Classification Of Liver Cancer Subtypes Based On Hierarchical Integration Deep Flexible Neural Forest Algorithm

Molecular typing of HCC can help patients to carry out personalized treatment to improve the cure rate of HCC.Therefore,more and more HCC subtypes based on molecular level have been studied.The rapid development of the new generation of genomic technology has generated and accumulated tens of thousands of hepatocellular carcinoma genomic data,which has created favorable conditions for people to explore the mechanism of hepatocellular carcinoma in all aspects and at various levels.However,hepatocellular carcinoma omics data usually have the characteristics of small sample size and high dimension.Therefore,in the classification of liver cancer subtypes,the classification model is disturbed by noise caused by high latitude,resulting in low accuracy of the classification results obtained.Since liver cancer subtypes have different features’ expressions in different omics data,we fused multiple omics data to study the impact on the classification of liver cancer subtypes.However,there is a certain correlation between different omics data,which makes it difficult for data fusion.To solve the above problems,genomic data composed of multiple omics including copy number variation,DNA methylation,m RNA expression and mi RNA expression of liver cancer were studied in this thesis.By analyzing the genomic data of HCC and considering the dependence of omics data,a hierarchical classification method was designed for small-scale data sets,which effectively improved the classification accuracy of HCC subtypes.First,in view of the high dimension and complexity of liver cancer data,a hierarchical stacked noise reduction encoder model is proposed in this thesis.Firstly,the intermediate feature representation of each omics data is learned by using stack denoising autoencoder.Secondly,all intermediate feature representations learned are fused into another layer of autoencoder to learn more complex advanced feature representations.Finally,the complex data representation learned is used as the input of Softmax layer to obtain the final classification result of liver cancer subtypes.The hierarchical stack denoising encoder not only takes into account the inherent statistical properties of various types of data but also maintains the correlation of different omics data,which ensures the diversity of omics data and realizes effective dimensionality reduction learning for data and better integrates multiple omics data.The experimental results showed that the hierarchical stack denoising encoder can improve the features’ expression ability of omics data to a certain extent,and then improved the classification accuracy of the classification model.Secondly,in view of the low accuracy of the existing models in the classification of liver cancer subtypes,a deep flexible neural forest model is proposed in this thesis.The model is an integration of the flexible neural tree.Firstly,the flexible neural tree is integrated in each layer of the network to deal with multiple classification problems.Secondly,in order not to add additional parameters,the cascade structure is adopted to deepen the depth of the network.The model can automatically determine the flexible neural tree structure by tree structure optimization algorithm,and adaptively adjust the number of the whole cascade hierarchy,so that it is suitable for small-scale liver cancer genomic data.Experimental results showed that the proposed model is better than support vector machine,random forest,deep forest and other classification models in classification accuracy,and better solved the problem of liver cancer subtype classification.