Nerve morphometry

Morphometric studies of nerves or fiber tracts involve information about alterations in nerve bundle size, number or size of the axons. They have been shown to be of great value in detecting developmental or pathological abnormalities [70,25,116,78,40]. They have also been broadly used in experimental nerve research [120,18].
Most techniques used for estimating nerve and fiber parameters are based on highly time consuming manual measures. For instance, in the study below, the sciatic nerve contains approximately 15000 myelinated fibers, included in 85 images of $1850 \times 1234$ pixels, or 500 MBytes of raw data. Therefore, the information is usually sampled by selecting only a few of the images in the nerve cross-section. Unfortunately, the large variability in fiber distribution according to function, i.e. sensitive or motor nerves, or to the specie specificity, precludes the selection of a sampling pattern that is reasonably representative of the nerve. Torch [158] estimates that sampling schemes involving less than 50% of the images provide an unreliable measure of myelinated fiber distribution.
Alternatively, an automatic image analysis tool solves the problem by allowing to examine all the available material. Algorithms are usually divided in two steps. First, the image is analyzed with a local operator that classifies pixels between the various tissues types. For this stage, Jain [79], Garbay [62] or Thiran [153] rely on thresholding, sometimes preceded by filtering. Secondly, the image is analyzed at the structural level using a variety of tools such as region growing segmentation [79], grouping of edge elements [62], or mathematical morphology [153]. Unfortunately, none of these methods can handle multi-part objects such as axons surrounded by the myelin sheath.
Alternatively, Amini [2], Fok [55] or Elmoatoz [43] rely on active contour models, or snakes, to handle both local and structural analysis in one step. After detecting candidates through a global tool such as the Hough transform, each region of interest is processed individually with an explicit active contour model evolving towards the real contours of the cell. Unfortunately, such methods tend to be too computationally expensive for the large data sets required by a full study. Also, it is unclear whether any of these models could handle the large size variability encountered in nerve fibers.