ESTRO 37 Abstract book

S1197

ESTRO 37

consistent dose delivery during each fraction of treatment across the junctions. The maximum doses calculated at the junction were higher in the CSI plans without FIF compared to those with FIF technique. Conclusion This paper hence proves that FIF technique is better in planning craniospinal irradiation. EP-2163 Deep Neural Networks vs Medical Physicists: An IMRT QA case study. Y. Interian 1 , G. Valdes 2 , R. Vincent 1 , C. Joey 2 , K. Vasant 2 , M. Olivier 2 , E. Gennatas 2 , T. Solberg 2 1 University of San Francisco, MS in Analytics, San Francisco, USA 2 University of California UCSF, Radiation Oncology, San Francisco CA, USA Purpose or Objective To compare the performance of deep neural networks against a technique designed by domain experts in the prediction of gamma passing rates for IMRT QA. Material and Methods 498 Intensity Modulated Radiation Therapy (IMRT) plans representative of all treatment sites were developed in Eclipse version 11 and delivered on Clinac iX or TrueBeam Linacs. 3%/3 mm local dose/distance-to-agreement (DTA) for these plans were obtained using a commercial 2D diode array. 78 physicist-designed metrics that describe different aspects of plan complexity were extracted using the MLC positions and MUs per control point. A generalized Poisson regression model, previously developed, was used to predict DTA using 78 features as input. Separately, fluence maps (not expert designed features) calculated for each plan were used as inputs to a convolution neural network (CNN) and trained by a computer scientist with no knowledge of medical physics, Figure 1. The CNNs were trained using TensorFlow and Keras. A modern architecture, inspired by the convolutional blocks of the VGG-16 ImageNet model, was constructed and implemented. To prevent overfitting and boost performance of the CNNs, synthetic data were generated by rotating and translating the fluence maps during training. Dropout, batch normalization, and data augmentation were utilized to help train the model. The performance of the CNNs was compared to the generalized Poisson regression model that used the 78 expert designed features. Results Deep neural networks without domain knowledge achieved comparable performance to a baseline system designed by domain experts. If an ensemble of deep neural networks is built, then the baseline model (not an ensemble) is outperformed, Figure 2. Conclusion Convolutional neural networks with transfer learning can predict IMRT QA passing rates without the use of expert domain knowledge. assurance (RTQA) in Irish oncology trials N. Wallace 1 , C. Skourou 2 , B. O'Neill 1 , L. O'Sullivan 3 , S. French 4 , M. Cunningham 5 1 St Luke's Radiation Oncology Network- Dublin- Ireland, Beaumont Centre, Dublin 9, Ireland 2 St Luke's Radiation Oncology Network- Dublin- Ireland, Physics, Dublin 9, Ireland 3 Clinical Trials Ireland, Clinical Trials, Dublin, Ireland 4 St Luke's Radiation Oncology Network, Dublin, Ireland 5 St Luke's Radiation Oncology Network, Radiation Oncology, Dublin, Ireland Purpose or Objective RTQA is a critical component in any multicentre radiotherapy trial. It ensures that each participating centre conforms to the standards set by the trial EP-2164 Optimising radiation therapy quality

Results For a sample workflow, the planning CT acquired was stored together with the DICOM small animal acquisition context. The entire pre-treatment data and delivery data were stored in the system, while the platform utilized SmART-ATP to automatically calculate a delivered dose after CBCT acquisition. Conclusion The developed preclinical data management platform supplies an infrastructure for preclinical research enabling data warehouse functionality capable of storing data from various sources. Furthermore, the architecture will support the analysis of large data-sets stored in the platform using image processing plug-ins of custom models with the goal to create a foundation to translate preclinical results into clinical trials and make them available globally. EP-2162 FiF Technique with Intrafractionally Modulated Junction shifts for CSI Planning with 3D- CRT. S.H. A. Ali 1 , H. Nazim 1 , R. Gohar 2 , J. Mallick 1 1 Ziauddin University Hospital, Radiation Oncology, Karachi, Pakistan 2 Aga Khan University Hospital, Radiation Oncology, Nairobi, Kenya Purpose or Objective To plan craniospinal irradiation with ‘‘field-in-field’’ (FIF) homogenization technique in combination with daily, intra-fractional modulation of the field junctions, to minimize the possibility of spinal cord overdose. Photon- based techniques for craniospinal irradiation (CSI) may result in dose inhomogeneity within the treatment volume and usually require a weekly manual shift of the field junctions to minimize the possibility of spinal cord overdose. Nowadays field-in-field technique is used to feather out the dose inhomogeneity caused by multiple fields. We have started using this technique after acquiring advanced technology machines in recent years. Material and Methods 16 patients (2 adults, 14 children) treated with 3D-CRT for craniospinal irradiation were retrospectively chosen for this analysis. These patients were planned and treated during 2016-2017. Contouring of Brain and Spine Cord and organ at risk were already done and planning done on Eclipse TM Treatment Planning System (Varian). All of these patients were planned Lateral cranio-cervical fields and posterior spinal fields were planned using a forward-planned, FIF technique. Field junctions were automatically modulated and custom-weighted for maximal homogeneity within each treatment fraction. Dose volume histogram (DVH) was used for analysis of results. A corresponding plan without FIF technique was planned and maximum dose at the junction was noted for each patient with both plans and the readings were evaluated. Results Plan inhomogeneity improved with FIF technique. Planning with daily modulated junction shifts provided