Multitask gated autoencoder-based brain MRI harmonization framework

Multicenter MRI studies improve statistical power in brain research by pooling cross-center samples. However, batch effects from scanner/parameter differences cause systematic biases, resulting in unreliable conclusions. Traditional linear methods fail to harmonize nonlinear batch effects, whereas current deep learning approaches often suffer from artifacts, label dependency, and feature confounding. We propose a novel U-shaped multitask-gated autoencoder (MGA) framework to achieve efficient cross-center MRI data harmonization by dynamically suppressing batch effects while preserving biological features. Specifically, multitask gating adaptively suppresses batch-related features overlooked by normalization methods, dual cross-attention mitigates artifact–biology confounding in generative models, and multilevel supervision disentangles biological and batch features without requiring additional annotations. Experiments on multicenter large-scale datasets, independent external validation cohorts, and cross-pathology datasets demonstrate that the MGA achieves superior performance in harmonized cross-site classification tasks, with both classification metrics and image quality indices outperforming mainstream conventional methods and deep generative models. The MGA framework offers a reliable harmonization solution for multicenter MRI, enhancing the consistency of imaging data and supporting more accurate clinical diagnosis.