CN115393362A - Method, equipment and medium for selecting automatic glaucoma recognition model - Google Patents

Method, equipment and medium for selecting automatic glaucoma recognition model Download PDF

Info

Publication number
CN115393362A
CN115393362A CN202211332335.4A CN202211332335A CN115393362A CN 115393362 A CN115393362 A CN 115393362A CN 202211332335 A CN202211332335 A CN 202211332335A CN 115393362 A CN115393362 A CN 115393362A
Authority
CN
China
Prior art keywords
model
feature
glaucoma
source domain
vector
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.)
Granted
Application number
CN202211332335.4A
Other languages
Chinese (zh)
Other versions
CN115393362B (en
Inventor
张健
戴梅
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.)
Central South University
Original Assignee
Central South 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 Central South University filed Critical Central South University
Priority to CN202211332335.4A priority Critical patent/CN115393362B/en
Publication of CN115393362A publication Critical patent/CN115393362A/en
Application granted granted Critical
Publication of CN115393362B publication Critical patent/CN115393362B/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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/761—Proximity, similarity or dissimilarity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/7715—Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/87—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using selection of the recognition techniques, e.g. of a classifier in a multiple classifier system
    • 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
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Medical Informatics (AREA)
  • Artificial Intelligence (AREA)
  • Multimedia (AREA)
  • Databases & Information Systems (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a method, equipment and a medium for selecting an automatic glaucoma identification model, wherein the method comprises the following steps: acquiring a pre-training model library; selecting a common retinal image dataset as a source domain dataset
Figure DEST_PATH_IMAGE002
Taking the glaucoma fundus image data set as a target domain data set
Figure DEST_PATH_IMAGE004
(ii) a Measure each model in
Figure 401623DEST_PATH_IMAGE002
And
Figure 115501DEST_PATH_IMAGE004
inter-migratability: using model extraction
Figure 426397DEST_PATH_IMAGE002
And
Figure 556027DEST_PATH_IMAGE004
the feature vector of the middle sample is subjected to bilinear transformation, and the obtained high-dimensional feature vector is subjected to low-dimensional mapping to obtain a feature set
Figure DEST_PATH_IMAGE006
、
Figure DEST_PATH_IMAGE008
(ii) a Computing
Figure 934181DEST_PATH_IMAGE006
And
Figure 553381DEST_PATH_IMAGE008
the distance is used for representing the mobility of the current prediction model for automatically identifying the source domain and the target domain; and selecting the model with the strongest mobility for training, and automatically identifying the glaucoma. The prediction model selected by the invention does not need a glaucoma sample label and has better automatic glaucoma identification effect.

Description

Method, equipment and medium for selecting automatic glaucoma recognition model
Technical Field
The invention belongs to the technical field of deep learning, and particularly relates to a method, equipment and medium for selecting an automatic glaucoma identification model based on mobility measurement.
Background
Recent advances in deep learning have been applied to different medical fields for early detection or prediction of certain abnormalities. In the field of ophthalmology, medical image analysis using deep learning methods has made significant progress. Among the major ophthalmic abnormalities, glaucoma is a common and serious one, which can lead to irreversible loss of vision. At present, the automatic identification research of glaucoma at home and abroad is mainly based on glaucoma prior characteristics and stealth characteristics based on deep learning. Where classification of glaucoma is based on a large number of manually screened features, this has the advantage of being somewhat targeted. However, this also results in a high time cost, since a large amount of labor is required to screen the classification features; the second screened feature may be affected by subjective factors of the screening personnel, so that the screened feature is not accurate enough; the third model is difficult to generalize, and large-scale glaucoma fundus data is difficult to obtain because medical data relates to privacy of patients and data barrier problems among hospitals in China are still serious.
In order to solve the problem of lack of effective labeling of medical data, researchers propose a new solution, namely a transfer learning method. Transfer learning is a learning method that mimics the human visual system in the process of performing a new task in a particular domain, using a large amount of prior knowledge in other relevant domains. It is expected that the model will train an ideal recognition effect when the number of medical image data sets is small. And the method also has higher automatic identification performance on a new test data set. However, different retinal fundus image datasets, due to different scanners, image resolution, light source intensity and parameter settings, result in images with significant differences in appearance. Resulting in a large degradation of performance on the target domain data set for identifying well-performing deep learning models on the source domain data set. In current migration learning applications, finding the optimal migration strategy still requires time-consuming experimentation and domain knowledge. The migratability metric for the model can quantitatively reveal how easily it is to migrate knowledge learned from the source domain data to the target domain data. And providing guidance for selecting the transfer learning model. Therefore, model migratability measurement research is of great significance to the wide and efficient application of migration learning in glaucoma automatic identification. At present, the model mobility measurement research mainly comprises the following methods, which have breakthroughs and face certain limitations:
(1) Model migratability metrics based on empirical studies: taskonomy evaluates migration performance by retraining the source model for each target task. Expensive training calculations are required.
(2) Model migratability measurements based on analytical methods: h-score analytically assesses migratability by solving the HGR maximum correlation problem. The NCE measures migratability at a particular setting using conditional entropy. LEEP constructs an empirical predictor by estimating the joint distribution of pre-training and target label space, predicts the virtual label distribution of target data in source label space, and calculates the empirical condition distribution of the target label of a given virtual label. The performance of the empirical predictor is used to evaluate the pre-trained model. Although fast in computation, the a priori method is not accurate and is applied specifically to image classification of supervised pre-trained models. On the one hand, these methods have strict assumptions about the data, and on the other hand, work poorly in cross-domain settings.
(3) Mobility metric based on Optimal Transport (OT): the migratability is described as a linear combination of domain differences and task differences. In the calculation process, part of target data set data with known labels is needed to be used as observation samples.
(4) Attribution graph-based heterogeneous depth model migratability measures: and calculating a data attribution map of each trained model in the model base to the detection data set by utilizing the existing depth model attribution method. The mobility metric of the model is measured by the similarity of the attributed graphs. The method needs to establish a detection data set, and an author collects the part of data through various network picture search engines, so that a large amount of additional cost is increased; the requirement on the detection data is strict, and the quality of the detection data has great influence on the measurement effect of the model mobility; and calculating the attribution graph of each model in the model library to the target data, wherein when the number of the models is large, the storage of the attribution graphs and the calculation cost of the distance are not negligible.
(5) A zero sample image retrieval method and device based on Hash coding and graph attention mechanism are disclosed: from the macroscopic view, the method utilizes the selected model to extract picture features, compares the unknown label image with all known label images in the database one by one in the model application stage, and selects the label of the image with the minimum Hamming distance of the Hash code between the unknown label image and the database as the prediction classification result of the unknown label image. This model selection approach, lacking a priori estimates, does not enable a measure of the model's own migratability before the model is used for actual migration behavior. From the microscopic perspective, the method realizes classification by comparing each picture in the database one by one, and the classification performance of the model in practical application is greatly limited by the richness of the known label image library of the database.
Due to the characteristics of transfer learning, the application effect of the transfer learning method in the field of automatic identification of glaucoma fundus images is closely related to model selection. Searching for a model with optimal recognition performance to realize the migration behavior requires a great deal of experiments and information. Learning the model with an imperfect recognition accuracy even after consuming a lot of computing resources, is a waste of computing resources. This lack of an evaluated model selection approach leaves the field of automatic glaucoma identification with greater uncertainty. Therefore, the mobility of the deep learning model is measured, and the method has guiding significance for the application of the fundus image recognition model. The model migratability is affected by many factors, such as the size of the data set, the model optimization method, etc. The capability of the deep learning model to the feature extraction module of the image is a main factor influencing the migration capability of the model. How to measure the ability of the deep learning model to the feature extraction module of the image and how to compare different models under the same criterion is then a key issue in evaluating model mobility. The current research aiming at the model mobility measurement mainly faces the following problems: the source model needs to be retrained, and expensive training calculation is consumed; strict assumption is made on data, and the working effect in cross-domain setting is poor; the partial target domain data sample label needs to be known.
Disclosure of Invention
Aiming at the defects of the existing model mobility measurement method, the invention provides a method, equipment and medium for selecting an automatic glaucoma identification model based on mobility measurement, which do not need a glaucoma sample label and have better automatic glaucoma identification effect.
In order to achieve the technical purpose, the invention adopts the following technical scheme:
a method for glaucoma auto-identification model selection based on migratability metrics, comprising:
step 1, obtaining a pre-training model library obtained by training on a standard data set
Figure DEST_PATH_IMAGE001
(ii) a Wherein
Figure 830351DEST_PATH_IMAGE002
Are respectively asNA plurality of different pre-training models;
step 2, selecting a public retina image data set as a source domain data set
Figure DEST_PATH_IMAGE003
Comprises that
Figure 791354DEST_PATH_IMAGE004
A common retinal image sample, noted
Figure DEST_PATH_IMAGE005
(ii) a Using the glaucoma fundus image dataset as a target domain dataset
Figure 32979DEST_PATH_IMAGE006
Comprises thatnA sample of a glaucoma fundus image
Figure DEST_PATH_IMAGE007
(ii) a For each pre-training model, its data set in the source domain is measured in steps 3-5
Figure 104840DEST_PATH_IMAGE003
With the target domain data set
Figure 748311DEST_PATH_IMAGE006
Migratability for automatic identification;
step 3, extracting a source domain data set by using a pre-training model
Figure 349057DEST_PATH_IMAGE003
And a target domain data set
Figure 343558DEST_PATH_IMAGE006
Carrying out bilinear transformation on the extracted feature vector to obtain a high-dimensional feature vector;
step 4, performing count sketch mapping on each high-dimensional feature vector obtained in the step 3 to obtain a characterization source domain data set
Figure 687951DEST_PATH_IMAGE003
Source domain feature set of
Figure 451508DEST_PATH_IMAGE008
And characterizing the target domain dataset
Figure 223155DEST_PATH_IMAGE006
Target domain feature set of
Figure DEST_PATH_IMAGE009
；
Step 5, calculating a source domain feature set
Figure 704952DEST_PATH_IMAGE008
Central feature of
Figure 354501DEST_PATH_IMAGE010
And a set of target features
Figure DEST_PATH_IMAGE011
Central feature of
Figure 972564DEST_PATH_IMAGE012
Then calculate
Figure 180692DEST_PATH_IMAGE010
And with
Figure 884206DEST_PATH_IMAGE012
And using the distance to characterize the current prediction model versus the source domain data set
Figure DEST_PATH_IMAGE013
With the target domain data set
Figure 570402DEST_PATH_IMAGE006
The mobility of automatic identification is carried out, and the mobility of the model with the minimum distance is ultra strong;
step 6, selecting a pre-training model with the strongest mobility, and using a source domain data set
Figure 308551DEST_PATH_IMAGE003
And training, and taking the model obtained by training as an automatic glaucoma identification model.
Further, theNThe different pre-training models are all heterogeneous deep learning models.
Further, the computation formula for performing bilinear transformation on the feature vector is as follows:
Figure 422000DEST_PATH_IMAGE014
in the formula (I), the compound is shown in the specification,
Figure DEST_PATH_IMAGE015
the feature vectors extracted for the pre-trained models,
Figure 612810DEST_PATH_IMAGE016
is a feature vector
Figure 102697DEST_PATH_IMAGE017
The transposed vector of (a) is provided,
Figure DEST_PATH_IMAGE018
a matrix obtained by bilinear transformation;
then the matrix is divided
Figure 695353DEST_PATH_IMAGE018
All elements are spliced into a length of
Figure 245283DEST_PATH_IMAGE019
High-dimensional feature vector of
Figure DEST_PATH_IMAGE020
。
Further, the specific process of performing count sketch mapping on each high-dimensional feature vector is as follows:
(1) Self-defining a projection dimension d of the count sketch transformation function;
(2) Randomly generating an array
Figure 188968DEST_PATH_IMAGE021
And
Figure DEST_PATH_IMAGE022
wherein
Figure 216967DEST_PATH_IMAGE021
Slave array
Figure 664129DEST_PATH_IMAGE023
The assignment is randomly drawn,
Figure 384960DEST_PATH_IMAGE022
assigning values from an array {1, -1} of random samples; initializationdZero vector of dimension
Figure 284783DEST_PATH_IMAGE024
；
(3) Computing
Figure 615008DEST_PATH_IMAGE025
Obtained bydDimension vector
Figure DEST_PATH_IMAGE026
The feature vector is obtained by mapping; wherein
Figure 182255DEST_PATH_IMAGE027
As high-dimensional feature vectors
Figure 73988DEST_PATH_IMAGE020
ToiAnd (4) a component.
Further, the source domain feature set
Figure 461107DEST_PATH_IMAGE008
Central feature of
Figure 830909DEST_PATH_IMAGE010
The calculating method comprises the following steps:
(1) Feature set for source domain
Figure DEST_PATH_IMAGE028
Setting the number of clusters
Figure 252663DEST_PATH_IMAGE029
Order cluster
Figure DEST_PATH_IMAGE030
(ii) a Wherein
Figure 580876DEST_PATH_IMAGE031
Respectively source domain data sets
Figure 455291DEST_PATH_IMAGE003
InmThe samples are mapped by the count sketchmA feature vector;
(2) From
Figure 628783DEST_PATH_IMAGE008
Randomly selecting 1 feature vector as initial mean vector
Figure DEST_PATH_IMAGE032
；
(3) From
Figure 170623DEST_PATH_IMAGE008
Randomly selects 1 feature vector
Figure 404158DEST_PATH_IMAGE033
Adding to clusters
Figure DEST_PATH_IMAGE034
；
(4) Updating mean vector
Figure 765869DEST_PATH_IMAGE035
At the same time will
Figure 477473DEST_PATH_IMAGE033
From cluster C and source domain feature set
Figure 873820DEST_PATH_IMAGE008
Removing; wherein
Figure 278256DEST_PATH_IMAGE032
And
Figure 861684DEST_PATH_IMAGE010
the mean vector is the vector before and after updating;
(5) Repeating the steps (3) and (4) until the source domain feature set
Figure 878444DEST_PATH_IMAGE008
Empty, mean vector at this time
Figure 129297DEST_PATH_IMAGE010
Is the source domain feature set
Figure 439056DEST_PATH_IMAGE008
Central feature of
Figure 775359DEST_PATH_IMAGE010
；
The set of target domain features
Figure 94345DEST_PATH_IMAGE011
Central feature of
Figure 199704DEST_PATH_IMAGE012
Computing method, and source domain feature set
Figure 680364DEST_PATH_IMAGE008
Central feature of
Figure 238384DEST_PATH_IMAGE010
The calculation method is the same.
In a further aspect of the present invention,
Figure 361061DEST_PATH_IMAGE010
and
Figure 320927DEST_PATH_IMAGE012
the Kanbera distance between them is calculated as:
Figure DEST_PATH_IMAGE036
in the formula (I), the compound is shown in the specification,
Figure 238067DEST_PATH_IMAGE037
is composed of
Figure 548963DEST_PATH_IMAGE010
And
Figure 209751DEST_PATH_IMAGE012
the Kanbera distance between the two can be determined,
Figure DEST_PATH_IMAGE038
and
Figure 24124DEST_PATH_IMAGE039
respectively represent
Figure DEST_PATH_IMAGE040
And
Figure 643324DEST_PATH_IMAGE012
to (1)iDimension feature, d is the dimension of the feature vector obtained by the count sketch mapping.
An electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor is enabled to implement the method for selecting model for glaucoma automatic identification based on migratability metrics according to any of the above technical solutions.
A computer-readable storage medium, on which a computer program is stored, which when executed by a processor implements the method for selecting a model for glaucoma automatic identification based on migratability measures according to any of the above-mentioned aspects.
Advantageous effects
1 in the prior art, a plurality of source models need to be retrained when transfer learning is carried out, and the transferability of each model is evaluated according to the recognition precision of a target image after the actual transfer behavior of the model, however, in the field of medical images, the data quantity of fundus data sets is small, the difference between data is large, and the application effect of the transfer learning in the field of automatic glaucoma recognition is limited. According to the method, the model mobility is measured at the stage of extracting the image characteristic from the convolution basis of the deep learning model, the source model does not need to be retrained, the model does not need to be actually migrated, the deep learning model with better migration performance can be selected for automatically identifying the glaucoma fundus image, and the glaucoma sample label is not needed and the better automatic glaucoma identification effect is achieved.
2. Aiming at the problems that the feature dimension extracted from the convolution base of the learning model with different depths is high, the characterization capability is not strong, the measurement cannot be realized and the like, the invention provides the method for generating the joint representation by using the bilinear feature and enhancing the characterization capability of the feature vector; approximating a kernel function by using a count sketch transformation function, and mapping high-dimensional bilinear features to the same vector space with relatively low dimensionality; the distance between the converted image feature vectors is measured by the Kanbera quantity, so that the model mobility can be reflected, and guidance is provided for model selection in the migration learning application process. On one hand, compared with other measurement methods, the Kanbera distance is suitable for measuring the distance between two points in a vector space, is sensitive to the value change close to 0 (more than or equal to 0) and is suitable for the model mobility measurement scene, and on the other hand, the method is low in calculation cost and does not need extra space storage cost.
3. Aiming at the problems that in the field of model mobility measurement, strict assumptions are made on data of a source domain and data of a target domain, and measurement effects are poor in cross-domain setting, the method for performing mobility measurement on a plurality of pre-training models does not strictly assume data of the source domain and the target domain, and measurement effects are not influenced in cross-domain setting.
Drawings
FIG. 1 is a diagram of a full flow analysis of a method according to an embodiment of the present application.
Detailed Description
The following describes embodiments of the present invention in detail, which are developed based on the technical solutions of the present invention, and give detailed implementation manners and specific operation procedures to further explain the technical solutions of the present invention.
Example 1
The embodiment provides a glaucoma automatic identification model selection method based on migratability measurement, which is shown in fig. 1 and includes the following steps:
step 1, obtaining a pre-training model library obtained by training on a standard data set ImageNet
Figure 910357DEST_PATH_IMAGE001
(ii) a Wherein
Figure 109257DEST_PATH_IMAGE002
Are respectively asNA different pre-training model, in this embodimentNA heterogeneous deep learning model.
Step 2, selecting a public retina image data set as a source domain data set
Figure 43715DEST_PATH_IMAGE003
Comprises that
Figure 302658DEST_PATH_IMAGE004
A common retinal image sample, noted
Figure 549663DEST_PATH_IMAGE005
(ii) a Using the cyan light fundus image dataset as a target domain dataset
Figure 817834DEST_PATH_IMAGE006
N cyan light fundus image samples, and recording
Figure 606798DEST_PATH_IMAGE007
(ii) a For each pre-training model, its data set in the source domain is measured in steps 3-5
Figure 36642DEST_PATH_IMAGE003
With the target domain data set
Figure 278268DEST_PATH_IMAGE006
For automatic identification.
The common retinal image set can be Drishti-GS, RIM-ONE-R1, R2, R3, REGUGGE and the like, and Drishti-GS is selected in the embodiment and comprises 101 retinal images and mask marks of optical disc and optical cup for detecting glaucoma.
Step 3, extracting a source domain data set by using a pre-training model
Figure 84550DEST_PATH_IMAGE003
And a target domain data set
Figure 728021DEST_PATH_IMAGE006
And then carrying out bilinear transformation on the extracted feature vector to obtain a high-dimensional feature vector.
The feature vectors of each common retinal image and each cyan fundus image are recorded as
Figure 594346DEST_PATH_IMAGE041
Wherein
Figure DEST_PATH_IMAGE042
Representing feature vectors
Figure 323267DEST_PATH_IMAGE015
S-dimensional feature of (1).
For feature vector
Figure 933240DEST_PATH_IMAGE015
Performing bilinear transformation to obtain a value ofMatrix of s
Figure 696797DEST_PATH_IMAGE018
Namely:
Figure 468444DEST_PATH_IMAGE014
then the matrix is divided into
Figure 684661DEST_PATH_IMAGE018
All elements are spliced to obtain the product with the length of
Figure 832746DEST_PATH_IMAGE019
High-dimensional feature vector of
Figure 450809DEST_PATH_IMAGE020
。
Step 4, performing count sketch mapping on each high-dimensional feature vector obtained in the step 3 to obtain a characterization source domain data set
Figure 393357DEST_PATH_IMAGE003
Source domain feature set of
Figure 831292DEST_PATH_IMAGE008
And characterizing the target domain dataset
Figure 783067DEST_PATH_IMAGE006
Target domain feature set of
Figure 255637DEST_PATH_IMAGE011
. The embodiment specifically includes:
(1) And (5) defining the projection dimension d of the count sketch transformation function. The appropriate setting of d depends on the amount of training data, memory budget and task difficulty. In this embodiment d =8000 is sufficient to achieve near maximum accuracy.
(2) Randomly generating an array
Figure 136131DEST_PATH_IMAGE021
And
Figure 61361DEST_PATH_IMAGE022
in which
Figure 551249DEST_PATH_IMAGE021
Slave array
Figure 143904DEST_PATH_IMAGE023
The assignment is randomly extracted and,
Figure 428255DEST_PATH_IMAGE022
assigning values from an array {1, -1} of random samples; initializationdZero vector of dimension
Figure 840782DEST_PATH_IMAGE024
。
(3) Computing
Figure 134360DEST_PATH_IMAGE025
Obtained bydDimension vector
Figure 581521DEST_PATH_IMAGE026
The feature vector obtained by mapping is low in dimensionality and high in representation; wherein
Figure 364670DEST_PATH_IMAGE027
As a high-dimensional feature vector
Figure 530072DEST_PATH_IMAGE020
To (1)iAnd (4) a component.
Integrating source domain data sets
Figure 361762DEST_PATH_IMAGE003
Passing through the model
Figure 663430DEST_PATH_IMAGE043
Extracting features from the convolution basis, mapping a count sketch function to finally obtain a feature set which is expressed as
Figure 555163DEST_PATH_IMAGE028
In which
Figure 942282DEST_PATH_IMAGE033
Representing source domain data
Figure DEST_PATH_IMAGE044
The characteristics of (1). Target domain data set
Figure 341777DEST_PATH_IMAGE006
Passing through the model
Figure 497951DEST_PATH_IMAGE043
Extracting features from the convolution base, mapping the count sketch function to finally obtain a feature set expressed as
Figure 560585DEST_PATH_IMAGE045
In which
Figure DEST_PATH_IMAGE046
Representing target domain data
Figure 700580DEST_PATH_IMAGE047
The characteristics of (1).
Step 5, calculating a source domain characteristic set
Figure 608493DEST_PATH_IMAGE008
Central feature of
Figure 884753DEST_PATH_IMAGE010
And a set of target features
Figure 118289DEST_PATH_IMAGE009
Central feature of
Figure DEST_PATH_IMAGE048
Then calculate
Figure 745579DEST_PATH_IMAGE040
And with
Figure 457183DEST_PATH_IMAGE048
The Kancperra distance therebetween, and using the distance to characterize the current prediction model versus the source domain data set
Figure 853529DEST_PATH_IMAGE003
With the target domain data set
Figure 257966DEST_PATH_IMAGE006
And the mobility of the automatic identification is realized, and the mobility of the model with the minimum distance is ultra strong.
Wherein the source domain feature set
Figure 841394DEST_PATH_IMAGE008
Central feature of
Figure 622268DEST_PATH_IMAGE010
The calculating method comprises the following steps:
(1) Feature set for source domain
Figure 607542DEST_PATH_IMAGE028
Setting the number of clustered clusters
Figure 182880DEST_PATH_IMAGE029
Order cluster
Figure 253604DEST_PATH_IMAGE030
(ii) a Wherein
Figure 572590DEST_PATH_IMAGE031
Respectively source domain data sets
Figure 179414DEST_PATH_IMAGE013
InmThe samples are mapped by the count sketchmA feature vector;
(2) From
Figure 660074DEST_PATH_IMAGE008
Randomly selecting 1 feature vector as initial mean vector
Figure 483673DEST_PATH_IMAGE032
；
(3) From
Figure 340771DEST_PATH_IMAGE008
Randomly selects 1 feature vector
Figure 566216DEST_PATH_IMAGE033
Adding to cluster C;
(4) Updating mean vectors
Figure 483356DEST_PATH_IMAGE035
At the same time will
Figure 263093DEST_PATH_IMAGE033
From cluster C and source domain feature set
Figure 923882DEST_PATH_IMAGE008
Removing; wherein
Figure 3833DEST_PATH_IMAGE032
And
Figure 91875DEST_PATH_IMAGE010
the mean vector is the vector before and after updating;
(5) Repeating the steps (3) and (4) until the source domain feature set
Figure 624487DEST_PATH_IMAGE008
Empty, mean vector at this time
Figure 823388DEST_PATH_IMAGE010
Is the source domain feature set
Figure 757846DEST_PATH_IMAGE008
Central feature of
Figure 16789DEST_PATH_IMAGE010
。
Target domain feature set
Figure 36697DEST_PATH_IMAGE011
Central feature of
Figure 304868DEST_PATH_IMAGE012
Computing method, and source domain feature set
Figure 828253DEST_PATH_IMAGE008
Central feature of
Figure 258097DEST_PATH_IMAGE010
The calculation method comprises the following steps:
(1) Targeting domain feature sets
Figure 765302DEST_PATH_IMAGE045
Setting the number of clusters
Figure 571584DEST_PATH_IMAGE029
Order cluster
Figure 949475DEST_PATH_IMAGE030
(ii) a Wherein
Figure 585774DEST_PATH_IMAGE049
Respectively target domain feature set
Figure 49116DEST_PATH_IMAGE011
The n samples are mapped by the count sketchnA feature vector;
(2) From
Figure 659089DEST_PATH_IMAGE011
Randomly selecting 1 feature vector as initial mean vector
Figure DEST_PATH_IMAGE050
；
(3) From
Figure 422646DEST_PATH_IMAGE011
Randomly selects 1 feature vector
Figure 194293DEST_PATH_IMAGE051
Adding to cluster C;
(4) Updating mean vector
Figure 410511DEST_PATH_IMAGE052
At the same time will
Figure 558595DEST_PATH_IMAGE051
From cluster C and target domain feature set
Figure 176658DEST_PATH_IMAGE011
Removing; wherein
Figure 384786DEST_PATH_IMAGE050
And
Figure 822720DEST_PATH_IMAGE012
the mean vector is the vector before and after updating;
(5) Repeating the steps (3) and (4) until the target domain feature set
Figure 774496DEST_PATH_IMAGE009
Empty, mean vector at this time
Figure 247065DEST_PATH_IMAGE012
Is the target domain feature set
Figure 626094DEST_PATH_IMAGE009
Central feature of (2)
Figure 816904DEST_PATH_IMAGE012
。
In addition, the first and second substrates are,
Figure 306791DEST_PATH_IMAGE010
and
Figure 633867DEST_PATH_IMAGE012
the Kanbera distance between them is calculated as:
Figure 918218DEST_PATH_IMAGE036
in the formula (I), the compound is shown in the specification,
Figure 596324DEST_PATH_IMAGE037
is composed of
Figure 391367DEST_PATH_IMAGE010
And
Figure 838529DEST_PATH_IMAGE012
the Kanbera distance between the two can be determined,
Figure 293781DEST_PATH_IMAGE038
and
Figure 193604DEST_PATH_IMAGE039
respectively represent
Figure 25294DEST_PATH_IMAGE010
And
Figure 592541DEST_PATH_IMAGE012
to (1) aiThe dimensional characteristics of the image data are measured,dthe dimensions of the feature vector obtained for the count sketch map.
For each pre-training model, the method obtained according to the step 3-5
Figure 218695DEST_PATH_IMAGE010
And
Figure 605814DEST_PATH_IMAGE012
the Kancperla distance between the two can be used for representing the source domain data set of the current prediction model
Figure 241194DEST_PATH_IMAGE003
With the target domain data set
Figure 662949DEST_PATH_IMAGE006
The automatic identification is performed. And the smaller the Kanbera distance, the more migratability of the pre-trained model is indicated.
Step (ii) of6, selecting the pre-training model with the strongest mobility, and using the labeled source domain data set
Figure 725582DEST_PATH_IMAGE003
And training, and taking the model obtained by training as an automatic glaucoma identification model.
Example 2
An electronic device comprising a memory and a processor, the memory having stored therein a computer program that, when executed by the processor, causes the processor to implement the method for automatic glaucoma recognition model selection based on migratability metrics of embodiment 1.
Example 3
A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the method for automatic glaucoma identification model selection based on migratability metrics according to embodiment 1.
The above embodiments are preferred embodiments of the present application, and those skilled in the art can make various changes or modifications without departing from the general concept of the present application, and such changes or modifications should fall within the scope of the claims of the present application.

Claims (8)

1. A method for selecting an automatic glaucoma identification model based on a mobility metric, comprising:
step 1, obtaining a pre-training model library obtained by training on a standard data set
Figure 96951DEST_PATH_IMAGE001
(ii) a Wherein
Figure 784065DEST_PATH_IMAGE002
Are respectively asNA plurality of different pre-training models;
step 2, selecting a public retina image data set as a source domain data set
Figure 829381DEST_PATH_IMAGE003
Comprises that
Figure 490170DEST_PATH_IMAGE004
A common retinal image sample, noted
Figure 304542DEST_PATH_IMAGE005
(ii) a Using the glaucoma fundus image dataset as a target domain dataset
Figure 392584DEST_PATH_IMAGE006
Comprises thatnA sample of a glaucoma fundus image
Figure 925196DEST_PATH_IMAGE007
(ii) a For each pre-training model, its data set in the source domain is measured in steps 3-5
Figure 389676DEST_PATH_IMAGE003
With the target domain data set
Figure 58554DEST_PATH_IMAGE006
Migratability for automatic identification;
step 3, extracting a source domain data set by using a pre-training model
Figure 317497DEST_PATH_IMAGE003
And a target domain data set
Figure 337406DEST_PATH_IMAGE006
Carrying out bilinear transformation on the extracted feature vector to obtain a high-dimensional feature vector;
step 4, performing count sketch mapping on each high-dimensional feature vector obtained in the step 3 to obtain a characterization source domain data set
Figure 339997DEST_PATH_IMAGE003
Source domain feature set of
Figure 128962DEST_PATH_IMAGE008
And characterizing the target domain dataset
Figure 558806DEST_PATH_IMAGE006
Target domain feature set of
Figure 66011DEST_PATH_IMAGE009
；
Step 5, calculating a source domain characteristic set
Figure 606713DEST_PATH_IMAGE008
Central feature of
Figure 515763DEST_PATH_IMAGE010
And a set of target features
Figure 116509DEST_PATH_IMAGE009
Central feature of
Figure 845431DEST_PATH_IMAGE011
Then calculate
Figure 189824DEST_PATH_IMAGE010
And
Figure 953381DEST_PATH_IMAGE011
the Kancperra distance therebetween, and using the distance to characterize the current prediction model versus the source domain data set
Figure 492072DEST_PATH_IMAGE012
With the target domain data set
Figure 442710DEST_PATH_IMAGE006
The mobility of automatic identification is carried out, and the mobility of the model with the minimum distance is ultra strong;
step 6, selecting the pre-stage with the strongest mobilityTraining models using source domain datasets
Figure 590795DEST_PATH_IMAGE003
And training, and taking the model obtained by training as an automatic glaucoma identification model.
2. The method for glaucoma model selection with automatic recognition according to claim 1, wherein the model is selected from the group consisting of a model for glaucoma model selection, and a model for glaucoma model selectionNThe different pre-training models are all heterogeneous deep learning models.
3. The method for selecting a model for automatic glaucoma recognition according to claim 1, wherein the computation formula for bilinear transformation of the feature vectors is:
Figure 208858DEST_PATH_IMAGE013
in the formula (I), the compound is shown in the specification,
Figure 416986DEST_PATH_IMAGE014
the feature vectors extracted for the pre-trained models,
Figure 854920DEST_PATH_IMAGE015
is a feature vector
Figure 541116DEST_PATH_IMAGE014
The transposed vector of (a) is,
Figure 279265DEST_PATH_IMAGE016
a matrix obtained by bilinear transformation;
then the matrix is divided into
Figure 658294DEST_PATH_IMAGE016
All elements are spliced into a length of
Figure 849104DEST_PATH_IMAGE017
Is highDimensional feature vector
Figure 73412DEST_PATH_IMAGE018
。
4. The method for selecting a model for automatic glaucoma recognition according to claim 1, wherein the specific process of performing count sketch mapping on each high-dimensional feature vector is as follows:
(1) Self-defining a projection dimension d of the count sketch transformation function;
(2) Randomly generating an array
Figure 400488DEST_PATH_IMAGE019
And
Figure 950418DEST_PATH_IMAGE020
wherein
Figure 628524DEST_PATH_IMAGE019
Slave array
Figure 922102DEST_PATH_IMAGE021
The assignment is randomly extracted and,
Figure 103685DEST_PATH_IMAGE020
assigning values from an array {1, -1} of random samples; initializationdZero vector of dimension
Figure 558937DEST_PATH_IMAGE022
；
(3) Computing
Figure 724339DEST_PATH_IMAGE023
Obtained bydDimension vector
Figure 556029DEST_PATH_IMAGE024
The feature vector is obtained by mapping; wherein
Figure 857697DEST_PATH_IMAGE025
As a high-dimensional feature vector
Figure 483851DEST_PATH_IMAGE018
To (1)iAnd (4) a component.
5. The method for selecting a model for automatic recognition of glaucoma according to claim 1, wherein the set of source domain features
Figure 635084DEST_PATH_IMAGE026
Central feature of
Figure 270465DEST_PATH_IMAGE027
The calculation method comprises the following steps:
(1) Feature set for source domain
Figure 426639DEST_PATH_IMAGE028
Setting the number of clustered clusters
Figure 223694DEST_PATH_IMAGE029
Order cluster
Figure 98109DEST_PATH_IMAGE030
(ii) a Wherein
Figure 537181DEST_PATH_IMAGE031
Respectively source domain data sets
Figure 813441DEST_PATH_IMAGE012
InmThe samples are mapped by the count sketchmA feature vector;
(2) From
Figure 781397DEST_PATH_IMAGE026
Randomly selecting 1 feature vector as initial mean vector
Figure 143109DEST_PATH_IMAGE032
；
(3) From
Figure 120292DEST_PATH_IMAGE026
Randomly selects 1 feature vector
Figure 251059DEST_PATH_IMAGE033
Adding to cluster C;
(4) Updating mean vector
Figure 389916DEST_PATH_IMAGE034
At the same time will
Figure 238924DEST_PATH_IMAGE033
From cluster C and source domain feature set
Figure 754219DEST_PATH_IMAGE026
Removing; wherein
Figure 739492DEST_PATH_IMAGE032
And
Figure 314830DEST_PATH_IMAGE027
the mean vector is the vector before and after updating;
(5) Repeating the steps (3) and (4) until the source domain feature set
Figure 651133DEST_PATH_IMAGE026
Empty, mean vector at this time
Figure 970119DEST_PATH_IMAGE027
Is the source domain feature set
Figure 809899DEST_PATH_IMAGE026
Central feature of
Figure 556138DEST_PATH_IMAGE027
；
The set of target domain features
Figure 114159DEST_PATH_IMAGE035
Central feature of
Figure 738300DEST_PATH_IMAGE036
Computing method, and source domain feature set
Figure 432587DEST_PATH_IMAGE026
Central feature of
Figure 349727DEST_PATH_IMAGE027
The calculation method is the same.
6. The glaucoma automatic recognition model selection method according to claim 1,
Figure 660623DEST_PATH_IMAGE027
and
Figure 55832DEST_PATH_IMAGE036
the Kanbera distance between them is calculated as:
Figure 870204DEST_PATH_IMAGE037
in the formula (I), the compound is shown in the specification,
Figure 223825DEST_PATH_IMAGE038
is composed of
Figure 490859DEST_PATH_IMAGE027
And
Figure 955338DEST_PATH_IMAGE036
the kanperla distance therebetween is increased by the distance,
Figure 624217DEST_PATH_IMAGE039
and
Figure 148739DEST_PATH_IMAGE040
respectively represent
Figure 168648DEST_PATH_IMAGE027
And
Figure 905659DEST_PATH_IMAGE036
to (1) aiThe dimensional characteristics of the image data are measured,dthe dimensions of the feature vector obtained for the count sketch map.
7. An electronic device comprising a memory and a processor, the memory having stored therein a computer program, wherein the computer program, when executed by the processor, causes the processor to carry out the method according to any one of claims 1 to 6.
8. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the method according to any one of claims 1 to 6.
CN202211332335.4A 2022-10-28 2022-10-28 Method, equipment and medium for selecting automatic glaucoma recognition model Active CN115393362B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211332335.4A CN115393362B (en) 2022-10-28 2022-10-28 Method, equipment and medium for selecting automatic glaucoma recognition model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211332335.4A CN115393362B (en) 2022-10-28 2022-10-28 Method, equipment and medium for selecting automatic glaucoma recognition model

Publications (2)

Publication Number Publication Date
CN115393362A true CN115393362A (en) 2022-11-25
CN115393362B CN115393362B (en) 2023-02-03

Family

ID=84115167

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211332335.4A Active CN115393362B (en) 2022-10-28 2022-10-28 Method, equipment and medium for selecting automatic glaucoma recognition model

Country Status (1)

Country Link
CN (1) CN115393362B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110070535A (en) * 2019-04-23 2019-07-30 东北大学 A kind of retinal vascular images dividing method of Case-based Reasoning transfer learning
CN110378366A (en) * 2019-06-04 2019-10-25 广东工业大学 A kind of cross-domain image classification method based on coupling knowledge migration
CN113344016A (en) * 2020-02-18 2021-09-03 深圳云天励飞技术有限公司 Deep migration learning method and device, electronic equipment and storage medium
US20210369195A1 (en) * 2018-04-26 2021-12-02 Voxeleron, LLC Method and system for disease analysis and interpretation
CN114724231A (en) * 2022-04-13 2022-07-08 东北大学 Glaucoma multi-modal intelligent recognition system based on transfer learning

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210369195A1 (en) * 2018-04-26 2021-12-02 Voxeleron, LLC Method and system for disease analysis and interpretation
CN110070535A (en) * 2019-04-23 2019-07-30 东北大学 A kind of retinal vascular images dividing method of Case-based Reasoning transfer learning
CN110378366A (en) * 2019-06-04 2019-10-25 广东工业大学 A kind of cross-domain image classification method based on coupling knowledge migration
CN113344016A (en) * 2020-02-18 2021-09-03 深圳云天励飞技术有限公司 Deep migration learning method and device, electronic equipment and storage medium
CN114724231A (en) * 2022-04-13 2022-07-08 东北大学 Glaucoma multi-modal intelligent recognition system based on transfer learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
吴星等: "人工智能眼底分析技术对青光眼病灶的诊断价值研究", 《解放军医学院学报》 *
徐志京等: "青光眼眼底图像的迁移学习分类方法", 《计算机工程与应用》 *

Also Published As

Publication number Publication date
CN115393362B (en) 2023-02-03

Similar Documents

Publication Publication Date Title
Stacke et al. Measuring domain shift for deep learning in histopathology
Chen et al. Source-free domain adaptive fundus image segmentation with denoised pseudo-labeling
Zhang et al. Quantifying facial age by posterior of age comparisons
US20220237788A1 (en) Multiple instance learner for tissue image classification
Putra et al. Enhanced skin condition prediction through machine learning using dynamic training and testing augmentation
Prabhu et al. Few-shot learning for dermatological disease diagnosis
CN113454733A (en) Multi-instance learner for prognostic tissue pattern recognition
Sainju et al. Automated bleeding detection in capsule endoscopy videos using statistical features and region growing
WO2019015246A1 (en) Image feature acquisition
Filipovych et al. Semi-supervised cluster analysis of imaging data
Zakazov et al. Anatomy of domain shift impact on U-Net layers in MRI segmentation
Mitchell-Heggs et al. Neural manifold analysis of brain circuit dynamics in health and disease
Zhang et al. Feature-transfer network and local background suppression for microaneurysm detection
Naqvi et al. Feature quality-based dynamic feature selection for improving salient object detection
Zhou et al. Adaptive weighted locality-constrained sparse coding for glaucoma diagnosis
He et al. A selective overview of feature screening methods with applications to neuroimaging data
Marinescu et al. A vertex clustering model for disease progression: application to cortical thickness images
Voon et al. Evaluating the effectiveness of stain normalization techniques in automated grading of invasive ductal carcinoma histopathological images
CN115393362B (en) Method, equipment and medium for selecting automatic glaucoma recognition model
Wang et al. Signal subgraph estimation via vertex screening
Fernández et al. Diffusion methods for aligning medical datasets: Location prediction in CT scan images
Kamraoui et al. Popcorn: Progressive pseudo-labeling with consistency regularization and neighboring
Yesilbek11 et al. SVM-based sketch recognition: which hyperparameter interval to try?
Islam et al. Unic-net: Uncertainty aware involution-convolution hybrid network for two-level disease identification
Ghoshal et al. Bayesian deep active learning for medical image analysis

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