ESTRO 37 Abstract book

ESTRO 37

S560

Conclusion The 5-mm-thick Y-90 solution seems appropriate for the suggested system. To deliver 3 Gy with the suggested system, treatment time of 18.5 min would be required. PO-1003 Feasibility of a machine learning QA system for failure detection in IORT with a mobile accelerator. R. Ayala 1 , R. Fabregat 2 , M. Alarcia 1 , J. Vilanova 1 , M.J. García 1 , G. Ruiz 1 , S. Gómez 1 , R. Jiménez 1 , M.A. López 1 1 Hospital General Universitario Gregorio Marañón, Servicio de Dosimetría y Radioprotección, Madrid, Spain 2 Hospital Universitario Marqués de Valdecilla, Servicio de Radiofísica y Protección Radiológica, Santander, Spain Purpose or Objective The aim of this work was to investigate the applicability of a failure detection system using information from log files and using machine learning tools to detect anomalies. Log files analysis has been a topic of interest for QA in conventional linacs but has not been addressed for mobile ones probably due to a more complicated process to obtain them. We focused on machine items that are involved in the correct functioning of the system. Drifts in those item values will presumably correspond with future linac failures or unexpected behaviors. In intraoperative radiotherapy (IORT) linac availability is a critical problem since the patient cannot wait for the machine to be repaired. A system that can anticipate possible failures is therefore of extreme importance. Material and Methods Linac This study has been performed with a Sordina Liac 12MeV model, a mobile electron linac with four energies: 6, 8, 10 and 12MeV. It is provided with 7 circular PMMA applicators of different diameter and 4 different beveled endings (0º, 15º, 30º and 45º). Log files The Liac accelerator does not have a control system PC interface nor a record and verify connection but Sordina has developed a software to monitor several machine parameters via RS232 cable. It is then possible to save a text file containing dosimetric information as well as some inner parameters of the machine. An in- house software has been developed to read log files, visualize and store them for further analysis (figure 1).

We have chosen to analyze cathode pulse voltage (VKAT), magnetron current , magnetron electromagnet current, thyratron reverse current (IINV) and temperature. Four different models have been fitted, one per energy. Results The SVM has a binary output which corresponds to whether the input data is contained in the learned frontier or not. Several parameters have been used which affect directly the effectiveness of the system. Overfitting the model will lead to false error detections, while underfitting renders the model insensitνive. Parameter ν in scikit-learn SVM implementation takes into account the fraction of training data that will be considered out of the boundaries and γ will transform the frontier to better fit the data. Some examples in a simplified 2D model can be shown in figure 2.

Conclusion It is indeed possible to build a QA system for a mobile IORT accelerator which is able to detect and/or anticipate linac failures. Care has to be taken selecting the parameters ν and γ of the model to avoid over or underfitting. It is important to train the model with a reasonable amount of data to properly calculate the anomaly frontier. PO-1004 Assessment Of A Commercially Available Algorithm For Deformable Image Registration M.A. De La Casa 1 , D. Zucca 1 , J. García 1 , J. Martí 1 , P. Fernández-Letón 1 1 Hospital Universitario Madrid Sanchinarro - Grupo Hospital de Madrid, Radiofísica y Protección Radiológica, Madrid, Spain Purpose or Objective The aim of this work is to evaluate the accuracy and invertibility of ANACONDA, one of the algorithms for deformable image registration (DIR) bundled in the commercial treatment planning system RayStation. Material and Methods We designed a home-made phantom using a rectal balloon and the compact bone and trabecular bone inserts from the CIRS phantom. We attached radiopaque markers to the surface of the balloon. These elements were submerged in a water tank in a position similar to that of the organs in the pelvic region. We acquired nine CT image series of the phantom varying the filling of the rectal balloon. The filling varied in volume, with 50, 70 and 90 cm 3 , as well as in composition, with three different mixtures of water and iodinated contrast: 20% of contrast, 5% and 1%. The image series were acquired with slice thickness 1.5 mm. In each imageset the balloon and the bone inserts were

SVM A one-class support vector machine is an unsupervised learning model that is able to trace the boundaries of a set of parameters, a new set of data that would be analyzed with the SVM will be categorized as either in or out of the boundaries. In this case we have chosen the scikit-learn python implementation with a radial basis function kernel (RBF). K(x,x') = exp( -γ ||x-x'||²)