Efficient interactive 3D Livewire segmentation of complex objects with arbitrary topology

Abstract

We present a novel interactive method based on a 3D Livewire approach for segmenting complex objects of arbitrary topologies. Our proposed technique automatically and seamlessly handles objects with branchings, concavities, protrusions, and non-spherical topologies with minimal user-input. Given sparse interactively segmented contours on orthogonal slices, our proposed method determines Livewire seedpoints on all slices in the third orthogonal direction, which are used to mimic user-guided segmentation. In doing so, our method pre-processes these points to increase algorithm robustness, and uses a novel seedpoint sorting method using ideas from L-system's Turtle algorithm. Moreover, we present a segmentation tool based on our proposed framework and demonstrate the robustness of our approach on real medical data. Results highlight the superior performance of the proposed method with validation tests on synthetic and real MRI and CT data, with segmentation reproducibility exceeding 95% and segmentation task time decreasing to less than 20% when compared to performing 2D Livewire on each volume slice.

Spotlight: Automated Confidence-Based User Guidance for Increasing Efficiency in Interactive 3D Image Segmentation

“…Users specify seeds along the boundary of the object, enforcing the segmentation to go through it. The reader is referred to the T-LW paper [12] for a detailed explanation of the algorithm. In this work, the T-LW user interaction mechanism remains the same but with the addition of a much needed "suggest plane" feature, which invokes Spotlight to locate a 2D LW contouring plane chosen optimally in regions of maximal segmentation ambiguity, which in turn improves accuracy and speeds up the interactive segmentation process.…”

“…We build Spotlight on top of an enhanced version of the T-LW algorithm [12]. Although other base methods such as graph cuts and random walker exist and may be adopted (see the Conclusions section), we chose T-LW because it allows for the more natural interaction experience of contouring within 2D slice planes.…”

“…In order to support oblique contouring slices, we develop a continuous analytical representation of the turtle map instead of the discrete binary bitmap of [12] that is restricted to the pixels in the standard three orthogonal views. We achieve this by taking the line segments produced by intersecting each user contour with P m (the red line segment in Fig.…”

“…The T-LW algorithm [12] automatically generates Livewire contours for an unvisited plane P m based on the existing sparse set of interactivelyseeded 2D contours. A plane P m will contain a turtle map, which is a set of line segments resulting from intersecting P m with the user-generated contours in other planes not parallel to P m (Fig.…”

“…Armstrong et al [1] suggested a 3D analog of a 2D Livewire, namely a "Live Surface", which can be localized in real-time through an optimized graph cuts algorithm. The more recent turtle map 3D Livewire algorithm [12], or T-LW for short, accepts a sparse set of user-specified planar contours as input, which it automatically combines to seed unvisited planes and subsequently generate a new dense set of 3D contours.…”