Abstract book - ESTRO meets Asia

S35 ESTRO meets Asia 2018

Abstract withdrawn

prediction accuracy of ER and OS in patients with EC, which would direct towards integrative system-based approaches with prognostic implications for personalized RT. OC-088 Planning CT-based radiomics can differentiate grade2 from grade 3 soft tissue sarcomas J.C. Peeken 1 , M. Bernhofer 2 , M.B. Spraker 3 , D. Pfeiffer 4 , A. Thamer 1 , M. Shouman 1 , F. Nüsslin 1 , N.A. Mayr 1 , B. Rost 1 , M.J. Nyflot 3 , S.E. Combs 1 1 Klinikum rechts der Isar- TU München, Department of Radiation Oncology, München, Germany 2 Technische Universität München TUM, Department for Bioinformatics and Computational Biolog y, Munich, Germany 3 University of Washington, Department of Radiation Oncology, Seattle, Germany 4 Klinikum rechts der Isar- TU München, Department of Radiology, München, Germany Purpose or Objective Treatment of soft tissue sarcomas (STS) is vastly influenced by tumor grading determined in invasive biopsies. Pre-therapeutic imaging may constitute an alternative way to characterize STS prior to therapy for optimal risk assessment. In this work, we sought to determine whether advanced quantitative imaging features ("Radiomics") of planning CT scans could be used to determine tumor grading non-invasively. Material and Methods Planning CT scans (+/- contrast agent), clinical information (AJCC staging and age) and French Fédération Nationale des Centres de Lutte Contre le Cancer (FNCLCC) tumor grading were determined from two retrospective high-grade STS cohorts of 76 patients from the Technical University of Munich (TUM) and 70 patients from the University of Washington, Seattle (UW). After manual segmentation of STS on planning CT scans under consideration of MRI studies, preprocessing steps including isotropic resampling and intensity discretization were applied. All over, 1358 radiomic features including volume, intensity histogram, texture and wavelet features were calculated using the pyradiomics package in python. 20 STS were segmented three times by independent experts. Features unstable to delineations variances with an intra-class coefficient below 0.8 were excluded. The TUM cohort was used for feature reduction and training of a machine learning (ML) model to differentiate tumor grade 2 and 3. The machine learning classifier was built using a Random Forest consisting of 500 Random Trees. Imbalance of endpoints was countered by re-weighting. Training performance stability was determined by bootstrapping the TUM cohort, validation performance by bootstrapping the UW cohort (1000 rounds, each). Results 633 features were stable against delineation variances. Tumor grade were distributed similarly between both cohorts; grade 2: TUM: 27 patients (35.5%), UW: 28 patients (40.0%); grade 3: TUM: 49 patients (64.4%), UW: 42 patients (60.0%). ML model building yielded a radiomic classifier differentiating grade 2 from grade 3 STS with an area under the receiver operator characteristic curve (AUC) of 0.63 without reaching statistical significance (95% confidence interval (CI): 0.48 - 0.76). Validation on the UW cohort showed an improved performance with an AUC of 0.69 significantly different from random (95% CI: 0.56 - 0.80). For comparison, a clinical model did not show a significant prediction of tumor grading (AUC-training: 0.64 (95%CI 0.48 - 0.76), AUC-validation: 0.54 (95% CI: 0.41-0.70)). Over all patients the radiomic score was significantly associated with tumor grading in univariate logistic regression (p=.038, odds ratio 24.0, 95% CI: 1.19- 483.7).

Purpose or Objective Most radiogenomics studies focused on identifying imaging phenotypes as a surrogate to mirror a pre-selected set of molecular tumor parameters in cancer patients, which may not fully characterize heterogeneous biological processes within the tumor region, thus compromising the power of radiogenomics to determine prognosis and guide treatment. In this study, we developed a radiogenomics framework to synergistically integrate CT radiomics signatures with gene profiles to predict 1-year early recurrence (ER) and overall survival (OS) in patients with esophageal cancer (EC) receiving concurrent chemoradiotherapy (CRT). Material and Methods Forty-nine patients diagnosed with EC who were treated with CRT were included in this study. All patients received CT scan before therapy and 468 radiomics features were extracted from tumor regions using texture analysis. Related tumor samples obtained via endoscopic biopsy were also used for gene profiling (1046 genes) via exome sequencing technology. The Mann-Whitney test was applied to assess the differences of individual radiomics and genomics features between ER and non-ER groups. The 4-fold cross validation method and the least absolute shrinkage and selection operator (LASSO) regression were employed to select optimal features from both imaging and gene features. Radiomics and genomics signatures were then built separately based on selected features. The multivariable logistic regression was used to construct a radiogenomics model via integrating radiomics with genomics signatures. The utilities of the radiomics signature, genomics signature and radiogenomics model in predicting 1-year ER were evaluated by ROC method individually. The Kaplan-Meier and log-rank tests were also used to evaluate their effectiveness in prediction of OS. Results Ten radiomics features and thirty gene expression levels showed significant differences between ER and non-ER groups. Two radiomics features (LLH_GLCM_corrm, LHH_GLSZM_SZLGE) and four genomics features (DFFA, LOC101060524, chr9: 67183179-67219393, CT45A8) were optimally selected to build the radiomics and genomics signatures separately, whose individual values in predicting 1-year ER equal to an AUC of 0.717 and 0.808, respectively. The radiogenomics model constructed showed an improved predictive value with an AUC of 0.832. Furthermore, the radiomics signature showed no significant value in prediction of OS. Instead, both the genomics signature alone and the radiogenomics model showed a significant predictive capability for OS.

The Kaplan-Meier analysis result for the radiomics signature (A), gene signature (B) and integration model (C). Conclusion A radiogenomics model integrating CT radiomics signature with gene characteristics was established to improve the