CN108090501A - Based on plate experiment and the bacteriostatic level recognition methods of deep learning - Google Patents
Based on plate experiment and the bacteriostatic level recognition methods of deep learning Download PDFInfo
- Publication number
- CN108090501A CN108090501A CN201711190762.2A CN201711190762A CN108090501A CN 108090501 A CN108090501 A CN 108090501A CN 201711190762 A CN201711190762 A CN 201711190762A CN 108090501 A CN108090501 A CN 108090501A
- Authority
- CN
- China
- Prior art keywords
- model
- drug
- paper
- deep learning
- sensitive test
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
Abstract
The invention discloses a kind of based on plate experiment and the bacteriostatic level recognition methods of deep learning, comprise the following steps：S1, drug sensitive test paper position is determined；S2, inhibition zone size is measured；S3, construction sample set：To the sample set of deep learning Construction of A Model tape label, for training pattern；S4, structure and training pattern：Structure can be used for identifying the model of drug sensitive test paper, and be trained with training sample set；S5, Model Identification drug sensitive test paper is used；S6, the bacteriostatic level for judging drug：The medicament categories gone out by Model Identification inquire about corresponding antibacterial standard from database, then determine that the diameter of inhibition zone belongs to which of standard section, so as to obtain the bacteriostatic level of drug.The method of the present invention so that operating personnel need to only shoot plate image, and the size of the medicament categories and its corresponding inhibition zone belonging to each drug sensitive test paper can be just identified from image so that automation bacteriostatic level identification is possibly realized.
Description
Technical field
The present invention relates to the technical field of bacteriostatic level identification, more particularly, to one kind based on plate experiment and deep learning
Bacteriostatic level recognition methods.
Background technology
Plate is also culture dish, is a kind of labware for being used for microorganism or cell culture, by a plane disc
The bottom of shape and a lid composition, are generally made of glass or plastics.Culture dish material is essentially divided into two classes, predominantly plastics and
Glass, the adhere-wall culture that can be used for vegetable material, microculture and zooblast of glass may also be used.Plastics
May be polythene material, once property and it is nonexpondable, be suitble to laboratory inoculation, line, the operation of separation of bacterial,
It can be used for the culture of vegetable material.
Drug sensitive test abbreviation drug sensitivity test (or resistance test).It is intended to understand pathogenic microorganism to various antibiotic
Sensitive (or tolerance) degree, clinical rational to be instructed to select the microbiology test of antibiotic medicine.If a kind of antibiotic with
The dosage of very little can inhibit, kill pathogenic bacteria, then claim this kind of pathogenic bacteria to the antibiotic " sensitivity ".Conversely, then it is known as " unwise
Sense " or " drug resistance ".To understand pathogenic bacteria to which kind of antibiotic sensitivity, with the rational use of medicines, blindness is reduced, should often carry out drug
Sensitization test.Substantially method is for it：The sample containing pathogenic bacteria is taken from the infection site of patient, is seeded in appropriate culture medium
On, it is cultivated under certain condition；The scraps of paper (i.e. susceptibility piece) for speckling with a certain amount of various antibiotic respectively are attached to culture medium simultaneously
Result is observed in surface (as shown in Figure 1), culture after a certain period of time.Since pathogenic bacteria are different to the sensitivity of various antibiotic,
Just occur different size of inhibition pathogen growth around susceptibility piece and form " sky circle ", be known as inhibition zone (as shown in Figure 1).
Inhibition zone size and pathogenic bacteria are proportional to the sensitivity of various antibiotic.Then can be directed to according to result of the test
Select antibiotic to property.Abuse of antibiotics at present causes anti-medicine bacterium to increase or even because long-term largely using broad-spectrum antibiotic, kills
Hinder internal microflora, lose the mutual restrictive function of microorganism, so that some non-causes rare or under normal circumstances
Germ amount reproduction causes the situation of so-called " superinfection " to occur repeatedly, and artificial difficulty is caused to treatment.Therefore, advocate
Using drug sensitivity test, adhere to that the rational use of medicines is particularly significant.
With the raising of computer performance and the propulsion of correlation theory so that the neutral net of training very multilayer becomes
It may.By stacked multilayer network struction deep neural network, more abstract expression is concentrated with learning training, here it is us
Known depth learning technology.Depth learning technology is quickly grown, in image procossing, speech recognition and natural language processing
The fields of grade achieve gratifying achievement, are the research hotspots in these current fields.Wherein in image processing field, behave oneself best
Deep learning model is convolutional neural networks (CNN).Convolutional neural networks are mainly made of convolutional layer and pond layer, while are also wrapped
The full articulamentum corresponding to classical neural network is included, main frame is as shown in Figure 2.
1) convolutional layer (Convolution)
Convolutional layer is part mostly important in a convolutional neural networks, each in convolutional layer different from full articulamentum
A fritter of the input of a node from last layer neutral net, such as the rectangle of 3*3.Convolutional layer passes through in neutral net
Each fritter carries out deep analysis so as to obtain the higher feature of level of abstraction.Most important part is filter in convolutional layer
(filter), as shown in figure 3, a sub- node matrix equation in current layer neutral net can be converted into next layer of nerve net by it
A unit-node matrix on network, i.e. length and width are all 1, but the node matrix equation that depth is unlimited.
2) pond layer (Pool)
In convolutional neural networks often between convolutional layer plus a special network structure, it is referred to as pond layer.Pond layer
The size of matrix can be effectively reduced, so as to reduce the parameter in last full articulamentum.Similar to convolutional layer, pond layer is also
It is completed by the structure of mobile one similar filter, but the calculating in the filter of pond layer is simple maximum
Or average value computing.It is referred to as maximum pond layer (max pooling) using the pond layer of maxima operation, is grasped using average value
The pond layer of work is referred to as averagely pond layer (average pooling).The filter of pond layer also has one compared with convolutional layer
Point is different, is exactly that the filter that convolutional layer uses is across entire depth, and the filter that pond layer uses only influences one
Node in depth, as shown in the figure 4.
Full articulamentum (Fully connected)
Full articulamentum, as its name suggests, i.e., each node of next layer neutral net is by whole in current layer neutral net
A node matrix equation converts to obtain, that is to say, that and next layer each node and all nodes of current layer have connection weight,
As shown in Figure 5.
To identify the bacteriostatic level of drug, need to distinguish in plate experiment the affiliated drug of drug sensitive test paper and its generate
The size of inhibition zone, afterwards according to medicament categories in its criterion of database search, according to criterion by inhibition zone size
Obtain the bacteriostatic level of drug.Traditionally, operating personnel need naked eyes to go to see the medicine name printed on drug sensitive test paper clearly, are used in combination
The survey tools such as vernier caliper measure the size of inhibition zone.These operations belong to cumbersome manual operation, occupy operator
The working time of member, consume great effort.If there is more easily and rapidly bacteriostatic level recognition methods, artificial behaviour can be saved
The spent time on work, then the work efficiency of operating personnel will be greatly improved, and can allow operating personnel that energy is more
Ground is placed on analysis experimental result rather than cumbersome measure stage.
The content of the invention
The shortcomings that it is a primary object of the present invention to overcome the prior art and deficiency provide a kind of based on plate experiment and deep
The bacteriostatic level recognition methods of study is spent, using depth learning technology, operating personnel's mobile phone or other digital equipments is only needed to clap
Microwell plate is taken the photograph, while microwell plate herbal species class and concentration are provided, it is corresponding just to provide drug automatically in the short period of time
Minimum inhibitory concentration.
In order to achieve the above object, the present invention uses following technical scheme：
The present invention is based on plate experiment and the bacteriostatic level recognition methods of deep learning, comprise the following steps：
S1, drug sensitive test paper position is determined：Original image is pre-processed, determines drug sensitive test paper position, convenient for measuring
Its inhibition zone size, and cut into single drug sensitive test paper image, the input as deep learning model；
S2, inhibition zone size is measured：The size of the inhibition zone is to be needed in plate experiment in the corresponding antibacterial standard of drug
The index to be used measures the diameter of inhibition zone using image processing techniques；
S3, construction sample set：To the sample set of deep learning Construction of A Model tape label, for training pattern；
S4, structure and training pattern：Structure can be used for identifying the model of drug sensitive test paper, and be instructed with training sample set
Practice；
S5, Model Identification drug sensitive test paper is used：Model training complete after operation, at this time model can receive training set
Image in addition judges the medicament categories belonging to it；
S6, the bacteriostatic level for judging drug：The medicament categories gone out by Model Identification inquire about corresponding suppression from database
Bacterium standard, then determine that the diameter of inhibition zone belongs to which of standard section, so as to obtain the bacteriostatic level of drug.
As preferred technical solution, in step S1, original image includes an entire plate, it is necessary to find each in plate
The position of drug sensitive test paper, could carry out the operation of follow-up corresponding inhibition zone identification and scraps of paper identification, and specific method is：
First, border circular areas where finding drug sensitive test paper using Hough transformation image processing techniques, after determining region, you can
Identification as inhibition zone size；
Secondly, in order to use the Model Identification scraps of paper, it is also necessary to cut down this block border circular areas, form single scraps of paper figure
Picture, as the input of model, for identifying the scraps of paper.
As preferred technical solution, the drug sensitive test paper is made by following methods：
By medicine name by the full whole filter paper of spacing line-spacing print of setting so that card punch no matter on filter paper which
Position is punched, and the disk for laying can be allowed to include entire medicine name.
As preferred technical solution, in step S2, by measuring inhibition zone size come the bacteriostatic level for the drug that quantizes,
It is described to identify that inhibition zone size is specially：
Inhibition zone is centered on drug sensitive test paper, therefore the center of circle by image processing techniques it has been determined that only need to scan susceptibility
Around piece, the radius of inhibition zone is obtained.
As preferred technical solution, in step S3, construction sample set includes construction authentic specimen collection and analog sample collection：
S3.1, authentic specimen collection：Here sample is the image under true shooting, only need to manually mark the affiliated drug of the scraps of paper
Species, sample can be successfully constructed；
S3.2, analog sample collection：Since the production method of drug sensitive test paper it is known that so can simulate similar on computers
Drug sensitive test paper image, being used for of the analog sample collection help deep learning model primary learning to arrive with sufficiently large sample set
The main feature of scraps of paper image helps model further to excavate scraps of paper image with the smaller authentic specimen collection of quantity again afterwards
Feature, and overcome the noise under true environment.
As preferred technical solution, in step S4, structure model includes the description below：
Thought based on transfer learning, using Google on ImageNet data sets trained Inception-v3 moulds
Type, the last full articulamentum of Inception-v3 models exports 1000 nodes, after Softmax layers, output
Model is moved in turbidity identification problem per a kind of probability in 1000 class samples in ImageNet data sets, is replaced
The full articulamentum of last layer of Inception-v3 models with Softmax layer, make new full articulamentum output node be 2, then
After Softmax layers, muddy and limpid difference probability can be exported, and the model other parts after building are kept and original
Inception-v3 models are identical.
As preferred technical solution, Inception structures are the important components in Inception-v3 models, it
Different convolutional layers is combined together by way of in parallel so that neutral net has different visual field scales, further
Improve the ability to express of model；
Concat layers are simple splicing layer, for multiple neural network node matrixes to be spliced into a node matrix equation；
Dropout layers are a kind of good model over-fitting issue-resolutions, it is based on the whole neutral net mould of training
Type, and it is average entirely gather as a result, particularly, with certain probability discard portion god in its each forward direction computing in training
Through member, only retain remaining neuron output operation result, that is, the output result for the neuron being rejected is 0；
Softmax layers are used to the output of neutral net becoming a probability distribution；
Assuming that the output of original neutral net is y1,y2,…,yn, then by Softmax layers processing after output be：
From formula as can be seen that by Softmax layer treated newly export meet being required for probability distribution.
As preferred technical solution, in step S3, using the method using substep training come training pattern, specifically include
Following step：
First, using analog sample collection training pattern, model primary learning is made to the main feature of scraps of paper image；
Secondly, reuse authentic specimen collection and continue training pattern, finely tune model parameter, make model learning to true picture
Feature can be applied to real scene；
During training, selection keeps the parameter constant of the all-network layer before last layer of full articulamentum, only updates most
The parameter of the full articulamentum of later layer.
As preferred technical solution, in step S5, the method for identifying drug sensitive test paper is：
S5.1, the probability distribution using the affiliated medicament categories of drug sensitive test paper in model prediction image；
S5.2, the medicament categories for obtaining maximum probability.
As preferred technical solution, in step S6, the method for judging the bacteriostatic level of drug is：
S6.1, antibacterial circle diameter unit conversion, i.e., be converted to millimeter by pixel；
S6.2, corresponding criterion in database is searched for according to medicine name；
S6.3, the antibacterial degree of corresponding drug is obtained by antibacterial circle diameter according to criterion.
The present invention basic principle be：
By handling plate image, inhibition zone size and the scraps of paper that can accurately obtain drug sensitive test paper in plate are corresponding
Medicament categories, and then obtain the bacteriostatic level of drug.Using this method, operating personnel need to only shoot the plate in bacteriostatic test,
The bacteriostatic level of drug can be just identified quickly.Compared with traditional recognition methods, method operation proposed in this paper is easier, and result
Reliably, while without special equipment, it is suitble to be generalized to most productive experiment place.In addition, the present invention also proposes that one kind is new
Drug sensitive test paper production method, be concentrated mainly on the innovation of text printout thinking on the scraps of paper, the manufacture difficulty of the scraps of paper can be reduced,
Production cost is reduced, and based on depth learning technology so that the accuracy rate of word is still maintained at higher water on the identification scraps of paper
It is flat.
Compared with prior art, the present invention having the following advantages that and advantageous effect：
1) convenience：Bacteriostatic level is identified so that operating personnel can be known as long as plate is shot based on deep learning
Not as a result, convenient and time-saving.
2) accuracy is high：The affiliated medicament categories of the scraps of paper are identified using deep learning and convolutional neural networks, in accuracy rate
Surmount traditional images Processing Algorithm, it is highly reliable.
3) sample is less needed for model：Utilize transfer learning thought structure and training pattern so that model is without magnanimity sample
This can also train good effect.
4) substep training：It is smaller true to reuse quantity collection for the larger analog sample collection training pattern of first usage quantity collection
Full pattern this collection training pattern so that model can learn on analog sample collection to preliminary feature, and it is smaller in quantity to reduce model
Authentic specimen collection on learning characteristic difficulty.
5) scraps of paper production cost is small：Use the scraps of paper production method that medicine name need not be accurately positioned so that the life of the scraps of paper
Produce cost reduction.And with the help of depth learning technology, identify that the accuracy rate of the new scraps of paper is still higher.
Description of the drawings
Fig. 1 is the schematic diagram of plate, inhibition zone and susceptibility piece；
Fig. 2 is the structure diagram of convolutional neural networks in the prior art；
Fig. 3 is convolutional layer filter (filter) structure diagram in Fig. 2；
Fig. 4 is layer filter schematic diagram in pond in Fig. 2；
Fig. 5 is full articulamentum schematic diagram in Fig. 2；
Fig. 6 is the flow chart the present invention is based on plate experiment and the bacteriostatic level recognition methods of deep learning；
Fig. 7 is present invention determine that sheet position and the cutting scraps of paper are used for model prediction schematic diagram；
The making schematic diagram of Fig. 8 drug sensitive test papers of the present invention；
Fig. 9 is present invention structure model structure schematic diagram；
Figure 10 is transfer learning schematic diagram of the present invention；
Figure 11 (a), Figure 11 (b) are the dropout schematic diagrames after the front and rear conversion of conversion respectively；
Figure 12 is the flow chart that the present invention uses the Model Identification scraps of paper；
Figure 13 is present invention determine that drug bacteriostatic level flow chart；
Figure 14 is the schematic diagram of present invention identification bacteriostatic level example 1；
Figure 15 is the schematic diagram of present invention identification bacteriostatic level example 2；
Figure 16 is the schematic diagram of present invention identification bacteriostatic level example 3.
Specific embodiment
With reference to embodiment and attached drawing, the present invention is described in further detail, but embodiments of the present invention are unlimited
In this.
Embodiment
As shown in fig. 6, the present embodiment is tested based on plate and the bacteriostatic level recognition methods of deep learning, including following steps
Suddenly：
S1, drug sensitive test paper position is determined：Original image is pre-processed, determines drug sensitive test paper position, convenient for measuring
Its inhibition zone size, and cut into single drug sensitive test paper image, the input as deep learning model；
Above-mentioned original image includes an entire plate, therefore needs to find the position of each drug sensitive test paper in plate, Cai Nengjin
The operation of the follow-up corresponding inhibition zone identification of row and scraps of paper identification.Specific method can use the image processing techniques such as Hough transformation
Border circular areas where finding drug sensitive test paper, after determining region, you can the identification as inhibition zone size；And in order to use Model Identification
The scraps of paper, it is also necessary to this block border circular areas be cut down, form single scraps of paper image, as the input of model, as shown in Figure 7.
Wherein, drug sensitive test paper is also the key point of the present embodiment, traditional drug sensitive test paper production method, by drug
Title be imprinted on the middles of the scraps of paper, be conveniently operated personnel's naked eyes and recognize, the problem of this method is to print toward big filter paper
It needs to be accurately positioned during upper medicine name, the position of medicine name on alignment filter paper is also required to when breaking into disk with card punch afterwards
It puts.This causes difficulty of the scraps of paper in production to increase, and increases cost of manufacture.For this purpose, the present invention proposes a kind of new scraps of paper system
Make method, production difficulty can be reduced in scraps of paper manufacturing process without positioning.It is closed specifically, this method presses medicine name
The full whole filter paper of spacing line-spacing print of reason so that card punch no matter which position punching on filter paper, can allow lay and
Disk include entire medicine name, as shown in Figure 8.Using this production method, filtered without accurately determining medicine name
Position on the scraps of paper, using, without the medicine name being deliberately aligned on filter paper, reducing the difficulty of making during card punch.
The drug sensitive test paper that this method makes with the naked eye can still distinguish, and in field of image recognition, be imprinted on compared with title
The scraps of paper of middle, identification difficulty want bigger, may be relatively low with traditional image recognition algorithm accuracy rate.For this purpose, the present invention makes
Drug sensitive test paper is identified with depth learning technology so that rate of accuracy reached has arrived sufficiently high level.
S2, inhibition zone size is measured：The size of the inhibition zone is to be needed in plate experiment in the corresponding antibacterial standard of drug
The index to be used measures the diameter of inhibition zone using image processing techniques；Bacteriostatic test in culture dish is big by measuring inhibition zone
The small bacteriostatic level come the drug that quantizes, so identification inhibition zone size is necessary for this method.Specifically, due to
Inhibition zone is centered on drug sensitive test paper, so the center of circle is it has been determined that remaining need scan medicine by image processing techniques
Around quick, the radius of inhibition zone is obtained.
S3, construction sample set：To the sample set of deep learning Construction of A Model tape label, for training pattern；
Step S3 main purposes are to identify the Construction of A Model sample set of the scraps of paper, since authentic specimen is very limited, institute
Two kinds of sample sets are constructed with this method, one kind is authentic specimen collection, and another kind is analog sample collection.
S3.1, authentic specimen collection：Here sample comes from the image under true shooting, only need to manually mark belonging to the scraps of paper
Medicament categories, sample can be successfully constructed.This is optimal sample, can correspond to real application scenarios, but sample is more difficult
It gathers, the requirement for reaching trained deep learning model is not enough in quantity.
S3.2, analog sample collection：Due to known to the production method of the scraps of paper, it is possible to simulate on computers similar
Scraps of paper image.This method can obtain countless scraps of paper images, although the later stage can manually add noise, (such as image revolves at random
Turn, random variation of brightness of image etc.), but the lower noise generated of true shooting can not be simulated completely.So analog sample collection
Main application is to help deep learning model primary learning to the main feature of scraps of paper image with sufficiently large sample set, afterwards
With the smaller authentic specimen collection of quantity model is helped further to excavate the feature of scraps of paper image again, and overcome under true environment
Noise.
S4, structure and training pattern：Structure can be used for identifying the model of drug sensitive test paper, and be instructed with training sample set
Practice；
According to the sample set obtained in step S3, it now is possible to deep learning model is built to train, in order to accelerate to train,
The present invention builds our model using the thought of transfer learning (Transfer Learning)；Transfer learning, it is possible to understand that
For trained model in a problem by simply adjusting, is made new similar with former problem to ask it is suitable for one
Topic.Similar to the mankind's study that can have ridden a bicycle by motorcycle, may learn than accent starts study by motorcycle
It is fast and good, transfer learning by the ability of trained extraction feature in a problem move to one it is similar new the problem of
On so that applied to it is new the problem of difficulty of the model in training significantly reduce.
Thought based on transfer learning, using Google on ImageNet data sets trained Inception-v3 moulds
Type, as shown in Figure 10, the last full articulamentum of the model export 1000 nodes, after Softmax layers, output
Per a kind of probability in 1000 class samples in ImageNet data sets.As shown in Figure 10, identified for model is moved to turbidity
In problem, the full articulamentum of last layer of the model and Softmax layer are replaced, the output node for making new full articulamentum is 2, then
After Softmax layers, muddy and limpid difference probability can be exported.And model other parts keep identical with master mould, therefore, it repaiies
Model structure after changing is as shown in Figure 9.
Deep learning model is since its scale of model is larger, to obtain good behaviour in process problem, it is necessary to sea
The labeled data of amount, such as ImageNet image classification datas are concentrated with 1,200,000 mark pictures, so could be by 152 layers
Accuracy of the model training of ResNet to 96.5%.Even if there is so more labeled data, the training time is also required to several days very
To the time in several weeks.At this point, in order to solve the problems, such as that training sample is less longer with the training time, it is possible to utilize transfer learning
Trained large-scale deep learning model is simply adjusted, such as simply replaces last layer and connect entirely by thought
Layer, can be applied to the problem of new, afterwards less sample only be needed to be trained, and be showed with regard to that can obtain model good enough.
S4.1, tectonic model：
There are many common point for such issues that identified due to image, be all need to detect the profile of objects in images with it is some
Specific structure, so it is reasonable that one trained model moves to another new image in original image problem
Identification problem can play certain effect, particularly when original image identification problem is more more complicated than new problem.ImageNet
The categorical measure of classification task is 1000, and the task more than the present invention is more complicated, so the training on ImageNet data sets
Good model is powerful enough in the ability for capturing picture key feature, and this model can be moved to our classification task
On, after being finely adjusted, it is possible to obtain a pretty good model of effect.
Therefore, the thought of model of the invention based on transfer learning, the training on ImageNet data sets using Google
Good Inception-v3 models replace the full articulamentum of last layer of this pre-training model and Softmax layers, can
Input sample is divided into two classes (muddy with limpid), the needs of so as to meet us.Model structure and transfer learning such as Fig. 9, figure
Shown in 10.
Below in Fig. 9 Inception structures, Concat layers, Dropout layers, Softmax layers make detailed analysis.
1) Inception structures；
Inception structures are the important components in Inception-v3 models.It passes through different convolutional layers
Mode in parallel is combined together so that neutral net has different visual field scales, further improves the expression energy of model
Power.
2) Concat layers；
Concat layers are a simple splicing layer, for multiple neural network node matrixes to be spliced into a node
Matrix usually in third dimension, i.e., splices in depth.
3) Dropout layers；
Dropout layers are a kind of good model over-fitting issue-resolutions, it is based on the whole neutral net mould of training
Type, and the average result entirely gathered.Particularly, it training when each forward direction computing in certain probability discard portion god
Through member, only retain remaining neuron output operation result, that is, the output result for the neuron being rejected is 0, as Figure 11 (a),
Shown in Figure 11 (b).
4) Softmax layers
Softmax layers of purposes is the output of neutral net becoming a probability distribution.
Assuming that the output of original neutral net is y1,y2,…,yn, then by Softmax layers processing after output be：
From formula as can be seen that by Softmax layer treated newly export meet being required for probability distribution.
The output of neutral net become probability distribution can let us intuitively understand neutral net perform classification appoint
Certainty factor during business judges that a problem of sample is much respectively, is mapped to us for different classes of probability is exactly muddy
Turbid and limpid probability is much respectively.Meanwhile the output of neutral net becomes probability distribution so that we can pass through intersection
Entropy calculates the gap between the probability distribution of prediction and the probability distribution of true answer, and thus, it is possible to definition, we train required
Loss function.
S4.2, training pattern；
After the completion of model construction, the labeled data obtained using a upper flow is trained model, due to true sample
This collection is less, so scheme of this method using substep training.First using analog sample collection training pattern, model is made tentatively to learn
Practise the main feature of scraps of paper image；Authentic specimen collection is reused afterwards and continues training pattern, is finely tuned model parameter, is made model
The feature of true picture is practised, real scene can be applied to.During training, selection is kept before the full articulamentum of last layer
All-network layer parameter constant, only update the parameter of last layer of full articulamentum.
The reason for so handling is：The negligible amounts of sample may be not enough to train whole network, be susceptible to serious
Over-fitting；Due to the parameter constant during the training period of the all-network layer before last layer of full articulamentum, so entirely
In training stage, a sample enters network model, and to the last before one layer of full articulamentum, the result being calculated is constant,
Still can keep in this as a result, such meter of each sample before last layer of full articulamentum is reached during the training period
Calculating just only needs to calculate once, this greatly accelerates trained speed.
S5, Model Identification drug sensitive test paper is used：Model training complete after operation, at this time model can receive training set
Image in addition judges the medicament categories belonging to it.
After the completion of model training, it is possible to identify the medicament categories belonging to the scraps of paper.When a piece of paper picture enters model
Afterwards, model will export its prediction result to the affiliated medicament categories of image, as a result be expressed as the probability of each possible drug.This
Invention can find out certainty factor of the model to prediction result from the probability of possible drug, therefore the drug of maximum probability is model
The affiliated drug of the scraps of paper thought；
As shown in figure 15, it is as follows using the Model Identification scraps of paper：
S5.1, the probability distribution with the affiliated medicament categories of drug sensitive test paper in model prediction image；
S5.2, the medicament categories for obtaining maximum probability.
S6, the bacteriostatic level for judging drug：The medicament categories gone out by Model Identification inquire about corresponding suppression from database
Bacterium standard, then determine that the diameter of inhibition zone belongs to which of standard section, so as to obtain the bacteriostatic level of drug.
In drug sensitive test, it is necessary to according to the size of inhibition zone as the standard for judging antibacterial degree height, such as CLSI/
Under NCCLs standards, it is muting sensitive (R) that the antibacterial circle diameter of Ofloxacin, which is less than or equal to 14 millimeters, 15 to 17 millimeters be in quick (I),
It is Gao Min (S) more than or equal to 18 millimeters.But different drug sensitive test papers has different criterion, it is necessary to according to medicine name
It goes to inquire about corresponding data information.
The step of front, has been obtained for the antibacterial circle diameter of drug sensitive test paper, corresponding medicine name, at this time can be with
Judge drug bacteriostatic level according to this two information.As shown in figure 16, it is as follows：
S6.1, antibacterial circle diameter unit conversion (pixel is converted to millimeter).
S6.2, corresponding criterion in database is searched for according to medicine name.
S6.3, the antibacterial degree of corresponding drug is obtained by antibacterial circle diameter according to criterion.
As figure-it is shown, the image that the present embodiment is tested using 3 plates identifies bacteriostatic level as example using this method,
Wherein S is Gao Min, and I is quick in being, R is muting sensitive.
The present invention can accurately obtain the inhibition zone size and the scraps of paper of drug sensitive test paper in plate by handling plate image
Corresponding medicament categories, and then obtain the bacteriostatic level of drug.Using this method, operating personnel need to be only shot in bacteriostatic test
Plate can just identify the bacteriostatic level of drug quickly.Compared with traditional recognition methods, method operation proposed by the present invention is simpler
Just, and reliable results, while without special equipment, it is suitble to be generalized to most productive experiment place.
Above-described embodiment is the preferable embodiment of the present invention, but embodiments of the present invention and from above-described embodiment
Limitation, other any Spirit Essences without departing from the present invention with made under principle change, modification, replacement, combine, simplification,
Equivalent substitute mode is should be, is included within protection scope of the present invention.
Claims (10)
1. based on plate experiment and the bacteriostatic level recognition methods of deep learning, which is characterized in that comprise the following steps：
S1, drug sensitive test paper position is determined：Original image is pre-processed, determines drug sensitive test paper position, convenient for measuring its suppression
Bacterium circle size, and cut into single drug sensitive test paper image, the input as deep learning model；
S2, inhibition zone size is measured：The size of the inhibition zone is to need to use in the corresponding antibacterial standard of drug in plate experiment
The index arrived measures the diameter of inhibition zone using image processing techniques；
S3, construction sample set：To the sample set of deep learning Construction of A Model tape label, for training pattern；
S4, structure and training pattern：Structure can be used for identifying the model of drug sensitive test paper, and be trained with training sample set；
S5, Model Identification drug sensitive test paper is used：Model training complete after operation, at this time model can receive beyond training set
Image, judge the medicament categories belonging to it；
S6, the bacteriostatic level for judging drug：The medicament categories gone out by Model Identification inquire about corresponding antibacterial mark from database
Standard, then determine that the diameter of inhibition zone belongs to which of standard section, so as to obtain the bacteriostatic level of drug.
2. according to claim 1 based on plate experiment and the bacteriostatic level recognition methods of deep learning, which is characterized in that step
In rapid S1, original image includes an entire plate, it is necessary to find the position of each drug sensitive test paper in plate, could carry out follow-up phase
The operation of the inhibition zone identification and scraps of paper identification answered, specific method are：
First, border circular areas where finding drug sensitive test paper using Hough transformation image processing techniques, after determining region, you can be used as
The identification of inhibition zone size；
Secondly, in order to use the Model Identification scraps of paper, it is also necessary to cut down this block border circular areas, form single scraps of paper image, make
For the input of model, for identifying the scraps of paper.
3. according to claim 1 or claim 2 existed based on plate experiment and the bacteriostatic level recognition methods of deep learning, feature
In the drug sensitive test paper is made by following methods：
By medicine name by the full whole filter paper of spacing line-spacing print of setting so that card punch no matter on filter paper which position
Punching can allow the disk for laying to include entire medicine name.
4. according to claim 1 based on plate experiment and the bacteriostatic level recognition methods of deep learning, which is characterized in that step
In rapid S2, by measuring inhibition zone size come the bacteriostatic level for the drug that quantizes, the identification inhibition zone size is specially：
Inhibition zone be centered on drug sensitive test paper, therefore the center of circle it has been determined that only need to pass through image processing techniques scan susceptibility piece week
It encloses, obtains the radius of inhibition zone.
5. according to claim 1 based on plate experiment and the bacteriostatic level recognition methods of deep learning, which is characterized in that step
In rapid S3, construction sample set includes construction authentic specimen collection and analog sample collection：
S3.1, authentic specimen collection：Here sample is the image under true shooting, only need to manually mark the affiliated drug kind of the scraps of paper
Class, sample can be successfully constructed；
S3.2, analog sample collection：Since the production method of drug sensitive test paper it is known that so can simulate similar medicine on computers
Quick scraps of paper image, the analog sample collection are used for sufficiently large sample set help deep learning model primary learning to the scraps of paper
The main feature of image helps model further to excavate the spy of scraps of paper image with the smaller authentic specimen collection of quantity again afterwards
Sign, and overcome the noise under true environment.
6. according to claim 1 based on plate experiment and the bacteriostatic level recognition methods of deep learning, which is characterized in that step
In rapid S4, structure model includes the description below：
Thought based on transfer learning, using Google on ImageNet data sets trained Inception-v3 models,
The last full articulamentum of Inception-v3 models exports 1000 nodes, after Softmax layers, exports ImageNet numbers
According to concentrating in 1000 class samples per a kind of probability, model is moved in turbidity identification problem, replaces Inception-v3
The full articulamentum of last layer of model with Softmax layer, the output node for making new full articulamentum is 2, then through Softmax layers
Afterwards, muddy and limpid difference probability can be exported, and the model other parts after building are kept and original Inception-v3 models
It is identical.
7. according to claim 6 based on plate experiment and the bacteriostatic level recognition methods of deep learning, which is characterized in that
Inception structures are the important components in Inception-v3 models, and different convolutional layers is passed through side in parallel by it
Formula is combined together so that neutral net has different visual field scales, further improves the ability to express of model；
Concat layers are simple splicing layer, for multiple neural network node matrixes to be spliced into a node matrix equation；
Dropout layers are a kind of good model over-fitting issue-resolutions, it is based on the whole neural network model of training, and
It is average entirely gather as a result, particularly, with certain probability discard portion neuron in its each forward direction computing in training,
Only retain remaining neuron output operation result, that is, the output result for the neuron being rejected is 0；
Softmax layers are used to the output of neutral net becoming a probability distribution；
Assuming that the output of original neutral net is y1,y2,…,yn, then by Softmax layers processing after output be：
<mrow>
<mi>S</mi>
<mi>o</mi>
<mi>f</mi>
<mi>t</mi>
<mi>m</mi>
<mi>a</mi>
<mi>x</mi>
<msub>
<mrow>
<mo>(</mo>
<mi>y</mi>
<mo>)</mo>
</mrow>
<mi>i</mi>
</msub>
<mo>=</mo>
<msubsup>
<mi>y</mi>
<mi>i</mi>
<mo>&prime;</mo>
</msubsup>
<mo>=</mo>
<mfrac>
<msup>
<mi>e</mi>
<msub>
<mi>y</mi>
<mi>i</mi>
</msub>
</msup>
<mrow>
<msubsup>
<mi>&Sigma;</mi>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>n</mi>
</msubsup>
<msup>
<mi>e</mi>
<msub>
<mi>y</mi>
<mi>j</mi>
</msub>
</msup>
</mrow>
</mfrac>
</mrow>
From formula as can be seen that by Softmax layer treated newly export meet being required for probability distribution.
8. according to claim 1 based on plate experiment and the bacteriostatic level recognition methods of deep learning, which is characterized in that step
In rapid S3, come training pattern, following step is specifically included using using the trained method of substep：
First, using analog sample collection training pattern, model primary learning is made to the main feature of scraps of paper image；
Secondly, reuse authentic specimen collection and continue training pattern, finely tune model parameter, make model learning to the spy of true picture
Sign, can be applied to real scene；
During training, selection keeps the parameter constant of the all-network layer before last layer of full articulamentum, only updates last
The parameter of the full articulamentum of layer.
9. according to claim 1 based on plate experiment and the bacteriostatic level recognition methods of deep learning, which is characterized in that step
In rapid S5, the method for identifying drug sensitive test paper is：
S5.1, the probability distribution using the affiliated medicament categories of drug sensitive test paper in model prediction image；
S5.2, the medicament categories for obtaining maximum probability.
10. according to claim 1 based on plate experiment and the bacteriostatic level recognition methods of deep learning, which is characterized in that
In step S6, the method for judging the bacteriostatic level of drug is：
S6.1, antibacterial circle diameter unit conversion, i.e., be converted to millimeter by pixel；
S6.2, corresponding criterion in database is searched for according to medicine name；
S6.3, the antibacterial degree of corresponding drug is obtained by antibacterial circle diameter according to criterion.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711190762.2A CN108090501B (en) | 2017-11-24 | 2017-11-24 | Bacteriostatic degree identification method based on plate experiment and deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711190762.2A CN108090501B (en) | 2017-11-24 | 2017-11-24 | Bacteriostatic degree identification method based on plate experiment and deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108090501A true CN108090501A (en) | 2018-05-29 |
| CN108090501B CN108090501B (en) | 2020-08-18 |
Family
ID=62172931
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711190762.2A Active CN108090501B (en) | 2017-11-24 | 2017-11-24 | Bacteriostatic degree identification method based on plate experiment and deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108090501B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108981597A (en) * | 2018-07-04 | 2018-12-11 | 昆明金域医学检验所有限公司 | A kind of inhibition zone automatic measuring instrument |
| CN109002855A (en) * | 2018-07-20 | 2018-12-14 | 长沙湘丰智能装备股份有限公司 | A kind of identification method of the fermentation of black tea degree based on convolutional neural networks |
| CN110243731A (en) * | 2019-07-03 | 2019-09-17 | 苏州新实医疗科技有限公司 | Antibacterial circle diameter dynamic measurement method, measuring device and readable storage medium storing program for executing |
| CN110288596A (en) * | 2019-07-03 | 2019-09-27 | 苏州新实医疗科技有限公司 | Antibacterial circle diameter method for fast measuring, measuring device and readable storage medium storing program for executing |
| CN110838088A (en) * | 2018-08-15 | 2020-02-25 | Tcl集团股份有限公司 | Multi-frame noise reduction method and device based on deep learning and terminal equipment |
| US11373287B2 (en) | 2018-09-06 | 2022-06-28 | Accenture Global Solutions Limited | Digital quality control using computer visioning with deep learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20060062440A1 (en) * | 2001-06-14 | 2006-03-23 | The Regents Of The University Of California | High-throughput, dual probe biological assays based on single molecule detection |
| CN102880880A (en) * | 2012-09-24 | 2013-01-16 | 河海大学常州校区 | Method and equipment for automatically identifying medicine sensitivity reaction on basis of fuzzy neural network |
| US20150056605A1 (en) * | 2013-08-21 | 2015-02-26 | Purdue Research Foundation | Identification of blood based metabolite biomarkers of pancreatic cancer |
| CN104749180A (en) * | 2015-03-25 | 2015-07-01 | 北京浩辰星月科技有限公司 | System for visually identifying and analyzing as well as identifying drug allergy and drug allergy detecting method |
| CN106238112A (en) * | 2016-08-27 | 2016-12-21 | 宋波 | A kind of micro-fluidic chip and the application in the qualification and drug sensitive experiment of pathogen thereof |
-
2017
- 2017-11-24 CN CN201711190762.2A patent/CN108090501B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20060062440A1 (en) * | 2001-06-14 | 2006-03-23 | The Regents Of The University Of California | High-throughput, dual probe biological assays based on single molecule detection |
| CN102880880A (en) * | 2012-09-24 | 2013-01-16 | 河海大学常州校区 | Method and equipment for automatically identifying medicine sensitivity reaction on basis of fuzzy neural network |
| US20150056605A1 (en) * | 2013-08-21 | 2015-02-26 | Purdue Research Foundation | Identification of blood based metabolite biomarkers of pancreatic cancer |
| CN104749180A (en) * | 2015-03-25 | 2015-07-01 | 北京浩辰星月科技有限公司 | System for visually identifying and analyzing as well as identifying drug allergy and drug allergy detecting method |
| CN106238112A (en) * | 2016-08-27 | 2016-12-21 | 宋波 | A kind of micro-fluidic chip and the application in the qualification and drug sensitive experiment of pathogen thereof |
Non-Patent Citations (2)
| Title |
|---|
| MAJID MASSO等: "A Novel Sequence-Structure Approach for Accurate Prediction of Resistance to HIV-1 Protease Inhibitors", 《2007 IEEE 7TH INTERNATIONAL SYMPOSIUM ON BIOINFORMATICS AND BIOENGINEERING》 * |
| 贺靖冬等: "半自动细菌分析专家系统在细菌药敏试验中的应用", 《检验医学》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108981597A (en) * | 2018-07-04 | 2018-12-11 | 昆明金域医学检验所有限公司 | A kind of inhibition zone automatic measuring instrument |
| CN109002855A (en) * | 2018-07-20 | 2018-12-14 | 长沙湘丰智能装备股份有限公司 | A kind of identification method of the fermentation of black tea degree based on convolutional neural networks |
| CN110838088A (en) * | 2018-08-15 | 2020-02-25 | Tcl集团股份有限公司 | Multi-frame noise reduction method and device based on deep learning and terminal equipment |
| CN110838088B (en) * | 2018-08-15 | 2023-06-02 | Tcl科技集团股份有限公司 | Multi-frame noise reduction method and device based on deep learning and terminal equipment |
| US11373287B2 (en) | 2018-09-06 | 2022-06-28 | Accenture Global Solutions Limited | Digital quality control using computer visioning with deep learning |
| CN110243731A (en) * | 2019-07-03 | 2019-09-17 | 苏州新实医疗科技有限公司 | Antibacterial circle diameter dynamic measurement method, measuring device and readable storage medium storing program for executing |
| CN110288596A (en) * | 2019-07-03 | 2019-09-27 | 苏州新实医疗科技有限公司 | Antibacterial circle diameter method for fast measuring, measuring device and readable storage medium storing program for executing |
| CN110243731B (en) * | 2019-07-03 | 2021-09-24 | 苏州新实医疗科技有限公司 | Method and device for dynamically measuring diameter of bacteriostatic zone and readable storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108090501B (en) | 2020-08-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108090501A (en) | Based on plate experiment and the bacteriostatic level recognition methods of deep learning | |
| CN107133960A (en) | Image crack dividing method based on depth convolutional neural networks | |
| CN108596757A (en) | A kind of personal credit file method and system of intelligences combination | |
| CN109977780A (en) | A kind of detection and recognition methods of the diatom based on deep learning algorithm | |
| CN106651830A (en) | Image quality test method based on parallel convolutional neural network | |
| CN108288035A (en) | The human motion recognition method of multichannel image Fusion Features based on deep learning | |
| CN106874956A (en) | The construction method of image classification convolutional neural networks structure | |
| CN109829399A (en) | A kind of vehicle mounted road scene point cloud automatic classification method based on deep learning | |
| CN108805833B (en) | Miscellaneous minimizing technology of copybook binaryzation ambient noise based on condition confrontation network | |
| CN107330446A (en) | A kind of optimization method of depth convolutional neural networks towards image classification | |
| CN107154048A (en) | The remote sensing image segmentation method and device of a kind of Pulse-coupled Neural Network Model | |
| CN103996195A (en) | Image saliency detection method | |
| CN108681706A (en) | A kind of double source remotely-sensed data semantic segmentation method | |
| CN108090502A (en) | Minimum inhibitory concentration recognition methods based on deep learning | |
| CN109919012A (en) | A kind of indicative microorganism image-recognizing method of sewage treatment based on convolutional neural networks | |
| CN109102498A (en) | A kind of method of cluster type nucleus segmentation in cervical smear image | |
| CN106803124A (en) | Field migration extreme learning machine method based on manifold canonical and norm canonical | |
| CN106777402A (en) | A kind of image retrieval text method based on sparse neural network | |
| US11807551B2 (en) | Systems and methods for treating wastewater | |
| CN109214407A (en) | Event detection model, calculates equipment and storage medium at method, apparatus | |
| CN110276445A (en) | Domestic communication label category method based on Inception convolution module | |
| CN110288515A (en) | The method and CNN coloring learner intelligently coloured to the microsctructural photograph of electron microscope shooting | |
| CN109978074A (en) | Image aesthetic feeling and emotion joint classification method and system based on depth multi-task learning | |
| CN114782982A (en) | Marine organism intelligent detection method based on deep learning | |
| CN103354073B (en) | A kind of LCD color deviation correction method |
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 |