CN114612448A - Fundus blood vessel segmentation method based on rapid label extraction and self-adaptive topology enhancement - Google Patents

Fundus blood vessel segmentation method based on rapid label extraction and self-adaptive topology enhancement Download PDF

Info

Publication number
CN114612448A
CN114612448A CN202210265279.0A CN202210265279A CN114612448A CN 114612448 A CN114612448 A CN 114612448A CN 202210265279 A CN202210265279 A CN 202210265279A CN 114612448 A CN114612448 A CN 114612448A
Authority
CN
China
Prior art keywords
blood vessel
image
skeleton
label
fundus
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210265279.0A
Other languages
Chinese (zh)
Inventor
李锋
石亦恒
刘丽
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Donghua University
Original Assignee
Donghua University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Donghua University filed Critical Donghua University
Priority to CN202210265279.0A priority Critical patent/CN114612448A/en
Publication of CN114612448A publication Critical patent/CN114612448A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30041—Eye; Retina; Ophthalmic
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30101—Blood vessel; Artery; Vein; Vascular

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Eye Examination Apparatus (AREA)

Abstract

The invention discloses an eyeground blood vessel segmentation method based on rapid label extraction and self-adaptive topology enhancement, which can realize automatic extraction of blood vessels in an eyeground image and does not depend on manual labels of experts in a model training process. The realization process is as follows: (1) a response map is acquired with an optimal directional gradient flux filter on the fundus image. (2) And (5) segmenting the main blood vessel on the response map by adopting an adaptive threshold method. (3) The local maxima are used to trace out the vascular skeleton on the response map to complement the small vessels. (4) Local information around the blood vessel skeleton is examined to eliminate false positives at the optic disc and pathological tissue margins. (5) And taking the result of the steps as a label, and training a deep learning model by combining an adaptive topology enhancement loss function. The method automatically acquires the training labels, eliminates the requirement of the deep learning algorithm on the artificial labels, and increases the accuracy of the deep learning algorithm through the self-adaptive topology enhancement loss function.

Description

Fundus blood vessel segmentation method based on rapid label extraction and self-adaptive topology enhancement
Technical Field
The invention belongs to the field of medical image processing, and particularly relates to automatic extraction of an eyeground blood vessel label to train an eyeground blood vessel deep learning segmentation model and realize automatic segmentation of an eyeground blood vessel.
Background
A large number of researches show that fundus images obtained through non-invasive examination can provide important basis for diagnosis of diseases such as diabetic retinopathy, cardiovascular diseases, hypertension and the like. The segmentation of retinal blood vessels plays a crucial role in clinical diagnosis for the analysis of subsequent vascular properties (such as tortuosity and width). However, manually segmenting blood vessels from fundus images is heavily dependent on the experience of the physician, and is a very time-consuming task. Therefore, a high-precision and high-speed automatic blood vessel segmentation method is urgently needed.
The fundus image blood vessel segmentation algorithm is always concerned by scholars at home and abroad, and a new method is always provided. Generally, existing segmentation algorithms for fundus blood vessels can be divided into three major categories: (1) segmentation methods based on conventional image processing generally perform segmentation according to morphological features of blood vessels, such as a matched filtering method, a blood vessel tracking method, a morphological method and a multi-scale method. (2) The segmentation method based on machine learning does not require a rule to be defined in advance, but discriminates between blood vessels and non-blood vessels by learning. (3) The segmentation method based on deep learning is a branch of machine learning method, and the segmentation of blood vessels is completed by training a specific deep network model through data.
The segmentation methods are further classified into a supervised method and an unsupervised method, and of the existing deep learning fundus segmentation methods, almost all of them are supervised methods. These supervised methods imply the need for a large number of fundus pictures, manually annotated by medical experts, to support the training of deep learning models. And a great deal of manpower and material resource cost is consumed for carrying out pixel-by-pixel class marking on the picture. In order to overcome this problem, some unsupervised deep learning methods are proposed. However, the unsupervised method is generally inferior to the supervised learning method in segmentation effect.
Considering that blood vessels are a linear topological structure, some studies propose topologically enhanced deep learning methods. These topological enhancement methods have a certain enhancement effect on the topological consistency of the blood vessel segmentation result, but do not distinguish the difference between the thick blood vessel and the thin blood vessel.
Disclosure of Invention
In order to solve the strong dependence of a supervised deep learning segmentation method on artificial labels, the invention provides a deep learning fundus blood vessel segmentation algorithm based on rapid label extraction and self-adaptive topology enhancement, and training labels are automatically generated by utilizing the steps of optimal directional gradient flux filtering, blood vessel skeleton tracking and the like, so that the consumption of manpower and material resources for artificial labeling is greatly reduced. In addition, the difference between the thick blood vessel and the thin blood vessel is considered, and a loss function with enhanced self-adaptive topology is provided to improve the blood vessel continuity and the sensitivity to the thin blood vessel of the deep learning model blood vessel segmentation result.
In order to achieve the above object, the method of the present invention comprises the steps of:
step 1: performing optimal directional gradient flux filtering on the training set picture to obtain an optimal directional gradient flux response, wherein the optimal directional gradient flux response comprises an optimal blood vessel metric value M of each pixel point and a corresponding blood vessel direction;
step 2: using self-adaptive threshold method on the blood vessel metric value in the optimal directional gradient flux response to divide the image into a main blood vessel structure chart;
and step 3: searching blood vessels on the blood vessel measurement value by using a local maximum algorithm to obtain a blood vessel skeleton graph s;
and 4, step 4: counting the average gray value difference of neighborhoods on two sides of each blood vessel skeleton s, and deleting the blood vessel skeleton with large difference to eliminate false positive at the edge of the optic disc and the pathological tissue;
and 5: using the blood vessel skeleton as constraint, deleting false positive in the main blood vessel structure diagram to obtain a coarse blood vessel label gthick；
Step 6: supplementing the missing small blood vessel part in the main blood vessel diagram by using a blood vessel skeleton, carrying out AND operation with the main blood vessel structure diagram, and carrying out a morphological algorithm (opening and closing operation) to perfect the image to obtain a fine binary blood vessel label g; at the same time, a thin blood vessel label g is obtainedthinAnd thick vessel label gthick；
And 7: introducing the fundus picture and the corresponding fine binary vessel image g as training set labels into a deep learning model together, and training the deep learning model specially used for fundus image vessel segmentation by matching with a self-adaptive topology enhancement loss function;
and 8: and storing the learned model, inputting the picture into the model and outputting a segmentation result when a new fundus image blood vessel segmentation task needs to be processed.
Further, in step 1, the optimal directional gradient flux response is a symmetric matrix Q (x, r), and two eigenvalues λ about the position x can be obtained by similar diagonalization1(x,r),λ2(x, r) and corresponding feature vector ω1(x,r),ω2(x, r); then the optimal directional gradient flux response can be decomposed in the following way:
Q(x,r)＝λ1(x,r)ω1(x,r)ω1 T(x,r)+λ2(x,r)ω2(x,r)ω2 T(x,r)
let lambda1≤λ2Then for one position x of the training set picture, the feature vector ω1Representing the vascular normal direction, ω2A vessel metric value m (x) representing the vessel direction, position x, is calculated according to the following formula:
Figure BDA0003552368720000021
wherein R isscaleIs a series of different scales for multi-scale detection and r is the corresponding scale.
Further, in step 2, the influence of the non-uniform light intensity and contrast on the segmentation is avoided by an adaptive threshold method. The adaptive thresholding method calculates the average value in a window of pixels around each pixel, and calculates the appropriate threshold for each pixel given the vessel foreground proportion. And then segmenting by a threshold method to segment the blood vessel measurement into a binary image, wherein the foreground is a main blood vessel image.
Further, in step 3, each pixel point in the image is checked, and for one pixel point x, the direction omega of the blood vessel is the direction of the pixel point x2In the orthogonal direction of
Figure BDA0003552368720000031
Above, take two points a, B at a distance ζ from point x:
Figure BDA0003552368720000032
Figure BDA0003552368720000033
searching local maximum values on line segments of the connection lines A and B; the local maximum point is set to 1, otherwise to 0. Traversing each pixel in the image to obtain a local maximum value image; removing the intersection points in the local maximum map in the 8-neighborhood to obtain the local maximum map without the intersection points; and finally, only remaining skeletons with the skeleton length larger than a given threshold value 20 in the local maximum value image without the intersection points to obtain the blood vessel skeleton image without background noise.
Further, in step 4, firstly, the neighborhoods on both sides of each skeleton are intercepted. Specifically, for a single blood vessel skeleton, a neighborhood with the radius of l and each skeleton pixel as the center is taken, and a union of all the neighborhood pixels is taken to obtain the neighborhood of the skeleton. The neighborhood of the framework can be divided into two parts by the framework curve and the extension lines of two end points of the framework curve, so that the neighborhoods on two sides of the framework can be obtained. Then, the average gray values of two sides of each blood vessel framework are calculated by counting the gray values of all pixels in the neighborhood, summing the gray values and dividing the sum by the number of the pixels in the neighborhood. And finally, deleting the skeletons with the average gray value difference of two sides exceeding a given threshold value of 0.3, so that the false positive removal of the blood vessel skeletons is completed.
Further, in step 5, based on the skeleton, the main vessel map pixels farther away from the skeleton are excluded, which is equivalent to only remaining the vessel pixels in the neighborhood of the skeleton in the main vessel map, and the coarse vessel label g is obtainedthick。
Further, in step 6, the blood vessel skeleton is expanded by two pixel widths and then is labeled with a thick blood vessel label gthickPerforming AND operation, and opening the merged imageEliminating fine noise in the image (firstly corroding the image and then expanding), and then eliminating holes in the image through closing operation (firstly expanding the image and then corroding), and smoothing the edges of blood vessels; wherein the thin blood vessel label gthin＝g-gthick。
Further, in step 7, the fundus image is subjected to model prediction to obtain a prediction result y, and the difference between y and the label g is calculated through the proposed loss function L. The loss function takes into account not only the correctness of each pixel, but also the differences between the thick and thin vessels, the weights of which are emphasized by the parameters α and β, respectively:
Figure BDA0003552368720000034
where y is the predicted value, g is the tag value,
Figure BDA0003552368720000035
is the hadamard product. L is a radical of an alcoholbceIs a discrete cross entropy loss function:
Figure BDA0003552368720000041
where N represents the total number of pixels, giAnd yiRespectively representing the label and the predicted value of the ith pixel.
The invention has the beneficial effects that: the blood vessels in the fundus image are automatically extracted as labels by using an optimal directional gradient flux-based method to train a deep learning model, and meanwhile, the influence of pathological regions, low-quality imaging regions and optic disc edges in the fundus image on the blood vessel labels is eliminated. The deep learning model is trained by the automatically extracted labels, so that the dependence of the deep learning model on the artificial labels is greatly improved, and manpower and material resources are saved. In addition, the adaptive topology enhancement loss function provided by the method can effectively improve the vessel continuity and the sensitivity to the tiny vessels in the segmentation result.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Fig. 2a is a fundus image.
Fig. 2b is a thick and thin vessel label map automatically generated in conjunction with fig. 2 a.
FIG. 3a is another fundus image
Fig. 3b is a schematic diagram of the result of vessel segmentation performed on fig. 3a by combining with a deep learning model.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
As shown in fig. 1, the deep learning fundus blood vessel segmentation algorithm based on fast tag extraction and adaptive topology enhancement comprises the following steps:
step 1: performing optimal directional gradient flux filtering on the training set picture to obtain an optimal directional gradient flux response, wherein the optimal directional gradient flux response comprises an optimal blood vessel metric value M of each pixel point and a corresponding blood vessel direction;
step 2: using an adaptive threshold method on the blood vessel metric value in the optimal directional gradient flux response to segment the image into a main blood vessel structure chart;
and step 3: searching blood vessels on the blood vessel measurement value by using a local maximum algorithm to obtain a blood vessel skeleton graph s;
and 4, step 4: counting the average gray value difference of neighborhoods on two sides of each blood vessel skeleton s, and deleting the blood vessel skeleton with large difference to eliminate false positive at the edge of the optic disc and the pathological tissue;
and 5: using the blood vessel skeleton as constraint, deleting false positive in the main blood vessel structure diagram to obtain a coarse blood vessel label gthick；
Step 6: supplementing the missing small blood vessel part in the main blood vessel diagram by using a blood vessel skeleton, carrying out AND operation with the main blood vessel structure diagram, and carrying out a morphological algorithm (opening and closing operation) to perfect the image to obtain a fine binary blood vessel label g; at the same time, a thin blood vessel label g is obtainedthinAnd thick vessel label gthick(ii) a As shown, FIG. 2b is the vessel label of FIG. 2a, wherein the thicker lines represent the thicker vessel labels and the thinner lines represent the thinner vessel labelsA blood vessel label.
And 7: introducing the fundus picture and the corresponding fine binary vessel image g as training set labels into a deep learning model together, and training the deep learning model specially used for fundus image vessel segmentation by matching with a self-adaptive topology enhancement loss function;
and 8: and storing the learned model, inputting the picture into the model and outputting a segmentation result when a new fundus image blood vessel segmentation task needs to be processed. The result of prediction after the fundus image of fig. 3a is input to the model is fig. 3 b.
Further, in step 1, the optimal directional gradient flux response is a symmetric matrix Q (x, r), and two eigenvalues λ about the position x can be obtained by similar diagonalization1(x,r),λ2(x, r) and corresponding feature vector ω1(x,r),ω2(x, r); then the optimal directional gradient flux response can be decomposed in the following way:
Q(x,r)＝λ1(x,r)ω1(x,r)ω1 T(x,r)+λ2(x,r)ω2(x,r)ω2 T(x,r)
let lambda1≤λ2Then for one position x of the training set picture, the feature vector ω1Representing the vascular normal direction, ω2A vessel metric value m (x) representing the vessel direction, position x, is calculated according to the following formula:
Figure BDA0003552368720000051
wherein R isscaleIs a series of different scales for multi-scale detection and r is the corresponding scale.
Example 1:
in step 1, a multi-scale detection range R is takenscale＝[1,8]。
Further, in step 2, the influence of the non-uniform light intensity and contrast on the segmentation is avoided by an adaptive threshold method. The adaptive thresholding method calculates the average value in a window of pixels around each pixel, and calculates the appropriate threshold for each pixel by a given vessel foreground proportion. And then segmenting the blood vessel measurement into a binary image by a threshold value method, wherein the foreground is a main blood vessel image.
Further, in step 3, each pixel point in the image is checked, and for one pixel point x, the direction omega of the blood vessel is the direction of the pixel point x2In the orthogonal direction of
Figure BDA0003552368720000052
Above, take two points a, B at a distance ζ from point x:
Figure BDA0003552368720000053
Figure BDA0003552368720000054
searching local maximum values on line segments of the connection lines A and B; the local maximum point is set to 1, otherwise to 0. Traversing each pixel in the image to obtain a local maximum value image; removing the intersection points in the local maximum map in the 8-neighborhood to obtain the local maximum map without the intersection points; and finally, only remaining skeletons with the skeleton length larger than a given threshold value 20 in the local maximum value image without the intersection points to obtain the blood vessel skeleton image without background noise.
Example 2:
in step 3, the search range ζ is taken to be 10.
Further, in step 4, firstly, the neighborhoods on both sides of each skeleton are intercepted. Specifically, for a single blood vessel framework, a neighborhood of the framework is obtained by taking each framework pixel as a center and taking a neighborhood of which the radius is l and taking a union of neighborhoods of all pixels. The neighborhood of the framework can be divided into two parts by the framework curve and the extension lines of two end points of the framework curve, so that the neighborhoods on two sides of the framework can be obtained. Then, the average gray values of two sides of each blood vessel framework are calculated by counting the gray values of all pixels in the neighborhood, summing the gray values and dividing the sum by the number of the pixels in the neighborhood. And finally, deleting the skeletons with the average gray value difference of two sides exceeding a given threshold value of 0.3, so that the false positive removal of the blood vessel skeletons is completed.
Further, in step 5, based on the skeleton, the main vessel map pixels farther away from the skeleton are excluded, which is equivalent to only remaining the vessel pixels in the neighborhood of the skeleton in the main vessel map, so as to obtain the coarse vessel label gthick。
Further, in step 6, the blood vessel skeleton is expanded by two pixel widths and then is labeled with a thick blood vessel label gthickPerforming AND operation, eliminating fine noise in the combined image through open operation (firstly corroding the image and then expanding), and eliminating the cavity in the image through close operation (firstly expanding the image and then corroding), and smoothing the edge of the blood vessel; wherein the thin blood vessel label gthin＝g-gthick。
Further, in step 7, the fundus image is subjected to model prediction to obtain a prediction result y, and the difference between y and the label g is calculated through the proposed loss function L. The loss function takes into account not only the correctness of each pixel, but also the differences between the thick and thin vessels, the weights of which are emphasized by the parameters α and β, respectively:
Figure BDA0003552368720000061
where y is the predicted value, g is the tag value,
Figure BDA0003552368720000062
is the hadamard product. L isbceIs a discrete cross entropy loss function:
Figure BDA0003552368720000063
where N represents the total number of pixels, giAnd yiRespectively representing the label and the predicted value of the ith pixel.
Example 4:
in step 7, the training set, the verification set and the test set are divided according to the proportion of 10:1: 10.
Finally, it should be noted that: the described embodiments are only some embodiments of the invention, not all embodiments. Based on the embodiments of the present invention, those skilled in the art may modify the technical solutions described in the foregoing embodiments or substitute equivalent technical features to obtain other embodiments, which belong to the protection scope of the present invention.

Claims (7)

1. An eyeground blood vessel segmentation method based on rapid label extraction and self-adaptive topology enhancement is characterized by comprising the following steps:
step 1: performing optimal directional gradient flux filtering on the training set picture to obtain an optimal directional gradient flux response, wherein the optimal directional gradient flux response comprises an optimal blood vessel metric value M of each pixel point and a corresponding blood vessel direction;
step 2: using an adaptive threshold method on the blood vessel metric value in the optimal directional gradient flux response to segment the image into a main blood vessel structure chart;
and step 3: searching blood vessels on the blood vessel measurement value by using a local maximum algorithm to obtain a blood vessel skeleton s;
and 4, step 4: counting the average gray value difference of neighborhoods on two sides of each blood vessel skeleton s, and deleting the blood vessel skeleton with large difference to eliminate false positive at the edge of the optic disc and the pathological tissue;
and 5: using the blood vessel skeleton as constraint, deleting false positive in the main blood vessel structure diagram to obtain a coarse blood vessel label gthick；
Step 6: supplementing the missing small blood vessel part in the main blood vessel map by using a blood vessel skeleton, carrying out AND operation with the main blood vessel structure map, and carrying out a morphological algorithm to perfect the image to obtain a fine binary blood vessel label g; at the same time, a thin blood vessel label g is obtainedthinAnd thick vessel label gthick；
And 7: introducing the fundus picture and the corresponding fine binary vessel image g as training set labels into a deep learning model together, and training the deep learning model specially used for fundus image vessel segmentation by matching with a self-adaptive topology enhancement loss function;
and 8: and (4) storing the deep learning model obtained in the step (7), inputting the picture into the model and outputting a segmentation result when a new fundus image blood vessel segmentation task needs to be processed.
2. An fundus blood vessel segmentation method according to claim 1, wherein in said step 1: the method for obtaining the optimal directional gradient flux response by performing the optimal directional gradient flux filtering on the training set picture comprises the following steps:
the optimal directional gradient flux response is a symmetric matrix Q (x, r), whose two eigenvalues λ about position x can be obtained by similar diagonalization1(x,r),λ2(x, r) and corresponding feature vector ω1(x,r),ω2(x, r); then the optimal directional gradient flux response can be decomposed in the following way:
Q(x,r)＝λ1(x,r)ω1(x,r)ω1T(x,r)+λ2(x,r)ω2(x,r)ω2 T(x,r)
let lambda1≤λ2Then for one position x of the training set picture, the feature vector ω1Represents the vascular normal direction, ω2A vessel metric value m (x) representing the vessel direction, position x, is calculated according to the following formula:
Figure FDA0003552368710000011
wherein R isscaleIs a series of different scales for multi-scale detection and r is the corresponding scale.
3. An fundus blood vessel segmentation method according to claim 1, wherein in said step 2: using an adaptive threshold method on the blood vessel metric value in the optimal directional gradient flux response to segment the image into a main blood vessel structure chart; the method comprises the following steps:
the influence of nonuniform light intensity and contrast on segmentation is avoided by an adaptive threshold method; calculating the average value in a pixel window around each pixel point by using an adaptive threshold method, and calculating a proper threshold value for each pixel according to a given blood vessel foreground proportion; and then segmenting by a threshold method to segment the blood vessel measurement into a binary image, wherein the foreground is a main blood vessel image.
4. An fundus blood vessel segmentation method according to claim 1, wherein in said step 3: searching blood vessels on the blood vessel measurement value by using a local maximum algorithm to obtain a blood vessel skeleton graph s; the method comprises the following steps:
checking each pixel point in the training set picture, and for one pixel point x, in the blood vessel direction omega2In the orthogonal direction of
Figure FDA0003552368710000021
Taking two points A, B at a distance ζ from the point x:
Figure FDA0003552368710000022
Figure FDA0003552368710000023
searching local maximum values on line segments of the connection lines A and B; setting the local maximum value point as 1, otherwise, setting the local maximum value point as 0; traversing each pixel in the training set picture to obtain a local maximum value picture; removing the intersection points in the local maximum map in the 8-neighborhood to obtain the local maximum map without the intersection points; and finally, only remaining skeletons with the skeleton length larger than a given threshold value 20 in the local maximum value image without the intersection points to obtain the blood vessel skeleton image without background noise.
5. An fundus blood vessel segmentation method according to claim 1, wherein in said step 4: counting the average gray value difference of neighborhoods on two sides of each blood vessel skeleton s, and deleting the blood vessel skeleton with large difference to eliminate false positive at the edge of the optic disc and pathological tissues; the method comprises the following steps:
firstly, intercepting neighborhoods on two sides of each framework, particularly a single blood vessel framework, wherein the neighborhoods with the pixel of each framework as the center and the radius of l are taken, and a union set of all the pixel neighborhoods is taken to obtain the neighborhoods of the frameworks; the neighborhood of the framework can be divided into two parts by the framework curve and the extension lines of two end points of the framework curve, so that the neighborhoods on two sides of the framework can be obtained; then, the average gray values of two sides of each blood vessel skeleton are calculated by counting the gray values of all pixels in the neighborhood, summing the gray values and dividing the sum by the number of the pixels in the neighborhood. And finally, deleting the skeletons with the average gray value difference of two sides exceeding a given threshold value of 0.3, so that the false positive removal of the blood vessel skeletons is completed.
6. An fundus blood vessel segmentation method according to claim 1 wherein in said step 6: supplementing the missing small blood vessel part in the main blood vessel map by using a blood vessel skeleton, carrying out AND operation with the main blood vessel structure map, and carrying out a morphological algorithm to perfect the image to obtain a fine binary blood vessel label g; at the same time, a thin blood vessel label g is obtainedthinAnd thick vessel label gthick(ii) a The method comprises the following steps:
expanding the blood vessel skeleton by two pixel widths and then matching the expanded blood vessel skeleton with a thick blood vessel label gthickPerforming AND operation, and eliminating tiny noise in the combined image through an opening operation, wherein the opening operation is to corrode and expand the image; then eliminating the cavity in the image through a closing operation, wherein the closing operation is to expand the image and then corrode the image; at the same time, the vessel edges are smoothed; wherein the thin blood vessel label gthin＝g-gthick。
7. An fundus blood vessel segmentation method according to claim 1, wherein in said step 7: the fundus image and the corresponding fine binary blood vessel label g are used as training set labels and are transmitted into a deep learning model together, and the deep learning model special for fundus image blood vessel segmentation is trained by matching with a self-adaptive topology enhancement loss function; the method comprises the following steps:
obtaining a prediction result y by model prediction of the fundus image, and calculating the difference between the y and a fine binary blood vessel label g through a proposed loss function L; the loss function takes into account not only the correctness of each pixel, but also the differences between the thick and thin vessels, the weights of which are emphasized by the parameters α and β, respectively:
Figure FDA0003552368710000031
where y is the predicted value, g is the tag value,
Figure FDA0003552368710000032
is the product of Hadamard, LbceIs a discrete cross entropy loss function:
Figure FDA0003552368710000033
where N represents the total number of pixels, giAnd yiRespectively representing the label and the predicted value of the ith pixel.
CN202210265279.0A 2022-03-17 2022-03-17 Fundus blood vessel segmentation method based on rapid label extraction and self-adaptive topology enhancement Pending CN114612448A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210265279.0A CN114612448A (en) 2022-03-17 2022-03-17 Fundus blood vessel segmentation method based on rapid label extraction and self-adaptive topology enhancement

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210265279.0A CN114612448A (en) 2022-03-17 2022-03-17 Fundus blood vessel segmentation method based on rapid label extraction and self-adaptive topology enhancement

Publications (1)

Publication Number Publication Date
CN114612448A true CN114612448A (en) 2022-06-10

Family

ID=81865123

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210265279.0A Pending CN114612448A (en) 2022-03-17 2022-03-17 Fundus blood vessel segmentation method based on rapid label extraction and self-adaptive topology enhancement

Country Status (1)

Country Link
CN (1) CN114612448A (en)

Similar Documents

Publication Publication Date Title
CN107644420B (en) Blood vessel image segmentation method based on centerline extraction and nuclear magnetic resonance imaging system
WO2022063199A1 (en) Pulmonary nodule automatic detection method, apparatus and computer system
CN112716446B (en) Method and system for measuring pathological change characteristics of hypertensive retinopathy
WO2021203795A1 (en) Pancreas ct automatic segmentation method based on saliency dense connection expansion convolutional network
Wen et al. A novel statistical cerebrovascular segmentation algorithm with particle swarm optimization
CN111598875A (en) Method, system and device for building thyroid nodule automatic detection model
CN111899244B (en) Image segmentation method, network model training method, device and electronic equipment
Ren et al. An unsupervised semi-automated pulmonary nodule segmentation method based on enhanced region growing
Hiremath et al. Automatic detection of follicles in ultrasound images of ovaries using edge based method
EP3975117A1 (en) Image segmentation method and apparatus, and training method and apparatus for image segmentation model
CN109087310B (en) Meibomian gland texture region segmentation method and system, storage medium and intelligent terminal
Rampun et al. Automated 2d fetal brain segmentation of mr images using a deep u-net
CN113537408A (en) Ultrasonic image processing method, device and equipment and storage medium
Wen et al. Review of research on the instance segmentation of cell images
CN116188485A (en) Image processing method, device, computer equipment and storage medium
Mohan et al. Exudate localization in retinal fundus images using modified speeded up robust features algorithm
CN113516644B (en) Intravascular optical coherence tomography branch blood vessel identification method
CN117373070B (en) Method and device for labeling blood vessel segments, electronic equipment and storage medium
CN111401102A (en) Deep learning model training method and device, electronic equipment and storage medium
CN116452579B (en) Chest radiography image-based pulmonary artery high pressure intelligent assessment method and system
CN114612448A (en) Fundus blood vessel segmentation method based on rapid label extraction and self-adaptive topology enhancement
CN115841472A (en) Method, device, equipment and storage medium for identifying high-density characteristics of middle cerebral artery
CN114419061A (en) Method and system for segmenting pulmonary artery and vein blood vessels
Chen et al. Centerline constrained minimal path propagation for vessel extraction
Ashame et al. Abnormality Detection in Eye Fundus Retina

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination