CN-121982419-A - Tumor medical image classification method based on double-index collaborative evaluation and zero sample evolution NAS

CN121982419ACN 121982419 ACN121982419 ACN 121982419ACN-121982419-A

Abstract

The invention provides a tumor medical image classification method of a zero sample evolution NAS based on double-index collaborative evaluation, which comprises the steps of collecting tumor medical image data, constructing a data set containing tumor focus images and normal tissue images, constructing an expanded search space based on a cell structure, setting steady-state evolution algorithm parameters, carrying out steady-state evolution initialization to obtain an initial candidate framework, using a double-index comprehensive score calculated by information rate and FIM stability as a comprehensive index for evaluating the current candidate framework, updating the candidate framework through the steady-state evolution algorithm based on a dynamic optimization mechanism until the maximum evolution algebra is reached, obtaining an optimal framework, training the optimal framework by utilizing the data set, and carrying out tumor medical image classification by utilizing the trained optimal framework. The invention reduces the dependence on the labeling data, reduces the consumption of computing resources, and improves the diagnosis efficiency and accuracy.

Inventors

YAN LI
WANG JINGLIN
LI CHAO
QU BOYANG
CHAI XUCHAO
LIU ZHONGYUN

Assignees

中原工学院

Dates

Publication Date: 20260505
Application Date: 20260205

Claims (10)

1. A tumor medical image classification method based on zero sample evolution NAS of double-index collaborative evaluation is characterized in that, S1, collecting tumor medical image data and constructing a data set containing tumor focus images and normal tissue images; S2, constructing an expanded search space based on a cell structure, wherein the search space comprises a basic operation set, the cell structure and a network integral structure; s3, setting steady-state evolution algorithm parameters, and carrying out steady-state evolution initialization to obtain initial candidate frameworks; S4, updating the candidate architecture by a steady state evolution algorithm based on a dynamic optimization mechanism by taking the double-index comprehensive score calculated by the information rate and the FIM stability as a comprehensive index for evaluating the current candidate architecture until the maximum evolution algebra is reached, and acquiring an optimal architecture; And S5, training the optimal architecture by using the data set obtained in the step S1, and classifying tumor medical images by using the trained optimal architecture.
2. The method for classifying tumor medical images based on zero sample evolution NAS with double-index collaborative evaluation according to claim 1, wherein the data set comprising tumor focus images and normal tissue images is constructed and comprises data preprocessing operations including image scaling, random cropping, random horizontal flipping, tensor conversion, normalization processing and Cutout operations.
3. The method for classifying tumor medical images based on zero sample evolution NAS with dual-index collaborative evaluation according to claim 2, wherein the basic operation set comprises at least 11 effective operations, namely 3X 3 average pooling operation, 3X 3 maximum pooling operation, jump connection operation, 3X 3 depth separable convolution, 5X 5 depth separable convolution, 3X 3 hole convolution, 5X 5 hole convolution, mobileNetV2 basic convolution, ECA attention convolution, CAM attention convolution and shuffle feature convolution.
4. The method for classifying tumor medical images based on zero-sample evolution NAS with dual-index collaborative evaluation according to claim 3, wherein the cell structure includes normal cells for preserving the feature map size and reduce cells for shrinking the feature map size, each of the normal cells and the reduce cells includes 4 computation nodes, each computation node generates an output feature by an operation combination of 2 precursor nodes, and finally the output features of the 4 computation nodes are spliced as an output of each cell.
5. The method for classifying tumor medical images based on zero sample evolution NAS of double-index collaborative evaluation according to claim 4, wherein the network overall structure sequentially comprises an input layer, a stem layer, a plurality of cells, global average pooling and a full connection layer, wherein the stem layer sequentially comprises 3x3 convolution and batch normalization operation, and the plurality of cells divide normal cells and reduce cells according to a preset proportion; The method comprises the steps of uniformly sampling different candidate frameworks with preset scales from a search space, performing gene coding on each candidate framework in a mode of normal cell operation+precursor node combination and reduce cell operation+precursor node combination, expressing precursor nodes by odd-numbered genes, and expressing cell operation by even-numbered genes.
6. The method for classifying tumor medical images based on zero sample evolution NAS for double-index collaborative evaluation according to any one of claims 1 to 5, wherein the method for calculating a double-index composite score is as follows: s411, calculating the energy duty ratio of the main component through singular value decomposition based on the initial weight matrix of each layer of the candidate architecture, and taking the energy duty ratio as an information rate index; s412, generating a random input sample simulating the noise distribution of the tumor medical image, acquiring a gradient vector through forward propagation and backward propagation, constructing a gradient matrix, calculating a condition number, and taking the reciprocal of the condition number as a FIM stability index; And S413, performing equal weight weighting on the information rate index and the FIM stability index to obtain a double-index comprehensive score.
7. The method for classifying tumor medical images based on zero sample evolution NAS with dual-index collaborative evaluation according to claim 6, wherein updating candidate architecture by a steady state evolution algorithm based on a dynamic optimization mechanism comprises: S421, parent selection, namely selecting 2 individuals from the current candidate architecture population randomly by adopting tournament selection, and selecting the individuals with higher double-index comprehensive scores as parent architectures; s422, crossover operation, namely randomly selecting connection gene crossover or operation gene crossover according to preset probability; s423, mutation operation, namely randomly selecting a plurality of gene positions according to preset probability to carry out connection gene mutation or operation gene mutation; S424, population updating, namely carrying out genetic validity verification on the current child framework generated after mutation, carrying out zero sample evaluation on the current child framework based on the double-index comprehensive score, obtaining the double-index comprehensive score of the current child framework, directly adding the evaluated current child framework into the current candidate framework population, eliminating 1 individual with the lowest double-index comprehensive score if the population size is larger than the preset size, and carrying out archiving updating to realize dynamic optimization updating of the population.
8. The method for classifying tumor medical images based on zero-sample evolution NAS with double-index collaborative evaluation according to claim 7, wherein the connecting genes are crossed, comprising randomly selecting a preset number of odd-numbered genes for exchange; The operation genes are crossed, comprising randomly selecting a preset number of even genes for exchange; The junction gene variation comprises random substitution within the range of [0, precursor node maximum index ]; the operator variation includes random substitution within the range of [0, maximum index of the operator set ].
9. The method for classifying tumor medical images based on zero sample evolution NAS with double-index collaborative evaluation according to claim 7 or 8, characterized in that the calculation of the principal component energy ratio by singular value decomposition as an information rate index based on the initial weight matrix of each layer of the candidate architecture includes: s4111, extracting a weight matrix of each layer in the neural network corresponding to the candidate architecture For neural network (N) Weighting matrix of layers Singular value decomposition is carried out; s4112, calculating total information capacity of the weight matrix based on singular values of the weight matrix: Wherein, the method comprises the steps of, Is the first of the weight matrix The number of singular values is chosen to be, The theoretical maximum rank of the weight matrix; S4113 before selection And (3) calculating the energy ratio of each main component: Wherein, the method comprises the steps of, Is the first The principal component energy ratio of the layer; s4114 energy duty cycle for all layers Taking the average value to obtain the information efficiency index of the whole candidate architecture: Wherein, the method comprises the steps of, Representing the total number of layers of the candidate architecture.
10. The method for classifying tumor medical images based on the zero sample evolution NAS of the double-index collaborative evaluation according to claim 7 or 8, wherein generating random input samples simulating noise distribution of tumor medical images, acquiring gradient vectors by forward propagation and backward propagation, constructing a gradient matrix and calculating condition number, taking the reciprocal of the condition number as FIM stability index, comprises: s4121 random generation The method comprises the steps of inputting noise samples, simulating tumor image noise distribution, carrying out forward propagation and backward propagation on each sample once to obtain gradient vectors, adopting a cross entropy loss function as a loss function, and then constructing a gradient matrix: ; Wherein, the The i-th input noise sample is the gradient vector obtained after forward propagation and backward propagation of the network, Is the mean vector of the N gradient vectors, The transpose is represented by the number, Is a gradient matrix after decentralization; s4122 pair gradient matrix Singular value decomposition is carried out; S4123 calculating maximum singular value And minimum non-zero singular values As a condition number ; S4124 against condition number And taking the reciprocal to obtain the FIM stability index.

Description

Tumor medical image classification method based on double-index collaborative evaluation and zero sample evolution NAS Technical Field The invention relates to the technical field of tumor medical image classification, in particular to a tumor medical image classification method. Background The tumor is a serious disease seriously threatening human health, and early accurate diagnosis is a key for improving survival rate of patients. The tumor medical image is a core basis for tumor screening and diagnosis, and the clinical needs to rely on a professional doctor to manually judge the size, shape, position and other characteristics of a tumor focus in the image. However, the process is time-consuming and labor-consuming, is influenced by factors such as experience of doctors, subjective judgment and the like, is easy to cause missed diagnosis and misdiagnosis, and is difficult to meet the large-scale screening requirement especially in areas with scarce medical resources. The development of the deep learning technology provides an effective path for automatic classification of tumor medical images, and models such as convolutional neural networks can learn focus features from massive labeling data, so that automatic classification is realized. However, the tumor medical image has the characteristics of various focus forms, fuzzy boundaries, large individual difference, rare labeling data, uneven data distribution and the like, and the high-performance neural network architecture is designed manually aiming at the specific task, so that the high-performance neural network architecture needs to have professional knowledge in the fields of deep learning and medical imaging at the same time, and the threshold is extremely high. And the manual design process needs a large number of trial and error experiments, and has the problems of large consumption of computing resources, long development period and the like. The neural network architecture searching technology can automatically explore the optimal network architecture through an algorithm without relying on manual priori knowledge, and provides a feasible scheme for solving the problems. The invention patent with publication number of CN113256593A discloses a tumor image detection method based on task self-adaptive neural network architecture search, and the task self-adaptive neural network architecture search method is used for constructing a self-adaptive neural network by using a small amount of tumor detection information, and a cascade multi-target detection network is combined, so that the problem of precision reduction of the deep convolution neural network in clinical medicine tumor image detection is solved, high-efficiency tumor image detection is realized, and the method is superior to the existing automatic detection method. However, the method still has the remarkable limitation that the traditional neural network architecture search needs to carry out complete training evaluation on candidate architectures, has high calculation cost, and is difficult to adapt to the scene of scarce tumor medical image data and high clinical requirements. Disclosure of Invention Aiming at the technical problems that the calculation cost is high and the application scene of the tumor medical image is difficult to adapt when the traditional tumor medical image classification adopts the neural network architecture for searching, the invention provides the tumor medical image classification method of the zero sample evolution NAS based on the double-index collaborative evaluation, the network architecture is not required to be designed manually, the evaluation can be completed only through initial weight and gradient information based on the innovative double-index collaborative evaluation strategy and the steady state evolution algorithm, the dependence on labeling data is reduced, the searching period is greatly shortened, the high-performance neural network of the automatic searching adaptation task can be realized, the accurate classification of the tumor medical image is realized, the clinician is assisted to rapidly judge the focus property, the diagnosis efficiency and accuracy are improved, and the medical resources are saved. In order to achieve the above purpose, the technical scheme of the invention is realized as follows: a tumor medical image classification method of a zero sample evolution NAS based on double-index collaborative evaluation comprises the following steps: S1, collecting tumor medical image data and constructing a data set containing tumor focus images and normal tissue images; S2, constructing an expanded search space based on a cell structure, wherein the search space comprises a basic operation set, the cell structure and a network integral structure; s3, setting steady-state evolution algorithm parameters, and carrying out steady-state evolution initialization to obtain initial candidate frameworks; S4, updating the candidate architec