SU-E-J-96: Predictive Neural Network for Parotid Glands Deformation Using IGRT and Dose Warping Systems

Purpose: Tumor shrinkage and modification of patient anatomy may occur within the weeks of therapy. Dose warping algorithm and IGRT can usefully determinate morphological variations occurring during treatments and predict possible challenge or requirement of re-planning. Using metaanalysis and developing Predictive Neural Network (PNN) is possible to collect information for perspective treatment evaluation of inappropriate dose delivery or to re-planning care course. Methods: 23 H'N patients, treated using Tomotherapy Unit, were post-processed using deformable hybrid algorithms to generate contours evolution throughout the course. 750 MVCT were elaborated focusing on parotids. For each deformable image registration were generated deformable ROIs and re-mapped to dose grid to accumulate dose. A nomogram was developed to evaluate Volume (V) and Dose (D) variations. Using Moving Average (MA) function and PNN were identified the weeks when re-plan should be predictable. Results: The weekly nomogram provided the percentage of patients affected by V and D deviation respect to the first day of treatment. During 1st week of treatment 95% of patients have a ΔV=10% and ΔD=2%. In the 6th week, only 70% of cases remain inside that range. The PNN using Cluster Analysis and Support Vector Machines, highlighted the requirement of warping methods in clinical practice. Validated by 8 patients test cases, PNN allow to classify 2 statistical cluster related to original plan and re-plan requirements. MA with period of 3 and 5 days, predict days of treatment where to locate physics replan. Best statistical benefits can be achieved with evaluation in the 4th week. Conclusion: Body morphing and organ motion during therapy can affect to the dose distribution and induce unexpected or late toxicities. Hybrid deformable registration allows to study the time to re-plan or to adapt treatments. Monitoring large patient's database during treatments, thereby rising accuracy in nomograms, PNN and reproducibility of treatments. The research is partially co-funded by the MoH (GR-2010-2318757) and Tecnologie Avanzate S.r.1.(Italy)