Almutairi, Cootes, Kadler: Tracking Collagen Fibres

Serial block-face scanning electron microscopy (SBF-SEM) has become an important modality for examining small structures in biology, and new instruments are capable of gathering large volumes of data. Analysing these datasets manually is difficult because of the number of individual images involved. Objects of interest often span many slices, and must be tracked by the user through these slices to understand their full 3D structure. In this paper we focus on analysing the shape and structure of collagen fibres. There can be over 10,000 fibres in a single image, and ideally each would be tracked through the volume in order to estimate their lengths and how they form into bundles. Manual annotation of such data has proved impractical. Here we describe an automated system which detects the fibres and tracks them through large image volumes. We show that multi-scale normalised cross correlation is effective for finding candidates in a single image, and that false positive matches can be eliminated using a random forest classifier. Tracking is performed by linking candidates from one image to the next. We describe the system in detail, including experiments assessing the accuracy and reliability of the approach when tracking through hundreds of image slices in a volume.