Abstract
Purpose: To investigate if radiomics signatures generated from longitudinal CT scans could predict IRE treatment effectiveness and outcomes in patients with locally advanced pancreatic cancer (LAPC). Methods: A cohort of 50 (60% male, mean [SD] age 60.7 [8.7] years) LAPC patients treated with IRE were retrospectively selected. Preoperative and 12-week follow-up CT scans were reviewed by two radiologists for tumor segmentation. A total of 2078 features were extracted: shape (n = 16), texture (n = 68), filter (n = 1892), intensity (n = 18), and local texture (n = 84). Principal component analysis (PCA) was applied to develop composite radiomics features. Composite signatures and clinically relevant radiomics features were correlated with time to recurrence (TTR), time to local recurrence (TTLR), time to distant recurrence (TTDR), recurrence-free survival (RFS) and overall survival (OS). Risk stratification performance was evaluated using hazard ratios (HRs), and significance was evaluated using the log-rank test. Results: Statistically significant separation between high and low patient TTR risk groups was observed in the following: gray-level co-occurrence matrix (HR = 2.65, p < 0.01, median survival difference = 6.6 mo); composite radiomics features derived from the following feature groups: all radiomics features (HR = 2.27, p = 0.01, median survival difference = 6.4 mo), intensity features (HR = 3.13, p < 0.01, median survival difference = 14.0 mo), and filter features (HR = 2.27, p = 0.01, median survival difference = 6.4 mo). Conclusions: Pre-treatment radiomics signatures were significantly associated with LAPC patient outcomes. The observed correlations used pre-treatment CT scans, implying that the features predict the individual risk of disease recurrence.