Pixel classification

As pointed out by Trier [160], locally adaptive threshold methods are more robust than global ones. In our case, this is particularly important since tissue processing and dye preparation often lead to inhomogeneous staining. This is observed as a smooth variation of the average luminance over the image. Thus, the threshold level for a pixel is chosen on the basis of the histogram of a subimage around it, typically a square of $25 \times 25 \mu m$ that contains a couple of axons. To maintain a low computational cost, the histogram analysis is only performed for a few locations of the window, and the threshold levels are bi-linearly interpolated in between.
The subimages contain three types of tissue: the myelin, the endoneurium and the axons. This corresponds to 2 or 3 lobes in the histogram, depending on whether the endoneurium presents any degree of coloration differentiating it from the axons. We use a simple heuristic to select a threshold level between the first two lobes. The 15^th and 50^th percentile in the histogram of the grey levels are considered typical values of the first and second lobes. The mean of these two values is taken as the threshold level.