ESTRO 2020 Abstract book

S907 ESTRO 2020

including different quantization algorithms, quantization bins, and normalization. Forty-two texture features were calculated from five texture features classes, including gray-level run length matrix (GLRLM), gray-level co- occurrence matrix (GLCOM), neighborhood gray-level difference matrix (NGTDM), and gray-level size zone matrix (GLSZM). The first fraction CBCT texture features were correlated to the default CBCT (DCBCT) used for clinically (iterative reconstruction type, standard convolution filter, and medium noise suppression) using the spearman correlation. The first fraction CBCT and pCT texture features were also correlated using the spearman correlation. Results The texture feature from the DCBCT and the various reconstruction and pre-processing settings have a strong correlation across a variety of reconstruction and pre- processing settings. The median spearman R revealed a moderate correlation between the pCT and CBCT texture features. In terms of pre-processing settings, equal quantization algorithms with 256 quantization bins produced the most consistent and highest correlations across all texture features between pCT and CBCT images. In term of reconstruction settings, the most consistent and highest correlations between pCT across all texture features occur for sharp convolution filters and very high noise suppression. Six texture features were found to be highly correlated (R>0.85 and P<0.05) across many different reconstruction and pre-processing methods. Conclusion This work demonstrated that prostate CBCT radiomics texture features are robust to many different reconstruction and pre-processing settings and correlate highly across various different reconstruction and pre- processing settings. Such texture features are the most suitable for CBCT-Based radiomics texture feature analysis and model development for prostate cancer.

down- and upsampling path with a layer size of 3 and a growth rate of 12 followed by a transition down or transition up layer, respectively. The first convolutional layer resulted in a feature map size of 32 and the bottleneck involved 4 layers. As loss function the dice coefficient was chosen.

Results A semantic segmentation network was trained to segment the GTV. The network achieved an overall dice coefficient of 0.82/0.72 for training/validation and of 0.70 for the independent test set. Figure 1 shows the dice coefficient in dependence of the axial tumor area for the test set. It can be seen that the network is not able to identify sarcomas with relatively small size. For axial tumor areas smaller than 50 cm² the mean dice coefficient is 0.31. STS of larger size are identified with an accuracy comparable to a medical expert achieving a mean dice coefficient of 0.90 above axial GTV area of 80 cm². Conclusion It was shown that the deep learning framework was capable of identifying patterns in CT images characteristic for soft tissue sarcomas in an automatic fashion. Such a network could be applied to automatic GTV segmentation which would help to speed up the process of STS treatment planning. PO-1580 CBCT-Based Radiomics of Prostate Cancer N. Dogan 1 , R. Delgadillo 1 , C. Ford 1 , F. Yang 1 , M. Studenski 1 , A. Pollack 1 , M. Abramowitz 1 1 University of Miami- Sylvester Comprehensive Cancer Center, Department of Radiation Oncology, Miami- Florida, USA Purpose or Objective Prostate Cancer is the sixth leading cause of male cancer death in the world. Recent publications have demonstrated that a variety of radiomics features extracted from different imaging modalities may have predictive value in prostate cancer. Currently, no studies have investigated the usefulness of imaging features extracted from Cone Beam CT (CBCT) images for patients treated with prostate cancer. The purpose of this work is to test the reproducibility of CBCT-Based radiomics texture features to reconstruction and pre-processing settings. Material and Methods Radiomics texture features were extracted from the gross tumor volume (GTV) from both the planning CT (pCT) and the first fraction CBCT for 20 prostate cancer patients. Parallel CBCT data sets were generated from the raw image projection data using different reconstruction types (iterative and standard), convolution filters (Sharp, Standard, and Smooth), and noise suppression filters (very high, high, medium, low, and very low). Texture features were extracted using different pre-processing settings