ESTRO 38 Abstract book

S1166 ESTRO 38

the treatment delivery, they can be used to predict the delivered dose distribution, which can be compared to the planned one to ensure that the patient receives the correct treatment. Therefore, in this work we proposed the application of machine-learning algorithms to allow the reconstruction of 2D and further 3D absorbed dose maps based on EPID signal obtained from both, Varian and Elekta machines. Material and Methods A supervised ANN algorithm consisting of a non-recurrent feed-forward multilayer model was employed, which works according to two phases - learning and recognition (like humans). In order to reconstruct 2D absorbed dose distribution from treatment delivery, during the learning phase, the creation of “neurons” was done by linking input and output data, taken from EPID images, and absorbed dose distributions from the TPS, respectively. Once the learning was over, new EPID signal (input data) was used to predict the delivered absorbed dose distribution (as output) in the recognition phase, to be compared to the TPS for treatment verification. The latter was done by means of global 2%/2mm gamma index (10% threshold). Although the learning phase was more time consuming while setting the ANN model, during the recognition phase, the model operated in an instantaneous way. Datasets from different EPID and TPS types were included, so that the developed model could be extended to Varian and Elekta machines. EPID images were taken from an aSi- 1000 EPID on a Varian Clinac 23iX and an iViewGT on Elekta Synerg during 6MV treatment delivery, with a dose rate of 600 MU/min and 400 MU/min, respectively, while 2D dose distributions of IMRT plans were calculated in Eclipse TM and Pinnacle TM , at the maximum depth dose in a water phantom. Results Learning was performed using 10 and 4 input/output datasets, respectively, from IMRT treatments. All of the used image datasets (both EPID inputs and absorbed dose distribution outputs) consisted of 384×512 and 1024×1024 pixels, respectively. In Figure 1 the result from one of the evaluated treatment plans is presented, showing the similarity between the absorbed dose distributions obtained by the implemented ANN and the originally planned. Gamma passing rates > 98% were obtained for the evaluated brain cases from Varian and Elekta instances, highlighting the ANN capability to predict the absorbed dose distribution based on EPIDs with different machines.

variation, with respect to the PTV volume. The values obtained from the point dose in a deviation range of +/- 3% have corresponded to an approximate 8-13 range of the MF. The same range of MF is observed in figure 1, for those average gamma values that are less than 0.5, obtaining more points that fail as the criterion's rigor increases. When evaluating the approval rate with 2 different criteria, figure 2, the dispersion range of the points is much higher with the strictest criterion. It is not only that the approval rate is lower, which is not necessarily unfavourable, but points are more dispersed making it more difficult to establish an evaluation criterion based on a stable averaged behaviour. In addition, the values distribution of the approval rate was evaluated with Monte Carlo simulation. Minimum Gumbel type distribution was the best fit obtained.

Conclusion Regarding compliance of the objective, with the obtained data, it could be determined that most achievable criterion in DQA prostate SBRT with the use of this technology could be: ΔD 2% / DTA 3 mm in relative with an approval rate greater than 90% keeping an average Gamma of less than 0.5 and a point dose value in +/- 3% between TPS calculated and ionization chamber measurement. Due to the high dispersion of the ΔD 2% / DTA 2 criterion, used with cones, it should not be used for the MLC, data shows that it is not possible to establish an evaluation criterion based on a stable behaviour averaging. EP-2108 Varian and Elekta quality assurance using artificial neural network based on portal imaging F. Chatrie 1 , A. Vasseur 2 , A.R. Barbeiro 3 , F. Younan 4 , J. Mazurier 4 , M.V. Le Lann 5 , X. Franceries 3 1 CRCT- UMR 1037- INSERM- Université Toulouse III-Paul Sabatier- LAAS-CNRS, Disco, Toulouse, France ; 2 Institut de Cancérologie de Bourgogne, Physique Médicale, Auxerre, France ; 3 CRCT- UMR 1037- INSERM- Université Toulouse III-Paul Sabatier, Team 15, Toulouse, France ; 4 Atrium- Clinique Pasteur – Groupe Oncorad-Garonne, Service de Radiothérapie, Toulouse, France ; 5 Institut National des Sciences Appliquées- LAAS-CNRS, Disco, Toulouse, France Purpose or Objective Artificial neural networks (ANN) applied to external beam radiation therapy, can be of great interest, especially for quality assurance (QA) based on electronic portal imaging device (EPID). As EPID images contain information about

Conclusion The ANN algorithm implemented in this work was able to perform IMRT treatment verification based on EPID, showing excellent gamma index results for the evaluated cases. Furthermore it was able to predict dose distributions with the same ANN model, depending of the training data, but regardless the machine type, which makes it extensible to Varian and Elekta QA. EP-2109 Can we improve the dosimetric values with the experience? our results with vmat in lung cancer F.J. Luna Tirado 1 , M. Rincón 1 , D. Gonsalves 1 , L. Guzmán 1 , M. Montero 1 , J.M. Penedo 1 , A. Ilundain 1 , J. Olivera 1 , E. López 1