Automated segmentation of complex patterns in biological tissues: Lessons from stingray tessellated cartilage

Abstract

IntroductionMany biological structures show recurring tiling patterns on one structural level or the other. Current image acquisition techniques are able to resolve those tiling patterns to allow quantitative analyses. The resulting image data, however, may contain an enormous number of elements. This renders manual image analysis infeasible, in particular when statistical analysis is to be conducted, requiring a larger number of image data to be analyzed. As a consequence, the analysis process needs to be automated to a large degree. In this paper, we describe a multi-step image segmentation pipeline for the automated segmentation of the calcified cartilage into individual tesserae from computed tomography images of skeletal elements of stingrays.MethodsBesides applying state-of-the-art algorithms like anisotropic diffusion smoothing, local thresholding for foreground segmentation, distance map calculation, and hierarchical watershed, we exploit a graph-based representation for fast correction of the segmentation. In addition, we propose a new distance map that is computed only in the plane that locally best approximates the calcified cartilage. This distance map drastically improves the separation of individual tesserae. We apply our segmentation pipeline to hyomandibulae from three individuals of the round stingray (Urobatis halleri), varying both in age and size.ResultsEach of the hyomandibula datasets contains approximately 3000 tesserae. To evaluate the quality of the automated segmentation, four expert users manually generated ground truth segmentations of small parts of one hyomandibula. These ground truth segmentations allowed us to compare the segmentation quality w.r.t. individual tesserae. Additionally, to investigate the segmentation quality of whole skeletal elements, landmarks were manually placed on all tesserae and their positions were then compared to the segmented tesserae. With the proposed segmentation pipeline, we sped up the processing of a single skeletal element from days or weeks to a few hours.

High-Throughput Segmentation of Tiled Biological Structures using Random-Walk Distance Transforms

“…Our proposed workflow—incorporating a series of common image analysis tools—is capable of segmenting quite different types of structurally-complex, biological datasets with a speed and quality that allows statistical analysis of huge numbers of data points. Our three case studies demonstrate the efficacy of the workflow, while also highlighting several particular features for consideration: – Our segmentation of stingray tesserae produced results quite similar to those obtained in a previous work (Knötel et al. 2017), where a specific distance transform was deemed necessary to account for particular aspects of tissue morphology.…”

“…Hence, in this case, the Euclidean distance transform is unsuitable for the purpose of separating tiled objects. For this reason, Knötel et al. (2017) developed a specialized distance map which measures the shortest distance only in the plane that approximates the local orientation of the object at each voxel (e.g., the plane of the surface in a single layer tiling).…”

“…Tesserae are typically hundreds of micrometers wide and thick and cover the surface of nearly the entire skeleton (excepting the centra of the vertebrae), meaning that even a small skeletal element such as the stingray hyomandibula in Fig. 3B (∼1.5 cm long) is armored by several thousand tesserae (Knötel et al. 2017).…”

“…For the current work, we scanned a single hyomandibula from a Haller’s stingray ( Urobatis halleri , sub-adult male, 11 cm disc width) with a Skyscan 1172 desktop µCT scanner (Bruker μCT, Kontich, Belgium), with scanning parameters as described in previous studies (Seidel et al. 2016; Knötel et al. 2017).…”

“…No post-processing was done for this dataset. The whole segmentation process took <2 h as opposed to several days that would be needed when doing the processing manually (Knötel et al. 2017).…”