ID #1130 Quantifying the impact of growth and pathology on spatial normalisation for paediatric dose-response late effect research

Background

Voxel based analysis (VBA) enables investigation of dose-response relationships across a population by spatially normalising (SN) patient anatomies and dose distributions into a common reference space via image registration[1]. However, the age of the reference anatomy, alongside pathology, can influence registration accuracy, especially in younger (<12 y.o.) children[2]. This study quantifies the impact of age and pathology on SN accuracy in paediatric brain VBA.

Methods

T1-weighted MR images from 103 patients (46 medulloblastoma, 57 ependymoma; age 1–24y) and 50 age- and sex-matched healthy volunteers (HVs) from the PING dataset (3–21y)[3] were analysed. Seven brain structures were delineated using FastSurfer[4]: brainstem, cerebellum, hypothalamus, bilateral hippocampi, and bilateral ventricles. Scans were registered using the ANTs SyN algorithm[5] to three population-averaged paediatric templates (1.75–2.25y, 4.5–8.5y, 7–11y)[6]. Failed registrations were identified via Jacobian determinant analysis and accuracy quantified using mean distance-to-agreement (mDTA).

Results

Four registrations failed, all from individuals >12y to the 1.75-2.25y reference. The cerebellum showed the greatest errors (median mDTA: 7.5 mm medulloblastoma, 2.2 mm ependymoma, 1.7 mm HVs), likely reflecting the impact of tumour presence and surgery. All other structures had median mDTA <2.3mm. A weak but significant association (Spearman correlation) between age difference and mDTA was observed across all groups (ρ = 0.11, p = 0.02). This was most evident for HVs (ρ = 0.46, p < 0.001). For medulloblastoma (ρ = 0.18, p = 0.04) and, particularly, ependymoma (ρ=-0.06, p = 0.45), this relationship may have been obscured by pathology and surgical effects. Median mean mDTA across all structures was greatest for the 1.75-2.25y reference (1.57±0.03mm). 4.5-8.5y (1.48±0.04mm) and 7-11y (1.49±0.07mm) were similar, however 7-11y showed greater variability (standard deviation; 1.75-2.25y:0.71mm, 4.5-8.5y:0.93mm, 7-11y:1.41mm).

Conclusion

Paediatric brain registration accuracy is influenced by age mismatch and disease-related anatomical variability. A mid-childhood (4.5–8.5y) atlas provided the most robust performance across ages and diagnoses, but diagnosis-specific anatomical differences remain a source of registration variability.

1. Green, A., et al. Front Oncol. 2020;10:1178. doi:10.3389/fonc.2020.01178

2. Peterson, MR, et al. J Neuro. Pediatr. 2021;28(4):458–68. doi:10.3171/2021.2.PEDS201006.

3. Jernigan, TL, et al. Neuroimage. 2015;124:1149–54. 10.1016/j.neuroimage.2015.04.057

4. Henschel, L, et al. Neuroimage. 2020;219:117012. doi:10.1016/j.neuroimage.2020.117012.

5. Avants, BB, et al. Med Image Anal. 2008;12(1):26–41. doi:10.1016/j.media.2007.06.004.

6. Fonov, VS, et al. NeuroImage, Volume 47, Supplement 1, 2009, S102. doi.org/10.1016/S1053- 8119(09)70884-5