Localizing target structures in ultrasound video - a phantom study

Abstract

The problem of localizing specific anatomic structures using ultrasound (US) video is considered. This involves automatically determining when an US probe is acquiring images of a previously defined object of interest, during the course of an US examination. Localization using US is motivated by the increased availability of portable, low-cost US probes, which inspire applications where inexperienced personnel and even first-time users acquire US data that is then sent to experts for further assessment. This process is of particular interest for routine examinations in underserved populations as well as for patient triage after natural disasters and large-scale accidents, where experts may be in short supply. The proposed localization approach is motivated by research in the area of dynamic texture analysis and leverages several recent advances in the field of activity recognition. For evaluation, we introduce an annotated and publicly available database of US video, acquired on three phantoms. Several experiments reveal the challenges of applying video analysis approaches to US images and demonstrate that good localization performance is possible with the proposed solution.

Keywords: Dynamic textures; Ultrasound imaging; Video analysis.

Figure 1

Localization of anatomical structures (e.g., structures A and B) by moving the Ultrasound probe along a path (dark red) on the human body.

Figure 2

Illustration of the acquisition fan — on a schematic liver model (dark red) — when moving the US probe according to two common probe movements: translation and tilting.

Figure 3

Generative model for DTs and KDTs.

Figure 4

Clinical data from one patient, acquired during a radio-frequency ablation of a liver tumor. The probe sweeps from left to right and back (see arrows) on the liver surface. The top row shows a video of the probe movement; the middle row shows a subset of the actual US frames with a blood vessel moving in and out of the imaging plane (highlighted); the bottom row shows the same set of frames, synthesized from a KDT model with 10 states (using a RBF kernel) that was estimated from the original US sequence.

Figure 5

Composition of the observation matrix Y — for the first sliding window — either based on (a) raw pixel-intensities or (b) mid-level feature histograms computed from sub-sequences within (in this example, the subsequences do not overlap).

Figure 6

Schematic illustration of the US phantom (left); three images of one of our test phantoms at different viewing angles (right).

Figure 7

Acquisition of the search path and template sequences (left). Definition of the earliest and latest localization times t₊, t₋ (right).

Figure 8

Annotation example for a search path; bounding boxes around the structure of interest, i.e., a loop, (positioned using VATIC [Vondrick et al., 2012]) are shown in yellow. Frame numbers 17 and 37 correspond to positions t₊ and t₋ (best viewed in color).

Figure 9

Definition of false positive/negatives and true positives/negatives in the context of the localization problem, illustrated on a toy example with three target structures. The similarity curve shows the similarity measure $d_t^p$ (cf. Sect. 4.4) for a search path containing structure B.

Figure 10

Precision/Recall (PR) curves for the best configuration of BoWDyn (8 states, 400 codewords) and IntDyn (5 states), on all nine search paths.