ESTRO 37 Abstract book

S754

ESTRO 37

61%; p=0.001), with metacronous vs synchronous metastases (2y, OS 98% vs 64%, p=0.001), chemotherapy or not after SABR (2y, OS 87% vs 53%, p=0.02). 5y OS for patients treated for breast metastases was 88%, for colorectal cancer 52% and for NSCLC 62%. No significant differences on LC was detected between different primary tumors.LC rates appears relate only to BED value. Toxicity was mild and not exceeded the grade 2. Conclusion The SABR appears as safe and effective therapy with a good rate of LC of pulmonary metastases using BED value exceeding 100Gy ( only 36pts relapsed in the treatment field with a total number of 647 treated lesions) . The influence on OS appears to be related to the time of metastases appearance (synchronous vs metacronous), primary tumor controlled or not and the use of chemotherapy after SABR. The rate of OS, confirms the possibility to use of SABR with curative intent in well selected oligometastatic patients. The better rates of OS occurs in pts treated for breast and colorectal cancer. The majority of failures was represented by extra thoracic spread, leading to necessity of more effective systemic therapies. EP-1380 Can radiomic features describe lung semantic features in NSCLC patients? E.E.C. De Jong 1 , W. Van Elmpt 2 , S. Rizzo 3 , R.T.H. Leijenaar 1 , T. Refaee 1 , L.E.L. Hendriks 4 , B. Reymen 2 , A.M.C. Dingemans 4 , P. Lambin 1 1 Maastricht University Medical Centre, The D-lab: Decision Support for Precision Medicine- GROW-School for Oncology and Developmental Biology- Maastricht Comprehensive Cancer Centre, Maastricht, The Netherlands 2 Maastricht University Medical Centre, GROW-School for Oncology and Developmental Biology- Department of Radiation Oncology- MAASTRO Clinic, Maastricht, The Netherlands 3 European Institute of Oncology, Department of Radiology, Milan, Italy 4 Maastricht University Medical Centre, GROW-School for Oncology and Developmental Biology- Department of Pulmonology, Maastricht, The Netherlands Purpose or Objective Survival rates of lung cancer are still quite low. Finding patient specific tumor characteristics will lead to more personalized treatments. Qualitative image characteristics such as cavitation (see Figure 1), ground- glass opacity or air bronchogram, have been shown to be prognostic in non-small cell lung cancer (NSCLC). However, since these semantic features typically are not automatically calculated, they suffer from interobserver variability. Radiomic features may well describe these semantic features. Therefore, in this study, we developed a model to predict these semantic features based on a combination of quantitative radiomic features, which are automatically extracted from CT images. Figure 1 : CT images of A) a tumor without cavitation and B) a tumor with cavitation.

Material and Methods On thoracic CT scans of 235 NSCLC patients (center A, training), experienced radiologists scored 13 semantic features (see table 1) for all patients. Because we assume that tumors show characteristics related to features in the whole lung we extracted 1417 features from the primary tumor, using in-house developed radiomics software. For validation, a dataset of 113 NSCLC patients from another center (center B) was used. Spearman’s correlation coefficient (≥0.85) was used to determine radiomic features that were strongly correlated. Per correlation pair, the feature with the highest average correlation was excluded to remove redundant information. Next, an ANOVA analysis was performed to select the significant radiomic features per semantic feature and model performance of a logistic regression was tested by calculating the c-index. The logistic regression model was validated in the dataset of center B. Results In total 130 radiomics were selected after removal of redundance and used for the logistic model to predict semantic features. The semantic feature models consisted of 1-15 radiomic features. Model performance is shown in table 1. Table 1 : Performance of the radiomic models predicting semantic features for the training- and validation dataset. Semantic feature c-index (training) c-index (validation) Gender 0.604 0.549