to the mode in the cards of above-mentioned steps 10 into
Row explanation, such as Fig. 3, Fig. 3 are a kind of Three dimensional convolution processing schematics provided by the embodiment of the present invention.When the one or four dimensional feature
When output channel number is 2048, the process of three dimensional convolution kernel processing includes:
Step 103-1, the one or four dimensional feature is handled with the three dimensional convolution kernel of 1x1x1, obtaining output channel number is
256 the three or four dimensional feature.
Step 103-2, by the three or four dimensional feature successively use the three dimensional convolution kernel of 5x1x3,5x3x1,7x1x3,7x3x1 into
Row processing obtains the four or four dimensional feature that output channel number is 256；The four or four dimensional feature is repaired using amendment linear unit
Just.
Step 103-3, revised four or four dimensional feature is up-sampled, obtains the 5th that output channel number is 256
Four dimensional features.
Up-sampling can be realized by bilinear interpolation, in order to be twice the spatial resolution increasing of data.
Step 103-4, the attribute mapping layer that size is corresponded in the May 4th dimensional feature and Xception network structure is carried out
Matrix is added, and obtains the six or four dimensional feature that output channel number is 256.
Since above-mentioned steps 101 have carried out two-dimensional convolution, above-mentioned matrix in three-dimensional primarily to carry out convolution.
Step 103-5, the six or four dimensional feature is handled with the three dimensional convolution kernel of two 3x3x3, obtains output channel
The seven or four dimensional feature that number is 256；The seven or four dimensional feature is modified using linear revision unit.
Step 103-6, revised seven or four dimensional feature is handled with the three dimensional convolution kernel of 1x1x1, obtains output
Port number is the two or four dimensional feature of corresponding class categories number.
Step 103-1 to step 103-6 realizes the conversion and processing of 2 d-to-3 d, and obtains output channel number and be
Two or four dimensional feature of class categories number, by calculating available class probability, i.e., each pixel is the probability of Lung neoplasm.
At present, three dimensional convolution kernel is to be most suitable for the mode of processing CT image data at present.
The model training method of above-mentioned Lung neoplasm image detection improves computational efficiency, reduces and wants to hardware resource
It asks, manufacturing cost can be reduced and shortens the detection time of Lung neoplasm.
In order to realize that Lung neoplasm detects, the present invention proposes that a kind of Lung neoplasm image detecting method, such as Fig. 4, Fig. 4 are the present invention
A kind of pulmonary nodule detection method schematic diagram provided by embodiment.The Lung neoplasm image detecting method includes:
Step 200 is divided into N/2 according to by convolutional network and successively extracts adjacent N layer CT image data to all CT
Image data, which extracts, to be completed, and is pre-processed to CT image data, and Lung neoplasm characteristic image is obtained.
Step 201, the three-dimensional feature data that Lung neoplasm characteristic image is obtained by Xception network structure.
Step 202, the three-dimensional feature data for stacking Lung neoplasm characteristic image obtain the one or four dimensional feature.
Step 203 carries out three dimensional convolution kernel processing to the one or four dimensional feature, obtains the two or four dimensional feature.
Step 204, according to the two or four dimensional feature, calculate class probability.
By obtaining every CT image, the detection to entire CT image is realized, is successively extracted according to interval N/2 to be detected
CT image is detected and is judged, such detection mode improves computational efficiency, reduces the requirement to hardware.
Optionally, in order to which the result for detecting Lung neoplasm is more accurate, the process of optimum results on the basis of Fig. 1, such as
Fig. 5, Fig. 5 are a kind of schematic diagrames of Optimization Prediction result provided by the embodiment of the present invention.The Lung neoplasm image detecting method is also
Include:
Step 205, according to class probability, every N layers of CT image data is predicted, the prediction result of Lung neoplasm is obtained.
Step 206 takes the average value of the prediction result of N layer CT image data adjacent twice to detect most as Lung neoplasm
Whole prediction result.
For example, as N=32, successively extract it is adjacent 32 (specific data can be set according to actual hardware resource,
It is the bigger the better) a sample is for detecting, and intermediate sample is using there is the overlapping region comprising 16 samples, that is to say, that the sample of extraction
This serial number [0 ..., 31], [16 ..., 47], [32 ..., 63] etc., there are two predicted values for middle section sample (such as
[16 ..., 31] number tomography first time can be drawn into, can be also drawn into for the second time) the average value used twice as most
Terminate fruit.
It is detected by adjacent CT image data, in conjunction with the average treatment to prediction result, can further improve lung
The accuracy rate of nodule detection.
In order to realize the model training method of above-mentioned Lung neoplasm image detection, the embodiment of the present invention also proposes a kind of Lung neoplasm
The model training apparatus of image detection, such as Fig. 6, Fig. 6 are a kind of models of Lung neoplasm image detection provided in an embodiment of the present invention
Training device schematic diagram.The model training apparatus 300 of Lung neoplasm image detection includes: the first acquisition module 301 and the first processing
Module 302.
First obtains module 301, for obtaining CT Lung neoplasm data by convolutional network.
First processing module 302, for pre-processing CT Lung neoplasm data；First, which obtains module 301, obtains Lung neoplasm feature
Image.
First acquisition module 301 is also used to obtain the three-dimensional feature of Lung neoplasm characteristic image by Xception network structure
Data；First processing module 302 is also used to stack the three-dimensional feature data of Lung neoplasm characteristic image, and first obtains module 301 also
For obtaining the one or four dimensional feature.First processing module 302 is also used to carry out three dimensional convolution kernel processing to the one or four dimensional feature, the
One acquisition module 301 is also used to obtain the two or four dimensional feature.First processing module 302 is also used to according to the two or four dimensional feature, meter
Calculate class probability；Class probability is the probability that each pixel is Lung neoplasm；When class probability meets the condition of convergence of model,
First acquisition module 301 is also used to obtain the Lung neoplasm image detection model trained.
Model is trained using two-dimensional convolution+three dimensional convolution kernel processing network structure, more meets processing CT image
The requirement of data, while the computational efficiency of model training can be improved, reduce the configuration requirement to hardware.
Optionally, in order to guarantee model training can obtain one preferably as a result, first processing module 302 is also used to basis
The two-dimensional image data pre-training Xception network structure of open source.
Pre-training can enable Xception network structure obtain more accurate result when handling data.
Optionally, when class probability is unsatisfactory for the condition of convergence of model, the first acquisition module 301 is also used to reacquire
CT Lung neoplasm data；First processing module 302 is also used to execute transfer training to Xception network structure.
The use of transfer training is that Xception network structure is trained again, can further improve computational efficiency,
The requirement to hardware is reduced, the defect that sample size is less in actual conditions is overcome.
In order to realize above-mentioned Lung neoplasm image detecting method, the embodiment of the present invention also proposes a kind of Lung neoplasm image detection
Device, such as Fig. 7, Fig. 7 are a kind of Lung neoplasm image detection device that the embodiment of the present invention proposes, Lung neoplasm image detection device
400 include: the second acquisition module 401 and Second processing module 402；
Second obtains module 401, successively extracts adjacent N layer CT image for being divided into N/2 according to by convolutional network
Data are extracted to all CT image datas and are completed；Second processing module 402 is used for pre-treatment CT image data；Second obtains mould
Block 401 is also used to obtain Lung neoplasm characteristic image.
Second acquisition module 401 is also used to obtain the three-dimensional feature of Lung neoplasm characteristic image by Xception network structure
Data；Second processing module 402 is also used to stack the three-dimensional feature data of Lung neoplasm characteristic image, and second obtains module 401 also
For obtaining the one or four dimensional feature；
Second processing module 402 is also used to carry out the one or four dimensional feature three dimensional convolution kernel processing, and second obtains module 401
It is also used to obtain the two or four dimensional feature；Second processing module 402 is also used to calculate class probability according to the two or four dimensional feature；Class
Other probability is the probability that each pixel is Lung neoplasm.
Above-mentioned Lung neoplasm image detection device has used two-dimensional convolution+three dimensional convolution kernel processing model framework, improves
The computational efficiency of model, while to realize that same function provides possibility on the hardware of low configuration.
In order to improve the accuracy rate of Lung neoplasm image detection, Second processing module 402 is also used to according to class probability, to every
N tomographic image data are predicted, the prediction result of Lung neoplasm detection is obtained.
Since every N/2 layer data can be extracted twice, Second processing module 402 is also used to take N layer CT adjacent twice to scheme
The final prediction result detected as the average value of the prediction result of data as Lung neoplasm.
Even preferably Lung neoplasm prediction same as the prior art can also be obtained on the hardware of lower configuration by realizing
As a result, improving the computational efficiency of Lung neoplasm image detection.
The embodiment of the present invention also proposes a kind of computer readable storage medium, is stored thereon with computer program, computer
When program is read out by the processor and runs, above-mentioned Lung neoplasm image detecting method and model training method are realized.
In conclusion the invention discloses a kind of Lung neoplasm image detecting method, model training method, device and storages to be situated between
Matter is related to medical image processing field.The model training method of the Lung neoplasm image detection includes: by convolutional network to CT lung
Tubercle data are pre-processed, and Lung neoplasm characteristic image is obtained；Lung neoplasm characteristic image is obtained by Xception network structure
Three-dimensional feature data；The three-dimensional feature data of Lung neoplasm characteristic image are stacked, the one or four dimensional feature is obtained；It is four-dimensional special to first
Sign carries out three dimensional convolution kernel processing, obtains the two or four dimensional feature；According to the two or four dimensional feature, class probability is calculated；Class probability
It is the probability that each pixel is Lung neoplasm；When class probability meets the condition of convergence of model, the Lung neoplasm trained is obtained
Image detection model.Using two-dimensional convolution+three dimensional convolution kernel framework, computational efficiency is improved, is reduced to hardware resource
It is required that.
The foregoing is merely alternative embodiments of the invention, are not intended to restrict the invention, for the skill of this field
For art personnel, the invention may be variously modified and varied.All within the spirits and principles of the present invention, made any to repair
Change, equivalent replacement, improvement etc., should all be included in the protection scope of the present invention.
Claims (10)
1. a kind of model training method of Lung neoplasm image detection, which is characterized in that the described method includes:
CT Lung neoplasm data are pre-processed by convolutional network, obtain Lung neoplasm characteristic image；
The three-dimensional feature data of the Lung neoplasm characteristic image are obtained by Xception network structure；
The three-dimensional feature data of the Lung neoplasm characteristic image are stacked, the one or four dimensional feature is obtained；
Three dimensional convolution kernel processing is carried out to the one or four dimensional feature, obtains the two or four dimensional feature；
According to the two or four dimensional feature, class probability is calculated；The class probability is the probability that each pixel is Lung neoplasm；
When the class probability meets the condition of convergence of model, the Lung neoplasm image detection model trained is obtained.
2. the model training method of Lung neoplasm image detection according to claim 1, which is characterized in that pass through described
Before Xception network structure obtains the three-dimensional feature data of the Lung neoplasm characteristic image, the method also includes:
According to Xception network structure described in the two-dimensional image data pre-training of open source.
3. the model training method of Lung neoplasm image detection according to claim 2, which is characterized in that the method is also wrapped
It includes:
When the class probability is unsatisfactory for the condition of convergence of model, then the CT Lung neoplasm data are reacquired, to described
Xception network structure executes transfer training.
4. the model training method of Lung neoplasm image detection according to claim 1, which is characterized in that described to described
One or four dimensional features carry out three dimensional convolution kernel processing, obtain the two or four dimensional feature that output channel number is corresponding class categories number,
Include:
One or four dimensional feature is handled with the three dimensional convolution kernel of 1x1x1, obtains the three or four that output channel number is 256
Dimensional feature；The output channel number of one or four dimensional feature is 2048；
Three or four dimensional feature is successively handled using the three dimensional convolution kernel of 5x1x3,5x3x1,7x1x3,7x3x1, is obtained
Taking output channel number is 256 the four or four dimensional feature；The four or four dimensional feature is modified using amendment linear unit；
Four or four dimensional feature described after amendment is up-sampled, the May 4th dimensional feature that output channel number is 256 is obtained；
The attribute mapping layer that size is corresponded in the May 4th dimensional feature and the Xception network structure is subjected to matrix phase
Add, obtains the six or four dimensional feature that output channel number is 256；
Six or four dimensional feature is handled with the three dimensional convolution kernel of two 3x3x3, obtains that output channel number is 256
Seven or four dimensional features；The seven or four dimensional feature is modified using the linear revision unit；
Revised seven or four dimensional feature is handled with the three dimensional convolution kernel of 1x1x1, acquisition output channel number is phase
Two or four dimensional feature of the class categories number answered.
5. a kind of Lung neoplasm image detecting method, which is characterized in that the method is applied to appoint by the claim 1-4
The Lung neoplasm image detection model trained that method described in meaning one obtains, which comprises
It is extracted by convolutional network according to being divided into N/2 and successively extract adjacent N layer CT image data to all CT image datas
It completes, and the CT image data is pre-processed, obtain Lung neoplasm characteristic image；
The three-dimensional feature data of the Lung neoplasm characteristic image are obtained by Xception network structure；
The three-dimensional feature data of the Lung neoplasm characteristic image are stacked, the one or four dimensional feature is obtained；
Three dimensional convolution kernel processing is carried out to the one or four dimensional feature, obtains the two or four dimensional feature；
According to the two or four dimensional feature, class probability is calculated；The class probability is the probability that each pixel is Lung neoplasm.
6. Lung neoplasm image detecting method according to claim 5, which is characterized in that the method also includes:
According to the class probability, the N layers of CT image data is predicted, obtains the prediction result of Lung neoplasm detection；
The average value of the prediction result of the N layers of adjacent twice CT image data is taken to detect as Lung neoplasm final pre-
Survey result.
7. a kind of model training apparatus of Lung neoplasm image detection characterized by comprising first processing module and first obtains
Module；
Described first obtains module, for obtaining CT Lung neoplasm data by convolutional network；
The first processing module, for pre-processing the CT Lung neoplasm data；Described first, which obtains module, obtains Lung neoplasm spy
Levy image；
The three-dimensional that the first acquisition module is also used to obtain the Lung neoplasm characteristic image by Xception network structure is special
Levy data；
The first processing module is also used to stack the three-dimensional feature data of the Lung neoplasm characteristic image, and described first obtains mould
Block is also used to obtain the one or four dimensional feature；
The first processing module is also used to carry out the one or four dimensional feature three dimensional convolution kernel processing, and described first obtains mould
Block is also used to obtain the two or four dimensional feature；
The first processing module is also used to calculate class probability according to the two or four dimensional feature；The class probability is every
A pixel is the probability of Lung neoplasm；
When the class probability meets the condition of convergence of model, the first acquisition module is also used to obtain the lung knot trained
Save image detection model.
8. a kind of Lung neoplasm image detection device characterized by comprising Second processing module and second obtains module；
Described second obtains module, successively extracts adjacent N layer CT image data for being divided into N/2 according to by convolutional network
It extracts and completes to all CT image datas；
The Second processing module, for pre-processing the CT image data；The second acquisition module is also used to obtain lung knot
Save characteristic image；
The three-dimensional that the second acquisition module is also used to obtain the Lung neoplasm characteristic image by Xception network structure is special
Levy data；
The Second processing module is also used to stack the three-dimensional feature data of the Lung neoplasm characteristic image, and described second obtains mould
Block is also used to obtain the one or four dimensional feature；
The Second processing module is also used to carry out the one or four dimensional feature three dimensional convolution kernel processing, and described second obtains mould
Block is also used to obtain the two or four dimensional feature；
The Second processing module is also used to calculate class probability according to the two or four dimensional feature；The class probability is every
A pixel is the probability of Lung neoplasm.
9. Lung neoplasm image detection device according to claim 8, which is characterized in that the Second processing module is also used to
According to the class probability, the N layers of CT image data is predicted, obtains the prediction result of Lung neoplasm detection；Described
Two processing modules are also used to take the average value of the prediction result of the N layers of adjacent twice CT image data as Lung neoplasm
The final prediction result of detection.
10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer program
When being read out by the processor and running, as the method according to claim 1 to 6 is realized.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910374504.2A CN110084810B (en) | 2019-05-06 | 2019-05-06 | Pulmonary nodule image detection method, model training method, device and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910374504.2A CN110084810B (en) | 2019-05-06 | 2019-05-06 | Pulmonary nodule image detection method, model training method, device and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110084810A true CN110084810A (en) | 2019-08-02 |
| CN110084810B CN110084810B (en) | 2021-11-05 |
Family
ID=67418839
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910374504.2A Active CN110084810B (en) | 2019-05-06 | 2019-05-06 | Pulmonary nodule image detection method, model training method, device and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110084810B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110751183A (en) * | 2019-09-24 | 2020-02-04 | 东软集团股份有限公司 | Image data classification model generation method, image data classification method and device |
| CN112529894A (en) * | 2020-12-22 | 2021-03-19 | 徐州医科大学 | Thyroid nodule diagnosis method based on deep learning network |
| CN113808111A (en) * | 2021-09-18 | 2021-12-17 | 广州幻境科技有限公司 | Three-dimensional virtual reconstruction method and system for medical image |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101515365A (en) * | 2009-03-25 | 2009-08-26 | 沈阳东软医疗系统有限公司 | Method for automatically separating adherent hyaline-vascular type lung nodule in CT image |
| CN107016665A (en) * | 2017-02-16 | 2017-08-04 | 浙江大学 | A kind of CT pulmonary nodule detection methods based on depth convolutional neural networks |
| CN107563383A (en) * | 2017-08-24 | 2018-01-09 | 杭州健培科技有限公司 | A kind of medical image auxiliary diagnosis and semi-supervised sample generation system |
| CN108537784A (en) * | 2018-03-30 | 2018-09-14 | 四川元匠科技有限公司 | A kind of CT figure pulmonary nodule detection methods based on deep learning |
| CN108921195A (en) * | 2018-05-31 | 2018-11-30 | 沈阳东软医疗系统有限公司 | A kind of Lung neoplasm image-recognizing method neural network based and device |
| CN109003260A (en) * | 2018-06-28 | 2018-12-14 | 深圳视见医疗科技有限公司 | CT image pulmonary nodule detection method, device, equipment and readable storage medium storing program for executing |
| CN109492547A (en) * | 2018-10-24 | 2019-03-19 | 腾讯科技（深圳）有限公司 | A kind of tubercle recognition methods, device and storage medium |
| CN109685768A (en) * | 2018-11-28 | 2019-04-26 | 心医国际数字医疗系统(大连)有限公司 | Lung neoplasm automatic testing method and system based on lung CT sequence |
-
2019
- 2019-05-06 CN CN201910374504.2A patent/CN110084810B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101515365A (en) * | 2009-03-25 | 2009-08-26 | 沈阳东软医疗系统有限公司 | Method for automatically separating adherent hyaline-vascular type lung nodule in CT image |
| CN107016665A (en) * | 2017-02-16 | 2017-08-04 | 浙江大学 | A kind of CT pulmonary nodule detection methods based on depth convolutional neural networks |
| CN107563383A (en) * | 2017-08-24 | 2018-01-09 | 杭州健培科技有限公司 | A kind of medical image auxiliary diagnosis and semi-supervised sample generation system |
| CN108537784A (en) * | 2018-03-30 | 2018-09-14 | 四川元匠科技有限公司 | A kind of CT figure pulmonary nodule detection methods based on deep learning |
| CN108921195A (en) * | 2018-05-31 | 2018-11-30 | 沈阳东软医疗系统有限公司 | A kind of Lung neoplasm image-recognizing method neural network based and device |
| CN109003260A (en) * | 2018-06-28 | 2018-12-14 | 深圳视见医疗科技有限公司 | CT image pulmonary nodule detection method, device, equipment and readable storage medium storing program for executing |
| CN109492547A (en) * | 2018-10-24 | 2019-03-19 | 腾讯科技（深圳）有限公司 | A kind of tubercle recognition methods, device and storage medium |
| CN109685768A (en) * | 2018-11-28 | 2019-04-26 | 心医国际数字医疗系统(大连)有限公司 | Lung neoplasm automatic testing method and system based on lung CT sequence |
Non-Patent Citations (2)
| Title |
|---|
| RAUL VICTOR MEDEIROS DA NÓBREGA: "Lung Nodule Classification via Deep Transfer Learning in CT Lung Images", 《2018 IEEE 31ST INTERNATIONAL SYMPOSIUM ON COMPUTER-BASED MEDICAL SYSTEMS (CBMS)》 * |
| 杨晶晶等: "基于深度卷积神经网络算法的肺结节检测模型", 《数学建模及其应用》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110751183A (en) * | 2019-09-24 | 2020-02-04 | 东软集团股份有限公司 | Image data classification model generation method, image data classification method and device |
| CN112529894A (en) * | 2020-12-22 | 2021-03-19 | 徐州医科大学 | Thyroid nodule diagnosis method based on deep learning network |
| CN112529894B (en) * | 2020-12-22 | 2022-02-15 | 徐州医科大学 | Thyroid nodule diagnosis method based on deep learning network |
| CN113808111A (en) * | 2021-09-18 | 2021-12-17 | 广州幻境科技有限公司 | Three-dimensional virtual reconstruction method and system for medical image |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110084810B (en) | 2021-11-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111739075B (en) | Deep network lung texture recognition method combining multi-scale attention | |
| CN104484886B (en) | A kind of dividing method and device of MR images | |
| US20190361972A1 (en) | Method, apparatus, device for table extraction based on a richly formatted document and medium | |
| CN108428229A (en) | It is a kind of that apparent and geometric properties lung's Texture Recognitions are extracted based on deep neural network | |
| CN108603922A (en) | Automatic cardiac volume is divided | |
| CN107590797A (en) | CT image pulmonary nodule detection method based on three-dimensional residual error neural network | |
| CN110084810A (en) | A kind of Lung neoplasm image detecting method, model training method, device and storage medium | |
| CN109886944B (en) | A kind of white matter high signal intensity detection and localization method based on multichannel chromatogram | |
| CN112329871B (en) | Pulmonary nodule detection method based on self-correction convolution and channel attention mechanism | |
| CN113221731B (en) | Multi-scale remote sensing image target detection method and system | |
| WO2020093435A1 (en) | Abdominal image segmentation method, computer device, and storage medium | |
| CN113468996A (en) | Camouflage object detection method based on edge refinement | |
| CN110930378A (en) | Emphysema image processing method and system based on low data demand | |
| CN116805360B (en) | Obvious target detection method based on double-flow gating progressive optimization network | |
| CN111680755A (en) | Medical image recognition model construction method, medical image recognition device, medical image recognition medium and medical image recognition terminal | |
| CN112287884A (en) | Examination abnormal behavior detection method and device and computer readable storage medium | |
| CN117611599B (en) | Blood vessel segmentation method and system integrating centre line diagram and contrast enhancement network | |
| Wang et al. | SLMS-SSD: Improving the balance of semantic and spatial information in object detection | |
| CN114511759A (en) | Method and system for identifying categories and determining characteristics of skin state images | |
| CN116664590B (en) | Automatic segmentation method and device based on dynamic contrast enhancement magnetic resonance image | |
| CN117953286A (en) | Phase identification method and device, phase identification model training method and device | |
| CN112927203A (en) | Glioma patient postoperative life prediction method based on multi-sequence MRI global information | |
| Asha et al. | Segmentation of Brain Tumors using traditional Multiscale bilateral Convolutional Neural Networks | |
| CN113361494B (en) | Self-service method and self-service system based on face recognition | |
| Astono et al. | [Regular Paper] Adjacent Network for Semantic Segmentation of Liver CT Scans |
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 |