Beyond case-control study in neuroimaging for psychiatric disorders: Harmonizing and utilizing the brain images from multiple sites

Abstract

Recent magnetic resonance imaging (MRI) research has advanced our understanding of brain pathophysiology in psychiatric disorders. This progress necessitates re-evaluation of the diagnostic system for psychiatric disorders based on MRI-based biomarkers, with implications for precise clinical diagnosis and optimal therapeutics. To achieve this goal, large-scale multi-site studies are essential to develop a standardized MRI database, with the analysis of several thousands of images and the incorporation of new data. A critical challenge in these studies is to minimize sampling and measurement biases in MRI studies to accurately capture the diversity of disease-derived biomarkers. Various techniques have been employed to consolidate datasets from multiple sites in case-control studies. Traveling subject harmonization stands out as a powerful tool that can differentiate measurement bias from sample variety and sampling bias. A non-linear statistical model for a normative trajectory across the lifespan also strengthens the database to mitigate sampling bias from known factors such as age and sex. These approaches can enhance the alterations between psychiatric disorders and integrate new data and follow-up scans into existing life-course trajectory, enhancing the reliability of machine learning classification and subtyping. Although this approach has been developed using T1-weighted structural image features, future research may extend this framework to other modalities and measures. The required sample size and methodological establishment are needed for future investigations, leading to novel insights into the brain pathophysiology of psychiatric disorders and the development of optimal therapeutics for bedside clinical applications. Sharing big data and their findings also need to be considered.

Keywords: Harmonization; Life-course trajectory; Magnetic resonance image; Normative modeling.