ESTRO 2021 Abstract Book

S1532

ESTRO 2021

case of overlap). To account for the PT extent, the distance was scaled with the radius of the volume-equivalent sphere. Results

Fig.2: Axial slice of the frequency weighted cumulative status (fwCS) map of the S1 cohort (left) and decreased survival areas (right). In both surgery and RCT cohorts, DSA could be identified which were located proximal to the mediastinum (Fig.2). For the surgery cohort the anatomical high-risk regions were located at right the main bronchus whereas for RCT cohort they further extended in CC direction. In the validation cohorts, the model based on distance to DSA achieved performance: AUC RCT2 [95%CI]=0.67[0.55-0.78] and AUC RCT3 =0.59[0.39-0.79] for RCT patients, but showed worse performance for surgery cohort (AUC S2 =0.52[0.30-0.74]). Closer distance to DSA was associated with worse outcome. Conclusion This explanatory analysis quantifies the value of PT location for OS prediction based on cumulative status maps. Closer distance of PT to a high-risk region at the right proximal bronchi was associated with worse prognosis in the RCT cohort. PO-1804 Local radiomics and clinical variables to predict radiation-induced pneumonitis in NSCLC patients D. Vuong 1,2 , C. Brink 2,3 , M. Bogowicz 1 , T. Schytte 4,3 , O. Hansen 4,3 , S. Long Krogh 2 , M. Guckenberger 1 , S. Tanadini-Lang 1 1 University Hospital Zurich and University of Zurich, Department of Radiation Oncology, Zurich, Switzerland; 2 Odense University Hospital, Laboratory of Radiation Physics, Department of Oncology, Odense, Denmark; 3 University of Southern Denmark, Department of Clinical Research, Odense, Denmark; 4 Odense University Hospital, Department of Oncology, Odense, Denmark Purpose or Objective Pneumonitis is a severe side effect of radiotherapy (RT). Radiomics, the extraction of a large number of quantitative features from medical images, has shown the potential to predict pneumonitis from CT imaging of healthy lung tissue. So far, this approach has limitedly accounted for regional variation in the lung. Local radiomics, i.e. radiomics from subvolumes, allows to capture regional variations in image-based heterogeneity and potentially provides additional value for pneumonitis prediction. The aim of this study was to predict pneumonitis by inclusion of local radiomics. Materials and Methods Planning CT scans of 418 locally advanced stage III NSCLC patients treated with RT from 2007-2013 were retrospectively collected. Using an in-house developed radiomics software (Z-Rad), 17 intensity and 137 texture features were extracted from subvolumes (11x11x11 voxels) of each lung side separately. An unsupervised cluster analysis was performed independently for ipsi- and contralateral lung, which was used to assign patients to a cluster based on majority voting. The dataset was randomly divided into training and test set using a 75%/25% split. Pneumonitis was clinically scored in four levels (CTCAE v4); thus, the outcome was modeled using ordinal logistic regression. The best feature subset was selected based on the performance of 5-fold cross-validation with 40 replicates. In addition to local radiomics clusters, clinical and dosimetric parameters were included into the modeling (age,gender,forced expiratory volume,forced vital capacity,performance status,log tumor volume,T,N,histology,lung side,surgery prior to RT,mean lung dose). The best- performing subset within the cross-validation was chosen as the final model for dyspnoea prediction. Results On average, 42(±15) and 53(±17) radiomics subvolumes were found for ipsi- and contralateral lung, respectively. Based on 30 and 28 uncorrelated local radiomics features, 3 clusters were identified as optimal for each lung side. In only 52.1% patients, the cluster assignment agreed between ipsi- and contralateral lung.

Fig.1: Predicted cumulative probability shown versus mean lung dose. Shaded area indicate 95% confidence interval. Among all clinical predictors and local radiomics clusters, only mean lung dose was selected for the final model (mean lung dose regression constant: 0.09Gy^-1,95%CI:0.03-0.16, Fig.1).