Main Session
Sep
29
PQA 05 - Physics
Presenter(s)
Anil Sethi, PhD, FAAPM - Loyola University Medical Center, Maywood, IL
A. Sethi1, M. Perepitchka2, M. Robertson2, Y. Abdelal1, T. Montenegro1, S. Zand2, A. Joseph2, W. Small Jr1, and T. Refaat1; 1Department of Radiation Oncology, Stritch School of Medicine, Cardinal Bernardin Cancer Center, Loyola University Chicago, Maywood, IL, 2Stritch School of Medicine, Cardinal Bernardin Cancer Center, Loyola University Chicago, Maywood, IL
Purpose/Objective(s):
To develop a predictive model to assess need for treatment plan adaptation. Clinical implementation is expected to result in improved patient selection, timely intervention, and efficient resource utilization.Materials/Methods:
Retrospective plans from 23 pancreatic SBRT patients treated on an MR linac (138 image sets) were analyzed for PTV/OAR doses and ranked by OAR proximity to PTV. Per-fraction patient data sets were constructed linking daily MR anatomy, adaptive dose metric (ADM), and clinical decision (adapt plan vs. treat-as-planned). ADM captured anatomical changes by combining organ motion, organ deformation, overlap index (OI) and dice similarity coefficient (DSC). A novel proximity index, touching distance (TD) was developed and compared to conventional position indicators to assess impact of organ motion on model predictions. Geometric and dosiomic features combined with machine learning (ML) were integrated into a unified training set for adaptive modeling within a five-phase predictive framework (pre-processing, tuning, performance, prediction, and clinical decision-tree). Adaptive model incorporated regularized logistic-regression baselines with Bayesian hyperparameters (class-weight, ?, a) fine-tuning to identify OARs most likely to cause plan adaptation. Model performance was evaluated with organ specific area under ROC curve (AUROC), sensitivity, specificity, accuracy and odds ratio (OR) to indicate features most associated with adaptation likelihood.Results:
Organ specific dose threshold for duodenum OAR was typically exceeded at ~1-2mm approach to PTV, stomach and small bowel at ~3-4mm, and large bowel at = 4mm with wide inter-patient variability. A mixed-effects negative-binomial model correlated each 1mm of margin buffer between OAR-PTV to ~7% fewer adapted fractions (incidence rate ratio, IRR = 0.93, p = 0.011). Predictive TD model performance (AUROC =0.755) was superior to conventional distance/similarity metrics as per average OAR sensitivity (0.880 vs 0.496), specificity (0.880 vs 0.639), and accuracy (0.875 vs 0.620). Organ specific analysis also favored TD model with significantly higher predictive powers for stomach (0.803 vs 0.606), duodenum (0.770 vs 0.545), small bowel (0.833 vs 0.548), and large bowel (0.706 vs 0.484). Fraction specific TD was the strongest indicator of treatment plan adaptation with an OR = 3.14 (CI = 1.97 – 5.02, p < 0.01) followed by inter-fractional change in TD (?TD) and OI with OR of 1.54 (p < 0.01) and 1.16 (p = 0.73) respectively.Conclusion:
A powerful adaptive dose metric based on a new proximity index was developed and evaluated in assessing need for treatment plan adaptation. Clinical implementation will likely improve treatment efficiency and efficacy for pancreatic cancer patients. The predictive tool can be readily generalized to other treatment sites and image modalities used in adaptive radiotherapy.