ESTRO 38 Abstract book

S576 ESTRO 38

simulation CT datasets were collected. For each dataset, six CTVs were contoured on MRI per International Spine Radiosurgery Consortium (ISRC) consensus guidelines. The bulk density override function of the integrated MR-linac treatment planning system was used to assign a tissue- equivalent density (1.02 g/cc) to all non-bony anatomy except lung-equivalent (0.25 g/cc) as appropriate. The spinal column and adjacent rib bones were assigned a standard bone density override value of 1.12 g/cc. Corresponding CT datasets were fused as secondary images to MRI datasets. For each PTV, three plans were created and compared by dose volume histogram analysis: (1) bulk density override plan, (2) bulk density override plan recalculated on the registered CT; (3) reoptimized plan using CT alone. Results PTV coverage changed by an average of 2.5% when bulk density override plans were compared to plans using the real relative electron densities. The maximum dose to 0.035cc of the spinal cord changed by an average of 0.24 Gy (corresponding to 0.9% of the constraint dose of 28 Gy), with a maximum observed deviation of 0.83 Gy. All plans achieved maximum cord doses <28 Gy, with a maximum of 21.44 Gy). Similar differences were observed for 3, 5, 7, and 10 mm expansions around the cord and 100% and 50% conformality indices, indicating that bulk density overrides provided for robustness equivalent to CT-based plans (Figure 1).

guidelines. These delineations served as reference labels for evaluating the automated contours. The Deep Learning Contouring (DLC) model (DLC Expert TM , WorkflowBox 2.0, Mirada Medical Ltd) was trained on 63 CTs for left-, and 52 for right-sided breast cancer. The remaining 15 and 16 CTs, respectively, served as a test set for evaluation. For the same test patients, atlas based automatic segmentation (ABAS) was also carried out using atlases formed by ten representative manually contoured pairs from each training set, i.e. two sets of auto-contours were generated for each patient: DLC and ABAS. The ABAS contours provide a benchmark for evaluating the DLC predictions. The accuracy of each method was assessed by quantitative measures against reference contours. Quantitative measures were obtained by comparing the automatic with the manual delineations, using the Dice similarity coefficient (DSC) and the median surface distance (SD). Results The DSC for most CTV’s, in both left and right patients, showed that DLC out-performed ABAS. The figure shows a scatter-plot comparing matched data and DSC. The surface distances for DLC were comparable or lower than for ABAS contours, again suggesting improved accuracy.

Figure: Scatter plots of Dice scores for ABAS and DLC.

Conclusion The tested deep learning contouring algorithm is a promising tool for improving clinical target volume delineation for breast cancer. To further validate this, we plan to assess the clinical acceptability of the automatic contours by using a modified version of the Turing Test to compare them against manually drawn contours (www.autocontouring.com) PO-1037 In silico analysis of MR-only planning for simulation-free MR-guided spine SBRT O. Green 1 , L. Henke 1 , S. Rudra 1 , A. Price 1 , S. Mutic 1 , C. Robinson 1 1 Washington University School of Medicine, Radiation Oncology- Physics Division, Saint Louis, USA Purpose or Objective Spine SBRT is a proven treatment modality for primary and metastatic tumors. The proximity of the spinal cord necessitates the acquisition of either MR or CT myelography imaging. The time between diagnostic imaging, simulation, and treatment can stretch to upward of two weeks, potentially reducing the probability of favorable outcome. Application of an established online- adaptive radiotherapy workflow using a hybrid MR-linac system may allow for reduction in time to treatment. In this study, we sought to evaluate the dosimetric reliability of applying our current online adaptive radiotherapy method to enable a simulation-free workflow for spine SBRT patients, wherein the patient is imaged on the hybrid MR-linac unit, a pre-selected base plan is adapted, and treatment commences while the patient remains on-table, using a bulk density override MR-only planning approach. Material and Methods Four patient datasets were chosen to be representative of typical patient habitus observed in the clinic (1 obese, 1 over-weight, 2 non-obese; 2 male, 2 female). For each, MRI datasets from previous MR-linac treatments as well as

Figure 1. Bulk density (A-1) plan (A-2) compared to CT- based (B-1) plan (B-2), with 100%, 80%, and 50% isodoses (cyan, pink, and yellow) shown. Conclusion Using a bulk density override for spine SBRT based on MR- linac images was found to have acceptable accuracy for a simulation-free, MR-only process. A Phase I clinical trial evaluating the feasibility and safety of simulation-free MR- guided spine SBRT is underway. PO-1038 MR-only Radiation Therapy: a silent patient- friendly workflow using a light-weight, flexible coil C. Cozzini 1 , C. Bobb 2 , M. Engström 3 , S. Kaushik 4 , R. Molthen 2 , D. Rettman 2 , V. Goruganti 5 , W. Chiang 5 , F. Wiesinger 1 1 GE Healthcare, Applied Science Laboratory Europe, Garching bei München, Germany ; 2 GE Healthcare, Healthcare Imaging, Waukesha, USA ; 3 GE Healthcare, Healthcare Europe, Stockholm, Sweden ; 4 GEGlobal Research, GE Corporate, Bangalore, India ; 5 NeoCoil, MRcoil, Pewaukee, USA Purpose or Objective The ability of using a single imaging system for tumor delineation and dose calculation for Radiation Therapy (RT) planning is highly appealing in terms of clinical workflow simplification and patient experience. MRI is known for its superior soft tissue contrast when compared to CT and silent Zero Echo Time (ZTE) MR imaging was