Direct Reconstruction of Crossing Muscle Fibers in the Human Tongue Using a Deep Neural Network

Abstract

The human tongue is made entirely of muscle fibers that either group in a single direction or cross orthogonally in pairs. Reconstructing the muscle fiber orientations throughout the tongue can be beneficial for better understanding the tongue’s function in speaking, swallowing, and breathing. Diffusion weighted imaging (DWI) can quantify the anisotropic diffusion of water in muscle and has been used to image tongue muscles. To resolve crossing muscle fibers in the tongue, high angular resolution diffusion imaging (HARDI) can be used to capture the complex fiber configuration. However, existing fiber reconstruction methods, which primarily focus on white matter tracts in the human brain, do not account for the orthogonal nature of the crossing muscle fibers in the tongue. In this paper, we propose a deep convolutional neural network to directly reconstruct the crossing muscle fiber orientations in the tongue from HARDI. The network takes the spherical harmonics (SH) coefficients of the HARDI signals as input and estimates both the SH coefficients of the fiber orientation distribution function (fODF) and the fiber orientations as outputs. Signals from neighboring voxels are incorporated in each estimate to encourage spatial consistency and a novel separation loss is used to encourage orthogonality of the crossing fibers. The network predicts the fiber orientations in a fully automatic manner, without setting a threshold to extract peaks. The proposed method provides superior quantitative performance compared to two state-of-the-art methods when evaluated on synthetic tongue data with different noise levels. Application to post-mortem human tongue data revealed the complex muscle fibers of the human tongue and showed qualitative improvements over the competing methods.