Page - 99 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Active contour models for individual keratin filament tracking Dmytro Kotsur1, Rudolf E. Leube2, Reinhard Windoffer2 and Julian Mattes3 Abstract—As a major component of the cytoskeleton, keratin filaments form a branched network, which plays a significant role in the mechanical response, motion and dynamics of the cell. They undergo a complex dynamic lifecycle, which we aim to investigate by tracking individual filaments. In this paper we introduce an active contour-based tracking algorithm to analyze the motion of individual keratin filaments in sequences of confocal images. The algorithm combines parametric active contours (snakes) with Lukas-Kanade’s algorithm for optical flow calculation. We define an image preprocessing workflow to compute robustly the external energy of the snake and we impose an additional structural constraint for controlling the length of the contour. I. INTRODUCTION The cytoskeleton plays a main role in cellular motility and dynamics, which in turn is of high relevance for vital and also for pathological processes, such as wound healing and tumor metastasis [5]. As a major component of the cytoskeleton, keratin filaments form a branched network and are essential for the mechanical response to external forces. Biophysical investigation and analysis of different types of keratin filaments requires their localization and the extraction of their motion in the time-sequences of consecutive confocal images. As it was shown previously [7], [3], [4], this problem can be successfully approached for separated individual actin filaments. However, applying this approach to tracking of keratin filaments within a branched network may lead to additional complications and errors, as for example, uncontrolled growth of the snake. In this paper we introduce a tracking algorithm based on stretching open active contours [3] to analyze the global motion features of individual keratin filaments within their network. We define an image preprocessing workflow to calculate robustly the “external energy” of the snake and impose an additional structural constraint for controlling the length of the contour. II. TRACKING ALGORITHM In this section, we first define our active contour model as a minimization problem. Then, we introduce an “external energy” based on the image and impose a contour length constraint to control snake growth. Finally, we combine all steps together and present an overall tracking procedure. 1Dmytro Kotsur is with Software Competence Center Hagenberg GmbH (SCCH), Austria,Dmytro.Kotsur@scch.at 2Rudolf E. Leube and Reinhard Windoffer are with MOCA, Institute of Molecular and Cellular Anatomy, RWTH Aachen University, Aachen, Germany, {rleube, rwindoffer}@ukaachen.de 3Julian Mattes is with MATTES Medical Imaging GmbH, Hagenberg, Austria,Julian.Mattes@mattesmedical.at A. Parametric snakes: active contour models We define a filament as a parametric curve x(s) = [x(s),y(s)],s∈ [0,1]. According to [2], the position of the filament within a frame in a time-sequence is obtained by minimizing the following so-called “energy” functional: E= ∫ 1 0 1 2 ( α|x′(s)|2+β|x′′(s)|2)+Eext(x(s))ds (1) whereα and β are parameters which control the stretching and bending resistance of the curve, correspondingly. This problem is solved by reducing (1) to a differential equation and applying an iterative scheme with an artificial time variable t: xt(s,t)=αxss(s,t)+βxssss(s,t)−∇Eext(x(s,t)) (2) The impact of the “external energy” Eext or the gradient of “external energy”∇Eext is crucial in this problem, because the convergence of a snake considerably depends on this term. B. External energy and structural constraints In Xu et al. [6] the gradient of the “external energy”∇Eext is replacedby thevectorfieldv(x,y)=[u(x,y),v(x,y)],which minimizes the functional: E= ∫ ∫ µ ( u2x+u 2 y+v 2 x+v 2 y ) +|∇f|2|v−∇f|2dxdy (3) where f(x,y) is the intensity of the pixel at the position (x,y), |•| is the Euclidean norm and µ is the regularization (smoothness) parameter. The vector field v(x,y) is called gradient vector flow (GVF). In this case the evolution of the snake on a single frame is defined as follows: xt(s,t)=αxss(s,t)+βxssss(s,t)−v(x(s,t)) (4) It is shown in [6] that GVF has a larger capture range, compared to the vector field given by∇Eext defined in [2]. It also improves the snake convergence in case of high concavities. However, the intensity variation along a filament may be high, which leads to additional errors during snake convergence. Therefore, we preprocess images applying the following pipeline of filters: Gaussian smoothing; Hessian ridge enhancement; gamma contrast correction. The drawback of the snake algorithm itself as defined in [2] is that the open-ended contour (Fig. 1C) tends to shrink over time (Fig. 1D). To overcome this, we use a stretching term for open ends as defined in [7]. However, it may lead to overgrowth of the contour (Fig. 1E). We it this by processing endpoints separately. We define an additional distance-based “energy” potential for the branching and end points of the 99