Extracting skeletal muscle fiber fields from noisy diffusion tensor data

Diffusion Tensor Imaging (DTI) allows the non-invasive study of muscle fiber architecture but musculoskeletal DTI suffers from low signal-to-noise ratio. Noise in the computed tensor fields can lead to poorly reconstructed muscle fiber fields. This paper describes an algorithm for producing denoised muscle fiber fields from noisy diffusion tensor data as well as its preliminary validation. The algorithm computes a denoised vector field by finding the components of its Helmholtz-Hodge decomposition that optimally match the diffusion tensor field. A key feature of the algorithm is that it performs denoising of the vector field simultaneously with its extraction from the noisy tensor field. This allows the vector field reconstruction to be constrained by the architectural properties of skeletal muscles. When compared to primary eigenvector fields extracted from noisy synthetic data, the denoised vector fields show greater similarity to the ground truth for signal-to-noise ratios ranging from 20 to 5. Similarity greater than 0.9 (in terms of fiber direction) is observed for all signal-to-noise ratios, for smoothing parameter values greater than or equal to 10 (larger values yield more smoothing). Fiber architectures were computed from human forearm diffusion tensor data using extracted primary eigenvectors and the denoised data. Qualitative comparison of the fiber fields showed that the denoised fields were anatomically more plausible than the noisy fields. From the results of experiments using both synthetic and real MR datasets we conclude that the denoising algorithm produces anatomically plausible fiber architectures from diffusion tensor images with a wide range of signal-to-noise ratios.