CN109522968A - A kind of focal zone detection method and system based on serial double Task Networks - Google Patents
A kind of focal zone detection method and system based on serial double Task Networks Download PDFInfo
- Publication number
- CN109522968A CN109522968A CN201811441059.9A CN201811441059A CN109522968A CN 109522968 A CN109522968 A CN 109522968A CN 201811441059 A CN201811441059 A CN 201811441059A CN 109522968 A CN109522968 A CN 109522968A
- Authority
- CN
- China
- Prior art keywords
- sample
- focal zone
- detection
- training
- candidate regions
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Biomedical Technology (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Quality & Reliability (AREA)
- Medical Informatics (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention discloses a kind of focal zone detection method based on serial double Task Networks, is related to Medical Imaging Technology field, including training part and detection part；Focal zone sample is collected first in training part, and the real focal zone of handmarking's sample, then the characteristics of detection e-learning real focal zone after pre-detection is carried out to sample, and marked lesion candidate regions, repeatedly training constructs new detection network, finally, the characteristics of using the screening real focal zone sample of e-learning and lesion candidate regions sample, the similarity relationships for finding the two export corresponding testing result according to handmarking's result, and repeatedly training constructs new screening network；In detection part, then the detection network of building completion and screening network output test result are directly utilized.Invention additionally discloses a kind of focal zone detection systems, combine with above-mentioned focal zone detection method, can efficiently and accurately detection image focal zone, false detection rate can be reduced.
Description
Technical field
The present invention relates to Medical Imaging Technology field, specifically a kind of focal zone inspection based on serial double Task Networks
Survey method and system.
Background technique
Faster R-CNN (wherein R corresponds to " Region (region) ") is examined based on deep learning R-CNN list of target
Survey the best way.Using the training of VOC2007+2012 training set, VOC2007 test set test mAP reaches 73.2%, target inspection
The speed of survey can achieve 5 frame per second.
Clinical medicine is all the key points and difficulties of people's research from ancient times to the present.In new century, scientific and technological is constantly progressive pole
The earth promotes clinical medical development.
In recent years, country pays much attention to the development of artificial intelligence technology.Artificial intelligence and clinical medical crossing research are more
It is to have obtained the extensive concern of country, enterprise and scientific research institution.Medical image processing is an important research direction in the field,
The artificial intelligence technology for being intended to study optimization, which analyzes medical image, carrys out adjuvant clinical diagnosis.Under the technical system, disease
Stove detection is the basis of analysis.However, medical image is different from general natural image, lesion and its hetero-organization usually have compared with
Big similitude, so that false detection rate is higher.
Therefore, how it to be directed to the higher problem of existing method false detection rate, designs a kind of focal zone detection method of low false detection rate,
For reducing entreprise cost, the competitiveness for enhancing product has great importance.
Summary of the invention
The present invention is directed to the demand and shortcoming of current technology development, provides a kind of disease based on serial double Task Networks
Stove area detection method and system.
A kind of focal zone detection method based on serial double task convolutional neural networks of the invention, solves above-mentioned technology and asks
Topic the technical solution adopted is as follows:
A kind of focal zone detection method based on serial double Task Networks, the detection method are serially completed by two networks,
Specifically comprise the following steps:
S100, training part:
S110, the focal zone sample of various diseases is collected as training sample, the real focal zone of handmarking's training sample；
The characteristics of S120, detection e-learning real focal zone, and pre-detection is carried out to training sample, by lesion candidate regions
It detected, repeatedly after training, complete the building of detection network；
S130, network is screened to input using real focal zone sample and lesion candidate regions sample as sample, screens network
For learning to detect the similarity relationships of real focal zone sample and lesion candidate regions sample, the real focal zone sample of handmarking
When the characteristics of with lesion candidate regions sample, is identical, screening network output result queue be 1, the real focal zone sample of handmarking with
When the characteristics of lesion candidate regions sample difference, screening network output result queue be 0, to multiple samples to being trained after, it is complete
At the building of screening network；
S200, detection part:
S210, sample to be detected is inputted to the detection network progress pre-detection that building is completed, detection network identity is to be detected
The lesion candidate regions of sample；
The characteristics of S220, the screening e-learning real focal zone sample completed based on building, real disease is also learnt
The similarity relationships of stove area sample and lesion candidate regions sample have constructed the sample to be detected for being marked with lesion candidate regions input
At screening network further detect, screen network output test result 1 or 0.
Optionally, involved training sample includes pulmonary lesions area sample, breast lesion area sample, thyroid gland focal zone sample
Sheet, uterine lesion area sample, five class sample of brain lesion area, the quantity of every class sample are not less than 20,000 parts, every class sample respectively into
Row training.
Optionally, involved detection network is based on PVAnet detection framework, and PVAnet detection framework is to Faster-rcnn
Improved target detection model introduces C.ReLU, Inc, eption, HyperNet and residual module, for improving
Detection accuracy and detection speed.
Optionally, involved screening network utilizes fully-connected network, and there are two input terminals for tool, artificial to mark in training part
Remember that real focal zone sample and lesion candidate regions sample input full convolutional neural networks respectively, completes detection.
Optionally, in training part, the specific steps of building screening network include:
S131, the lesion candidate regions sample of the real focal zone sample and detection network identity of handmarking is carried out respectively
Coding, and belong to the real focal zone sample of the same focal zone sample, lesion candidate regions sample coding having the same；
S132, by with identical coding real focal zone sample and lesion candidate regions sample be known as true sample pair, will have
There are any real focal zone sample of different coding and any lesion candidate regions sample to be known as pseudo- sample pair；
S133, true sample is screened into network to input, screening e-learning detects the similarities and differences of true sample pair, learns multiple
After the similarities and differences of true sample pair, the similarity relationships of true sample pair are obtained, and be 1 by the output result queue for screening network；
S134, pseudo- sample is screened into network to input, screening e-learning detects the similarities and differences of pseudo- sample pair, obtains pseudo- sample
This pair of similarity relationships, then, according to the similarity relationships for having obtained true sample pair, by exclusive method or overlay method to true sample
This pair of similarity relationships optimize, and are 0 by the output result queue for screening network；
S135, the building for completing screening network.
Based on above-mentioned detection method, the present invention also provides a kind of focal zone detection system based on serial double Task Networks,
The system includes training part and detection part.
Training department divides
Collection module, for collecting the focal zone sample of various diseases as training sample；
Mark module, the real focal zone for handmarking's training sample；
Training building module one the characteristics of for learning real focal zone, and carries out pre-detection to training sample, repeatedly pre-
Detection obtains the lesion candidate regions of training sample, completes the building of detection network model；
Training building module two is more to carrying out for the sample to real focal zone sample and lesion candidate regions sample composition
Secondary training obtains the similarity relationships of real focal zone sample and lesion candidate regions sample, then, is belonging to a training sample
Real focal zone sample and lesion candidate regions sample constitute sample clock synchronization, by the testing result of output be labeled as 1, be not belonging to
The real focal zone sample and lesion candidate regions sample of the same training sample constitute sample clock synchronization, by the testing result mark of output
It is denoted as 0, repeatedly after study, label, completes the building of screening network model.
Detection part includes:
Network model is detected, for carrying out pre-detection to sample to be examined, and marks the lesion candidate regions of sample to be examined；
Screen network model, for detecting the lesion candidate regions of sample to be examined, the spy that has based on real focal zone sample
Point and the similarity relationships that really focal zone sample and lesion candidate regions sample should meet, export the detection of sample to be examined
As a result 1 or 0.
Optionally, involved training sample includes pulmonary lesions area sample, breast lesion area sample, thyroid gland focal zone sample
Sheet, uterine lesion area sample, five class sample of brain lesion area, manually classify to these five types of samples, the quantity of every class sample
Not less than 20,000 parts, every class sample is trained respectively.
Optionally, involved training department point includes study module, the characteristics of for learning inhomogeneity focal zone sample, and will
Similar focal zone sample is delivered to mark module；
The detection part includes categorization module, for being classified according to the characteristics of focal zone sample to sample to be examined,
And similar sample is sent to detection network model simultaneously, only one primary sample class is sent to detection network model.
Optionally, involved detection network model is based on PVAnet detection framework, and PVAnet detection framework is to Faster-
The improved target detection model of rcnn, introduces C.ReLU, Inc, and eption, HyperNet and residual module are used for
Improve detection accuracy and detection speed；The screening network model utilizes fully-connected network, and there are two input terminals for tool, in training department
Point, the real focal zone sample of handmarking and lesion candidate regions sample input full convolutional neural networks respectively, complete detection.
Optionally, involved training building module two specifically includes:
Coding unit, the lesion candidate regions sample for real focal zone sample and detection network identity to handmarking
It is encoded respectively, and the real focal zone sample, the lesion candidate regions sample that belong to the same focal zone sample are having the same
Coding；
Structural unit, for will there is the real focal zone sample equally encoded and lesion candidate regions sample architecture is true sample
This is right, by with different coding real focal zone sample and lesion candidate regions sample architecture be pseudo- sample pair；
Learning training unit one, for learning the similarities and differences of true sample pair, after the similarities and differences for learning multiple true samples pair, instruction
Get out the similarity relationships of true sample pair；
Learning training unit one, for learning the similarities and differences of pseudo- sample pair, after the similarities and differences for learning multiple pseudo- samples pair, instruction
Get out the similarity relationships of pseudo- sample pair；
Optimize construction unit, for optimizing the similarity relationships of true sample pair according to the similarity relationships of pseudo- sample pair, and
1 is exported to the testing result of true sample pair, 0 is exported to the testing result of pseudo- sample pair, completes the building of screening network model.
A kind of focal zone detection method and system based on serial double Task Networks of the invention, has compared with prior art
Beneficial effect be:
1) focal zone detection method of the invention includes training part and detection part；In training part firstly the need of collection
Focal zone sample, and the real focal zone of handmarking's sample, to sample after the characteristics of then detecting e-learning real focal zone
This progress pre-detection, and marked lesion candidate regions, repeatedly train the process, can construct the detection suitable for focal zone pre-detection
Network, finally, utilizing the real focal zone sample of screening same sample of e-learning and lesion candidate regions sample, different samples
The characteristics of this real focal zone sample and lesion candidate regions sample, find real focal zone sample and lesion candidate regions sample
The testing result of the real focal zone sample of same sample and lesion candidate regions sample is labeled as 1 by similarity relationships, will not
It is labeled as 0 with the real focal zone sample of sample and the testing result of lesion candidate regions sample, is repeatedly trained in screening network
The process can construct the screening network suitable for focal zone detection；In detection part, then the detection of building completion is directly utilized
Network handles sample this progress pre-detection, marks the lesion candidate regions of sample to be examined, the screening network completed followed by building
Sample to be examined is further detected, and output test result 1 or 0；In the case where early period, training sample was sufficiently large, this detection side
The detection image focal zone of method energy efficiently and accurately reduces false detection rate；
2) focal zone detection system of the invention is combined with detection method, and same includes training part and detection part；
Pass through collection module, mark module, training building module one, the training building building detection network model of module two in training part
With screening network model；Then pass through the detection network model of building in detection part and screen network model and completes sample to be examined
Detection, and output test result 1 or 0；This detection system has the advantages that detection is quick, accuracy in detection is high, can reduce mistake
Inspection rate；
3) focal zone detection method of the invention and system be not limited only to pulmonary lesions area sample, breast lesion area sample,
The detection of thyroid gland focal zone sample, uterine lesion area sample, five class sample of brain lesion area, as long as by training early period, also
It can be adapted for the detection of other position focal zone samples.
Detailed description of the invention
Attached drawing 1 is the flow diagram of the embodiment of the present invention one；
Attached drawing 2 is the flow diagram of building screening network in the embodiment of the present invention one；
Attached drawing 3 is the structure connection block diagram of the embodiment of the present invention two.
Each label information indicates in attached drawing:
10, collection module, 20, mark module, 30, training building module one,
40, training building module two, 50, detection network model, 60, screening network model,
70, study module, 80, categorization module；
41, coding unit, 42, structural unit, 43, learning training unit one,
44, learning training unit two, 45, optimization construction unit.
Specific embodiment
The technical issues of to make technical solution of the present invention, solving and technical effect are more clearly understood, below in conjunction with tool
Body embodiment is checked technical solution of the present invention, is completely described, it is clear that described embodiment is only this hair
Bright a part of the embodiment, instead of all the embodiments.Based on the embodiment of the present invention, those skilled in the art are not doing
All embodiments obtained under the premise of creative work out, all within protection scope of the present invention.
Embodiment one:
With reference to attached drawing 1, the present embodiment proposes a kind of focal zone detection method based on serial double task convolutional neural networks,
A kind of focal zone detection method based on serial double Task Networks, the detection method are serially completed by two networks,
Specifically comprise the following steps:
S100, training part:
S110, the focal zone sample of various diseases is collected as training sample, the real focal zone of handmarking's training sample；
The characteristics of S120, detection e-learning real focal zone, and pre-detection is carried out to training sample, by lesion candidate regions
It detected, repeatedly after training, complete the building of detection network；
S130, network is screened to input using real focal zone sample and lesion candidate regions sample as sample, screens network
For learning to detect the similarity relationships of real focal zone sample and lesion candidate regions sample, the real focal zone sample of handmarking
When the characteristics of with lesion candidate regions sample, is identical, screening network output result queue be 1, the real focal zone sample of handmarking with
When the characteristics of lesion candidate regions sample difference, screening network output result queue be 0, to multiple samples to being trained after, it is complete
At the building of screening network；
S200, detection part:
S210, sample to be detected is inputted to the detection network progress pre-detection that building is completed, detection network identity is to be detected
The lesion candidate regions of sample；
The characteristics of S220, the screening e-learning real focal zone sample completed based on building, real disease is also learnt
The similarity relationships of stove area sample and lesion candidate regions sample have constructed the sample to be detected for being marked with lesion candidate regions input
At screening network further detect, screen network output test result 1 or 0.
Training sample includes pulmonary lesions area sample, breast lesion area sample, thyroid gland focal zone sample, uterine lesion area
The quantity of sample, five class sample of brain lesion area, every class sample is not less than 20,000 parts, and every class sample is trained respectively.
It detects network and is based on PVAnet detection framework, PVAnet detection framework is examined to the improved target of Faster-rcnn
Model is surveyed, C.ReLU, Inc, eption, HyperNet and residual module, for improving detection accuracy and inspection are introduced
Degree of testing the speed.
It screening network and utilizes fully-connected network, there are two input terminals for tool, in training part, the real focal zone sample of handmarking
This inputs full convolutional neural networks with lesion candidate regions sample respectively, completes detection.
With reference to attached drawing 2, the specific steps in training part, building screening network include:
S131, the lesion candidate regions sample of the real focal zone sample and detection network identity of handmarking is carried out respectively
Coding, and belong to the real focal zone sample of the same focal zone sample, lesion candidate regions sample coding having the same；
S132, by with identical coding real focal zone sample and lesion candidate regions sample be known as true sample pair, will have
There are any real focal zone sample of different coding and any lesion candidate regions sample to be known as pseudo- sample pair；
S133, true sample is screened into network to input, screening e-learning detects the similarities and differences of true sample pair, learns multiple
After the similarities and differences of true sample pair, the similarity relationships of true sample pair are obtained, and be 1 by the output result queue for screening network；
S134, pseudo- sample is screened into network to input, screening e-learning detects the similarities and differences of pseudo- sample pair, obtains pseudo- sample
This pair of similarity relationships, then, according to the similarity relationships for having obtained true sample pair, by exclusive method or overlay method to true sample
This pair of similarity relationships optimize, and are 0 by the output result queue for screening network；
S135, the building for completing screening network.
The focal zone detection method of the present embodiment includes training part and detection part；In training part firstly the need of collection
Focal zone sample, and the real focal zone of handmarking's sample, to sample after the characteristics of then detecting e-learning real focal zone
This progress pre-detection, and marked lesion candidate regions, repeatedly train the process, can construct the detection suitable for focal zone pre-detection
Network, finally, utilizing the real focal zone sample of screening same sample of e-learning and lesion candidate regions sample, different samples
The characteristics of this real focal zone sample and lesion candidate regions sample, find real focal zone sample and lesion candidate regions sample
The testing result of the real focal zone sample of same sample and lesion candidate regions sample is labeled as 1 by similarity relationships, will not
It is labeled as 0 with the real focal zone sample of sample and the testing result of lesion candidate regions sample, is repeatedly trained in screening network
The process can construct the screening network suitable for focal zone detection；In detection part, then the detection of building completion is directly utilized
Network handles sample this progress pre-detection, marks the lesion candidate regions of sample to be examined, the screening network completed followed by building
Sample to be examined is further detected, and output test result 1 or 0；In the case where early period, training sample was sufficiently large, this detection side
The detection image focal zone of method energy efficiently and accurately reduces false detection rate.
Embodiment two:
With reference to attached drawing 3, the present embodiment proposes a kind of focal zone detection system based on serial double Task Networks, the system packet
Include training part and detection part.
Training department divides
Collection module 10, for collecting the focal zone sample of various diseases as training sample；
Mark module 20, the real focal zone for handmarking's training sample；
Training building module 1 the characteristics of for learning real focal zone, and carries out pre-detection to training sample, repeatedly
Pre-detection obtains the lesion candidate regions of training sample, completes the building of detection network model 50；
Training building module 2 40, for the sample to real focal zone sample and lesion candidate regions sample composition to progress
Repeatedly training, obtains the similarity relationships of real focal zone sample and lesion candidate regions sample, then, is belonging to a trained sample
This real focal zone sample and lesion candidate regions sample constitutes sample clock synchronization, and the testing result of output is labeled as 1, is not being belonged to
Sample clock synchronization is constituted in the real focal zone sample and lesion candidate regions sample of the same training sample, by the testing result of output
Labeled as 0, repeatedly after study, label, the building of screening network model 60 is completed.
Detection part includes:
Network model 50 is detected, for carrying out pre-detection to sample to be examined, and marks the lesion candidate regions of sample to be examined；
Screening network model 60 is had for detecting the lesion candidate regions of sample to be examined based on real focal zone sample
Feature and the similarity relationships that really focal zone sample and lesion candidate regions sample should meet, export the inspection of sample to be examined
Survey result 1 or 0.
Involved training sample includes pulmonary lesions area sample, breast lesion area sample, thyroid gland focal zone sample, uterus
Focal zone sample, five class sample of brain lesion area, manually classify to these five types of samples, and the quantity of every class sample is not less than 2
Ten thousand parts, every class sample is trained respectively.
Involved training department point includes study module 70, the characteristics of for learning inhomogeneity focal zone sample, and will be similar
Focal zone sample is delivered to mark module 20.
Involved detection part includes categorization module 80, for being divided according to the characteristics of focal zone sample sample to be examined
Class, and similar sample is sent to detection network model 50 simultaneously, only one primary sample class is sent to detection network model 50.
Involved detection network model 50 is based on PVAnet detection framework, and PVAnet detection framework is to Faster-rcnn
Improved target detection model introduces C.ReLU, Inc, eption, HyperNet and residual module, for improving
Detection accuracy and detection speed；It screening network model 60 and utilizes fully-connected network, there are two input terminals for tool, in training part, people
Work marks real focal zone sample and lesion candidate regions sample inputs full convolutional neural networks respectively, completes detection.
Involved training building module 2 40 specifically includes:
Coding unit 41, the lesion candidate regions sample for real focal zone sample and detection network identity to handmarking
This is encoded respectively, and belongs to the real focal zone sample of the same focal zone sample, lesion candidate regions sample with identical
Coding；
Structural unit 42, for will there is the real focal zone sample equally encoded and lesion candidate regions sample architecture is true
Sample pair, by with different coding real focal zone sample and lesion candidate regions sample architecture be pseudo- sample pair；
Learning training unit 1, for learning the similarities and differences of true sample pair, after the similarities and differences for learning multiple true samples pair,
Training obtains the similarity relationships of true sample pair；
Learning training unit 1, for learning the similarities and differences of pseudo- sample pair, after the similarities and differences for learning multiple pseudo- samples pair,
Training obtains the similarity relationships of pseudo- sample pair；
Optimize construction unit 45, for optimizing the similarity relationships of true sample pair according to the similarity relationships of pseudo- sample pair,
And to the testing result of true sample pair output 1,0 is exported to the testing result of pseudo- sample pair, completes the structure of screening network model 60
It builds.
The focal zone detection system of the present embodiment is combined with detection method, and same includes training part and detection part；
Pass through collection module 10, mark module 20, training building module 1, the training building building detection of module 2 40 in training part
Network model 50 and screening network model 60；Then pass through the detection network model 50 and screening network model of building in detection part
60 complete the detection of sample to be examined, and output test result 1 or 0；This detection system is quick with detection, accuracy in detection is high
Advantage can reduce false detection rate
Based on embodiment one, embodiment two, focal zone detection method of the invention and system are not limited only to pulmonary lesions area
The detection of sample, breast lesion area sample, thyroid gland focal zone sample, uterine lesion area sample, five class sample of brain lesion area,
As long as can be applicable to the detection of other position focal zone samples by training early period.
Use above specific case elaborates the principle of the present invention and embodiment, these embodiments are
It is used to help understand core of the invention technology contents, the protection scope being not intended to restrict the invention, technical side of the invention
Case is not limited in above-mentioned specific embodiment.Based on above-mentioned specific embodiment of the invention, those skilled in the art
Without departing from the principle of the present invention, any improvement and modification to made by the present invention should all be fallen into of the invention special
Sharp protection scope.
Claims (10)
1. a kind of focal zone detection method based on serial double Task Networks, which is characterized in that the detection method is by two networks
It is serial to complete, specifically comprise the following steps:
S100, training part:
S110, the focal zone sample of various diseases is collected as training sample, the real focal zone of handmarking's training sample；
The characteristics of S120, detection e-learning real focal zone, and pre-detection is carried out to training sample, lesion candidate regions are detected
Out, the building of detection network repeatedly is completed after training；
S130, network is screened to input using real focal zone sample and lesion candidate regions sample as sample, screening network is used for
Study detects the similarity relationships of real focal zone sample and lesion candidate regions sample, the real focal zone sample of handmarking and disease
When the characteristics of stove candidate regions sample is identical, screening network output result queue is 1, the real focal zone sample of handmarking and lesion
When the characteristics of candidate regions sample difference, screening network output result queue be 0, to multiple samples to being trained after, complete sieve
The building of network selection network；
S200, detection part:
S210, sample to be detected is inputted to the detection network progress pre-detection that building is completed, detects network identity sample to be detected
Lesion candidate regions；
The characteristics of S220, the screening e-learning real focal zone sample completed based on building, real focal zone is also learnt
The similarity relationships of sample and lesion candidate regions sample complete the sample to be detected input building for being marked with lesion candidate regions
Screening network further detects, and screens network output test result 1 or 0.
2. a kind of focal zone detection method based on serial double Task Networks according to claim 1, which is characterized in that institute
State training sample include pulmonary lesions area sample, breast lesion area sample, thyroid gland focal zone sample, uterine lesion area sample,
The quantity of five class sample of brain lesion area, every class sample is not less than 20,000 parts, and every class sample is trained respectively.
3. a kind of focal zone detection method based on serial double Task Networks according to claim 1, which is characterized in that institute
State detection network be based on PVAnet detection framework, PVAnet detection framework be to the improved target detection model of Faster-rcnn,
Introduce C.ReLU, Inc, eption, HyperNet and residual module, for improving detection accuracy and detection speed.
4. a kind of focal zone detection method based on serial double Task Networks according to claim 3, which is characterized in that institute
Screening network is stated using fully-connected network, there are two input terminals for tool, in training part, the real focal zone sample of handmarking and disease
Stove candidate regions sample inputs full convolutional neural networks respectively, completes detection.
5. a kind of focal zone detection method based on serial double Task Networks according to claim 1, which is characterized in that
Training part, the specific steps that network is screened in building include:
S131, the lesion candidate regions sample of the real focal zone sample and detection network identity of handmarking is compiled respectively
Code, and belong to the real focal zone sample of the same focal zone sample, lesion candidate regions sample coding having the same；
S132, by with identical coding real focal zone sample and lesion candidate regions sample be known as true sample pair, will have not
It is known as pseudo- sample pair with any real focal zone sample of coding and any lesion candidate regions sample；
S133, true sample is screened into network to input, screening e-learning detects the similarities and differences of true sample pair, learns multiple true samples
After this pair of similarities and differences, the similarity relationships of true sample pair are obtained, and be 1 by the output result queue for screening network；
S134, pseudo- sample is screened into network to input, screening e-learning detects the similarities and differences of pseudo- sample pair, obtains pseudo- sample pair
Similarity relationships, then, according to the similarity relationships for having obtained true sample pair, by exclusive method or overlay method to true sample pair
Similarity relationships optimize, and by screen network output result queue be 0；
S135, the building for completing screening network.
6. a kind of focal zone detection system based on serial double Task Networks, which is characterized in that the system include training part and
Detection part；
The training department divides
Collection module, for collecting the focal zone sample of various diseases as training sample；
Mark module, the real focal zone for handmarking's training sample；
Training building module one the characteristics of for learning real focal zone, and carries out pre-detection, multiple pre-detection to training sample
It obtains the lesion candidate regions of training sample, completes the building of detection network model；
Training building module two, for the sample to real focal zone sample and lesion candidate regions sample composition to repeatedly being instructed
Practice, obtains the similarity relationships of real focal zone sample and lesion candidate regions sample, then, belonging to the true of a training sample
Positive focal zone sample and lesion candidate regions sample constitute sample clock synchronization, and the testing result of output is labeled as 1, be not belonging to it is same
The real focal zone sample and lesion candidate regions sample of a training sample constitute sample clock synchronization, and the testing result of output is labeled as
0, repeatedly after study, label, complete the building of screening network model；
The detection part includes:
Network model is detected, for carrying out pre-detection to sample to be examined, and marks the lesion candidate regions of sample to be examined；
Screen network model, for detecting the lesion candidate regions of sample to be examined, had the special feature that based on real focal zone sample, with
And the similarity relationships that really focal zone sample and lesion candidate regions sample should meet, export the testing result 1 of sample to be examined
Or 0.
7. a kind of focal zone detection system based on serial double Task Networks according to claim 6, which is characterized in that institute
State training sample include pulmonary lesions area sample, breast lesion area sample, thyroid gland focal zone sample, uterine lesion area sample,
Five class sample of brain lesion area, manually classifies to these five types of samples, and the quantity of every class sample is not less than 20,000 parts, every class sample
This is trained respectively.
8. a kind of focal zone detection system based on serial double Task Networks according to claim 7, which is characterized in that institute
Stating training department point includes study module, the characteristics of for learning inhomogeneity focal zone sample, and similar focal zone sample is conveyed
To mark module；
The detection part includes categorization module, for being classified according to the characteristics of focal zone sample to sample to be examined, and it is same
Class sample is sent to detection network model simultaneously, only one primary sample class is sent to detection network model.
9. a kind of focal zone detection system based on serial double Task Networks according to claim 6, which is characterized in that institute
It states detection network model and is based on PVAnet detection framework, PVAnet detection framework is to the improved target detection of Faster-rcnn
Model introduces C.ReLU, Inc, eption, HyperNet and residual module, for improving detection accuracy and detection
Speed；The screening network model utilizes fully-connected network, and there are two input terminals for tool, and in training part, handmarking is really sick
Stove area sample and lesion candidate regions sample input full convolutional neural networks respectively, complete detection.
10. a kind of focal zone detection system based on serial double Task Networks according to claim 6, which is characterized in that
The training building module two specifically includes:
Coding unit, the lesion candidate regions sample for real focal zone sample and detection network identity to handmarking are distinguished
It is encoded, and belongs to the real focal zone sample of the same focal zone sample, lesion candidate regions sample coding having the same；
Structural unit, for will there is the real focal zone sample equally encoded and lesion candidate regions sample architecture is true sample
It is right, by with different coding real focal zone sample and lesion candidate regions sample architecture be pseudo- sample pair；
Learning training unit one, for learning the similarities and differences of true sample pair, after the similarities and differences for learning multiple true samples pair, trained
The similarity relationships of true sample pair out；
Learning training unit one, for learning the similarities and differences of pseudo- sample pair, after the similarities and differences for learning multiple pseudo- samples pair, trained
The similarity relationships of pseudo- sample pair out；
Optimize construction unit, for optimizing the similarity relationships of true sample pair according to the similarity relationships of pseudo- sample pair, and to true
The testing result output 1 of sample pair exports 0 to the testing result of pseudo- sample pair, completes the building of screening network model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811441059.9A CN109522968A (en) | 2018-11-29 | 2018-11-29 | A kind of focal zone detection method and system based on serial double Task Networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811441059.9A CN109522968A (en) | 2018-11-29 | 2018-11-29 | A kind of focal zone detection method and system based on serial double Task Networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109522968A true CN109522968A (en) | 2019-03-26 |
Family
ID=65793636
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811441059.9A Pending CN109522968A (en) | 2018-11-29 | 2018-11-29 | A kind of focal zone detection method and system based on serial double Task Networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109522968A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110610202A (en) * | 2019-08-30 | 2019-12-24 | 联想(北京)有限公司 | Image processing method and electronic equipment |
| CN111861966A (en) * | 2019-04-18 | 2020-10-30 | 杭州海康威视数字技术股份有限公司 | Model training method and device and defect detection method and device |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130170718A1 (en) * | 2012-01-03 | 2013-07-04 | Samsung Electronics Co., Ltd. | Lesion diagnosis apparatus and method to determine regularity of shape of lesion |
| CN105139390A (en) * | 2015-08-14 | 2015-12-09 | 四川大学 | Image processing method for detecting pulmonary tuberculosis focus in chest X-ray DR film |
| CN107103601A (en) * | 2017-04-14 | 2017-08-29 | 成都知识视觉科技有限公司 | A kind of cell mitogen detection method in breast cancer points-scoring system |
| CN107330263A (en) * | 2017-06-26 | 2017-11-07 | 成都知识视觉科技有限公司 | A kind of method of area of computer aided breast invasive ductal carcinoma histological grading |
| CN108491828A (en) * | 2018-04-20 | 2018-09-04 | 济南浪潮高新科技投资发展有限公司 | A kind of parking site detecting system and method for the pairwise similarity PVAnet based on level |
| CN109241967A (en) * | 2018-09-04 | 2019-01-18 | 青岛大学附属医院 | Thyroid ultrasound automatic image recognition system, computer equipment, storage medium based on deep neural network |
-
2018
- 2018-11-29 CN CN201811441059.9A patent/CN109522968A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130170718A1 (en) * | 2012-01-03 | 2013-07-04 | Samsung Electronics Co., Ltd. | Lesion diagnosis apparatus and method to determine regularity of shape of lesion |
| CN105139390A (en) * | 2015-08-14 | 2015-12-09 | 四川大学 | Image processing method for detecting pulmonary tuberculosis focus in chest X-ray DR film |
| CN107103601A (en) * | 2017-04-14 | 2017-08-29 | 成都知识视觉科技有限公司 | A kind of cell mitogen detection method in breast cancer points-scoring system |
| CN107330263A (en) * | 2017-06-26 | 2017-11-07 | 成都知识视觉科技有限公司 | A kind of method of area of computer aided breast invasive ductal carcinoma histological grading |
| CN108491828A (en) * | 2018-04-20 | 2018-09-04 | 济南浪潮高新科技投资发展有限公司 | A kind of parking site detecting system and method for the pairwise similarity PVAnet based on level |
| CN109241967A (en) * | 2018-09-04 | 2019-01-18 | 青岛大学附属医院 | Thyroid ultrasound automatic image recognition system, computer equipment, storage medium based on deep neural network |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111861966A (en) * | 2019-04-18 | 2020-10-30 | 杭州海康威视数字技术股份有限公司 | Model training method and device and defect detection method and device |
| CN111861966B (en) * | 2019-04-18 | 2023-10-27 | 杭州海康威视数字技术股份有限公司 | Model training method and device and defect detection method and device |
| CN110610202A (en) * | 2019-08-30 | 2019-12-24 | 联想(北京)有限公司 | Image processing method and electronic equipment |
| CN110610202B (en) * | 2019-08-30 | 2022-07-26 | 联想(北京)有限公司 | Image processing method and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109886273B (en) | CMR image segmentation and classification system | |
| Mamalakis et al. | DenResCov-19: A deep transfer learning network for robust automatic classification of COVID-19, pneumonia, and tuberculosis from X-rays | |
| Sharma et al. | Covid-MANet: Multi-task attention network for explainable diagnosis and severity assessment of COVID-19 from CXR images | |
| CN110991536B (en) | Training method of early warning model of primary liver cancer | |
| CN108511056A (en) | Therapeutic scheme based on patients with cerebral apoplexy similarity analysis recommends method and system | |
| CN108511055A (en) | Ventricular premature beat identifying system and method based on Multiple Classifier Fusion and diagnostic rule | |
| CN106980815A (en) | Facial paralysis objective evaluation method under being supervised based on H B rank scores | |
| Qiu et al. | Multi-label detection and classification of red blood cells in microscopic images | |
| CN110415818A (en) | A kind of intelligent pediatric disease interrogation system and method based on observable illness | |
| CN109522968A (en) | A kind of focal zone detection method and system based on serial double Task Networks | |
| CN113610118A (en) | Fundus image classification method, device, equipment and medium based on multitask course learning | |
| CN113397485A (en) | Scoliosis screening method based on deep learning | |
| Bajaj et al. | Classification And Prediction of Brain Tumors and its Types using Deep Learning | |
| CN109190699A (en) | A kind of more disease joint measurement methods based on multi-task learning | |
| Li et al. | Lesion-aware convolutional neural network for chest radiograph classification | |
| Jabbar et al. | A Lesion-Based Diabetic Retinopathy Detection Through Hybrid Deep Learning Model | |
| Zhao et al. | Pulmonary nodule detection based on multiscale feature fusion | |
| Mvoulana et al. | Fine-tuning Convolutional Neural Networks: a comprehensive guide and benchmark analysis for Glaucoma Screening | |
| Hadi et al. | A lightweight CORONA-NET for COVID-19 detection in X-ray images | |
| Ganeshkumar et al. | Two-stage deep learning model for automate detection and classification of lung diseases | |
| Gupta et al. | Class-specific hierarchical classification of HEp-2 cell images: The case of two classes | |
| Orlando et al. | Learning to detect red lesions in fundus photographs: An ensemble approach based on deep learning | |
| CN115482927B (en) | Children's pneumonia diagnostic system based on little sample | |
| Sumon et al. | Using Deep Learning Systems for Imaging Methods for Recognising Brain Tumors | |
| Xiang et al. | Application of convolutional neural network algorithm in diagnosis of chronic cough and tongue in children with traditional Chinese medicine |
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: 20190326 |