CN110929775A - Convolutional neural network weight optimization method for retinopathy classification - Google Patents

Convolutional neural network weight optimization method for retinopathy classification Download PDF

Info

Publication number
CN110929775A
CN110929775A CN201911127264.2A CN201911127264A CN110929775A CN 110929775 A CN110929775 A CN 110929775A CN 201911127264 A CN201911127264 A CN 201911127264A CN 110929775 A CN110929775 A CN 110929775A
Authority
CN
China
Prior art keywords
frog
weight
value
loss
convolutional neural
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
Application number
CN201911127264.2A
Other languages
Chinese (zh)
Other versions
CN110929775B (en
Inventor
丁卫平
任龙杰
孙颖
鞠恒荣
丁帅荣
曹金鑫
张毅
冯志豪
李铭
文万志
胡彬
赵理莉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nantong University
Original Assignee
Nantong University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Nantong University filed Critical Nantong University
Priority to CN201911127264.2A priority Critical patent/CN110929775B/en
Publication of CN110929775A publication Critical patent/CN110929775A/en
Application granted granted Critical
Publication of CN110929775B publication Critical patent/CN110929775B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • 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
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10024—Color image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30041—Eye; Retina; Ophthalmic

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)
  • Eye Examination Apparatus (AREA)

Abstract

The invention relates to the field of medical information intelligent processing, in particular to a convolutional neural network weight optimization method for retinopathy classification. Firstly, acquiring a fundus image training set and a multi-lesion label corresponding to the fundus image training set; searching an optimal initial weight through a single swarm leaping algorithm, then constructing a convolution layer, a pooling layer and a full-link layer in a convolutional neural network, and taking the optimal initial weight as a parameter for the first forward propagation calculation; respectively carrying out cross entropy loss calculation on four predicted values of four pathological changes in retina and a true value, summing to obtain a loss value, judging whether the loss value is abnormal, if so, generating a frog group around a previous forward propagation weight, and searching for an optimal frog updating network weight; otherwise, updating the network weight by adopting a gradient descent algorithm; and finally optimizing the final weight. The invention can effectively improve the accuracy of fundus image multi-lesion detection and has stronger application value to retinal diseases and adjuvant therapy.

Description

Convolutional neural network weight optimization method for retinopathy classification
Technical Field
The invention relates to the field of medical information intelligent processing, in particular to a convolutional neural network weight optimization method for retinopathy classification.
Background
Color fundus images are the most basic examination modality for diagnosing ophthalmic diseases. Meanwhile, the fundus images can enable people to discover various eye diseases such as glaucoma, optic neuritis, macular degeneration and the like as early as possible, and the treatment is convenient, timely and effective. Early diagnosis and timely treatment can effectively reduce the prevalence rate. However, because the number of Chinese population is huge and the number of ophthalmologists is relatively limited, and a large amount of time is consumed for diagnosing the ophthalmic diseases by only depending on the doctors, other methods for large-scale screening are urgently needed. The computer-aided diagnosis not only can greatly reduce the workload of doctors, but also has the advantages of objectivity, rapidness, accuracy and the like.
With the rapid increase of the amount of annotated data and the great improvement of the performance of the graphics processor unit in recent years, the research of the convolutional neural network is rapidly started, and the advantages in the field of image classification are also reflected. However, due to the complexity of the fundus image, the gradient descent algorithm is easy to fall into a local optimal solution by adopting the traditional convolutional neural network to carry out multi-label classification on the fundus image. Therefore, a new method is urgently needed to effectively deal with the complexity of the fundus image and improve the accuracy of detecting the retinopathy.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides
In order to achieve the purpose, the invention adopts the following technical scheme:
a convolutional neural network weight optimization method for retinopathy classification comprises the following steps:
step 1, inputting a fundus image training set and a label, wherein the training set is X ═ X (X)1，x2，...，xn) N is 1, 2, 3 …, and the label is B (B)1，b2，…，bn) 1, 2, 3 …, label b corresponding to fundus imageiOne-hot coding is carried out to obtain a true value y _ truei；
Step 2, initializing the weight parameters of the convolutional neural network, generating m frogs by adopting standard normal distribution, sequencing according to adaptive values, and finding out the optimal frogs fbWorst and worst frog fwContinuously updating the worst frog position and reordering until the single frog leaping algorithm meets the convergence condition to obtain the global optimal frog fqA 1 is to fqAs the initial weight of the convolutional neural network;
step 3, inputting the jth fundus image in the training set into a convolutional neural network forward propagation calculation model, performing multilayer convolution and pooling, finally performing two-layer full-connection calculation, and outputting to obtain the jth predicted value y _ predictj；
Step 4, the jth true value y _ true is processedjAnd the j-th predicted value y _ predictjPerforming softmax calculation for softmax (y _ predict)j) Performing cross entropy loss calculation to obtain the jth loss value lossj；
Step 5, judging the jth loss value lossjIf the abnormal weight is judged, correcting the abnormal weight by adopting a single swarm leaping algorithm, and taking the corrected weight as a new weight of the network, otherwise, continuing to execute the next step;
step 6, judging whether the network reaches the end condition, if so, generating a frog cluster around the initial frog by taking the final weight of the network as the initial frog, and carrying out frog jump optimization to obtain the global optimal frog fqbThe weight parameter is the weight parameter of the final training of the algorithm;
step 7, the trained weight parameters are brought into a network, fundus images are input, two-layer full-connection calculation is finally carried out through multilayer convolution and pooling, and the predicted value y for detecting various pathological changes in retina is outputr。
Further, as a preferred technical solution of the present invention, the specific steps of step 2 are as follows:
step 2.1, determining the number m of frogs, and generating all frogs by adopting standard normal distribution, wherein the standard normal distribution formula is as follows:
Figure BDA0002277252030000021
in formula (1), μ represents a position parameter, σ represents a scale parameter, x represents a random variable, and each frog f comprises an ownership value parameter in the network;
2.2, constructing a convolutional neural network forward propagation calculation model, and randomly selecting a small number w of fundus images from the fundus image training set as reference images;
step 2.3, the frog without loss value calculation is brought into a convolutional neural network forward propagation calculation model to calculate the loss value of the frog, wherein the loss value calculation function is a fitness function of a single group frog leaping algorithm, and the loss calculation formula is as follows:
Figure BDA0002277252030000022
in the formula (2), p represents a network output value, t represents a true value, s represents the dimension of each group of lesion labels, and b represents the number of types of retinopathy needing to be detected simultaneously. Sequencing all frogs in ascending order according to fitness function to obtain the optimal frogs fbWorst and worst frog fw；
And 2.4, updating the position of the worst frog through a position updating function, wherein the position updating function is as follows:
D＝(fb+fp-fw)×Rand(0，1.2) (3)
fnew＝fw+D (4)
Figure BDA0002277252030000031
in the formula (3), fpRepresenting an offset, the dimension of which is the same as the dimension of each frog; in the formula (4), fnewRepresenting the updated frog; in the formula (5), fpiDenotes fpThe value in the ith dimension;
and 2.5, judging whether the single frog leap algorithm meets the convergence condition, if so, stopping the algorithm, and taking the value of the optimal frog as the initial weight of the convolutional neural network, otherwise, turning to the step 2.3.
Further, as a preferred technical solution of the present invention, the specific steps of step 5 are as follows:
step 5.1, judging the loss value loss calculated at the j-1 th timej-1And judging the loss value loss calculated at the j-th timejAbsolute value of the difference of | lossj-1-lossjIf the | is larger than the threshold epsilon, executing the next step, otherwise, not executing the frog leap calculation;
step 5.2, taking the weight value in the j-1 th network as the initial frog wbAround wbThe frog group is generated according to the following formula:
wij＝wbj+0.0001×Rand(-1，1)(1＜i≤c-1) (6)
in the formula (6), wijIndicates the ith cyan of generationThe value in the jth dimension of the frog, c representing the total number of frogs;
step 5.3, constructing a convolutional neural network forward propagation calculation model, and randomly selecting a small number w of fundus images from a training image library as reference images;
step 5.4, bringing the frogs with the loss values not calculated into a convolutional neural network forward propagation calculation model, calculating the loss values of the frogs and sequencing the frogs;
step 5.5, updating the worst frog position, calculating the updated loss value, and reordering the frog group;
step 5.6, judging whether the single swarm frog-leaping algorithm meets the convergence condition, if so, ending the frog-leaping calculation, and obtaining the overall optimal frog fqbAnd (4) taking the corrected new weight value into the network, otherwise, turning to the step 5.4.
Further as a preferred technical solution of the present invention, the convolutional neural network forward propagation computation model in step 3 includes a convolutional layer, a pooling layer, and a fully connected layer.
Compared with the prior art, the convolutional neural network weight optimization method for classifying the retinopathy has the following technical effects by adopting the technical scheme:
the invention provides a convolutional neural network weight optimization method for retinopathy classification, which optimizes the initialization weight, the abnormal weight and the final weight of a convolutional neural network by adopting a single swarm frog-leaping algorithm, can effectively deal with the complexity of fundus images, and improves the detection accuracy of retinopathy.
The invention can search the optimal initial value of the network through the single swarm frog-leap algorithm, reduce the initial loss value of the convolutional neural network for classifying the multiple lesions of the fundus image, improve the network execution efficiency, correct the abnormal weight through the single swarm frog-leap algorithm, effectively avoid the convolutional neural network from falling into the local optimization, optimize the final weight, and effectively improve the classification accuracy. The method can further improve the detection efficiency of fundus image lesion, is applied to pathological detection in an electronic medical record system, and provides effective intelligent medical service for multi-level comprehensive intelligent decision support of human diseases.
Drawings
FIG. 1 is a schematic flow diagram of the present invention;
fig. 2 is a schematic diagram of an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings.
As shown in fig. 1, a convolutional neural network weight optimization method for retinopathy classification includes the following steps:
step 1, inputting a fundus image training set and a label, wherein the training set is X ═ X (X)1，x2，...，xn) N is 1, 2, 3 …, and the label is B (B)1，b2，…，bn) 1, 2, 3 …, label b corresponding to fundus imageiOne-hot coding is carried out to obtain a true value y _ truei；
Step 2, initializing the weight parameters of the convolutional neural network, generating m frogs by adopting standard normal distribution, sequencing according to adaptive values, and finding out the optimal frogs fbWorst and worst frog fwContinuously updating the worst frog position and reordering until the single frog leaping algorithm meets the convergence condition to obtain the global optimal frog fqA 1 is to fqAs the initial weight of the convolutional neural network;
step 3, inputting the jth fundus image in the training set into a convolutional neural network forward propagation calculation model, performing multilayer convolution and pooling, finally performing two-layer full-connection calculation, and outputting to obtain the jth predicted value y _ predictj；
Step 4, the jth true value y _ true is processedjAnd the j-th predicted value y _ predictjPerforming softmax calculation for softmax (y _ predict)j) Performing cross entropy loss calculation to obtain the jth loss value lossj；
Step 5, judging the jth loss value lossjIf the abnormal weight is judged, correcting the abnormal weight by adopting a single swarm leaping algorithm, taking the corrected weight as a new weight of the network, otherwise,continuing to execute the next step;
step 6, judging whether the network reaches the end condition, if so, generating a frog cluster around the initial frog by taking the final weight of the network as the initial frog, and carrying out frog jump optimization to obtain the global optimal frog fqbThe weight parameter is the weight parameter of the final training of the algorithm;
step 7, the trained weight parameters are brought into a network, fundus images are input, two-layer full-connection calculation is finally carried out through multilayer convolution and pooling, and the predicted value y for detecting various pathological changes in retina is outputr。
The specific steps of step 2 are as follows:
step 2.1, determining the number m of frogs, and generating all frogs by adopting standard normal distribution, wherein the standard normal distribution formula is as follows:
Figure BDA0002277252030000051
in formula (1), μ represents a position parameter, σ represents a scale parameter, x represents a random variable, and each frog f comprises an ownership value parameter in the network;
2.2, constructing a convolutional neural network forward propagation calculation model, and randomly selecting a small number w of fundus images from the fundus image training set as reference images;
step 2.3, the frog without loss value calculation is brought into a convolutional neural network forward propagation calculation model to calculate the loss value of the frog, wherein the loss value calculation function is a fitness function of a single group frog leaping algorithm, and the loss calculation formula is as follows:
Figure BDA0002277252030000052
in the formula (2), p represents a network output value, t represents a true value, s represents the dimension of each group of lesion labels, and b represents the number of types of retinopathy needing to be detected simultaneously. Sequencing all frogs in ascending order according to fitness function to obtain the optimal frogs fbWorst and worst frog fw；
And 2.4, updating the position of the worst frog through a position updating function, wherein the position updating function is as follows:
D＝(fb+fp-fw)×Rand(0，1.2) (3)
fnew＝fw+D (4)
Figure BDA0002277252030000053
in the formula (3), fpRepresenting an offset, the dimension of which is the same as the dimension of each frog; in the formula (4), fnewRepresenting the updated frog; in the formula (5), fpiDenotes fpThe value in the ith dimension;
and 2.5, judging whether the single frog leap algorithm meets the convergence condition, if so, stopping the algorithm, and taking the value of the optimal frog as the initial weight of the convolutional neural network, otherwise, turning to the step 2.3.
The specific steps of step 5 are as follows:
step 5.1, judging the loss value loss calculated at the j-1 th timej-1And judging the loss value loss calculated at the j-th timejAbsolute value of the difference of | lossj-1-lossjIf the | is larger than the threshold epsilon, executing the next step, otherwise, not executing the frog leap calculation;
step 5.2, taking the weight value in the j-1 th network as the initial frog wbAround wbThe frog group is generated according to the following formula:
wij＝wbj+0.0001×Rand(-1，1)(1＜i≤c-1) (6)
in the formula (6), wijRepresenting the value of the ith frog generated in the jth dimension, and c representing the total number of the frogs;
step 5.3, constructing a convolutional neural network forward propagation calculation model, and randomly selecting a small number w of fundus images from a training image library as reference images;
step 5.4, bringing the frogs with the loss values not calculated into a convolutional neural network forward propagation calculation model, calculating the loss values of the frogs and sequencing the frogs;
step 5.5, updating the worst frog position, calculating the updated loss value, and reordering the frog group;
step 5.6, judging whether the single swarm frog-leaping algorithm meets the convergence condition, if so, ending the frog-leaping calculation, and obtaining the overall optimal frog fqbAnd (4) taking the corrected new weight value into the network, otherwise, turning to the step 5.4.
The convolutional neural network forward propagation calculation model in the step 3 comprises a convolutional layer, a pooling layer and a full-link layer.
Firstly, acquiring a fundus image training set and a multi-lesion label corresponding to the fundus image training set; searching an optimal initial weight through a single swarm leaping algorithm, then constructing a convolution layer, a pooling layer and a full-link layer in a convolutional neural network, and taking the optimal initial weight as a parameter for the first forward propagation calculation; respectively carrying out cross entropy loss calculation on four predicted values of four pathological changes in retina and a true value, summing to obtain a loss value, judging whether the loss value is abnormal, if so, generating a frog group around a previous forward propagation weight, and searching for an optimal frog updating network weight; otherwise, updating the network weight by adopting a gradient descent algorithm; and finally optimizing the final weight.
As shown in FIG. 2, in the concrete implementation, step 1, label biIs a 1 × 4 array, and obtains a 1 × 8 true value y _ true through one-hot codingi；
Step 2, initializing the weight parameters of the convolutional neural network, generating 15 frogs by adopting standard normal distribution, sequencing according to adaptive values, and finding out the optimal frogs fbWorst and worst frog fwContinuously updating the worst frog position and reordering until the algorithm meets the convergence condition to obtain the global optimum frog fqA 1 is to fqAs the initial weight of the convolutional neural network;
step 3, setting the jth sheet in the training set to be 500, 400, 3]The fundus image of (1) is input into the algorithm model, and the size of the convolution kernel of the first layer is [5, 5, 3, 32 ]]Offset is [32 ]]Obtaining a [1, 500, 400, 32 ] by the first layer convolution]Is pooled to obtain [1, 250, 200, 32 ]]Of a second layer convolution kernel size of [5, 5, 32, 64]Offset is of magnitude [64 ]]Obtaining a [1, 250, 200, 64 ] by the second layer convolution]Then pooling to obtain [1, 125, 100, 64 ]]The parameters of the last two fully-connected layers of the matrix are respectively one [ 125X 100X 64, 8 ]]And one [8, 8 ]]Of offset size [8 ]]Outputting to obtain the j-th predicted value y _ predictjHas a size of [1, 8 ]]；
Step 4, the jth true value y _ true is processedjAnd the j-th predicted value y _ predictjPerforming softmax calculation for softmax (y _ predict)j) Performing cross entropy loss calculation to obtain the jth loss value lossj；
Step 5, judging whether the loss value of the jth time is abnormal, if so, correcting the abnormal weight by adopting a single swarm leaping algorithm, taking the corrected weight as a new weight of the network, and otherwise, continuing to execute the next step;
step 6, judging whether the network reaches the end condition, if so, taking the final weight of the network as an initial frog, generating a frog cluster around the initial frog, wherein the size of the frog cluster is 20, and carrying out frog jump optimization to obtain a global optimal frog fqbNamely, the weight parameter of the algorithm which is finally trained is obtained;
and 7, substituting the trained weight into a network, inputting an eye fundus image, performing multilayer convolution and pooling, finally performing two-layer full-connection calculation, and outputting a predicted value y for detecting various pathological changes in retinarHas a size of [1, 8 ]]。
The invention optimizes the initialization weight, the abnormal weight and the final weight of the convolutional neural network by adopting a single swarm frog-leaping algorithm, can effectively deal with the complexity of fundus images, and improves the accuracy of detecting retinopathy.
The invention can search the optimal initial value of the network through the single swarm frog-leap algorithm, reduce the initial loss value of the convolutional neural network for classifying the multiple lesions of the fundus image, improve the network execution efficiency, correct the abnormal weight through the single swarm frog-leap algorithm, effectively avoid the convolutional neural network from falling into the local optimization, optimize the final weight, and effectively improve the classification accuracy. The method can further improve the detection efficiency of fundus image lesion, is applied to pathological detection in an electronic medical record system, and provides effective intelligent medical service for multi-level comprehensive intelligent decision support of human diseases.
The above-mentioned embodiments, objects, technical solutions and advantages of the present invention are further described in detail, it should be understood that the above-mentioned embodiments are only illustrative of the present invention, and are not intended to limit the scope of the present invention, and any person skilled in the art should understand that equivalent changes and modifications made without departing from the concept and principle of the present invention should fall within the protection scope of the present invention.

Claims (4)

1. A convolutional neural network weight optimization method for retinopathy classification is characterized by comprising the following steps:
step 1, inputting a fundus image training set and a label, wherein the training set is X ═ X (X)1，x2，...，xn) 1, 2, 3, labeled B ═ B1，b2，...，bn) 1, 2, 3., label b corresponding to the fundus imageiOne-hot coding is carried out to obtain a true value y _ truei；
Step 2, initializing the weight parameters of the convolutional neural network, generating m frogs by adopting standard normal distribution, sequencing according to adaptive values, and finding out the optimal frogs fbWorst and worst frog fwContinuously updating the worst frog position and reordering until the single frog leaping algorithm meets the convergence condition to obtain the global optimal frog fqA 1 is to fqAs the initial weight of the convolutional neural network;
step 3, inputting the jth fundus image in the training set into a convolutional neural network forward propagation calculation model, performing multilayer convolution and pooling, finally performing two-layer full-connection calculation, and outputting to obtain the jth predicted value y _ predictj；
Step 4, the jth true value y _ true is processedjAnd j (th) preliminaryMeasured value y _ predictjPerforming softmax calculation for softmax (y _ predict)j) Performing cross entropy loss calculation to obtain the jth loss value lossj；
Step 5, judging the jth loss value lossjIf the abnormal weight is judged, correcting the abnormal weight by adopting a single swarm leaping algorithm, and taking the corrected weight as a new weight of the network, otherwise, continuing to execute the next step;
step 6, judging whether the network reaches the end condition, if so, generating a frog cluster around the initial frog by taking the final weight of the network as the initial frog, and carrying out frog jump optimization to obtain the global optimal frog fqbThe weight parameter is the weight parameter of the final training of the algorithm;
step 7, the trained weight parameters are brought into a network, fundus images are input, two-layer full-connection calculation is finally carried out through multilayer convolution and pooling, and the predicted value y for detecting various pathological changes in retina is outputr。
2. The convolutional neural network weight optimization method for retinopathy classification as claimed in claim 1, wherein the specific steps of the step 2 are as follows:
step 2.1, determining the number m of frogs, and generating all frogs by adopting standard normal distribution, wherein the standard normal distribution formula is as follows:
Figure FDA0002277252020000011
in formula (1), μ represents a position parameter, σ represents a scale parameter, x represents a random variable, and each frog f comprises an ownership value parameter in the network;
2.2, constructing a convolutional neural network forward propagation calculation model, and randomly selecting a small number w of fundus images from the fundus image training set as reference images;
step 2.3, the frog without loss value calculation is brought into a convolutional neural network forward propagation calculation model to calculate the loss value of the frog, wherein the loss value calculation function is a fitness function of a single group frog leaping algorithm, and the loss calculation formula is as follows:
Figure FDA0002277252020000021
in the formula (2), p represents a network output value, t represents a true value, s represents the dimension of each group of lesion labels, and b represents the number of types of retinopathy needing to be detected simultaneously. Sequencing all frogs in ascending order according to fitness function to obtain the optimal frogs fbWorst and worst frog fw；
And 2.4, updating the position of the worst frog through a position updating function, wherein the position updating function is as follows:
D＝(fb+fp-fw)×Rand(0，1.2) (3)
fnew＝fw+D (4)
Figure FDA0002277252020000022
in the formula (3), fpRepresenting an offset, the dimension of which is the same as the dimension of each frog; in the formula (4), fnewRepresenting the updated frog; in the formula (5), fpiDenotes fpThe value in the ith dimension;
and 2.5, judging whether the single frog leap algorithm meets the convergence condition, if so, stopping the algorithm, and taking the value of the optimal frog as the initial weight of the convolutional neural network, otherwise, turning to the step 2.3.
3. The convolutional neural network weight optimization method for retinopathy classification as claimed in claim 1, wherein the specific steps of the step 5 are as follows:
step 5.1, judging the loss value loss calculated at the j-1 th timej-1And judging the loss value loss calculated at the j-th timejAbsolute value of the difference of | lossj-1-lossjIf l is greater than the threshold epsilon, the next step is performed, otherwise,the calculation of the frog leap is not executed;
step 5.2, taking the weight value in the j-1 th network as the initial frog wbAround wbThe frog group is generated according to the following formula:
wij＝wbj+0.0001×Rand(-1，1)(1＜i≤c-1) (6)
in the formula (6), wijRepresenting the value of the ith frog generated in the jth dimension, and c representing the total number of the frogs;
step 5.3, constructing a convolutional neural network forward propagation calculation model, and randomly selecting a small number w of fundus images from a training image library as reference images;
step 5.4, bringing the frogs with the loss values not calculated into a convolutional neural network forward propagation calculation model, calculating the loss values of the frogs and sequencing the frogs;
step 5.5, updating the worst frog position, calculating the updated loss value, and reordering the frog group;
step 5.6, judging whether the single swarm frog-leaping algorithm meets the convergence condition, if so, ending the frog-leaping calculation, and obtaining the overall optimal frog fqbAnd (4) taking the corrected new weight value into the network, otherwise, turning to the step 5.4.
4. The method of claim 1, wherein the convolutional neural network weight optimization for retinopathy classification comprises a convolutional layer, a pooling layer and a full-link layer.
CN201911127264.2A 2019-11-18 2019-11-18 Convolutional neural network weight optimization method for retinopathy classification Active CN110929775B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911127264.2A CN110929775B (en) 2019-11-18 2019-11-18 Convolutional neural network weight optimization method for retinopathy classification

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911127264.2A CN110929775B (en) 2019-11-18 2019-11-18 Convolutional neural network weight optimization method for retinopathy classification

Publications (2)

Publication Number Publication Date
CN110929775A true CN110929775A (en) 2020-03-27
CN110929775B CN110929775B (en) 2022-08-12

Family

ID=69853297

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911127264.2A Active CN110929775B (en) 2019-11-18 2019-11-18 Convolutional neural network weight optimization method for retinopathy classification

Country Status (1)

Country Link
CN (1) CN110929775B (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111612096A (en) * 2020-06-01 2020-09-01 南通大学 Large-scale fundus image classification system training method based on Spark platform
CN112016434A (en) * 2020-08-25 2020-12-01 安徽索贝数码科技有限公司 Lens motion identification method based on attention mechanism 3D residual error network
CN112331313A (en) * 2020-11-25 2021-02-05 电子科技大学 Automatic grading method for sugar net image lesions based on label coding
CN112884144A (en) * 2021-02-01 2021-06-01 上海商汤智能科技有限公司 Network quantization method and device, electronic equipment and storage medium
CN113971367A (en) * 2021-08-27 2022-01-25 天津大学 Automatic design method of convolutional neural network framework based on shuffled frog-leaping algorithm
WO2022016912A1 (en) * 2020-07-20 2022-01-27 南通大学 Full convolutional genetic neural network optimization method for infant brain medical record image segmentation

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107403195A (en) * 2017-07-28 2017-11-28 中南大学 A kind of meteorologic parameter Intelligent Fusion processing method of carrying robot identification floor
CN109886143A (en) * 2019-01-29 2019-06-14 上海鹰瞳医疗科技有限公司 Multi-tag disaggregated model training method and equipment

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107403195A (en) * 2017-07-28 2017-11-28 中南大学 A kind of meteorologic parameter Intelligent Fusion processing method of carrying robot identification floor
CN109886143A (en) * 2019-01-29 2019-06-14 上海鹰瞳医疗科技有限公司 Multi-tag disaggregated model training method and equipment

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
晁萍瑶，王小敏: "基于ＳＦＬＡ优化的ＲＢＦ神经网络入侵检测算法", 《软件导刊》 *

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111612096A (en) * 2020-06-01 2020-09-01 南通大学 Large-scale fundus image classification system training method based on Spark platform
WO2022016912A1 (en) * 2020-07-20 2022-01-27 南通大学 Full convolutional genetic neural network optimization method for infant brain medical record image segmentation
US11593942B2 (en) 2020-07-20 2023-02-28 Nantong University Fully convolutional genetic neural network method for segmentation of infant brain record images
CN112016434A (en) * 2020-08-25 2020-12-01 安徽索贝数码科技有限公司 Lens motion identification method based on attention mechanism 3D residual error network
CN112331313A (en) * 2020-11-25 2021-02-05 电子科技大学 Automatic grading method for sugar net image lesions based on label coding
CN112331313B (en) * 2020-11-25 2022-07-01 电子科技大学 Automatic grading method for sugar net image lesions based on label coding
CN112884144A (en) * 2021-02-01 2021-06-01 上海商汤智能科技有限公司 Network quantization method and device, electronic equipment and storage medium
CN113971367A (en) * 2021-08-27 2022-01-25 天津大学 Automatic design method of convolutional neural network framework based on shuffled frog-leaping algorithm
CN113971367B (en) * 2021-08-27 2024-07-12 天津大学 Automatic convolutional neural network framework design method based on shuffled frog-leaping algorithm

Also Published As

Publication number Publication date
CN110929775B (en) 2022-08-12

Similar Documents

Publication Publication Date Title
CN110929775B (en) Convolutional neural network weight optimization method for retinopathy classification
CN111340819B (en) Image segmentation method, device and storage medium
Wang et al. DenseNet-201-based deep neural network with composite learning factor and precomputation for multiple sclerosis classification
US7905599B2 (en) Methods for diagnosing glaucoma utilizing combinations of FD-OCT measurements from three anatomical regions of the eye
CN112700434B (en) Medical image classification method and classification device thereof
US11488309B2 (en) Robust machine learning for imperfect labeled image segmentation
Singh et al. Collaboration of features optimization techniques for the effective diagnosis of glaucoma in retinal fundus images
Dipu et al. Ocular disease detection using advanced neural network based classification algorithms
CN113096806A (en) Disease prediction system based on medical image algorithm and block chain
Yang et al. Classification of diabetic retinopathy severity based on GCA attention mechanism
Kaya Feature fusion-based ensemble CNN learning optimization for automated detection of pediatric pneumonia
Singh et al. Optimized convolutional neural network for glaucoma detection with improved optic-cup segmentation
Mai et al. Few-shot transfer learning for hereditary retinal diseases recognition
Kim et al. Web applicable computer-aided diagnosis of glaucoma using deep learning
WO2021021657A1 (en) System and method for ai-based eye condition determinations
Sambyal et al. Aggregated residual transformation network for multistage classification in diabetic retinopathy
Jayachandran et al. Multi-dimensional cascades neural network models for the segmentation of retinal vessels in colour fundus images
Yang et al. Multi-classification of fundus diseases based on DSRA-CNN
CN118397280A (en) Endoscopic gastrointestinal tract image segmentation and recognition system and method based on artificial intelligence
CN116934777A (en) Uncertainty-guided retina blood vessel image segmentation method and system
Ishwerlal et al. Lung disease classification using chest X ray image: An optimal ensemble of classification with hybrid training
US20230037183A1 (en) Storage medium, adjustment method, and information processing apparatus
Uyar et al. Abc-based weighted voting deep ensemble learning model for multiple eye disease detection
Zhang et al. PCANet: pyramid context-aware network for retinal vessel segmentation
CN109637633B (en) Method for diagnosing breast cancer state based on big data and machine learning

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