Multi-Level PET and CT Fusion Radiomics-based Survival Analysis of NSCLC Patients

Abstract

To provide a comprehensive characterization of intra-tumor heterogeneity, this study proposes multi-level multimodality radiomic models derived from 18F-FDG PET and CT images by feature- and image-level fusion. Specifically, we developed fusion radiomic models to improve overall survival prediction of NSCLC patients. In this work, a NSCLC dataset including patients from two different institutions (86 patients used as training and 95 patients used as testing) was included. By extracting 225 features from CT, PET, and fused images, radiomics analysis was used to build single-modality and multimodality models where the fused images are constructed by 3D-wavelet transform fusion (WF). Two models were also developed using two feature-level fusion strategies of feature concatenation (ConFea) and feature averaging (AvgFea). Cox proportional hazard (Cox PH) regression was utilized for survival analysis. Spearman's correlation was utilized as a measure of redundancy, and then best combination of 10 most related features (ranked by univariate Cox PH) were fed into multivariate Cox model. Moreover, the median prognostic score captured from training cohort was used as an untouched threshold in the test cohort to stratify patients into low- and high-risk groups. The difference between groups was assessed using log-rank test. Among all models, WF (C-index=0.708) had the highest index and significantly outperformed CT and PET (C-index = 0.616, 0.572, respectively). Image-level fusion model also outperformed feature-level fusion models ConFea and AvgFea (C-indices = 0.581, 0.641, respectively). Our results demonstrate that multimodal radiomics models especially models fused at image-level have the potential to improve prognosis by combining information from different tumor characteristics, including anatomical and metabolic captured by different imaging modalities.