Abstract
As an analytic pipeline for quantitative imaging feature extraction andanalysis, radiomics has grown rapidly in the past a few years. Recent studiesin radiomics aim to investigate the relationship between tumors imagingfeatures and clinical outcomes. Open source radiomics feature banks enable theextraction and analysis of thousands of predefined features. On the other hand,recent advances in deep learning have shown significant potential in thequantitative medical imaging field, raising the research question of whetherpredefined radiomics features have predictive information in addition to deeplearning features. In this study, we propose a feature fusion method andinvestigate whether a combined feature bank of deep learning and predefinedradiomics features can improve the prognostics performance. CT images fromresectable Pancreatic Adenocarcinoma (PDAC) patients were used to compare theprognosis performance of common feature reduction and fusion methods and theproposed risk-score based feature fusion method for overall survival. It wasshown that the proposed feature fusion method significantly improves theprognosis performance for overall survival in resectable PDAC cohorts,elevating the area under ROC curve by 51% compared to predefined radiomicsfeatures alone, by 16% compared to deep learning features alone, and by 32%compared to existing feature fusion and reduction methods for a combination ofdeep learning and predefined radiomics features.