Abstract Book

S1174

ESTRO 37

distance (ASD), Dice index (DSC) and Jaccard index (JSC). Results The HD, ASD, DSC, JSC (mean±SD) were 28.1±18.9mm, 4.39±7.19mm, 0.732±0.157 and 0.598±0.167 for validation dataset; And these indices was 14.9±7.62mm, 2.67±1.46mm, 0.713±0.126 and 0.568±0.148 between two human radiation oncologists. T-test suggested there is no statistically significant difference between automated segmentation and manual segmentation considering DSC (p=0.61), JSC (p=0.47) and ASD (p=0.29). However, significant difference was found for HD (p=0.0029), which was sensitive to extreme value. Those results were summarized in Table 1.

Table 1: Average values obtained with their standard deviations (STD) and tolerances. Conclusion Image quality of EPID images was evaluated using the Standard Imaging QC-3 phantom. Several parameters were selected for time evaluation in order to set tolerances for future QA comparison. It is expected that we would be able to anticipate strong variations and lack of calibration of EPID devices as of deviations from the set tolerances in subsequent measurements. EP-2131 Deep learning based rectum tumor auto segmentation on MRI T2 image J. Lu 1 , J. Wang 1 , W. Hu 1 , Z. Zhang 1 , L. Shen 1 , Y. Sun 1 , G. Qing 1 1 Fudan University Cancer Hospital, Radiation Oncology, Shanghai, China Purpose or Objective Manually contouring gross tumor volume (GTV) is a crucial and time-consuming process in rectum cancer radio- therapy. This study aims to develop a deep learning based rectum tumor auto segmentation algorithm on MRI T2 images. Material and Methods Ninety-five patients with locally advanced (cT3-4 and/or cN1-2) rectal cancer treated with neoadjuvant chemoradiotherapy followed by surgery were enrolled in this study. The MRI T2 images were acquired before treatment on a 3.0 T MR scanner. A very deep convolutional networks (VGG16) based full convolution network (FCN) was established as training model (Figure 1). Considering the 3D structure of the MRI image, 5 image slices were simultaneously input into the network.

Conclusion This study showed that a simple deep learning neural network can perform a human comparable segmentation for rectum cancer based on MRI T2 images. EP-2132 Repeatability and reproducibility of radiomic features: results of a systematic review A. Traverso 1 , L. Wee 1 , A. Dekker 1 , R. Gillies 2 1 Maastricht Radiation Oncology MAASTRO clinic, Purpose or Objective Broadly generalizable predictions from radiomics-assisted models are tainted by concerns about whether the explanatory features are reproducible and repeatable. We performed a systematic review of published peer- reviewed studies that specifically tested the repeatability and/or reproducibility of radiomic features. The primary objective of the review was to provide a qualitative synthesis of factors affecting feature reproducibility and repeatability. Material and Methods The PubMed index was searched using combinations of the broad Haines and Ingui filters along with a set of text words specific to cancer, radiomics (including texture analyses), reproducibility and repeatability. Two reviewers working entirely independently and the review outcomes have been reported in compliance with PRISMA guidelines. Studies included either medical imaging of human subject and/or of radiological phantoms. Detailed analysis on quality of reporting in retrieved studies pointed towards a number of deficiencies. Qualitative synthesis of radiomic feature stability was grouped by cancer type and imaging modality. Results After screening process, the qualitative synthesis was derived from 41 studies, out of which 35 involved human subjects and 6 were exclusively concerning radiological phantoms. The PRISMA flow diagram is shown in Figure 1. Repeatability and reproducibility of radiomic features are sensitive in varying degrees to methodological details such as image acquisition settings, image reconstruction algorithm, digital image pre-processing and the software used to extract features. Intra-class and concordance correlations were the most widely used statistical metric, but arbitrarily selected cut-offs were not consistent. In general, most of the studies did not report adequate information regarding the software used to compute features, nor about the image pre-processing steps. Few studies made either the images or the computed features openly accessible. Overall, first-order radiomic features Radiotherapy, Maastricht, The Netherlands 2 Moffitt Cancer Center, Imaging, Tampa, USA

A two-phase training process was implemented to increase training efficiency. The existence of GTV on an image slice was identified in phase 1, and segmented in phase 2. Data were randomly separated into training (90%) and validation (10%) dataset for 10 folder cross- validation. The training dataset were input for FCN model training and the model performances were assessed on validation dataset. The segmentations were performed by an experienced radiation oncologist (reader 1). Additionally, 20 patients were double contoured by an independent radiologist (reader 2) for performance evaluation. Four indices were calculated to evaluate the similarity of automated and manual segmentation, including Hausdorff distance (HD), average surface