CN111815563B

CN111815563B - Retina optic disc segmentation method combining U-Net and region growing PCNN

Info

Publication number: CN111815563B
Application number: CN202010524252.XA
Authority: CN
Inventors: 徐光柱; 陈莎; 林文杰; 雷帮军; 石勇涛; 周军; 刘蓉; 王阳
Original assignee: Chongqing Bio Newvision Medical Equipment Ltd
Current assignee: Chongqing Bio Newvision Medical Equipment Ltd
Priority date: 2020-06-10
Filing date: 2020-06-10
Publication date: 2024-04-09
Anticipated expiration: 2040-06-10
Also published as: CN111815563A

Abstract

A retina optic disc segmentation method combining U-Net and region growing PCNN carries out graying treatment on retina optic disc dataset pictures; performing CLAHE processing on the data set picture subjected to the graying processing, and enhancing the contrast between the optic disc and the background in the retina optic disc image; blocking the retinal optic disc image; building, training and rough picture extraction of a U-Net neural network model; building a region growing PCNN neural network model; retinal optic disc segmentation was performed using region growing PCNN. The invention provides a rough extraction method of a U-Net retina video disc image based on improvement, which remarkably suppresses the background, highlights the video disc area and increases the picture contrast by the rough extraction, thereby improving the picture quality of a data set; on the other hand, an improved region growing PCNN-based video disc image segmentation method is provided, and the segmentation performance of the PCNN is improved by changing a seed selection mode, a PCNN initial ignition threshold selection mode and a region growing ending condition, so that complete video disc segmentation is realized.

Description

Retina optic disc segmentation method combining U-Net and region growing PCNN

Technical Field

The invention relates to the technical field of video disc identification, in particular to a retina video disc segmentation method combining U-Net and region growing PCNN.

Background

Glaucoma is extremely dangerous and is the primary cause of irreversible blindness, and blindness is very easy to occur in the early stage without intervention or treatment. The changes in the size, shape and depth of the optic disc are closely related to the occurrence, absence and degree of symptoms of glaucoma. Therefore, the identification of the optic disc plays a vital role in the diagnosis and treatment of glaucoma, and has great research value. However, a doctor needs to read a large number of fundus images to make a diagnosis, which is a time-consuming and tedious process and is easily affected by subjective experience and fatigue. The risk of missing certain detailed information in fundus images when a doctor is tired increases, so that missed diagnosis and misdiagnosis occur, and judgment is not suitable for large-scale glaucoma screening by the doctor. In summary, the task of disc segmentation in current retinal images presents a number of difficulties:

1) The contrast between optic disc and background in fundus image is low. Due to the influence of the acquisition equipment and the acquisition environment such as uneven illumination and the like and the influence of physiological changes of people on the video disc, the conditions of low contrast and poor picture quality are caused;

2) The video disc images obtained by most of the current segmentation methods have misjudgment and insufficient detail information, so that the segmentation effect is poor.

Therefore, it is important to design a precise segmentation scheme for the retinal optic disc.

A common method in conventional retinal optic disc segmentation methods is image segmentation based on edge detection. Document [1] Chinese patent 'video disc segmentation method integrating fundus image edge information and brightness information' (application number: 2016091015463. X) utilizes an Isotopic Sobel operator to extract vertical edge and horizontal edge images of fundus images after contrast enhancement, positions video disc horizontal coordinates according to an edge difference curve, then extracts blood vessel density characteristics and video disc brightness characteristics, multiplies the blood vessel density characteristics and the video disc brightness characteristics to obtain characteristic curves related to the video disc vertical coordinates, the coordinate corresponding to the maximum value on the curve is the video disc vertical coordinates, then locates video disc center coordinates, and utilizes a CV level set model to segment video discs.

Document [2] Chinese patent "a method and apparatus for positioning a video disc" (application number: 201710405139.8) improves algorithms such as fuzzy convergence, edge detection, bright area detection, template matching, etc., each algorithm generates one or more outputs, fuses all the outputs to position the video disc, then uses the video disc positioning and edge detection results as initial parameters for video disc segmentation, determines the rough contour of the video disc by ellipse fitting, fits the deformable contour to the edge of the video disc, and finally completes video disc segmentation. However, the traditional fundus image segmentation algorithm has an unsatisfactory segmentation effect in a complex scene, and cannot obtain higher segmentation precision for fundus images with lower contrast and poorer image quality so as to realize accurate judgment by a computer-aided system.

The deep learning technology is rapidly developed, and the method has great advantages in the field of medical image processing. Numerous researchers have applied deep learning techniques to the task of optic disc segmentation in fundus images. Document [3] Chinese patent 'MCASPP neural network fundus image cup and optic disc segmentation model based on Attention mechanism' (application number: 201910711320.0) extracts a first image feature of an input image through a feature extraction module, an Attention mapping module extracts a second image feature, a first feature is obtained according to a high-level feature, a low-level feature and the second image feature in the first image feature, a multi-scale cavity convolution module is used for carrying out convolution on the high-level feature for multiple times to obtain a second feature, and an output module is used for obtaining a optic disc segmentation result according to the first feature and the second feature.

Document [4] Chinese patent 'a method, a device and a storage medium for determining a video cup and video disc segmentation model' (application number: 201910964739.7) inputs images into a full convolution neural network to position video discs, introduces a space constraint module and a combined loss semantic segmentation algorithm to build a video disc and video disc video segmentation network, and obtains video disc ROI segmentation results. The method loses a plurality of useful information by extracting the characteristics layer by layer, and the fundus image data set has large picture quality difference and insufficient quantity, so that the finally learned parameters of the model can not completely characterize the video disc, and the requirements of medical staff can not be met.

In the field of deep learning, there are numerous network models, wherein the neural network described in the document [5]Olaf Ronneberger,Philipp Fischer,Thomas Brox.U-Net: convolutional Networks for Biomedical Image segment. MICCAI,2015:234-241 ] is friendly to the medical image processing of rare data sets, and exhibits great advantages.

Document [6] Chinese patent 'method for segmenting retinal vessel image by combining multi-scale characteristic convolutional neural network' (application number: 201810635753.8) utilizes U-Net network to segment fundus image map, the obtained predictive segmented image is transmitted to a discriminator network generating an countermeasure network to identify true or false, the obtained loss is returned to the updating weight of U-Net network model, and finally the optic disc segmentation model is obtained.

Document [7] Chinese patent "a video disc segmentation method based on focus color retina fundus image" (application number: 201811142896.1) establishes two U-Net models to perform blood vessel detection and video disc detection respectively, and obtains a blood vessel probability map and a video disc probability map. Firstly, acquiring a main vessel fitting straight-line graph from a vessel probability graph, acquiring a probability graph of the main vessel fitting straight-line graph, selecting a video disc region from a video disc communication region, and estimating the center and the radius of a video disc, thereby dividing the video disc. In the methods, the optic disc is segmented by using a hard threshold value, so that low-gray-level pixels at the edge of the optic disc are mistakenly segmented into the background, high-gray-level pixels are mistakenly segmented into the optic disc, segmentation accuracy is reduced, judgment of medical staff is interfered, the U-Net is the result of segmentation by using the hard threshold value, and red is mistakenly segmented into the background of the optic disc.

Still other researchers have improved U-Net models. The U-Net encoder section described in document 8]Shuang Yu,Ygesan Kanagasingam,et al.Robust optic disc and cup segmentation with deep learning for glauxoma detection,Computerized Medical Image and Graphics,2019,74:61-71 is modified to have been pre-trained to the ResNet34 network, the decoder section is unmodified, and the subsequent results are morphologically processed. As ResNet34 parameters are frozen during U-Net model training and cannot be updated, performance indexes such as segmentation result accuracy and the like are reduced.

Document [9]Artem Seveastopolsky,Optic Disc and Cup Segmentation Methods for Glaucoma Dection with Modification of U-Net Convolution Neural Network, pattern Recognition and Image Analysis,2017,27 (3): 618-624 describes reducing the number of convolution kernels in the U-Net convolution layers, all of which are 64, thus increasing the training speed, but extracting feature images from each layer is too few, resulting in low final accuracy and insignificant improvement. And these improved methods have finally not escaped the problem of hard threshold segmentation of the segmentation probability map.

At present, a plurality of self-adaptive threshold algorithms can be used for image binarization segmentation, and the common self-adaptive algorithm is an OTSU algorithm as described in document [10]Nobuyuki Otsu,A threshold selection method from gray level histograms,IEEE Trans.syst.man Cybern,1979,9 (1): 62-66; the maximum entropy threshold segmentation method is described in, for example, documents [11]Abutaleb A S,Automatic Thresholding of Gray-Level Picture Using Two-Dimensional Entropies, pattern Recognition,1989,47 (1): 22-32 ].

The OTSU algorithm uses a threshold to divide the image into two parts, the target and background, and then maximizes the variance between the two classes of pixels, where the threshold is the segmentation threshold. The maximum entropy threshold segmentation method also divides the image into a target part and a background part, calculates the information entropy of the two parts, and maximizes the information entropy to obtain a threshold. Although the method is fast, the thresholds are global thresholds, and when the proportion of the target to the background is great, the inter-class variance function and the maximum entropy function can show multiple peaks, so that the obtained threshold segmentation effect is poor. While PCNN has multiplicative coupling that allows the image display to accept local neighborhood information and dynamic pulse threshold characteristics that allow each neuron of the PCNN to have a different threshold and to change as iterations proceed.

Disclosure of Invention

On the one hand, aiming at the conditions that the quality of the data set pictures is uneven and the contrast is low, and the follow-up processing requirements are difficult to meet, the U-Net model is improved, the invention provides an improved U-Net retina video disc image rough extraction method, by the rough extraction, the background can be obviously restrained, the video disc area is highlighted, the noise interference is weakened, the picture contrast is increased, and the quality of the data set pictures is improved. U-Net is a deep convolutional neural network, so that the U-Net has strong generalization capability.

On the other hand, the invention provides a video disc image segmentation method based on the improved region growing PCNN aiming at the situation that the information of the video disc edge region is easy to lose, defect and misjudge after the U-Net output is segmented by a hard threshold, and the segmentation performance of the PCNN is improved by changing a seed selection mode, a PCNN initial ignition threshold selection mode and a region growing ending condition, so that the segmentation of the complete video disc is realized.

The technical scheme adopted by the invention is as follows:

a retinal optic disc segmentation method combining U-Net with region growing PCNN, comprising the steps of:

step 1: graying the retina optic disc data set picture;

step 2: performing CLAHE processing on the data set picture subjected to the gray processing in the step 1, enhancing the contrast between the optic disc and the background in the retina optic disc image,

step 3: blocking the retinal optic disc image;

step 4: building, training and rough picture extraction of a U-Net neural network model;

step 5: building a region growing PCNN neural network model;

step 6: retinal optic disc segmentation was performed using region growing PCNN.

The step 3 comprises the following steps:

step 3.1: the partitioning of the training set pictures is to randomly select 48X 48 pictures from the pictures;

step 3.2: the test set picture blocking is to randomly select 48X 48 image blocks in the image;

step 3.3: judging that the central pixel of the block is not in the fundus image area, if yes, reserving and selecting the next block; if not, discarding and re-selecting are carried out, so that all pixels of the image block are ensured to be in the fundus image area.

The invention relates to a retina video disc segmentation method combining U-Net and region growing PCNN, which has the technical effects that:

1: aiming at the situations of low contrast, strong noise interference and poor quality of the current data set picture, the invention provides a video disc image rough extraction method.

2: aiming at the problem that the prior algorithm has insufficient segmentation precision, the hard threshold is used for segmenting the video disc for U-Net, so that the pixels at the edge of the video disc are easy to be classified by mistake, and the improved region growing PCNN is used for segmenting the video disc of the rough extraction image. The seed point, the PCNN initial ignition threshold selection mode and the region growing ending condition are changed, so that the segmentation effect is better than that of the region growing PCNN, and the misclassified pixels are fewer.

Drawings

FIG. 1 is a diagram of U-Net hard threshold misclassified pixels.

Fig. 2 is a general flow chart of the present invention.

FIG. 3 (a) is a gray scale diagram before CLAHE processing;

fig. 3 (b) is a grayscale image after the CLAHE process.

FIG. 4 (a) is an original view of a test set image;

fig. 4 (b) is a block diagram of a test set image.

Fig. 5 is a diagram of a U-net network structure.

Fig. 6 is a block diagram of an encoder.

Fig. 7 is a structural diagram of a decoder.

Fig. 8 is a rough extraction disc map.

Fig. 9 is a PCNN structure diagram.

FIG. 10 (a) is a plot of raw region growing PCNN under-growth labels;

fig. 10 (b) is a diagram of the original region growing PCNN segmentation result.

Fig. 11 is a modified PCNN flowchart.

FIG. 12 (a) is a graph comparing the segmentation results of U-Net and PCNN (U-Net+hard threshold results);

FIG. 12 (b) is a graph showing the comparison of the segmentation results of U-Net and PCNN (U-Net+PCNN segmentation results).

FIG. 13 is a graph of quantitative comparison of U-Net+hard threshold versus U-Net+PCNN partitioning algorithm performance across a data set.

Detailed Description

step 1: the retinal optic disc dataset pictures are subjected to graying processing, and all the pictures are extracted in proportion into red, green and blue three-channel images x=0.299r+0.587g+0.114b, namely graying processing, as shown in fig. 3 (a).

Step 2: and (3) performing CLAHE processing on the data set picture subjected to the gray processing in the step (1) to enhance the contrast between the video disc and the background in the retina video disc image, as shown in fig. 3 (b).

Step 3: blocking the retinal optic disc image;

step 5: building a region growing PCNN neural network model;

The step 3 specifically comprises the following steps:

step 3.2: the test set picture segmentation is to randomly select 48×48 image blocks in an image, and fig. 4 (a) is an original image of the test set image;

step 3.3: judging that the central pixel of the block is not in the region frame of fig. 4 (b), if yes, reserving and selecting the next block; if not, discarding and re-selecting are performed, so that all pixels of the image block are ensured to be in the fundus image area, and the image block is divided as shown in fig. 4 (b).

The step 4 specifically comprises the following steps:

step 4.1: aiming at the problem of uneven picture quality of a data set, the invention builds the U-Net neural network model for coarse extraction, thereby improving the picture quality. The U-Net neural network model is a full convolution neural network with a U-shaped symmetrical structure, and FIG. 5 is a schematic diagram of the U-Net neural network model. The left half of fig. 5 is the encoder and the right half is the decoder. Each encoder consists of a first convolution layer, a first batch normalization layer, a first activation layer, a first regularization layer, and a maximum pooling layer, as shown in fig. 6, that enable path contraction of the input image during training of the network to capture global information. The output image is scaled down to 1/4 of the input image per pass through one encoder. The activation function is the leak-ReLU function commonly used for deep learning at present. The function is evolved based on a ReLU function, and the leak-ReLU function imparts a small non-zero slope to all negative values compared to the ReLU function, so that the negative axis information is not completely lost.

The decoder consists of an upsampling layer, a second convolution layer, a second batch normalization layer, a second activation layer, a second regularization layer, as shown in fig. 7. The output image is enlarged to 4 times of the input image every time the up-sampling unit passes. And then the two-dimensional convolution of the convolution layer is carried out, and the batch standardization layer stretches the convolution layer and outputs the stretched convolution layer to normal distribution. The leakage-ReLU function performs nonlinear mapping, and the regularization layer discards the neuron activation value with a certain probability. The output layer activation function is a softmax function.

TABLE 1 hidden layer parameters

Encoder with a plurality of sensors	Feature map size	Decoder	Feature map size	Convolution kernel size
					Layer_1	48×48	Layer_1	6×6	3×3
Layer_2	24×24	Layer_2	12×12	3×3
					Layer_3	12×12	Layer_3	24×24	3×3
Layer_4	6×6	Layer_4	48×48	3×3

Compared with the original U-Net model, the invention has some improvements:

firstly, the encoder and the decoder of the invention have 4 lateral convolutions respectively, 3 lateral convolutions are adopted, and 2 lateral convolutions are adopted by a common U-Net model, so that the depth of the U-Net network is increased, and more detail information of video discs can be extracted and stored by 3 lateral convolutions, the video discs are highlighted and noise is suppressed.

Secondly, the input of the activation function is forced to be pulled to the standard normal distribution by using the batch standardization layer, so that the input value of the nonlinear transformation function falls into a region sensitive to input, the gradient dispersion problem frequently occurring in deep learning is avoided, and the network convergence speed can be accelerated.

Third, the present invention uses a regularization layer. The method has the function of discarding the activation value of the convolutional layer neuron with a certain probability, reducing the function among nodes of each layer and preventing the phenomenon of model overfitting. The batch normalization layer and the regularization layer are used together, so that the model training speed is increased, and the model generalization capability and robustness are enhanced.

Step 4.2: during the model training phase, during each round of training, we selected 90% of the data in the training set for training, and the remaining 10% of the data was used for verification. The error is calculated in the model using a cross entropy cost function, then the cost is minimized using a random gradient descent approach, and then the update weights and biases are back-propagated.

Step 4.3: the preliminarily processed test set pictures are input into a trained improved U-Net model for rough extraction of the video disc, and fig. 8 is a rough extraction process of the video disc.

The step 5 specifically comprises the following steps:

the invention constructs the region growing PCNN aiming at the binary segmentation of the gray level image. PCNN is proposed by Eckhorn according to the mammalian visual cortex model, and belongs to the third generation of neural networks. The structure of PCNN is shown in fig. 9.

The PCNN model consists of an accept domain, a modulate domain, a pulse generation domain. The receiving domain consists of a connection input L and a feedback input F; the modulation domain mainly generates an internal activity item U; the pulse generation domain is composed of a threshold regulator and a pulse generator. When the internal activity term U is greater than the dynamic threshold θ, the neuron fires, y=1. The mathematical expression for the region growing PCNN is:

L[n]＝∑ _kl W _ijkl Y _kl [n-1]-d (1)

U[n]＝F[n]{1+β _n L[n]} (2)

wherein L is connection input, U is internal activity item, F is feedback input, Y is pulse output, θ is dynamic threshold, W is connection weight, β is connection strength coefficient, d is inhibition item, P is recording neuron ignition matrix, W and Ω are dynamic threshold values under different conditions, and n is iteration number.

The value of the connection coefficient beta in the region growing PCNN is varied. This captures more neurons and the upper limit of β is T, which acts to prevent excessive disc area growth by limiting the upper limit of β. Region growing PCNN also introduces a fast-connect mechanism so that neurons of the same connected domain fire simultaneously. The region growing PCNN firstly selects a reliable video disc region as a seed through a seed point threshold value, and then realizes the automatic growth of the video disc region through a connection strength coefficient and a stop condition in the PCNN.

Compared with the image segmentation method by using the region growing PCNN, the invention has the following improvement:

first, the seed point selection mode is improved. The improved PCNN selects from the input image that is greater than the seed point threshold u ₀ The pixel with the largest connected area is selected as the seed point. Whereas the original region growing PCNN is selected from the input image to be larger than the seed point threshold value u ₀ As seed pixels, and the minimum value in the seed pixels is set as the initial firing threshold θ of PCNN. Since in the picture after U-Net rough extraction, there is still some noise interference in the fundus image edge region, affecting the accuracy of the PCNN segmentation result, but these regions are much smaller in area than the optic disc region and are not connected, the pixel with the largest connected region is now used as a seed point.

Second, the stop condition for region growing is changed. Fig. 10 (a) is a map of the original region growing PCNN growth deficiency label. The original region growing PCNN algorithm is a case where, when region growing is abnormal while iterative region growing is continuously performed, an outer contour irregularity occurs for a video disc. That is, the resulting optic disc area is not elliptical in shape, having a depression in its outer contour, as shown in fig. 10 (b). This is mainly due to the excessively dark portions of the target area. In this case, the stop condition for PCNN area growth needs to be adjusted accordingly. Aiming at the problem, the invention obtains an empirical threshold Z of the edge area ratio by counting the proportion of the area of the optic disc area to the edge pixel point in the training set in the fundus image. When the PCNN is used for segmentation and the original region growing and stopping condition is reached, if the edge area ratio of the obtained result is larger than the threshold value, the fact that the video disc segmented region is incomplete can be judged, and the stopping condition should be revised again to enable the video disc segmented region to continue growing.

Thirdly, changing the PCNN neuron firing threshold selection mode. Calculating the area of the seed point and the threshold delta _θ In contrast, if the PCNN ignition initiation threshold is greater than this threshold, the PCNN ignition initiation threshold is set to θ=max (F) -0.05, and if it is small, θ=max (F) -0.5. Because the communication area is large and the PCNN initial ignition threshold value is small, the area overgrowth is caused, the background is mistaken for the video disc, the communication area is small and the PCNN initial ignition threshold value is too large, the area growth is insufficient, the video disc is mistakenly classified as the background, and the video disc area is reduced. While the original PCNN initial firing threshold is selected by selecting either the minimum seed point or the maximum non-firing neuron as the threshold, the connected region typically includes a optic disc region, which is likely to cause excessive region growth.

The step 6 specifically comprises the following steps:

step 6.1: setting initial beta of the connection strength coefficient, and increasing step delta of the connection strength coefficient _β Initial value beta of maximum value of connection strength coefficient _max ，β _max For selecting the threshold value u of the seed point ₀ Threshold z of disc image to determine whether region growth is sufficient to limit disc edge overgrowthThreshold T, threshold delta of (a) _θ Selecting a PCNN initial ignition threshold;

step 6.2: defining the output term of PCNN as y=0 and defining a matrix Y of the same order as Y ₁ ＝1；

Step 6.3: the extracted gray image S _y Normalization process as input of external excitation of PCNN, i.e. f=s _y ；

Step 6.4: selecting from the input image a gray value greater than the seed point threshold value u ₀ The pixel with the largest connected area is selected as the seed point,

step 6.5: calculating the area of the maximum communication area and determining the threshold delta _θ In comparison with the prior art. If the PCNN neuron initial firing threshold is greater than the threshold, the PCNN neuron initial firing threshold is set to θ=max (F) -0.05, whereas θ=max (F) -0.5;

step 6.6: calculating the ratio x of the disk edge pixels to the disk area in the Y, comparing x with Z, and executing the next step if the ratio is large; otherwise, ending the program and outputting Y;

step 6.7: comparison of beta _max And T, if beta _max >T, ending the program and outputting Y; otherwise, executing the next step;

step 6.8: comparison of beta and beta _max If beta is less than or equal to beta _max Then the next step is carried out; otherwise, executing the step 6.12;

step 6.9: comparing Y and gamma ₁ If the two types are different, executing the next step; if the two are the same, executing the step 6.11;

step 6.10: assigning the value of Y to Y ₁ And performing a quick connection, returning to step 6.9;

step 6.11: beta increase delta _β ，Y ₁ Resetting to 1, and returning to the step 6.8;

step 6.12: beta _max Delta beta is increased, and the process returns to the step 6.6.

The invention provides a retina video disc segmentation method combining U-Net and region growing PCNN, and the structure of the retina video disc segmentation method is shown in fig. 5 and 9.

1): the method uses the improved U-Net model as a preprocessing neural network model to extract the optic disc region, and the structure of the optic disc region is shown in figure 5. The U-Net model is improved and is used for a preprocessing link of the video disc extraction process, so that the video disc area is effectively enhanced, and the background area is restrained, thereby being convenient for providing stable and reliable input for the subsequent segmentation link.

2): according to the method, the U-Net is combined with the improved PCNN to perform the retina video disc segmentation task, and compared with the method for segmenting the video disc area by using the U-Net model only, the accuracy is higher, and the video disc information is more complete. And by adding seed region initial range selection conditions and edge area ratio limiting conditions, effective control of PCNN region growing process is realized, so that optimal video disc region segmentation output is obtained.

3) FIG. 13 is a graph of quantitative comparison of U-Net+hard threshold versus U-Net+PCNN partitioning algorithm performance across a data set. Compared with the segmentation result of the U-Net in FIG. 12 (a) by using a hard threshold, the segmentation algorithm of the U-Net+PCNN in FIG. 12 (b) has the defects of sensitivity, specificity, accuracy and accuracy, and particularly has obvious increase in accuracy, so that the improvement of combining the U-Net and the region-growing PCNN reduces false positive, improves segmentation accuracy and has good application prospect.

Claims

1. A retinal optic disc segmentation method combining U-Net with region growing PCNN, characterized by comprising the steps of:

step 1: graying the retina optic disc data set picture;

step 3: blocking the retinal optic disc image;

step 5: building a region growing PCNN neural network model;

the step 5 specifically comprises the following steps:

for binary segmentation of gray images, a region growing PCNN model is constructed, wherein the PCNN model consists of a receiving domain, a modulating domain and a pulse generating domain, and the receiving domain consists of a connecting input L and a feedback input F; the modulation domain mainly generates an internal activity item U; the pulse generation domain consists of a threshold regulator and a pulse generator, when the internal activity item U is larger than the dynamic threshold theta, the neuron fires, Y=1, and the mathematical expression of the region growing PCNN model is as follows:

L[n]＝∑ _kl W _ijkl Y _kl [n-1]-d(1)

U[n]＝F[n]{1+β _n L[n]}(2)

the value of a connection coefficient beta in the region growing PCNN model is changed, so that more neurons are captured, the upper limit of beta is T, the function of the region growing PCNN model is to prevent the excessive growth of the optic disc region by limiting the upper limit of beta, a quick connection mechanism is also introduced into the region growing PCNN model, so that the neurons of the same connection region are excited simultaneously, the region growing PCNN model firstly selects a reliable optic disc region as a seed through a seed point threshold value, and then the automatic growth of the optic disc region is realized through the connection strength coefficient and a stop condition in the PCNN;

step 6: performing retinal optic disc segmentation using region growing PCNN;

the step 6 comprises the following steps:

step 6.1: setting initial beta of the connection strength coefficient, and increasing step delta of the connection strength coefficient _β Initial value beta of maximum value of connection strength coefficient _max ，β _max Is added by a step-size delta beta,threshold value u for selecting seed points ₀ Threshold z of the video disc image to determine whether the region growth is sufficient, threshold T to limit excessive growth of the video disc edge, threshold delta _θ Selecting a PCNN initial ignition threshold; step 6.2: defining the output term of PCNN as y=0 and defining a matrix Y of the same order as Y ₁ ＝1；

Step 6.3: the extracted gray image S _y Normalization process as input of external excitation of PCNN, i.e. f=s _y The method comprises the steps of carrying out a first treatment on the surface of the Step 6.4: selecting a gray value greater than a seed point threshold value u from the input image ₀ The pixel with the largest connected area is selected as the seed point,

step 6.5: calculating the area of the maximum communication area and determining the threshold delta _θ Comparing; if the PCNN neuron initial firing threshold is greater than the threshold, the PCNN neuron initial firing threshold is set to θ=max (F) -0.05, whereas θ=max (F) -0.5;

step 6.7: comparison of beta _m)x And T, if beta _max >T, ending the program and outputting Y; otherwise, executing the next step;

step 6.9: comparing Y and Y ₁ If the two types are different, executing the next step; if the two are the same, executing the step 6.11;

2. The method for retinal optic disc segmentation in combination with region growing PCNN according to claim 1, wherein: the step 3 comprises the following steps:

3. The method for retinal optic disc segmentation in combination with region growing PCNN according to claim 1, wherein: the step 4 comprises the following steps:

step 4.1:

aiming at the problem of uneven picture quality of a data set, a U-Net neural network model is built for coarse extraction, the picture quality is improved, the U-Net neural network model is a full convolution neural network with a U-shaped symmetrical structure, and the full convolution neural network comprises 4 encoders and decoders, wherein the number of the encoders and the decoders is 3 times of transverse convolution;

each encoder consists of a first convolution layer, a first batch of standardization layers, a first activation layer, a first regularization layer and a maximum pooling layer, the input images can be subjected to path shrinkage in the training of the network so as to capture global information, and each encoder is used for outputting the images to be reduced to 1/4 of the input images, and the adopted activation function is a Leaky-ReLU function; the decoder consists of an up-sampling layer, a second convolution layer, a second batch of standardization layers, a second activation layer and a second regularization layer, wherein the output image is expanded to 4 times of the input image after passing through the up-sampling layer once, then the output image is subjected to two-dimensional convolution by the second convolution layer, the second batch of standardization layers stretch the second convolution layer to be output to normal distribution, the leakage-ReLU function carries out nonlinear mapping, the second regularization layer discards neuron activation values according to set probability, and the output layer activation function is a softmax function;

step 4.2: in the model training stage, 90% of data in a training set are selected for training during each round of training, the rest 10% of data are used for verification, a cross entropy cost function is used for calculating errors in a U-Net neural network model, then a random gradient descent mode is used for minimizing cost, and then updating weights and biases are back propagated;

step 4.3: and inputting the preliminarily processed test set pictures into a trained improved U-Net neural network model, and performing optic disk rough extraction.