A Comparative Study on the Influence of Probe Placement on Quality Assurance Measurements in B-mode Ultrasound by Means of Ultrasound Phantoms

2.1. Ultrasound Signal Processing

To better understand the experimental results expressed in this work on quality assurance parameters, it is appropriate to touch on the echo processing from the ultrasound probe to the gray level on the monitor: every ultrasound image is produced by sending an ultrasound wave into a medium and collecting the reflected energy (i.e. ultrasound echoes); in 2D imaging (B-mode) reflected echoes amplitudes are converted into grayscale values of pixels in the ultrasound image. The main steps of the conversion above are outlined in Fig. (1).

For each ultrasound pulse transmitted into tissues, the array probe receives a sequence of echoes that contains information related to the variation of the acoustical properties of the tissue. The time delays depend on the distance from the echo source to the probe emitting surface. Ultrasound echoes are converted into electric signals by piezoelectric transducers PT within the array probe.

In Fig. (1), the scanning is performed by the beam former, which sets the time delay for each piezoelectric transducer PT within the array probe and sums echo signals in phase to build a RF line of sight¹. The electric signal above (e.g. about 90dB Dynamic Range) is firstly logarithmically amplified by means of the Time Gain Compensation (TGC) to compensate the attenuation of the investigated medium (e.g. about 50dB), then it is filtered and demodulated in order to increase the SNR (signal processing section in Fig. 1). In spite of the TGC action, amplified echo amplitudes change over a wide range (e.g. approximately a 100:1 ratio between the strongest and the weakest echo [2]) . To display them a compression is needed (to match the lower dynamic range of the display device, e.g. 20dB).

A scheme of the dynamic compression of echoes is shown in Fig. (2): a same difference Δx in echo signal from TGC can be log-compressed using three different compression curves (a), (b) and (c) to three different echo level intervals Δz_a, Δz_b and Δz_c in the image matrix. The stronger is the compression of Δx, the wider is the echo range Δz (i.e. gray level) in the image with a higher average value, therefore, echoes are displayed with higher contrast and brightness. This is important because by compression the low amplitude echoes can be enhanced in the image, as well as image noise. Then compressed echoes are arranged in the image matrix by means of the scan converter, and before storing it some processing is applied in order to improve image quality and details (pre-processing ²).

After image storage, collected data are interpolated and converted into brightness by means of a post-processing curve and digital-to-analog conversion (D/A): the ultrasound digital image is finally displayed on the monitor of the scanner. Images used in the present work are stored in a lossless format to avoid nonlinearities due to format compression.

2.2. Ultrasound Phantom

In the present work two Point Spread Function (PSF) Phantoms models are used for image quality assessment of a real time B-mode scanner, GAMMEX 405 GSX LE (phantom A) and GAMMEX 1425A (Phantom B). Both of them made by a tissue mimicking material embedded with nylon wires (whose transversal sections are considered as point target in the ultrasound image), low scatter targets (cystic masses) and gray scale targets (only phantom A), whose characteristics are shown in Table 1.

From Table 1, it can be observed that both phantoms have the same speed of sound and very similar pin targets arrangement (along both vertical and horizontal directions with the same spacing) but very different attenuation (phantom A attenuation is almost 40 percent higher than phantom B) and dimensions (phantom B is bigger and heavier than phantom A, with a wider transversal section).

2.3. Measurement Set-up

The measurement setup includes an ultrasound scanner Philips HD3, an array probe on the ultrasound phantom and a notebook pc for data acquisition and processing. Tests are performed with two different models of array probe (convex and linear array) on two tissues mimicking phantoms³ (see Table 1 for specifications) by a single operator. In particular, our study focuses on evaluation of the follow image quality parameters: (a) accuracy in distance measurements (both vertical and horizontal directions) [3, 5], (b) high contrast spatial resolution (both lateral and axial resolution) [5] and (c) maximum depth of signal visualization (or maximum depth of penetration) [6-8, 18]. The ultrasound probe is applied on the phantom by hand and its position is manually adjusted to clearly display test objects in the US image (no holder is used), then a single image is acquired and exported on a peripheral in a TIFF or DICOM format (lossless): for each phantom the whole procedure is repeated from 11 to 16 times. Therefore for a same test (Tables 2 and 3) a group of 11 up to 16 ultrasound images is provided and processed by an in-house software package, developed in a Matlab environment [3, 5-10, 15] to evaluate the variation of the parameters (a), (b) and (c) due to the probe manipulation on the phantom surface and whose operating principles are explained in the following paragraphs. The use of self-produced software is mainly due to economic reasons and transparent management of data processing, although comparable products are available on the market [17]. Settings are chosen to display the maximum number of test objects at the central frequency of the ultrasound probe (working frequency). Moreover all images are acquired within ±2° tilt angle, at medium overall gain and linear post processing while persistence, edge enhancement and other image processing are set to minimum.

To understand the settings better in Table 2 some definitions are reported below [5-9]:

• Nominal Frequency (of a transducer): acoustic working frequency of a transducer as quoted by the designer or manufacturer. In our study the ultrasound scanner provides a nominal frequency range, i.e. 2-5 MHz and 5-9MHz for the convex and linear array probe respectively.

• Field of View (FoV): area in the ultrasonic scan plane from which ultrasound information is acquired to produce one image frame. Since for each probe the diagnostic image size (in pixel) is quite constant, a FoV variation is associated to a different scale factor (mm to pixel ratio) and could influence other image characteristics, i.e. spatial resolution, accuracy in distance measurements.

• Focus Depth: nominal depth of the transmission focus in the ultrasound image. Each transmission focus is usually marked on the diagnostic image.

• Dynamic range (global): ratio of the maximum to the minimum echo-signal amplitude, even with changes of settings, that a scanner can process without distortion of the output signal.

• Local Dynamic Range: ratio, expressed in decibels, of the minimum echo amplitude that yields the maximum gray level in the digitized image to that of the echo that yields the lowest gray level at the same location in the image and the same settings. Local Dynamic Range influences image contrast and can be controlled by the operator (i.e. through dynamic range or compression settings).

• Gray Scale Mapping Function (GSMF): relationship between echo amplitudes and gray levels on the image. From GSMF the Local Dynamic Range can be estimated.

• Overall gain: the overall gain control amplifies all echo signals equally, irrespective of when they return to the transducer. Increasing the overall gain cannot improve significantly the Signal to Noise Ratio (SNR), because noise and signal are both amplified.

• Transmit Power: the transmit power controls the amplitude of pulses transmitted by the transducer. As the transmit power increases, echoes amplitudes grow, usually with a better Signal to Noise Ratio (SNR). On the other hand, by reducing the transmit power the exposure of the patient to ultrasound decreases and so the risk of any adverse effect (ALARA principle).

• Pre-processing - edge enhancement: a spatial high-pass filter to enhance the appearance of anatomical features in the ultrasound image.

• Pre-processing – persistence: it is a frame averaging to suppress random image noise.

• Pre-processing – zoom (read zoom): it is a way of magnifying a part of the ultrasound image, to overcame the inefficient use of the screen area when the imaged depth is larger than the imaged width and anatomical features appear small on the display.

• Post-processing: mapping functions that relate echo amplitudes to displayed brightness levels.

2.4. Accuracy in Distance Measurements

Vertical and horizontal distances are measured “from peak to peak” of the wires sections within the two phantoms and compared with their nominal values by an in-house software [3]: in particular, after the ultrasound image is processed in a workstation, first, the scale factor is automatically calculated from FoV then the operator is asked to choose n test object pairs by clicking on the image without care about centering each target because their barycentric coordinates are automatically calculated within a Region of Interest (ROI). These latter are then used for measuring the distance between each pair of wires, so errors due to operator subjectivity can be neglected. Therefore, the measurement distance relative error e_k=|s_rk−s_mk|/s_rk×100 between the nominal (s_rk) and the measured distance (s_mk) in the image is calculated for each pair (k=1, 2, ...n) along both horizontal and vertical distances [3, 11]. For the k-th distance and for each test, the accuracy in distance measurements is calculated by averaging relative error values e_k over the images of the same group Fig. (3).

Since distance accuracy depends on the difference between the effective acoustic velocities of the medium and the acoustic velocity used to calibrate the scanner, as well as on image scan conversion, scanning geometry and pixel dimensions [3], a dependence on FoV settings is expected with different results between horizontal and vertical distances.

2.5. High Contrast Spatial Resolution

High-contrast resolution characteristics can be evaluated locally from the size of the point-spread function (PSF), that is the characteristic response of the imaging system to a high-contrast point target: PSFs can be displayed by imaging high reflection targets that are smaller than one wavelength, as the nylon wires sections in the US phantoms that have been used in this work. In diagnostic ultrasounds the PSF is not isotropic, with different lateral and axial dimensions in the FoV, depending on distance from the transducer emitting surface: in other words, spatial resolution features change with position and depth in the US image, therefore many measurements of the lateral and axial diameters of the displayed wires sections are performed at different positions and depths (by the in-house software for all the settings in Table 2) to achieve the ultrasound system’s lateral and axial resolutions. In particular every wire section k (point target) in the US image is processed by a threshold algorithm at Full Width Half Maximum to evaluate both its maximum length PSF_Zmax(k) and width PSF_Xmax(k), whose values are assigned as high contrast axial and lateral resolution respectively Fig. (4).

Since PSF measurement depends also on pixel dimension and contrast between point target and tissue background, a dependence on FoV settings and phantom attenuation is expected with different results between horizontal and vertical dimensions.

2.6. Maximum Depth of Penetration

Maximum depth of penetration or Maximum Depth of Signal Visualization (DSV_max) is the maximum depth at which echo signals from scattering within a tissue-mimicking phantom can be detected [6-8, 15, 18]. In particular, DSV_max is achieved from a depth profile of gray levels within a ROI_DP of 30 pixel width⁴ in the US image, by means of a threshold at 2dB [8] above the mean value displayed in a 10×10 pixel ROI_n positioned at the end of the FoV⁵ (Fig. 5). To evaluate the +2dB threshold, the relationship between gray levels and echo amplitudes (dB) in the image is needed [8]: to this aim a GSMF is provided by means of the method proposed in [10] for each probe and dynamic range setting in Table 2. The method above is implemented in a specific software tool developed by the authors [15] to provide measurement results of DSV_max.

Since DSV_max measurement depends on SNR ratio (between speckle signal and electronic noise), a dependence on dynamic range settings and phantom attenuation is expected.

3.1. Accuracy in Distance Measurements

Accuracy in distance measurements is evaluated for the settings in Table 2 and 3 as expressed in section 2.4: measurements ranges and their uncertainty are shown in Tables 4 and 5, diagrams are plotted in Fig. (6).

In Fig. (6) results for the convex probe are shown: for vertical measurements no significant differences between tests are observed, while for horizontal distances (Fig. 6b and Fig. 6d) discrepancies between phantoms can be noticed. Moreover a slight difference in the horizontal distance relative error between tests (i.e. test 2, test 4) seems to confirm a FoV influence on results at lower dynamic ranges. To evaluate the test influences on measurement reproducibility a Fisher test (p-value<0.05) has been performed for results at horizontal-60mm and vertical-80mm distance: although in vertical distance measurements no significant differences can be observed among variances, in horizontal ones homoscedasticity should be rejected for both of phantoms, therefore a variation in measurement reproducibility related to test settings can be supposed. Measurements uncertainty ranges from ±0.3 to ±1.2 percent for horizontal distances, ±0.2 to ±0.7 percent for vertical distances. A hypothesis test has been done to verify whether the differences in accuracy measurements are statistically significant: results of the Kruskal-Wallis non parametric method and the Dunnet’s test post hoc analysis are reported in Tables 6 and 7 and a significant difference among test is found for both phantoms depending on distances.

Results for the linear array probe are shown in Fig. (7) and no significant differences are noticed for most of the measurements (see also Tables 8 and 9), nevertheless the F-test (p-value<0.05) indicates no equality of variances on horizontal (30mm, both of the phantoms) and vertical distances (20 mm, phantom A). Measurements uncertainty ranges from: ±0.1 to ±0.6 percent for horizontal distances, ±0.1 to ±5.5 percent for vertical distances. As seen for the convex array probe, a hypothesis test is performed on accuracy measurements for the linear array and results are shown in Table 10.

For the convex array transducer the differences between tests in horizontal distance can be attributed to phantom tissue mimicking material and mostly to FoV settings (Fig. 6b and Fig. 6d), because it influences mm to pixel ratio (i.e. scale factor) and the distortion of imaged nylon wires sections. In fact at lower FoVs, wires are selected near to the image boundaries, where distortion artifacts are stronger and become more evident for higher contrasts (i.e. Dynamic Range settings). All the above causes are likely to affect the measurement reproducibility, because of the enhancement of slight differences in the acquired images and their processing by the analysis software. Analogous considerations can be done for the linear array transducer, where high noise levels add dispersion to results. In particular, a singular dispersion is observed for vertical distance error e_v in test 4 of Fig. (7a), likely due to a combined effect of the higher attenuation in phantom A and the increased noise with Dynamic Range reduction, that involved a reduced pin target visibility and software failures in barycentric calculus.

3.2. High Contrast Spatial Resolution

Lateral and axial resolution at FWHM are evaluated for settings in Table 2 and 3. Convex array measurements results are plotted in Fig. (8) and summarized in Tables 11 and 12.

In particular no significant variations between tests have been noticed for most of lateral and axial measurements of the convex array probe (Fig. 8), nevertheless some discrepancies between test can be observed in phantom A. A Fisher test (p-value<0.05) has been performed on parameter variances at 100 mm (both for lateral and axial resolution) and suggests that for results in Fig. (8c and Fig. 8d) (phantom B) homoscedasticity should be excluded. Measurements uncertainty ranges from: ±0.3 mm to ±1.9 mm for lateral resolution, ±0.2 mm to ±1.0 mm for axial resolution. Results of the hypothesis tests on spatial measurements are shown for both probes in Tables 13, 14 and 17.

Linear array spatial measurements are shown in Fig. (9) and no significant discrepancies can be noticed for most of the measurements, nevertheless no equality of variances, on both axial and lateral resolution values with the phantom B, are suggested by the F-test. Measurements results and uncertainty ranges are shown in Tables 15 and 16: uncertainties range from ±0.1 mm to ±1.4 mm for lateral resolution, from ±0.1 mm to ±0.4 mm for axial resolution. Differences between tests can be likely due to electrical noise level (whose display depends on Dynamic Range settings), phantom tissue mimicking material (attenuation) and FoV variation, because of its influence on both the scale factor and the pixelation error. As in distance evaluations, measurement reproducibility likely depends on all the above causes, because of the enhancement of differences among tested images.

3.3. Maximum Depth of Penetration

DSV_max is evaluated for settings 1 and 2 in Table 2 and 3: results are plotted in Fig. (10) and reported in Tables 18 and 19.

In Fig. (10a) a significant disagreement between phantoms is noticed for both the settings in the convex array probe: higher penetration depths are related to phantom B and measurements uncertainty ranges from ±1 mm to ±4 mm. On the other hand the F-test (p-value<0.05) suggests that homoscedasticity should be refused for phantom B tests. In Fig. (10b) a considerable discrepancy in results between tests and phantom has been observed for the linear array probe: higher penetration depths are consistent to each other and related to settings 2 for both of the phantoms and measurements uncertainty ranges from ±1 mm to ±10 mm, while variances should be considered non-homogeneous between tests for both of phantoms. In the convex array DSV_max seems to depend mostly on phantom model than test settings, on the other hand it is more settings dependent in linear array examinations while differences between phantom tests are not significant. As in the previous sections, a hypothesis test (Kruskal-Wallis) has been carried out to verify whether differences in maximum depth of penetration measurements are statistically significant: p-values are reported in Table 20 and showed significant differences between tests for both the phantoms.