Abstract Book

S1155

ESTRO 37

EP-2102 Correlating MRI radiomics with biology during therapy in cancer models: a tool for drug development? H. Woodruff 1 , F.X. Blé 2 , A. Jochens 3 , R.T.H. Leijenaar 3 , A. Ibrahim 3 , K.M. Brindle 4 , K. Heinzmann 5 , D.J.O. McIntyre 4 , P. Lambin 3 1 MAASTRO Clinic, Radiation Oncology, Maastricht, The Netherlands 2 Imaging Labs- Precision Medicine and Genomics- AstraZeneca, IMED Biotech Unit, Cambridge, United Kingdom 3 The D-Lab: Decision Support for Precision Medicine- GROW - School for Oncology and Developmental Biology- Maastricht Comprehensive Cancer Centre- Maastricht University Medical Centre, Oncology, Maastricht, The Netherlands 4 University of Cambridge, CRUK Cambridge Institute, Cambridge, United Kingdom 5 Imperial College London, Department of Surgery and Cancer, London, United Kingdom Purpose or Objective To investigate whether MRI radiomic signatures extracted from fast spin echo (FSE) images, diffusion weighted (DW) images, and apparent diffusion coefficient (ADC) maps as imaging biomarkers could be used to monitor therapy- dependent changes in tumor biology in preclinical models of pancreatic and lung cancer. Material and Methods A total of 89 mice from two centers (Cancer Research UK (CRUK), AstraZeneca (AZ)) underwent therapy and MR images were acquired at varying time points and doses. After the final imaging session, mice were euthanized and the tumour processed for histological evaluation. CRUK: 45 mice, including genetically engineered mice with spontaneous pancreatic adenocarcinoma (KPC, N=13) driven by mutant KRAS and p53 and mice with subcutaneous allografts of a KPC-derived cell line (K8484, N=32) were treated with vehicle or gemcitabine at 100 mg/kg. AZ: 44 immunodeficient mice with PC9 lung adenocarcinoma xenografts were treated daily with vehicle or gefitinib at 6.25 mg/kg. For both datasets features consisting of first-order statistics, shape, local intensity, texture, fractal dimension, multi-scale filters and textures were extracted from FSE (CRUK only), DW images (AZ only) as well as ADC maps. Radiomics features were correlated with tumor type (CRUK only), drug dosage at varying time points as well as histopathology results (AZ only) for Hematoxylin/Eosin (H&E), pEGFR, cleaved Caspase 3 (CC3), and Ki67 staining. Drug response stratification models were built using random forest (RF) classification. The value of MRI biomarkers was explored by correlating radiomics features with histopathology via multiple linear regression models, including RF and support vector machines. Feature selection and a ten-fold cross validation were performed to evaluate model performance and to avoid overfitting. Results CRUK: An RF classifier model was able to differentiate between the tumor types from pre-treatment images (AUC FSE =0.97, AUC ADC =0.92) as well as between the treated and control arms for the responsive K8484 tumors after treatment (AUC FSE =0.65, AUC ADC =0.77), while no significant difference due to treatment could be established for unresponsive KPC mice (AUCs~0.5). AZ: An RF model was able to separate the treated and control groups after 2 doses of gefitinib (AUC DWI,b700 =0.88, AUC ADC =0.75), while tumour volume changes were observed only after 3 doses of treatment. Linear regression models of ADC features exhibited significant (p<0.05) correlation with Ki67 (Pearson r = 0.71), H&E cell density and % necrotic area (r = 0.62 and 0.66, respectively), and pEGFR (r = 0.45) while no correlation was found for the CC3 histology score (r = 0.2, p = 0.47) in the AZ dataset.

iteration number and manufacturer’s post-filter width (p>0.05). 21 features were not robust (p<0.05). After standardization, range of SNR values was reduced (before: 9.2-27.7; after: 20.7-33.7). Filtering was highly efficient to harmonize 3 protocols on phantom data (p>0.05, 36 features) where the only varying parameter was manufacturer’s post-filter width. For more complex variations (PSF +/-), no benefit was observed. On original patient data, histogram features were robust (p>0.05). Differences in textural feature values were highly variable depending on PET scanner. Conclusion For the first time we proposed and validated on phantom data a simple method of harmonization. This method was more efficient to compensate for a difference induced by a variation of manufacturer post-filter width than a more complex parameter. For patient acquisitions, a combination of features calculated on original data and on filtered data appears necessary for a full set of robust features. EP-2101 Radiomic CT Features for Evaluation of PD- L1, CD8+TILs and Foxp3+TILs Expression in Stage I NSCLC Q. Wen 1 , W. Linlin 1 , Z. Jian 1 , B. Tong 1 , Y. Yong 1 , S. Xindong 1 , Y. Jinming 1 1 Shandong Cancer Hospital Affiliated to Shandong University- Shandong University, Radiation Oncology, Jinan, China Purpose or Objective Radiomic can quantify tumor phenotypic characteristics non-invasively and apply features algorithms to computed tomography (CT) images. In this study, we investigated the association between CT-based radiomic features and programmed death-ligand 1 (PD-L1), CD8+ tumor- infiltrating lymphocytes (TILs) and forkhead box protein 3+ (FOXP3+) TILs expression in patients with surgically resected stage I non-small cell lung cancer (NSCLC). Material and Methods A total of 96 patients with surgically resected NSCLC were included in the institutional review board-approved retrospective study and performed immunohistochemistry of PD-L1, CD8+TILs and FOXP3+ TILs. Clinical and demographic factors were obtained from medical records. 127 radiomic features coming from 5 different feature categories (tumor shape, intensity histogram, gray-level co-occurrence matrix, run length matrix, wavelet texture) were extracted from segmented volumes of each tumor. 48 out of 127 were considered as independent features and were performed in this analysis. Results In our univariate analysis, PD-L1 expression was significantly correlated with male sex (P = 0.003), squamous carcinoma (P < 0.001), never smoking statys (P = 0.015). And 8 radiomic features was detected a statistically significant difference between positive PD-L1 expression group and negative PD-L1 expression group in univariate analysis. A multiple logistic regression model illustrated that adding radiomic feature to clinical factors might improve the predictive value, due to the AUC increasing from 0.628 to 0.714 (P < 0.001). In addition, there was no radiomic features and clinical variables had significant correlated with CD8+ TILs and FOXP3+ TILs Radiomic features based on computed tomography of NSCLC could provide useful information regarding tumor phenotype, and the model was made up of radiomic features and clinical data could be predict the expression of PD-L1 non-invasively. expression. Conclusion