Background. Advances in three-dimensional (3D) shape capture technology have made powerful shape analyses, such as geometric morphometrics, more feasible. While the highly accurate micro-computed tomography (μCT) scanners have been the “gold standard,” recent improvements in 3D surface scanner resolution may make this technology a faster, more portable, and cost-effective alternative. Several studies have already compared the two scanning devices but all use relatively large specimens such as human crania. Here we perform shape analyses on Australia’s smallest rodent species to test whether a 3D surface scanner produces similar results to a μCT scanner.
Methods. We captured 19 delicate mouse crania with a μCT scanner and a 3D surface scanner for geometric morphometrics. We ran multiple Procrustes ANOVAs to understand how variation due to scan device compared to other sources of variation such as biologically relevant sources and operator error. We quantified operator error with morphological disparity and repeatability. Finally, we tested whether the different scan datasets could detect intra-specific variation using cross-validation classification. Shape patterns were visualized with Principal Component Analysis (PCA) plots.
Results. In all Procrustes ANOVAs, regardless of factors included, differences between individuals contributed the most to total variation. This is also reflected in the way individuals disperse on the PCA plots. Including only the symmetric component of shape increased the biological signal relative to variation due to device and due to error. 3D scans create a higher level of operator error as evidenced by a greater spread of their replicates on the PCA, a higher morphological disparity, and a lower repeatability score. However, in the test for small intra-specific differences, the 3D scan and μCT scan datasets performed identically.
Discussion. Compared to μCT scans, we find that even very low resolution 3D scans of very small specimens are sufficiently accurate to capture variation at the level of interspecific differences. We also make three recommendations for best use of low resolution data. First, we recommend analyzing the symmetric component of shape to decrease signal from operator error. Second, using 3D scans generates more random error due to increased landmarking difficulty, therefore be conservative in landmark choice and avoid multiple operators. Third, using 3D scans introduces a source of systematic error relative to μCT scans, therefore do not combine them when possible and especially in studies with little variation. Our findings support increased use of low resolution 3D images for most morphological studies; they are likely applicable to low resolution scans of large specimens made in a medical CT scanner, for example. As most vertebrates are relatively small, we anticipate our results to bolster more researchers designing affordable large scale studies on small specimens with 3D surface scanners.
A neural model of 3D shape-from-texture: Multiple-scale filtering, cooperative-competitive grouping, and 3D surface filling-in