ESTRO 36 Abstract Book

S127 ESTRO 36 2017 _______________________________________________________________________________________________

errors below 2.5Gy. For proton H&N plans, a dataset size of at least 173 plans resulted in all mean errors below 2.5Gy. Dataset sizes are shown in Table 1. Shown visually in Figure 1, using predictive modeling of the plan outcome, re-planning a lung SBRT case resulted in improved dose to critical structures while maintaining coverage to the PTV, compared to the clinically-developed and treated plans. Error Target (Gy) 5 4 3 2.5 2 Lung dataset size 16 46 61 69 74 HN Proton size 169 130 153 173 192 HN Photon size 68 136 145 121 136 Table 1: Plan datasets required for desired dose accuracy.

deteriorate them by more than 1Gy/1% compared to the original plan. Given the 5 parameters of interest and the 90% criteria, (1- (0.9) 5 )≈40% patients were expected to fall outside the prediction range. Results Figure 1 shows the results of the model validation.

Figure 1 Comparison of clinical plan developed without (A) and with (B) predictive modeling. Conclusion We have demonstrated the ability to predict dosimetric indices. These results have clinical implications that extend from decision making to planning workflow improvement to quality improvement. OC-0254 Prospective validation of independent DVH prediction for QA of automatic treatment planning Y. Wang 1 , B.J.M. Heijmen 1 , S.F. Petit 1 1 Erasmus MC - Cancer Institute, Radiation Oncology, Rotterdam, The Netherlands Purpose or Objective In our institute, fully automated, knowledge-based treatment planning is used in routine clinical practice. For the majority of patients, this is expected to result in high quality treatment plans. However, technical and procedural issues might result in suboptimal plans for some patients that might go undetected. In this study, we prospectively investigated the clinical usefulness of an independent DVH prediction tool to detect outliers in treatment plan quality for prostate cancer patients. Material and Methods All prostate cancer patients treated from January 2015 till half September 2016 with the full prescribe ed dose delivered to the prostate only or to the prostate+seminal vesicles were included in the study. They were treated with an automatically generated VMAT or dMLC plan. The QA method was based on overlap volume histogram and principal component analysis and is fully independent of the planning method. The model was trained with 50% of the patients treated in 2015 (N=22) and validated on the other 50% (N=21). We focused on 5 different dose metrics: rectum D mean , V 65 , V 75 ; anus D mean and bladder D mean . Next, to study the clinical usefulness of treatment planning QA, the QA model was applied prospectively for the patients treated in 2016 (N=50). Patients for which at least one of the five dose metrics fell outside the 90% prediction confidence interval (CI) were further improved by manual plan adjustments (‘re-planning’). The re- planning goals were to keep or improve all dose metrics of interest within or lower than the 90% CI, and anyway not

17 Patients from the prospective cohort were classified as outliers, including all four patients with metal hips, which were excluded from further analysis. The remaining outliers 13/46 (28.3%) were re-planned and for all the re- planning requirements (above) were met. As shown in Figure 2, the new plans were moderately superior to the clinical plans for rectum D mean (average improvement 0.9Gy, max. improvement 3.0Gy, p =0.009), V 65 (2.4%, max 4.2%, p =0.001), anus D mean (1.5Gy, max 6.8Gy, p= 0.004), and bladder D mean (1.7Gy, max 5.1Gy, p =0.001). The rectum V 75 of the new plans was slightly higher than with the original plan (0.2 %, p =0.028). No significant differences were found in PTV conformity or the femoral head D max .