Session Item

Auto-segmentation tools have been widely implemented in neuroimaging research, enabling extensive brain segmentations to be obtained with little to no manual interaction. Applying these tools in paediatric radiotherapy research could enable analyses that include a wider range of structures than are routinely delineated, be of benefit for standardising contours in multi-centre studies and allow extensive dose-effect studies. These tools are developed in adults, so their applicability in children with cancer is unclear due to age-related differences and the presence of the tumour and other pathology. This study compares contours from three auto-segmentation tools in healthy children and in children with brain tumours.

We examined T1-weighted MRIs from 40 healthy children (age 5.0-16.4 years, median 9.3 years) and 40 children/young adults with brain tumours (including medulloblastoma, low-grade glioma and astrocytoma; age 1.8-25.2 years, median 8.9 years). Segmentations of 15 subcortical structures (accumbens, amygdala, caudate, hippocampus, pallidum, putamen and thalamus bilaterally, and brainstem) were generated by 3 open-source packages: FreeSurfer v7.2.0, the FMRIB Software Library v6.0.5 FIRST tool (FSL), and the Computational Anatomy Toolbox v12.8 (CAT). Failed segmentations are reported but excluded from further analyses. We assessed consistency between each package via comparison of each structure’s centre-of-mass (CoM), Dice similarity coefficient (DSC), 95% Hausdorff distance and average contour distance. We performed ANOVA to evaluate differences between each pairwise software comparison for each similarity metric, and t-tests to compare differences between healthy children and children with brain tumours.

Visual contour quality was acceptable (Figure 1). Segmentation failed in 11 cases (9 FSL, 1 FreeSurfer, 1 FreeSurfer/FSL), predominantly due to atypical anatomy e.g. enlarged ventricles, or poor scan quality. CoM discrepancies and DSC scores revealed significant differences (p <0.05) between FSL contours and both CAT and FreeSurfer, but not between CAT and FreeSurfer. FSL contours were significantly different from FreeSurfer in average distance analyses and from CAT in Hausdorff distance analyses. We found lower DSC scores, larger CoM and contour distances, and larger standard deviations within each metric for every structure in children with brain tumours compared to healthy children. The difference was significant in analysis considering all structures (Table 1).

The greater magnitude and variation in similarity metrics in children with brain tumours suggests auto-segmentation tools perform worse than in healthy children. Contour differences remained within 4mm in children with brain tumours. FreeSurfer and CAT were the most consistent and showed the fewest failures, and therefore show promise for use in paediatric radiotherapy research. Further work validating against clinical contours is needed.