Randomly selected images of astrocytes arranged by increasing hypertrophy score from left to right. The cell outline is shown in yellow, and the soma outline in red. GFAP is shown in red and DAPI in blue. Scale bar 10µm.
Example of how the model arrives at a particular classification decision for the cell shown in inset. The value of each feature is shown on the left (z-normalized, gray) and the effect on the decision is shown on the right (larger values tend toward hypertrophy). In this example, the large number of branching points in the skeleton contributes the most towards the decision to classify as hypertrophic.
To obtain a finer quantification of the astrocytic morphology, our model also outputs a hypertrophy score that is linearly dependent on the morphological features, such as the cell size. Since many morphological features have been found to be linearly correlated with functional markers, we reasoned that this score would provide a sensitive metric of the astrocytic phenotype.
For each cell, we can explain how the model arrives at a particular decision on its classification and hypertrophy score. For example, for the cell shown at the bottom of the figure, although its processes were thin and its soma not larger than average, it has a high number of branching points in its processes and a low fraction of the area in the soma. These features result in a high hypertrophy score and the cell is thus classified as having a hypertrophic morphology.