CN112184697B - Diabetic retinopathy grading deep learning method based on drosophila optimization algorithm - Google Patents
Diabetic retinopathy grading deep learning method based on drosophila optimization algorithm Download PDFInfo
- Publication number
- CN112184697B CN112184697B CN202011103209.2A CN202011103209A CN112184697B CN 112184697 B CN112184697 B CN 112184697B CN 202011103209 A CN202011103209 A CN 202011103209A CN 112184697 B CN112184697 B CN 112184697B
- Authority
- CN
- China
- Prior art keywords
- annihilation
- subgroup
- formula
- iteration
- subgroups
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 71
- 238000004422 calculation algorithm Methods 0.000 title claims abstract description 69
- 238000005457 optimization Methods 0.000 title claims abstract description 53
- 206010012689 Diabetic retinopathy Diseases 0.000 title claims abstract description 23
- 241000255581 Drosophila <fruit fly, genus> Species 0.000 title claims abstract description 22
- 238000013135 deep learning Methods 0.000 title claims abstract description 17
- 230000007246 mechanism Effects 0.000 claims abstract description 51
- 230000008569 process Effects 0.000 claims abstract description 37
- 238000012937 correction Methods 0.000 claims abstract description 18
- 241000255588 Tephritidae Species 0.000 claims description 24
- 238000012549 training Methods 0.000 claims description 22
- 230000006870 function Effects 0.000 claims description 19
- 238000011156 evaluation Methods 0.000 claims description 18
- 238000004364 calculation method Methods 0.000 claims description 12
- 206010012601 diabetes mellitus Diseases 0.000 claims description 10
- 238000013136 deep learning model Methods 0.000 claims description 8
- 230000003902 lesion Effects 0.000 claims description 8
- 238000013527 convolutional neural network Methods 0.000 claims description 6
- 201000010099 disease Diseases 0.000 claims description 6
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 claims description 6
- 238000004891 communication Methods 0.000 claims description 5
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 claims description 4
- 230000002431 foraging effect Effects 0.000 claims description 4
- 238000012360 testing method Methods 0.000 claims description 4
- 208000017442 Retinal disease Diseases 0.000 claims description 3
- 206010038923 Retinopathy Diseases 0.000 claims description 3
- 230000001174 ascending effect Effects 0.000 claims description 3
- 239000000047 product Substances 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 238000012546 transfer Methods 0.000 claims description 3
- 238000013526 transfer learning Methods 0.000 claims description 3
- 230000006399 behavior Effects 0.000 claims description 2
- 230000001105 regulatory effect Effects 0.000 claims description 2
- 238000006467 substitution reaction Methods 0.000 claims description 2
- 239000013589 supplement Substances 0.000 claims description 2
- 230000003044 adaptive effect Effects 0.000 claims 1
- 238000002790 cross-validation Methods 0.000 abstract description 6
- 238000010586 diagram Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 241000255925 Diptera Species 0.000 description 2
- 235000013305 food Nutrition 0.000 description 2
- 230000019637 foraging behavior Effects 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000035772 mutation Effects 0.000 description 2
- 230000002028 premature Effects 0.000 description 2
- 230000002123 temporal effect Effects 0.000 description 2
- 206010017577 Gait disturbance Diseases 0.000 description 1
- BQCADISMDOOEFD-UHFFFAOYSA-N Silver Chemical compound [Ag] BQCADISMDOOEFD-UHFFFAOYSA-N 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000000977 initiatory effect Effects 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000010845 search algorithm Methods 0.000 description 1
- 229910052709 silver Inorganic materials 0.000 description 1
- 239000004332 silver Substances 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 238000005309 stochastic process Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30041—Eye; Retina; Ophthalmic
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Business, Economics & Management (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Medical Informatics (AREA)
- Quality & Reliability (AREA)
- Bioinformatics & Computational Biology (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Economics (AREA)
- Geometry (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Entrepreneurship & Innovation (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Computer Hardware Design (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a drosophila optimization algorithm-based diabetic retinopathy hierarchical deep learning method, which realizes an annihilation and escape mechanism of a dynamic subgroup in a searching process of a multi-subgroup for the first time, and organically fuses a multi-subgroup, column-dimensional search and Gaussian boundary correction mechanism, thereby improving the capability of searching a global optimal solution and reducing the risk of trapping into a local optimal solution. Comparison of MALBFOA-DL with the benchmark model on the recall data through 10 rounds of cross validation on the diabetic retinopathy image dataset used by the model, it is evident that the MALBFOA-DL model compared with the same-level benchmark model (using VGG-16) yielded better results almost every round.
Description
Technical Field
The invention relates to the field of artificial intelligence, in particular to a drosophila nonequilibrium problem optimization algorithm enhanced by multiple mechanisms, and more particularly to a diabetic retinopathy grading deep learning method based on the drosophila optimization algorithm.
Background
The original fruit fly optimization algorithm (hereinafter referred to as FOA) is a global optimization algorithm which is inspired by foraging behaviors of fruit flies for foraging by using vision and smell and further develops the performance. The algorithm model simulates the situation that fruit flies in the real world find food and fruit fly individuals search for things respectively, and achieves the situation that the whole population completes iterative approach search to food in a relatively efficient mode through cooperation and information exchange of the group. In practical application, the original algorithm and the related improved algorithm have the advantages of simple structure, few control parameters and easy understanding. However, the convergence speed and the solution quality of the original FOA and various improved algorithms on the multimode problem, the asymmetric problem and the complex problem are still not ideal, the problem of low mutation probability in the process of finding the optimal solution exists, the model search space is limited, and the situation that the original FOA algorithm and the improved algorithms are difficult to find the local optimal solution and cannot find the global optimal solution or find a plurality of optimal solutions occurs.
Meanwhile, the diabetic retinopathy fundus data volume is large (> 50G), and the category data volume is highly unbalanced, so that when the existing various deep learning methods are adopted for optimization, the selection of hyper-parameters is extremely difficult, and the optimal selection is difficult.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides a fruit fly algorithm (MALBEAA) reinforced based on various systems (a column dimension mechanism, a boundary rebound mechanism, a multi-subgroup cooperation mechanism and a subgroup annihilation mechanism), which can effectively improve the convergence speed of the FOA algorithm, improve the mutation probability in the direction and step length in the solution optimization process, avoid removing the local optimal solution and improve the quality of the solution; the deep learning method (MALBEAA-DL) optimized by the improved consequence fly optimization algorithm also has a good effect on the grading of the lesion grade of the diabetic fundus lesion image data set.
The technical scheme for realizing the purpose of the invention is as follows:
a diabetic retinopathy hierarchical deep learning method based on a drosophila optimization algorithm comprises the following steps:
(1) Space search mechanism based on column-dimensional flight
The step distribution of the step size of the column dimension is adopted in the dim-dimension search space, and is expressed as the following formula (1):
where dim is the dimension of the search space, β represents an important column dimension index for adjusting stability, and u and v follow a standard normal distribution, as shown in equation (2):
wherein, the son and mother N outside the number indicates that the probability distribution defined by the current formula is the standard normal distribution.
In the formula σ u And σ v Can be calculated from equation (3):
wherein Γ represents a gamma function;
(2) Boundary-crossing correction mechanism based on Gaussian distribution
By observing the behavior of the fruit flies in the foraging process, some fruit flies in the fruit fly population fly for a longer distance, part of parameters can cross the boundary determined by the upper boundary ub and the lower boundary lb of the search space in the algorithm, the position coordinate X of the fruit fly individual is corrected back to the correct range, and the transboundary correction of the fruit flies beyond the range is realized through the following formula (4):
wherein lb and ub are the upper and lower bounds of the search space, respectively, in which the dimension is dim; when the fruit fly crosses the boundary, the correction is immediately carried out to return to the boundary, or the correction process is moved to the range within the boundary, and the formula is expressed as the formula (5):
d t ~N(0,1,dim) (5)
wherein d represents the return distance and t represents the fruit fly number crossing the boundary;
(3) Enhancement mechanism based on multi-subgroup cooperation
1) The improved sample division method comprises the following steps: dividing the population into M individual subgroups of Drosophila with the same number, wherein the coordinate X of the ith object of the mth subgroup m,i The initialization process of (2) is as in equation (6):
in the formula: x axis,m For the initial coordinates of the m-th subgroup, X, obtained randomly in the search space 0 Is the center of the search space obtained by equation (7); r is 0 To the initial search radius according to equation (8),represents a dot product
X 0 =(lb(:)+ub(:))/2 (7)
R 0 =(ub(:)-lb(:))/2 (8)
Where (: indicates that the same operation is performed on elements in lb and ub that correspond to the same dimension.
2) And (3) improving the calculation process of the discriminant variable: a method for randomly obtaining coordinate points searched in different subgroups as judgment variables by using a column-dimension-based flight algorithm, as shown in formula (9):
in the formula: r is the search radius in each iteration according to an equation, and the calculation process is as the formula (10); radio _ levy is a scale factor of a column-dimensional random process described by an equation, and the calculation process is as formula (11):
R=R 0 *((nit-it)/nit)^pa (10)
nit represents the total iteration number, the total iteration number is calculated according to an annihilation mechanism, it represents the currently executed iteration number, and Pa is an iteration scale factor of a search space:
3) Improving the subgroup coordination mechanism: an improved subgroup cooperation mechanism defined by formula (12) is adopted among subgroups to be used as a supplement of the multi-subgroup mechanism, so as to search the optimal solution after each iteration of the algorithm:
will coordinate X new Substituting into fitness function to calculate communication smell value Fit Communitcate If Fit Communitcate If the global fitness SmellBest is less than the obtained global fitness SmellBest, fit is used Communitcate Updating global fitness and assigning BestPos value to X new And (4) completing the optimal position updating:
(4) Algorithm operation optimization mechanism based on subgroup annihilation
1) Initializing subgroup annihilation parameters: setting an iterative ratio stopcopies for subgroup annihilation; setting annihilation proportion killRatio of M subgroups in each annihilation iteration process; randomly selecting a part of annihilated molecule groups according to a fixed proportion reliveRatio, reviving the molecule groups into normal subgroups and continuing iteration; with the above parameters, the total iteration parameter nit obtained from the annihilation mechanism can be represented by equation (14):
2) Evaluation of annihilation efficiency: (ii) updating the ascending order of the rate in descending order of the fitness, the subgroups being sorted in descending order of the annihilation time, the subgroups with higher rank among them being added to the annihilation group (kills); annihilation evaluation will be performed after nit 'iterations (nit' = nit × 10%);
3) Annihilation is carried out: after nit 'iterations (nit' = nit × 10%), annihilation operation of subgroups is performed, subgroups belonging to the annihilation array stop further optimization, annihilation parameters of the subgroups are updated, and iteration times of an evaluation function are reduced to obtain a better optimization direction;
4) Annihilation escape: while the subgroup annihilation operation is performed in step 3), a part of the annihilated subgroup is randomly selected at a fixed ratio (reliveRatio) according to the equation to reactivate and participate in the iterative optimization again, in the same manner as in equation (15):
in the formula: the revive is a subgroup revived again and participating in iteration, and theta represents the difference set of the revive and the iteration;
(5) Integrating the steps (1) - (4) into the existing FOA, and carrying out operation optimization according to an MALBEAA algorithm;
(6) In the MALBEAA algorithm, the time complexity is analyzed, the iteration times of nit are estimated, and a subgroup annihilation mechanism is executed;
(7) MALBEA-DL algorithm
Three key hyper-parameters in the deep convolutional neural network CNN were optimized using the MALBFOA model to generate a new deep learning framework MALBFOA-DL:
in a diabetes fundus lesion grading deep learning model based on a drosophila optimization algorithm, a pre-trained acceptance v3 is used as a basis, the last 2-4 layers of connection layers and a classifier are subjected to transfer training, and a model capable of better finishing classification of a diabetes retinopathy image data set is obtained;
1) The data set was processed to examine the distribution and severity of the eyes according to the following scale: 0-no DR, 1-mild, 2-moderate, 3-severe, 4-proliferative DR, the clinician distinguished the presence of diabetic retinopathy in each image, and classified each image as to whether it is the left or right eye and which grade falls within the scale; the distribution of the severity of diabetic retinopathy is very uneven and polarized, with: the number of samples for stage 1, stage 2, stage 3, stage 4, and stage 5 are 25810, 2443, 5292, 873, and 708, respectively;
2) The diabetes fundus lesion hierarchical deep learning model based on the drosophila optimization algorithm is used, the transfer learning of the Incepisov 3 model is completed for the last 2-4 layers in a trainable mode, and the depth features are extracted from the original image; in model training, randomly dividing data into training sets and testing sets according to the ratio of 9:1, dividing the training sets into 10 rounds, balancing the distribution of the training sets, dividing each training set into five types according to the severity of diseases, and then performing substitution method sampling in different types; the sample size for each disease category was 1000.
The important column dimension index for regulating stability in the step (1) is 1.5.
The moving distance of the fruit fly in the process of returning to the boundary in the step (2) follows the Dim-dimensional Gaussian distribution, and the probability density of the Gaussian distribution is expressed by a formula (16):
in the formula, dist i The values are randomly distributed, and X and Y represent the abscissa and ordinate, respectively, of each drosophila in the current reference system. Modeling in FOA algorithmJudgment value of odor concentration of target (S) i ) The definition is shown in formula (17), and according to the definition of the original model, the value range of Si is larger than 0:
S i =1/Dist i (17)
step (2) the correction process is represented by formula (18):
the MALBFOA algorithm of step (5) is as follows:
object: minimum target odor concentration
The fitness function f (x), the maximum number of evaluations MaxFEs, the overall size popSize, the dimension dim, the search range lb, ub,
and outputting the optimal odor concentration SmellBest and the optimal position BestPos.
(1) Initializing parameters: popsize, maxFEs, M, subgroup annihilation rate killRatio, original location of search space R0, original radius of search space R1, iterative scaling factor Pa of search space;
(2) Initializing initial positions X of M subgroups axis,m (1≤m≤M)
For m=1 to M
For i=1 to popsize
Initializing the ith position X of the mth subgroup by equation (1) m,i ;
Mixing X m,i Substituting the fitness function into the Smell to calculate the Smell;
End For
[bestSmell,bestIndex]=min(Smell);
X axis,m =X(bestIndex);
groupBestSmell m =bestSmell;
groupBestPos m =X(bestIndex);
End For
[SmellBest,BestPos]=min(groupBestSmell);
(3) Evaluating an annihilation mechanism according to equation (14), calculating the iteration number nit, and initializing the iteration number of the optimal olfactory array IterationSmell of nit;
(4) Calculating a scale factor Radio _ levy according to a column dimension random process described by a formula (11);
(5) Iterative optimization:
while(iterationit<nit)
as described above, the annihilation of the subgroups is carried out with the annihilation rates killRatio of the M subgroups to obtain an annihilation array kill;
dynamically updating the radius R of each iteration according to the formula (10);
R=R 0 ×((nit-it)/nit) pa
For m=1to M
if m in annihilation array kills
Continuing;
End if
For i 1 to popsize
updating X according to the column dimension mechanism described in equation (9) m,i ;
Correcting the out-of-range object according to formula (18);
calculating Smellvalue, and determining the number of current objects in fitness function
End For
[bestSmell,bestIndex]=min(Smell);
X axis,m =X(bestIndex);
groupBestSmell m =bestSmell;
groupBestPos m =X(bestIndex);
Updating annihilation parameters of the mth subgroup;
End For
[SmellBest,BestPos]=min(groupBestSmell);
obtaining the inter-subgroup communication condition according to the description of a formula (12);
will substitute for the coordinate X new Substituting into fitness function to calculate the value of the alternating odor Fit communitcate .
If Fit communitcate <SmellBest
SmellBest=Fit communitcate ;
BestPos=X new ;
End
iteration=iteration+1;
IterationSmell m =SmellBest;
End While
(6) Min is returned (IterationSmell).
The invention has the beneficial effects that: in the algorithm, an annihilation and escape mechanism of the dynamic subgroup in the searching process of the multiple subgroups is realized for the first time, and the multiple subgroups, column dimension searching and a Gaussian boundary correction mechanism are organically integrated, so that the capability of searching for the global optimal solution is improved, and the risk of trapping into the local optimal solution is reduced. Specifically, each subgroup searches solution space independently according to a column-dimensional flight mechanism and a Gaussian cross-boundary correction mechanism, the quality of the solution space is improved, the search efficiency of the solution space is optimized, and the problem that an original model is easy to fall into local optimum is solved; on the basis, a subgroup annihilation mechanism is adopted to prune the model optimization process, which subgroup continues to participate in iteration (optimization direction) is dynamically determined, and which subgroups are annihilated, so that more evaluation time and calculation capacity are saved, and the method is used in the direction with stronger convergence and more optimization, so that the convergence speed of the algorithm is improved, and the aim of improving the efficiency of the algorithm is fulfilled; the boundary Gaussian correction mechanism can further save the problem of computation waste of the search algorithm on the boundary, and the search direction of the solution is optimized towards the boundary in a Gaussian distribution mode. Comparison of MALBFOA-DL with the benchmark model on the recall data through 10 rounds of cross validation on the diabetic retinopathy image dataset used by the model, it is evident that the MALBFOA-DL model compared with the same-level benchmark model (using VGG-16) yielded better results almost every round.
Drawings
FIG. 1 is a flow chart of the MALBEPA algorithm according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a parameter optimization of a Deep Learning (DL) classification model using a proposed MALBOFA model according to an embodiment of the present invention;
FIG. 3 is a flow chart of the MALBFOA-DL model according to an embodiment of the present invention;
FIG. 4 is an architecture diagram of the overall framework of the MALBEFA-DL model according to the embodiment of the present invention;
FIG. 5 is a graph of a comparison of the recall rate data between MALBFOA-DL and the benchmark model by performing 10 rounds of cross validation on a diabetic retinopathy data set according to an embodiment of the present invention.
Detailed Description
The invention is further illustrated by the following figures and examples, but is not limited to the content of the invention.
Example (b):
a diabetic retinopathy grading deep learning method based on a drosophila optimization algorithm comprises the following steps:
1. space search mechanism based on column dimension flight
The column-dimensional flight mechanism is generally used for improving the meta-heuristic algorithm, and is essentially a random non-gaussian traversal, and the traversal characteristics are as follows: most traversal steps are distributed over a smaller range, while in a few cases a large adjustment occurs. The specific stride distribution is similar to the reported road-seeking situation of the drunk people going home. Drunk persons typically walk with unstable stance and stumbling strides, but most can still return home with little disorientation, with their course of action following roughly a rank dimensional mechanism, with few exceptions. From the above example, the following dimensional flight mechanism has good optimization properties when the target object is under an irregular moving model.
The step distribution (in the dim-dimensional search space) of the column-dimensional step adopted by the method can be expressed as the following formula:
where dim is the dimension of the search space, and β represents an important column dimension index (1.5 in common) for adjusting stability, s represents the step size, and u and v follow a standard normal distribution, as shown in equation (2) below:
wherein sigma u And σ v The specific value can be calculated by formula (3). Equation (3) is as follows:
in formula (3), Γ represents a gamma function.
2. Boundary-crossing correction mechanism based on Gaussian distribution
By observing the foraging behavior of the flies, it can be seen that some flies in the fly population may fly a greater distance during foraging, and in particular in the optimization algorithm, it is reflected in that some of the parameters may cross the boundary defined by the upper (ub) and lower (lb) bounds of the search space. In this case, it is necessary to correct the position coordinates (X) of the individual drosophila (i.e., the model optimization object) back to the correct range. In general, correction for out of range fruit flies can be achieved by the following equation (4):
where lb and ub are the upper and lower bounds of the search space, respectively, and may be in multiple dimensions (e.g., dim dimensions). In previous researches on drunk people, the drunk people can immediately turn back once finding that the drunk people are in a strange place in the process of returning home. In most cases an intoxicated person will move back a small distance and then seek home again, but occasionally it will happen that he suddenly moves back a large distance. Inspired by this, when a fruit fly crosses a boundary, it should also immediately return to the boundary, even moving directly to a range a little inside the boundary. It is generally considered that the moving distance of such a return process approximately follows a gaussian distribution in the dim dimension, which is expressed as formula (5):
d t ~N(0,1,dim) (5)
in the formula: d represents the return distance and t represents the individual drosophila crossing the border.
The new correction process in the correction with the above-described movement distance can be represented by equation (6).
As is known today, this is the first time that this method is employed in the improvement of the FOA algorithm to correct the model optimization object.
3. Multi-subgroup enhancement mechanism
In order to solve the problem that the FOA algorithm may be difficult to obtain a global optimal solution due to being trapped in a local optimal solution, in the method of the present invention, a new multi-subgroup enhancement mechanism is proposed, and three improvement points involved therein will be specifically described below:
(1) Improved sample partitioning method
The method enhances the diversity of the solution and realizes effective exploration of the whole designated area by dividing the population into M subgroups (usually 10) with the same number of drosophila individuals, thereby avoiding the situation of falling into local optimum or premature (premature convergence). The initialization process of the coordinates (Xm, i) of the ith object of the mth subgroup is as shown in formula (7):
where Xaxis, m is the initial coordinate of the m-th subgroup obtained randomly in the search space. X0 is the center of the search space obtained by equation (8). R0 is the initial search radius according to equation (9),representing dot products (multiplication by one item)
X 0 =(lb(:)+ub(:))/2 (8)
R 0 =(ub(:)-lb(:))/2 (9)
(2) Improved discriminant variable calculation:
in fact, it is obvious that the values of Disti in equation (10) are randomly distributed over a large range. However, the odor concentration judgment value (Si) of the model target in the FOA algorithm (defined in formula (11)) is limited to a smaller range, and more importantly, according to the definition of the original model, the value range of Si is constantly larger than 0:
S i =1/Dist i (11)
this results in the model being able to search for an optimal solution only in a space where all coordinate values are greater than 0 in different dimensions, resulting in the design space and solution having moved or rotated and being asymmetric. This is the main reason why the FOA algorithm is inefficient and even difficult to solve the optimization problem.
In order to solve the problem, the method disclosed by the patent uses a method for randomly obtaining coordinate points searched in different subgroups based on a column-dimensional flight algorithm as judgment variables. As in equation (12):
in the formula: r is the search radius in each iteration according to the equation, which is calculated as equation (13); radio _ levy is a scale factor of a column-dimensional stochastic process described by the equation, and the calculation process is as in equation (14):
R=R 0 *((nit-it)/nit)^pa (13)
in the formula: nit represents the total number of iterations; it represents the number of currently executed iterations, and a calculation formula thereof is given below; pa is an iterative scaling factor of the search space with a constant value (typically 4).
By using the method, the calculation process of the judgment variable can be effectively simplified, and the proposed symmetric and asymmetric optimization problems are solved, so that the calculation efficiency of the original FOA algorithm is improved, and the applicability of the algorithm is effectively improved.
(3) Improved subgroup coordination mechanism
In order to further improve the capability of the model in finding the global optimal solution. The patented method is intended to employ an improved subgroup cooperation mechanism defined by equation (15). It is used as a complement to the multi-subgroup enhancement mechanism described above to search for the optimal solution after each iteration of the algorithm.
The coordinates Xnew are substituted into the fitness function to calculate the communication scent value fitcommurtate. If the FitCommunicate is less than the obtained global fitness SmellBest, then the global fitness is updated using the FitCommunicate, and the BestPos value is assigned to Xnew to complete the optimal location update, as shown in equation (16).
4. Algorithm operation optimization mechanism based on subgroup annihilation
In most cases, the temporal complexity of the evolutionary optimization algorithm is mainly reflected in the number of times the evaluation function is called. On a fixed number of evaluations (MaxFE), it is worth thinking how to save valuable computing power and get more optimization directions.
The entire subgroup annihilation process will be described in detail from the following steps:
(1) The subgroup annihilation parameters are initialized. First, the iteration ratio stopcopies for the subgroup annihilation is set (typically 90% of the total number of iterations). Next, the annihilation ratio killRatio (4/5) for the M subgroups is set at each annihilation iteration process. Third, a fraction of the population of molecules that are annihilated is randomly selected at a fixed ratio (reliveRatio), then revived to a normal subgroup and the iteration continues. With these three parameters, the total iteration parameter nit from the annihilation mechanism can be represented by equation (17):
(2) Annihilation efficiency assessment. The method uses three parameters of fitness value, update rate (the fewer the fitness is improved, the larger the update rate) and the annihilation time of a subgroup as criteria for evaluating whether a subgroup is to be annihilated. First, all subgroups are sorted in descending order of adaptability, ascending order of update rate, descending order of annihilation time of the subgroups, and the subgroup with the higher rank (highest killRatio) among them is added to the annihilation group (kills). Annihilation evaluation will be performed after nit 'iterations (nit' = nit × 10%).
(3) Annihilation is performed. After nit 'iterations (nit' = nit × 10%), the annihilation operation of the subgroup will be performed. Subgroups (kills) belonging to the annihilation array will stop further optimization and update their annihilation parameters, thereby saving computational resources (i.e., reducing the number of iterations of the evaluation function) for the remaining subgroups with more convergent performance, and obtaining a better optimization direction.
(4) Annihilation escapes. In order to increase the activity of the subgroups, the model optimizes the robustness of the quality and expands the search space of the solution, while the subgroup annihilation operation is performed in step three, a part of the annihilated subgroups will be randomly selected according to the equation in a fixed ratio (reliveRatio) to revive and participate in the iterative optimization again. The specific selection mode is as formula (18):
in the formula: revive represents the subgroup revived again and participating in the iteration; Θ represents the difference between the two.
5. MALBEFA algorithm process
In this section, the four mechanisms described above are integrated into the existing classical FOA, and the entire MALBFOA operation optimization process is described in detail according to algorithm 1. Fig. 2 shows a flow chart of the MALBFOA algorithm. The pseudo code for MALBFOA is shown in algorithm 1.
6. Algorithm time complexity analysis
The computational complexity of MABLFOA is mainly determined by the maximum number of evaluations (MaxFEs), which is a constant and depends in part on the population (popsize), the dimension size (dim), the number of subgroups (M), and the maximum number of evaluations (MaxFE). The iteration number nit is determined by MaxFE, M and the annihilation rate (killRatio) of the mth subgroup according to equation (11). With a fixed number of evaluations (MaxFE), the efficiency of the proposed MABLFOA is mainly in dynamically determining which subgroups are still alive (optimization direction) and which are destroyed, thus saving more evaluation time (computing power) for a more optimization direction.
In the proposed algorithm, the temporal complexity analysis is mainly focused on three steps: initialization, estimating the number of iterations of nit and performing the subgroup annihilation mechanism. Wherein O (initiation) = O (M × pop size × dim) × onerandom time, O (Levy flight) = O (nit × M × pop size × dim) × oneLevyTime, O (timing front flight) = O (M × pop size + nit × M × pop size × (1-kill ratio)) × onefixedtime. oneFindnestaTime represents the time to perform an evaluation (calculate fitness value) and oneLevyTime represents the time required to perform a column dimension operation. Therefore, we can derive O (MABLFOA) = O (M × popsize × dim) × oneRandTime + O (nit × M × popsize × dim) × oneLevyTime + O (M × popsize + nit × M × popsize × (1-killRatio)) × oneFitnessTime. In general, applying the proposed algorithm to a more complex evaluation function (one training time is larger) will save more training time.
Application of MALBEPA algorithm in diabetic retinopathy hierarchical deep learning model optimization
In this example, we optimized three key hyper-parameters in the deep Convolutional Neural Network (CNN) using the proposed MALBFOA model to generate a new deep learning framework (MALBFOA-DL). And the recommended MALBEA-DL is adopted under the detection data set of the diabetic retinopathy, so that a better effect is achieved. In order to verify the effectiveness of the method, ten-fold cross verification method is adopted to compare the method. The above-mentioned MALBEA-DL Framework experiment is carried out on Linux 64 bit (Red Hat 4.8.5-11)Performed on Xeon Silver 4116CPU @2.10GHz,250GB RAM and 32G NVIDIA GPU (V100), the correlation model algorithm was encoded using Pycharm 2017.3.4 (professional edition) using as input a color diabetic retinopathy fundus image dataset (size about 82.23 GB) sorted by severity level.
In the diabetes fundus lesion grading deep learning model based on the drosophila optimization algorithm, a pre-trained acceptance v3 is used as a basis, the last 2-4 layers of connection layers and a classifier are subjected to transfer training, and a model capable of better finishing classification of a diabetes retinopathy image data set is obtained.
The data set needs to be processed first, examining the distribution (left and right eye) and severity of the eyes (graded by 5 levels in the original data). According to the following scale: 0-no DR, 1-mild, 2-moderate, 3-severe, 4-proliferative DR, the clinician distinguished the presence of diabetic retinopathy in each image, and classified each image as to whether it was left or right and to which grade in the above scale it belongs. The distribution of diabetic retinopathy severity (level) is very uneven and polarized, with: the number of samples for stage 1, stage 2, stage 3, stage 4 and stage 5 are 25810, 2443, 5292, 873 and 708, respectively.
And then, the diabetic fundus lesion hierarchical deep learning model based on the drosophila optimization algorithm provided by the patent is used, the transfer learning of the Incepisionv 3 model is completed on the last 2-4 layers in a trainable mode, and the depth features are extracted from the original image. In the training of the model, we trained 9:1 randomly dividing the data into training sets and testing sets, dividing the training sets into 10 rounds, balancing the distribution of the training sets, dividing each training set into five categories according to the severity of diseases, and then performing alternative method sampling in different categories. The sample size for each disease category was 1000. The final equilibrium results are shown in figure 2.
FIG. 3 shows the model framework of the above proposed MALBFOA-DL. The model consists of two main parts: the first part dynamically adjusts the parameter n (number of nodes in the complete connectivity layer) in the Deep Learning (DL) model by MALBFOA. In the second section, the DL with optimized parameters was passed through 10 rounds of cross-validation to verify the classification Accuracy (ACC) of the model. In model training, 90% of the data set is used for training the model, 10% of the data set is used for testing the model, then 10 rounds of cross validation are carried out, and the average value is used as the final prediction accuracy; fig. 4 shows an architecture diagram of the overall framework of the MALBFOA-DL model.
Comparison of MALBFOA-DL with the benchmark model on the recall data through 10 rounds of cross validation on the diabetic retinopathy image dataset used by the model, it is evident that the MALBFOA-DL model compared with the same-level benchmark model (using VGG-16) yielded better results almost every round. Particularly at levels 2, 4 and 5, the recall rate of the reference model is nearly 0 in this range. Specifically as shown in FIG. 5 on The left (The benchmark model); in contrast to the MALBEFA-DL model, better results than the original model were obtained in this range, as shown in FIG. 5 (chart subject MALBEFA-DL), and the effectiveness of the improved method was visually confirmed.
Claims (3)
1. A diabetic retinopathy grading deep learning method based on a drosophila optimization algorithm is characterized by comprising the following steps: the method comprises the following steps:
(1) Space search mechanism based on column-dimensional flight
A step distribution of column-dimensional step sizes is adopted in the dim-dimensional search space, which is expressed as the following formula (1):
where dim is the dimension of the search space, β represents an important column dimension index for adjusting stability, and u and v follow a standard normal distribution, as shown in equation (2):
wherein, the son-mother N outside the number indicates that the probability distribution defined by the current formula is standard normal distribution;
in the formula σ u And σ v Can be calculated from equation (3):
wherein Γ represents a gamma function;
(2) Boundary-crossing correction mechanism based on Gaussian distribution
By observing the behavior of the fruit flies in the foraging process, some fruit flies in the fruit fly population fly out a longer distance, part of parameters can cross the boundary determined by the upper boundary ub and the lower boundary lb of the search space in the algorithm, the position coordinate X of the fruit fly individual is corrected back to the correct range, and the transboundary correction of the fruit flies out of the range is realized through the following formula (4):
wherein lb and ub are the upper and lower bounds of the search space, respectively, and wherein lb and ub are dim dimensions; when the fruit flies cross the boundary, the correction is immediately carried out to return to the boundary, or the correction process moves to the range within the boundary, and the formula is expressed as the formula (5):
d t ~N(0,1,dim) (5)
wherein d represents the return distance and t represents the fruit fly number crossing the boundary;
(3) Enhancement mechanism based on multi-subgroup cooperation
1) The improved sample division method comprises the following steps: dividing the population into M individual subgroups with the same number of drosophila, wherein the coordinate X of the ith object of the mth subgroup m,i The initialization process of (2) is as in equation (6):
in the formula: x axis,m For the initial coordinate, X, of the m-th subgroup randomly obtained in the search space 0 Is the center of the search space obtained by equation (7); r 0 To the initial search radius according to equation (8),representing dot products
X 0 =(lb(:)+ub(:))/2 (7)
R 0 =(ub(:)-lb(:))/2 (8)
2) And (3) improving the calculation process of the discriminant variable: a method for randomly obtaining coordinate points searched in different subgroups as judgment variables by using a column-dimension-based flight algorithm, as shown in formula (9):
in the formula: r is the search radius in each iteration according to the equation, and the calculation process is as the formula (10); radio _ levy is a scale factor of a column-dimensional random process described by the equation, and the calculation process is as in formula (11):
R=R 0 *((nit-it)/nit)^pa (10)
nit represents the total iteration number, the total iteration number is calculated according to an annihilation mechanism, it represents the currently executed iteration number, and Pa is an iteration scale factor of a search space:
3) Improving subgroup coordination mechanism: an improved subgroup cooperation mechanism defined by formula (12) is adopted among subgroups to be used as a supplement of the multi-subgroup mechanism, so as to search the optimal solution after each iteration of the algorithm:
will coordinate X new Substituting into fitness function to calculate communication smell value Fit Communitcate If Fit Communitcate If the global fitness SmellBest is less than the obtained global fitness SmellBest, fit is used Communitcate Updating global fitness and assigning BestPos value to X new And (4) completing the optimal position updating:
(4) Algorithm operation optimization mechanism based on subgroup annihilation
1) Initializing subgroup annihilation parameters: setting an iterative ratio stopcopies for subgroup annihilation; setting annihilation proportion killRatio of M subgroups in each annihilation iteration process; randomly selecting a part of annihilated molecule groups according to a fixed proportion reliveRatio, reviving the molecule groups into normal subgroups and continuing iteration; with the above parameters, the total iteration parameter nit obtained from the annihilation mechanism can be represented by equation (14):
2) Evaluation of annihilation efficiency: all subgroups are sorted according to an adaptive descending order and an ascending order of update rate, and the subgroups are sorted according to a descending order of annihilation time, wherein the subgroup with higher grade is added into an annihilation group kills; annihilation assessment will be performed after nit 'iterations, where nit' = nit 10%;
3) Annihilation is carried out: after nit' times of iteration, annihilation operation of subgroups is carried out, further optimization of subgroups in the annihilation array is stopped, annihilation parameters of the subgroups are updated, and the number of iteration times of an evaluation function is reduced to obtain a better optimization direction;
4) Annihilation escape: while the subgroup annihilation operation is performed in step 3), randomly selecting a part of annihilated subgroups with a fixed ratio according to the equation, reactivating the annihilated subgroups and participating in iterative optimization again, wherein the selection is as shown in formula (15):
in the formula: the revive is a subgroup revived again and participating in iteration, and theta represents the difference set of the revive and the iteration;
(5) Integrating the steps (1) - (4) into the existing FOA, and carrying out operation optimization according to an MALBEAA algorithm; the MALBFOA algorithm is as follows:
object: minimum target odor concentration
The fitness function f (x), the maximum number of evaluations MaxFEs, the overall size popSize, the dimension dim, the search range lb, ub,
and outputting the optimal odor concentration SmellBest and the optimal position BestPos.
1) Initializing parameters: popsize, maxFEs, M, subgroup annihilation rate killRatio, original location of search space R0, original radius of search space R1, iterative scaling factor Pa of search space;
2) Initializing initial positions X of M subgroups axis,m Wherein M is more than or equal to 1 and less than or equal to M
For m=1 to M
For i=1 to popsize
Initializing the ith position X of the mth subgroup by equation (9) m,i ;
X is to be m,i Substituting the fitness function to calculate Smell;
End For
[bestSmell,bestIndex]=min(Smell);
X axis,m =X(bestIndex);
groupBestSmell m =bestSmell;
groupBestPos m =X(bestIndex);
End For
[SmellBest,BestPos]=min(groupBestSmell);
3) Evaluating an annihilation mechanism according to equation (14), calculating the iteration number nit, and initializing the iteration number of the optimal olfactory array IterationSmell of nit;
4) Calculating a scale factor Radio _ levy according to a column dimension random process described by a formula (11);
5) Iterative optimization:
while(iterationit<nit)
as described above, the annihilation of the subgroups is carried out with the annihilation rates killRatio of the M subgroups to obtain an annihilation array kill;
dynamically updating the radius R of each iteration according to the formula (10);
R=R 0 ×((nit-it)/nit) pa
For m 1 to M
if m in annihilation array kills
Continuing;
End if
For i 1 to popsize
updating X according to the column dimension mechanism described in equation (9) m,i ;
Calculating Smellvalue, and determining the number of current objects in fitness function
End For
[bestSmell,bestIndex]=min(Smell);
X axis,m =X(bestIndex);
groupBestSmell m =bestSmell;
groupBestPos m =X(bestIndex);
Updating annihilation parameters of the mth subgroup;
End For
[SmellBest,BestPos]=min(groupBestSmell);
obtaining the inter-subgroup communication situation according to the description of formula (12);
will substitute for the coordinate X new Substituting into fitness function to calculate the value of the alternating odor Fit communitcate .
If Fit communitcate <SmellBest
SmellBest=Fit communitcate ;
BestPos=X new ;
End
iteration=iteration+1;
IterationSmell m =SmellBest;
End While
6) Return min (IterationSmell);
(6) In the MALBEOA algorithm, time complexity is analyzed, iteration times of nit are estimated, and a subgroup annihilation mechanism is executed;
(7) MALBEFA-DL algorithm
Three key hyperparameters in the deep convolutional neural network CNN are optimized using the MALBFOA model to generate a new deep learning framework MALBFOA-DL:
in a diabetes fundus lesion grading deep learning model based on a drosophila optimization algorithm, a pre-trained acceptance v3 is used as a basis, the last 2-4 layers of connection layers and a classifier are subjected to transfer training, and a model capable of better finishing classification of a diabetes retinopathy image data set is obtained;
1) The data set was processed to examine the distribution and severity of the eyes according to the following scale: 0-no DR, 1-mild, 2-moderate, 3-severe, 4-proliferative DR, the clinician distinguished the presence of diabetic retinopathy in each image, and classified each image as to whether it was left or right and to which grade in the above scale it belongs; the distribution of diabetic retinopathy severity is very uneven and polarized, with: the number of samples of the 1 st, 2 nd, 3 rd, 4 th and 5 th stages is 25810, 2443, 5292, 873 and 708, respectively;
2) The diabetes fundus lesion hierarchical deep learning model based on the drosophila optimization algorithm is used, the transfer learning of the Incepisov 3 model is completed for the last 2-4 layers in a trainable mode, and the depth features are extracted from the original image; in model training, randomly dividing data into training sets and testing sets according to the ratio of 9:1, dividing the training sets into 10 rounds, balancing the distribution of the training sets, dividing each training set into five types according to the severity of diseases, and then performing substitution method sampling in different types; the sample size for each disease category was 1000.
2. The drosophila optimization algorithm-based diabetic retinopathy hierarchical deep learning method according to claim 1, which is characterized in that: the important column dimension index for regulating stability in the step (1) is 1.5.
3. The drosophila optimization algorithm-based diabetic retinopathy hierarchical deep learning method according to claim 1, which is characterized in that: the moving distance of the fruit fly in the process of returning to the boundary in the step (2) follows the Dim-dimensional Gaussian distribution, and the probability density of the Gaussian distribution is expressed by a formula (16):
in the formula, dist i The values are randomly distributed, X and Y respectively represent the abscissa and the ordinate of each fruit fly in the current reference system; odor concentration determination value (S) of model target in FOA algorithm i ) The definition is shown in a formula (17), and the value range of Si is large according to the definition of the original modelIn the ratio of 0:
S i =1/Dist i (17)
step (2) the correction process is represented by equation (18):
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011103209.2A CN112184697B (en) | 2020-10-15 | 2020-10-15 | Diabetic retinopathy grading deep learning method based on drosophila optimization algorithm |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011103209.2A CN112184697B (en) | 2020-10-15 | 2020-10-15 | Diabetic retinopathy grading deep learning method based on drosophila optimization algorithm |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112184697A CN112184697A (en) | 2021-01-05 |
CN112184697B true CN112184697B (en) | 2022-10-04 |
Family
ID=73949645
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011103209.2A Active CN112184697B (en) | 2020-10-15 | 2020-10-15 | Diabetic retinopathy grading deep learning method based on drosophila optimization algorithm |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112184697B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108021916A (en) * | 2017-12-31 | 2018-05-11 | 南京航空航天大学 | Deep learning diabetic retinopathy sorting technique based on notice mechanism |
CN108615051A (en) * | 2018-04-13 | 2018-10-02 | 博众精工科技股份有限公司 | Diabetic retina image classification method based on deep learning and system |
CN110837803A (en) * | 2019-11-07 | 2020-02-25 | 复旦大学 | Diabetic retinopathy grading method based on depth map network |
WO2020176039A1 (en) * | 2019-02-26 | 2020-09-03 | Ngee Ann Polytechnic | System and method for classifying eye images |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP3616120A1 (en) * | 2017-04-27 | 2020-03-04 | Retinascan Limited | System and method for automated funduscopic image analysis |
-
2020
- 2020-10-15 CN CN202011103209.2A patent/CN112184697B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108021916A (en) * | 2017-12-31 | 2018-05-11 | 南京航空航天大学 | Deep learning diabetic retinopathy sorting technique based on notice mechanism |
CN108615051A (en) * | 2018-04-13 | 2018-10-02 | 博众精工科技股份有限公司 | Diabetic retina image classification method based on deep learning and system |
WO2020176039A1 (en) * | 2019-02-26 | 2020-09-03 | Ngee Ann Polytechnic | System and method for classifying eye images |
CN110837803A (en) * | 2019-11-07 | 2020-02-25 | 复旦大学 | Diabetic retinopathy grading method based on depth map network |
Non-Patent Citations (3)
Title |
---|
A SAR Image Classification Algorithm Based on Multi-Feature Polarimetric Parameters Using FOA and LS-SVM;Shiyu Luo等;《IEEE Access》;20191204;第7卷;175259-175276 * |
基于迁移学习的糖尿病视网膜病变辅助诊断算法;黄义劲等;《中华实验眼科杂志》;20190810;第37卷(第08期);603-607 * |
融合注意力机制和高效网络的糖尿病视网膜病变识别与分类;张子振等;《中国图象图形学报》;20200812;第25卷(第08期);1708-1718 * |
Also Published As
Publication number | Publication date |
---|---|
CN112184697A (en) | 2021-01-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Jia et al. | A new hybrid seagull optimization algorithm for feature selection | |
WO2021155706A1 (en) | Method and device for training business prediction model by using unbalanced positive and negative samples | |
AU2018101317A4 (en) | A Deep Learning Based System for Animal Species Classification | |
CN106560848B (en) | Novel neural network model for simulating biological bidirectional cognitive ability and training method | |
CN108898213B (en) | Adaptive activation function parameter adjusting method for deep neural network | |
CN111105045A (en) | Method for constructing prediction model based on improved locust optimization algorithm | |
Fu et al. | An ensemble unsupervised spiking neural network for objective recognition | |
CN111126549A (en) | Double-star spectrum fitting method based on strategy improved goblet and sea squirt intelligent algorithm | |
CN114861838B (en) | Intelligent classification method for pulsatile neural brains based on neuron complex dynamics | |
CN116015967B (en) | Industrial Internet intrusion detection method based on improved whale algorithm optimization DELM | |
CN114724705A (en) | Esophageal squamous carcinoma survival prediction method based on improved ant lion algorithm and BP neural network | |
Wang et al. | Optimizing deep transfer networks with fruit fly optimization for accurate diagnosis of diabetic retinopathy | |
Abdull Hamed et al. | Quantum-inspired particle swarm optimization for feature selection and parameter optimization in evolving spiking neural networks for classification tasks | |
CN112184697B (en) | Diabetic retinopathy grading deep learning method based on drosophila optimization algorithm | |
Pradhan | Cardiac image-based heart disease diagnosis using bio-inspired optimized technique for feature selection to enhance classification accuracy | |
CN116886398B (en) | Internet of things intrusion detection method based on feature selection and integrated learning | |
Leinonen | Improvements to short-term weather prediction with recurrent-convolutional networks | |
CN115759346B (en) | Carbon emission prediction method, device and equipment based on AO algorithm | |
Mousavirad et al. | How effective are current population-based metaheuristic algorithms for variance-based multi-level image thresholding? | |
CN108491923A (en) | Based on the pest image-recognizing method for improving wolf pack algorithm optimization Bayesian network | |
Mohamed et al. | Adaptive trainer for multi-layer perceptron using artificial gorilla troops optimizer algorithm | |
CN110534160B (en) | Method for predicting protein solubility by convolutional neural network | |
Boniecki | The Kohonen neural network in classification problems solving in agricultural engineering | |
Demosthenous et al. | Deep reinforcement learning for improving competitive cycling performance | |
CN117784615B (en) | Fire control system fault prediction method based on IMPA-RF |
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 | ||
EE01 | Entry into force of recordation of patent licensing contract | ||
EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20210105 Assignee: Guilin Qinghong Technology Co.,Ltd. Assignor: GUILIN University OF ELECTRONIC TECHNOLOGY Contract record no.: X2023980044782 Denomination of invention: A Hierarchical Deep Learning Method for diabetes Retinopathy Based on Drosophila Optimization Algorithm Granted publication date: 20221004 License type: Common License Record date: 20231031 |