Radiomics Feature Activation Maps as a New Tool for Signature Interpretability.

摘要: Introduction: In the field of personalized medicine, radiomics has shown its potential to support treatment decisions. However, limited feature interpretability hampers introduction into clinics. Here, we propose a new methodology create activation maps, which allows identify spatial-anatomical locations responsible for signature based on local radiomics. The feasibility this technique will be studied histological subtype differentiation (adenocarcinoma versus squamous cell) in non-small cell lung cancer (NSCLC) using computed tomography (CT) Materials and Methods: Pre-treatment CT scans were collected from multi-centric Swiss trial (training, n=73, IIIA/N2 NSCLC, SAKK 16/00) an independent cohort (validation, n=31, IIIA/N2/IIIB NSCLC). Based gross tumor volume (GTV), four peritumoral region interests (ROI) defined: lung_exterior (expansion lung), iso_exterior soft tissue), gradient (GTV border region), GTV+Rim iso_exterior). For each ROI, 154 radiomic features extracted in-house developed software implementation (Z-Rad, Python v2.7.14). Features robust against delineation variability served as input multivariate logistic regression analysis. Model performance was quantified area under receiver operating characteristic curve (AUC) verified 5-fold cross validation internal validation. Local ROI non-overlapping 3x3x3 voxel patches previously marked GTV or rim. A binary map created patient median global value training. ratios activated/non-activated rim compared between subtypes (Wilcoxon test). Results: Feature stability moderate (49.7% 74.5% GTV+Rim, respectively). Iso_exterior, gradient, showed good performances prediction (AUCtraining=0.68-0.72 AUCvalidation=0.73-0.74) whereas models failed model maps that texture distribution differed significantly (p=0.0481) but not (p=0.461). Conclusion: exploratory study, radiomics-based NSCLC predominantly indicating can useful tracing back spatial location regions activation.