Structured Proxy Features Enhance NSCLC Survival Prediction from CT

A recent preprint on arXiv (2608.00446) introduces structured proxy features aimed at enhancing survival predictions for non-small cell lung cancer (NSCLC) through pretreatment CT scans. Lung cancer is responsible for approximately 1.8 million deaths each year globally, with NSCLC being the predominant form. Despite advancements in treatment, accurately stratifying survival rates remains difficult due to the intratumoral heterogeneity that traditional descriptors fail to address. Existing radiomic and deep learning techniques often treat imaging features as isolated, overlooking the structured relationships among tumor characteristics. This study investigates the potential of six simulation-derived features, which are intended to represent the interplay between heterogeneity and morphology, to improve multimodal survival predictions by complementing CT data, radiomics, and clinical factors. A radiomic-parameterized cellular automaton produces proxy features for growth rate and necrosis ratio from baseline CT, employing entropy and sphericity for low-dimensional descriptor calculation. The findings were presented as a cross-type preprint on arXiv.