WO2023240674A1

WO2023240674A1 - Fundus image quality optimization method based on deep learning

Info

Publication number: WO2023240674A1
Application number: PCT/CN2022/100938
Authority: WO
Inventors: 高荣玉; 张�杰
Original assignee: 潍坊眼科医院有限责任公司
Priority date: 2022-06-15
Filing date: 2022-06-24
Publication date: 2023-12-21
Also published as: CN115018799A; CN115018799B

Abstract

The present invention relates to the technical field of fundus image analysis, and in particular, to a fundus image quality optimization method based on deep learning, which can quickly recognize a fundus image, accurately position a lesion position of a fundus disease, and optimize a processing process of the fundus image. The method comprises the following steps: step 1, acquiring the fundus image, and performing sharpening and grayscale processing on the fundus image; step 2, performing circle center determination on the image subjected to the grayscale processing in step 1, and performing concentric ring band segmentation according to the determined circle center, so as to obtain a plurality of groups of concentric equidistant ring bands; and uniformly segmenting each group of concentric ring bands in the radial direction into a plurality of groups of pixel blocks; step 3, calculating an average grayscale value of each group of pixel blocks in step 2, and performing grayscale value assignment on the pixel blocks to obtain a circular numerical matrix; and step 4, performing convolution calculation on the circular numerical matrix in step 3, and obtaining a corresponding lesion area image according to a convolution calculation result.

Description

A fundus image quality optimization method based on deep learning

Technical field

The present invention relates to the technical field of fundus image analysis, and in particular to a fundus image quality optimization method based on deep learning.

Background technique

Fundus images are an important reference for examining diseases of the vitreous body, retina, choroid and optic nerve. Many systemic diseases such as hypertension and diabetes can cause fundus lesions, so fundus images are an important diagnostic data;

Existing fundus image processing methods use traditional matrix convolution calculations. Due to the particular shape of the fundus image, when using matrix convolution calculations for identification, the positioning efficiency and positioning accuracy of the lesion location are relatively low.

Contents of the invention

In order to solve the above technical problems, the present invention provides a fundus image quality optimization method based on deep learning that can quickly identify fundus images, accurately locate fundus disease lesions, and optimize the processing process of fundus images.

A fundus image quality optimization method based on deep learning of the present invention includes the following steps:

Step 1. Obtain the fundus image and perform sharpening and grayscale processing on the fundus image;

Step 2: Determine the center of the circle on the grayscale processed image in step 1, and segment concentric rings based on the determined center of the circle to obtain several groups of concentric and equidistant rings; and divide each group of concentric rings evenly along the radial direction. into several groups of pixel blocks;

Step 3. Calculate the average grayscale value of each group of pixel blocks in step 2, and perform grayscale assignment on this pixel block to obtain a circular array;

Step 4: Perform convolution calculation on the circular array in step 3, and obtain the corresponding lesion area image based on the convolution calculation result.

Further, in step 2, the circle center of the grayscale processed fundus image is confirmed through MATLAB. The circle center positioning procedure includes the following steps:

B＝imread(‘fundus image’); % read the original image

A=im2bw(B);% binarization

[x,y]=find(A==0);%Coordinate set of edge pixels in the fundus image

center_x=min(x)+(max(x)-min(x))/2;

center_y=min(y)+(max(y)-min(y))/2;

center＝[center_x,center_y];% This is the center coordinate of the fundus image

Among them, fundus image formats include BMP, GIF, HDF, JPEG, PCX, PNG, TIFF and XWD.

Further, in step 3, the average gray value of each group of pixel blocks is calculated through MATLAB. The calculation program includes the following steps:

II1=imread(‘pixel block’); % read image

II2=imread(‘pixel block’); % read image

I1=rgb2gray(II1);

I2=rgb2gray(II2);

startX＝350; endX＝400; % Set the starting coordinate and end coordinate of the abscissa of the pixel block

startY＝300; endY＝350; % Set the starting coordinate and end coordinate of the vertical coordinate of the pixel block

aver1=mean(mean(I1(startY:endY,startX:endX)))% select the average gray value of the pixel block

aver2＝mean(mean(I2(startY:endY,startX:endX)))%90 degree selected pixel block average gray value

p1=(aver1-aver2)/(aver1+aver2)% This is the average gray value of the pixel block.

Further, the convolution calculation in step 4 includes the following steps:

S1. Enter the circular array after grayscale assignment in step 3;

S2. Use the preliminary convolution kernel to traverse and calculate the circular array in S1 to obtain the primary convolution result;

S3. Map and compare the primary convolution results in S2 with the pre-stored primary convolution results of the normal fundus image one by one, and extract the circular array corresponding to the part where the absolute value of the difference in the primary convolution result is greater than the preset threshold. Pixel blocks, i.e. abnormal blocks in fundus images;

S4. Use the positioning convolution kernel to traverse and calculate the abnormal blocks extracted in S3 to obtain the secondary convolution result;

S5. Map and compare the lesion convolution results in S4 and the pre-stored secondary convolution results of the normal fundus image one by one, and extract the circles corresponding to the parts of the secondary convolution results where the absolute value of the difference is greater than the preset threshold. The array area is used to obtain the image of the lesion area in the fundus image.

Furthermore, the preliminary screening convolution kernel adopts a sector-shaped structure with the same radius as the circular array. When the preliminary screening convolution kernel performs the traversal calculation, the center of the preliminary screening convolution kernel coincides with the center of the circular array, clockwise. Rotate, each rotation angle is the arc angle corresponding to the pixel block. Each rotation can obtain a set of convolution results, until the circular array is traversed, and the calculated single-ring array is the primary convolution result.

Furthermore, the positioning convolution kernel adopts a partial ring structure with the same radius as the circular array. When the positioning convolution kernel performs traversal calculation, the center of the positioning convolution kernel coincides with the center of the circle where the abnormal block is located, and rotates clockwise. , convolute and calculate the pixel blocks in the same annulus in the abnormal block one by one until all the annulus in the same abnormal block are calculated, that is, the second-level convolution result is obtained.

Furthermore, the annulus width in step 2 can be adjusted according to different fundus lesions. The smaller the lesion range, the narrower the annulus width.

Further, a fundus image template is preset based on the starting position relative to the optic papilla during convolution calculation. In step 2, the angle of the fundus image sample is adjusted based on the fundus image template until the sample optic papilla coincides with the template optic papilla.

Compared with the existing technology, the beneficial effects of the present invention are: due to the circular characteristics of the fundus image edge, the traditional matrix convolution calculation is optimized and designed into a circular matrix, and a sector-shaped convolution kernel and a partial annular convolution are used. The convolution calculation of the convolution kernel can more accurately locate the location of fundus lesions; at the same time, the lateral movement and vertical movement of the convolution kernel in the matrix convolution calculation are designed as rotational movements, which can improve the convolution calculation efficiency, thereby improving The positioning and extraction rate of the fundus image lesion area; the second-order convolution calculation is performed through the initial screening convolution kernel and the positioning convolution kernel. Compared with directly using the positioning convolution kernel to perform the first-order convolution calculation on the circular array, the calculation It is faster and achieves the "1+1<1" effect in terms of time; through the above settings, this method can quickly identify fundus images, accurately locate fundus disease lesions, and optimize the processing of fundus images.

Description of the drawings

Figure 1 is the logic flow chart of the first-level convolution operation;

Figure 2 is a logic flow chart of the two-level convolution operation.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all the embodiments.

Example 1

As shown in Figures 1 and 2, the fundus image quality optimization method based on deep learning of the present invention includes the following steps:

Step 2. Use MATLAB to confirm the center of the circle on the grayscale processed fundus image. The circle center positioning procedure includes the following steps:

B＝imread(‘fundus image’); % read the original image

A=im2bw(B);% binarization

[x,y]=find(A==0);%Coordinate set of edge pixels in the fundus image

center_x=min(x)+(max(x)-min(x))/2;

center_y=min(y)+(max(y)-min(y))/2;

Among them, the fundus image formats are BMP, GIF, HDF, JPEG, PCX, PNG, TIFF and The group of rings is evenly divided into several groups of pixel blocks, as shown in the fundus image in Figure 1;

Step 3. Calculate the average gray value of each group of pixel blocks through MATLAB. The calculation procedure includes the following steps:

II1=imread(‘pixel block’); % read image

II2=imread(‘pixel block’); % read image

I1=rgb2gray(II1);

I2=rgb2gray(II2);

p1=(aver1-aver2)/(aver1+aver2)% This is the average gray value of the pixel block;

Assign the obtained average gray value of the pixel block to the corresponding fundus image in Figure 1 to obtain a circular array;

Step 4. Perform convolution calculation on the circular array in S3 through the following steps:

S1. Enter the circular array after grayscale assignment in step 3;

S2. Use the preliminary screening convolution kernel to traverse and calculate the circular array in S1. The preliminary screening convolution kernel adopts a sector-shaped structure with the same radius as the circular array. When performing the traversal calculation, the preliminary screening convolution kernel makes the initial screening The center of the convolution kernel coincides with the center of the circular array. Rotate clockwise. Each rotation angle is the arc angle corresponding to the pixel block. A set of convolution results can be obtained with each rotation until the circular array is traversed. , the calculated single-ring array is the primary convolution result;

S4. Use the positioning convolution kernel to traverse the abnormal blocks extracted in S3. The positioning convolution kernel adopts a partial ring structure with the same radius as the circular array. When performing the traversal calculation, the positioning convolution kernel uses The center of the circle of the kernel coincides with the center of the circle where the abnormal block is located. Turn clockwise to convolve and calculate the pixel blocks in the same annulus in the abnormal block one by one until all the annulus in the same abnormal block are calculated, that is, the second level is obtained. convolution result;

In this embodiment, due to the rounded edge characteristics of the fundus image, the traditional matrix convolution calculation is optimized and designed into a circular matrix, and a sector-shaped convolution kernel and a partial annular convolution kernel are used to perform the convolution calculation. , can more accurately locate the location of fundus lesions; at the same time, the lateral movement and vertical movement of the convolution kernel in the matrix convolution calculation are designed as rotational movements, which can improve the convolution calculation efficiency, thereby improving the positioning and extraction of the fundus image lesion area. Speed; the second-order convolution calculation is performed through the initial screening convolution kernel and the positioning convolution kernel. Compared with directly using the positioning convolution kernel to perform the first-order convolution calculation on the circular array, the calculation speed is faster, in terms of time To achieve the effect of "1+1<1", the initial screening convolution kernel and positioning convolution kernel in this method are obtained through deep learning of a large number of normal fundus images; through the above settings, this method can quickly identify fundus images. Accurately locate fundus disease lesions and optimize the processing of fundus images.

Example 2

Step 2: Calculate the starting position relative to the optic papilla based on the convolution, preset the fundus image template, rely on the fundus image template to adjust the angle of the fundus image obtained in step 1 until the sample optic papilla coincides with the template optic papilla, and then pass MATLAB confirms the center of the circle on the grayscale processed fundus image. The circle center positioning procedure includes the following steps:

B＝imread(‘fundus image’); % read the original image

A=im2bw(B);% binarization

[x,y]=find(A==0);%Coordinate set of edge pixels in the fundus image

center_x=min(x)+(max(x)-min(x))/2;

center_y=min(y)+(max(y)-min(y))/2;

Among them, the fundus image formats are BMP, GIF, HDF, JPEG, PCX, PNG, TIFF and The group of annulus is evenly divided into several groups of pixel blocks. The width of the concentric and equidistant annulus and the arc angle of the pixel block can be adjusted according to different fundus lesions. The smaller the lesion range, the narrower the width of the annulus, and the arc angle of the pixel block. The smaller the shape angle, conversely, the larger the lesion range, the wider the annulus width, and the larger the arc angle of the pixel block;

II1=imread(‘pixel block’); % read image

II2=imread(‘pixel block’); % read image

I1=rgb2gray(II1);

I2=rgb2gray(II2);

S1. Enter the circular array after grayscale assignment in step 3;

In this embodiment, setting concentric and equidistant ring widths and adjustable arc angles of pixel blocks can make this method suitable for image positioning of different fundus lesions while ensuring the positioning speed, and improve the flexible operation of this method. At the same time, preliminary angle adjustment of the fundus image can reduce errors in the calculation process and improve the accuracy of the method.

The above are only preferred embodiments of the present invention. It should be noted that those of ordinary skill in the art can also make several improvements and modifications without departing from the technical principles of the present invention. These improvements and modifications It should also be regarded as the protection scope of the present invention.

Claims

A fundus image quality optimization method based on deep learning, which is characterized by including the following steps:

Step 1. Obtain the fundus image and perform sharpening and grayscale processing on the fundus image;

Step 2: Determine the center of the circle on the grayscale processed image in step 1, and segment concentric rings based on the determined center of the circle to obtain several groups of concentric and equidistant rings; and divide each group of concentric rings evenly along the radial direction. into several groups of pixel blocks;

Step 3. Calculate the average grayscale value of each group of pixel blocks in step 2, and perform grayscale assignment on this pixel block to obtain a circular array;

Step 4: Perform convolution calculation on the circular array in step 3, and obtain the corresponding lesion area image based on the convolution calculation result.
A fundus image quality optimization method based on deep learning as claimed in claim 1, characterized in that in step 2, the circle center of the grayscale processed fundus image is confirmed through MATLAB, and the circle center positioning program includes the following steps:

B＝imread(‘fundus image’); % read the original image

A=im2bw(B);% binarization

[x,y]=find(A==0);%Coordinate set of edge pixels in the fundus image

center_x=min(x)+(max(x)-min(x))/2;

center_y=min(y)+(max(y)-min(y))/2;

center＝[center_x,center_y];% This is the center coordinate of the fundus image

Among them, fundus image formats include BMP, GIF, HDF, JPEG, PCX, PNG, TIFF and XWD.
A fundus image quality optimization method based on deep learning as claimed in claim 1, wherein in step 3, the average gray value of each group of pixel blocks is calculated through MATLAB, and the calculation program includes the following steps:

II1=imread(‘pixel block’); % read image

II2=imread(‘pixel block’); % read image

I1=rgb2gray(II1);

I2=rgb2gray(II2);

startX＝350; endX＝400; % Set the starting coordinate and end coordinate of the abscissa of the pixel block

startY＝300; endY＝350; % Set the starting coordinate and end coordinate of the vertical coordinate of the pixel block

aver1=mean(mean(I1(startY:endY,startX:endX)))% select the average gray value of the pixel block

aver2＝mean(mean(I2(startY:endY,startX:endX)))%90 degree selected pixel block average gray value

p1=(aver1-aver2)/(aver1+aver2)% This is the average gray value of the pixel block.
A fundus image quality optimization method based on deep learning as claimed in claim 1, characterized in that the convolution calculation in step 4 includes the following steps:

S1. Enter the circular array after grayscale assignment in step 3;

S2. Use the preliminary convolution kernel to traverse and calculate the circular array in S1 to obtain the primary convolution result;

S3. Map and compare the primary convolution results in S2 with the pre-stored primary convolution results of the normal fundus image one by one, and extract the circular array corresponding to the part where the absolute value of the difference in the primary convolution result is greater than the preset threshold. Pixel blocks, i.e. abnormal blocks in fundus images;

S4. Use the positioning convolution kernel to traverse and calculate the abnormal blocks extracted in S3 to obtain the secondary convolution result;

S5. Map and compare the lesion convolution results in S4 and the pre-stored secondary convolution results of the normal fundus image one by one, and extract the circles corresponding to the parts of the secondary convolution results where the absolute value of the difference is greater than the preset threshold. The array area is used to obtain the image of the lesion area in the fundus image.
A fundus image quality optimization method based on deep learning as claimed in claim 4, characterized in that the preliminary screening convolution kernel adopts a sector-shaped structure with the same radius as the circular array, and the preliminary screening convolution kernel performs traversal calculations. When , make the center of the circle of the initial screening convolution kernel coincide with the center of the circular array, rotate clockwise, and the angle of each rotation is the arc angle corresponding to the pixel block. A set of convolution results can be obtained with each rotation until the The circular array is traversed, and the calculated single-ring array is the primary convolution result.
A fundus image quality optimization method based on deep learning as claimed in claim 5, characterized in that the positioning convolution kernel adopts a partial ring structure with the same radius as the circular array, and the positioning convolution kernel performs traversal calculations. When , make the center of the circle positioning the convolution kernel coincide with the center of the circle where the abnormal block is located, rotate clockwise, and convolve the pixel blocks in the same annulus in the abnormal block one by one until all the annulus in the same abnormal block are calculated When completed, the second-level convolution result is obtained.
A fundus image quality optimization method based on deep learning as claimed in claim 1, characterized in that the annular band width in step 2 can be adjusted according to different fundus lesions. The smaller the lesion range, the narrower the annular band width.
A fundus image quality optimization method based on deep learning as claimed in claim 1, characterized in that the fundus image template is preset according to the starting position relative to the optic nerve head during convolution calculation, and in step 2, the fundus image template is preset. The image template adjusts the angle of the fundus image sample until the sample optic head coincides with the template optic head.