ESTRO 2024 - Abstract Book

S3802

Physics - Image acquisition and processing

ESTRO 2024

Material/Methods:

An online survey was sent out to all ESTRO members in April 2023. It consisted of 98 questions divided into 4 sections, focusing on the generation and usage sCT algorithms, as shown in table 1. The second section (18 questions) was directed towards developers of sCT algorithms, while the other sections targeted both developers and users. In section 3, 5 main anatomical regions were identified to focus on: pelvis, brain, thorax, head-and-neck and abdomen. For each specific site, 11 questions were created to investigate the clinical commissioning, use and routinely QA, in terms of image and dosimetric metrics. Lastly, section 4 focuses on the clinical implementation in terms of QA practices and safety checks.

Table 1 – Survey Sections and associated question topics.

Results:

By June 2023, a total of 100 centers, predominantly from Europe (>85%), participated in the survey. Response rates varied across the 4 sections and reflected diverse clinical experiences (14% MRgRT, 69% MRI only Workflow, 17% both) and roles either as developers (17%), users (68%), or both (15%). The primary sites of sCT utilization were pelvis (60%), brain (15%), abdomen (11%), thorax (8%) and head-and-neck (6%). Notably, sCT was predominantly employed for normo-fractionated treatments (68%), followed by photon SBRT (40%), and palliative cases (28%). Proton therapy is rarely used (4%). The most widely used neural network architectures, in sequence, were conditional Generative Adversarial Networks (cGAN), and GANs. These networks predominantly operated on 1.5 T and 3 T MRI images, with T2 images accounting for over 50% of input, and T1 for 40%. Some experiences were also reported on 0.35 T with TRUFI images (10%). Less than half of developers employed paired images for network training. Moreover, only 20% conducted image selection,