Complementary correlation and fusion of multi-modal data features for NSCLC intelligent recognition

Abstract:
Recent studies have showed that the integrative analysis of CT images and genomic data can significantly improve the prediction performance of tumor diagnosis and prognosis. However, tumor classification based on the complementary correlation between CT images and genomic data has rarely been studied. In addition, when using multi-modal data features for tumor classification, most of the existing methods concatenate the multi-modal features directly, which ignore the importance of different modal features. Based on the above considerations, we construct an integrative imaging genomics classification framework for non-small cell lung cancer (NSCLC). Specifically, an adaptive weighted feature fusion method based on category contribution rate is proposed to distinguish the importance of different modal features adaptively in the classification process. Candidate’s pathogenic genes selected by correlation analysis have the potential to be suitable markers in distinguishing different subtypes of NSCLC. Our experimental results verify that our method has a better classification performance for multi-modal features than other related methods.