ESTRO 38 Abstract book

S225 ESTRO 38

L9cc and Dx. I-Model-2 included L3cc, L9cc and volume of duodenum as explanatory variables to establish a linear equation in three variables to predict the response variable, Dx. For validation, the difference between the predicted dose (D3cc, D9cc) from these three models and the achieved dose of clinical plans, was defined as delta dose (Gy) to evaluate the prediction accuracy.

development. Comprehensive QA procedures and tolerances were developed to evaluate pCT dose calculation accuracy and gold fiducial marker positions for image guidance prior to treatment. 3D Gamma evaluation was performed between pCT calculated plan dose and recalculation on QA CT following image registration. Distances of each of three gold fiducial markers to the centroid and isocentre were compared between pCT and QA CT scan. Results Isocentre dose differences between pCT and QA CT were (mean=-0.1%, SD=1.1%). The 3D Gamma dose comparison pass-rates were (mean=99.3%, SD=0.7%) with mean gamma (mean=0.239, SD=0.074) for 2%,2 mm criteria with 20% low dose threshold and QA CT as reference dose distribution. Results were similar for the two centres using two different scanners. All gamma comparisons exceeded the 90% pass-rate tolerance with a minimum gamma pass-rate of 97.6%. In all cases the gold fiducial markers were correctly identified on MRI and the distances of all seeds to centroid were within the tolerance of 1.0 mm of the distances on QA CT (mean=0.1 mm, SD = 0.5 mm). Differences in distances to isocentre were larger (mean=1.8 mm, SD=5.4 mm). This reflects displacements due to the image registration of QA CT to the pCT scan to copy the plan to the QA CT. All MRI-only treatment plans passed the QA criteria. All 17 prospective MRI-only plans were approved by the radiation oncologist and used for patient treatment. Conclusion To our knowledge this is the first multi-centre study examining prospectively the implementation of MRI-only radiotherapy planning. The results to date support the hypothesis that MRI-only prostate planning can be implemented safely and accurately. Alternatives to comparison to QA CT scans for quality assurance of MRI- only planning are currently under development. PV-0429 A machine learning method to improve duodenum dose prediction for pancreatic cancer radiotherapy Z. Feng 1 , K. Ding 1 1 John Hopkins School of Medicine, Department of Radiation Oncology and Molecular Radiation Sciences, Baltimore, USA Purpose or Objective Our goal is to use machine learning based method to improve the accuracy of overlap-volume-histogram (OVH) based duodenum dose prediction for pancreatic cancer radiotherapy. We have previously proposed a method to reduce duodenum dose using biodegradable pancreas- duodenum spacer and accurate dose prediction is an important step to understand if a patient needs spacer. Material and Methods A database of OVH and dose-volume-histogram (DVH) metrics, was collected from the stereotactic body radiation therapy (SBRT) plans of 230 previous patients with unresectable pancreatic cancer (33Gy in 5 fractions). OVH metrics L9cc and L3cc were defined as the tumor volume expansion distance at which 9cc and 3cc volume of the duodenum overlap with tumor. DVH metrics D9cc and D3cc of the duodenum were defined as the dose to 9cc and 3cc of the duodenum. We randomly selected 180 patients in the database as the training group and the rest of the 50 patients as testing and validation group. Our previously published prediction model, a linear regression model (O-M) between Lx and Dx, where x=3cc and 9cc, served as the baseline for comparison. For machine learning based method, we used multivariate regression model with Least Absolute Shrinkage and Selection Operator (LASSO). We included OVH data (L9cc, L3cc, duodenum volume) as explanatory variables to predict Dx of duodenum: I-Model-1 was a linear equation in two variables which is to model the relationship between L3cc,

Results For the training group, the ANOVA results of these models showed that the predicted accuracy of D9cc was significantly improved by including one or two parameters data in the IM-1(p=0.048) and IM-2(p=0.014) compared with O-M. Furthermore, by comparing between IM-1 and IM-2, adding volume of duodenum could lead to a significantly improved fit over the I-Model-1(p= 0.0087). For the testing group, the goodness of fitting between predicted dose and clinical dose was higher in IM- 1(r 2 =0.366 and 0.289 for D3cc and D9cc, respectively) and IM-2(r 2 =0.366 and 0.324 for D3cc and D9cc, respectively), compared to the O-M (r 2 =0.358 and 0.282 for D3cc and D9cc, respectively). All the predicted doses of D3cc and D9cc were in corresponding predicted ranges of three types predicted models. The root mean square error (RMSE) of D3 is lower in IM-1(1.50) and IM-2(1.50) compared to the O-M(1.51). IM-2 produces the lowest RMSE value of D9(1.31) than O-M(1.35) and IM-1(1.35).

Conclusion By utilizing machine learning method, the multivariate regression models can more accurately predict achievable duodenum planning dose. PV-0430 automated IMRT planning integrating knowledge-based model with Auto-Planning for cervical cancer C. Tao 1 , B. Liu 1 , C. Li 1 , J. Zhu 1 , J. Lu 1 , Y. Yin 1 1 Shandong Cancer Hospital Affiliated to Shandong University, Department of Radiation Oncology, Jinan, China Purpose or Objective In this study, a fully automated hybrid IMRT planning platform, integrating knowledge-based model and Auto- Planning engine, were established by Pinnacle scripts and Python codes for cervical cancer. And the advantages of this hybrid planning platform were investigated.