CN109785963A - Lung neoplasm screening algorithm based on depth learning technology - Google Patents
Lung neoplasm screening algorithm based on depth learning technology Download PDFInfo
- Publication number
- CN109785963A CN109785963A CN201910038443.2A CN201910038443A CN109785963A CN 109785963 A CN109785963 A CN 109785963A CN 201910038443 A CN201910038443 A CN 201910038443A CN 109785963 A CN109785963 A CN 109785963A
- Authority
- CN
- China
- Prior art keywords
- tubercle
- lung neoplasm
- sign
- block
- screening
- 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
Landscapes
- Image Analysis (AREA)
- Apparatus For Radiation Diagnosis (AREA)
Abstract
The invention discloses a kind of Lung neoplasm screening algorithm based on depth learning technology, it is that candidate nodule generates task first, for extracting a large amount of doubtful tubercles it is characterized by: being three phases, in the case where certain false positive rate, guarantee 99.5% or more tubercle recall rate；Followed by False Positive Reduction task filters out real tubercle for reducing false positive rate；Tubercle is qualitative, for diagnosing good pernicious, the observation sign of Lung neoplasm, provides diagnosis basis for doctor.This method uses chest CT as input, accurately screening can go out contained Lung neoplasm, and auxiliary doctor improves diagosis efficiency, solves medical industry doctor and lack that demand is more, medical level imbalance, misdiagnosis rate height the problems such as time-consuming.
Description
Technical field
The present invention relates to a kind of Lung neoplasm screening algorithms, are specifically a kind of Lung neoplasm sieves based on deep learning technology
Look into algorithm.
Background technique
Lung neoplasm screening task is substantially the target detection problems of computer vision field, and popular solution has two
Kind: the algorithm of target detection of conventional machines learning model, the algorithm of target detection based on deep learning method.
The method that traditional target detection generally uses sliding window combining classification device mainly includes three steps:
Certain a part in figure is framed as candidate region using various sizes of sliding window；Extract the relevant view in candidate region
Feel feature, such as Harr feature, HOG feature etc.；It is identified using classifier, such as common SVM model.Such side
Method accuracy rate is not high, and main cause is that traditional characteristic needs a large amount of expert's domain knowledge that can form effective Feature
Set, and be difficult to extract structure feature complicated inside data, the transfer learning property of model algorithm is also bad.
With the development of deep learning technology, more and more mechanisms and the object detection task for organizing to apply it to medical image
In.For chest CT Lung neoplasm screening task there are mainly two types of thinking, the first is respectively to each layer of scanning in CT
Picture carries out the tubercle screening of 2D image, finally merges the screening results of all scanning slices, provides final screening results,
Since this method does not utilize the three-dimensional structure characteristic of tubercle, there are obvious disadvantages in screening accuracy.Second method makes
With more fully image information, lung tissue is split first from CT image, is then directed to using FPN network
The lung tissue extracted carries out the tubercle screening of 3-dimensional image, and effect is often preferable, but the screening efficiency and sieve of existing algorithm
Looking into effect still has biggish room for promotion, this is because lung tissue's cutting procedure is more time-consuming, and the segmentation of lung edge is not
The loss of important information is accurately caused, it is more significant there are influencing when large volume of lesion especially on lung wall, in addition, micro-
Lesser tubercle as low as 3mm, greatly to 80mm, in the case where considering video memory and computational efficiency, FPN network is such to dimensional variation
Big target detection effect is bad.
Summary of the invention
Therefore, in order to solve above-mentioned deficiency, the present invention provides a kind of Lung neoplasm screening based on deep learning technology herein
Algorithm.This patent proposes a kind of efficient deep learning method, and this method uses chest CT as input, can be compared with subject to
True screening goes out contained Lung neoplasm, and auxiliary doctor improves diagosis efficiency, solve medical industry doctor lack demand is more, medical level not
Balance, misdiagnosis rate high the problems such as time-consuming.
The invention is realized in this way constructing a kind of Lung neoplasm screening algorithm based on deep learning technology, feature exists
In: it is that candidate nodule generates task first for three phases, for extracting a large amount of doubtful tubercles, certain false positive rate the case where
Under, guarantee 99.5% or more tubercle recall rate；Followed by False Positive Reduction task, for reducing
False sun rate, filters out real tubercle；Tubercle is qualitative, for diagnosing good pernicious, the observation sign of Lung neoplasm, provides and examines for doctor
Disconnected foundation.
The Lung neoplasm screening algorithm based on deep learning technology according to the present invention, it is characterised in that: candidate nodule is raw
At: task is extracted for candidate nodule, the advantage of models coupling Unet, RetinaNet use FocalLoss when training,
Realize the tubercle screening model based on 3D image.Process can be divided into four-stage: to the pretreatment of image, using backbone network
Network carries out the non-maximum restraining of feature extraction, multiple dimensioned candidate nodule pre-generatmg, candidate nodule；Specially；
(1) pretreatment of image: since input CT is the image of 100-700 layers of 512 × 512 sizes, the number of plies mostly
Difference causes model not train directly, and 3D image is excessive limits the complexity of model, or even can not construct based on master
The 3D tubercle screening network of video card video memory is flowed, it is essential to the pretreated work of CT progress as a result, to obtain model
It can be used for training and the data mode predicted；
(2) backbone network: tubercle screening backbone network uses the Unet based on three-dimensional ResNet, mainly includes two parts:
Contract-ing Path and Expansive Path.Wherein Contracting Path is a three dimensional stress
ResNet, it is one after the last one Block that totally 4 Block, each Block, which are before max-pooling layers,
Layer CNN, and each Block is made of multi-layer C NN.The structure and Expansive Path of Expansive Path
On the contrary, only the input of each Block is previous Block input and the symmetrical Block of Expansive Path
Sum of output does not use max-pooling between Block, but uses deconvolution progress upsampling；
(3) multiple dimensioned candidate nodule pre-generatmg: 4 Block in backbone network Expansive Path are exported respectively
4 three-dimensional feature figures, for each characteristic pattern, respectively by the CNN network of similar Retina-Head, carry out classification and
Location prediction；
(4) it non-maximum restraining: obtains that Non- can be used by after the candidate nodule of entire CT difference patch pre-generatmg
Boxes of the Maximum-Suppression by degree of overlapping higher than threshold is merged, and inhibits redundancy box, most
Candidate nodule is obtained eventually.
The Lung neoplasm screening algorithm based on deep learning technology according to the present invention, it is characterised in that: False
Positive Reduction task is divided into two stages: candidate nodule pretreatment, tubercle classification.Preprocessing process cuts lung
Image patch near tubercle, region physics size is fixed, then size needed for scaling to tubercle disaggregated model.Tubercle classification
Model uses CNN network, and the prediction result ensamble of task is generated with candidate nodule, obtains better Lung neoplasm inspection
Survey effect.
The Lung neoplasm screening algorithm based on deep learning technology according to the present invention, it is characterised in that: tubercle is qualitative；It is logical
Cross the analysis to perinodal voxel, tubercle classification, tubercle edge, reality ingredient edge, sign of lobulation, spicule sign, blood vessel at
As sign, dizzy sign, graininess, honeycomb sign, vacuole sign, scar sample, interlobar fissure is related, growth long axis is consistent with bronchus and pleura
More than equal ten observations sign in close relations is diagnosed, and provides the good pernicious and specific cancer parting information of tubercle.
The present invention has the advantage that the present invention provides a kind of Lung neoplasm screening calculation based on deep learning technology herein
Method.This patent proposes a kind of efficient deep learning method, and this method uses chest CT as input, can be more accurate
Screening go out contained Lung neoplasm, auxiliary doctor improves diagosis efficiency, solves medical industry doctor and lack that demand is more, medical level injustice
Weighing apparatus, misdiagnosis rate high the problems such as time-consuming.This patent have the following advantages that and the utility model has the advantages that
1) remove the links such as lung areas segmentation, feature extraction, improve operation efficiency.
2) 3-D image target identification, in conjunction with more dimensional informations, improvement effect.
3) End-to-End deep learning model possesses bigger space in terms of automatically adjusting according to data,
Increase model entirety compatible degree, and there is good transfer learning property.
4) False Positive Reduction and candidate nodule generate the integrated of two tasks, judge more accurate.
5) a variety of image enhancement means make model have good applicability for the tubercle in different size, direction.
6) Multi-Task model sharing indicates, improves generalization ability.
Detailed description of the invention
Fig. 1 is backbone network schematic diagram；
Fig. 2 is residual error network diagram；
Fig. 3 is splicing network diagram.
Specific embodiment
Below in conjunction with attached drawing 1- Fig. 3, the present invention is described in detail, technical solution in the embodiment of the present invention into
Row clearly and completely describes, it is clear that described embodiments are only a part of the embodiments of the present invention, rather than whole realities
Apply example.Based on the embodiments of the present invention, those of ordinary skill in the art are obtained without making creative work
Every other embodiment, shall fall within the protection scope of the present invention.
The present invention provides a kind of Lung neoplasm screening algorithm based on deep learning technology by improving herein；As shown in the figure；
Lung neoplasm screening PROBLEM DECOMPOSITION is three phases by this patent, is that candidate nodule generates task first, a large amount of doubtful for extracting
Tubercle guarantees 99.5% or more tubercle recall rate in the case where certain false positive rate；Followed by False Positive
Reduction task filters out real tubercle for reducing false positive rate；Tubercle is qualitative, for diagnosing the good evil of Lung neoplasm
Property, observation sign, provide diagnosis basis for doctor.
One, candidate nodule generate: for candidate nodule extract task, the advantage of models coupling Unet, RetinaNet,
FocalLoss is used when training, realizes the tubercle screening model based on 3D image.Process can be divided into four-stage: to figure
The pretreatment of picture carries out feature extraction, the non-very big suppression of multiple dimensioned candidate nodule pre-generatmg, candidate nodule using backbone network
System.
(1) pretreatment of image: since input CT is the image of 100-700 layers of 512 × 512 sizes mostly,
Number of plies difference causes model not train directly, and 3D image is excessive limits the complexity of model, or even can not construct base
It is essential to the pretreated work of CT progress as a result, in the 3D tubercle screening network of mainstream video card video memory, to obtain
The data mode that model can be used for training and predict.
(2) backbone network: tubercle screening backbone network uses the Unet based on three-dimensional ResNet, and main includes two
Point: Contract-ing Path and Expansive Path.Wherein Contracting Path is a three dimensional stress
ResNet, it is one after the last one Block that totally 4 Block, each Block, which are before max-pooling layers,
Layer CNN, and each Block is made of multi-layer C NN.The structure and Expansive Path of Expansive Path
On the contrary, only the input of each Block is previous Block input and the symmetrical Block of Expansive Path
Sum of output does not use max-pooling between Block, but uses deconvolution progress upsampling.
(3) multiple dimensioned candidate nodule pre-generatmg: 4 Block difference in backbone network Expansive Path
4 three-dimensional feature figures are exported, for each characteristic pattern, respectively by the CNN network of similar Retina-Head, are carried out
Classification and location prediction.
(4) it non-maximum restraining: obtains to use after the candidate nodule of entire CT difference patch pre-generatmg
Boxes of the Non-Maximum-Suppression by degree of overlapping higher than threshold is merged, and inhibits redundancy
Box finally obtains candidate nodule.
(2) False Positive Reduction:
False Positive Reduction task is divided into two stages: candidate nodule pretreatment, tubercle classification.Pretreatment
Image patch near process tailoring Lung neoplasm, region physics size are fixed, then needed for scaling to tubercle disaggregated model it is big
It is small.Tubercle disaggregated model uses CNN network, and the prediction result ensamble of task is generated with candidate nodule, obtains more preferable
Lung neoplasm detection effect.
(3) tubercle is qualitative: by the analysis to perinodal voxel, in tubercle classification, tubercle edge, reality ingredient side
Edge, sign of lobulation, spicule sign, blood vessel imaging sign, dizzy sign, graininess, honeycomb sign, vacuole sign, scar sample, interlobar fissure is related, grows
Long axis is consistent with bronchus, more than ten observation signs are diagnosed with pleura is in close relations etc., and provides that tubercle is good pernicious and tool
Body cancer parting information.
The sample distributions of most of observation signs are unbalanced, or even the case where 80:1 occur, rare sample also more difficult receipts
Collection, building machine learning model, which classifies to different observation signs, respectively can not reach desired effect.This patent uses
Multi-Task model, while classifying to observation sign, shared backbone network feature extraction, only it is independent calculate prediction and
Loss, substantially increases the generalization ability and effect of model.
This patent have the following advantages that and the utility model has the advantages that
1) remove the links such as lung areas segmentation, feature extraction, improve operation efficiency.
2) 3-D image target identification, in conjunction with more dimensional informations, improvement effect.
3) End-to-End deep learning model possesses bigger space in terms of automatically adjusting according to data,
Increase model entirety compatible degree, and there is good transfer learning property.
4) False Positive Reduction and candidate nodule generate the integrated of two tasks, judge more accurate.
5) a variety of image enhancement means make model have good applicability for the tubercle in different size, direction.
6) Multi-Task model sharing indicates, improves generalization ability.
The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention.
Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, of the invention
It is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein phase one
The widest scope of cause.
Claims (4)
1. a kind of Lung neoplasm screening algorithm based on deep learning technology, it is characterised in that: be three phases, be candidate knot first
Generation task is saved, for extracting a large amount of doubtful tubercles, in the case where certain false positive rate, guarantees 99.5% or more tubercle inspection
Extracting rate；Followed by False Positive Reduction task filters out real tubercle for reducing false positive rate；Knot
Save it is qualitative, for diagnose Lung neoplasm it is good it is pernicious, observation sign, provide diagnosis basis for doctor.
2. the Lung neoplasm screening algorithm based on deep learning technology according to claim 1, it is characterised in that: candidate nodule is raw
At: task is extracted for candidate nodule, the advantage of models coupling Unet, RetinaNet use FocalLoss when training,
Realize the tubercle screening model based on 3D image；Process can be divided into four-stage: to the pretreatment of image, using backbone network
Network carries out the non-maximum restraining of feature extraction, multiple dimensioned candidate nodule pre-generatmg, candidate nodule；Specially；
(1) pretreatment of image: since input CT is the image of 100-700 layers of 512 × 512 sizes, the number of plies mostly
Difference causes model not train directly, and 3D image is excessive limits the complexity of model, or even can not construct based on master
The 3D tubercle screening network of video card video memory is flowed, it is essential to the pretreated work of CT progress as a result, to obtain model
It can be used for training and the data mode predicted；
(2) backbone network: tubercle screening backbone network uses the Unet based on three-dimensional ResNet, mainly includes two parts:
Contract-ing Path and Expansive Path；Wherein Contracting Path is a three dimensional stress
ResNet, it is one after the last one Block that totally 4 Block, each Block, which are before max-pooling layers,
Layer CNN, and each Block is made of multi-layer C NN；The structure and Expansive Path of Expansive Path
On the contrary, only the input of each Block is previous Block input and the symmetrical Block of Expansive Path
Sum of output does not use max-pooling between Block, but uses deconvolution progress upsampling；
(3) multiple dimensioned candidate nodule pre-generatmg: 4 Block in backbone network Expansive Path are exported respectively
4 three-dimensional feature figures, for each characteristic pattern, respectively by the CNN network of similar Retina-Head, carry out classification and
Location prediction；
(4) it non-maximum restraining: obtains that Non- can be used by after the candidate nodule of entire CT difference patch pre-generatmg
Boxes of the Maximum-Suppression by degree of overlapping higher than threshold is merged, and inhibits redundancy box, most
Candidate nodule is obtained eventually.
3. the Lung neoplasm screening algorithm based on deep learning technology according to claim 1, it is characterised in that: False
Positive Reduction task is divided into two stages: candidate nodule pretreatment, tubercle classification；In advance
Treatment process cuts the image patch near Lung neoplasm, and region physics size fixes, then scaling is to tubercle disaggregated model
Required size；Tubercle disaggregated model uses CNN network, and the prediction result ensamble of task is generated with candidate nodule,
Obtain better Lung neoplasm detection effect.
4. the Lung neoplasm screening algorithm based on deep learning technology according to claim 1, it is characterised in that: tubercle is qualitative；
By the analysis to perinodal voxel, at tubercle classification, tubercle edge, reality ingredient edge, sign of lobulation, spicule sign, blood vessel
Imaging sign, dizzy sign, graininess, honeycomb sign, vacuole sign, scar sample, interlobar fissure is related, growth long axis is consistent with bronchus and chest
Film more than equal ten observations sign in close relations is diagnosed, and provides the good pernicious and specific cancer parting information of tubercle.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910038443.2A CN109785963A (en) | 2019-01-16 | 2019-01-16 | Lung neoplasm screening algorithm based on depth learning technology |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910038443.2A CN109785963A (en) | 2019-01-16 | 2019-01-16 | Lung neoplasm screening algorithm based on depth learning technology |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109785963A true CN109785963A (en) | 2019-05-21 |
Family
ID=66499407
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910038443.2A Pending CN109785963A (en) | 2019-01-16 | 2019-01-16 | Lung neoplasm screening algorithm based on depth learning technology |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109785963A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110570425A (en) * | 2019-10-18 | 2019-12-13 | 北京理工大学 | Lung nodule analysis method and device based on deep reinforcement learning algorithm |
| CN111445946A (en) * | 2020-03-26 | 2020-07-24 | 北京易康医疗科技有限公司 | Calculation method for calculating lung cancer genotyping by using PET/CT (positron emission tomography/computed tomography) images |
| CN112116558A (en) * | 2020-08-17 | 2020-12-22 | 您好人工智能技术研发昆山有限公司 | CT image pulmonary nodule detection system based on deep learning |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107977966A (en) * | 2017-12-20 | 2018-05-01 | 点内（上海）生物科技有限公司 | Based on artificial intelligence pernicious and infiltration degree aided diagnosis method good to Lung neoplasm |
| CN108198179A (en) * | 2018-01-03 | 2018-06-22 | 华南理工大学 | A kind of CT medical image pulmonary nodule detection methods for generating confrontation network improvement |
| CN108648172A (en) * | 2018-03-30 | 2018-10-12 | 四川元匠科技有限公司 | A kind of CT figure Lung neoplasm detecting systems based on 3D-Unet |
| CN108670285A (en) * | 2018-06-05 | 2018-10-19 | 胡晓云 | A kind of CT pulmonary tuberculosis detection artificial intelligence diagnosis and therapy system |
-
2019
- 2019-01-16 CN CN201910038443.2A patent/CN109785963A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107977966A (en) * | 2017-12-20 | 2018-05-01 | 点内（上海）生物科技有限公司 | Based on artificial intelligence pernicious and infiltration degree aided diagnosis method good to Lung neoplasm |
| CN108198179A (en) * | 2018-01-03 | 2018-06-22 | 华南理工大学 | A kind of CT medical image pulmonary nodule detection methods for generating confrontation network improvement |
| CN108648172A (en) * | 2018-03-30 | 2018-10-12 | 四川元匠科技有限公司 | A kind of CT figure Lung neoplasm detecting systems based on 3D-Unet |
| CN108670285A (en) * | 2018-06-05 | 2018-10-19 | 胡晓云 | A kind of CT pulmonary tuberculosis detection artificial intelligence diagnosis and therapy system |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110570425A (en) * | 2019-10-18 | 2019-12-13 | 北京理工大学 | Lung nodule analysis method and device based on deep reinforcement learning algorithm |
| CN110570425B (en) * | 2019-10-18 | 2023-09-08 | 北京理工大学 | Pulmonary nodule analysis method and device based on deep reinforcement learning algorithm |
| CN111445946A (en) * | 2020-03-26 | 2020-07-24 | 北京易康医疗科技有限公司 | Calculation method for calculating lung cancer genotyping by using PET/CT (positron emission tomography/computed tomography) images |
| CN112116558A (en) * | 2020-08-17 | 2020-12-22 | 您好人工智能技术研发昆山有限公司 | CT image pulmonary nodule detection system based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US8111896B2 (en) | Method and system for automatic recognition of preneoplastic anomalies in anatomic structures based on an improved region-growing segmentation, and commputer program therefor | |
| CN109063710A (en) | Based on the pyramidal 3D CNN nasopharyngeal carcinoma dividing method of Analysis On Multi-scale Features | |
| CN110310281A (en) | Lung neoplasm detection and dividing method in a kind of Virtual Medical based on Mask-RCNN deep learning | |
| CN110458249A (en) | A kind of lesion categorizing system based on deep learning Yu probability image group | |
| CN108257135A (en) | The assistant diagnosis system of medical image features is understood based on deep learning method | |
| Alilou et al. | A comprehensive framework for automatic detection of pulmonary nodules in lung CT images | |
| CN108288271A (en) | Image detecting system and method based on three-dimensional residual error network | |
| CN109785963A (en) | Lung neoplasm screening algorithm based on depth learning technology | |
| CN110517253B (en) | Method for classifying benign and malignant pulmonary nodules based on 3D multi-target feature learning | |
| CN111144474B (en) | Multi-view, multi-scale and multi-task lung nodule classification method | |
| CN109363698A (en) | A kind of method and device of breast image sign identification | |
| CN109685768A (en) | Lung neoplasm automatic testing method and system based on lung CT sequence | |
| Romero et al. | End-to-end discriminative deep network for liver lesion classification | |
| CN110555836A (en) | Automatic identification method and system for standard fetal section in ultrasonic image | |
| CN110853011A (en) | Method for constructing convolutional neural network model for pulmonary nodule detection | |
| US20210248747A1 (en) | Organs at risk auto-contouring system and methods | |
| CN110349167A (en) | A kind of image instance dividing method and device | |
| CN112365973A (en) | Pulmonary nodule auxiliary diagnosis system based on countermeasure network and fast R-CNN | |
| Mobiny et al. | Lung cancer screening using adaptive memory-augmented recurrent networks | |
| Feng et al. | Deep learning for chest radiology: a review | |
| Duggan et al. | A technique for lung nodule candidate detection in CT using global minimization methods | |
| CN113421240A (en) | Mammary gland classification method and device based on ultrasonic automatic mammary gland full-volume imaging | |
| CN109671055A (en) | Pulmonary nodule detection method and device | |
| CN115953393A (en) | Intracranial aneurysm detection system, equipment and storage medium based on multitask learning | |
| Mridha et al. | A comprehensive survey on the progress, process, and challenges of lung cancer detection and classification |
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 |