ESTRO 2020 Abstract book

S886 ESTRO 2020

8 and 15 fractions. These represent practical dose-volume limits that could be implemented to limit the incidence of esophageal complications.

features) for survival prediction of NSCLC patients. All quantitative imaging feature models showed stronger predicting power than the TNM stage model, among which pooling-based radiomic signature outperformed deep learning radiomics signature. The pooling-based feature selection method was proved to be able to find robust and strongly-predictive features. When dealing with deep learning radiomics, we found the pooling feature selection method showed slightly better performance than fitting it directly to CNN classification model. PO-1545 Dose-volume factors predicting esophageal after SBRT for ultra-central lung tumors A. Jackson 1 , C. Wang 2 , E. Yorke 1 , D. Gelblum 2 , A. Apte 1 , J. Yang 1 , A. Rimner 2 , A. Wu 2 1 Memorial Sloan Kettering Cancer Center, Medical Physics, New York- NY, USA ; 2 Memorial Sloan Kettering Cancer Center, Radiation Oncology, New York- NY, USA Purpose or Objective The safety of SBRT for ultra-central tumors remains unclear. We studied dose-volume correlates of esophageal toxicity after ultra-central SBRT. Material and Methods 88 patients with lung tumors abutting proximal bronchial tree (PBT) or PTV overlapping esophagus (n=76 and 23; 11 met both criteria) were included. All had 5, 8 or 15 fractions of image-guided RT with BED ≥84Gy (α/β=10 Gy). Primary endpoint was grade ≥2 esophageal toxicity (E2+, defined by CTCAE v4 criteria). Dose-volume histograms (DVHs) using linear-quadratic equivalent doses in 2 Gy fractions (LQED) were calculated for esophagus (esoph) with α/β= 3 Gy. Variables tested for correlation with E2+ using Cox proportional hazards (CPH) models were: equivalent (equiv) dose D to absolute volume V of esoph (D V ) for volumes from 0-70 cc; absolute vol V of esoph receiving at least equiv dose D (V D ) for doses from 5-80 Gy; mean, max, and min equiv doses to the esoph; age, sex, KPS, smoking history, anti-angiogenic agents, stage, GTV size, and whether GTV abutted esoph. Multivariate CPH models were tested, including the most significant D V and V D variables, using a step-up procedure. For significant variables, the log rank test was used to assess thresholds for clinical use. Results Median followup was 14.3 months (mos). There were 14 cases of E2+ (16%), mostly acute esophagitis. There were 2 cases of grade 3 (one severe esophagitis, one esophageal bleed) and 2 cases of grade 4 toxicity (both fistulae). One fistula occurred in a patient with antiangiogenic exposure. Median time to E2+ was 0.78 mos after SBRT (range: 0- 13.8 months). E2+ was significantly correlated with: D V for volumes from 0-10 Gy, with most significant volume at 3cc (p= 1.4 X 10 -6 ), Fig 1; V D for doses from 25-80 Gy with most significant equivalent dose at 60 Gy (p=6.6 X 10 -8 ); mean and max equivalent dose to the esoph (p=1.8x10 -5 and 0.016, respectively); GTV abutting esoph (p=0.0024). There were no significant multivariate models. Incidence of E2+ was significantly different with D 3cc < or > the median of 25.7 Gy (p=0.011), Fig 2, with actuarial rates of 4.5% vs. 20.5% at 6 mos and 4.5% vs. 33% at 18 mos. Incidence of E2+ was significantly different in patients with V60 Gy > 0cc (p = 0.041), with actuarial rates of 8.5% vs. 17.1% at 6 mos and 8.5% vs. 30.2% at 18 mos. Conclusion We observed a 16% rate of grade ≥2 esoph toxicity after ultra-central SBRT. Incidence of E2+ was significantly correlated with LQED to 3cc of esoph, and absolute vol receiving LQED 60 Gy. This suggests that limiting D3cc to <25.7Gy results in 18-mo actuarial incidence and a raw rate of G2+ <5%; and that limiting max LQED to 60Gy results in an 18-mo actuarial incidence and a raw rate of G2+ <8.5%. With α/β=3Gy, an LQED of 25.7Gy corresponds to 18.9, 22.2, and 26.8Gy in 5, 8 and 15 fractions while an LQED of 60 Gy corresponds to 32.0, 38.4, and 48.3Gy in 5,

PO-1546 KRAS mutation prediction of mCRC patients based on pre-treatment liver metastasis MRI images W. Zou 1 , J. Wang 1 , Y. Wang 1 , Z. Zhang 1 1 Fudan University Shanghai Cancer Center, Radiation Oncology, Shanghai, China Purpose or Objective KRAS mutation state plays a significant role in the guidance of mCRC (metastatic colorectal cancer) treatment. To evaluate the performance of radiomic signature of liver metastasis MRI image in predicting KRAS mutation state of mCRC patients. Material and Methods One-hundred-forty-six mCRC patients were enrolled into this study. These patients received treatment from year 2015 -2018 in Shanghai Cancer Center. Their pretreatment liver metastasis MR image and clinical base information were collected. KRAS mutation state are obtained from the medical records, based on the biopsy results from either the primary lesion or the liver metastases. Three region of interests (ROIs), including liver metastases and two surrounding rings (ring1 0.5cm or 1cm wide), were segmented by one radiation oncologist. Radiomics features were calculate on MRI images. After radiomics feature extraction, we use LASSO to develop our prediction model. The AUC value of was calculated on validation dataset to evaluate the model performance (figure 1).