Fig. 6: PXCT density estimation schematic.

A A 2D cross-section of the PXCT reconstruction volume of a blue scale of J. orithya male; the boxed region is selected for the demonstration. Each voxel’s reconstructed phase density is shown in grayscale (darker means denser). Red arrows in each panel indicate the next key transformation step. In this panel, the red arrow shows the rotation direction of the volume (B). The volume is rotated by the global mean tilt so that the lower lamina of the entire scale (which extends beyond this field of view) is, as much as possible, aligned with the x-y plane. Consequently, the trabeculae are conveniently aligned along the vertical z-axis for column-wise analysis. Red arrows indicate that the volume will be laterally partitioned into multiple 1D (z-axis) voxel columns next. C Using peak-finding, we separately detect the position of the lower lamina in each 1D column (indicated by cyan markers). Red arrows indicate that these detected peaks will be computationally aligned in the next step, since we cannot remove the local undulations in the lower lamina of the entire scale with a single rotation. D Local undulations in the lower lamina are flattened via peak alignment. Red arrows indicate that the lower lamina will be segmented from the upper lamina features within each 1D column in the next step. E The upper lamina feature vectors are reduced and clustered into four categories—window, ridge, trabeculae, and cross-rib—using PCA followed by K-means clustering. F The 1D profiles of the lower lamina and the upper lamina features of each column are then fitted to Gaussian functions to reduce partial volume effects and estimate thickness and phase density. This fitting is separately done on every 1D column throughout the volume. Shown here are examples from one column per cluster (ridge, trabeculae, and cross-rib).