AI Deformable Dose Accumulation and Patient-Specific Uncertainty for MR-Guided Radiotherapy of Locally Advanced Pancreatic Cancer
Abstract
Purpose
Dose accumulation using deformable image registration (DIR) enables a more accurate evaluation of delivered dose in adaptive radiotherapy, allowing evaluation of discrepancies between planned and delivered treatments. However, different DIR algorithms produce varying deformation vector fields, potentially affecting accumulated dose distributions. This study evaluated dose variability when accumulating dose on different reference fractions for each DIR algorithm, as well as differences between originally planned and accumulated plans in pancreatic MR-guided adaptive radiotherapy (MRgART).
Methods
Eight patients treated with 5-fraction pancreatic SBRT on a 1.5T MR-Linac were retrospectively analyzed. Dose accumulation was performed using ProRSeg, a progressively refined joint registration-segmentation network, with three registration modes: (1) standard intensity-based registration, (2) AI-contour guided registration using predicted segmentations, and (3) clinician-contour guided registration using ground truth contours. For each patient, dose was accumulated onto each of the 5 treatment fractions as reference anatomy to assess dose variability. Also, the duodenum-stomach normal tissue complication probability (NTCP) was calculated using the Lyman-Kutcher-Burman model with the following parameters: n=0.10, m=0.21, TD50=56 Gy, α/β=3 Gy for both the original reference plan and the delivered accumulated plan.
Results
Duodenum-stomach dose metrics varied depending on which fraction was used as the accumulation reference. Median D0.035cc/D5cc variability across fractions was 6.16/3.44 Gy [2.46-17.87/0.67-8.87] for intensity-based registration, 6.80/2.59 Gy [3.53-13.63/0.92-9.35] for AI-contour guided, and 4.47/1.95 Gy [2.81-14.23/0.63-4.39] for clinician-contour guided registration. NTCP differences between accumulated treatment and original reference plan showed median deltas of +3.9% [-2.8, +58.7], +1.4% [-2.1, +37.0], and +1.5% [-2.7, +4.0], for each DIR algorithm respectively.
Conclusion
Contour-guided DIR methods reduced dose variation between accumulated reference fractions, indicating a more robust estimating of deformation in these more complex and mobile structures like the duodenum-stomach. Contour-guided based NTCP values provided a low deviation between accumulated to initial reference plan, suggesting that adaptation is properly reflecting intended treatment plan.