ESTRO 2020 Abstract book

S357 ESTRO 2020

voxel level. In the present study, machine learning approaches were used to develop such model on PET/MRI data of ectopic human HNC xenografts grown in immuno-

deficient nude mice. Material and Methods

Combined PET/MRI data of n=42 HNC xenografts was analyzed including dynamic FMISO-PET (90min), diffusion- weighted (DWI) and dynamic contrast-enhanced (DCE) MRI. Animal data was processed on a voxel level (~117000 voxels) and randomly split into training (80%) and test cohorts (20%). Using K-means clustering of noise-filtered FMISO-PET time-series data, tumor voxels were categorized into five different hypoxia-associated curve types (Fig.1B). These curve types were used as response variable for classification. Three MRI parameters served as predictors: ADC and projection coefficients c1 and c2 of the first two principal components obtained by principal component analysis of the DCE-MRI data. Gradient-boosted decision trees (GBDT) and artificial neural networks (ANN) were trained and assessed by 5-fold cross-validation. ROC- AUC served as a metric to assess predictive performance. The best performing model in training was evaluated in the independent test cohort. Results Training resulted in a similar cross-validation based ROC- AUC of 0.80±0.08 (mean±SD, averaged across classes) for both GBDT and ANN. GBDT was favored over ANN as the final model for better interpretability and yielded mean ROC-AUC of 0.80±0.09 in the independent test cohort. A true vs. predicted tumor map is shown in Fig.1C-D. Parameter importance of ADC and DCE-MRI projection coefficients c1 and c2 was 20%, 39% and 41%, respectively. Conclusion A machine learning classifier for the voxel-wise prediction of hypoxia-associated dynamic FMISO-PET information based on multi-parametric MRI was trained and successfully tested on small animal data. The results give new insights into the relation between hypoxia PET and functional MRI and motivate a translation to a clinical setting for the development of novel MRI based therapy adaptation strategies. Fig. 1 : A: FMISO-PET/MR image for a xenografted tumor of the independent test cohort (hind leg, axial view). B: Five different hypoxia-associated classes of FMISO-PET uptake curves found in the training cohort by K-means clustering of FMISO-PET voxel data. Hypoxia classes are presented as mean temporal uptake (bold) ± one SD (shaded). C: Map of the intratumoral distribution of the hypoxia classes for the same slice as shown in A. D: Predicted map by machine learning of multi-parametric DWI and DCE-MRI data.

Conclusion The evolution of the hypoxic compartment during radiotherapy has predictive value for the outcome. The changes in the tumour hypoxia during the first two weeks of the treatment, in particular, have the potential to predict LRR and therefore might be used for adaptive treatment approaches. OC-0584 Intratumoral prediction of dynamic FMISO-PET information by machine learning of multi-parametric MRI R. Winter 1 , S. Leibfarth 1 , S. Boeke 1,2,3 , M. Krueger 4 , P. Mena-Romano 1 , E. Cumhur Sezgin 2 , G. Bowden 4 , J. Cotton 4 , B. Pichler 3,4 , D. Zips 2,3 , D. Thorwarth 1,3 1 University Hospital Tübingen, Radiation Oncology- Section for Biomedical Physics, Tübingen, Germany ; 2 University Hospital Tübingen, Radiation Oncology, Tübingen, Germany ; 3 German Cancer Consortium DKTK- and German Cancer Research Center DKFZ, partner site Tübingen, Heidelberg, Germany ; 4 Werner Siemens Imaging Center, Preclinical Imaging and Radiopharmacy, Tübingen, Germany Purpose or Objective Hypoxia imaging via dynamic FMISO-PET is an important prognostic marker for radiation therapy response of patients with head-and-neck cancer (HNC). In routine clinical practice, however, FMISO-PET is often subject to limited availability and time-consuming imaging protocols. Its substitution by MRI would therefore be an attractive alternative. We therefore hypothesize that a multi- parametric model comprising diffusion and perfusion related MRI parameters can predict dynamic FMISO-PET at