ESTRO 38 Abstract book

S161 ESTRO 38

Oncology, Amsterdam, The Netherlands ; 12 University Medical Center Utrecht, Department of Radiotherapy, Utrecht, The Netherlands ; 13 The Netherlands Cancer Institute, Department of Head and Neck Oncology and Surgery, Amsterdam, The Netherlands Purpose or Objective Radiomics has been shown a promising prognostic biomarker for different tumor types. However, one of the major challenges in radiomics is collection of data from large, preferably multicenter cohorts, which is important for reliable model training. Sharing data between the hospitals is restricted by legal and ethical regulations. Distributed learning is a technique to train models on multicenter data without data leaving the hospitals. In this study we tested the feasibility of distributed learning with radiomics data in the context of overall survival prediction in head and neck cancer patients. Material and Methods Pretreatment, contrast-enhanced CT images were collected from 1005 head and neck cancer patients in 5 different centers Tab. 1. All patients underwent definitive radiochemotherapy. 981 radiomic features were extracted from the tumor region using Z-Rad software implementation. Two years overall survival was chosen as endpoint. For comparison, both feature selection and final classification was performed in a centralized and distributed manner. Five different models were trained, four datasets were always used for training while one of the datasets was left out for validation. The maximum relevance minimum redundancy (MRMR) method was used in the feature selection step. The MRMR was performed over 100 bootstrap samples, but in the distributed setting the selection rate was averaged over the different datasets. The final models for 2y-OS was trained with logistic regression. For the distributed solution the grid binary logistic regression was implemented. In the validation dataset, the receiver operating characteristics were compared between the models trained in the centralized and distributed manner using DeLong test (p<0.05). Additionally, patients were split into two risk groups based on median prediction from the training cohort and the misclassification error rate between the centralized and distributed models was calculated.

Conclusion We have shown that both feature selection and classification are feasible in distributed manner for radiomics data. This opens new possibility for training more reliable radiomics models by gaining access to larger multi-institutional data. PV-0313 Ventilation functional lung volumes obtained from SPECT and 4D-CT do not identify the same voxels. T. Nyeng 1 , L. Hoffmann 1 , K.P. Farr 2 , A.A. Khalil 2 , C. Grau 2 , D.S. Møller 1 1 Aarhus University Hospital, Medical Physics Department of Oncology, Aarhus C, Denmark; 2 Aarhus University Hospital, Department of Oncology, Aarhus C, Denmark Purpose or Objective Several imaging modalities are available to obtain functional lung (FL) information. The overall lung toxicity may be decreased by reducing dose to the highly functional lung tissue during radiotherapy (RT), at the expense of added dose to the less functional lung regions. Recently, methods using 4D-CT and deformable registration have been shown to produce a map of ventilated lung regions. We have compared 4D-CT derived ventilation volumes with SPECT derived ventilation volumes, generally considered the gold standard for lung ventilation imaging. Material and Methods Seventeen non-small cell lung cancer patients had a 4D- CT scan (10 bins phase sorted), and a ventilation SPECT- CT (V-SPECT) scan prior to RT treatment. For each patient, the total lung volume (V LUNG ) was delineated on the exhale phase of the 4D-CT (4D-ex) and on the CT-part of the V-SPECT. The inhale and exhale phases of the 4D- CT were deformably registered using a free form intensity based deformable registration algorithm (DRA) (MIMv6.7) and the Jacobian determinant of the deformation matrix was used to segment expanding regions within the V LUNG . The best ventilated third of the lungs (V FL-4D ) was segmented and transferred to the 4D-ex. The best ventilated third of the lungs as defined using the V-SPECT (V FL-SPECT ) was also segmented. The V-SPECT was then deformably registered to the 4D-ex, using the same DRA. V FL-SPECT was propagated to the 4D-ex and the two FL volumes were compared in terms of overlap fraction (OF) with respect to V FL-SPECT , (V FL-SPECT ∩ V FL-4D )/V FL-SPECT . The fraction of the total lung volume indicated as inferiorly functional by both methods, V nonFL-SPECT and V nonFL-4D , was calculated as (V non-FL-SPECT ∩ V non-FL-4D )/V LUNG .

Results In the centralized training, 16-18 features were selected in the MRMR step. 7-11 features were selected in distributed setting, but 78-100% of those features overlapped with features selected centrally. Fig. 1 shows the area under the curve (AUC) for models trained in the centralized and distributed manner. No difference was observed between AUC for models with features selected with two methods, indicating no loss of prognostic power with smaller feature set. Also, final logistic regression coefficients resulted in similar predictions for both methods. No significant difference in AUC was observed. The misclassification rate was 5-12% and considered only patients close to splitting threshold. For the center E model validated poorly, however these was a subgroup of solely HPV negative patients.