Overview of our method to quantify microglial activation based on morphology (left panel), and an example showing activated (white) and non-activated (orange) microglia (right panel).
Shapley (SHAP) values describing how much each morphological parameter influences the decision of the machine-learning classifier to categorize a given microglial cell as activated or non-activated.
Representative examples of individual microglia with increasing activation score from left to right. The soma of each cell is highlighted in red and the processes are highlighted in yellow. The cells show considerably different morphological features as they become more activated. Scale bars are 10µm.
Activated microglia have characteristic morphological features that distinguish them from resting microglia (e.g. shorter and denser processes, ameboid-like shape, etc.). We have used supervised learning to develop a machine-learning (ML) computer vision model that classifies the activation state of each microglial cell based on these morphological characteristics.
To classify the activation state of microglia, we first segment the cell body using the Iba-1 channel, and we use the colocalized DAPI staining to delineate the soma and to separate individual microglia. Each microglial cell is skeletonized and its processes are separated from the soma to extract a high number of morphological features, such as the cell size, the soma size, the number of branching points in the skeleton, etc. These features are then used by a machine-learning model that assigns an activation score to each cell, and then classifies that cell as activated or non-activated.
The Shapley values represent the additive effects of the different features, and can be used to explain how the machine learning model classifies a cell. Each point in the graph represents the value of the parameter of a given cell indicated by color, with red corresponding to values higher than average, and its impact on the decision, where lower Shapley values tend towards classifying cells as activated. For example, an ameboid-like microglial cell (not elongated) with lots of processes (low fraction of area occupied by soma) and high solidity will be classified as activated.
The machine-learning model outputs an activation score between 0 and 1 for each microglia, where 1 represents the most activated state. It is possible to binarize the prediction by thresholding the activation at 0.5. We can also use the continuous activation score directly to highlight fine differences in morphology that may not be captured with a simple binary prediction.