Signal & Image Processing

Aberration Correction

Every conventional beamformer assumes sound crosses the body at one fixed speed. A fat layer near the probe breaks that assumption, and the error compounds with depth. This project replaces the geometric delay with one marched through an actual speed-of-sound map.

1540m/s
The one speed
conventional DAS assumes
10%
Speed-of-sound variation
across human tissues
~11cm
Imaging depth validated,
twice the prior art
0 → 6
Point targets localized
in the steepest fat case

The aberration bench

The four simulated media from the paper. Add a fat layer, tilt it, and watch where conventional delay-and-sum thinks the targets are. Then switch the beamformer.

transducer array · 39.4 mm fat layer · ~1400 m/s background · 1540 m/s depth
Fat layer at the superficial depth, average 1400 m/s
True positions of the three cysts and ten point targets
Apparent positions under the selected beamformer
Assumed path vs refracted path of one transmit ray
Mean GDS
0.7
Targets localized
7of 10
gCNR · cyst CY1
0.79

Geometry after Figs. 1 and 3 of the paper; displacements are drawn to illustrate the failure mode, not to scale. The three readouts are the measured values from Table I for the selected medium and beamformer.

Every scanner assumes 1540 metres per second

Delay-and-sum (DAS) receive beamforming reconstructs each pixel by delaying echoes for the round trip between transducer and tissue, weighting them, and summing across the array. Estimating that round trip requires the speed of sound in the medium, and conventional DAS settles the question by fiat: a spatially constant 1540 m/s everywhere, regardless of what the sound actually crossed.

Human tissue is heterogeneous, with speed-of-sound variations of up to 10%. The mismatch produces delay errors that accumulate with depth, degrading resolution, contrast, and the geometric accuracy of the image. A superficial fat layer, common in liver and abdominal imaging, is the classic offender: sound crosses it at roughly 1400 m/s, not 1540, and everything below it lands in the wrong place.

Refraction-corrected delay estimation exists in the literature, but earlier work either confined the correction to specific regions such as the skull, or to specific transmit schemes such as synthetic aperture, and demonstrated it only to about 5 cm of depth. To the best of the authors’ knowledge, this is the first work to consider the effect of speed-of-sound on image quality for deeper imaging, validated here to roughly 11 cm.

March the wavefront, not the straight line

Instead of dividing a Euclidean distance by 1540, each delay is computed as the travel time of a wavefront propagating through the actual speed map, obtained by solving the eikonal equation with the fast marching algorithm.

|∇τ′(x, z)| = 1 / c(x, z)   ·   the eikonal equation, solved by fast marching over the speed-of-sound map c
FM RUN 1 · τfoc
Aperture centre to focus
Source at the transmit centre (xt, zt), destination at the focal point (xf, zf). The first leg of the transmit delay.
FM RUN 2 · τfoc_p
Focus to every pixel
Source at the focal point, destination at the pixel (xp, zp). Negated for pixels above the focus, exactly as in the virtual-source model.
FM RUN 3 · τrx
Element to every pixel
Source at each receiving element (xi, zi), destination at the pixel. The receive delay, independent of transmit.
SUM · τp = τtx + τrx
Into standard DAS
The refraction-corrected round trip replaces the geometric delay. Apodization and summation stay untouched.

Redrawn from Fig. 2 of the paper. Each run inputs the speed-of-sound map with a source and destination; the correction works with ordinary focused transmissions.

Four media, one fat layer, three tilts

Four k-Wave simulations share the same three cysts and ten point targets. M1 has no fat layer. M2 adds a straight one; M3 and M4 incline it by 10 and 25 degrees, adding a lateral gradient to the delay error.

Speed-of-sound maps of the four simulated media M1 to M4, showing no fat layer, a straight fat layer, and fat layers inclined at 10 and 25 degrees

Fig. 3 of the paper. Speed-of-sound maps of the four media modeled with the k-Wave ultrasound toolbox: (a) M1, no fat layer; (b) M2, straight fat layer; (c) M3, inclined at 10°; (d) M4, inclined at 25°.

Centre frequency
3MHz
Array pitch / width
0.3 / 39.4mm
Grid, step
38.5×120mm, 75 µm
Focal depth
60mm
Focal lines
128
Tx apodization
HanningF# 2
Fat layer SoS
~1400m/s avg
Background SoS
1540m/s avg

What a speed map buys back

With conventional delays, cysts and pins drift as the fat layer tilts, and CY1 visibly shears. With fast marching delays computed on the (smoothed, median-filtered) ground-truth speed maps, the shifts and the shear are corrected.

Beamformed images of media M1 to M4: top row with conventional DAS showing shifted cysts and pins, bottom row with the proposed fast marching DAS showing corrected positions

Fig. 4 of the paper. Beamformed images at 60 dB dynamic range. Top row: conventional DAS for (a) M1 to (d) M4; the yellow box marks the shearing CY1, the blue box the shifting CY3, the red box the shifting pins. Bottom row (e)–(h): the proposed fast marching DAS corrects the distortions.

Table I of the paper · mean GDS over 10 point targets and gCNR for cyst CY1
Medium Mean GDS · conventional Mean GDS · fast marching gCNR · conventional gCNR · fast marching
M1 · no fat 0.70.7 0.790.76
M2 · straight 0.10.7 0.770.77
M3 · 10° 0.20.9 0.750.75
M4 · 25° 00.6 0.670.75

GDS scores a point target 1 if it sits within one wavelength of its true position, 0 otherwise, averaged over the ten targets. gCNR measures the overlap of intensity distributions inside and outside the cyst against its true location.

With the fat layer inclined at 25 degrees, conventional DAS localizes none of the ten point targets. The fast marching delays recover six, hold at least seven of ten in every other medium, and lift the worst-case gCNR from 0.67 to 0.75.

What this page does not claim

The correction shown here uses the ground-truth speed-of-sound maps from the simulator, smoothed and median filtered before feeding the fast marching solver. In a clinic no such map is handed to you; pairing the method with an approximate speed-of-sound estimation technique is the stated path forward, and the clinical case, overcoming fat-induced aberrations in liver and abdominal imaging, depends on it.

The study is simulation only, with no human or animal subjects. On an Intel Core i7-8700 workstation the delay estimation takes about 90 seconds per image. Experimental validation, approximate speed-of-sound maps, and acceleration of the fast marching algorithm are named as future work in the paper.

The paper behind this page

IEEE EMBC 2023

Fast Marching based Tissue Adaptive Delay Estimation for Aberration Corrected Delay and Sum Beamforming in Ultrasound Imaging

Asif M S · Gayathri Malamal · Madhavanunni A N · Vikram Melapudi · Rahul V · Abhijit Patil · Rajesh Langoju · Mahesh Raveendranatha Panicker
2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pp. 1–4. Asif M S and Gayathri Malamal contributed equally. A collaboration between the Center for Computational Imaging, Department of Electrical Engineering, IIT Palakkad, and GE Healthcare Bangalore.