ESTRO meets Asia 2024 - Abstract Book

S316

Physics – Image acquisition and processing

ESTRO meets Asia 2024

for the brainstem, orbits, parotid glands and cochleae (Bonferroni-corrected p-value=0.0083). The mDTA results for the brain and lacrimal glands were also superior for ANTs, but a larger sample size is required to establish significance. For both methods large outliers were observed for the lacrimal glands for one of the patients (mDTA>70mm), likely linked to contouring variation for this structure in this one patient.

Conclusion:

ANTs produced better mDTA metric results than the method employed in the study from Wilson and colleagues for all common structures (brainstem and cochleae) 3 , an encouraging indication of the applicability of ANTs for paediatric IBDM. Better image-based metrics and poorer contour-based results for Voxelmorph indicate over fitting 7 . The performance deficit of Voxelmorph could be attributed to bias in training data (only adult HN cases). However, gathering a representative paediatric dataset with which to train learning-based registrations is challenging. The primary benefit of learning-based approaches is speed, however, execution time is non-critical for the IBDM research setting, where higher quality registration is preferable. Here, the learning-based method proved inferior to an established algorithm, which itself proved to be a viable candidate for paediatric IBDM purposes. Therefore, future efforts should focus on optimising parameters of this existing algorithm for spatial normalisation use in IBDM of paediatric patient cohorts.

Keywords: paediatric RT, image registration, deep learning

References:

1. Amy Colori et al., “Paediatric radiotherapy in the United Kingdom: an evolving subspecialty and a paradigm for integrated teamworking in oncology”, 2023, BJR, 97 (1153), 21-30, doi: 10.1093/bjr/tqad028