Tubular Enhanced Geodesic Active Contours for Continuum Robot Detection using 3D Ultrasound

Abstract

Three dimensional ultrasound is a promising imaging modality for minimally invasive robotic surgery. As the robots are typically metallic, they interact strongly with the sound waves in ways that are not modeled by the ultrasound system's signal processing algorithms. Consequently, they produce substantial imaging artifacts that can make image guidance difficult, even for experienced surgeons. This paper introduces a new approach for detecting curved continuum robots in 3D ultrasound images. The proposed approach combines geodesic active contours with a speed function that is based on enhancing the "tubularity" of the continuum robot. In particular, it takes advantage of the known robot diameter along its length. It also takes advantage of the fact that the robot surface facing the ultrasound probe provides the most accurate image. This method, termed Tubular Enhanced Geodesic Active Contours (TEGAC), is demonstrated through ex vivo intracardiac experiments to offer superior performance compared to conventional active contours.

Fig. 1

Continuum robot inserted inside a porcine heart. (a) Experiment set up. (b) 3D ultrasound image showing both the left atrium and the curved robot.

Fig. 2

Block diagram of the Tubular Enhanced Geodesic Active Contour (TEGAC) algorithm

Fig. 3

Original image with artifacts: CTA, Comet Tail Artifacts; DSL, diffractive side lobe; and RAA, Range ambiguity artifacts. The images show 2D Maximum Intensity Projection (MIP) views for 3DUS volumetric images. (a) MIP Top View, b) MIP Front View, (c) MIP Side View.

Fig. 4

Tubular enhancement filtering results in 2D Maximum Intensity Projection (MIP) views for 3DUS volumetric images. (a) MIP Top View, (b) MIP Front View, (c) MIP Side View.

Fig. 5

Raw robot images and overlaid TEGAC results. The 2D slices correspond to front, top and side views. The crosshair locations indicate tubular object. (a) 2D slices of volumetric image. (b) 2D slices of volumetric image overlaid with the evolved contour from TEGAC.

Fig. 6

Comparison between the conventional GAC and TEGAC algorithms. The 2D slices correspond to front, top and side views. The crosshair locations indicate tubular object. (a) Speed image overlaid with resulting active contour from conventional GAC; (b) Speed image overlaid with resulting active contour from TEGAC algorithm.

Fig. 7

3D rendering of detected robot using conventional GAC and TEGAC algorithms. (a) Conventional GAC. (b) TEGAC algorithm.

Fig. 8

DICE metric comparing manual segmentation, the TEGAC detection algorithm and the conventional Geodesic Active Contour (GAC) algorithm.