ESTRO 2021 Abstract Book

S412

ESTRO 2021

radiotherapy. Future work will incorporate HarMonAE into deep-learning models for treatment personalisation in radiotherapy for glioblastoma.

Proffered papers: Proffered papers 33: Outcome modelling 1

OC-0526 Deep learning based time-to-event prediction for a large multicentric cohort of H&N cancer patients E. Lombardo 1,2 , C. Kurz 1,2 , S. Marschner 1,3 , M. Avanzo 4 , V. Gagliardi 4 , G. Fanetti 5 , G. Franchin 5 , J. Stancanello 6 , S. Corradini 1 , M. Niyazi 1,7 , C. Belka 1,7 , K. Parodi 2 , M. Riboldi 2 , G. Landry 1,2 1 University Hospital, LMU Munich, Radiation Oncology, Munich, Germany; 2 Faculty of Physics, Ludwig- Maximilians-Universität München, Medical Physics, Garching, Germany; 3 German Cancer Consortium , DKTK , Munich, Germany; 4 Centro di Riferimento Oncologico di Aviano (CRO) IRCCS, Medical Physics, Aviano, Italy; 5 Centro di Riferimento Oncologico di Aviano (CRO) IRCCS, Radiation Oncology, Aviano, Italy; 6 Guerbet, SA, Villepinte, France; 7 German Cancer Consortium, DKTK, Munich, Germany Purpose or Objective Recent studies have shown that convolutional neural networks (CNNs) are able to predict clinical outcomes after radiotherapy (RT) for cancer patients, using only their segmented pre-treatment CT images. The aims of this study were to build 2D/3D CNNs for distant metastasis (DM) prognosis of H&N cancer patients, to extend the models by incorporating censoring information to produce time-dependent predictions and to assess their performance using several independent testing cohorts. Moreover, we investigated the role of image texture for the CNNs by performing a binary masking experiment on the input CTs. Materials and Methods We implemented image based 2D and 3D-CNNs, a clinical covariates based artificial neural network (ANN) and 2D/3D CNNs combined with clinical covariates (CNN+Clinical). All models were extended with a previously published survival model for neural networks to incorporate censoring information and output distant DM-free probability curves for every patient. CTs from 294 patients in four different Canadian hospitals, available on the cancer imaging archive (TCIA), were used for training and 3-fold cross-validation (CV). As independent testing cohorts, we used 136 patients treated at MAASTRO clinic (from TCIA), 497 patients treated at Princess Margaret Cancer Center (PMH) (from TCIA) and 110 patients treated at Centro di Riferimento Oncologico (CRO). All 1037 patients received RT or chemo-RT as primary treatment. The CNN input consisted of CTs masked using the primary and lymph node gross tumor volume (GTV) contours whereas the ANN used seven selected clinical variables such as tumor volume or TNM stage. For the binary masking experiment, we set all voxels inside the GTV to uniform intensity. We evaluated results by Harrell’s Concordance Index (HCI) and by looking at stratification into high and low-risk groups using the log-rank test. Results All networks achieved good HCIs on two out of three testing sets (Tab. 1), being able to reproduce the good performance reached on the Canadian CV cohort. Results on the PMH cohort were significantly better than 0.5 but at best 0.69, which might be explained with certain unknown clinical background for this cohort, e.g. missing information on surgery. We did not observe a significant drop of HCIs when using a binary masked image input. Regarding stratification capability, the best CNN models in terms of average testing HCI (2D- CNN+Clinical and 3D-CNN) and the baseline ANN were able to significantly stratify all testing cohorts. However, as visible in the Kaplan-Meier plots (Fig. 1), the difference between high and low-risk groups is less visible for the ANN compared to a CNN.