Abstract Book

S1171

ESTRO 37

Conclusion Half-size amplitude binning reduced the reconstructed motion underestimation with no loss in 4DMRI quality compared with equal-size amplitude binning. EP-2127 Feasibility of Automatic Multi-Atlas Based Cardiac Segmentation in Planning CT R.N. Finnegan 1,2 , J.A. Dowling 1,3,4,5,6 , L. Holloway 1,2,7,8,9 , J. Otton 7 , E.S. Koh 7 , C. Luo 6 , A. Satchithanandha 6 , P. Atluri 10 , S. Tang 8,11 , G.P. Delaney 6,7,8,11 , V. Batumalai 2,7,8 , D.I. Thwaites 1 1 University of Sydney, School of Physics- Institute of Medical Physics, Sydney, Australia 2 Ingham Institute of Applied Medical Research, Medical Physics Research, Liverpool, Australia 3 The Australian e-Health and Research Centre, CSIRO Health and Biosecurity, Herston, Australia 4 University of Wollongong, Faculty of Engineering and Information Sciences, Wollongong, Australia 5 University of Newcastle, School of Mathematical and Physical Sciences, Newcastle, Australia 6 University of New South Wales, Faculty of Medicine, Sydney, Australia 7 University of New South Wales, South Western Sydney Clinical School, Sydney, Australia 8 Liverpool & Macarthur Cancer Therapy Centres, Department of Radiation Oncology, Liverpool, Australia 9 University of Wollongong, Centre for Medical Radiation Physics, Wollongong, Australia 10 University of Texas, Department of Molecular and Cell Biology, Dallas, USA 11 Ingham Institute of Applied Medical Research, Clinical Radiation Oncology, Liverpool, Australia Purpose or Objective Toxicity to cardiac and coronary structures is an important late morbidity for patients undergoing left- sided breast radiotherapy. Past studies [1] have relied on estimates of cardiac doses assuming standardised anatomy. Development of an automatic cardiac segmentation tool would allow individualised RT dose estimates to the whole heart and cardiac substructures. Material and Methods A dataset consisting of 20 planning computed tomography (CT) images were manually contoured by 3 independent observers; twenty structures (including whole heart, 4 chambers, coronary arteries and valves) were contoured for each image. Each observer followed a protocol based on a published reference atlas [2], and the contours were validated by a cardiologist before use in this study. To develop and validate an automatic multi-atlas- based segmentation framework a 'leave-one-out” cross- validation strategy was employed. Given a segmented atlas image it is possible to transfer the contours onto a given unsegmented target image by finding a transformation that registers the two images. This image registration is performed by first aligning the atlas and target image using a rigid transform and calculating a locally optimal transformation using deformable image registration. The manual segmentation is propagated using the same transformation and gives a potential segmentation of the target image. A comparison of two popular deformable image registration schemes: free- form (B-spline) and symmetric diffeomophic demons- based registration was investigated, with the demons producing superior registration accuracy as measured by image similarity metrics. After registration of a number of atlas images the set of propagated contours are combined using a patch-based locally weighted label fusion scheme. Measures of volume and surface accuracy (Dice similarity coefficient (DSC) and mean absolute surface distance (MASD), respectively) were used to compare automatic segmentation to the consensus segmentation from manual contours. Evaluation of the performance of the

parameters between strategies (Wilcoxon’s signed rank test; α=0.05). Results Strategy A reduced the underestimation of the reconstructed displacement significantly by on average 1.3% (range: 0%–4%) of the true motion amplitude, compared with strategy B. The average RC improved significantly by 1.3% from 94.2 % to 95.5 % when adopting strategy A; The average IBV remained the same, 1.6 mm. The average S was virtually the same for both strategies A (0.85) and B (0.86) (Figure 2).