ESTRO 38 Abstract book

S513 ESTRO 38

focused on predictors that were strong (but not necessarily the strongest) and did not have much redundancy. Feature selection was based on statistical stability of features in the teaching set. Corrections for multiple hypothesis testing were applied whenever appropriate. Results The table summarizes our findings. When including texture features in addition to non-texture features, the best AUC for a single feature improved in both the training set and the testing set. When combining three texture features into a model, the best AUC in the training set was better than for a single feature. However, this improvement was not seen in the testing set. The likely explanation is that while standardization reduces the differences between the two datasets, it cannot eliminate them.

our model performed (Fig. 2). This method was also used to predict DLCO and FVC % predicted 12 months post-RT, with r-squared values of 71% and 74% respectively. A 10- fold cross-validation indicated that the model was over fitted using all 12 PCs, therefore 4 PCs with a p value greater than 0.05 were removed to resolve this issue.

Fig. 2: Measured vs predicted FEV 1 % predicted 12 months post-RT. Predicted values were obtained using PCAmix regression ( ▲ ) and MLR regression ( ■ ) models. Conclusion The PCAmix regression method shows great potential in combing PFTs, dosimetry and demographic parameters for the prediction of lung function post-RT for lung cancer patients. This method must be validated with a larger sample size in the future. PO-0949 Improved external validation performance of predictive radiomics models using statistical methods A. Chatterjee 1 , M. Vallières 1 , A. Dohan 2 , I. Levesque 1 , Y. Ueno 2 , S. Saif 2 , C. Reinhold 2 , J. Seuntjens 1 1 McGill University, Medical Physics, Montreal, Canada ; 2 McGill University, Radiology, Montreal, Canada Purpose or Objective In radiomics, a predictive model may underperform on data from independent institutions. Region-of-interest contouring variability and image acquisition differences are two possible causes. While models can be made robust by inter-institutional data pooling, it is unlikely that the combined dataset reflects the global distribution of radiomic features. We aimed to create a statistical methodology to improve outcome prediction on external datasets of uterine adenocarcinoma patients with endpoints of (a) lymphovascular space invasion (LVSI), and (b) FIGO stage, grouped as early (IA) and advanced. Material and Methods The central idea, developed by us in earlier work, involves (a) creating balanced training and testing sets by under- sampling the majority class, and (b) standardizing training and testing sets separately. Standardization makes a feature distribution have zero mean and unit standard deviation. Standardizing features separately for each dataset reduces feature variability between datasets. Teaching set (used for training and validation) contained 94 samples (Hospital X) and testing set comprised 63 samples (Hospital Y). 6 different MRI sequences were available for each patient. Extracted features followed Image Biomarker Standardisation Initiative recommendations. They were divided into non-texture (e.g., morphological, histogram-based) and texture (e.g., matrix-based), to see if any benefit resulted from using texture features, which are harder to interpret. The 2 prediction approaches were: (i) using single features, and (ii) combining only three features, using basic machine learning tools (e.g., Naïve Bayes) to avoid over-training. Single feature selection focused on picking the strongest predictive features whereas combined feature selection

Conclusion Our methodology yielded model performances that are statistically significant and match (FIGO stage) or surpass (LVSI) the performance of an expert radiologist. Due to the statistical nature of the approach, it can be applied to diverse scenarios. We are applying this technique when modeling the risk of distant metastasis in head and neck squamous cell carcinoma datasets collected from 6 different hospitals. PO-0950 Determining the radiodensity range for data- driven quantification of radiation-induced lung fibrosis L.M. Wang 1 , A. Chatterjee 2 , A. Semionov 3 , J. Tsui 4 , S. Lee 5 , I. Yang 4 , Y. Al Bulushi 4 , J. Seuntjens 2 , N. Ybarra 6 1 McGill University, Medical Physics Unit, Montreal, Canada ; 2 Cedar's Cancer Centre, Medical Physics Unit, Montreal, Canada ; 3 Royal Victoria Hospital, Radiology, Montreal, Canada ; 4 Cedar's Cancer Centre, Radiation Oncology, Montreal, Canada ; 5 Memorial Sloan Kettering Cancer Centre, Memorial Sloan Kettering Cancer Centre, New York, USA ; 6 Montreal University Healthcare Centre - Research Institute, Radiobiology, Montreal, Canada Purpose or Objective Many data-driven strategies have been developed to quantify radiation-induced lung fibrosis (RILF) in patients after thoracic radiotherapy (RT). These methods make use of the unique distributions of voxel radio-densities, measured in Hounsfield Units (HU), in computed tomography (CT) images of the lung. These unique HU distributions can indicate the presence of certain interstitial diseases through measurable variations. However, to our knowledge, there is no empirical and quantitative evidence to support exact HU ranges that best characterize and delineate the radio-density