Evaluation of neuroCombat and deep learning harmonization for multi-site magnetic resonance neuroimaging in youth with prenatal alcohol exposure

A multi-university research team has published a comparative evaluation of methods to harmonize MRI data collected across different sites and scanners, a critical challenge in neuroimaging studies of conditions like prenatal alcohol exposure (PAE). The study, accepted for ISBI 2026, specifically tested a deep learning approach called HACA3 against a established statistical method known as neuroCombat. The goal was to remove technical noise from scanner differences while preserving the biological signals related to PAE in a pediatric cohort aged 7 to 21.

The researchers analyzed data from three unique MRI scanners, processing the images with both HACA3 and neuroCombat before extracting brain volume metrics using the MaCRUISE pipeline. Their key finding was that while the AI-based HACA3 method qualitatively improved inter-site contrast in the raw images, the statistical neuroCombat method was more effective at quantitatively reducing site-related variance in the final volume measurements. Notably, HACA3's performance in preserving the biological signal of interest required a subsequent application of a statistical method like neuroCombat to reach its full potential.

This work represents a crucial validation step for applying deep learning harmonization tools like HACA3 to pediatric populations, where such methods had not been thoroughly tested. The results suggest a hybrid approach—using AI for initial image enhancement followed by statistical correction—may offer the best path forward for large-scale, multi-site neuroimaging studies aiming to detect subtle brain changes.