ESTRO 2020 Abstract Book

S1094 ESTRO 2020

and PDR afterloaders, in single acute fractions and multiple hourly fractions (or pulses); all experiments were performed twice in triplicate. α and β were determined using HDR and T 1/2 using PDR. Cell colonies were counted to determine the clonogenic survival after irradiation. α/β and T 1/2 were estimated for each cell line through minimum chi square estimation. Single fraction experiments were used to determine α/β of the cells while multiple fraction experiments determined T 1/2 . Results Calculated parameter values are shown in Table 1. The α/β of the cells were found to be smaller than the conventional value utilized for treatment planning and the T 1/2 values were generally larger than the clinically used value. Analogous experiments using different sources (e.g. single acute fractions of Cs-137 and Ir-192) yielded mostly similar radiobiological parameters, suggesting that the differences between the sources (e.g. radiation energy) did not significantly impact survival. The α/β of CaSki cells varied between Cs-137 and Ir-192 experiments.

Conclusion Differences between the clinically used and experimentally determined values for α/β (5 Gy maximum) and T 1/2 (2.5 hours maximum), highlight possible uncertainties in planned treatment doses. The values obtained experimentally suggest differences of over 15% in the delivered dose, which could potentially impact patient outcomes. OC-1042 Deep-learning based Physician’s Preference Prediction For HDR Brachytherapy Of Cervical Cancer Y. Gonzalez 1 , C. Shen 1 , C. Wang 1 , K. Albuquerque 1 , X. Jia 1 1 UT Southwestern Medical Center, Radiation Oncology, Dallas, USA Purpose or Objective To automate the treatment planning process of HDR brachytherapy (HDRBT) for cervical cancers to improve plan quality, consistency and planning efficiency, many studies have developed automated methods to generate a plan aiming at achieving a pre-defined plan score. However, the plan score does not necessarily reflect a physician’s intention during treatment planning. With recent advances in deep learning, we propose a deep- learning method that learns a physician’s preference when approving a plan for HDRBT of cervical cancer treated with a tandem-and-ovoid applicator. Material and Methods We collected approved treatment plans of 220 treatment fractions from 45 patients treated by a single radiation oncologist. We set up one dose prediction neural network (DPNN) and one preference prediction neural network (PPNN) formed under an adversarial framework. The DPNN predicts EQD2 of CTV D90 and OAR D2cc from patient anatomy. The PPNN predicts the probability of a plan being acceptable to the physician based on the specific anatomy. Training of the networks was achieved in two steps. First, we individually trained each network in a supervised learning process. In particular, training the PPNN required clinically unapproved plans, which were generated by randomly perturbing dwell time of approved plans. Second, we jointly trained the two networks with the DPNN aiming at best predicting doses for each patient,

Potential impacts of the experimental results are shown in Figure 1. Deviation in either parameter value could significantly alter the calculated radiobiological dose. A change in either α/β or T 1/2 , within the range of our experimentally derived values, could introduce differences between conventionally equivalent HDR and PDR BT ( ΔD ) of up to 10 Gy EQD2 and 13 Gy EQD2 for the treatment schedules considered, respectively.