Fig 3 - uploaded by Nikolaus Obholzer
Content may be subject to copyright.
(a) Multithreading with non separable output leads to higher memory consumption and additional processing to merge the temporary outputs from each thread into the final output. (b) Streaming allow to decrease the memory consumption by processing sub regions of the data at a time. (c) streaming and multithreading allow to process all data in parallel with no memory overhead compared to single threaded processing, but supposes that both input and output can be divided into independent sub-regions.
Source publication
We present a high performance variant of the popular geodesic active contours which are used for splitting cell clusters in microscopy images. Previously, we implemented a linear pipelined version that incorporates as many cues as possible into developing a suitable level-set speed function so that an evolving contour exactly segments a cell/nuclei...
Citations
We are developing an approach called in toto imaging whose goal is to track all the cell movements and divisions that give rise to an embryo. We can also capture protein expression and localization throughout development using GFP transgenics. Our long term goal is to integrate these data into a “Digital Fish” that shows how the genetic circuits encoded in the genome turn an egg into an embryo. We have a two pronged approach. We use confocal and 2-photon, time-lapse microscopy to capture very high spatial and temporal resolution movies of developing zebrafish embryos which permit single cell tracking but for only a portion of an embryo. We also use a robotic, 96-well plate based system that can screen 5000 embryos per day but at much lower resolution. Both approaches generate ~100,000 images per experiment. We are developing software systems for analyzing these large image sets.
In confocal microscopy, target objects are labeled with fluorescent
markers in the living specimen, and usually appear with irregular
brightness in the observed images. Also, due to the existence of
out-of-focus objects in the image, the segmentation of 3-D objects in
the stack of image slices captured at different depth levels of the
specimen is still heavily relied on manual analysis. In this paper, a
novel Bayesian model is proposed for segmenting 3-D synaptic objects
from given image stack. In order to solve the irregular brightness and
out-offocus problems, the segmentation model employs a likelihood using
the luminance-invariant 'wavelet features' of image objects in the
dual-tree complex wavelet domain as well as a likelihood based on the
vertical intensity profile of the image stack in 3-D. Furthermore, a
smoothness 'frame' prior based on the a priori knowledge of the
connections of the synapses is introduced to the model for enhancing the
connectivity of the synapses. As a result, our model can successfully
segment the in-focus target synaptic object from a 3D image stack with
irregular brightness.
Accurate segmentation of cracked body from three-dimensional (3D) industrial Computed Tomography (CT) images is an important step in the process of crack measurement and automatic recognition. In this paper we present a fast method for the segmentation of cracked body. The improved algorithm incorporates wavelet transform and Chan and Vese (C-V) model as key components. The 3D wavelet transform is applied for detecting rough edges. Then region growing is used to find a suitable region which contains cracked body. Based on the resulting volume data, 3D C-V model is used to capture the edges of cracked body. The improved method can locate rough regions by using wavelet modulus maxima, which not only reduces the amount of data C-V model processed, but also provides initial contour surface that can accelerate the convergence speed of C-V model. Experimental results illustrate our method can accurately detect the cracked surface, as well as give computational savings of segmentation which satisfy the demand of defects detection of industrial CT.
