LU500959B1

LU500959B1 - Rough set neural network method for segmentation of fundus retinal blood vessel images

Info

Publication number: LU500959B1
Application number: LU500959A
Authority: LU
Inventors: Ying Sun; Hengrong Ju; Ming Li; Yi Zhang; Zhihao Feng; Weiping Ding; Jie Wan; Jinxin Cao
Original assignee: Univ Nantong
Priority date: 2020-06-18
Filing date: 2021-04-12
Publication date: 2022-05-04
Also published as: WO2021253939A1; LU500959A1; CN111815574B; CN111815574A

Abstract

Provided is a rough set neural network method for segmentation of fundus retinal blood vessel images, which includes the following steps: S10. image preprocessing: obtaining an enhanced fundus retinal blood vessel image based on the rough set; S20. establishing a U- 10 net neural network model; S30. performing optimization training for the U-net neural network model by means of particle swarm optimization (PSO), to obtain a PSO-U-net neural network model; and S40. after image enhancement preprocessing on a to-be-tested color fundus retinal blood vessel image by means of the rough set theory, segmenting the to-be-tested color fundus retinal blood vessel image by using the PSO-U-net neural 15 network model.

Description

ROUGH SET NEURAL NETWORK METHOD FOR SEGMENTATION OF FUNDUS RETINAL BLOOD VESSEL IMAGES

TECHNICAL FIELD The present disclosure relates to the field of medical image processing technologies, and more particularly to a rough set neural network method for segmentation of fundus retinal blood vessel images.

BACKGROUND The health of retinal blood vessels in fundus images 1s of great significance and value for doctors in the early diagnosis of the diabetic cardiovascular disease and various ophthalmic diseases. However, due to the complex structure of the retinal blood vessels and their susceptibility to light in the acquisition environment, the manual segmentation of the retinal blood vessels in clinical practice is not only a huge workload but also requires a high level of experience and skill for the medical staff. In addition, there may be differences in the segmentation result of the same fundus image by different medical staff. Therefore, manual segmentation no longer meets the clinical needs.

With the continuous development of the computer technology, automatic segmentation of the fundus retinal blood vessel image using artificial intelligence technology can effectively assist in early diagnosis and decision making of ophthalmic diseases, which has become a research hotspot for domestic and foreign scholars. The convolutional neural network model in deep learning has unique superiority in medical image processing with its special structure of local perception and parameter sharing. Image information has strong spatial complexity and correlation, and it is likely to encounter problems such as incompleteness and uncertainty during image processing. Therefore, applying the rough set theory to image processing can attain better results than the conventional method in many situations.

SUMMARY To solve the foregoing problem, the present disclosure provides a rough set neural network method for segmentation of fundus retinal blood vessel images, which can reduce the workload of medical staff, avoid differences in a segmentation result of the same fundus image due to differences in experience and skills between the medical staff, and efficiently implement segmentation of a color fundus retinal blood vessel image, thus achieving high segmentation accuracy and efficiency.

To achieve the foregoing objective, the present disclosure adopts the following technical solution: A rough set neural network method for segmentation of fundus retinal blood vessel images is provided, which includes the following steps: S10. image preprocessing: performing image enhancement preprocessing for each standard-RGB color fundus retinal blood vessel image with a size of MxMx3 by using a rough set theory, to obtain an enhanced fundus retinal blood vessel image based on the rough set; S20. establishing a U-net neural network model to segment the enhanced fundus retinal blood vessel image based on the rough set to obtain a segmented image, and using an error between the segmented image and a standard segmentation tagged image corresponding to the standard-RGB color fundus retinal blood vessel image as an error function of the established U-net neural network, to obtain the U-net neural network model; S30. performing optimization training for the U-net neural network model by means of particle swarm optimization (PSO); by using the enhanced fundus retinal blood vessel images based on the rough set as particles, obtaining an optimal population of particles through continuous iteration of the particle swarm; and adjusting parameters of the U-net neural network by means of gradient descent, to obtain a PSO-U-net neural network model; and S40. after image enhancement preprocessing on a to-be-tested color fundus retinal blood vessel image by means of the rough set theory, segmenting the to-be-tested color fundus retinal blood vessel image by using the PSO-U-net neural network model.

Further, the U-net neural network model includes an input layer, a convolutional layer, a ReLU non-linear layer, a pooling layer, a deconvolution layer, and an output layer.

Further, step S10 includes the following sub-steps: S11. storing each standard-RGB color fundus retinal blood vessel image with a size of MxMx3 as three matrices with uniform sizes of MxM, which are respectively recorded as R*, G* and B*, where each value in the matrices indicates a component value of each color of each pixel point in the three channels; and establishing an HIS model by using the matrices R*, G* and B*, where H denotes hue, S denotes saturation, and I denotes brightness:

1 @& Bo | Lu -GHHR -BN | AD where & = arcous pees (1) 4 3 rent ent py oy Sf Foe ne Jr BH {2} {= ! (R467 +B) (3)

S12. the brightness component I being equivalent to a grayscale graph of the fundus retinal blood vessel image and regarded as an image information system, and performing image preprocessing by means of the rough set theory; using a two-dimensional fundus retinal image with a size of MxM as a universe U of discourse, where each pixel point x in the fundus retinal image indicates an object in U, and a gray level of the pixel point x is recorded as f(m, n), (m, n) denoting that the pixel point x is located in the mth row and nth column; determining two condition attributes of the fundus retinal blood vessel grayscale graph as cı and ca, namely, C={c1, c2}, where c1 denotes the gray level attribute of the pixel point, and an attribute value is c:={0, 1}; and cz is recorded as the noise property which indicates an absolute value of the difference between average gray levels of two adjacent sub-blocks, and an attribute value is c:= {0, 1}; a decision attribute D indicating the classification of the pixel point, and D= {d1,d2,d3,d4}, where di denotes a relatively bright noise-free region, da denotes a bright-area edge noise region, ds denotes a relatively dark noise-free region, and ds denotes a dark-area edge noise region, thus constructing a fundus retinal blood vessel image information system (U, CUD); S13. determining a grayscale threshold œ; denoting the gray level of the pixel point x in the mth row and nth column in

> Uas Sc ln, n), where if fc, (nm, n) meets Re,= {x€U]| f,(m,n)>a}, c1=1, which indicates that the gray level of the pixel point x falls within [a+1, 255], and thus, the pixel point is classified as the equivalence class of Rc,, which indicates that the pixel point belongs to a bright set in the fundus retinal blood vessel images; or otherwise, c1=0, which indicates that the gray level of the pixel point x falls within [0, a], and thus, the pixel point is classified as the equivalence class of Rc, which indicates that the pixel point belongs to a dark set in the fundus retinal blood vessel images; S14. determining a noise threshold J; dividing the fundus retinal blood vessel image into sub-blocks by 2x2 pixels, where fc,(m,

n) denotes an absolute value of the difference between average values of pixel gray levels of adjacent sub-blocks, namely, fc,=int|avg(S;;)-avg(S:=1,+1)|, avg(Sij) denoting an average pixel value of the sub-block Sij, 1<i<Z-1, and 1<j<Z-1; and if fo,(m,n) meets Ro, ={XEU] Jo, (m,n)>ß}, c2=1, which indicates that the pixel point x has noise and is classified as the equivalence class of Rc), that is, the pixel belongs to an edge noise set; or otherwise, c2=0, which indicates that the pixel point x does not have noise and is classified as the equivalence class of Re,» that 1s, the pixel belongs to a noise-free set; S15. determining the set to which the pixel point belongs according to the foregoing two condition attributes c1 and ca; and by using the condition attributes as a decision basis, performing decision classification for the pixel point, and dividing the original fundus retinal blood vessel image P into sub-images; and based on the gray level attribute cı and the noise attribute cs, dividing the original image into a relatively bright noise-free sub-image Pi, a bright-area edge noise sub-image P», a relatively dark noise-free sub-image Ps, and a dark-area edge noise sub-image P4 according to the condition attribute ca; completing the relatively bright noise-free sub-image Pi, that is, separately using the grayscale threshold œ and the noise threshold B to fill in all the relatively dark and noisy pixel positions, to form P1'; and completing the relatively dark noise-free sub-image Ps, that is, separately using the grayscale threshold a and the noise threshold B to filling in all the relatively bright and noisy pixel positions, to form P3'; and S16. performing enhanced transformation for Pı' and P3': performing histogram equalization transformation for P1', performing histogram exponential transformation for P3', and superimposing images obtained after histogram 5 transformation of Pı' and P3', to obtain an enhanced fundus retinal blood vessel image P'; and further performing normalization for the enhanced fundus retinal blood vessel image P' according to the following formula (4): x i x) 4 ROFIFE = (4) thus obtaining the enhanced fundus retinal blood vessel image based on the rough set, where x; denotes the value of the ith pixel point in the fundus retinal blood vessel image, and min(x) and max(x) respectively denote a minimum value and a maximum value of the pixel of the fundus retinal blood vessel image.

Further, step S20 includes the following sub-steps: S21. performing feature extraction for the enhanced fundus retinal blood vessel images based on the rough set by means of downsampling, performing a convolution operation twice for the input fundus retinal blood vessel images by using a convolution kernel with a size of 3x3, and performing non-linear transformation by using the ReLU activation function; then, performing 2x2 pooling operation and repeating it four times, where in the first 3x3 convolution operation after each pooling operation, the number of 3x3 convolutional kernels exponentially increases; and afterwards, performing 3x3 convolution operation twice, to continue the foregoing operation related to feature extraction by means of downsampling, where calculation for the convolutional layer is shown as follows: x= SOK +55) {5) M; denoting an input feature map set, x;" denoting the jth feature map in the nth layer, Ki; denoting a convolution kernel function, f() denoting the activation function and the ReLU function being selected as the activation function, and bj” denoting an offset parameter; and calculation for the pooling layer 1s shown as follows: x = FU dou 3 BT) (6) denoting a weight constant of the feature map of the downsampling layer, and down() denoting a downsampling function; S22. performing operations by means of upsampling: first, performing 3x3 deconvolution operation twice, copying and cropping images in the maximum pooling layer, and splicing the processed images and the images obtained after deconvolution; then, performing 3x3 convolution operation and repeating it four times, where in the first 3x3 convolution operation after each splicing operation, the number of 3x3 convolutional kernels exponentially decreases; and finally, performing 3x3 convolution operation twice and 1x1 convolution operation once, thus completing the upsampling process; and S23. after the upsampling and downsampling process, calculating an error between the segmented image obtained by using the U-net neural network and the standard segmentation tagged image corresponding to the standard-RGB color fundus retinal blood vessel image by means of forward calculation, where an error function is shown as follows: Cl Sr _ es A Es 229 out (= > tre) (7) where T denotes the number of fundus image samples input into the U-net neural network, y _out:(1) denotes the gray level of the ith pixel point in the tth fundus retinal image sample output by the U-net neural network, and y_true(i) denotes the gray level of the ith pixel point in the ith fundus retinal image tag.

Further, an error threshold is set in the sub-step S23 and is equal to 0.1, and when the error 1s not greater than the error threshold, the required U-net neural network model is obtained; or when the error is greater than the error threshold, backpropagation is performed according to a gradient descent algorithm to adjust a network weight, and then steps S21 to S22 are repeated to perform forward calculation until the error is not greater than the error threshold.

Further, step S30 includes the following sub-steps: S31. randomly selecting few H fundus images from a training set of the enhanced fundus retinal blood vessel images based on the rough set as reference images, and denoting a particle swarm Q as Q= (Qi, Qa, ..., Qn), where H denotes the number of particles in the particle swarm Q and is consistent with the number of the selected fundus images, each bit of each particle represents a link weight or threshold, and an encoding manner of the ith particle Qi is Qi= {Qi1, Qi, ..., Qin}, n denoting a total number of the link weights or thresholds; initializing acceleration constants 61 and 02, and an initial value of the inertia weight w; and initializing each particle position vector Yi= {vil, Yi2, …, Yin} and particle velocity vector Vi= {vi1, via, ..., Vin} into random numbers in the interval of [0,1], where n denotes the number of parameters in the U-net model; S32. separately completing the downsampling process and the upsampling process for each particle in the U-net model; calculating the fitness of each particle by using the error function of the U-net neural network as a fitness function of the particle swarm, and arranging the particles in an ascending order, to obtain an optimal position pbest of each particle and an optimal position gbest of the whole particle swarm; S33. if a minimal value in the error threshold range 1s reached, it indicating that the training has converged, and then stopping running; or otherwise, continuously updating the position and velocity of each particle according to the following formulas (8) and (9):

VAE WV pro; * rand (0 {phestix in) +02 * rand () * (gbestio-x in) (8) X in Kit Vin (9) where vin and xin respectively denote the current position and velocity of the particle 1, V'in and x'in respectively denote the updated velocity and position of the particle 1, w denotes the inertia weight, 61 and 02 are acceleration constants, and rand() is a random function in the interval of [0,1]; S34. sending the updated particles back to the U-net neural network to update a link weight to be trained, performing the upsampling and downsampling process again, and calculating its error; and S35. splitting the obtained optimal position gbest of the particle swarm and mapping it to the weight and threshold of the U-net neural network model, thus completing the whole PSO-based process of optimization of the U-

net neural network weight. The foregoing technical solution of the present disclosure has the following advantages compared to the prior art: The rough set neural network method for segmentation of fundus retinal blood vessel images in the present disclosure reduces the workload of medical staff, avoids differences in a segmentation result of the same fundus image due to differences in experience and skills between the medical staff, and efficiently implements segmentation of a color fundus retinal blood vessel image, thus achieving high segmentation accuracy and efficiency.

BRIEF DESCRIPTION OF THE DRAWINGS The technical solution of the present disclosure and its beneficial effects will be apparent by the following detailed description of specific embodiments of the present disclosure with reference to the accompanying drawings.

FIG. 1 is a flowchart of a rough set neural network method for segmentation of fundus retinal blood vessel images in an embodiment of the present disclosure; FIG. 2 is a detailed flowchart of a rough set neural network method for segmentation of fundus retinal blood vessel images in an embodiment of the present disclosure; and FIG. 3 is a structural diagram of a U-net neural network model in an embodiment of the present disclosure.

DETAILED DESCRIPTION OF THE EMBODIMENTS The technical solution in the embodiments of the present disclosure is clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure. Apparently, the described embodiments are some rather than all of the embodiments of the present disclosure. Based on the embodiments of the present disclosure, other embodiments acquired by those of ordinary skill in the art without creative effort all belong to the protection scope of the present disclosure.

This embodiment provides a rough set neural network method for segmentation of fundus retinal blood vessel images, which, as shown in FIGs. 1 and 2, includes the following steps: S10. Image preprocessing: Image enhancement preprocessing 1s performed for each standard-RGB color fundus retinal blood vessel image with a size of MxMx3 by using a rough set theory, to obtain an enhanced fundus retinal blood vessel image based on the rough set. S20. A U-net neural network model is established to segment the enhanced fundus retinal blood vessel image based on the rough set to obtain a segmented image, and an error between the segmented image and a standard segmentation tagged image corresponding to the standard-RGB color fundus retinal blood vessel image 1s used as an error function of the established U-net neural network, to obtain the U-net neural network model. S30. Optimization training 1s performed for the U-net neural network model by means of PSO; by using the enhanced fundus retinal blood vessel images based on the rough set as particles, an optimal population of particles is obtained through continuous iteration of the particle swarm; and parameters of the U-net neural network are adjusted by means of gradient descent, to obtain a PSO-U-net neural network model. S40. After image enhancement preprocessing on a to-be-tested color fundus retinal blood vessel image by means of the rough set theory, the to-be-tested color fundus retinal blood vessel image is segmented by using the PSO-U-net neural network model.

Step S10 includes the following sub-steps: S11. Fach standard-RGB color fundus retinal blood vessel image with a size of MxM*3 is stored as three matrices with uniform sizes of MxM, which are respectively recorded as R*, G* and B*, where each value in the matrices indicates a component value of each color of each pixel point in the three channels; and an HIS model is established by using the matrices R*, G* and B*, where H denotes hue, S denotes saturation, and I denotes brightness: . ep (OR GREEN | = hero pag whee De ee hos CD Sef em RAA) (23 i= : (R + +8) (3)

S12. The brightness component I is equivalent to a grayscale graph of the fundus retinal blood vessel image and is regarded as an image information system, and image preprocessing is performed by means of the rough set theory.

A two-dimensional fundus retinal image with a size of MxM is used as a universe U of discourse, where each pixel point x in the fundus retinal image indicates an object in U, and a gray level of the pixel point x is recorded as f(m, n), (m, n) denoting that the pixel point x is located in the mth row and nth column.

Two condition attributes of the fundus retinal blood vessel grayscale graph are determined as ci and cz, namely, C= {ci, ca}, where c1 denotes the gray level attribute of the pixel point, and an attribute value is cı= {0, 1}; and cz is recorded as the noise property which indicates an absolute value of the difference between average gray levels of two adjacent sub-blocks, and an attribute value is co= {0, 1}. A decision attribute D indicates the classification of the pixel point, and D= {di,d2,d3,ds}, where di denotes a relatively bright noise-free region, d» denotes a bright-area edge noise region, ds denotes a relatively dark noise-free region, and ds denotes a dark-area edge noise region, thus constructing a fundus retinal blood vessel image information system (U, CUD). S13. A grayscale threshold a is determined.

The gray level of the pixel point x in the mth row and nth column in U is denoted as fe, (m, n); and if fc (m, n) meets Re,= {xEU] fe,(m,n)>a}, c1=1, which indicates that the gray level of the pixel point x falls within [o+1, 255]. Thus, the pixel point is classified as the equivalence class of R. , which indicates that the pixel point belongs to a bright set in the fundus retinal blood vessel images.

Otherwise, cı =0, which indicates that the gray level of the pixel point x falls within [0, a], Thus, the pixel point is classified as the equivalence class of Re, which indicates that the pixel point belongs to a dark set in the fundus retinal blood vessel images.

S14. A noise threshold ß is determined.

The fundus retinal blood vessel image is divided into sub-blocks by 2x2 pixels, where f,(m, n) denotes an absolute value of the difference between average values of pixel gray levels of adjacent sub-blocks, namely, Je, =intlavg(S;;)-avg(Siz1,+1)|, where avg(Sij) denotes an average pixel value of the sub-

block Si I<i<=-1, and 1<j<5-1. If f.,(m,n) meets R,={xEU] f,(m,n)>B}, c2=1, which indicates that the pixel point x has noise and is classified as the equivalence class of Ke, that is, the pixel belongs to an edge noise set. Otherwise, c:=0, which indicates that the pixel point x does not have noise and is classified as the equivalence class of Re,» that 1s, > the pixel belongs to a noise-free set. S15. The set to which the pixel point belongs is determined according to the foregoing two condition attributes cı and cz; and by using the condition attributes as a decision basis, decision classification is performed for the pixel point, and the original fundus retinal blood vessel image P is divided into sub-images. Based on the gray level attribute cı and the noise attribute cs, the original image is divided into a relatively bright noise-free sub- image Pi, a bright-area edge noise sub-image P», a relatively dark noise-free sub-image Ps, and a dark-area edge noise sub-image P4. The relatively bright noise-free sub-image P1 is completed. That is, the grayscale threshold a and the noise threshold B are separately used to fill in all the relatively dark and noisy pixel positions, to form P1'. The relatively dark noise-free sub-image P; is completed. That is, the grayscale threshold a and the noise threshold B are separately used to fill in all the relatively bright and noisy pixel positions, to form Ps".

S16. Perform enhanced transformation for Pi and Ps": Histogram equalization transformation is performed for P1', histogram exponential transformation is performed for P+', and images obtained after histogram transformation of P1' and P+' are superimposed, to obtain an enhanced fundus retinal blood vessel image P'; and further normalization is performed for the enhanced fundus retinal blood vessel image P' according to the following formula (4): x — min x) x AOE = ee {4} Thus, the enhanced fundus retinal blood vessel image based on the rough set is obtained,

where xi denotes the value of the ith pixel point in the fundus retinal blood vessel image, and min(x) and max(x) respectively denote a minimum value and a maximum value of the pixel of the fundus retinal blood vessel image.

As shown in FIG. 3, the U-net neural network model includes an input layer, a convolutional layer, a ReLU non-linear layer, a pooling layer, a deconvolution layer, and an output layer. Step S20 includes the following sub-steps: S21. Feature extraction is performed for the enhanced fundus retinal blood vessel images based on the rough set by means of downsampling, a convolution operation is performed twice for the input fundus retinal blood vessel images by using a convolution kernel with a size of 3x3, and non- linear transformation is performed by using the ReLU activation function. Then, 2x2 pooling operation is performed and repeated four times, where in the first 3x3 convolution operation after each pooling operation, the number of 3x3 convolutional kernels exponentially increases. Afterwards, 3x3 convolution operation is performed twice, to continue the foregoing operation related to feature extraction by means of downsampling, Calculation for the convolutional layer is shown as follows: x; = f à a” A K: + £7) {3 where M; denotes an input feature map set, x;* denotes the jth feature map in the nth layer, K{} denotes a convolution kernel function, f() denotes the activation function and the ReLU function is selected as the activation function, and bj" denotes an offset parameter. Calculation for the pooling layer 1s shown as follows: x = FL dom) +B) {6) where ß denotes a weight constant of the feature map of the downsampling layer, and down() denotes a downsampling function.

S22. Operations are performed by means of upsampling. First, 3x3 deconvolution operation is performed twice, images in the maximum pooling layer are copied and cropped, and the processed images and the images obtained after deconvolution are spliced.

Then, 3x3 convolution operation 1s performed and repeated four times, where in the first 3x3 convolution operation after each splicing operation, the number of 3x3 convolutional kernels exponentially decreases.

Finally, 3x3 convolution operation 1s performed twice and 1x1 convolution operation is performed once, thus completing the upsampling process.

S23. After the upsampling and downsampling process, an error between the segmented image obtained by using the U-net neural network and the standard segmentation tagged image corresponding to the standard-RGB color fundus retinal blood vessel image is calculated by means of forward calculation, where an error function is shown as follows: Es Ze SE Su He {7} where T denotes the number of fundus image samples input into the U-net neural network, y_out:(1) denotes the gray level of the ith pixel point in the tth fundus retinal image sample output by the U-net neural network, and y_true((1) denotes the gray level of the ith pixel point in the ith fundus retinal image tag.

Step S30 includes the following sub-steps: S31. Few H fundus images are randomly selected from a training set of the enhanced fundus retinal blood vessel images based on the rough set as reference images, and a particle swarm Q is denoted as Q=(Q1, Q, ..., Qn), where H denotes the number of particles in the particle swarm Q and is consistent with the number of the selected fundus images, each bit of each particle represents a link weight or threshold, and an encoding manner of the ith particle Qi is Qi= {Qi1, Qi, ..., Qin}, n denoting a total number of the link weights or thresholds.

Acceleration constants 61 and 62, and an initial value of the inertia weight w are initialized; and each particle position vector Yi= {yi1, Yi2, ..., Vin} and particle velocity vector Vi= {vi1, vi, ..., Vin} are initialized into random numbers in the interval of [0,1], where n denotes the number of parameters in the U-net model.

S32. The downsampling process and the upsampling process are separately completed for each particle in the U-net model; the fitness of each particle is calculated by using the error function of the U-net neural network as a fitness function of the particle swarm, and the particles are arranged in an ascending order, to obtain an optimal position pbest of each particle and an optimal position gbest of the whole particle swarm. S33. If a minimal value in the error threshold range is reached, it indicates that the training has converged, and then running is stopped; or otherwise, the position and velocity of each particle are continuously updated according to the following formulas (8) and (9): voemmwy tay * vand (Of {phéstis-x ad +02 * rand (OO + {gbestia-X is) (8) £ exit nn (9) In the foregoing formulas, Vin and xin respectively denote the current position and velocity of the particle 1, V'in and x'in respectively denote the updated velocity and position of the particle 1, w denotes the inertia weight, 61 and 02 are acceleration constants, and rand() 1s a random function in the interval of [0,1]. S34. The updated particles are sent back to the U-net neural network to update a link weight to be trained, the upsampling and downsampling process is performed again, and its error is calculated. S35. The obtained optimal position gbest of the particle swarm is split and mapped to the weight and threshold of the U-net neural network model, thus completing the whole PSO-based process of optimization of the U-net neural network weight.

The above merely describes exemplary embodiments of the present disclosure, and does not limit the patent protection scope of the present disclosure. Any equivalent structures or process transformations made by using the description and the accompanying drawings of the present disclosure can be applied directly or indirectly in other related technical fields, and all fall within the patent protection scope of the present disclosure.

Claims

1. A rough set neural network method for segmentation of fundus retinal blood vessel images, comprising the following steps: S10. image preprocessing: performing image enhancement preprocessing for each standard-RGB color fundus retinal blood vessel image with a size of MxMx3 by using a rough set theory, to obtain an enhanced fundus retinal blood vessel image based on the rough set; S20. establishing a U-net neural network model to segment the enhanced fundus retinal blood vessel image based on the rough set to obtain a segmented image, and using an error between the segmented image and a standard segmentation tagged image corresponding to the standard-RGB color fundus retinal blood vessel image as an error function of the established U-net neural network, to obtain the U-net neural network model; S30. performing optimization training for the U-net neural network model by means of particle swarm optimization (PSO); by using the enhanced fundus retinal blood vessel images based on the rough set as particles, obtaining an optimal population of particles through continuous iteration of the particle swarm; and adjusting parameters of the U-net neural network by means of gradient descent, to obtain a PSO-U-net neural network model; and S40. after image enhancement preprocessing on a to-be-tested color fundus retinal blood vessel image by means of the rough set theory, segmenting the to-be-tested color fundus retinal blood vessel image by using the PSO-U-net neural network model.

2. The rough set neural network method for segmentation of fundus retinal blood vessel images according to claim 1, wherein the U-net neural network model comprises an input layer, a convolutional layer, a ReLU non-linear layer, a pooling layer, a deconvolution layer, and an output layer.

3. The rough set neural network method for segmentation of fundus retinal blood vessel images according to claim 2, wherein step S10 comprises the following sub-steps: S11. storing each standard-RGB color fundus retinal blood vessel image with a size of MxMx3 as three matrices with uniform sizes of MxM, which are respectively recorded as R*, G* and B*, wherein each value in the matrices indicates a component value of each color of each pixel point in the three channels; and establishing an HIS model by using the matrices R*, G* and B*, wherein H denotes hue, S denotes saturation, and I denotes brightness: mal 0 BEG oe aces | DEAN | Bevo, BSG | JR GAR GN F3) | 3 mt 5 n° ~~ Safe ETERS GB} (2) I= RG 48) (3) 3 S12. the brightness component I being equivalent to a grayscale graph of the fundus retinal blood vessel image and regarded as an image information system, and performing image preprocessing by means of the rough set theory; using a two-dimensional fundus retinal 1mage with a size of MxM as a universe U of discourse, wherein each pixel point x in the fundus retinal image indicates an object in U, and a gray level of the pixel point x is recorded as f(m, n), (m, n) denoting that the pixel point x is located in the mth row and nth column; determining two condition attributes of the fundus retinal blood vessel grayscale graph as c; and cz, namely, C={c1, cz}, wherein c1 denotes the gray level attribute of the pixel point, and an attribute value is c1={0, 1}; and c is recorded as the noise property which indicates an absolute value of the difference between average gray levels of two adjacent sub-blocks, and an attribute value is c= {0, 1}; a decision attribute D indicating the classification of the pixel point, and D= {d1,d2,d3,d4}, wherein d; denotes a relatively bright noise-free region, dz denotes a bright-area edge noise region, ds denotes a relatively dark noise-free region, and ds denotes a dark-area edge noise region, thus constructing a fundus retinal blood vessel image information system (U, CUD); S13. determining a grayscale threshold a; denoting the gray level of the pixel point x in the mth row and nth column in U as f1(m, n), wherein if fc1(m, n) meets Re1= {xEU] fc1(m,n)>a}, cı=1, which indicates that the gray level of the pixel point x falls within [a+1, 255]; and thus, the pixel point is classified as the equivalence class of Re1, which indicates that the pixel point belongs to a bright set in the fundus retinal blood vessel images; or otherwise, ¢1=0, which indicates that the gray level of the pixel point x falls within [0, a], and thus, the pixel point is classified as the equivalence class of Re, which indicates that the pixel point belongs to a dark set in the fundus retinal blood vessel images.

S14. determining a noise threshold PB; dividing the fundus retinal blood vessel image into sub-blocks by 2x2 pixels, wherein (m, n) denotes an absolute value of the difference between average values of pixel gray levels of adjacent sub-blocks, namely, Jo=intlavg(S;;)-avg(S=1;=1)|, avg(Sij) denoting an average pixel value of the sub-block Sj, I<i<=-1, and 1<j<5-1; and if f2(m,n) meets R2={xEU] foa(m,n}>B}, &;=1, which indicates that the pixel point x has noise and is classified as the equivalence class of Re, that is, the pixel belongs to an edge noise set; or otherwise, c2=0, which indicates that the pixel point x does not have noise and is classified as the equivalence class of Re that is, the pixel belongs to a noise-free set; S15. determining the set to which the pixel point belongs according to the foregoing two condition attributes cı and ca; and by using the condition attributes as a decision basis, performing decision classification for the pixel point, and dividing the original fundus retinal blood vessel image P into sub-images; and based on the gray level attribute c1 and the noise attribute ca, dividing the original image into a relatively bright noise-free sub-

image Pi, a bright-area edge noise sub-image P», a relatively dark noise-free sub-image Ps, and a dark-area edge noise sub-image P4; completing the relatively bright noise-free sub- image Pi, that is, separately using the grayscale threshold a and the noise threshold ß to fill in all the relatively dark and noisy pixel positions, to form P1'; and completing the relatively dark noise-free sub-image Ps, that is, separately using the grayscale threshold œ and the noise threshold ß to fill in all the relatively bright and noisy pixel positions, to form Ps’; and S16. performing enhanced transformation for P;' and Ps': performing histogram equalization transformation for P1', performing histogram exponential transformation for Ps', and superimposing images obtained after histogram transformation of P1' and P3', to obtain an enhanced fundus retinal blood vessel image P'; and further performing normalization for the enhanced fundus retinal blood vessel image P' according to the following formula (4): x — main x Ay rors = x ; me x} 4) thus obtaining the enhanced fundus retinal blood vessel image based on the rough set, wherein x; denotes the value of the ith pixel point in the fundus retinal blood vessel image, and min(x) and max(x) respectively denote a minimum value and a maximum value of the pixel of the fundus retinal blood vessel image.

4. The rough set neural network method for segmentation of fundus retinal blood vessel images according to claim 3, wherein step S20 comprises the following sub-steps: S21. performing feature extraction for the enhanced fundus retinal blood vessel images based on the rough set by means of downsampling, performing a convolution operation twice for the input fundus retinal blood vessel images by using a convolution kernel with a size of 3x3, and performing non-linear transformation by using the ReLU activation function; then, performing 2x2 pooling operation and repeating it four times, wherein in the first 3x3 convolution operation after each pooling operation, the number of 3x3 convolutional kernels exponentially increases; and afterwards, performing 3x3 convolution operation twice, to continue the foregoing operation related to feature extraction by means of downsampling, wherein calculation for the convolutional layer is shown as follows: x; fi 2x KB) (5) M; denoting an input feature map set, x;" denoting the jth feature map in the nth layer, Kj} denoting a convolution kernel function, f() denoting the activation function and the ReLU function being selected as the activation function, and bj denoting an offset parameter; calculation for the pooling layer 1s shown as follows: Xm {BT down x ye 8) {6) B denoting a weight constant of the feature map of the downsampling layer, and down() denoting a downsampling function: S22. performing operations by means of upsampling: first, performing 3x3 deconvolution operation twice, copying and cropping images in the maximum pooling layer, and splicing the processed images and the images obtained after deconvolution; then, performing 3x3 convolution operation and repeating it four times, wherein in the first 3x3 convolution operation after each splicing operation, the number of 3x3 convolutional kernels exponentially decreases; and finally, performing 3x3 convolution operation twice and 1x1 convolution operation once, thus completing the upsampling process; and S23. after the upsampling and downsampling process, calculating an error between the segmented image obtained by using the U-net neural network and the standard segmentation tagged image corresponding to the standard-RGB color fundus retinal blood vessel image by means of forward calculation, wherein an error function is shown as follows:

Em 37 2. ZU YOU suey (7) wherein T denotes the number of fundus image samples input into the U-net neural network, y_outi(i) denotes the gray level of the ith pixel point in the tth fundus retinal image sample output by the U-net neural network, and y_true((1) denotes the gray level of the ith pixel point in the ith fundus retinal image tag.

5. The rough set neural network method for segmentation of fundus retinal blood vessel images according to claim 4, wherein an error threshold is set in the sub-step S23 and is equal to 0.1, and when the error is not greater than the error threshold, the required U-net neural network model is obtained; or when the error is greater than the error threshold, backpropagation is performed according to a gradient descent algorithm to adjust a network weight, and then steps S21 to S22 are repeated to perform forward calculation until the error 1s not greater than the error threshold.

6. The rough set neural network method for segmentation of fundus retinal blood vessel images according to claim 5, wherein step S30 comprises the following sub-steps: S31. randomly selecting few H fundus images from a training set of the enhanced fundus retinal blood vessel images based on the rough set as reference images, and denoting a particle swarm Q as Q= (Q1, Q>, ..., Qu), wherein H denotes the number of particles in the particle swarm Q and is consistent with the number of the selected fundus images, each bit of each particle represents a link weight or threshold, and an encoding manner of the ith particle Qi 1s Qi= {Qi1, Qi2, ..., Qin}, n denoting a total number of the link weights or thresholds; initializing acceleration constants 61 and 02, and an initial value of the inertia weight w; and initializing each particle position vector Yi= {yii, Vi, ..., Vin} and particle velocity vector Vi= {vii Vi, …, Vin} into random numbers in the interval of [0,1], wherein n denotes the number of parameters in the U-net model;

S32. separately completing the downsampling process and the upsampling process for each particle in the U-net model; calculating the fitness of each particle by using the error function of the U-net neural network as a fitness function of the particle swarm, and arranging the particles in an ascending order, to obtain an optimal position pbest of each particle and an optimal position gbest of the whole particle swarm;

S33. if a minimal value in the error threshold range is reached, it indicating that the training has converged, and then stopping running; or otherwise, continuously updating the position and velocity of each particle according to the following formulas (8) and (9):

voy tor * rand (OO * (bests is) +04 * rand 0 * (gbestis-X us) (8) X n= Mint Vin (9)

wherein vin and xin respectively denote the current position and velocity of the particle 1, v'in and x'in respectively denote the updated velocity and position of the particle 1, w denotes the inertia weight, 61 and 62 are acceleration constants, and rand() is a random function in the interval of [0,1];

S34. sending the updated particles back to the U-net neural network to update a link weight to be trained, performing the upsampling and downsampling process again, and calculating its error; and S35. splitting the obtained optimal position gbest of the particle swarm and mapping 1t to the weight and threshold of the U-net neural network model, thus completing the whole PSO-based process of optimization of the U-net neural network weight.