JUST: A joint universal and specific brain templates generation framework

Constructing universal and specific brain templates from a population or its constituent cohorts, respectively, is essential for neuroimaging research. Generation-based methods have excelled in learning either universal or specific templates, while showing the potential for handling large-scale data. However, a unified framework that jointly learns universal and specific templates is still lacking, resulting in limited computational efficiency. Moreover, in the construction process, existing methods mainly rely on similarity optimization between specific templates and matched-cohort images, without explicitly leveraging distinguishing information from mismatched cohorts, which may limit the learned template specificity. To address these challenges, we propose a novel brain template generation framework, dubbed JUST, that efficiently synergizes universal and specific template learning and explicitly leverages template-cohort divergence to enhance the specificity during the learning process. Firstly, JUST unifies the previously independent constructions of universal and specific templates via a versatile conditional generator that produces deformations between a universal template and specific templates. Secondly, JUST involves formulating inherent correlations among specific templates and all the cohorts, subsequently using a contrastive loss to enforce template specificity. Additionally, we present the Template Distinguishability Score (TDS), an innovative metric that quantifies the specificity of a set of templates by computing deformation norms from cross-cohort registrations. Compared to baseline generation-based methods, JUST improves efficiency by 33.6% while achieving optimal performance on 14 out of 18 metrics, evaluated across visual quality, centrality, and specificity on two public brain datasets. We further validate the proposed framework and new evaluation metric in a clinical scenario, demonstrating their potential for downstream applications.