ESTRO 38 Abstract book

S489 ESTRO 38

comparison: 256s vs 570s) of proton arc delivery time compared to SPArc in prostate, lung and brain cancer respectively. Absolute dose measurements showed within 2% difference compared to the plans. 2D Gamma Index (3%/3mm) showed more than 97% passing rate in both SPArc and SPArc_seq.

Conclusion Anatomical robust optimization showed superior plan robustness in comparison with the classical approach in a comprehensive multi-scenario evaluation. Anatomical variations play an important role in the overall plan robustness together with setup and range uncertainties, therefore their effect should not be underestimated or neglected. PO-0916 Energy layer switching sequence optimization algorithm for an efficiency proton arc therapy delivery X. Ding 1 , X. Li 1 , G. Liu 1 , C. Stevens 1 , D. Yan 1 , P. Kabolizadeh 1 1 Beaumont Health, Proton Therapy Center, Royal Oak, USA Purpose or Objective Spot-Scanning Proton Arc therapy (SPArc) has been a great interest of the society because of the improved dosimetric outcome. Due to the magnetic field hysteresis, it costs significant time in the energy layer switching especially switching from low energy to high energy layers. Thus, we presented a new energy layer and delivery sequence optimized SPArc algorithm (SPArc_seq) to shorten the proton arc delivery time. Material and Methods SPArc_seq includes an energy layer sorting and control point re-sampling mechanism taking into account of proton arc delivery sequence through the gantry rotation. The SPArc_seq plan is optimized for high to low energy delivery sequence instead of random layer switching which was introduced in the original SPArc algorithm. Both SPArc and the novel SPArc_seq were tested on three kinds of disease sites: prostate, lung and brain cancer. Both plan group (SPArc and SPArc_seq) were delivered at a fixed 0 degree gantry angle simulating the arc delivery sequence and energy switching in clock-wise gantry rotation. Total actual delivery time was recorded and dose measurements were performed using a 2D ion chamber array device, MatriXXONE, at 3cm depth. Results With a similar proton arc plan quality, SPArc-seq optimized plan was able to successfully reduce 56% (beam- on time comparison: 330s vs 756s) , 52% (beam-on time comparison: 305s vs 635s ) and 55% (beam-on time

Conclusion The new SPArc_seq optimization algorithm is able to effectively reduce proton arc treatment delivery time by about half compared to the original SPArc algorithm. Such findings in the proton arc delivery efficiency improvement paves the road for future clinical implementations. PO-0917 Deep Convolutional Network with transfer learning for dose prediction in VMAT prostate treatments P.G. Franco 1 , E.M. Ambroa 2 , J. Perez-Alija 1 , M. Ribas 1 , M. Colomer 2 1 Hospital de la Santa Creu i Sant Pau, Medical Physics, Barcelona, Spain ; 2 Consorci Sanitari de Terrassa, Medical Physics Unit -Radiation Oncology, Terrassa, Spain Purpose or Objective We believe that the use of Deep convolutional neural networks (DCNN) with transfer learning (widely used in other science fields with good results) offers tremendous opportunities in the field of Medical Physics. We propose its use in radiotherapy to identify suboptimal plans and guide plan optimization. We have focused the study in the prediction of rectum dose-volume histograms (DVH) for We retrospectively collected data of 134 prostate patients treated with a simultaneous integrated boost VMAT technique (70 Gy and 50.4 Gy in 28 fractions to the prostate and lymph nodes respectively). QUANTEC dose- volume constraints were taken into account as a starting point for treatment planning. VGG-16 (DCNN, ImageNet Large Scale Visual Recognition Challenge 2014) was used as our network architecture. VGG-16 is a DCNN pre-trained with more than 1.2 million natural images of 1000 object categories (ImageNet image dataset). For image classification we dropped the last three fully connected VMAT prostate patients. Material and Methods