Multi-Brain

Multi-Brain: Unified segmentation of population neuroimaging data

The Multi-Brain (MB) model has the general aim of integrating a number of disparate image analysis components within a single unified generative modelling framework (segmentation, nonlinear registration, image translation, etc.). The model is described in Brudfors et al [2020], and it builds on a number of previous works. Its objective is to achieve diffeomorphic alignment of a wide variaty of medical image modalities into a common anatomical space. This involves the ability to construct a "tissue probability template" from a population of scans through group-wise alignment [Ashburner & Friston, 2009; Blaiotta et al, 2018]. Diffeomorphic deformations are computed within a geodesic shooting framework [Ashburner & Friston, 2011], which is optimised with a Gauss-Newton strategy that uses a multi-grid approach to solve the system of linear equations [Ashburner, 2007]. Variability among image contrasts is modelled using a much more sophisticated version of the Gaussian mixture model with bias correction framework originally proposed by Ashburner & Friston [2005], and which has been extended to account for known variability of the intensity distributions of different tissues [Blaiotta et al, 2018]. This model has been shown to provide a good model of the intensity distributions of different imaging modalities [Brudfors et al, 2019].