Repeatability and prognostic value of radiomic features : a study in esophageal cancer and nasopharyngeal carcinoma

Abstract

Objectives: To investigate whether radiomic features (RFs) repeatability and their prognostic value are study-specific. Materials and methods: This retrospective study included 234 esophageal cancer (EC) patients (contrast-enhanced computed tomography (CECT) and fluorine-18 fluorodeoxyglucose positron emission tomography (PET)), and 525 nasopharyngeal carcinoma (NPC) patients (CECT). Tumor, peritumor, and lymph node regions were defined as regions of interest. RF repeatability was assessed via perturbation analysis using intraclass correlation coefficients (ICC), with consistency and differences across cancer types, pathological regions, and modalities evaluated. The independent prognostic features common to both EC and NPC were screened from highly repeatable features using C-index and redundancy analysis. Results: CT-based RFs in NPC and PET-based RFs in EC demonstrated significantly higher repeatability compared to CT-based RFs in EC (median ICC: 0.886 vs 0.806; 0.897 vs 0.806; p < 0.05). While CT-based peritumoral features showed comparable repeatability to tumor features in EC (0.824 vs 0.806, p > 0.05), PET-based peritumoral features exhibited significantly lower repeatability than tumor features (0.819 vs 0.897, p < 0.05). CT-based lymph node features demonstrated significantly lower repeatability than tumor features in NPC (0.863 vs 0.886, p < 0.05). Nevertheless, the effects of bin count, feature class, and filter on repeatability demonstrated consistent patterns across different cancer types, imaging modalities, and pathological regions. Moreover, four common independent prognostic features effectively stratified both EC and NPC patients into high- and low-risk groups with significant survival differences (p < 0.05). Conclusions: RF repeatability might be affected by cancer type, pathological region, and imaging modality, while certain features maintain consistent prognostic performance across different cancer types. Critical relevance statement: The identification of high-repeatable pan-cancer prognostic radiomics features enables noninvasive patient risk stratification to guide personalized therapy, with cross-cancer consistency enhancing their applicability and convenience in clinical practice, thereby accelerating the integration of radiomics into precision oncology clinical workflows. Graphical abstract: [Figure not available: see fulltext.]