Semi-automated Intensity-based Image quality assessment with Gaussian Mixture Models

Abstract

Protocol development for X-ray micro-computed tomography (microCT) requires optimisation of several factors at once. Quantifying microCT image quality enables objective selection of the most suitable protocol for the user requirements. Signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) are image quality metrics describing the clarity of a feature in relation to the image noise. SNR and CNR are typically determined by comparing intensities of user-defined regions containing the feature of interest and the background. However, this method is tedious and impractical when comparing between many 3D image datasets. Reiter et al. proposed a region of interest (ROI)-independent method of determining CNR from image intensity distributions of microCT datasets (1). MicroCT intensities are proportional to material- dependent X-ray attenuation, so it is assumed that the overall image intensity for multi-material specimens is the sum of individual Gaussian distributions for each material. Reiter et al. calculate CNR from the means and variances of Gaussians fitted for each material, isolated from the overall image intensity distribution. However, their method is limited to well-separated Gaussians, which is often not applicable in microCT for biological specimens with overlapping image intensity distributions. Here we present a semi-automated tool for determining SNR and CNR from microCT data using Gaussian Mixture Models (GMMs). GMMs estimate the mean, variance and weight of a user-specified number of Gaussian components from the image intensity distribution. These properties are then used to calculate SNR and CNR between any combination of the Gaussian components. This tool was implemented using the Python scikit-learn library (2) and has a custom ImageJ user interface. Validation was performed in silico using a phantom image with pre-defined Gaussian distributions of image intensity for three materials. The phantom image used was a microCT scan of a paraffin-embedded murine colon from a custom-built microCT setup (Nikon Metrology, UK). Materials were segmented by thresholding and image intensities from known distributions of similar specimens were assigned to each pixel, according to the materials represented. The SNR and CNR calculated by GMMs was found to be within 2.5% of the phantom image values. In conclusion, fitting GMMs to the intensity distribution offers a repeatable and objective measurement of image quality for 3D microCT datasets, even from low contrast images with overlapping distributions. Furthermore, SNR and CNR between several material components can be calculated at once without additional user selection of ROIs for each material.

More information

Identifiers

ORCID for Charalambos Rossides: orcid.org/0000-0002-7547-0256

ORCID for Sylvia Pender: orcid.org/0000-0001-6332-0333

ORCID for Orestis L. Katsamenis: orcid.org/0000-0003-4367-4147

ORCID for Peter Lackie: orcid.org/0000-0001-7138-3764

ORCID for Philipp Schneider: orcid.org/0000-0001-7499-3576

Catalogue record

Export record