ESTRO 2020 Abstract book

S1020 ESTRO 2020

consuming. In this work, we propose a privacy preserving distributed liver tumor segmentation using deep learning. Material and Methods We previously implemented a novel U-Net model to solve the Liver Tumor Segmentation challenge (LiTS)[1]. We reproduced this work in distributed settings. We split the LiTS dataset into 69 patients for training and 22 for validation. The training set was further partitioned into three “partner subsets” (N=3) of 23 patients, simulating three medical partners. Training a model usually requires fragmenting the training set into data batches and then iterate over these batches [2]. In our setting, the model is trained on the first batch of each “partner subset” before iterating to the next batch (see Figure 1). Only model weights are saved to the cloud - thus preserving data privacy.

Figure1. Some examples of GTV-nx contours between manual delineation and automatic delineation from three models.

Figure2. Some examples of GTV-nd contours between manual delineation and automatic delineation from three models. Conclusion Firstly, our approach had higher DICE than previously reported study, this may cause by using single institution. Secondly,almost all reported studies have described the auto-segmentation of GTVnx, and few studies have simultaneously referred to the auto-segmentation of GTVnd. delineation of GTVnd still needs to be done manually, which makes contour delineation still time- consuming. we proposed a convolutional neural network to automatically segment GTVnx and GTVnd simultaneously, which improves the speed of delineation.Thirdly,the results suggest that in the field of deep learning, especially medical image analysis, although there is a lack of high- quality marker data, when the deep learning algorithm is good enough, it can also obtain very satisfactory performance. Obtaining lots of high-quality medical marker images is time consuming and difficult, and this problem may be avoided by improving the deep learning algorithm. PO-1744 Privacy preserving distributed liver tumor segmentation F. Zerka 1,2 , A. Vaidyanathan 1,2 , S. Barakat 2 , M. Benjamin 2 , L. Ralph T.H. 1,2 , W. Sean 2 , L. Philippe 1 1 Maastricht University, Precision Medicine, Maastricht, The Netherlands ; 2 Oncoradiomics, Research and Development, Liege, Belgium Purpose or Objective Segmentation of GTV is a major issue in radiotherapy. Automatic segmentation is the obvious way to go but it requires access to larger amounts of medical data. Medical data is greatly sensitive and highly protected by law and ethics; making the access of such data harder and time

Figure 1 (A) training of U-Net models sequentially in every iteration, the last model of the previous iteration is used to initialize the first model of the next iteration (B) detailed explanation of every iteration: the first cite trains a model using a batch “M”, the gradient of the resulting model is then saved to cloud, and then the next cite will import the that gradient and continue training using the corresponding batch of their data. Results We have compared the results of distributed and centralized approaches on the validation set. There is no significant difference between distributed and centralized training; with a dice score coefficient (DSC) of 0.70 and 0.72 respectively, and a sensitivity of 0.82 and 0.85 respectively. Conclusion We showed that distributed learning enables to train models as effectively as centralized approach. Moreover,