ESTRO 2024 - Abstract Book

S2990

Physics - Autosegmentation

ESTRO 2024

In summary, our proposed model showed a highly accurate LN CTVs segmentation for NPC RT. Our model allowed oncologists to select various LN level coverage and quickly obtain individual LN CTVs for each NPC patients. This promising work holds significant clinical applicability and is helpful to facilitate RT planning and workflow with ongoing efforts.

Keywords: nasopharyngeal carcinoma; clinical target volume

151

Digital Poster

Improving auto-contouring through decentralized and asynchronous medical image synthesis

Alexander Chebykin 1 , Peter Bosman 1,2 , Tanja Alderliesten 3

1 Centrum Wiskunde & Informatica, Evolutionary Intelligence, Amsterdam, Netherlands. 2 TU Delft, Electrical Engineering, Mathematics and Computer Science, Delft, Netherlands. 3 Leiden University Medical Center, Radiation Oncology, Leiden, Netherlands

Purpose/Objective:

Contouring organs at risk (OARs) in medical images is an important task in radiotherapy. Nowadays, great results for the automation of contouring can be obtained via deep learning (DL). However, large amounts of data are typically needed. The accumulation of data from multiple sources to create a centralized dataset is however challenging, involving privacy issues. We therefore study whether data that is synthesized via standard DL models with default hyperparameters can be useful for privacy-preserving data sharing and enhancing downstream auto contouring performance.

Material/Methods:

Experiments are retrospectively conducted on a dataset of T2-weighted MRI scans (Philips Intera, Ingenia, and Achieva 1.5T or 3T) of 109 cervical cancer patients who underwent brachytherapy. Annotations of 4 OARs (bladder, bowel, rectum, sigmoid) were available for each scan. We use 2D segmentation models and work with data slices (4134 in total). We use StyleGAN2 [1] with differentiable augmentations [2] as our Generative Adversarial Network (GAN) for synthetic data generation, and a U-Net [3] for auto-contouring. To emulate the data sharing process as it may occur in practice, the dataset available from one hospital is randomly split into two parts: “hospital A” (55 patients) and “hospital B” (54). The following procedure is executed for each hospital. Data is separated into test (10 patients), validation (10 patients), train (remaining patients). A GAN and a U-Net ( U-Net-real ) are trained, respectively, to synthesize images and to annotate the generated images as well as serve as a baseline. The GAN is used to generate 10,000 2D slices, which are then annotated by U-Net-real . A new U-Net, U-Net-syn , is trained on the synthetic data, and evaluated on the same real test set as U-Net-real .