ESTRO 37 Abstract book

ESTRO 37

S551

suggest caution in interpreting multicenter Radiomics studies or even studies carried out in a sole institution but involving several scanners. PO-0990 A Clinical Decision Support Tool Based on Active Appearance Modelling for Prostate Segmentation K. Cheng 1 , H. Li 2 , D. McLaren 3 , S. McLaughlin 4 , W. Nailon 5 1 Beijing University of Posts & Telecommunications, School of Electronic Engineering, Beijing- 100876, China 2 The University of Edinburgh, Institute for Digital Communications, Edinburgh, United Kingdom 3 Edinburgh Cancer Centre, Department of Clinical Oncology, Edinburgh, United Kingdom 4 Heriot Watt University, School of Engineering and Physical Sciences, Edinburgh, United Kingdom 5 Edinburgh Cancer Centre, Department of Oncology Physics, Edinburgh, United Kingdom Purpose or Objective In prostate cancer there is an ever-increasing amount of research into the use automatic segmentation methods for 1) identifying the gross tumour volume (GTV) for radiotherapy and 2) tracking disease pro gression on magnetic resonance imaging (MRI). The Active Appearance Model (AAM) has the potential to accurately segment the GTV and organs at risk (OARs) on MRI. However, when the conventional AAM model is used to assess post-brachytherapy MRI the model analyses the whole prostate GTV, which suffers from appearance distortions caused by the inhomogeneous distribution of disease and the pre-implanted brachytherapy seeds. The aim of this study was to improve the robustness and accuracy of the AAM model for prostate GTV definition on MRI scans of patients receiving brachytherapy. Material and Methods The training set used contained 50 pre-contoured transversal T2-weighted MR images of the prostate GTV. The image re-sampling process was guided by the image resolution and in this data set all training images were re-sampled into a 1x1x1 mm/pixel isotropic frame. To remove the appearance distortion caused by brachytherapy seeds and inhomogeneous foci distribution, a three-dimension (3D) volume was extracted at a distance of 10 pixels perpendicular to the GTV surface in both the interior and exterior directions. This 3D appearance patch inherently discards the distortion within the prostate gland and includes potential lesions in surrounding structures. The GTV appearance model interprets the GTV surface as the central surface of the appearance patch extracted, which deforms to minimise the appearance difference with the training set. Training was performed on 44 randomly selected cases and testing on the remaining 6 cases. The Dice similarity coefficient was used as a metric to evaluate the performance of the model. Results Figure 1 shows the proposed AAM search at different iterations against the ground truth GTV where the mesh color was calculated using a pairwise Euclidean distance (mm) based on the point correspondence. Table 1 shows a summary of the results obtained on the 6 test cases where the lowest Dice similarity was observed in the case with smallest GTV volume. In the cases with larger GTV volume the model performance significantly improved.

Figure 1: The color of the mesh indicates the distance error between the appearance model and the ground truth. Left: Iteration 5, differences 491; Middle: Iteration 10, differences 311; Right: Iteration 21, differences 197. 1 2 3 4 5 6

Clinical Vol (cm3)

45

42

28

87

65

58

Model Vol (cm3)

44

45

32

77

65

62

Dice

Similarity

0.88 0.77 0.84 0.79 0.93 0.89

Coefficient

Table 1 : Dice coefficient and volume obtained by the appearance model. Conclusion The proposed model has the potential to be used for automatically contouring the GTV after brachytherapy treatment. However, the 10 pixel margin used to extract the appearance patch needs to be further investigated to fit smaller GTV volumes. In addition with further work the approach has the potential to be used for tracking disease progression. PO-0991 Serial tumor radiomic features predict response of head and neck cancer treated with Radiotherapy H.E. Elhalawani 1 , A.S.R. Mohamed 1 , S. Volpe 1 , P. Yang 1 , S. Campbell 1 , R. Granberry 1 , R. Ger 1 , X. Fave 1 , L. Zhang 1 , G.E. Marai 2 , D. Vock 3 , G.M. Canahuate 4 , D. Mackin 1 , L. Court 1 , G.B. Gunn 1 , A. Rao 1 , C.D. Fuller 1 1 The University of Texas- MD Anderson Cancer Center, Radiation Oncology, Houston, USA 2 University of Illinois at Chicago- Chicago- Illinois- USA., Computer Science, Chicage, USA 3 University of Minnesota of Public Health- Minneapolis- Minnesota- USA., Biostatistics, Minneapolis, USA 4 University of Iowa- Iowa City- IA- USA, Electrical & Computer Engineering, Iowa City, USA Purpose or Objective Predicting ultimate tumor response before/during radiotherapy (RT) is key for risk stratification and subsequent treatment individualization. We analyzed radiomic features longitudinally for quantifying changes in tumoral structure in a cohort of head and neck cancer (HNC) patients. We studied how clinical and temporally- derived imaging features can be integrated into a multifactorial predictive tool of treatment outcome. Material and Methods HNC patients undergoing image-guided RT were included. Primary tumor response at the end of RT course per RECIST v1.1 was retrieved. Baseline patient and disease characteristics were recorded. A total of 155 in- treatment CT scans at days 1, 5, 10 and 15 were reclaimed. Primary gross tumor volumes were contoured. A total of 145 radiomics features were selected from the categories: intensity direct, neighborhood intensity difference (NID), grey-level co-occurrence matrix (GLCM), grey-level run length and shape, and analyzed