Tensor regularized total variation for third harmonic generation brain images

Abstract

Third harmonic generation (THG) microscopy is a label-free imaging technique that shows great potential to visualize brain tumor margins during surgery. However, the complexity of THG brain images makes image denoising challenging. Anisotropic diffusion filtering (ADF) has been recently applied to reconstruct the noise-free THG images, but the reconstructed edges are in fact smooth and the existing methods are time-consuming. In this work, we propose a robust and efficient scheme for ADF to overcome these drawbacks, by expressing an ADF model as a tensor regularized total variation (TRTV) model. First, the gradient magnitude of Gaussian at each point is used to estimate the first eigenvalue of the structure tensor, with which flat and non-flat areas can be distinguished. Second, tensor decomposition is performed only in non-flat areas. Third, the robust-to-outliner Huber norm is used for the data fidelity term to maintain image contrast. Finally, a recently developed primal-dual algorithm is applied to efficiently solve the resulting convex problem. Several experiments on THG brain images show promising results.