On the Fly Segmentation of Intravascular Ultrasound Images Powered by Learning of Backscattering Physics

Abstract

Intravascular ultrasound (IVUS) is commonly used as an adjunct to the imaging-based diagnosis of vascular plaques that have been primarily imaged with computed tomography angiography (CTA) or magnetic resonance angiography (MRA). Since speckle intensity in IVUS images are inherently stochastic in nature, clinicians face the critical challenge of identifying vessel boundaries. In this Introduction, we present a method for segmenting the lumen and external elastic laminae in IVUS images of coronary arteries via a two-stage framework. The first stage employs an understanding of the physics of the interaction between ultrasound and tissue as a statistical mechanical parametric framework, and ensemble learning of this parametric space to initialize lumen and external elastic luminae boundary contours. Subsequently, in the second stage, these initialized contours are solved using a random walker solution. While both stages are employed when segmenting individual IVUS frames, in the event of complete pullbacks, the two stages are employed on the first frame but only the second stage is employed on subsequent frames, using initialization propagated from the previously segmented frame as the basis for investigation. We have experimentally evaluated the approach using 77 IVUS frames acquired at 40 MHz and 10 IVUS pullbacks acquired at 20 MHz. Our approach obtains a Jaccard score of $0.89 \pm 0.14$ and $0.87 \pm 0.12$ for lumen and $0.88 \pm 0.09$ and $0.91 \pm 0.10$ for external elastic laminae segmentation on 40 and 20 MHz IVUS data, respectively, over a 10-fold cross-validation experiment. Our computationally lightweight implementation enables on-the-fly segmentation of IVUS frames of 40 MHz at 512 × 512 pixels and 20 MHz at 384 × 384 pixels on a PC-based system without any special accelerators within 1.2 ± 0.2s and 1.1 ± 0.2s, respectively.