ESTRO 2023 - Abstract Book

S1901

Digital Posters

ESTRO 2023

PO-2115 federated learning of decentralized AI for predicting radiation pneumonitis in lung cancer patients

Z. Zhang 1 , Z. Wang 1 , L. Zhao 2 , A. Dekker 1 , A. Traverso 1 , L. Wee 1

1 Maastricht University Medical Centre+, Department of Radiation Oncology (Maastro), Maastricht, The Netherlands; 2 Tianjin Medical University Cancer Institute and Hospital, Department of Radiation Oncology, Tianjin, China Purpose or Objective Radiation pneumonitis (RP) is one of the common side effects of radiotherapy in lung cancer. Our previous studies have shown that radiomics or deep learning approaches can effectively predict RP. However, global collaboration is challenging due to concerns about data privacy and protection, potentially solvable via federated learning (FL). We develop a federated deep learning model to predict RP. Materials and Methods CT images were retrospectively collected from our hospital from a total of 314 patients (dataset-1, D1) diagnosed with lung cancer. Two additional datasets were used: 153 (60 Gy group, D2) and 89 (74 Gy group, D3) patients from the clinical trial RTOG 0617. The deep learning ResNet based FL model (Model FL ) was constructed in privacy-preserving manner without data transfer from the three datasets which were placed in different countries. For comparison, two local training strategies were performed, where each dataset was trained individually (Model sep ), and all datasets were assigned to a single node (Model com ). Thirty percent of the samples from each dataset were used as validation sets, and the performance of the models was evaluated by receiver operating characteristic (ROC) curves. An online platform was developed to facilitate use by clinicians without programming knowledge.