ESTRO 2024 - Abstract Book

S3755

Physics - Image acquisition and processing

ESTRO 2024

The presence of metal artifacts in the sCTs had minimal effects on both the dose calculation and the patient positioning compared to the standard planning procedure based on CT scans. However, notable discrepancies in the dose distribution were observed in cases where the artifact altered the body shape and the treated lesion was in close proximity to it. In such cases, it is advisable to conduct a standard CT scan.

Keywords: synthetic CT, brain, metal implant artifact

387

Digital Poster

Influence of Sex and Age on Synthetic CT Scan Generation

Claudia Boily 1 , Jean-Paul Mazellier 2,1 , Alex Lallement 1 , Philippe Meyer 3,1

1 ICube Laboratory, University of Strasbourg, CNRS UMR 7357, Strasbourg, France. 2 IHU Strasbourg, R&D, Strasbourg, France. 3 ICANS, Department of Medical Physics, Strasbourg, France

Purpose/Objective:

This study aims to assess the ability to extend the generation of synthetic CT scans across different age groups and genders, to demonstrate the model’s robustness. Typically, studies on the generation of pseudo CT involve a small number of patients, typically ranging from 10 to 300[1]. In the present study, the database was substantial consisting of 922 head and neck (H&N) cancer patients (60,093 image slices), a much larger database than was usually used in synthetic CT studies.

Material/Methods:

The proposed algorithm, a cycle-consistency generative adversarial network (CycleGAN)[2] was used to learn the mapping from megavoltage computed tomography (MVCT) images to kilovoltage computed tomography (kVCT) images, generating synthetic kVCT scans (skVCT). No data augmentation techniques were employed. A deformable registration was applied to obtain paired MVCT and kVCT scans. The values of the scans were clipped to [-600,400] Hounsfield Units (HU) and scaled to [-1.0,1.0] and rescaled to [-600,400] HU for evaluation. The 922 patients database consists of two-thirds male patients and one-third women patients ranging from 2 to 98 years, with a median age of 62 years. The database was substantial, allowing for training on multiple different age groups and sexes and testing on others, thereby enabling the investigation of the impact of age and sex and demonstrating the model's robustness. Four different experiments were conducted, each involving training on different subsets of the database. In the first pair of experiments, one model was trained on the male population while the other model was trained on the women population. Both models were evaluated on both sexes. The objective of these experiments was to examine whether sex had any influence on the model's learning abilities and robustness. The second experiment consisted of training the model on the adult population and evaluating the pediatric population to investigate whether the anatomical variation has an impact on the model's ability to generate synthetic CT scans. At last, a model was also trained using