The Weakly supervised classification method of early-stage senile maculopathy based on deep learningTechnical field
The present invention relates to a kind of eye disease classification method more particularly to a kind of early-stage senile based on deep learning are yellowPinta dies down supervised classification method.
Background technique
Treating senile maculopathy is the third-largest blinding disease after cataract, glaucoma, occurs mainly in the age and is greater than55 the elderly group.Treating senile maculopathy clinically can be divided into early stage, mid-term, advanced stage three phases.Early-stage senile is yellowPinta change will not influence patient's vision, thus be not easy by Finding case.And if early-stage senile maculopathy cannot be timelyAnd effective treatment, often deteriorate, and then develop into mid-term or advanced stage, causes the loss of patient's vision.Conventional eyeground figureScreening be find the effective means of early stage maculopathy, but doctor artificial screening and lesions position label can expend it is a large amount ofTime.Therefore it in order to improve screening efficiency, researches and develops automatic early stage maculopathy screening system and has great importance.
Clinically, early-stage senile maculopathy chief complaint is that macular area glass-film wart occurs.It is showed on the figure of eyegroundThere is yellow-white spot for macular area,
According to the illness, the early stage maculopathy classification method of current main-stream needs to acquire and enough contains glass-filmThe eyeground figure of wart is then based on strong measure of supervision design classifier, and whether detection macular area glass-film wart occurs, and according to eyegroundThe position of the information flags glass-film wart such as color of figure.But the strong measure of supervision of mainstream needs largely to have lesions positionThe eyeground figure of precise marking is as training sample.The eyeground figure of these precise markings is difficult to largely obtain in reality, andAnd procurement cost is very high.
There is " the eye fundus image maculopathy dividing method research based on supervision description " in the prior art, by being based onSupervision description son study dividing method by by supervised learning in conjunction with characteristics of the underlying image, by broad sense low-rank matrix approximation sideMethod carries out dimensionality reduction to data, constructs manifold regularization item in conjunction with data label, to extract characteristics of image, and combines SVM pointsClass device can obtain certain segmentation effect.But this method, in extracting characteristic procedure, the iteration optimization for being related to matrix solves, meterCalculation amount is huge, very time-consuming in practical application, and relies on the label information of image pixel, has more lack in practical applicationsIt falls into.
Summary of the invention
Technical problem to be solved by the invention is to provide the depth of a kind of combination convolutional neural networks and attention networkLearn the Weakly supervised classification method of early-stage senile maculopathy, while guaranteeing precision, label training number can be effectively savedAccording to cost.
In order to solve the above technical problems, the technical scheme is that a kind of combination convolutional neural networks and attention netThe Weakly supervised classification method of early-stage senile maculopathy of network, includes the following steps:
Step 1 positions and intercepts macular area, builds convolutional neural networks, using in convolutional neural networks positioning eyeground figureHeart recessed position intercepts square area as candidate regions by side length of twice of disc diameter using central fovea center as origin；
Step 2, judges whether macular area glass-film wart occurs, judges whether macular area occurs using convolutional neural networksGlass-film wart carries out the detection of glass-film wart in the way of Weakly supervised, judges whether occur glass-film wart in the figure of eyeground；
Step 3 remembers lesions position to the eyeground icon for glass-film wart occur, carries out line using the intermediate result of step 2Property interpolation obtains final Pixel-level lesion marking result.
As a kind of perferred technical scheme, in said step 1, the convolutional neural networks include five convolution blocks,First convolution block, second convolution block, third convolution block and the 4th convolution block include two convolutional layers, the 5th volumeBlock includes a convolutional layer；
The convolution kernel size of all convolutional layers is 3*3, first convolution block, second convolution block, thirdThe port number of first convolutional layer of convolution block and the 4th convolution block is followed successively by 32,64,128 and 156, and step-length is 1；FirstA convolution block, second convolution block, third convolution block and the 4th convolution block the port number of second convolutional layer be 1,Step-length is 2；The port number of 5th convolution block is 512, step-length 1.
As a kind of perferred technical scheme, the output end of the convolutional neural networks is returned by Sigmoid functionReturn；For the picture of input after the convolutional neural networks, output channel 1, size reduce 16 times of grayscale image with respect to original image.
As a kind of perferred technical scheme, using the region 16*16 where central fovea centre coordinate in original image as positive sample,The each pixel of grayscale image represents the region 16*16 of corresponding position in original image as the probability of positive sample；Take the point of maximum probabilityCentral fovea centre coordinate of the corresponding original image regional center as prediction.
As a kind of perferred technical scheme, the calculation method of the disc diameter is to ask optic disk in all sample datas straightThe average value of diameter.
As a kind of perferred technical scheme, in said step 1, it in the network training stage, is first sat by central fovea centerMark generates the required true value of training；Original image is divided equally into the grid of several 16*16, each grid corresponds to one picture of true valueVegetarian refreshments, the corresponding true value pixel of grid where central fovea centre coordinate are 1, remaining is 0；After the true value for obtaining image, with predictionAs a result it calculates and intersects entropy loss, and update network weight until network convergence using stochastic gradient descent.
As a kind of perferred technical scheme, in the step 2, judge whether macular area the inspection of glass-film wart occursThe three first layers of survey grid network are two convolutional layers and maximum value pond layer, convolution kernel size are 3*3；The detection network introduces noteMeaning power branch, first carries out maximum value pond and convolution, then carry out linear interpolation up-sampling, finally obtains attention weight；It willAttention weight carries out multiplication weighting to the convolutional layer output of another branch road, then is added to obtain finally with the output of reel laminationOutput；
It is repeated twice the output that the above process obtains the detection network, the detection network output is added one 1The full articulamentum that the convolutional layer and output channel that channel, convolution kernel size are 1*1 are 2；Full articulamentum output is that image isThe no prediction result comprising glass-film wart.
As a kind of perferred technical scheme, in the step 2, loss function is using intersection entropy function, right value updateUsing stochastic gradient descent method.
As a kind of perferred technical scheme, in the step 3, the previous of full articulamentum described in step 2 is takenInput of the convolutional layer as step 3 obtains final lesion mark by traditional linear interpolation method by twice of output amplificationNote output result.
By adopting the above-described technical solution, the early-stage senile of a kind of combination convolutional neural networks and attention network is yellowPinta dies down supervised classification method, includes the following steps: step 1, positions and intercept macular area, build convolutional neural networks, adoptEyeground figure center recessed position is positioned with convolutional neural networks, using central fovea center as origin, is cut by side length of twice of disc diameterTake square area as candidate regions；Step 2, judges whether macular area glass-film wart occurs, is judged using convolutional neural networksWhether macular area there is glass-film wart, and the detection of glass-film wart is carried out in the way of Weakly supervised, judges whether occur in the figure of eyegroundGlass-film wart；Step 3 remembers lesions position to the eyeground icon for glass-film wart occur, carries out line using the intermediate result of step 2Property interpolation obtains final Pixel-level lesion marking result；This algorithm carries out classifier training and detection using Weakly supervised method,It only needs to provide whether eyeground figure glass-film wart information occurs, classifier can be trained without more specific location information, realized to morningThe correct classification of phase treating senile maculopathy eyeground figure, the algorithm can be effectively saved label training while guaranteeing precisionThe cost of data.
Detailed description of the invention
Fig. 1 is the positioning convolutional neural networks structure chart of the embodiment of the present invention；
Fig. 2 is that the embodiment of the present invention judges whether macular area glass-film wart network structure occurs.
Specific embodiment
With reference to the accompanying drawings and examples, the present invention is further explained.In the following detailed description, only pass through explanationMode describes certain exemplary embodiments of the invention.Undoubtedly, those skilled in the art will recognize,In the case where without departing from the spirit and scope of the present invention, described embodiment can be repaired with a variety of different modesJust.Therefore, attached drawing and description are regarded as illustrative in nature, and are not intended to limit the scope of the claims.
A kind of Weakly supervised classification method of early-stage senile maculopathy of combination convolutional neural networks and attention network, packetInclude following step:
Step 1 positions and intercepts macular area, builds convolutional neural networks, using in convolutional neural networks positioning eyeground figureHeart recessed position intercepts square area as candidate regions by side length of twice of disc diameter using central fovea center as origin；
As shown in Figure 1, the convolutional neural networks include five convolution blocks, first convolution block, second convolution block, theThree convolution blocks and the 4th convolution block include two convolutional layers, and the 5th convolution block includes a convolutional layer；All institutesThe convolution kernel size for stating convolutional layer is 3*3, first convolution block, second convolution block, third convolution block and the 4th volumeThe port number of first convolutional layer of block is followed successively by 32,64,128 and 156, and step-length is 1；First convolution block, secondThe port number of second convolutional layer of convolution block, third convolution block and the 4th convolution block is 1, and step-length is 2；5thThe port number of convolution block is 512, step-length 1.
The output end of the convolutional neural networks is returned by Sigmoid function；The picture of input passes through the volumeAfter product neural network, output channel 1, size reduce 16 times of grayscale image with respect to original image.With central fovea centre coordinate in original imageThe region 16*16 at place is positive sample, and the region 16*16 that each pixel of grayscale image represents corresponding position in original image is positive sampleThis probability；Take the corresponding original image regional center of the point of maximum probability as the central fovea centre coordinate of prediction.The optic disk is straightThe calculation method of diameter is to seek the average value of disc diameter in all sample datas.
In the network training stage, the required true value of training is first generated by central fovea centre coordinate；Original image is divided equally intoThe grid of several 16*16, each grid correspond to one pixel of true value, the corresponding true value of grid where central fovea centre coordinatePixel is 1, remaining is 0；After the true value for obtaining image, intersect entropy loss with prediction result calculating, and use stochastic gradient descentNetwork weight is updated until network convergence.
Step 2, judges whether macular area glass-film wart occurs, judges whether macular area occurs using convolutional neural networksGlass-film wart carries out the detection of glass-film wart in the way of Weakly supervised, judges whether occur glass-film wart in the figure of eyeground；
As shown in Fig. 2, judge macular area whether occur glass-film wart detection network three first layers be two convolutional layers withMaximum value pond layer, convolution kernel size are 3*3；The detection network introduces attention branch, first carry out maximum value pond withAnd convolution, then linear interpolation up-sampling is carried out, finally obtain attention weight；By attention weight to the volume of another branch roadLamination output carries out multiplication weighting, then is added to obtain final output with the output of reel lamination；
It is repeated twice the output that the above process obtains the detection network, the detection network output is added one 1The full articulamentum that the convolutional layer and output channel that channel, convolution kernel size are 1*1 are 2；Full articulamentum output is that image isThe no prediction result comprising glass-film wart.Loss function uses stochastic gradient descent method using entropy function, right value update is intersected.
Step 3 remembers lesions position to the eyeground icon for glass-film wart occur, carries out line using the intermediate result of step 2Property interpolation obtains final Pixel-level lesion marking result.
In the step 3, input of the previous convolutional layer of full articulamentum described in step 2 as step 3 is taken,By traditional linear interpolation method by twice of output amplification, final lesion marking output result is obtained.The process is no prisonProcess is superintended and directed, obtained result is the label of Pixel-level, i.e. every bit represents whether the pixel is glass-film wart.
It is compared with existing method, the difference of this method is:
1, the detection of traditional glass-film wart and lesion marking algorithm are strong measure of supervision, are needed to acquire largely with Pixel-levelThe sample data of mark.And the training process of the algorithm uses Weakly supervised method, it is only necessary to acquire the markup information of image level, and nothingNeed the mark of Pixel-level.The cost of mark sample data can substantially be saved.
2, similar Weakly supervised learning method mostly uses greatly traditional more case-based learnings, lesion of this method to obscurity boundaryIt is difficult to be effectively detected.And this method innovatively joined attention branch of a network in the detection network of glass-film wart and be used forThe lesions position of eyeground figure is extracted, attention network is superimposed with the convolutional layer again after weighting to convolutional layer, and which is to module of boundaryThe glass-film verrucosis stove of paste can also be effectively detected.
3. describing the dividing method of son study based on supervision, although there is stronger ability in feature extraction, algorithm is related to greatlyThe operation of moment matrix Optimized Iterative, calculation amount is huge, and relies on image pixel tag information, it is difficult to solve that at high cost, time-consumingProblem；In comparison, the deep learning of the application passes through figure in such a way that convolutional neural networks and attention network combineAs grade markup information, label time and cost can be substantially saved, realizes the Weakly supervised classification for the treatment of senile maculopathy.
As a kind of specific embodiment, residual error network is can be used as main frame in the detection network of glass-film wart, then plusEnter attention network and extracts lesions position.Constrained multi-instance learning method can be used in the detection segmentation network of glass-film wart.
The above shows and describes the basic principle, main features and advantages of the invention.The technology of the industryPersonnel are it should be appreciated that the present invention is not limited to the above embodiments, and the above embodiments and description only describe thisThe principle of invention, without departing from the spirit and scope of the present invention, various changes and improvements may be made to the invention, these changesChange and improvement all fall within the protetion scope of the claimed invention.The claimed scope of the invention by appended claims and itsEquivalent thereof.