Personalized White Matter Bundle Segmentation for Early Childhood

White matter segmentation methods from diffusion magnetic resonance imaging range from streamline clustering-based approaches to bundle mask delineation, but none have proposed a pediatric-specific approach. We hypothesize that a deep learning model with a similar approach to TractSeg will improve similarity between an algorithm-generated mask and an expert-labeled ground truth. Given a cohort of 56 manually labelled white matter bundles, we take inspiration from TractSeg’s 2D UNet architecture, and we modify inputs to match bundle definitions as determined by pediatric experts, evaluation to use k fold cross validation, the loss function to masked Dice loss. We evaluate Dice score, volume overlap, and volume overreach of 16 major regions of interest compared to the expert labeled dataset. To test whether our approach offers statistically significant improvements over TractSeg, we compare Dice voxels, volume overlap, and adjacency voxels with a Wilcoxon signed rank test followed by false discovery rate correction. We find statistical significance across all bundles for all metrics with one exception in volume overlap. After we run TractSeg and our model, we combine their output masks into a 60 label atlas to evaluate if TractSeg and our model combined can generate a robust, individualized atlas, and observe smoothed, continuous masks in cases that TractSeg did not produce an anatomically plausible output. With the improvement of white matter pathway segmentation masks, we can further understand neurodevelopment on a population level scale, and we can produce reliable estimates of individualized anatomy in pediatric white matter diseases and disorders.

💡 Research Summary

This paper addresses the critical gap in automated white‑matter (WM) tract segmentation for early childhood, where existing tools such as TractSeg are trained on adult high‑resolution diffusion MRI (dMRI) data and perform poorly on low‑resolution, anisotropic pediatric scans. The authors propose a pediatric‑specific deep‑learning pipeline that adapts the 2‑D UNet architecture of TractSeg but modifies three key components: (1) input features are FOD peaks derived from 1 mm isotropic resampled dMRI (b=0 and 750 s/mm² shells) to match the signal characteristics of 2–8 year‑old brains; (2) training employs subject‑level 5‑fold cross‑validation rather than a fixed train/validation/test split, mitigating over‑fitting given the modest cohort of 56 subjects; (3) the loss function is a spatially masked Dice loss that excludes background or missing‑annotation voxels from gradient computation, a strategy suited for sparse, expert‑derived labels.

The dataset consists of 56 preschool children (ages 2–8) from the Calgary Preschool cohort, each with expert‑annotated masks for 16 major WM bundles (corpus callosum sub‑parts, bilateral cingulum, SLF, IFO, ILF, uncinate, pyramidal tracts, fornix, etc.). After standard preprocessing (denoising, distortion correction, quality control via PreQual) and resampling to a common 1 mm isotropic grid, the authors compute FOD peaks, normalize them, and replace NaNs with zeros. Expert masks are binarized at a 0.5 threshold and aggregated into a multi‑channel 4‑D volume for simultaneous training.

The network mirrors TractSeg’s UNet: five encoder blocks, a bottleneck, and four decoder blocks with skip connections; each block contains two convolutions followed by batch normalization and ReLU. The final layer applies a sigmoid per channel and thresholds at 0.5 to produce binary bundle masks. Training uses Adam (lr = 1e‑3), runs for up to 250 epochs, and stops early if validation loss does not improve for 25 epochs. The best model per fold (selected by highest validation Dice) is used to infer the held‑out validation subjects, ensuring no data leakage.

Performance is evaluated with three metrics per bundle: Dice coefficient (voxel‑wise overlap), volume overlap (intersection‑over‑union), and volume overreach (extent of predicted volume beyond ground truth). The same metrics are computed for TractSeg run out‑of‑the‑box on the pediatric data. Because TractSeg’s adult‑trained atlas defines more fine‑grained sub‑bundles, the authors merge relevant TractSeg outputs (e.g., CC_3, CC_4, CC_5 → CC_Body) to create comparable masks. Statistical comparison uses Wilcoxon signed‑rank tests with false discovery rate (FDR) correction across bundles and metrics.

Results show that the proposed model significantly outperforms TractSeg on Dice and adjacency (volume overlap) for all 16 bundles (p < 0.05 after FDR). Volume overlap is also significantly better for every bundle except the right cingulum. Notably, TractSeg fails to generate any fornix voxels in 29 of 56 subjects, whereas the new model consistently produces plausible fornix masks. Visual inspection confirms smoother, more anatomically coherent segmentations, especially in small or low‑anisotropy tracts.

Beyond direct comparison, the authors combine the outputs of both models to construct a 60‑label individualized WM atlas. They augment the 16 expert‑derived bundles with additional TractSeg‑derived tracts (e.g., arcuate fasciculus, corticospinal tract, cerebellar peduncles, thalamic radiations) to achieve comprehensive coverage. In regions lacking ground truth, qualitative assessment shows that the combined masks retain anatomical plausibility and continuity, suggesting that the pediatric‑trained UNet can fill gaps where TractSeg alone fails.

The study concludes that a pediatric‑tailored UNet, leveraging masked Dice loss and robust cross‑validation, can deliver high‑quality WM bundle segmentation even with limited expert‑labeled data and lower‑quality dMRI. This advancement enables more accurate quantification of white‑matter development, facilitates population‑scale neurodevelopmental studies, and provides a foundation for individualized analyses in pediatric neurological disorders such as autism, ADHD, and neurofibromatosis type 1.