Cancer Prognosis Analysis Algorithms For Complex Multi-modality Scenarios

Cancer has become one of the major public health issues that threaten human health.With the development of medical imaging and detection technologies,abundant multi-modality data(such as genomic sequences,histopathological slices,clinical information)can be collected to provide a comprehensive multi-dimensional perspective for studying the mechanisms of cancer,thereby better guiding the prognosis analysis of cancer.However,current multi-modality prognosis methods still confront various challenges,such as how to effectively fuse heterogeneous multi-modality information,how to flexibly handle incomplete multi-modality data,and how to properly cope with the unique censoring problem.This thesis focuses on complex multi-modality scenarios and proposes a series of machine learning-based cancer prognosis analysis algorithms so as to achieve accurate prognostic prediction.The main research contents of this thesis are as follows:(1)In the scenario of complete multi-modality data,a multi-constraint latent representation(McLR)method is proposed to address the issues of insufficient feature fusion,high feature dimension,indirect utilization of survival time and so on.This method constructs a multi-modality shared space to fuse multi-modality information,learns low-dimensional latent representation in a model-driven manner,and also makes full use of ranking and regression information of survival time.(2)In the scenario of complete multi-modality data,an adaptive risk-aware sharable and individual subspace learning(Ada-RSIS)algorithm is proposed to address the issues of insufficient exploration of modality-sharable and modality-specific information,inattention to local properties,time-consuming parameter tuning and so on.This algorithm jointly learns modality-sharable and modality-specific subspaces to fully exploit complementary information in multi-modality data,introduces grouping co-expression learning to obtain risk-aware local representation,and employs an adaptive-weighted strategy to automatically tune hyper-parameters.(3)In the scenario of partially incomplete multi-modality data,a relation-aware shared representation learning(RaSR)method based on auxiliary clinical variables and incomplete multi-modality data is proposed to address the issues of incomplete multimodality data,noises or redundancy introduced by data imputation technologies,insufficient utilization of clinical variables and so on.This method obtains a compact and discriminative shared representation by relational regularization terms,dynamically regresses and combines multiple clinical attributes to help improve model’s performance,and uses a partial projection strategy to deal with incomplete multi-modality data.(4)In the scenario of multi-modality training and mono-modality testing data,a mutual-assistance learning(MaL)paradigm is proposed to address the issues that multi-modality models are not available and clinical data are difficult to meet strong distributional and parametric assumptions.This paradigm leverages the knowledge of multi-modality data in the training stage to reinforce the mono-modality representation learning for improving the generalization performance of mono-modality model in the testing stage,formulates mutual-assistance regression and ranking functions independent of strong hypotheses to simultaneously concentrate on the survival time and event occurrence order,and integrates representation learning and survival modeling into a unified mutual-assistance framework to strengthen their synergy.The experimental results on multiple datasets demonstrate the effectiveness of the above-mentioned models in predicting cancer prognosis under complex multi-modality scenarios and the potential of exploring prognostic markers.