ESTRO 2021 Abstract Book

S1417

ESTRO 2021

Conclusion Despite the limited size of the lung cancer datasets especially for GGO nodules.The proposed method has the power to synthetic the realistic GGO data to make the deep learning-based methods data to learn for “edge case”. PO-1691 A deeply supervised convolutional neural network ensemble for multilabel segmentation of pelvic OARs M. Lempart 1,2 , M.P. Nilsson 1 , J. Scherman 1 , M. Nilsson 3 , C. J. Gustafsson 1 , C.J. Gustafsson 2 , P. Munck af Rosenschöld 1,4 , L. E. Olsson 1,2 1 Skåne University Hospital, Department of Hematology, Oncology, and Radiation Physics, Lund, Sweden; 2 Lund University, Department of Translational Sciences, Medical Radiation Physics, Malmö, Sweden; 3 Lund University, Centre for Mathematical Sciences, Lund, Sweden; 4 Lund University, Department of Medical Radiation Physics, Lund, Sweden Purpose or Objective Accurate delineation of organs at risk (OAR) is a crucial step in radiation therapy (RT) treatment planning but is a manual and time-consuming process. Deep learning-based methods have shown promising results for medical image segmentation and can be used to accelerate this task. Nevertheless, it is rarely applied to complex structures found in the pelvis region, where manual segmentation can be difficult, costly and is not always feasible. The aim of this study was to train and validate a model, based on a modified U-Net architecture, for automated and improved multilabel segmentation of 10 pelvic OAR structures (total bone marrow, lower pelvis bone marrow, iliac bone marrow, lumosacral bone marrow, bowel cavity, bowel, small bowel, large bowel, rectum, and bladder). Materials and Methods For 143 patients, OARs were retrospectively contoured in planning CT volumes by two clinical experts. Five- fold cross validation training was performed by randomly cropping 3D patches with a size of 80x160x160 (depth x width x height) from CT volumes. All folds were trained for 1000 epochs using the Adam optimizer, a learning rate of 0.01 and batch size of 2. The segmentation task was treated as a multilabel problem, using sigmoid activations in the last model layer. Deep supervision was added to the network, generating secondary segmentation maps from lower resolution levels by bilinear interpolation. These were added to the main objective function (Dice + binary crossentropy) as weighted auxiliary losses, to speed up network convergence. The final model was established using ensembling, by averaging the output logits of the cross-validation models. Model performance was evaluated using the Dice coefficient and the 95 th percentile Hausdorff distance (HD95) for 15 independent test patients. Results Bone marrow structures were predicted accurately with mean Dice coefficients between 0.94–0.97 ± 0.00– 0.001(mean ± sd) and HD95 values between 2.41–3.70 ± 0.26–1.48mm (mean ± sd) (table 1) . Dice values > 90% were found for bowel cavity and the whole bowel structure, with a mean HD95 of 4.61 ± 1.16 (mean ± sd) and 3.43 ± 0.88 (mean ± sd) respectively. For small and large bowel, some outliers due to misclassification were observed, as it was difficult to distinguish the difference between the two structures, even for an experienced oncologist (data not shown). Model training took about 6 days per fold, while the average segmentation time per patient, including pre- and postprocessing, was about 4min.