ESTRO 2021 Abstract Book

S1537

ESTRO 2021

Eighty-six left-sided breast cancer patients from a single institute were included in the study. They were treated by 15- fraction whole breast radiotherapy up to 40.05 Gy with a concomitant boost to deliver 48 Gy to the tumor bed. There was no treatment of lymph nodes. Plans were manually created with the VoLO™ TPS (Tomotherapy, Accuray Inc, Sunnyvale, USA) with TomoDirect technique. For each patient, a second TomoDirect plan was generated, using an in-house system for automated plan generation [1], which was integrated in the RayStation TPS (Raysearch Laboratories AB, Stockholm, Sweden). The treating clinician simultaneously evaluated the two available plans for each patient, and selected the preferred plan or assigned a parity score. The SVM model was trained using for 70 arbitrarily selected patients: i) dosimetric parameters for target and OARs of both plans, ii) clinical characteristics like age (<50 or >50), iii) previous pulmonary disease (yes or not), iv) previous cardiovascular disease (yes or not), and v) the clinician’s plan preference . Training was performed through iterative optimization using leave-one-out cross validation. Finally, the 16/86 patients not used for training were used for final validation of the prediction model. Model predictions were analyzed in the light of repeated clinician evaluations on 25 patients to asses impact of intra-observer variations. Results The involved clinician preferred the autoplan for 39/86 (45%) patients, the clinical plan for 19 (22%) patients and scored plan parity for 29 (33%) patients. The SVM model achieved an AUC of 0.76 for the training patients, while Positive Predicted Values (PPV) for preference for Auto, preference for Manual and plan parity were 65.5%, 75% and 51.7%, respectively. For the independent validation set, True Positive Rates of 60% (6/10), 50% (1/2) and 50% (2/4) were found for Auto, Manual and parity, respectively. Only for 2/16 patients, the prediction was completely off; for one patient the SVM predicted Auto will the clinician’s ground truth was Manual, and for the other vice versa. In the repeat plan scoring, the clinician had a True Positive Rate of 50% (5/10), 80% (8/10) and 80% (4/5) respectively for Auto, Manual and parity with no predictions completely off.

Conclusion Clinicians’ judgements are crucial in daily plan evaluations as they represent an integrated view on all aspects of treatments plans (targets, OARs, other healthy tissues, homogeneity, conformality, low and high doses). Supervised learning of an SVM model to predict the clinician’s plan preference is encouraging and this tool could potentially become useful for aiding treatment planners and reducing clinicians’ workload. To increase model robustness, more patients will be available at the time of the congress. PO-1809 Experimental setup for pre-clinical demonstration of tissue sparing with scanning beam proton FLASH P. Poulsen 1 , M. Sitarz 1 , J.G. Johansen 1 , E. Kanouta 1 , C.E. Andersen 2 , C. Ankjærgaard 2 , C. Grau 1 , B.S. Sørensen 1 1 Aarhus University Hospital, Danish Center for Particle Therapy, Aarhus N, Denmark; 2 DTU, DTU Health Technology, Roskilde, Denmark Purpose or Objective Radiotherapy causes less normal tissue damage when delivered with ultra-high dose rates (FLASH, >40Gy/s) than with conventional dose rates (CONV). The healthy tissue sparing of FLASH has mainly been investigated with electron beams. FLASH delivered with proton beams may have broader clinical perspectives for treatment in all tumor depths. FLASH dose rates can be obtained at clinical proton facilities with the highest beam energies. However, the time pattern differs markedly from electron beam FLASH and pre-clinical data are still very scarce. We recently demonstrated a clear skin sparing effect of proton FLASH in mice using pencil beam scanning (PBS) (ESTRO 2021, submitted abstract). Here, we give a thorough characterization of the experimental setup used for this demonstration of proton FLASH. Materials and Methods The right hind leg of the mice was submerged in a water bath at 13.5cm depth and irradiated with a 250 MeV FLASH field (138 mice) or a 244 MeV CONV field (154 mice) (Fig 1A). The field size was 2cm x 3cm and the spot spacing 5mm. The dose profiles were measured with EBT-XD radiochromic films. The planned mouse leg doses were 31.7-54.4Gy (FLASH) and 24.1- 40.7Gy (CONV). The FLASH field was delivered with a requested nozzle beam current of 215nA. For FLASH, the monitor units (MU) scaling factor to compensate for ion recombination in the beam monitor chamber was determined daily with an Advanced Markus ionization chamber and once with calorimetry for comparison. For each FLASH irradiation, the dose rates