ESTRO 37 Abstract book

S1168

ESTRO 37

EP-2119 Single vs. Multi-atlas auto-segmentation for liver and lung SBRT: comparison of two systems D. Dechambre 1 , P. Berkovic 2 , Z.L. Janvary 3 , N. Jansen 1 , A.P. Coucke 1 , V. Baart 1 , A. Gulyban 1 1 C.H.U. - Sart Tilman, Radiotherapy department, Liège, Belgium 2 U.Z. Leuven, Radiation Oncology, Leuven, Belgium 3 National Institute of Oncology, Radiotherapy department, Budapest, Hungary Purpose or Objective Atlas-based auto-segmentation has the potential to reduce the staff workload while improving delineation consistency. Our aim was to evaluate the volumetric accuracy of two commercial systems for lung and liver stereotactic treatment 1) while completing the atlas (learning curve), 2) using the full atlas (performance) and 3) determining dose volume histogram parameter (DVH) variations between the auto-generated and the reference contours. Material and Methods Forty random liver and lung cases were selected (CT, RT_DOSE, RT_STRUCT). Two systems were used, the single-atlas based Raystation ('RS’, v5.0.2, Raysearch) and the multi-atlas based RTx ('MIR’, v1.6.3, Mirada Medical). The 1-5 th case was used as base atlas. The learning phase was completed in an incremental way, where each new auto-contours was generated using an atlas consisting of all former cases, until the 20 th case. Performance was evaluated using the complete atlas on another 20 cases. Analysis included the Dice Similarity Coefficient (DSC), Jaccard index (JI), commonly contoured volumes (CCV), volumetric ratios (VR) and 95% of the Hausdorff distance (HD95%). Furthermore using the dose matrix, DVHs were generated for all volumes and the differences of relevant organs at risk specific parameters were compared. Results For volumetric comparison (DSC, JI, and CCV, VR, HD95%) liver and lung showed 12 vs. 1 out of 20 and 16 vs. 20 out of 35 parameters that improved from the learning to the performance stages for MIR vs. RS respectively (Table. 1).

identifying the features affecting survival to aid improved treatment planning. There are advanced method using Data mining that help to find relation between the features and survival. The goal of research is to identify the GBM patient variables association of GBM using rule association mining. Material and Methods To calculate GBM outcome, 55 Iranian histological proven glioblastoma patients between 2012 and 2014 were included. Radiotherapy or chemotherapy (one or both) would be performed along with the surgery. The follow up of cutoff point was 12 months after diagnosing. 20 patients variables were used to calculate outcome of GBM patients that includes Age, Gender, KPS, Chemotherapy, Radiotherapy, Cyst, Enhancing rim, Margin, Solid ,Multifoculity, Satellites, Tumor cross midline (TCM), Edema cross midline (ECM), Side, Volume of necrosis(neccal), Volume of enhancement(Encal), Volume of non-enhancement (nEncal), Volume of edema(edemcal), major length of tumor , major width of tumor. According to VASARI research project, features were selected and categorized. Automated association rule mining, Apiori algorithm, was run by IBM SPSS Modeler 18. Obtained rules were evaluated by confidence and lift formulas beside conformation of neurosurgical physician. Results 55 glioblastoma patients [male/female: 29/26; median age: 56; alive/dead: 24/31; median survival: 9.12 months] were selected. By association rule mining separated rules were generated for alive and dead class. Redundant obtained rules were manually removed based on domain knowledge. Table 1 lists the 5 the best rules for alive status.

Table 1: Rules related Alive status

There are 5 the best rules for dead status in table 2. Table 2: Rules related dead status

Conclusion In this research we analyzed GBM Patients data by association rule mining and made rules related alive and dead status. We believe that such analysis can help to identify the factors affecting survival time furthermore aid the physician to make better decision to select proper treatment.

For the great vessel and left lung, MIR showed superior performance for all parameters but statistical significance was achieved for only half them. For the heart and right lung, MIR outperformed RS (p<0.05) for all parameter but the volume ratio. Similar results were