ESTRO 2021 Abstract Book

S1431

ESTRO 2021

probability of getting undesirable results. Nevertheless, most reported complexity scores demonstrate a lower or moderate correlation with GPR, having sub-optimal prediction performances. Contrastingly, several machine learning (ML) algorithms have been proposed using different datasets with customised complexity scores, increasing the prediction power and offering classification and regression solutions. However, it remains unclear if the complexity metrics considers all the potential parameters associated with GPR, or if a specific metric is suitable or not for a dataset or treatment units. For these reasons, it was intended to compare the GPR prediction power of ML algorithms using conventional calculated complexity metrics against a neural network model that extracts the information from high dimensional planning data stored in the DICOM files. Materials and Methods The data from 231 prostate plans were extracted to feed four ML algorithms, (I) random forest (RF), (II) XG- Boost, (III) simple neural network (1D-NN), and (IV) 3D-NN. All models performed binary classification to predict if one plan will pass (0 if GPR ³ 98%) or fail (1 if GPR <98%). The first three models used 175 features, including complexity metrics, radiomics features, and volume/dosimetric information of target volumes and organs at risk. The fourth model was dedicated to use non-calculated inputs, including the trajectory leaf maps (TLM) and monitor units (MU) per control point profiles (MUcp). The area under the curve (AUC) were calculated to measure the prediction (classification) performance of each model. Results The AUC values for RF, XG-Boost and 1-D NN were 0.83±0.16, 0.84±0.14, and 0.91±0.04, respectively. The AUC for 3D-NN with TLM and MUcp as inputs were 0.92±0.06. The AUC for 3D-NN model using TLM only were 0.84±0.14. The classification performance of 3D-NN is comparable with 1D-NN and superior to RF and XG- Boost. The features used in 3D-NN are suitable to propose GPR models strategies in pre-treatment verification protocols with less data preparation time. The MUcp increase the prediction power of the latter model. Conclusion Modulation complexity might be automatically inferred from deep learning algorithms instead of time- consuming calculations. This, would include missed information not considered before, and thus, increase the prediction power. Further investigations are needed to account datasets from different treatment units, anatomic region, and different dose per fraction. PO-1704 A method for proton pencil beam scanning treatment fraction and course integrity QA G. Guterres Marmitt 1 , A. Pin 2 , A. Hengeveld 1 , C.O. Ribeiro 1 , J.A. Langendijk 1 , S. Both 1 1 University Medical Center Groningen, Department of Radiation Oncology, Groningen, The Netherlands; 2 Ion Beam Applications, IBA research, Louvain-la-Neuve, Belgium Purpose or Objective In radiation therapy, it is crucial to ensure that the treatment dose delivered (TDD) matches the physician approved planned dose (PD). Uncertainties in the treatment planning system (TPS) dose calculations, machine delivery errors or plan data corruption during transfer to the Proton Therapy System (PTS) might lead to significant differences between intended and delivered doses. Moreover, in our model-based clinic, patient’s treatment is planned, assessed and delivered under NTCP indicators guidance using NTCP-models. Therefore, in addition to the log-based integrity and dose reconstruction pre-treatment plan QA, we propose and validate a two-step daily treatment delivery QA by means of plan transfer integrity check and delivered dose and NTCP The plan transfer integrity check consists of an automated verification of the plan parameters sent to PTS (machine file) prior to delivery compared to the TPS plan. Dose accumulation is performed per fraction, as delivery log files are uploaded and converted into the fraction log-plan. The fraction dose is calculated on the most recent verification CT image available and warped to the planning CT to determine the dose accumulated over treatment. MCsquare is used as an independent Monte Carlo dose calculation engine, and REGGUI for DVH and NTCP indicators evaluations. The treatment of 10 head and neck cancer patients was tracked with daily integrity checks and treatment fraction and course wise dose and NTCP accumulation in our in-house developed treatment QA platform (CAPTAIN). The delivered treatment robustness was assessed for target coverage (D98), ΔNTCP xerostomia and ΔNTCP dysphagia of the proton dose in relation to the photon planned dose. accumulation workflows. Materials and Methods

Results The workflows were automated and performed during the entire treatment course. No deviations were found in the plan transfer integrity check. The dose accumulated over the treatment fraction and course varied within 2.5% and 0.3% of prescription dose, respectively for D98; within 0.9% and 0.4%, respectively for NTCP Grade 3 xerostomia; and within 0.7% and 0.5%, respectively for NTCP Grade 3 dysphagia.