ESTRO 2024 - Abstract Book

S4820

Physics - Quality assurance and auditing

ESTRO 2024

(35 patients with the primary brain tumor glioblastoma) and Linköping university hospital (25 additional glioblastoma patients), Sweden. Most of the patients (about 60%) had undergone tumor resection prior to the radiotherapy, while the tumors were only biopsied for the other patients. The DICOM images of each patient were pre-processed in the software MICE Toolkit, and included registration between the MR volumes and the CT volume (as the manual segmentations were performed in the CT geometry), bias field correction and saving each volume in NIfTI format. After pre-processing in MICE Toolkit, the MR volumes were resampled to a voxel size of 1 x 1 x 2 cubic mm, and were then cut and zero padded to 256 x 256 x 140 voxels (using nibabel). The manual segmentations were extracted from the DICOM RTSTRUCT files, converted to volumes, and then resampled and resized to the same resolution and size as the MR volumes. As the manual segmentations for the regions GTV, CTV, and brainstem may overlap, priority was given to CTV over brainstem. The face of each patient was removed using in-house developed methods. The segmentation network used was a 3D U-Net based on the MONAI framework. It used the 4 MR volumes (T1w, T1w GD, T2w FLAIR, and T2w GD) as input channels, and output the segmentation maps for GTV, CTV, and brainstem. The network has a total of 1.1 million trainable parameters, and several types of 3D augmentation were used during training (e.g., random rotations and elastic deformations). The federated learning framework FEDn was used to perform the training, and federated averaging was used to aggregate the global model in each round. A federated training between Lund and Linköping (about 400 km apart) was performed using a total of 1000 rounds, and each site used 20% of the data for validation. The clients in Lund and Linköping both used an Nvidia RTX 3090 graphics card for training. The progress was monitored in Scaleout Studio (a web interface for FEDn).

Results:

The federated training curves are shown in Figures 1 and 2. The obtained validation Dice scores for GTV, CTV, and brain stem were approximately 0.5, 0.6, and 0.8, respectively. For the brain stem the two sites obtained a similar Dice score, while for CTV the Dice score varied between 0.7 and 0.5.