Pressure Based Segmentation in Volumetric Images

Abstract

Abstract. Analysing Roman coins found in archaeology sites has been traditionally done manually by an operator using volumetric image slices provided by a computed tomography scanner. In order to automate the counting process, a good segmentation for the coins has to be achieved to separate the touching surfaces of the coins. Separating touching surfaces in volumetric images has not yet attracted much attention. In this paper we propose a new method based on using a form of pressure to separate the intersecting surfaces. We analogise the background of the image to be filled with an ideal gas. The pressure at a point has an inverse relationship with the volume of homogeneous material surrounding it. By studying the pressure space, the locations of intersecting surfaces are highlighted and encouraging segmentation results are achieved. Our analysis concerns a selection of images, naturally demonstrating success, together with an analysis of the new technique's sensitivity to noise. IntroductionComputed tomography imaging is an increasingly popular source for information about three dimensional objects, with many applications ranging from medical to industrial. Scans can contain multiple objects with the same density or single objects containing smaller ones with similar density. The placement of the objects in the 3D space can be random and in some cases the surfaces of those objects touch which makes it difficult to separate them using conventional thresholding and segmentation techniques, motivating development of a higher level process. An example of such a problem is the CT scanned jar ( Fig. 1) which contains a set of Roman coins. This set of data contains coins with similar density randomly placed with different orientations and locations within the jar. The problem associated with this particular set of data relies in the high attenuation factor for the material from which the coins are made which in turn increases the chance of touching surfaces in the volumetric image especially for the coins in the centre of the jar. Separating objects with the same density and texture is challenging due to the absence of techniques for detecting in 3D the regions of intersection between the objects, impeding the possibility of counting the coins.Pressure Based Segmentation in Volumetric Images 239 Fig. 1. CT image of Roman coins inside a jarMany approaches have been developed to solve the problem of separating touching objects in two dimensional (2D) space. The two main application concern separating rice grains [1] and counting cells in microscope images [2]. A traditional approach involves thresholding, corner detection and joining points of interest to create a binary image of disconnected objects. On the other hand, there is no such technique for 3D image analysis. The literature provides some model-based methods that have been used to separate left and right lungs [3]. The search for regions of interest uses images where lungs intersect, to minimize the computational demands. Edge or surface detection can be appli...

How many specimens make a sufficient training set for automated 3D feature extraction?

“…Our results suggest extracting the internal structure of planktonic foraminifera poses a greater challenge compared to extraction of the external structure, evidenced in the measures of percentage calcite and the shape of the internal structure (see figures 5b, 6). This discrepancy in difficulty can be likely be attributed to material homogeneity, which is reflected in low contrast differences within CT scans [37, 38]. This challenge is particularly prominent in the case of planktonic foraminifera, where the internal structure contains sedimentary infill and nannofossil ooze with densities similar to external calcite [39, 40].…”

Challenges and opportunities in applying AI to evolutionary morphology

“…For example, when specimens being scanned are very dense, scans may not have a consistent perceived density (e.g. Alathari, 2015;Furat et al, 2019). Objects of similar densities may not be displayed at the same greyscale value through the scan, though the structural properties of the material will be evident.…”