Multi-covariate Imaging of Sub-resolution Targets

Abstract

Conventional B-mode ultrasound imaging assumes that targets consist of collections of point scatterers. Diffraction, however, presents a fundamental limit on a scanner's ability to resolve individual scatterers in most clinical imaging environments. Well-known optics and ultrasound literature has characterized these diffuse scattering targets as spatially incoherent and statistically stationary. In this paper, we apply a piecewise-stationary statistical model to diffuse scattering targets, in which the covariance of backscattered echoes can be described as the linear superposition of constituent components corresponding to echoes from distinct spatial regions in the field. Using this framework, we present Multi-covariate Imaging of Sub-resolution Targets (MIST), a novel estimation-based method to image the statistical properties of diffuse scattering targets, based on a decomposition of aperture domain spatial covariance. The mathematical foundations of the estimator are analytically derived, and MIST is evaluated in phantom, simulation, and in vivo studies, demonstrating consistent improvements in contrast-to-noise ratio and speckle statistics across imaging targets, without an apparent loss in resolution.

Fig. 1:

Estimator framework using three constituent models. The predicted intensity distribution for a narrowband array is segmented into distinct mainlobe and sidelobe spatial regions. A third covariance model corresponding to incoherent noise is represented by an identity matrix. The mainlobe, sidelobe, and incoherent noise covariance matrices are plotted at bottom, and a 1-D plot showing the average along the diagonal of each matrix is shown at right, with the superposition of the three components plotted in the dashed line.

Fig. 2:

MIST estimator images of simulated targets. A conventional B-mode image is compared to the main lobe estimate and side lobe estimate of a (a) point and (b) anechoic cyst target. 1-D lateral profiles at 50 mm depth are shown at right, where the y-axis has been split to display the full range of the lateral profiles.

Fig. 3:

B-Mode is compared to the 2-parameter (mainlobe + sidelobe) and 3-parameter (mainlobe + sidelobe + incoherent noise) MIST model across a range of channel signal-to-noise ratios for a 1 cm anechoic target. (a) Sample images at selected SNRs. (b) Contrast, CNR, and Speckle SNR are plotted against channel SNR.

Fig. 4:

(a) Comparison of B-Mode and MIST images across targets of varying native contrast at a channel SNR of 0 dB. (b) Plots of measured contrast vs. channel SNR demonstrate the stability of MIST across SNR. Across native contrast targets, MIST attains the true target contrast at a lower SNR than B-Mode.

Fig. 5:

Edge resolution simulations. (a) Sample B-Mode and MIST images of the edge phantom with a speckle generating and an anechoic region (−∞ dB contrast). (b) Averaged lateral profiles of B-Mode and MIST images for each set of phantom simulations. The native contrast of the imaging phantom is indicated by the axis on the right.

Fig. 6:

Approximately matched B-Mode and MIST images using a single sample (0λ) and axial averaging of the covariance matrix over a wavelength (1λ). (a) Simulation and (b) phantom images are shown with contrast and CNR plotted to compare performance for each lesion.

Fig. 7:

In vivo liver images comparing B-Mode, receive spatial compounding (3 apertures, 50% overlap), SLSC (Q=5%, 1λ kernel), and MIST (1λ). B-Mode, receive spatial compounding, and MIST images are shown on a 70 dB dynamic range. SLSC is shown on a [0-1] scale.

Fig. 8:

In vivo fetal images comparing B-Mode, receive spatial compounding (3 apertures, 50% overlap), SLSC (Q=5%, 1λ kernel), and MIST (1λ). B-Mode, receive spatial compounding, and MIST images are shown on a 60 dB dynamic range. SLSC is shown on a [0-1] scale.