Brain tumor grading evaluation method based on deep learning and digital twinning

Abstract

The invention relates to the field of brain tumor grading, in particular to a brain tumor grading evaluation method based on deep learning and digital twinning. According to the technical scheme, the method comprises the steps that brain tissue CT image data are acquired and preprocessed; constructing a tumor digital twinborn three-dimensional model, performing three-dimensional reverse reconstruction on the preprocessed two-dimensional CT image by adopting a filtered back projection algorithm to obtain corresponding three-dimensional data, and constructing a virtual simulation experiment environment and a tumor digital twinborn three-dimensional model by using the three-dimensional data UE and a visual development framework based on the UE; making a digital twin data set, selecting a deep learning model, and introducing the domain confrontation layer into the deep learning model for training; and after the deep learning model training is finished, adjusting a parameter group required for constructing the tumor digital twinborn three-dimensional model based on a genetic algorithm, obtaining an optimal parameter group, and dynamically updating the tumor digital twinborn three-dimensional model based on the optimal parameter group. The method is suitable for brain tumor grading.

Claims

1. A brain tumor grading and assessment method based on deep learning and digital twins, characterized by: Including: S1. Obtain the CT image data of brain tissue and perform preprocessing; S2. Construct a three-dimensional model of the tumor digital twin; S3. Construct a digital twin dataset based on the three-dimensional model of the digital twin; S4. Select a deep learning model, introduce a domain adversarial layer into the deep learning model, and train it based on the CT image data and the digital twin dataset; S5. After the training of the deep learning model is completed, adjust the parameter group required for constructing the three-dimensional model of the tumor digital twin based on the genetic algorithm, obtain the optimal parameter group, and dynamically update the three-dimensional model of the tumor digital twin based on the optimal parameter group.

2. The brain tumor grading and assessment method based on deep learning and digital twins according to claim 1 is characterized in that: Step S1 specifically includes: Perform isotropic resampling on the CT image through a three-dimensional linear interpolation algorithm, calculate the pixel value weights of the eight vertex coordinates of the cubic unit around the interpolation point, and normalize the voxel. The normalization formula is as follows: ； Where y(r) is the pixel value of the image to be interpolated at the interpolation point r, is the known image at the grid point The pixel value of represents the weight coefficient; Perform layer-by-layer smoothing on the CT image array using a two-dimensional Gaussian filtering algorithm, and then use a semantic segmentation model to segment the tumor area in the CT image.

3. The brain tumor grading and assessment method based on deep learning and digital twins according to claim 1 is characterized in that: Step S2 specifically includes: Perform three-dimensional inverse reconstruction on the preprocessed two-dimensional CT image using a filtered back-projection algorithm to obtain the corresponding three-dimensional data; Import the three-dimensional data into UE, and construct a virtual simulation experimental environment based on the visualization development framework of UE to construct a three-dimensional model of the tumor digital twin.

4. The brain tumor grading and assessment method based on deep learning and digital twins according to claim 1 is characterized in that: Step S3 specifically includes: Export the twin slices from UE, unify the facial direction and anteroposterior diameter, standardize the slice size, add masks to the tumor areas one by one, and use the labeled tumor grade as the label; Set a random seed to clean the data, and divide the training set, test set, and validation set according to the set ratio. The training set, test set, and validation set are used as the digital twin dataset, and the data in the digital twin dataset corresponds one-to-one with the original CT image data.

5. The brain tumor grading and assessment method based on deep learning and digital twins according to claim 1 is characterized in that: Step S4 specifically includes: The input data of the domain adversarial layer includes source domain data and target domain data. The source domain data is the original CT image data, and the target domain data is the digital twin dataset. Input the source domain data and target domain data into the feature extractor in the domain adversarial layer for feature extraction, and input the extracted features into the domain discriminator and the deep learning model in the domain adversarial layer at the same time. Through the adversarial training of maximizing the accuracy of the classifier of the deep learning model and minimizing the accuracy of the domain discriminator, continuously optimize the performance of the feature extractor.

6. The brain tumor grading and assessment method based on deep learning and digital twins according to claim 1 is characterized in that: Step S5 specifically includes: Define a fitness function, and the iteration aim is to maximize the fitness function. The fitness function is as follows: ； Where x represents the input feature, W represents the weight of the feature extractor, b represents the bias of the feature extractor, V represents the weight of the deep learning model, and c represents the bias of the deep learning model. Represents the loss function of the deep learning model; Selection: Initialize the population, let the population contain N individuals, and the fitness of the i-th individual is F i , the probability of individual i being selected is equal to its own fitness probability: ； In the formula, N represents the number of individuals in the population, F j represents the fitness of the jth individual, P i Indicates that the probability of individual i being selected is equal to its own fitness probability; A crossover point is randomly selected on the chromosomes of the two selected parent individuals, and then the chromosome segments after the crossover point are exchanged. Assuming there are two parent individuals and , the chromosome length is n; Crossover: Randomly select the crossover point position as k, 1 ≤ k < n. The two offspring individuals generated through single-point crossover operation are C1 and C2, and the expressions are as follows: ； ； Mutation: Set the mutation probability Decide whether to mutate a gene in the offspring. The mutation process is expressed as: ； Where: L i is the lower limit of the value of the i-th gene, U i is the upper limit of the value of the i-th gene, represents the i-th mutant individual; Generate the next generation population through selection, crossover, and mutation operations, and find the optimal parameter group after multiple iterations.

Images