Oct 4~6, 2017
An optimization method for correction of ultrasound probe-related contours to head-centric coordinates Wei-rong Chen1, Mark Tiede1, Shuwen Chen2 1 Haskins 2
Laboratories
The Chinese University of Hong Kong
[email protected],
[email protected],
[email protected] Ver: oct-06-2017
1
Challenge • One well-known challenge for using ultrasound in phonetic studies is the correction of probe/head movements. Tongue held still, probe moving
Relative movements between probe and head induce errors in ultrasound images.
2
Two solutions 1 Constraint method Head stabilizer headset (helmet) Example: Restrict probe/head/jaw movements. How: Challenges: 1. Difficult to guarantee purely no movements 2. Potentially affects naturalness of speech.
2 Correction method Example: How:
HOCUS system (Whalen et al, 2005) Track and transfer probe/head movements to ultrasound contours as correction
Challenges: 1. Difficult to define the origin of coordinates. 2. Accuracy is difficult to assess. Whalen, D. H., et al. (2005). "The Haskins Optically Corrected Ultrasound System (HOCUS).” JSLHR 48(3): 543-553.
3
Correction method • 3 tasks for correction method: - Mutual scaling (easy task) - Transformation of the orientations of the planes in coordinates (solved in HOCUS) Not settled yet! - Determining the mutual origin.
4
Objectives of the current work • This works demonstrates a simple way for correction methods to : 1. Calibrate the origin of coordinate system for ultrasound probe. 2. Assess the accuracy of the correction of probe/head movements. 5
“Floating ruler” experiment1 - 3D EMA (NDI WAVE) sensors on a ruler (to simulate speaker’s head) and probe - Manipulate ruler/probe movements. Pick initial guess of origin. - Optimize for the origin by aligning contours in ultrasound image (estimation) and in EMA (truth). - Use the optimal origin for the correction of ruler/probe movements.
1. Thanks
to Doug Whalen for this cool name!
6
Floating ruler experiment U1-U3: EMA sensors on probe RR1-RR3: EMA sensors on the upper part of ruler R1, R2: EMA sensors on the bottom part of ruler Ruler bottom edge: position of ruler bottom edge relative to R1R2, traced from photo image. Probe top surface: position of probe top surface relative to U1U3, traced from photo image.
7
Floating ruler experiment 6 settings: 1. Baseline: aligning ruler and probe in straight, centered position. 2. Rot_ccw: Rotate ruler by 20~25° counterclockwise. 3. Rot_cw: Rotate ruler by 20~25° clockwise. 4. Shift_L: Shift ruler to the left by approx. 2cm 5. Shift_R: Shift ruler to the right by approx. 2cm 6. Shift_D: Shift ruler down by approx. 1cm
8
Contour tracking in ultrasound images baseline
GOAL of head/probe movement correction: Accurately recover/transform the contours in 2-6 to fit the baseline, by removing probe/ruler movements 9
TASK: Co-register the origins in both coordinate systems Where is the origin [0,0] ?
10
STEP 1: Initial guess for the coordinate origin (based on our understanding of how ultrasound works)
11
Pick our best guess for the coordinate origins Ultrasound
EMA
12
STEP 2: Transfer probe movements to ruler (head) movements (Remove probe movements w.r.t. baseline)
13
After probe-movements are removed If the origin is correct: - ruler bottom contours in ultrasound and in EMA should perfectly align. Ultrasound:
EMA: 14
Calculate the error of alignments with initial guess
US = ultrasound
15
STEP 3: Find the optimal origin: 1. Define cost function J(θ) = mean distance between ruler bottom contours in ultrasound and in EMA, where θ = the position of origin. 2. Solve 𝜃 = 𝑎𝑟𝑔𝑚𝑎𝑥𝜃 𝐽(𝜃) using gradient descent.
16
Compare optimization results Initial guess:
Optimized: 17
STEP 4: Correct ruler/head movement with optimized origin 1. Calculate and store the ruler/head movements w.r.t baseline 2. Transfer ruler/head movements to ultrasound contours in 2-6 as correction.
18
Calculate and store the ruler movements w.r.t baseline For each setting in 2-6: Calculate the rigid body transformation (rotation R and translation T) from the baseline (setting1). B = baseline ruler position (setting 1), A(i) = ruler position of the setting i. B = R(i) x A(i) + T(i) R(i) = rotation from A(i) to B T(i) = translation from A(i) to B
19
FINAL RESULTS of correction: Update the origin in ultrasound by the optimal value Apply R and T to ultrasound contours to correct for probe/ruler movements
‘x’: initial guess of origin (0,0) ‘x’: optimal origin (-2.8, -2) Mean error after correction = 0.45 mm
20
RESULT: APPLY optimal origin to tongue contour correction:
21
Discussion • How accurate was our initial guess of origin? Actually not too bad. Initial guess [0,0] vs. optimal [-2.8, -2] in millimeter is not a huge difference. In most settings, the improvement of accuracy by the optimal origin is less than 1 mm.
22
Discussion • Why optimal origin and initial guess are different? • What is the source of this discrepancy? Some possibilities: Machine-internal process. Errors in hand-picking initial guess
23
Discussion • Since the initial guess is not bad, then why bother to find the optimal origin? Uncertain errors occur when hand-picking the initial guess. Uncertainties about the internal processes in ultrasound machine. Good initial guess for this machine doesn’t guarantee the same accuracy for another machine. The proposed “floating ruler” is a simple and cheap method of calibration. Just collect 6 frames only once for each probe in each machine. 24
Discussion • Alternatives for EMA? The use of EMA is for tracking probe/ruler(head) movements. Alternatives: - Optical tracking as in HOCUS system - Video-based landmark tracking (e.g., ‘blue-dot-tracking’) - 3D video camera (e.g., Microsoft Kinect) - 3D facial recognition technology (e.g., OpenFace)
25
Conclusion • The proposed “Floating ruler” method is a simple and necessary procedure for both calibration and accuracy assessment if head-correction for ultrasound contours is performed.
26
Future directions • Simplify and standardize the procedure. • As the transformational correction is only affected by ‘ rotation’, more settings with different probe angles need to be tested. • Apply the optimal origin to the head-correction for speech data.
27
Acknowledgements • We thank Doug Whalen, Christine Shadle, Emily Phillips, Kevin Roon for comments and helps. • This works is supported by NIH grant DC-002717 to Haskins Laboratories.
28
