S2466
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Digital Poster Highlight 3508 A framework for modelling typical brain development as a baseline for post-treatment changes in children and young adults Iolyn Chennell, Thomas Melichar, Eliana Vasquez Osorio, Marcel van Herk, Marianne Aznar, Angela Davey Division of Cancer Sciences, The University of Manchester, Manchester, United Kingdom Purpose/Objective: Children treated for brain tumours are at risk of late effects e.g. cognitive decline. Atypical development of brain substructures on MRI may be an early indicator of these late effects, but detecting such deviations requires understanding typical development. Large open-source MRI datasets allow for modelling development, but many cohorts only image a child at a single timepoint (STP). Whereas, multi-timepoint (MTP) datasets capture individual development. This study evaluates the suitability of modelling brain substructure development across childhood and adolescence using STP data alone and in combination with MTP data, to establish the best normative reference for oncology applications. Material/Methods: Brain MRI of 734 participants (1 timepoint, age 3-21 years) from the PING dataset1, and 1026 participants (3 timepoints, age 9–16 years) of the ABCD dataset2 were segmented using FastSurfer, and volumes of 20 brain structures were extracted3. Volume-age trajectories for each structure (stratified by sex) were modelled using multi-task Gaussian Processes with a shared mean trajectory across individuals, as implemented in MAGMA4. Models were first trained on STP data, and then on STP&MTP data. All models were evaluated on the MTP dataset using different strategies: the STP models on the full dataset in one pass, and the STP&MTP models via ten-fold cross- validation.Volume predictions at final timepoint were made for each participant and structure using the baseline volume and posterior mean trajectory. Prediction accuracy was assessed using z-scores based on posterior mean uncertainty, i.e., the number of standard deviations between the predicted and real
Overall, the STP models provided more statistically sound estimates than the STP&MTP models, for instance, for the hippocampus and lateral ventricles (Figure 2). It appears that adding MTP data reduced the posterior mean uncertainty (Figure 1B), which reduces flexibility to fit individual variation.
volume. Results:
Segmentations were generated and deemed plausible after visual inspection (Figure 1A). 14/20 structures increased in volume with age (e.g., cerebral white matter – Figure 1B&C), while 6 decreased (e.g., cerebral cortex).
Made with FlippingBook - Share PDF online