2514 - Longitudinal Radiomic Feature Instability in Auto-Segmented Gastrointestinal Organs-at-Risk Identifies Distinct Patient Phenotypes During Abdominal Radiotherapy

Purpose/Objective(s): Current GI toxicity risk assessment relies on single-timepoint dosimetric parameters. We hypothesized that quantifying inter-fraction radiomic instability from auto-segmented GI organs-at-risk (OARs) could reveal distinct patient phenotypes reflecting degrees of anatomical change, serving as an early imaging biomarker for adaptive intervention

Materials/Methods: Serial abdominal CTs from 96 patients (455 scans, multiple fractions) were analyzed. A U-Net with EfficientNet-B3 encoder auto-segmented stomach, small bowel, and large bowel (Dice: 0.93, 0.88, 0.89). Per scan, 38 radiomic features were extracted per organ: GLCM texture (contrast, correlation, energy, homogeneity), first-order statistics (mean, entropy, kurtosis, skewness), and shape descriptors (area, perimeter, bounding box). Inter-fraction variability was quantified via temporal coefficient of variation (CV). Patients were classified into radiomic drift phenotypes (Stable/Moderate/Unstable) by tertile stratification of a composite instability score. Wilcoxon signed-rank tests assessed first-to-last fraction changes. Spearman correlations evaluated radiomic-volumetric associations. Kruskal-Wallis tests validated phenotype separation.

Results: Substantial inter-fraction radiomic instability was observed: first-order CV 26-30%, GLCM 10-12%, shape 8-10%. Kurtosis was most unstable (CV: 61-75%). Three phenotypes showed highly significant separation (KW H=38–54, p<10??; Cohen's d=3.07, Stable vs Unstable). Composite instability scores were 8.1, 10.5, and 14.0 respectively. Six features showed significant first-to-last fraction changes (p<0.05), including large bowel intensity minimum (+44.9%, p=0.0001). Radiomic instability correlated with OAR volume variability (?=0.55, p<0.0001).

Summary of Radiomic Instability Metrics, Phenotype Characteristics, and Statistical Associations

Conclusion: Auto-segmented GI OARs exhibit significant, patient-specific inter-fraction radiomic instability reproducibly classified into distinct drift phenotypes. The correlation between radiomic and volumetric instability supports biological plausibility. These longitudinal signatures represent a novel biomarker that could trigger adaptive replanning. Prospective validation correlating radiomic drift phenotypes with CTCAE-graded GI toxicity is warranted.