ESTRO 38 Abstract book

S163 ESTRO 38

CTV definition. We present a model that addresses the limitations of small size datasets by providing a range of risks of microscopic involvement. References: [1] G. Sanguineti et al, IJROBP, V. 74(5), p.1356-64, 2009 [2] Y. Zhou et al, Int. J. of Approximate Reasoning, V. 55(5), p. 1252-1268, 2014

Conclusion We demonstrated a significant dependence of radiomic features on tumor volume in lung and head & neck CT images. Volume could be reconstructed as a linear combination of even moderately correlated features. Volume-independent features did not show substantial discriminating power. We highlight the importance of accounting for the volume effect when evaluating the predictive performance of imaging biomarkers. Unsupervised and supervised machine learning techniques provide powerful tools to investigate the robustness of radiomics models and relations / redundancies with clinical features. PV-0315 A risk assessment method including credible intervals for lymphatic metastatic spread for HNSCC B. Pouymayou 1 , O. Riesterer 1 , M. Guckenberger 1 , J. Unkelbach 1 1 Universitätsspital Zürich, Radiation Oncology, Zürich, Switzerland Purpose or Objective When treating head and neck squamous cell carcinoma (HNSCC) with radiotherapy, a large portion of the CTV aims at a prophylactic irradiation of the lymph node levels (LNL) at risk of harboring microscopic tumor despite the absence of visible metastases on imaging. We present a statistical model to estimate a probability range of ipsilateral microscopic involvement of LNL based on the patient's observed state of tumor progression, i.e. the location of metastases detected in imaging. Material and Methods We apply Bayesian Networks (BN) to model tumor progression. Each LNL is associated with 2 binary random variables (nodes on Fig. 1), the first corresponding to the hidden microscopic state of involvement and the second to the observable macroscopic state, i.e. whether metastases are detected on imaging for a LNL. The relationship between these two states is given by the sensitivity and specificity of the imaging modality. The model assumes that tumor cells can spread directly from the primary tumor (PT) to a LNL or from one LNL to the next. The possible spreading paths are given by the graph of the BN and are depicted by arrows on Fig. 1. In this work, we investigate ipsilateral lymphatic spread (levels Ib to IV) for oropharyngeal (T1-T2) tumors. For learning the BN network parameters, we reconstructed a training data set of 103 cases from a published neck dissection series [1]. To assess parameter accuracy, we apply multinomial parameter learning (MPL) [2] to estimate the posterior distribution of parameters. An inference algorithm retrieves the probabilities of microscopic involvement of LNL, given the observed macroscopic involvement. Results Fig. 1 shows the BN model together with the posterior distribution over the parameters describing lymphatic spread. For example, p02 is the probability of the primary tumor to spread to LNL II, and p23 is the probability to spread from II to III. As a result, we can define credible intervals (CI) similar to confidence intervals. We construct 3 models describing the microscopic spread: one where all the parameters are set to the CI lower bounds, one where they are set to the expectation of the distribution and one where they are set to the upper CI bound. Fig. 2 summarizes the ranges of microscopic risk of involvement for each LNL and different scenarios. For example, even for the most pessimistic model, the probability of microscopic involvement of level IV despite negative finding on imaging does not exceed 10% as long as level III is not harboring macroscopic metastases. Conclusion Bayesian networks provide a framework for combining patient characteristics with population based patterns of lymphatic progression and thereby guide an individualized

PV-0316 Deep Learning Based Automatic Grading of Colon Cancer in Digitized Histopathology Images S. Chen 1 , J. Wang 1 , W. Hu 1 , Z. Zhang 1 , M. Zhang 2 , M. Xu 2 , D. Huang 2 , W. Sheng 2 1 FuDan University Shanghai Cancer center, Radiation Oncology, Shanghai, China ; 2 FuDan University Shanghai Cancer center, Pathology, Shanghai, China Purpose or Objective The status of tumor differentiation is a crucial reference for cancer diagnosis and prognostication. However, the tumor differentiation grading can be difficult and time- consuming for pathologists. In this study we manage to use deep learning algorithm to automatic grading colon cancer. Material and Methods The workflow was presented in figure 1. 169 whole slide images (WSI) were enrolled into this study. These data were come from The Cancer Genome Atlas (TCGA). Two-