ESTRO meets Asia 2024 - Abstract Book

S315

Physics – Image acquisition and processing

ESTRO meets Asia 2024

395

Digital Poster

Learning-based versus established image registration methods for paediatric voxel-based analysis

Edward G A Henderson, Eliana M Vasquez Osorio, Angela Davey, Marianne C Aznar

Division of Cancer Sciences, The University of Manchester, Manchester, United Kingdom

Purpose/Objective:

Radiotherapy is an important part of treatment for many children diagnosed with cancer 1 , but can lead to life altering late effects 2 . New voxel-based analysis techniques, such as image-based data mining (IBDM), can be used to discover radiation-sensitive anatomical regions and have been shown to be applicable to paediatric populations 3 (Wilson). Spatial normalisation, using non-rigid registration (NRR), is a key step of IBDM to align the anatomy of the patient cohort to a common frame of reference. NRR is particularly challenging within paediatric populations where large heterogeneity in patient size, development and treatment position is expected. The application of deep learning for image registration is becoming widespread but its suitability for IBDM is unknown. The objective of this pilot study was to establish a registration pipeline and assess the performance of two NRR methods, 1) a popular learning-based method and 2) an established iterative optimisation-based method, evaluating their relative usefulness as a spatial normalisation method for IBDM of paediatric cohorts. Planning CT scans and organ-at-risk contours from eight paediatric patients treated with proton therapy for head and neck (HN) rhabdomyosarcoma were gathered (age range: 3.2-13.4 years). A reference patient with the median brain volume was chosen from the dataset as recommended by Wilson and colleagues 3 . Two open-source NRR methods were compared for spatial normalisation: Voxelmorph , a popular learning-based approach built using a convolutional neural network 4 (trained for inter-patient registration in an adult population of 34 HN CTs), and the iterative optimisation-based Symmetric Normalization (SyN) algorithm within the Advanced Normalization Tools ( ANTs ) package 5 . Registration performance was assessed using image- and contour-based similarity metrics: the normalised cross correlation (NCC) computed between each fixed and registered image pair 6 ; and the mean distance-to-agreement (mDTA) calculated for each of the brainstem, brain, lacrimal glands, orbits, parotid glands and cochleae following registration to the reference patient. Wilcoxon signed-rank tests were used to compare the mDTA results. Additionally, the regularity (i.e. the absence of tissue “folding”) of the deformations was assessed using the Jacobian determinant. Material/Methods:

Results:

Overall, the regularity of deformations produced by both methods was excellent, with (nearly) all Jacobian determinant values larger than zero (100% for ANTs and >99.99% for Voxelmorph), i.e., no folding was introduced.

Example registrations are shown in Figure 1. Summary statistics of the NCC and mDTA results are shown in Figure 2. The NCC was slightly better for Voxelmorph than ANTs. However, ANTs significantly outperformed Voxelmorph