Forest: White Matter Fiber Orientation Estimation Using Deep Learning in Diffusion MRI

For a reliable neural connectivity analysis in the human brain using diffusion magnetic resonance imaging (dMRI), an accurate estimation of the local white matter fiber orientation is required. The fiber orientation is typically measured in terms of orientation distribution function (ODF). In the absence of ground truth, most of the ODF estimation methods rely on prior assumptions about the physical models of the diffusion signal that may severely degrade the performance of fiber orientation estimation. Further, using simple mathematical models like diffusion tensor imaging (DTI), fiber orientation estimation becomes inaccurate in regions with multiple fiber crossings. Recently, deep learning is being used for the estimation of fiber orientations. However, this only allows possible orientations to belong to a predefined set of directions on the sphere. This leads to unavoidable discretization errors as the fiber direction can only be estimated from the available set of discrete points. To overcome this problem, we propose the idea of estimating peak fiber directions directly from the diffusion signal. In this work, we present a branched multilayer perceptron (MLP) model ‘FOREST’, for the estimation of prominent white matter fiber directions in a voxel. The model architecture is chosen carefully for estimating a fiber direction, especially in the voxels with multiple fibers. We compared the proposed FOREST model with the state-of-the-art diffusion ODF (dODF) and fiber ODF (fODF) estimation methods used traditionally for fiber orientation estimation.