CN114898872A - Multi-mode probability distribution self-adaptive primary liver cancer pathological grading prediction method - Google Patents

Multi-mode probability distribution self-adaptive primary liver cancer pathological grading prediction method Download PDF

Info

Publication number
CN114898872A
CN114898872A CN202210390976.9A CN202210390976A CN114898872A CN 114898872 A CN114898872 A CN 114898872A CN 202210390976 A CN202210390976 A CN 202210390976A CN 114898872 A CN114898872 A CN 114898872A
Authority
CN
China
Prior art keywords
source
classification
nest
domain
bird
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210390976.9A
Other languages
Chinese (zh)
Inventor
崔恩铭
冯宝
刘昱
马长宜
龙晚生
徐坤财
侍江峰
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.)
Jiangmen Central Hospital
Guilin University of Aerospace Technology
Original Assignee
Jiangmen Central Hospital
Guilin University of Aerospace Technology
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 Jiangmen Central Hospital, Guilin University of Aerospace Technology filed Critical Jiangmen Central Hospital
Priority to CN202210390976.9A priority Critical patent/CN114898872A/en
Publication of CN114898872A publication Critical patent/CN114898872A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • 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
    • 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
    • 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
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • 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
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10081—Computed x-ray tomography [CT]
    • 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
    • G06T2207/30056—Liver; Hepatic
    • 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)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Medical Informatics (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Databases & Information Systems (AREA)
  • Multimedia (AREA)
  • Public Health (AREA)
  • Probability & Statistics with Applications (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Evolutionary Biology (AREA)
  • Pathology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to the technical field of CT image processing, in particular to a multi-modal probability distribution self-adaptive primary liver cancer pathology grading prediction method, which comprises the following steps: s1, a multi-source migration feature extraction network based on feature distribution dynamic alignment and classification layer alignment adopts a dynamic probability distribution measurement method, fully considers edge distribution and condition distribution difference of multi-source medical data, and realizes fine-grained alignment migration of multi-source heterogeneous features; s2, classification of multi-source migration features is achieved based on a liver MR image feature classification algorithm of the improved cuckoo optimization extreme learning machine. According to the method, on the basis of effectively solving the heterogeneous problem of multi-source data, multi-source medical data are fully fused, feature information is enriched, more robust and effective multi-source migration features are extracted, the number of hidden layer nodes of the extreme learning machine is subjected to self-adaptive optimization by adopting an improved cuckoo algorithm, a more stable and accurate classifier is trained, and effective classification of the multi-source migration features is realized.

Description

Multi-mode probability distribution self-adaptive primary liver cancer pathological grading prediction method
Technical Field
The invention relates to the technical field of CT image processing, in particular to a multi-modal probability distribution self-adaptive primary liver cancer pathology grading prediction method.
Background
The primary liver cancer is the fourth leading cause of cancer-related death in China, is the fourth most common malignant tumor and the second most cause of tumor death in China, and 50% of death cases occur in China globally, wherein Hepatocellular Carcinoma (HCC) accounts for 85% -90%.
Surgical resection is the most effective treatment method for HCC patients, and even though the diagnosis and treatment of liver cancer make many breakthrough progresses, the overall survival rate of liver cancer patients is still very low, the 5-year survival rate is only 18%, and the high recurrence rate after surgery is one of the most main factors for reducing the overall survival rate. According to statistics, the tumor recurrence and metastasis rate after 5 years of hepatoma resection reaches 40-70%.
The histopathological grading of HCC tumor is closely related to the prognosis of the patient, and is an important index for selecting a treatment scheme and evaluating survival prognosis. Most patients with high-grade HCC tumors are reported to have a higher risk of recurrence, a greater margin of safety is required for surgical resection, and a higher frequency of follow-up examinations is required after treatment; while low-grade HCC patients have a lower risk of recurrence. Therefore, accurate prediction of HCC pathological grading before treatment facilitates selection of treatment strategies for liver cancer patients and realization of accurate treatment for individuals.
At present, deep learning becomes the first-developed computer-aided diagnosis technology in medical image diagnosis classification due to its excellent characteristic learning ability, and is continuously applied to classification and prognosis of liver cancer. However, the performance of the deep learning model depends on a large amount of training data, and in clinical practice, the performance is limited by various conditions, such as different and extremely dispersed disease incidence, data existence manufacturers of multiple hospitals, non-uniform scanning parameters, and the like, so that the effective data set of the medical image is small. In order to improve the deep learning diagnosis performance under a small sample of medical data, the transfer learning technology is widely applied, wherein the transfer learning paradigm of performing parameter fine tuning on target data after pre-training based on an ImageNet source domain data set is the most common. However, studies have shown that the similarity between the source domain and the target domain is a key factor for determining the migration effect, and when the source domain and the target domain are low or irrelevant, the migration learning based on the pre-training fine tuning paradigm may be very inefficient. In addition, migration is performed based on the ImageNet single-source field, and the model mainly learns the basic texture features of the natural image, so that the discriminability of the model is probably mainly prone to only single-source-field representation, and the generalization performance is poor. In consideration of clinical practice, doctors often need to perform disease diagnosis by means of multiple means, for example, lesion screening is performed based on multi-sequence CT Images such as arterial phase and venous phase, then pathological diagnosis is performed based on full-field digital slice Images (WSI), and whether multi-modal data is used as source domain data for migration is better.
The introduction of multi-modal data brings rich features and must bring complex differences, and in order to reduce the differences among multi-source data and better perform feature migration, the multi-source migration method mainly has two learning paradigms: the method comprises the following steps of firstly, transforming and migrating the data based on geometric features and secondly, transforming and migrating statistical features, wherein the geometric feature transformation mainly focuses on the empty geometric space structure of the data, and the difference among multi-source domain data is eliminated by constructing a proper transformation subspace. For example, Li et al use a large open-source medical data set and a small local medical data set as source domain data, construct a multi-source domain feature subspace based on a singular value decomposition method, and then measure the distribution difference between the source domain data and the target domain data by Bregman divergence in the feature subspace, thereby implementing the assisted diagnosis of Alzheimer's disease. Zhang et al use multi-view contrast-enhanced ultrasound images (including arterial phase, venous phase, and lag phase) as source domain data, and B-mode ultrasound images as target domain data to achieve liver cancer diagnosis. The research designs a plurality of kernel functions aiming at different source domain data, maps the data to different high-dimensional spaces, and then measures the characteristic difference between a source domain and a target domain based on the mean square Euclidean distance to realize multi-source data migration. Compared with a geometric feature transformation method, the feature transformation method based on statistics is more widely applied. For example, Fang et al propose a Multi-source integrated migration learning framework (Multi-LSTM-DANN) that implements selective migration of Multi-source data based on a Maximum Mean Difference (MMD) metric of marginal probability distribution between different data domains. The edge Distribution and the condition Distribution of the complex multi-source data of Li et al and Wang are simultaneously evaluated by Joint Distribution Adaptation (JDA), so that the Distribution difference among the data is reduced, and the classification of medical images is realized.
The research is based on a multi-source migration technology, multi-modal medical data are fully utilized, more dimensional feature information is mined, and the performance of the model is improved, but the following problems still exist: (1) in the face of multi-modal medical images, a specific subspace transformation matrix is tried to be found, multi-source domain features are mapped to a public feature space and then are aligned, the dimensionality of subspace basis vectors is usually required to be set manually, and loss of data information in different degrees can be caused due to the fact that the multi-source domain features are subjective and different basis vector dimensions are provided. (2) In the face of multi-source heterogeneous images, the single probability distribution of the images is simply considered, or the edge distribution and condition distribution difference of data are aligned in the same position, the edge distribution (concerning a data generation mechanism) and condition distribution (concerning a specific downstream task) difference of data cannot be adaptively measured, and fine-grained classification migration is carried out.
Disclosure of Invention
In order to solve the problems, the invention provides a multi-mode probability distribution self-adaptive primary liver cancer pathological grading prediction method, which is characterized in that multi-source medical data are fully fused on the basis of effectively solving the heterogeneous problem of the multi-source data, feature information is enriched, more robust and effective multi-source migration features are extracted, an improved cuckoo algorithm is adopted to carry out self-adaptive optimization on the number of hidden layer nodes of an extreme learning machine, a more stable and accurate classifier is trained, and effective classification of the multi-source migration features is realized.
In order to achieve the purpose, the invention adopts the technical scheme that:
the multi-modal probability distribution self-adaptive primary liver cancer pathological grading prediction method comprises the following steps:
s1, a multi-source migration feature extraction network based on feature distribution dynamic alignment and classification layer alignment adopts a dynamic probability distribution measurement method, fully considers edge distribution and condition distribution difference of multi-source medical data, and realizes fine-grained alignment migration of multi-source heterogeneous features;
s2, classification of multi-source migration features is achieved based on a liver MR image feature classification algorithm of the improved cuckoo optimization extreme learning machine.
Further, in step S1, first, the multi-source data is input into the common feature space extraction F (), the common space output features are continuously mapped to the corresponding specific domain feature space h (), the probability distribution difference between the source domain and the target domain is dynamically aligned in the specific domain feature space, and the specific feature expression of each source domain and each target domain is found
Figure BDA0003595539960000043
Training domain-specific classifiers based on the specific features
Figure BDA0003595539960000044
Making a prediction of the target domain; and finally, constructing a total objective function according to a structure risk minimization principle by minimizing the distribution difference of the source domain and the target domain characteristics and the difference between classifiers of all classification layers as follows:
Figure BDA0003595539960000041
wherein f represents the total error of the network, the first term on the right side of the equal sign represents the probability distribution difference loss between the source domain and the target domain in the specific domain feature space, the second term represents the difference loss between the predicted values of the classifiers corresponding to different domains, the third term represents the cross entropy classification loss of the specific domain classifier, the fourth term represents the 2-norm constraint term of f, and alpha, beta and gamma are corresponding regularization parameters.
Further, in step S1, ResNet50 pre-trained based on ImageNet is used as a common feature extraction network, then common space output features are continuously mapped to a corresponding specific domain feature space, and the maximum mean difference is used in the specific domain feature space to dynamically measure the probability distribution difference between the source domain data and the target domain data, so that the input features of the classification layer meet the independent same distribution condition of the prediction model application.
Further, in step S1, the target domain is classified using the softmax function as the domain-specific classifier, and then its classification loss is recorded based on the cross entropy, that is:
Figure BDA0003595539960000042
wherein x represents the original target domain data, y represents its corresponding label, J () represents the cross entropy loss function, h j (F (x)) represents that the specific domain characteristics corresponding to the jth group of source domains extract the target domain image characteristics output by the sub-network;
in order to reduce the error classification of target samples near class boundaries, the classification difference of different sub-classifiers aiming at the same target data is measured based on Euclidean distance, and a classification layer loss value f is constructed disc As shown in the following formula:
Figure BDA0003595539960000051
by combining the formulas (9), (10), (11) and the structure risk minimization theory, the total loss function of the feature extraction network can be obtained as follows:
Figure BDA0003595539960000052
further, in step S2, the basic steps of the liver MR image feature classification algorithm of the improved cuckoo optimization extreme learning machine are as follows:
(1) initializing OCS algorithm parameters, wherein the OCS algorithm parameters mainly comprise the number of bird nests, algorithm parameters, the initialization positions of the bird nests and the like;
(2) setting initial probability Pa of bird nest discovery, randomly generating a group of initial bird nests according to the value range of the parameters, wherein the positions of the bird nests represent the number of nodes of the hidden layer, the classification accuracy is used as the adaptability value of each position of the bird nest, and the corresponding bird nest when the position is highest is used as the current optimal position of the bird nest;
(3) and updating the rest bird nests by adopting Le' vy flight, and dynamically adjusting the walking step length according to the following formula:
b i ＝b min +(b max -b min )d i
in the formula, b i Represents the current ith nest walk step length, b max And b min Respectively representing a maximum step size and a minimum step size, d i Is defined as follows:
Figure BDA0003595539960000053
wherein d is max Representing the maximum distance, s, of the optimal position from the position of the remaining bird's nest i Indicating the current ith bird nest position, s best Representing the current optimal bird nest position;
adding inertial weight to the nest-searching path and position updating position of cuckoos according to the following formula to obtain a group of new superior bird nest positions;
Figure BDA0003595539960000054
wherein τ is an inertial weight;
(4) with random numbers r ∈ [0,1 ] subject to uniform distribution]Comparing with Pa, retaining the bird nest with lower probability of finding, and randomly changing the bird nest with higher probability of finding to obtain a group of new bird nest positions s g ；
(5) For bird nest position s according to g Performing Gaussian disturbance to obtain a group of new birdsNest position s g '，
nest' ik ＝nest ik +η·R(0,1)
Wherein nest ik Is the k-dimensional variable of the individual i, R (0,1) is the sum of nest ik A standard gaussian variation random matrix of the same order, η being a coefficient, is used to control the search range of R (0,1), typically 0.618;
calculating its corresponding fitness value and comparing it with s g The fitness of each bird nest is compared, and a more optimal bird nest position is selected as s g ”；
(6) Judging the termination condition of the algorithm, solving the optimal bird nest and the corresponding adaptability value thereof, and outputting the optimal bird nest position as the optimal ELM hidden layer node number if the adaptability value meets the requirement; otherwise, repeating the steps (3) and (6) and continuing to execute;
(7) setting the number of nodes of an ELM hidden layer according to the optimal bird nest position, training a training set to obtain a classification model, predicting and verifying a prediction set, and solving the classification accuracy.
According to the method, on the basis of effectively solving the heterogeneous problem of multi-source data, multi-source medical data are fully fused, feature information is enriched, more robust and effective multi-source migration features are extracted, the number of hidden layer nodes of the extreme learning machine is subjected to self-adaptive optimization by adopting an improved cuckoo algorithm, a more stable and accurate classifier is trained, and effective classification of the multi-source migration features is realized.
Drawings
Other features, objects and advantages of the invention will become more apparent upon reading of the detailed description of non-limiting embodiments with reference to the following drawings:
fig. 1 is a CT image classification algorithm structure of an improved cuckoo optimization extreme learning machine based on dynamic alignment network features under multi-source heterogeneous data according to an embodiment of the present invention.
FIG. 2 is a multi-source migration feature extraction network framework based on feature distribution dynamic alignment and classification layer alignment in an embodiment of the present invention.
FIG. 3 illustrates specific convolution feature extraction in an embodiment of the present invention.
Fig. 4 is a liver CT classification model framework based on OCS optimization ELM in the embodiment of the present invention.
Detailed Description
The present invention will be described in detail with reference to specific examples. The following examples will assist those skilled in the art in further understanding the invention, but are not intended to limit the invention in any way. It should be noted that variations and modifications can be made by persons skilled in the art without departing from the concept of the invention. All falling within the scope of the present invention.
The structure of a CT image classification algorithm of an improved cuckoo optimization extreme learning machine based on dynamic alignment network characteristics under multi-source heterogeneous data is shown in FIG. 1, and the CT image classification algorithm mainly comprises a multi-source transfer characteristic extraction network based on characteristic distribution dynamic alignment and classification layer alignment and a characteristic classification algorithm based on the improved cuckoo optimization extreme learning machine.
1.1 Multi-Source migration feature extraction network based on feature distribution dynamic alignment and Classification level alignment
As shown in FIG. 2, assume that there are N sets of source domain medical data
Figure BDA0003595539960000071
And a set of target domain medical data D t ＝{X t ,Y t Therein of
Figure BDA0003595539960000072
And
Figure BDA0003595539960000073
indicating the data and label corresponding to the jth group of source domain samples,
Figure BDA0003595539960000074
and
Figure BDA0003595539960000075
representing the target domain sample and the label. The invention aims to firstly input multi-source data into a public characteristic space to extract F (), and continuously map the output characteristics of the public space toCorresponding specific domain feature space h (), dynamically aligning the probability distribution difference of the source domain and the target domain in the specific domain feature space, and searching the specific feature expression of each source domain and each target domain
Figure BDA0003595539960000076
Training domain-specific classifiers based on the specific features
Figure BDA0003595539960000077
A prediction is made of the target domain. And finally, constructing a total objective function according to a structure risk minimization principle by minimizing the distribution difference of the source domain and the target domain characteristics and the difference between classifiers of all classification layers as follows:
Figure BDA0003595539960000081
wherein f represents the total error of the network, the first term on the right side of the equal sign represents the probability distribution difference loss between the source domain and the target domain in the specific domain feature space, the second term represents the difference loss between the predicted values of the classifiers corresponding to different domains, the third term represents the cross entropy classification loss of the specific domain classifier, the fourth term represents the 2-norm constraint term of f, and alpha, beta and gamma are corresponding regularization parameters.
1.1.1 Domain-specific feature probability distribution dynamic alignment
The method takes ResNet50 pre-trained based on ImageNet as a public feature extraction network, then continuously maps the output features of a public space to a corresponding specific domain feature space, and then dynamically measures the probability distribution difference between source domain data and target domain data by adopting Maximum Mean value difference (MMD) in the specific domain feature space, so that the input features of a classification layer meet the independent same distribution condition applied by a prediction model.
The basic principle is as follows: there is a set of source domain data that satisfies a p probability distribution
Figure BDA0003595539960000082
And satisfying q target domain data
Figure BDA0003595539960000083
Then X s And X t The MMD of (A) is defined as:
Figure BDA0003595539960000084
where H denotes a regenerative nuclear Hilbert space (RKHS), and n ═ X s I denotes a set of source domain samples, m ═ X t I represents the number of samples in the target domain, phi () represents a mapping function that maps from the original space to RKHS, satisfying<φ(x)，φ(y)> H K (x, y), where k (x, y) is a gaussian kernel function, i.e.
k(x,y)＝exp(-||x-y|| 2 /2σ 2 ) (3)
Wherein, σ represents the size of the gaussian kernel, and in order to obtain a better measurement result, the invention uses a weighted average of a plurality of gaussian kernel functions to obtain a final result. The following equations (2) and (3) are combined to obtain:
Figure BDA0003595539960000091
in clinical practice, the complex multi-source data has large difference in probability distribution, for example, the medical data (CT and WSI data) generated by different mechanisms has large difference in marginal probability distribution, and the medical data corresponding to different learning tasks (liver cancer detection and lung cancer detection) has large difference in conditional probability distribution. In order to fully adapt to the edge distribution and the condition distribution of the data according to the data information, the MMD is taken as a basic measuring tool of the data distribution, a dynamic probability distribution measuring method is adopted, and the edge distribution and the condition distribution of the data are measured at the same time, and are defined as follows:
Figure BDA0003595539960000092
wherein, mu belongs to [0,1 ]]Representing a probabilistic adaptation factor, [ epsilon ] {0, 1} tableSample class, p (X) s ) And q (X) t ) Representing the edge distribution of source domain data and target domain data, p (Y) s |X s ) And q (Y) t |X t ) Representing the conditional distribution of the source domain data and the target domain data.
When mu → 0, the difference between the data of the source domain and the data of the target domain is larger, and the edge distribution adaptation is more important; when mu → 1, the similarity of the data from the source and the target domain is higher, the data distribution among the categories is more important, and the condition distribution needs to be analyzed with emphasis; when μ → 0.5, the edge distribution and conditional distribution representing data are equally important, and joint probability domain adaptation (JDA) mainly studies this work. Different from JDA, the invention adopts Wasserstein distance as the measuring means of data probability distribution to calculate the weight occupied by the edge distribution and the condition distribution between the source domain data and the target domain data and dynamically measure two distributions of the data.
The Wasserstein distance is based on the optimal transmission theory, aiming at adapting the difference between the probability distributions of the source domain and the target domain data at a minimum cost, and is defined as follows:
Figure BDA0003595539960000093
where Π (p, q) represents the set of all joint distributions that correspond to the combination of p and q distributions, E (x,y)ν And the lower bound of the expected value in all possible joint distributions is the Wasserstein distance.
Calculating the global Wasserstein distance of the source domain data and the target domain data based on the formula (6) as the weight occupied by the edge probability, and recording the weight as W g (ii) a Source domain X s And a target domain X t The Wasserstein distance of the middle l-th class data is taken as the conditional probability distribution of the data and is marked as W l Wherein W is l ＝W(Xs (l) ,X t(l) ) And further calculating a probability adaptation factor mu as shown in the following formula:
Figure BDA0003595539960000101
combining the formulas (5) and (7), obtaining the final specific domain feature probability distribution difference dynamic measurement method, which is shown as follows:
Figure BDA0003595539960000102
then the mean of the feature distribution differences corresponding to the N sets of source domain data is:
Figure BDA0003595539960000103
the probability adaptation factor starts from local information and global information of data, comprehensively considers conditional probability distribution and marginal probability distribution of source domain data and target domain data, balances data distribution difference caused by different data generation mechanisms (CT and WSI) and downstream tasks (liver cancer diseases and lung cancer diseases), and obtains more robust specific domain data characteristics.
1.1.2 Domain-specific classifier alignment
Based on the characteristics, the integrated learning idea is referred to, and N groups of sub-classifiers are trained by using the characteristics corresponding to N groups of source domain data
Figure BDA0003595539960000104
And respectively predicting the target data. The invention uses the softmax function as a specific domain classifier to classify the target domain, and then records the classification loss based on the cross entropy, namely:
Figure BDA0003595539960000105
wherein x represents the original target domain data, y represents its corresponding label, J () represents the cross entropy loss function, h j (F (x)) showing that specific domain feature extraction sub-network circuit output corresponding to jth group of source domainsTarget domain image features.
In addition, in order to reduce the error classification of target samples near class boundaries, the classification difference of different sub-classifiers for the same target data is measured based on Euclidean distance, and a classification layer loss value f is constructed disc As shown in the following formula:
Figure BDA0003595539960000111
by combining the formulas (9), (10), (11) and the structure risk minimization theory, the total loss function of the feature extraction network can be obtained as follows:
Figure BDA0003595539960000112
based on the total objective function, the feature extraction network is trained, and the output feature map of each convolution kernel in the network model represents the specific features of the liver cancer and the liver lymphoma. As shown in fig. 3, this feature consists of 22720 public domain convolution features and 2304 specific domain convolution features. In order to reduce feature Redundancy and improve the operation speed of a classification model, the top 10% of features with higher label correlation are selected for classification by adopting a Maximum correlation Minimum Redundancy (mRMR) algorithm.
1.2 liver MR image feature classification algorithm of improved cuckoo optimization extreme learning machine
1.2.1 extreme learning machine
Compared with the traditional feedback neural network, the learning precision and convergence rate of the extreme learning machine ELM based on small samples are improved. The basic principle is as follows: for L training samples
Figure BDA0003595539960000113
Wherein x i ＝[x i1 ,...,x in ] T ∈R n For input data, the target expected output value is o i ＝[o i1 ,...,o im ] T ∈R m The ELM output expression with K hidden layer neurons is:
Figure BDA0003595539960000114
wherein, y i ∈R N Is the actual output of the network for the ith case, ω i As the connection weight vector of the input layer neuron and the ith hidden layer neuron, b i For the biasing of the ith hidden layer neuron,
Figure BDA0003595539960000115
a connecting weight vector, ω, for the ith hidden layer neuron and the output neuron i x j Represents omega i And x j Is the activation function of the hidden layer, g ().
When the actual output can approach the desired output value of the network with zero error, i.e.
Figure BDA0003595539960000116
This time is:
Figure BDA0003595539960000117
for simplicity of representation, the matrix of equation (14) is represented as:
Figure BDA0003595539960000121
wherein G denotes an output matrix of the hidden layer node,
Figure BDA0003595539960000122
is the output layer weight matrix, O is the desired output vector of the neural network:
Figure BDA0003595539960000123
Figure BDA0003595539960000124
the above equation is equivalent to solving a linear system
Figure BDA0003595539960000125
The minimum 2-norm least square solution of (a) is, according to the generalized inverse theory:
Figure BDA0003595539960000126
wherein the content of the first and second substances,
Figure BDA0003595539960000127
is a Moore-Penrose generalized inverse matrix of G,
Figure BDA0003595539960000128
the ELM theory is based on an empirical risk minimization theory, and the network is easy to overfit and poor in stability. To improve model performance, based on the structure risk minimization theory, the present invention uses L 1 The norm regularization carries out sparse constraint on the output weight matrix, the robustness of the model is improved, and ELM solution is as follows:
Figure BDA0003595539960000129
solving equation (18) comprehensively:
Figure BDA00035955399600001210
where ζ is the regularization coefficient and I is the identity matrix.
Theoretically, when the number of hidden layer nodes and the number of samples of the ELM model are equal, the network can approximate the learned samples with zero error. However, due to the ELM network input layer weights ω i And is hiddenContaining a layer bias b i Due to the adoption of random generation, part of nodes fail or cannot meet the data requirement, and finally, the problems that the output of the ELM is easy to generate random fluctuation, the stability and the generalization capability are not ideal and the like are caused. Therefore, the adaptive selection of the number of hidden layer nodes is crucial to improving the model performance.
1.2.2 improved cuckoo algorithm
The invention adopts an improved cuckoo algorithm to adaptively solve the number of nodes of the ELM hidden layer, and the bird nest position represents the number of the nodes of the hidden layer, thereby constructing a more robust and effective classifier.
To simulate nesting behavior of cuckoos, 3 ideal states are set: (1) cuckoos lay only one egg at a time and randomly select a parasitic nest position to hatch it; (2) of a randomly selected set of nests, the best nest will be retained to the next generation; (3) the number n of available nests is fixed, and the probability that a nest owner can find a foreign bird egg is P a ∈[0,1]. On the basis of 3 ideal states, the position updating formula of the cuckoo nest searching is as follows:
Figure BDA0003595539960000131
wherein the content of the first and second substances,
Figure BDA0003595539960000132
and
Figure BDA0003595539960000133
indicating the positions of the ith bird nest in the z +1 th and z th generations; delta is step control quantity;
Figure BDA0003595539960000134
is a dot product; l (lambda) is a cuckoo random walk search mode, which obeys Le' vy distribution:
Le′vy～u＝t -λ 1≤λ≤3(22)
after the position is updated, a random number r belongs to [0,1 ]]And P a By contrast, if r > P a Then to s i z Go on and followThe machine changes, otherwise it does not. And finally reserving a group of bird nest positions with better test values.
However, the standard CS algorithm employs Le' vy flight to randomly generate the step size, which is not favorable for calculation, and a smaller step size will reduce the search speed, and a larger step size will reduce the search accuracy, and is lack of adaptivity. Aiming at the defects of low search speed, low optimization precision and the like of the CS algorithm, a self-adaptive step length adjustment strategy is introduced, meanwhile, the inertial weight combined with a Gaussian disturbance strategy is added to the CS algorithm, the search space is expanded, the search strength and speed of the algorithm are improved, the diversity of the change of the position of a bird nest is increased, and an optimized cuckoo search algorithm (OCS) is provided. Firstly, the adaptive step size adjustment strategy is as follows:
b i ＝b min +(b max -b min )d i (23)
in the formula, b i Represents the current ith nest walk step length, b max And b min Respectively representing a maximum step size and a minimum step size, d i Is defined as follows:
Figure BDA0003595539960000141
wherein d is max Representing the maximum distance, s, of the optimal position from the position of the remaining bird's nest i Indicating the current ith bird nest position, s best Indicating the current optimal bird nest position.
The method for adding the inertial weight is to introduce an inertial weight nonlinear decrement strategy in a brook nest searching path and position updating formula, namely:
Figure BDA0003595539960000142
Figure BDA0003595539960000143
τ is the inertial weight, i epoch Is the current iterationThe generation number. By introducing inertial weight, the search space of the CS algorithm can be expanded to a certain extent, so that the search space is more capable of searching a new region. The bigger the tau is, the more easily the CS algorithm jumps out the local optimum to carry out the global optimization; the smaller tau is, the better the CS algorithm is to carry out local optimization, thereby the convergence speed is faster. Meanwhile, in order to balance the global and local search capabilities of the algorithm, the value of the inertial weight τ is decreased as the number of iterations increases.
The specific method of the Gaussian disturbance strategy is as follows: after the step of finding the foreign bird eggs by the parasitic nest owner is finished, adding a one-dimensional Gaussian disturbance factor to the position of the parasitic nest, wherein the calculation method is as follows:
nest' ik ＝nest ik +η·R(0,1) (27)
wherein nest ik Is the k-dimensional variable of the individual i, R (0,1) is the sum of nest ik The standard gaussian variance random matrix of the same order, η is a coefficient, which is used to control the search range of R (0,1), and is usually 0.618. Adding Gaussian disturbance to obtain new nest' ik Then with nest ik Comparing every nest, reserving the position of the nest with better test value, and entering the next iteration.
1.2.3 liver MR image feature classification algorithm of improved cuckoo optimization extreme learning machine
Because ELM adopts random generation of input layer weight omega i And hidden layer bias b i And part of hidden layer nodes are caused to fail or not meet data distribution, and finally, the problems that the output of the ELM is easy to generate random fluctuation, the stability and the generalization capability are not ideal and the like are caused. Aiming at the problem, the method is different from the conventional empirical selection of the number of hidden layer nodes, adopts an OCS algorithm to adaptively select the number of hidden layer nodes, improves the stability and generalization capability of the model, and provides a liver CT classification model based on OCS optimized ELM, wherein a specific frame diagram is shown in FIG. 4:
the basic steps of the OCS-ELM classification model are as follows:
(1) initializing OCS algorithm parameters, wherein the OCS algorithm parameters mainly comprise the number of bird nests, algorithm parameters, the initialization positions of the bird nests and the like;
(2) setting initial probability Pa of bird nest discovery, randomly generating a group of initial bird nests according to the value range of the parameters, wherein the positions of the bird nests represent the number of nodes of the hidden layer, the classification accuracy is used as the adaptability value of each position of the bird nest, and the corresponding bird nest when the position is highest is used as the current optimal position of the bird nest;
(3) updating the rest bird nests by adopting Le' vy flight, dynamically adjusting the walking step length according to the formula (23), and adding inertia weight to the nest searching path and position updating position of the cuckoo according to the formula (25) to obtain a group of new better bird nest positions;
(4) with random numbers r ∈ [0,1 ] subject to uniform distribution]Comparing with Pa, retaining the bird nest with lower probability of finding, and randomly changing the bird nest with higher probability of finding to obtain a group of new bird nest positions s g ；
(5) According to the formula (27) for the bird nest position s g Gaussian disturbance is carried out to obtain a group of new bird nest positions s g ', calculating its corresponding fitness value and corresponding to s g The fitness of each bird nest is compared, and a more optimal bird nest position is selected as s g ”；
(6) Judging the termination condition of the algorithm, solving the optimal bird nest and the corresponding adaptability value thereof, and outputting the optimal bird nest position as the optimal ELM hidden layer node number if the adaptability value meets the requirement; otherwise, repeating the steps (3) and (6) and continuing to execute;
(7) setting the number of nodes of an ELM hidden layer according to the optimal bird nest position, training a training set to obtain a classification model, predicting and verifying a prediction set, and solving the classification accuracy.
The foregoing description has described specific embodiments of the present invention. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes and modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention.

Claims (5)

1. The multi-modal probability distribution self-adaptive primary liver cancer pathology grading prediction method is characterized by comprising the following steps of: the method comprises the following steps:
s1, a multi-source migration feature extraction network based on feature distribution dynamic alignment and classification layer alignment adopts a dynamic probability distribution measurement method, fully considers edge distribution and condition distribution difference of multi-source medical data, and realizes fine-grained alignment migration of multi-source heterogeneous features;
s2, classification of multi-source migration features is achieved based on a liver MR image feature classification algorithm of the improved cuckoo optimization extreme learning machine.
2. The method of claim 1, wherein the multi-modal probability distribution adaptive primary liver cancer pathology grading prediction method comprises: in step S1, first, the multi-source data is input into the common feature space extraction F (), the common space output features are continuously mapped to the corresponding specific domain feature space h (), the probability distribution differences between the source domain and the target domain are dynamically aligned in the specific domain feature space, and the specific feature expression of each source domain and each target domain is found
Figure FDA0003595539950000012
Training domain-specific classifiers based on the specific features
Figure FDA0003595539950000013
Making a prediction of the target domain; and finally, constructing a total objective function according to a structure risk minimization principle by minimizing the distribution difference of the source domain and the target domain characteristics and the difference between classifiers of all classification layers as follows:
Figure FDA0003595539950000011
wherein f represents the total error of the network, the first term on the right side of the equal sign represents the probability distribution difference loss between the source domain and the target domain features in the specific domain feature space, the second term represents the difference loss between the predicted values of the classifiers corresponding to different domains, the third term represents the cross entropy classification loss of the specific domain classifier, the fourth term represents the 2-norm constraint term of f, and alpha, beta and gamma are corresponding regularization parameters.
3. The method of claim 1, wherein the multi-modal probability distribution adaptive primary liver cancer pathology grading prediction method comprises: in step S1, ResNet50 pre-trained based on ImageNet is used as a common feature extraction network, then common space output features are continuously mapped to a corresponding specific domain feature space, and the maximum mean difference is used in the specific domain feature space to dynamically measure the probability distribution difference between the source domain data and the target domain data, so that the input features of the scoring layer satisfy the independent same distribution condition of the prediction model application.
4. The method of claim 1, wherein the multi-modal probability distribution adaptive primary liver cancer pathology grading prediction method comprises: in step S1, the target domain is classified by using the softmax function as the specific domain classifier, and then the classification loss is recorded based on the cross entropy, that is:
Figure FDA0003595539950000021
wherein x represents the original target domain data, y represents its corresponding label, J () represents the cross entropy loss function, h j (F (x)) represents that the specific domain characteristics corresponding to the jth group of source domains extract the target domain image characteristics output by the sub-network;
in order to reduce the error classification of target samples near class boundaries, the classification difference of different sub-classifiers aiming at the same target data is measured based on Euclidean distance, and a classification layer loss value f is constructed disc As shown in the following formula:
Figure FDA0003595539950000022
by combining the formulas (9), (10), (11) and the structure risk minimization theory, the total loss function of the feature extraction network can be obtained as follows:
Figure FDA0003595539950000023
5. the method of claim 1, wherein the multi-modal probability distribution adaptive primary liver cancer pathology grading prediction method comprises: in step S2, the basic steps of the liver MR image feature classification algorithm of the improved cuckoo optimization extreme learning machine are as follows:
(1) initializing OCS algorithm parameters, including the number of bird nests, algorithm parameters and the initialized positions of the bird nests;
(2) setting initial probability Pa of bird nest discovery, randomly generating a group of initial bird nests according to the value range of the parameters, wherein the positions of the bird nests represent the number of nodes of the hidden layer, the classification accuracy is used as the adaptability value of each position of the bird nest, and the corresponding bird nest when the position is highest is used as the current optimal position of the bird nest;
(3) and updating the rest bird nests by adopting Le' vy flight, and dynamically adjusting the walking step length according to the following formula:
b i ＝b min +(b max -b min )d i
in the formula, b i Represents the current ith nest walk step length, b max And b min Respectively representing a maximum step size and a minimum step size, d i Is defined as follows:
Figure FDA0003595539950000031
wherein d is max Representing the maximum distance, s, of the optimal position from the position of the remaining bird's nest i Indicating the current ith bird nest position, s best Representing the current optimal bird nest position;
adding inertial weight to the nest searching path and position updating position of the cuckoo according to the following formula to obtain a group of new superior bird nest positions;
Figure FDA0003595539950000032
wherein τ is an inertial weight;
(4) with random numbers r ∈ [0,1 ] subject to uniform distribution]Comparing with Pa, retaining the bird nest with lower probability of finding, and randomly changing the bird nest with higher probability of finding to obtain a group of new bird nest positions s g ；
(5) For bird nest position s according to g Gaussian disturbance is carried out to obtain a group of new bird nest positions s g' ，
nest' ik ＝nest ik +η·R(0,1)
Wherein nest ik Is the k-dimensional variable of the individual i, R (0,1) is the sum of nest ik A standard Gaussian variation random matrix of the same order, wherein eta is a coefficient and is used for controlling the search range of R (0, 1);
calculating its corresponding fitness value and comparing it with s g The fitness of each bird nest is compared, and a more optimal bird nest position is selected as s g” ；
(6) Judging the termination condition of the algorithm, solving the optimal bird nest and the corresponding adaptability value thereof, and outputting the optimal bird nest position as the optimal ELM hidden layer node number if the adaptability value meets the requirement; otherwise, repeating the steps (3) and (6) and continuing to execute;
(7) setting the number of nodes of an ELM hidden layer according to the optimal bird nest position, training a training set to obtain a classification model, predicting and verifying a prediction set, and solving the classification accuracy.
CN202210390976.9A 2022-04-14 2022-04-14 Multi-mode probability distribution self-adaptive primary liver cancer pathological grading prediction method Pending CN114898872A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210390976.9A CN114898872A (en) 2022-04-14 2022-04-14 Multi-mode probability distribution self-adaptive primary liver cancer pathological grading prediction method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210390976.9A CN114898872A (en) 2022-04-14 2022-04-14 Multi-mode probability distribution self-adaptive primary liver cancer pathological grading prediction method

Publications (1)

Publication Number Publication Date
CN114898872A true CN114898872A (en) 2022-08-12

Family

ID=82717037

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210390976.9A Pending CN114898872A (en) 2022-04-14 2022-04-14 Multi-mode probability distribution self-adaptive primary liver cancer pathological grading prediction method

Country Status (1)

Country Link
CN (1) CN114898872A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117096070A (en) * 2023-10-19 2023-11-21 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) Semiconductor processing technology abnormality detection method based on field self-adaption
CN117407698A (en) * 2023-12-14 2024-01-16 青岛明思为科技有限公司 Hybrid distance guiding field self-adaptive fault diagnosis method

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117096070A (en) * 2023-10-19 2023-11-21 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) Semiconductor processing technology abnormality detection method based on field self-adaption
CN117096070B (en) * 2023-10-19 2024-01-05 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) Semiconductor processing technology abnormality detection method based on field self-adaption
CN117407698A (en) * 2023-12-14 2024-01-16 青岛明思为科技有限公司 Hybrid distance guiding field self-adaptive fault diagnosis method
CN117407698B (en) * 2023-12-14 2024-03-08 青岛明思为科技有限公司 Hybrid distance guiding field self-adaptive fault diagnosis method

Similar Documents

Publication Publication Date Title
Biswas et al. State-of-the-art review on deep learning in medical imaging
Madani et al. Fast and accurate view classification of echocardiograms using deep learning
Wang et al. Multiple graph regularized nonnegative matrix factorization
Dehuri et al. An improved swarm optimized functional link artificial neural network (ISO-FLANN) for classification
US8423596B2 (en) Methods of multivariate data cluster separation and visualization
CN114898872A (en) Multi-mode probability distribution self-adaptive primary liver cancer pathological grading prediction method
EP3570288A1 (en) Method for obtaining at least one feature of interest
US20210312242A1 (en) Synthetically Generating Medical Images Using Deep Convolutional Generative Adversarial Networks
Venugopal et al. Privacy preserving generative adversarial networks to model electronic health records
Kumar et al. Future of machine learning (ml) and deep learning (dl) in healthcare monitoring system
Zhang et al. Triplet attention and dual-pool contrastive learning for clinic-driven multi-label medical image classification
Rani et al. HIOC: a hybrid imputation method to predict missing values in medical datasets
Pradhan et al. A deep learning-based approach for detection of lung cancer using self adaptive sea lion optimization algorithm (SA-SLnO)
Misra et al. Deep learning‐based multimodal fusion network for segmentation and classification of breast cancers using B‐mode and elastography ultrasound images
CN117422964A (en) Rectal cancer prediction method, system and equipment based on multi-mode data fusion
Epifano et al. A comparison of feature selection techniques for first-day mortality prediction in the icu
CN114708347A (en) Lung nodule CT image classification method based on self-adaptive selection dual-source-domain heterogeneous migration learning
Pölsterl et al. Scalable, axiomatic explanations of deep Alzheimer’s diagnosis from heterogeneous data
Madadi et al. Stacking Ensemble Learning in Deep Domain Adaptation for Ophthalmic Image Classification
Chaithanyadas et al. Detection of pancreatic tumor from computer tomography images using 3D convolutional neural network
Leena et al. Automatic Brain Tumor Classification via Lion Plus Dragonfly Algorithm
Tripathy et al. Brain Tumour Detection Using Convolutional Neural Network-XGBoost
Sharma et al. Image segmentation of X-Ray and optical images using U-Net/UNet++ based deep learning architecture
Chelluboina et al. CATDSNet: Computer Aided Tongue Diagnosis System for Disease Prediction Using Hybrid Extreme Learning Machine.
Krebs Robust medical image registration and motion modeling based on machine learning

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