CN111415350B - Colposcope image identification method for detecting cervical lesions - Google Patents

Colposcope image identification method for detecting cervical lesions Download PDF

Info

Publication number
CN111415350B
CN111415350B CN202010230980.XA CN202010230980A CN111415350B CN 111415350 B CN111415350 B CN 111415350B CN 202010230980 A CN202010230980 A CN 202010230980A CN 111415350 B CN111415350 B CN 111415350B
Authority
CN
China
Prior art keywords
attribute
sample
image
pixel
pixel block
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.)
Active
Application number
CN202010230980.XA
Other languages
Chinese (zh)
Other versions
CN111415350A (en
Inventor
孙蓬明
柳培忠
薛慧丰
顾培婷
董滨华
杜永兆
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.)
Fuzhou Zhongsheng Zhanrui Technology Co ltd
Original Assignee
Fujian Maternal And Child Care Service Centre
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 Fujian Maternal And Child Care Service Centre filed Critical Fujian Maternal And Child Care Service Centre
Priority to CN202010230980.XA priority Critical patent/CN111415350B/en
Publication of CN111415350A publication Critical patent/CN111415350A/en
Application granted granted Critical
Publication of CN111415350B publication Critical patent/CN111415350B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G06T3/4053—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/20—Image enhancement or restoration using local operators
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • 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
    • G06T2207/20016—Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
    • 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
    • G06T2207/20021—Dividing image into blocks, subimages or windows
    • 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
    • G06T2207/20024—Filtering details
    • 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
    • G06T2207/30096—Tumor; Lesion

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)

Abstract

The application relates to the technical field of image recognition, in particular to a colposcopy image recognition method for detecting cervical lesions. The method comprises the following steps: step 1, collecting a sample image; step 2, carrying out data preprocessing on the sample image; step 3, establishing a sample characteristic space according to the preprocessed sample image; step 4, constructing an optimal attribute predictor according to the sample feature space; and 5, classifying the image samples to be detected according to the optimal attribute predictor. According to the technical scheme, the attribute learning method is regularized by the class orthograph, the accuracy and the recall rate of attribute prediction are improved, and meanwhile, the learned attribute predictor can be used for rapidly and accurately identifying and classifying the colposcope image to be detected, so that a good auxiliary detection and diagnosis effect is achieved.

Description

Colposcope image identification method for detecting cervical lesions
Technical Field
The application relates to the technical field of image recognition, in particular to a colposcopy image recognition method for detecting cervical lesions.
Background
Cervical cancer in China is one of the most common gynecological malignant tumors, the incidence rate of the cervical cancer is second to breast cancer, and the cervical cancer is high in the second place of female malignant tumors, so that the physical and psychological health of female people is seriously influenced. Worldwide, there are 528,000 new cases of cervical cancer annually, with 266,000 deaths. With the increasing number of the cervical cancer and the number of the death people year by year and the younger occurrence of the disease population, the prevention, treatment and diagnosis of the cervical lesion/the cervical cancer are very important. The overall process for cervical cancer screening is complex and costly, often resulting in the inability to deploy more advanced cervical cancer screening technologies in resource-poor areas, and thus the incidence and mortality of cervical cancer remains high in underdeveloped areas. At present, colposcopy has become one of the important steps for clinically screening CIN and early cervical cancer, and directly influences the diagnosis and treatment scheme of patients.
With the rapid development of artificial intelligence, the related technology has been gradually applied to the identification and analysis process of cervical cancer clinical images and colposcopic images for auxiliary diagnosis and improvement of detection efficiency and accuracy. Therefore, how to realize an efficient and accurate image detection method as an auxiliary diagnostic tool for cervical cancer screening aiming at the collected colposcope images is a problem to be solved urgently.
Disclosure of Invention
The invention aims to provide a colposcopic image identification method for detecting cervical lesions, which is characterized in that cervical images are preprocessed through a self-adaptive image enhancement technology, regions of interest are effectively segmented, then characteristic information of cervical lesions is effectively extracted according to a multi-scale pyramid method with consistent phases, an optimal attribute predictor is obtained by utilizing a class positive definite hypergraph regularization attribute learning method, and accurate classification of the images to be detected is realized.
In order to solve the above technical problem, the present invention provides a colposcopic image recognition method for detecting cervical lesions, comprising the following steps:
step 1, collecting a sample image;
step 2, carrying out data preprocessing on the sample image;
step 3, establishing a sample characteristic space according to the preprocessed sample image;
step 4, constructing an optimal attribute predictor according to the sample feature space;
and 5, classifying the image samples to be detected according to the optimal attribute predictor.
Further, the step 2 of performing data preprocessing on the sample image includes the following steps:
step 21, carrying out self-adaptive enhancement filtering processing on the sample image;
and step 22, carrying out region segmentation on the sample image after the adaptive enhancement filtering processing.
Further, the step 21 of performing adaptive enhancement filtering processing on the sample image includes the following steps:
step 211, carrying out equalization processing on the sample image by using a variance equalization method to enable the sample image to reach the required average gray scale and variance;
step 212, dividing the sample image into pixel blocks which are w in width and height and are overlapped with each other, wherein two adjacent pixel blocks are separated by one pixel in the horizontal direction, two adjacent pixel blocks are separated by one pixel in the vertical direction, and w is an odd number which is more than or equal to 3;
step 213, calculating the direction field of each pixel block and performing smoothing processing;
step 214, calculating the reliability of each pixel block direction field, wherein the calculation formula is as follows:
Figure GDA0004054416110000021
wherein R is a set of all pixels in the pixel block, (i, j) is a central pixel of the pixel block, (i ', j') is any pixel of the pixel block except for the central pixel (i, j), n is a total number of pixels in the pixel block, θ (i, j) represents a directional field of the pixel block centered at (i, j), and θ (i ', j') represents a directional field of the pixel block centered at (i ', j');
step 215, for each pixel block, judging whether the reliability χ (i, j) of the pixel block is greater than a threshold T, executing step 216, otherwise, jumping to step 217 to continue executing;
step 216, in a pixel block set which takes a non-central pixel (i ', j') of the pixel block as a central pixel, judging the number of blocks which satisfy the reliability χ (i ', j') less than the threshold T, if the number of blocks is greater than or equal to the threshold Tn, setting the direction field of the pixel block as the average value of the direction fields of the pixel blocks which satisfy the reliability χ (i ', j') less than the threshold T in the pixel block set, otherwise, setting the pixel block as an unrecoverable pixel block;
step 217, for each pixel block except the unrecoverable pixel block, calculating coordinates (x ', y') of all pixels in the pixel block after rotation, wherein the calculation formula is as follows:
x'＝icosθ+jsinθ
y'＝-isinθ+jcosθ
wherein, i, j is coordinate values of pixels in the pixel block on x, y axes; θ is the directional field of the pixel block centered on the pixel;
step 218, for each pixel block except the unrecoverable pixel block, performing convolution operation on the gray values of all the rotated pixel points in the pixel block through a Gaussian function, and limiting the result value after convolution between [0,255] as the gray value of the central pixel of the pixel block;
steps 215 through 218 are repeated until all pixel block processing is completed.
According to the technical scheme, the image is divided into a plurality of overlapped pixel blocks, and the image is enhanced based on the reliability of the pixel block direction field, so that the image quality can be improved, and the image recognition and segmentation effect is enhanced.
Further, the step 22 of performing region segmentation on the sample image after the adaptive enhancement filtering process includes the following steps:
step 221, pre-dividing the sample image subjected to the adaptive enhancement filtering processing by using a mean shift algorithm to obtain a plurality of superpixels;
step 222, constructing a weighted Graph according to the pre-divided superpixels, namely constructing a Graph cuts division model as follows:
Figure GDA0004054416110000031
wherein, alpha represents the segmentation label value, the value of 0 or 1,0 represents the background, 1 represents the foreground, I represents the set of all the super-pixels obtained by pre-segmentation, D i (α) represents the feature similarity of the current superpixel i with respect to the segmentation index α, m and n represent any two different superpixels adjacent to the current superpixel i, S m,n (α) represents penalty values for adjacent superpixels m and n belonging to different index sets, γ being the balance coefficient;
and 223, performing global S-T minimum cut operation on the weighted Graph to minimize the Graph cuts segmentation model to obtain optimal image segmentation.
According to the technical scheme, the sample image is pre-segmented into the super pixels, and then the Graph cuts weighted segmentation model is adopted, so that the segmentation precision and accuracy can be improved.
Further, the feature similarity of the current super-pixel i with respect to the segmentation index α is to use color gaussian distance metric statistics of the region to characterize the super-pixel, and the gaussian distance metric calculation formula is:
Figure GDA0004054416110000032
wherein (C) T For the transpose operation of the matrix, tr () is the trace operation of the matrix, d is the spatial dimension of the Gaussian feature, i.e. the number of channels in the color space, m and n represent any two different superpixels adjacent to the current superpixel i, g m (x) Set of Gaussian parameters [ mu ] representing the correspondence of a superpixel m m ,∑m}，g n (x) Set of gaussian parameters μ representing the correspondence of a super-pixel n n ,∑n}，μ m Represents the mean of the color features in superpixel m, ∑ m represents the covariance of the color features in superpixel m,. Mu. n Represents the mean of the color features in the super-pixel n and Σ n represents the covariance of the color features in the super-pixel n.
Further, the step 3 of establishing a sample feature space according to the preprocessed sample image includes the following steps:
step 31, performing laplacian pyramid decomposition on the preprocessed sample image to obtain a layered image, wherein a decomposition iteration formula is expressed as:
Figure GDA0004054416110000041
Figure GDA0004054416110000042
LP l (b,c)＝L l-1 (b,c)-L l '(b,c)
wherein L is l First layer image of low-pass Gaussian pyramid, LP l Is the L-layer image, L 'of the Laplacian pyramid' l B represents the number of rows of the pyramid ith layer image, c represents the number of columns of the pyramid ith layer image, w (p, q) is a 3 x 3 low pass filter, p ∈ [ -2,2],q∈[-2,2]；
Step 32, respectively performing phase consistent feature extraction on each layer of image in the Laplacian pyramid to obtain a multi-scale pyramid phase consistent image;
step 33, fusing the multi-scale pyramid phase consistent images to obtain a sample characteristic diagram based on phase consistency;
and step 34, performing the same processing on all sample images, and forming a sample feature space by all sample feature maps.
According to the technical scheme, the sample image is subjected to Laplacian pyramid decomposition, phase consistent feature extraction is performed on each layer of image, and finally, the images are fused into a final sample feature map, so that feature extraction is more complete and accurate.
Further, the step 4 of constructing an optimal attribute predictor according to the sample feature space includes the following steps:
step 41, constructing an attribute hypergraph according to the sample feature space and the attribute label of each sample;
step 42, carrying out regularization hypergraph segmentation on the attribute hypergraph, and solving an optimal attribute predictor to ensure that the attribute relation loss of the attribute hypergraph is minimum and the error between a predicted attribute and an attribute label is minimum; the objective function of the attribute predictor is as follows:
Figure GDA0004054416110000043
wherein, the Trace function represents the Trace operation of the matrix, () T Is the transpose operation of matrix, eta is a nonnegative regularization parameter, lambda is a positive parameter, X is the sample feature space, Y is the sample attribute label matrix, B is the projection matrix, L H A Laplace matrix L representing the correspondence of said attribute hypergraph * And representing the corresponding Laplace matrix after the attribute hypergraph is added with the category information.
According to the technical scheme, a mode of the attribute hypergraph is adopted, the multivariate logical relation between the sample characteristic space and the attribute label of the sample can be accurately described, and meanwhile, the classification accuracy of the image to be detected can be realized based on the optimal attribute predictor solved by the regularization hypergraph segmentation method.
Further, the step 5 of classifying the image samples to be detected according to the optimal attribute predictor includes the following steps:
step 51, calculating an attribute predicted value of the image sample to be detected according to a calculation formula:
P k ＝sign(z k T B)
wherein, P k Representing attribute predicted values, z k B represents a projection matrix from the sample characteristic space to an attribute hypergraph embedding space for an image sample to be detected, and a sign () function represents a positive sign and a negative sign of each element of a vector;
step 52, performing normalization processing on the attribute predicted value, wherein a normalization formula is as follows:
Figure GDA0004054416110000051
wherein, the vector r k Expressing the predicted value of each attribute of the image sample to be detected after normalization, wherein rho is a positive parameter of the normalized control scale, P k Representing an attribute prediction value;
step 53, calculating the distance between the cervical lesion classified attribute list template and the predicted value of each normalized attribute of the image sample to be detected, and classifying the image sample to be detected, wherein the calculation formula is as follows:
Figure GDA0004054416110000052
wherein the function O (z) k ) The representation takes the class label with the minimum Euclidean distance, r k Representing the predicted value, t, of each normalized attribute of the image sample to be detected a A classification attribute list template representing cervical lesions.
Different from the prior art, the technical scheme of the invention has the following beneficial effects:
1. the image is divided into a plurality of overlapped pixel blocks, and the image is enhanced based on the reliability of the pixel block direction field, so that the image quality can be improved, and the image recognition and segmentation effect can be enhanced.
2. The sample image is pre-divided into super pixels, and then a Graph cuts weighting division model is adopted, so that the division precision and accuracy can be improved.
3. And performing Laplacian pyramid decomposition on the sample image, performing phase consistent feature extraction on each layer of image, and finally fusing the images into a final sample feature map, so that the feature extraction is more complete and accurate.
4. By adopting the attribute hypergraph mode, the multivariate logical relationship between the sample characteristic space and the attribute labels of the samples can be accurately described, and meanwhile, the classification accuracy of the to-be-detected images can be realized by the optimal attribute predictor solved based on the regularization hypergraph segmentation method.
Drawings
Fig. 1 is a flow chart of the steps of a colposcopic image recognition method for detecting cervical lesions of the present invention.
FIG. 2 is a flow chart of the steps of the present invention for data preprocessing of the sample image.
FIG. 3 is a flow chart of the steps of the present invention for creating a sample feature space from a preprocessed sample image.
FIG. 4 is a flow chart of the steps for constructing an optimal attribute predictor from the sample feature space according to the present invention;
FIG. 5 is a flowchart illustrating the steps of classifying an image sample to be detected according to the optimal attribute predictor.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Fig. 1 is a flow chart of steps of a colposcopic image recognition method for detecting cervical lesions, comprising the following steps:
step 1, collecting a sample image; the sample images referred to herein refer to a number of colposcopic images used to learn and construct training classification models. The specific acquisition mode is as follows: and selecting high-frame-frequency and high-resolution cervical image acquisition equipment, namely a colposcope, acquiring cervical images in real time, and uploading the images to a PC (personal computer) terminal for storage.
Step 2, carrying out data preprocessing on the sample image; fig. 2 is a flowchart of the steps of preprocessing the sample image according to the present invention, including the following steps:
step 21, carrying out self-adaptive enhancement filtering processing on the sample image; the purpose of this step is to enhance the image, remove noise in the image, make the edge sharp, and highlight the features in the image to improve the robustness of feature extraction.
Specifically, the technical scheme performs adaptive enhancement filtering processing on the sample image, and comprises the following steps:
and step 211, carrying out equalization processing on the sample image by adopting a variance equalization method, so that the processed sample image reaches the required average gray scale and variance.
Step 212, dividing the sample image into square pixel blocks which are w in width and height and are overlapped with each other, wherein two adjacent pixel blocks are separated by one pixel in the horizontal direction, two adjacent pixel blocks are separated by one pixel in the vertical direction, and w is an odd number which is more than or equal to 3; in a preferred embodiment w is set to 3, i.e. each pixel block is 3 pixels wide and high.
Step 213, calculating the direction field of each pixel block and performing smoothing processing.
Step 214, calculating the reliability of each pixel block direction field, wherein the calculation formula is as follows:
Figure GDA0004054416110000071
wherein R is a set of all pixels in the pixel block, (i, j) is a central pixel of the pixel block, (i ', j') is any pixel except for the central pixel (i, j) in the pixel block, n is a total number of pixels in the pixel block, θ (i, j) represents a directional field of the pixel block centered at (i, j), and θ (i ', j') represents a directional field of the pixel block centered at (i ', j').
Step 215, regarding each divided pixel block, if the reliability χ (i, j) of the pixel block is greater than the threshold T, determining that the obtained directional field is unreliable, and adjusting the obtained directional field according to the directional field of the peripheral area of the block, and executing step 216, otherwise, indicating that the calculated directional field of the pixel block is reliable, and skipping to step 217 to continue execution; in a preferred embodiment, the threshold may be set as: t = π/8.
Step 216, in a pixel block set taking a non-central pixel (i ', j') of the pixel block as a central pixel, judging the number of blocks satisfying that the reliability χ (i ', j') is smaller than a threshold T, if the number of blocks is larger than or equal to the threshold Tn, setting a direction field of the pixel block as an average value of direction fields of the pixel blocks satisfying that the reliability χ (i ', j') is smaller than the threshold T in the pixel block set, otherwise, setting the pixel block as an unrecoverable pixel block; in a specific embodiment, tn may be set to 4.
Step 217, for each pixel block except the unrecoverable pixel block, calculating coordinates (x ', y') of all pixels in the pixel block after rotation, wherein the calculation formula is as follows:
x'＝icosθ+jsinθ
y'＝-isinθ+jcosθ
wherein, i, j is coordinate values of pixels in the pixel block on x, y axes; θ is the directional field of the pixel block centered on the pixel.
And step 218, performing convolution operation on each pixel block except the unrecoverable pixel block and the gray values of all the rotated pixel points in the pixel block through a Gaussian function, and limiting the result value after convolution to be between [0 and 255] to be used as the gray value of the central pixel of the pixel block.
And repeating the steps 215 to 218 until all pixel blocks in the sample image are processed, so as to obtain the sample image after the adaptive enhancement filtering processing.
According to the technical scheme, the image is divided into a plurality of overlapped pixel blocks, and the image is enhanced based on the reliability of the pixel block direction field, so that the image quality can be improved, and the image recognition and segmentation effect is enhanced.
Step 22, performing region segmentation on the sample image after the adaptive enhancement filtering processing, including the following steps:
step 221, pre-dividing the sample image after the adaptive enhancement filtering processing by using a conventional mean shift algorithm, effectively dividing the enhanced sample image into a plurality of regions, and describing the regions as superpixels.
Step 222, constructing a weighted Graph according to the pre-divided superpixels, namely constructing a Graph cuts division model as follows:
Figure GDA0004054416110000081
wherein, alpha represents the segmentation label value, the value of 0 or 1,0 represents the background, 1 represents the foreground, I represents the set of all the super-pixels obtained by pre-segmentation, D i (α) represents the feature similarity of the current superpixel i with respect to the segmentation index α, m and n represent any two different superpixels adjacent to the current superpixel i, S m,n And (α) represents a penalty value for neighboring superpixels m and n belonging to different index sets, and γ is a balancing coefficient.
In a preferred embodiment, the "feature similarity of the current super-pixel i with respect to the segmentation index α" is to use color gaussian distance metric statistics of the region to characterize the super-pixel, and the gaussian distance metric calculation formula is:
Figure GDA0004054416110000082
wherein, the (A) T For the transpose operation of the matrix, tr () is the trace operation of the matrix, d is the spatial dimension of the gaussian feature, i.e. the number of channels in the color space, and each superpixel i can be represented as G i ＝{μ i ,∑ i In which, mu i Sum Σ i Corresponding to the mean and covariance, respectively, of the color features in a superpixel i, m and n representing any two different superpixels adjacent to the current superpixel i, g m (x) Set of Gaussian parameters [ mu ] representing the correspondence of a superpixel m m ,∑m}，g n (x) Set of gaussian parameters μ representing the correspondence of a super-pixel n n ,∑n}，μ m Represents the mean of the color features in superpixel m, ∑ m represents the covariance of the color features in superpixel m,. Mu. n Represents the mean of the color features in the super-pixel n and Σ n represents the covariance of the color features in the super-pixel n.
And 223, based on the maximum flow or minimum cut theory of the network flow, performing global S-T minimum cut operation on the weighted Graph to minimize the Graph cuts segmentation model, and obtaining optimal image segmentation.
According to the technical scheme, the sample image is pre-segmented into the super pixels, and then the Graph cuts weighted segmentation model is adopted, so that the segmentation precision and accuracy can be improved.
Step 3, establishing a sample characteristic space according to the preprocessed sample image; fig. 3 is a flowchart illustrating steps of establishing a sample feature space according to a preprocessed sample image according to the present invention, including the following steps:
step 31, performing laplacian pyramid decomposition on the preprocessed sample image to obtain a layered image, assuming that the input original sample image is L 0 And using it as the zero-th layer of Gaussian pyramid to the original image L 0 Filtering and interlacing and downsampling are carried out through a Gaussian low-pass filter L to obtain a first layer image L of a low-pass Gaussian pyramid 1 The first layer image L of the low-pass Gaussian pyramid is processed 1 Obtaining L through interpolation expansion filtering carried out by upsampling and band-pass filter H 1 ' and calculating the difference between the image and the original image to obtain the band-pass component, i.e. the first layer LP of Laplacian pyramid 1 The next-stage decomposition of the laplacian pyramid is performed on the obtained low-pass gaussian pyramid image, and the multi-scale decomposition is completed iteratively, wherein the decomposition iterative formula is expressed as:
Figure GDA0004054416110000091
Figure GDA0004054416110000092
LP l (b,c)＝L l-1 (b,c)-L l '(b,c)
where L is the number of levels of Gaussian pyramid L and Laplacian pyramid LP decomposition, e.g., L l First layer image of low-pass Gaussian pyramid, LP l Is the L-layer image, L 'of the Laplacian pyramid' l B represents the number of rows of the pyramid ith layer image, c represents the number of columns of the pyramid ith layer image, w (p, q) is a 3 x 3 low pass filter, p ∈ [ -2,2],q∈[-2,2]。
And step 32, respectively extracting phase consistent features of each layer of images in the Laplace pyramid, wherein in the Laplace pyramid, the low-resolution images are arranged on the upper layer, the high-resolution images are arranged on the lower layer, and the size of each layer of images is different, so that a group of multi-scale pyramid phase consistent images can be obtained.
Step 33, fusing the multi-scale pyramid phase consistent images to obtain a sample characteristic diagram based on phase consistency; the image fusion is to fuse the features and the details of each layer of image in the laplacian pyramid into one image. The specific mode is that contour features are extracted from large-scale phase consistent images, feature points are searched in neighborhoods of corresponding positions of small-scale phase consistent images to obtain accurate positioning of the feature points, and the feature points which exist in large scale but do not exist in small scale need to be reserved. The image features are tracked from coarse to fine, so that high-precision feature positioning can be obtained, and a fused image, namely a multi-scale pyramid feature map based on phase consistency, is obtained.
And step 34, performing the same processing on all sample images, and forming a sample feature space by all sample feature maps.
According to the technical scheme, the sample image is subjected to Laplacian pyramid decomposition, phase consistent feature extraction is performed on each layer of image, and finally, the images are fused into a final sample feature map, so that feature extraction is more complete and accurate.
Step 4, constructing an optimal attribute predictor according to the sample feature space; fig. 4 is a flowchart illustrating steps of constructing an optimal attribute predictor according to the sample feature space, according to the present invention, including the following steps:
step 41, constructing an attribute hypergraph according to the sample feature space and the attribute label of each sample;
for a given sample feature space, the attribute relationship between samples can be described by a hypergraph, and the attribute is a keyword used for describing the characteristic of each sample. In the hypergraph, each sample corresponds to a vertex in the hypergraph, and samples with the same attribute are classified into the same hypergraph. Because each sample has several attributes, a vertex can belong to multiple hyper-edges, and how many attributes there are in the sample space corresponds to how many hyper-edges there are. Because the hypergraph is used to describe the attribute relationships between sample data, it is called an attribute hypergraph.
And 42, performing regularized hypergraph segmentation on the attribute hypergraph, and solving an optimal attribute predictor to minimize the attribute relation loss of the attribute hypergraph and ensure the minimum error between the predicted attribute and the attribute label.
The technical scheme of the invention needs to construct an attribute predictor according to the attribute hypergraph and classify the image to be detected by utilizing the attribute predictor. Generally, such problems can be solved by regularized hypergraph segmentation, that is, for a given attribute hypergraph, a set of hypergraph vectors F exist to segment the vertex set of the attribute hypergraph into two subsets, so that the connection degree between samples with different attributes is as sparse as possible, and the connection degree between samples with the same attributes is as close as possible, that is, the attribute relationship between the samples is not destroyed as much as possible. Tangent vector F = [ F ] of the set of attribute hypergraphs 1,..., f m ]Called attribute prediction of samples, each element of the attribute hypergraph vector corresponds to the predicted value of all samples for the kth attribute.
On one hand, to ensure that the attribute relationship loss is minimal, i.e. the loss of the super-edge structure is minimized, the attribute relationship loss function is defined as: omega (F, G) * )＝Trace{F T L H F+λF T L * F, wherein F is a tangent vector needing to be solved, G is an attribute hypergraph, and G is * Class information is added to the attribute hypergraph and then a hypergraph, L, of the class relationship is depicted H Is Laplace matrix corresponding to the attribute hypergraph, λ is a positive parameter, L * Is a drawing G * The corresponding laplacian matrix, trace function represents the Trace operation of the matrix.
On the other hand, it is also required to ensure that the error between the attribute prediction F and the actual attribute label of the sample image is the minimum, and a specific error value can be measured by using the euclidean distance, and the calculation formula is as follows:
Figure GDA0004054416110000101
wherein, F is a tangent vector needing to be solved, also called a prediction attribute, G is an attribute hypergraph, and a matrix Y = [ Y ] 1 ,...,y m ]And representing an attribute label vector representing the sample images, wherein if all the sample images have the attribute, the element value corresponding to the attribute in the attribute label vector Y is 1, and otherwise, the number of the sample attributes is represented as-1,m.
Again, since the attribute predictor of the training sample image can be represented as a mapping problem from the feature sample space to the attribute hypergraph embedding space: f = X T B, wherein the matrix B = [ beta ] 1 ,...,β m ]Is a projection matrix from the feature sample space X to the attribute hypergraph embedding space. Thus, the problem solved by the hypergraph regularization attribute predictor can be converted into a mapping problem for solving the hypergraph embedding from a sample feature space to an attribute space, namely a problem for solving the optimal projection matrix B. Therefore, according to the above formulas, the final objective function can be defined as:
Figure GDA0004054416110000111
wherein, the Trace function represents the Trace operation of the matrix, () T Is the transpose operation of matrix, eta is a nonnegative regularization parameter, lambda is a positive parameter, X is the sample feature space, Y is the sample attribute label matrix, B is the projection matrix, L H A Laplace matrix L representing the correspondence of said attribute hypergraph * And representing the corresponding Laplace matrix after the attribute hypergraph is added with the category information. And solving an optimal solution according to the objective function of the attribute predictor to obtain a group of optimal attribute hypergraph projection matrixes B.
According to the technical scheme, the multivariate logical relation between the sample characteristic space and the attribute labels of the samples can be accurately described in an attribute hypergraph mode, and meanwhile, the classification accuracy of the to-be-detected images can be realized on the basis of the optimal attribute predictor solved by the regularization hypergraph segmentation method.
And 5, classifying the image samples to be detected according to the optimal attribute predictor. Fig. 5 is a flowchart illustrating the steps of classifying the image samples to be detected according to the optimal attribute predictor, and the method includes the following steps:
step 51, calculating an attribute predicted value of the image sample to be detected according to a calculation formula:
P k ＝sign(z k T B)
wherein, P k Representing attribute predicted values, z k For an image sample to be detected, B represents a projection matrix from the sample feature space to an attribute hypergraph embedding space, sign () function represents a positive sign and a negative sign of each element of a vector, and if P is the case k Is a positive number, and represents a cervical lesion sample z k The attribute of the sample with the cervical lesion characteristics is negative number, which represents the cervical lesion sample z k Does not have this property.
Step 52, utilizing a Sigmoid function to carry out normalization processing on the attribute predicted value to ensure that the value is between 0 and 1, wherein the normalization formula is as follows:
Figure GDA0004054416110000112
wherein, the vector r k Expressing the predicted value of each attribute of the image sample to be detected after normalization, wherein rho is a positive parameter of the normalized control scale, P k Representing an attribute prediction value; the normalized attribute prediction value can be regarded as the probability that a sample has a specific attribute; the posterior probability belonging to each test class is calculated according to the attribute existence probability r of a sample by using the Bayes rule, and the real class of the sample should have the maximum posterior probability.
Step 53, calculating the distance between the cervical lesion classified attribute list template and the predicted value of each normalized attribute of the image sample to be detected, and classifying the image sample to be detected, wherein the calculation formula is as follows:
Figure GDA0004054416110000113
wherein the function O (z) k ) The representation takes the class label with the minimum Euclidean distance, r k Representing the predicted value, t, of each normalized attribute of the image sample to be detected a A classification attribute list template representing cervical lesions. In a specific embodiment, an attribute list template is defined in advance for each type of lesion samples, each element of the template represents the prior probability that the type of samples have incompatible attributes, and the image samples to be detected are classified into the class with the minimum Euclidean distance by calculating the Euclidean distance between the predicted value after the normalization of each attribute of the image samples to be detected and each template.
The cervical image is preprocessed through the self-adaptive image enhancement technology, the region of interest is effectively segmented, the characteristic information of cervical lesions is effectively extracted according to a multi-scale pyramid method with consistent phases, classification is carried out by utilizing a class-positive definite hypergraph regularization attribute learning method, an attribute predictor is constructed, and the sample images to be detected are classified according to the attribute predictor. According to the experimental results, the average accuracy rate of extracting the lesion area by adopting the technical scheme of the invention reaches 93%, the average recall rate reaches 86%, and the effect of auxiliary detection and diagnosis can be well played.
The above embodiments are merely illustrative of the technical solutions of the present invention, and the present invention is not limited to the above embodiments, and any modifications or alterations according to the principles of the present invention should be within the protection scope of the present invention.

Claims (2)

1. A colposcopic image recognition method for detecting cervical lesions, comprising the steps of:
step 1, collecting a sample image;
step 2, carrying out data preprocessing on the sample image, and comprising the following steps:
step 21, performing adaptive enhancement filtering processing on the sample image, including the following steps:
step 211, carrying out equalization processing on the sample image by using a variance equalization method to enable the sample image to reach the required average gray scale and variance;
step 212, dividing the sample image into pixel blocks which are w in width and height and are overlapped with each other, wherein two adjacent pixel blocks are separated by one pixel in the horizontal direction, two adjacent pixel blocks are separated by one pixel in the vertical direction, and w is an odd number which is more than or equal to 3;
step 213, calculating the direction field of each pixel block and performing smoothing processing;
step 214, calculating the reliability of each pixel block direction field, wherein the calculation formula is as follows:
Figure FDA0004054416100000011
wherein R is a set of all pixels in the pixel block, (i, j) is a central pixel of the pixel block, (i ', j') is any pixel of the pixel block except for the central pixel (i, j), n is a total number of pixels in the pixel block, θ (i, j) represents a directional field of the pixel block centered at (i, j), and θ (i ', j') represents a directional field of the pixel block centered at (i ', j');
step 215, for each pixel block, judging whether the reliability χ (i, j) of the pixel block is greater than a threshold T, executing step 216, otherwise, jumping to step 217 to continue executing;
step 216, in a pixel block set which takes a non-central pixel (i ', j') of the pixel block as a central pixel, judging the number of blocks which satisfy the reliability χ (i ', j') less than the threshold T, if the number of blocks is greater than or equal to the threshold Tn, setting the direction field of the pixel block as the average value of the direction fields of the pixel blocks which satisfy the reliability χ (i ', j') less than the threshold T in the pixel block set, otherwise, setting the pixel block as an unrecoverable pixel block;
step 217, for each pixel block except the unrecoverable pixel block, calculating coordinates (x ', y') of all pixels in the pixel block after rotation, wherein the calculation formula is as follows:
x'＝icosθ+jsinθ
y'＝-isinθ+jcosθ
wherein, i, j is coordinate values of pixels in the pixel block on x, y axes; θ is the directional field of the pixel block centered on the pixel;
step 218, for each pixel block except the unrecoverable pixel block, performing convolution operation on the gray values of all the rotated pixel points in the pixel block through a Gaussian function, and limiting the result value after convolution between [0,255] as the gray value of the central pixel of the pixel block;
repeating the steps 215 to 218 until all pixel blocks are processed;
step 22, performing region segmentation on the sample image after the adaptive enhancement filtering processing, including the following steps:
step 221, pre-dividing the sample image subjected to the adaptive enhancement filtering processing by using a mean shift algorithm to obtain a plurality of superpixels;
step 222, constructing a weighted Graph according to the pre-divided superpixels, namely constructing a Graph cuts division model as follows:
Figure FDA0004054416100000021
wherein, alpha represents the segmentation label value, the value of 0 or 1,0 represents the background, 1 represents the foreground, I represents the set of all the super-pixels obtained by pre-segmentation, D i (α) represents the feature similarity of the current superpixel i with respect to the segmentation index α, m and n represent any two different superpixels adjacent to the current superpixel i, S m,n (α) represents the penalty value for adjacent superpixels m and n belonging to different label sets, γ is the equilibrium coefficient;
step 223, performing global S-T minimum segmentation operation on the weighted Graph to minimize the Graph cuts segmentation model to obtain optimal image segmentation;
step 3, establishing a sample characteristic space according to the preprocessed sample image, and the method comprises the following steps:
step 31, performing laplacian pyramid decomposition on the preprocessed sample image to obtain a layered image, wherein a decomposition iteration formula is expressed as:
Figure FDA0004054416100000023
Figure FDA0004054416100000022
LP l (b,c)＝L l-1 (b,c)-L l '(b,c)
wherein L is l First layer image of low-pass Gaussian pyramid, LP l Is the L-layer image, L 'of the Laplacian pyramid' l B represents the number of rows of the pyramid ith layer image, c represents the number of columns of the pyramid ith layer image, w (p, q) is a 3 x 3 low pass filter, p ∈ [ -2,2],q∈[-2,2]；
Step 32, respectively extracting phase consistent features of each layer of images in the Laplacian pyramid to obtain multi-scale pyramid phase consistent images;
step 33, fusing the multi-scale pyramid phase consistent images to obtain a sample characteristic diagram based on phase consistency;
step 34, performing the same processing on all sample images, and forming a sample feature space by all sample feature maps;
step 4, constructing an optimal attribute predictor according to the sample feature space, comprising the following steps:
step 41, constructing an attribute hypergraph according to the sample feature space and the attribute label of each sample;
step 42, carrying out regularization hypergraph segmentation on the attribute hypergraph, and solving an optimal attribute predictor to ensure that the attribute relation loss of the attribute hypergraph is minimum and the error between a predicted attribute and an attribute label is minimum; the objective function of the attribute predictor is as follows:
Figure FDA0004054416100000031
wherein, the Trace function represents the Trace operation of the matrix, () T Is the transpose operation of matrix, eta is a nonnegative regularization parameter, lambda is a positive parameter, X is the sample feature space, Y is the sample attribute label matrix, B is the projection matrix, L H A Laplace matrix L representing the correspondence of said attribute hypergraph * Representing a corresponding Laplace matrix after the attribute hypergraph is added with the category information;
and 5, classifying the image samples to be detected according to the optimal attribute predictor, and comprising the following steps of:
step 51, calculating an attribute predicted value of the image sample to be detected according to a calculation formula:
P k ＝sign(z k T B)
wherein, P k Representing attribute predicted values, z k B represents a projection matrix from the sample characteristic space to an attribute hypergraph embedding space for an image sample to be detected, and a sign () function represents a positive sign and a negative sign of each element of a vector;
step 52, performing normalization processing on the attribute predicted value, wherein a normalization formula is as follows:
Figure FDA0004054416100000032
wherein, the vector r k Expressing the predicted value of each attribute of the image sample to be detected after normalization, wherein rho is a positive parameter of the normalized control scale, P k Representing an attribute prediction value;
step 53, calculating the distance between the cervical lesion classified attribute list template and the predicted value of each normalized attribute of the image sample to be detected, and classifying the image sample to be detected, wherein the calculation formula is as follows:
Figure FDA0004054416100000033
wherein the function O (z) k ) The representation takes the class label with the minimum Euclidean distance, r k Representing the predicted value, t, of each normalized attribute of the image sample to be detected a A classification attribute list template representing cervical lesions.
2. The colposcopic image recognition method for detecting cervical lesions according to claim 1, wherein the "feature similarity of the current super pixel i with respect to the segmentation index α" is to use color gaussian distance measurement statistics of the region to characterize the super pixel, and the color gaussian distance measurement calculation formula is:
Figure FDA0004054416100000041
wherein, the (A) T For the transpose operation of the matrix, tr () is the trace operation of the matrix, d is the spatial dimension of the Gaussian feature, i.e. the number of channels in the color space, m and n represent any two different superpixels adjacent to the current superpixel i, g m Set of Gaussian parameters [ mu ] representing the correspondence of a superpixel m m ,∑ m }，g n Set of gaussian parameters μ representing the correspondence of a super-pixel n n ,∑ n }，μ m Means, Σ, representing the color characteristics in a superpixel m m Representing the covariance, μ, of the color features in the superpixel m n Representing the mean of the color features in the super-pixel n, sigma n Representing the covariance of the color features in the super-pixel n.
CN202010230980.XA 2020-03-27 2020-03-27 Colposcope image identification method for detecting cervical lesions Active CN111415350B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010230980.XA CN111415350B (en) 2020-03-27 2020-03-27 Colposcope image identification method for detecting cervical lesions

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010230980.XA CN111415350B (en) 2020-03-27 2020-03-27 Colposcope image identification method for detecting cervical lesions

Publications (2)

Publication Number Publication Date
CN111415350A CN111415350A (en) 2020-07-14
CN111415350B true CN111415350B (en) 2023-04-07

Family

ID=71494604

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010230980.XA Active CN111415350B (en) 2020-03-27 2020-03-27 Colposcope image identification method for detecting cervical lesions

Country Status (1)

Country Link
CN (1) CN111415350B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112419309B (en) * 2020-12-11 2023-04-07 上海联影医疗科技股份有限公司 Medical image phase determination method, apparatus, computer device and storage medium

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107240102A (en) * 2017-04-20 2017-10-10 合肥工业大学 Malignant tumour area of computer aided method of early diagnosis based on deep learning algorithm
CN108376565A (en) * 2018-02-13 2018-08-07 北京市神经外科研究所 A kind of image group prediction technique of glioma Ki-67 expressions
CN109741358A (en) * 2018-12-29 2019-05-10 北京工业大学 Superpixel segmentation method based on the study of adaptive hypergraph
CN110335267A (en) * 2019-07-05 2019-10-15 华侨大学 A kind of cervical lesions method for detecting area

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105930841B (en) * 2016-05-13 2018-01-26 百度在线网络技术（北京）有限公司 The method, apparatus and computer equipment of automatic semantic tagger are carried out to image

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107240102A (en) * 2017-04-20 2017-10-10 合肥工业大学 Malignant tumour area of computer aided method of early diagnosis based on deep learning algorithm
CN108376565A (en) * 2018-02-13 2018-08-07 北京市神经外科研究所 A kind of image group prediction technique of glioma Ki-67 expressions
CN109741358A (en) * 2018-12-29 2019-05-10 北京工业大学 Superpixel segmentation method based on the study of adaptive hypergraph
CN110335267A (en) * 2019-07-05 2019-10-15 华侨大学 A kind of cervical lesions method for detecting area

Also Published As

Publication number Publication date
CN111415350A (en) 2020-07-14

Similar Documents

Publication Publication Date Title
CN108830326B (en) Automatic segmentation method and device for MRI (magnetic resonance imaging) image
CN111985536B (en) Based on weak supervised learning gastroscopic pathology image Classification method
CN105427296B (en) A kind of thyroid gland focus image-recognizing method based on ultrasonoscopy low rank analysis
WO2022001571A1 (en) Computing method based on super-pixel image similarity
Kromp et al. Evaluation of deep learning architectures for complex immunofluorescence nuclear image segmentation
CN105957063B (en) CT image liver segmentation method and system based on multiple dimensioned weighting similarity measure
CN110458172A (en) A kind of Weakly supervised image, semantic dividing method based on region contrast detection
Pan et al. Cell detection in pathology and microscopy images with multi-scale fully convolutional neural networks
CN106056595A (en) Method for automatically identifying whether thyroid nodule is benign or malignant based on deep convolutional neural network
CN105931226A (en) Automatic cell detection and segmentation method based on deep learning and using adaptive ellipse fitting
CN107977661B (en) Region-of-interest detection method based on FCN and low-rank sparse decomposition
CN111754472A (en) Pulmonary nodule detection method and system
CN112686902B (en) Two-stage calculation method for brain glioma identification and segmentation in nuclear magnetic resonance image
CN106157330A (en) A kind of visual tracking method based on target associating display model
CN108932518A (en) A kind of feature extraction of shoes watermark image and search method of view-based access control model bag of words
CN111667491A (en) Breast mass image generation method with marginal landmark annotation information based on depth countermeasure network
CN114299383A (en) Remote sensing image target detection method based on integration of density map and attention mechanism
CN114140445A (en) Breast cancer pathological image identification method based on key attention area extraction
Xing et al. Machine learning and its application in microscopic image analysis
CN112488996A (en) Inhomogeneous three-dimensional esophageal cancer energy spectrum CT (computed tomography) weak supervision automatic labeling method and system
CN111415350B (en) Colposcope image identification method for detecting cervical lesions
CN108765384B (en) Significance detection method for joint manifold sequencing and improved convex hull
Shetty et al. Optimized deformable model-based segmentation and deep learning for lung cancer classification
Athinarayanan et al. Multi class cervical cancer classification by using ERSTCM, EMSD & CFE methods based texture features and fuzzy logic based hybrid kernel support vector machine classifier
Mulmule et al. Classification of overlapping cells in microscopic cervical images: a transfer learning approach

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
GR01 Patent grant
GR01 Patent grant
TR01 Transfer of patent right

Effective date of registration: 20240311

Address after: 350000, Building 8, Qishan Industrial Concentration Zone, Houshan Village, Nanyu Town, Minhou County, Fuzhou City, Fujian Province 320

Patentee after: Fuzhou Zhongsheng Zhanrui Technology Co.,Ltd.

Country or region after: China

Address before: 350001 No.18 daoshan Road, Fuzhou City, Fujian Province

Patentee before: FUJIAN MATERNAL AND CHILD CARE SERVICE CENTRE

Country or region before: China

TR01 Transfer of patent right