1075 - Auto-Segmented Organ-at-Risk Volume Variability as a Patient-Specific Predictive Biomarker for Gastrointestinal Toxicity Risk Stratification in Abdominal Radiotherapy

Purpose/Objective(s): Inter-fraction anatomical variability of GI organs-at-risk (OARs) is well-recognized but poorly quantified at patient level. Current practice applies uniform margins regardless of individual variability. We hypothesized that inter-fraction OAR volume variability, quantified via deep learning auto-segmentation across longitudinal imaging, could serve as a patient-specific biomarker for GI toxicity risk, enabling individualized margin adaptation.

Materials/Methods: The UW-Madison GI Tract dataset (96 patients, serial abdominal CTs across multiple fractions) was analyzed. A U-Net with EfficientNet-B3 encoder was trained on 38,964 annotated slices to auto-segment stomach, small bowel, and large bowel (mean Dice: 0.928). Per-organ volumes were computed per patient-fraction using predicted segmentations with patient-specific pixel spacing and 3mm slice thickness. Inter-fraction variability was quantified by coefficient of variation (CV%), volume range, and maximum inter-fraction delta. Patients were stratified into tertile-based risk groups (Low/Medium/High) using composite CV. Predictive performance was assessed via 5-fold cross-validated logistic regression with ROC analysis. Patient-specific PTV margins were derived using an adapted van Herk formula (M=2.5S+0.7s). Statistical comparisons used Kruskal-Wallis tests with post-hoc Mann-Whitney U and Spearman correlations.

Results: Substantial inter-fraction volume variability was observed: stomach CV 20.2±14.2%, small bowel 21.4±15.4%, large bowel 21.7±14.4%. Risk stratification showed significant separation (KW p<10?6; Cohen's d: 1.37-2.14). High-risk patients demonstrated 32.3% mean CV vs 11.4% for low-risk. Cross-validated ROC analysis: composite AUC=0.995±0.006; individual organ AUCs 0.807-0.857. Small bowel CV was the strongest predictor (LR coefficient=2.27). Adaptive margins showed 52-78% reductions for low-variability patients vs uniform worst-case (e.g., stomach: 4.9 vs 21.7mm).

Conclusion: Inter-fraction OAR volume variability quantified through auto-segmentation is a robust patient-specific biomarker for GI toxicity risk stratification. High-variability patients may benefit from adaptive replanning, while low-variability patients can safely receive reduced margins. This framework provides a clinically actionable trigger for individualized adaptive RT and warrants prospective validation with toxicity outcomes.

Volume Variability Metrics, Risk Stratification, Predictive Performance, and Adaptive Margins