How To - Do Ultrasound Calibration

Table of Contents

• Introduction
• Choice of Calibration Phantom
• Collecting Data for Ultrasound Calibration.
• Sorting Data
• The Calibration

Introduction

While NifTK is not a toolkit dedicated specifically to ultrasound image calibration, there are a few programs that may be useful. The aim in ultrasound calibration is to calculate the 6DOF transformation matrix, and 2 scale factors that relate pixels in an ultrasound image to the coordinate system of the tracking markers attached on the probe. Most methods rely on collecting synchronised pairs of ultrasound images and tracking data. A survey of methods can be found here.

Key issues for success are:

1. Detecting image features (e.g. a single pin or cross-wire) reliably and accurately in many ultrasound images.
2. Collecting ultrasound images and tracking information, that are accurately synchronised in time.

For the first point, there is no automatic feature detection in NifTK. The user must manually detect features by clicking on them. For the second issue, the user must be aware of the timings of each data feed, and optimise their experimental setup to reduce errors.

Choice of Calibration Phantom

We have tried a single pin, single cross-wire, multiple cross-wire and the PLUS fCal-2.0 phantom. A single, stationary, cross-wire seems (to Matt) to be the most workable example. This method is known to be tedious, but if done correctly it can be as accurate as a more elaborate setup. For example, you could grab each frame of data, where a frame means a matching pair of ultrasound image and tracking matrix, making sure that every piece of hardware is clamped stationary, thereby minimising movement and timing errors.

If you want to do automated N-Wire calibration, consider using the PLUS toolkit, as CMIC has a 3D printed PLUS fCal-2.0 phantom. For the remainder of this page, we assume that calibration is performed using a single pin or single cross-wire. Mathematically, these two methods are equivalent, as they both utilise a single invariant (stationary) point. This point must be stationary with respect to the tracking device. The potential advantage of a single cross-wire over a pin head is that it is easier to determine when the ultrasound beam is directly over the cross. If using a single pin, try pins with different sized heads, as you need to systematically pick the centre of the pin in the ultrasound image.

Collecting Data for Ultrasound Calibration.

In NifTK:

• When using Tracking Devices such as Spectra, and Ultrasonix machines, record data using the IGI Data Sources plugin.
• When using visual tracking, record using the appropriate custom plugin. The principal is the same.

Make sure:

• Cross-wire must be stationary with respect to the tracker. Glue/Clamp things if necessary.
• You are advised to move the ultrasound probe slowly. However 'slowly' is in relation to the speed of data collection. You need to collect data that accurately represents the cross-wire, and have an ultrasound image and tracking matrix that exactly capture that point, free of timing or movement errors.
• Scan the single cross-wire from as many angles as possible, covering all degrees of freedom
• Don't forget the 'depth' degree of freedom. i.e. Up/Down relative to the cross-wire. If the tracking markers must be above water, then the cross-wire can be imaged at several depths by incrementally filling the water bath.

If you clamp things stationary, you can save each individual image and tracking matrix manually. If you record data while moving freehand, you will need to us the next program niftkUltrasoundPinCalibrationSorter.

Sorting Data

Given:

• A directory of tracking matrices, each one with a filename that is <timestamp>.txt
• A directory of .png files, each one with a filename that is a <timestamp>.png
• The natural (i.e. by name) sort order must be the same, which it will be if the timestamps are correctly formed. Timestamps should be 19 digits.

Use the program

niftkUltrasoundPinCalibrationSorter --inputMatrixDirectory <dir1> --inputImageDirectory <dir2> --outputMatrixDirectory <dir3> --outputPointDirectory <dir4>

to quickly scan through data. The controls are:

• N - next image
• P - previous image
• left click - click on the ultrasound feature. You should click exactly once per frame, corresponding to the single pin or single cross-wire.

For each click, a single tracking matrix, and corresponding cross-wire pixel location is saved. If you click multiple times on each ultrasound image, each subsequent click over-writes the previous click on that image. The program checks if the timing difference between the ultrasound image and tracking matrix is suitable. The default tolerance is 40 milliseconds, but this can be adjusted using the argument –timingTolerance. Also, if you have a lot of ultrasound data, where there is no tracking data, you can use the –skipForward flag which allows the program at each 'N' key stroke to jump forward until the next ultrasound image that has tracking data within the given timing tolerance.

Use of this program results in well matched images and tracking matrices of the single cross-wire. It also means that you should be able to record data free-hand, without clamps.

The Calibration

There is a C++ program in NifTK called niftkUltrasoundPinCalibration. This was tested, and checked, but never seemed to work properly. It was concluded that the Levenberg Marquardt algorithm was not converging properly. Therefore the calibration must be done in MATLAB. MATLAB proved to have the most reliable non-linear optimisation routine, but the maths is identical to the C++ program.

You must:

1. Concatenate all tracking data into a single file e.g. matrices.txt
2. Concatenate all point data into a single file e.g. points.txt
3. The ordering of matrices in matrices.txt and points in points.txt must correspond.
4. These files should have at least 50 samples of data. i.e. matched 50 matrices, and 50 corresponding single points.
5. In MATLAB, addpath NifTK/Scripts/UltrasoundPinCalibration
6. The calibration routine is: niftkUltrasoundPinCalibrationFromFile.m

The MATLAB scripts reads a single file containing all the tracking matrices in order eg:

0.205221 0.245414 -0.947447 -1.43977
0.580367 -0.810001 -0.0841022 2.30268
-0.788073 -0.532607 -0.308659 167.379
0 0 0 1
0.211506 0.235885 -0.948485 4.94036
0.563073 -0.822621 -0.0790214 1.9959
-0.798883 -0.517352 -0.30681 169.94
0 0 0 1
etc

and also all the points in order eg:

180 162
253 166
etc

The parameters are:

>> help niftkUltrasoundPinCalibrationFromFile
 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

  Usage:
    [finalParams, sumsqs, residuals] = niftkUltrasoundPinCalibrationFromFile(initialGuess)
  where:
    initialGuess : parameters array [tx, ty, tz, rx, ry, rz, x, y, z, sx, sy]
                   where:
                   tx, ty, tz = translation in millimetres
                   rx, ry, rz = rotations in radians
                   x,y,z      = location of invariant point in millimetres
                   sx, sy     = scale factor (mm/pix)

There are also MATLAB evaluation routines such as

>> niftkUltrasoundPinCalibrationEvaluation - leave one out cross validation, with accuracy assessment against a gold standard location.