Cross-Age and Cross-Site Domain Shift Impacts on Deep Learning-Based White Matter Fiber Estimation in Newborn and Baby Brains

TL;DR

This work tackles domain shift in deep learning-based white matter FOD estimation across developing brains, focusing on cross-age and cross-site transfers between two infant cohorts. It compares Method of Moments (MoM) harmonization and fine-tuning (FT) using a U-Net-like backbone, with ground-truth FODs generated by MSMT-CSD from full multi-shell dMRI data. The study demonstrates that even small amounts of target-domain data can substantially mitigate domain shifts, with fine-tuning often yielding larger gains than MoM, and reveals that babies exhibit less microstructural variation than newborns, informing age-aware transfer strategies for pediatric diffusion MRI. These findings support practical, adaptable DL pipelines for robust FOD estimation in developing brains across diverse scanners and protocols.

Abstract

Deep learning models have shown great promise in estimating tissue microstructure from limited diffusion magnetic resonance imaging data. However, these models face domain shift challenges when test and train data are from different scanners and protocols, or when the models are applied to data with inherent variations such as the developing brains of infants and children scanned at various ages. Several techniques have been proposed to address some of these challenges, such as data harmonization or domain adaptation in the adult brain. However, those techniques remain unexplored for the estimation of fiber orientation distribution functions in the rapidly developing brains of infants. In this work, we extensively investigate the age effect and domain shift within and across two different cohorts of 201 newborns and 165 babies using the Method of Moments and fine-tuning strategies. Our results show that reduced variations in the microstructural development of babies in comparison to newborns directly impact the deep learning models' cross-age performance. We also demonstrate that a small number of target domain samples can significantly mitigate domain shift problems.

Figures (5)

• Figure 1: Diagram of the workflow separated into: (1) initial model training on the source dataset $S_{\text{train}}$, (2) MoM harmonization and model fine-tuning applied independently on the target training dataset $T_{\text{train}}$ with varying subject numbers ({1,2,5,10}), and (3) inference where the original model assesses harmonized data $T'_{\text{infr}}$, and the fine-tuned model evaluates the original target data $T_{\text{infr}}$.
• Figure 2: Comparative intra-site performance of DL models across age-specific training in dHCP (top) and BCP (bottom), showing AR and AE under 1/2-F configurations alongside the AFD Error. $\text{DL}_{\text{a}\shortrightarrow\text{b}}$ denotes models trained on "a" and tested on "b"; further fine-tuned on "b" when followed by $^{\mathsf{FT}}$.
• Figure 3: Mean FA within white matter area by postnatal age, with dHCP (×) shifted from post-menstrual to postnatal age and BCP ($\blacktriangle$), modeled with an $\arctan$ growth fit curve (red).
• Figure 4: Qualitative comparison between GT and cross-site estimated FODs on dHCP, visualized on FA map.
• Figure 5: Inter-site performance of BCP-trained models tested on dHCP (a, b), and dHCP-trained models tested on BCP (c, d), comparing MoM and FT methods using varying subjects (1, 2, 5, 10) from the target domain under single-fiber configuration, with cross-testing and self-testing serving as lower- and upper-performance bounds ("Control"), respectively.