Quantifying segmentation-driven measurement error in infant brain MRI and its impact on volume and fractal dimension reliability

Fractal dimension (FD) is increasingly used to quantify brain structural complexity, but its reliability in infant MRI depends critically on segmentation quality. Here, we establish segmentation-related reliability bounds for both volume and FD in infant brain MRI using the Baby Open Brains (BOB) dataset (71 scans, 1–9 months) with expert annotations as reference. Different automated segmentation methods (SynthSeg, SamSeg, FastSurfer, and a pre-trained MONAI UNest model) were applied. However, only SynthSeg and SamSeg produced anatomically plausible results and were therefore included in the quantitative evaluation using Dice, Intersection over Union, 95th-percentile Hausdorff distance, Normalised Mutual Information, and agreement analyses for regional volumes and fractal dimension (FD). SynthSeg outperformed SamSeg across all segmentation metrics (mean Dice > 0 . 8 for major regions) and produced volumes close to expert delineations (mean + 4 % [ − 28 % to 71%]). In contrast, SamSeg systematically overestimated ventricular and whole-brain volumes (mean + 76 % [ − 12 % to 190%]). Accuracy improved with age, consistent with increasing tissue contrast during myelination. Importantly, FD differences between automated and expert segmentations were region dependent, and Bland–Altman limits of agreement showed segmentation-related FD variability exceeding many reported between-group effects. Volume and FD deviations were positively correlated across structures, demonstrating direct propagation of segmentation bias into complexity metrics. Overall, within the methods tested, SynthSeg provided the most reliable volumetric and FD estimates in this infant cohort, while small volume and FD differences should be interpreted cautiously given segmentation-driven uncertainty.