Supplementary Figure 3: Automated Segmentation.
The goal of the automated segmentation procedure is to group pixels such that each segment within the SLM ‘ON’ area belongs to a single neuronal compartment (such as a spine), all segments are of a minimum integrated brightness, and the total number of compartments within the SLM ‘ON’ area is less than 1000. These constraints can only be approximately satisfied. To do this, the reference volume (left, gamma=0.5; small region shown for clarity) is classified into 4 classes (center) using a hand trained classifier in Ilastik’s autocontext mode. The classes are correspond to dark (background) voxels, spine heads, voxels with fluorescent label, and dendritic shafts. The voxel classes and image intensity are passed to a skeletonization-based algorithm that groups non-dark pixels into segments. The automated segmentation output (right) occasionally produces flaws such as fused or split spines, which have not yet been quantified. We have written an interface (segEdit in the software package) for manual curation of these errors, which was used for most datasets. This example automated segmentation is typical of all yGluSnFR fields of view imaged.
