Research On Brain Tumors Image Classification Based On Radiomics And Deep Learning

Brain tumor is a type of highly lethal tumors,accounting for about 5% of all body tumors,and their incidence has been on the rise in recent years.Due to the diverse lesion morphology and uncertain location of brain tumors,and the significant differences in the extent of lesions among patients,accurately classifying them presents a huge challenge.In recent years,brain tumor image research based on imaging genomics and deep learning has made some progress,but both methods still face pressing issues.The first is feature selection: imaging genomics methods obtain a large number of features in a highthroughput manner during feature extraction,but medical imaging is expensive and there are few research samples,leading to the "high-dimensional small samples" characteristics of imaging genomics data.Therefore,appropriate feature selection algorithms must be used to screen features to establish an effective classification model.The second is feature representation ability: imaging genomics extracts features using algorithm-defined by humans,which is less targeted to the target problem,while the feature performance extracted by deep learning methods is limited by the number of samples,both methods have insufficient representation ability.Therefore,only using imaging genomics features or deep learning features is difficult to accurately distinguish different types of brain tumors.Only by fusing different types of features can a more accurate classification model be established.Brain tumors are a type of highly lethal tumors,accounting for about 5% of all body tumors,and their incidence has been on the rise in recent years.Due to the diverse lesion morphology and uncertain location of brain tumors,and the significant differences in the extent of lesions among patients,accurately classifying them presents a huge challenge.In recent years,brain tumor image research based on imaging genomics and deep learning has made some progress,but both methods still face pressing issues.The first is feature selection: imaging genomics methods obtain a large number of features in a high-throughput manner during feature extraction,but medical imaging is expensive and there are few research samples,leading to the "high-dimensional smallsample" characteristics of imaging genomics data.Therefore,appropriate feature selection algorithms must be used to screen features to establish an effective classification model.The second is feature representation ability: imaging genomics extracts features using algorithm-defined by humans,which is less targeted to the target problem,while the feature performance extracted by deep learning methods is limited by the number of samples,both methods have insufficient representation ability.Therefore,only using imaging genomics features or deep learning features is difficult to accurately distinguish different types of brain tumors.Only by fusing different types of features can a more accurate classification model be established.Brain tumors are a type of highly lethal tumors,accounting for about 5% of all body tumors,and their incidence has been on the rise in recent years.Due to the diverse lesion morphology and uncertain location of brain tumors,and the significant differences in the extent of lesions among patients,accurately classifying them presents a huge challenge.In recent years,brain tumor image research based on imaging genomics and deep learning has made some progress,but both methods still face pressing issues.The first is feature selection: imaging genomics methods obtain a large number of features in a high-throughput manner during feature extraction,but medical imaging is expensive and there are few research samples,leading to the "highdimensional small-sample" characteristics of imaging genomics data.Therefore,appropriate feature selection algorithms must be used to screen features to establish an effective classification model.The second is feature representation ability: imaging genomics extracts features using algorithm-defined by humans,which is less targeted to the target problem,while the feature performance extracted by deep learning methods is limited by the number of samples,both methods have insufficient representation ability.Therefore,only using imaging genomics features or deep learning features is difficult to accurately distinguish different types of brain tumors.Only by fusing different types of features can a more accurate classification model be established.This thesis focuses on the above issues and conducts research on two brain tumor classification problems based on radiomics and deep learning methods,including the classification of low-grade glioma alpha-thalassemia/mental retardation syndrome Xlinked gene mutation status images and the classification of high-grade glioma and brain metastasis images.The relevant work is as follows:(1)This thesis proposes a modified grey wolf optimizer algorithm to address the problem of modeling high-dimensional small samples data in radiomics.The algorithm enhances the feature selection performance for high-dimensional small-sample data by introducing an initialization strategy based on correlation measurement,a competitive update strategy,and a head wolf differential evolution strategy.The initialization strategy based on correlation measurement is used to improve the quality of the initial population,the competitive update strategy balances global search and local search,and the head wolf differential evolution strategy reduces ineffective iterations of the population.Experimental results on multiple high-dimensional small-sample datasets show that compared with multiple high-dimensional data feature selection methods,the modified grey wolf optimizer algorithm in this thesis has significant advantages and is more suitable for feature selection problems in radiomics classification data.(2)To address the limited feature representation capability in radiomics,this thesis introduces a deep convolutional neural network to extract features of brain tumor images to compensate for the deficiency of traditional radiomics feature representation.Deep learning feature extraction is accomplished through multiple pre-trained networks,and the network with high classification accuracy is selected as the extraction module to extract features with good representation ability.To better utilize different types of features,this study designed various feature fusion methods to enhance the classification performance of the model.For feature fusion in multi-sequence deep learning,this thesis designs a feature fusion module based on cross-attention mechanism to fuse features from different sequences.Experiments show that compared to other commonly used featurelevel fusion methods,this module can fully exploit complementary features of different sequence images.This thesis also designs a two-layer Stacking structure to perform decision-level fusion of radiomics features and deep learning features,and experimental results demonstrate that this model can effectively combine the advantages of different features and achieve better classification performance.(3)In response to the brain tumor image classification problems in clinical research,this thesis uses the multiple methods proposed above to establish two brain tumor image classification models: Using magnetic resonance imaging and positron emission tomography images obtained from multiple centers of patients with low-grade glioma,a machine learning model based on radiomics was established to classify images based on the α-thalassemia/mental retardation syndrome X-linked gene mutation status in lowgrade glioma.The other is the deep learning-based radiomics classification model for high-grade glioma and brain metastasis based on magnetic resonance images.Experimental results show that the established models accurately classify the two types of brain tumor images and can provide decision support for clinical diagnosis.(4)This thesis designs and develops a brain tumor image classification prototype system based on deep learning radiomics,with feature selection,feature extraction,and model establishment as the main functional modules.This system can provide reference for the practical application of computer-aided brain tumor image classification.In summary,this thesis has addressed the defects in brain tumor image classification.To address the feature selection problem in radiomics data,a modified grey wolf optimization algorithm was proposed to remove redundant and irrelevant features.To enhance the feature representation capability,a convolutional neural network was used to extract deep learning features,and a feature fusion module based on cross-attention mechanism and Stacking method was proposed to combine different types of features.By combining multi-sequence radiomics features and deep learning features,a more comprehensive and accurate characterization of tumor images was achieved.The above methods optimized feature selection,feature extraction,and feature fusion,which improved the accuracy of brain tumor image classification and could potentially provide support for clinical decision-making.