Automated Hierarchical Time GainCompensation for Ultrasound Imaging

Ramin MoshaveghSupervisors:

Martin Christian Hemmsen, Bo Martinsand Jørgen Arendt Jensen

Center for Fast Ultrasound Imaging (CFU), Department of Electrical Enginering,Technical University of Denmark

SummaryBackground• Radiofrequency (RF) echoes are strongly attenuated by the tissues scanned.• Time gain compensation (TGC) is usually utilized to compensate for the acousticattenuation.

• Scanning rely on the interaction with the medical doctor to optimize the scansettings.

• Several adjustments on the keyboard of the modern scanners.

Objective• Decrease the adjustments done by a medical doctor on the ultrasound scanner andoptimize the quality of the scans.

Problem• Automatic time gain compensation used in ultrasound scanners weakens the edgesand over-gains large fluid collections such as urine bladder or gallbladder (anechoic regions).

Approach• Estimating the attenuation map using log spectral difference method to correct thegains inside the anechoic regions.

Future work• Using Deep Learning Architectures for segmentation and tracking of tissuesin ultrasound scans.

• Learning hierarchical features for scene segmentation.

What has been done previously for TGCTGC offsets the attenuation of ultrasound echo signals along the depth so that echoesbelonging to deep structures are more amplified compared to superficial echoes. Thisprovides more uniform signals to be displayed on the scanner.

−10−8−6−4−2 0 2 4 6 8 10

02468

101214

Lateral position [cm]

Axi

alpo

sitio

n[c

m]

0 51015

02468

101214

dB

−10−8−6−4−2 0 2 4 6 8 10

02468

101214

Lateral position [cm]

−8−6−4−2 0 2 4 6 8

02468

1012

Lateral position [cm]

Axi

alpo

sitio

n[c

m]

0 5101520

02468

1012

dB

−8−6−4−2 0 2 4 6 8

02468

1012

Lateral position [cm]

• Very good performance on abdominal scans of human liver and bladder.

• TGC over-gains the anechoic region ( inside the bladder).

Our approach: Latest results• Estimating the attenuation slopes, and generating the 2-D attenuation map usingthe spectral log difference of RF-data.

0.02 0.04 0.06 0.08 0.1 0.12 0.140

2

4

6

rfsignalnumber

Axial distance [m]

5 rf signal of (raw data)

3 3.2 3.4 3.6 3.8 4 4.2 4.4

−80

−75

−70

Frequency [MHz]

Log

PS

D[d

B]

DistalProximal

(a) Example Log PSD of a paired proximal anddistal segments, where the f0 = 3.75 MHz.

3 3.2 3.4 3.6 3.8 4 4.2 4.4−6

−4

−2

0

Frequency [MHz]

Log

PS

Ddi

ff[d

B] Log PSD diff

Fitted line

(b) Log PSD difference, and fitted line with slope=−0.502 dB/cm×MHz.

Illustration of how the attenuation value for a pair of proximal and distal segments are computed.

• The attenuation maps characterising the scanned media are computed and used tocorrect the mis-adjusted high gains inside the anechoic regions.

−6−5−4−3−2−1 0 1 2 3 4 5 6

012345678

Lateral position [cm]

Axi

alpo

sitio

n[c

m]

(a) Normalized attenuation map of a bladder scan.

−6−5−4−3−2−1 0 1 2 3 4 5 6

012345678

Lateral position [cm]

00.10.20.30.40.50.60.70.80.9

1

perunit

(b) Normalized attenuation map of a gallbladder scan.

Examples of normalized attenuation maps overlaying on B-mode images

• The original scan including a large collection of fluid (urine bladder) is firstcompensated using a TGC curve. The 2-D attenuation map is then applied to avoidover-gaining inside the bladder.

−5−4−3−2−10 1 2 3 4 5

01234567

Lateral position [cm]

Axi

alpo

sitio

n[c

m]

−80−70−60−50−40−30−20−100

dB

(a) Un-processed sagittal scan of human bladder.

0 1020

01234567

dB

(b) TGC

−5−4−3−2−10 1 2 3 4 5

01234567

Lateral position [cm]

−80−70−60−50−40−30−20−100

dB

(c) ATGC compensated Image (a) withcurve (b).

−5−4−3−2−10 1 2 3 4 5

01234567

Lateral position [cm]

Axi

alpo

sitio

n[c

m]

00.20.40.60.81

p-u

(d) Normalized attenuation map com-puted from image(a).

−5−4−3−2−10 1 2 3 4 5

01234567

Lateral position [cm]

Axi

alpo

sitio

n[c

m]

−80−70−60−50−40−30−20−100

dB

(e) AHTGC corrected image (c) with2-D map(d).

Illustration of how the proposed AHTGC algorithm is applied to a sagittal scan of human bladder. (a)Un-processesed scan. (b) TGC curve computed for image (a). (c) ATGC compensated image (a) withcurve (b). (d) 2-D attenuation map compuetd from image (a). (e) Attenuation corrected image (c)with 2-D map (d).

Results and Discussion• Matching Pairs of in vivo sequences, unprocessed and processed with the proposed AHTGC were visualized side by side and evaluated by two radiologists in terms of imagequality.

• Wilcoxon signed-rank test was used to evaluate whether radiologists preferred the processed sequences or the unprocessed data.• The results indicate that the average VAS score is positive ( p-value: 2.34×10−13) and estimated to be 1.01 (95% CI: 0.85; 1.16) favoring the processed data with the proposedAHTGC algorithm.

• The 2-D attenuation profiles also provide solid foundation for other processes like segmentation of the tissues.

Summer school on deep learning for image analysis, August 2014, Langeland, Denmark. Preprints from: jaj@elektro.dtu.dk