| Summary: | Micro-computed tomography (µCT) systems are widely used for high-resolution, nondestructive analysis of internal microvascular networks. When the scale of the targeted vessel approaches the imaging resolution limit, the level of noise becomes a limiting factor for accurate reconstruction. Denoising algorithms provided by vendors are often suboptimal for enhancing SNR of fine (vessel) features. Furthermore, the performance of existing methods has not been systematically analyzed in the context of final network reconstruction and graph model extraction. This work evaluates several standard and state-of-theart noise reduction techniques using both in silico and physical phantoms, and ex vivo rat coronary data for their ability to improve vascular network analysis. We compared five noise reduction approaches, including vendor-supplied (Gaussian smoothing), conventional (median filter) and advanced (i.e. wavelet filter with soft thresholding, block-matching collaborative filtering (BM3D), and isotropic and anisotropic total variation (TV) denoising) techniques. The latter two methods were chosen for their reported ability to preserve fine details, a prerequisite for a successful microvascular extraction. The full evaluation pipeline included the reconstruction from projection images, denoising, vascular segmentation and graph model extraction to be performed on all simulated and real image data sets. SNR, CNR and 3D NPS were quantified from denoised images, and where the ground truth was known, Sørensen-Dice coefficients, Jaccard index metrics were calculated as measures of segmentation error. The performance of the image denoising algorithms where the ground-truth was available has been assessed by computing the correlation coefficients between the residual images (obtained between the noise-free data and the denoised data) and the first derivative of the noise-free data were computed. Overall, simpler denoising techniques including the median and wavelet filters and the vendor-supplied implementations have been found to perform inadequately for segmentation of fine vessel features, particularly on real images. BM3D technique performed well in most of our tests, however ITV was the optimal choice for noise reduction and feature preservation in real data as shown by the extracted network models. Globally, ITV increased the SNR from 10.2 to 31.7 dB in a Shepp Logan phantom, doubled SNR and CNR values in a scanned physical phantom compared with BM3D, enabled the smallest vessels to be fully recovered in an in silicon phantom and achieved a near-ideal outcome in the rat coronary data.
|
|---|