An Automated Region-Selection Method for Adaptive ALARA Ultrasound Imaging

Abstract

The objective of this work was to develop an automated region of interest selection method to use for adaptive imaging. The As Low As Reasonably Achievable (ALARA) principle is the recommended framework for setting the output level of diagnostic ultrasound devices, but studies suggest that it is not broadly observed. One way to address this would be to adjust output settings automatically based on image quality feedback, but a missing link is determining how and where to interrogate the image quality. This work provides a method of region of interest selection based on standard, envelope-detected image data that is readily available on ultrasound scanners. Image brightness, the standard deviation of the brightness values, speckle signal-to-noise ratio, and frame-to-frame correlation were considered as image characteristics to serve as the basis for this selection method. Region selection with these filters was compared to results from image quality assessment at multiple acoustic output levels. After selecting the filter values based on data from twenty-five subjects, testing on ten reserved subjects’ data produced a positive predictive value of 94% using image brightness, speckle signal-to-noise ratio, and frame-to-frame correlation. The best-case filter values for using only image brightness and speckle signal-to-noise ratio had a positive predictive value of 97%. These results suggest that these simple methods of filtering could select reliable regions of interest during live scanning to facilitate adaptive ALARA imaging.

I. INTRODUCTION

To prioritize patient safety, US Food and Drug Administration (FDA) guidelines recommend following the As Low As Reasonably Achievable (ALARA) principle when using ultrasound [1]. In this framework, the end-user should limit the acoustic output to what is necessary to make a diagnosis. To help inform these decisions, safety indices, like the mechanical index (MI), are displayed on the ultrasound system [2]. However, studies have shown that surveyed ultrasound users generally do not know what these safety indices mean or where they are displayed [3], [4]. Furthermore, an eye-tracking study by Drukker et al. found that the indices were viewed in just 4.2% of scans and, even then, they were only checked once [5]. This represents a significant gap between ultrasound safety guidelines and clinical practice. The current system of ultrasound regulation in the United States depends on an informed user to respond to imaging conditions and make adjustments to the settings to avoid unnecessary exposure, but it appears that the acoustic output is often overlooked.

One approach to address this problem is building these adjustments into the scanner software. We have previously shown that an automated implementation of the ALARA principle for fetal ultrasound imaging can be performed with only a few lines of data and so fast as to be functionally invisible to the sonographer [6]. This method is based on measuring image quality with lag-one coherence (LOC) [7]. LOC is the spatial coherence of the data received by the ultrasound transducer on neighboring (lag-one) channels. These calculations can be done on a subset of the image in real time to provide image quality feedback [8]. To identify the ALARA acoustic output level, LOC is measured at several MI settings. As the MI is increased, the image quality initially increases dramatically, but then begins to approach an asymptote. These data can be fit with a logistic curve. Then, the MI at which the curve reaches 98% of its maximum is considered the “ALARA MI” [6], [7].

A limitation of these ALARA ultrasound studies is that the data were collected in a fixed image position, which relied on the sonographer to align the region with the structure of interest. If the structure of interest was not at the assumed location, which was at the lateral center and the focal depth, the optimization would be less reliable. Adding the additional tasks of positioning the region of interest (ROI) and triggering the ALARA acoustic output calculation to the obstetric ultrasound exam protocol would disrupt the clinical workflow. In a standard second trimester anatomy exam, dozens of structures are imaged [9], and it can take nearly an hour [10]. Part of this exam length is attributable to the many controls the ultrasound user has to adjust, including frequency, depth, compounding, focal depth, and dynamic range [11]. Incorporating ROI selection into the automated ALARA method would help make clinical implementation feasible.

Automated region selection has been attempted in other areas of ultrasound, such as shear wave speed measurements. Although these regions are generally manually selected when assessing liver fibrosis [12], choosing the shear wave elastography image (SWEI) ROI with the lowest standard deviation has also been proposed [13]. For the manual selection method, there are guidelines for where the user should place the ROI, and several measurements are taken to establish an average value for that patient [12].

Sophisticated techniques have been developed that can automate some ultrasound detection tasks [14]. For example, a fully automated segmentation approach has been proposed for breast ultrasound by Shan et al. [15]. It is a region growing method that employs automatic seed point selection with an artificial neural network for lesion detection. Image texture, intensity, and spatial characteristics are all used by this system, and it achieves a sensitivity of 93%. However, each case takes about 9.5 seconds to run, which is faster than the two other segmentation techniques it was compared to, snake and level-set, but it is not fast enough for our purposes.

In this study, local statistics of the beamformed data were used to establish a method of automated ROI selection for ALARA MI calculations. This technique uses the imaging data that are already collected during live scanning to inform where to perform the ALARA optimization. The images are subdivided into patches where the image statistics are calculated, specifically image brightness, standard deviation, speckle signal-to-noise ratio (SNR), and frame-to-frame correlation. These filters were assessed independently and in combination for their ability to predict where successful ALARA MI calculations could be performed. Additionally, preliminary clinical evaluation of the ROI selection method was conducted. The goal of this work is to facilitate ALARA ultrasound imaging through adjustments to the ultrasound system design. A real-time, automated implementation of ROI selection would be needed for ALARA acoustic output adjustments to fit into the clinical workflow and be performed throughout the scan. Fig. 1 shows how this ROI selection method could be used with the rest of the ALARA MI calculation steps to fully automate the system.

Fig. 1.

The steps involved in ALARA MI imaging using the proposed ROI selection method.

II. METHODS

A. Data Collection

As described in a previous paper [6], data were collected at the Duke University Hospital Fetal Diagnostic Center following an approved Duke Health Institutional Review Board (IRB) protocol that included obtaining informed consent from all participants. Thirty-five volunteers with healthy, singleton, second-trimester pregnancies were enrolled in this study. The first twenty-five volunteer data sets were used as training data, and the remaining ten were used for testing. Channel echo data were collected using a Verasonics Vantage 256 research ultrasound system (Verasonics, Kirkland, WA) and a C5-2v transducer. Pulse-inversion harmonic images were acquired with a transmit frequency of 2.36 MHz and an F/2.0 aperture. An expert fetal sonographer with more than twenty-five years of experience performed the scanning. Focal depths between 4 and 9 cm were used in this study. The focal depth was chosen for each data set by the sonographer depending on the location of the structure of interest. For each participant, there were three data sets that focused on the placenta and three data sets that focused on the fetal abdomen. The transducer was lifted off the subject’s abdomen between consecutive data sets to allow for independent samples.

Frames of channel data included 100 image lines and had a 50° span. Sixteen consecutive frames were collected at increasing Mechanical Index (MI) values, evenly spaced between 0.15 and 1.20. For the deepest focal depth (9 cm), a maximum MI of 1.06 was used because of system output limitations, so these swept-MI data sets only included fourteen frames. Immediately after these frames were acquired at increasing MI values, two frames were collected at a fixed MI of 0.8. This MI value corresponds to the default of the routine second trimester preset used at the Duke University Hospital Fetal Diagnostic Center. All data were collected at fourteen frames per second and used dynamic-receive beamforming.

B. Data Analysis

Clinical data were used to answer two questions: (1) With frames of swept-MI data, which regions worked well for ALARA optimization? (2) With two frames of fixed-MI data, can we predict what regions will work well? For the first question, data frames were processed offline to identify which regions would support ALARA optimization. This process, which served as a “gold standard” for this study, is more time-intensive and computationally demanding. It is only used for development in this study and is not proposed to be a part of the real-time system. The proposed real-time system would use the predictive method, which is based on standard B-mode data that could be quickly and easily accessed on clinical ultrasound systems. We assessed different statistics and combinations of statistics of this envelope-detected, fixed-MI data to see how well they matched the results from the full analysis.

1). Lag-one Coherence:

LOC is the image quality metric used in the ALARA acoustic output optimization method [6], [7]. It is a measure of the spatial coherence from the backscattered ultrasound signal received at neighboring transducer elements, and therefore channel data is required for this calculation. Long et al. [7] showed that the normalized spatial coherence of the data (modeled as the sum of signal (S) and uncorrelated noise (N)) from two transducer elements can be described as a function of their element separation (m) using the following formula:

$${\widehat{R}}_{S+N}[m]=\frac{\langle \left(S_i+N_i\right){\left(S_{i+m}+N_{i+m}\right)}^{*}\rangle }{\sqrt{\langle {\left|S_i+N_i\right|}^2\rangle \langle {\left|S_{i+m}+N_{i+m}\right|}^2\rangle }}$$ (1)

In this equation, i refers to the transducer element index. The normalized spatial coherence can also be written in terms of channel signal-to-noise ratio (SNR_c), or the ratio of the signal power to the noise power [7].

$${\widehat{R}}_{S+N}[m]=\{\begin{array}{ll}1\hfill & m=0\hfill \\ \frac{SN{R}_c}{1+SN{R}_c}{\widehat{R}}_S[m]\hfill & m\ne 0\hfill \end{array}$$ (2)

To calculate LOC, only neighboring transducer elements are included, so m = 1. LOC, or ${\widehat{R}}_{S+N}[1]$, reflects image quality degradation from reverberation, off-axis scattering, high-frequency aberration, and thermal noise [7], [16], [17], [18].

Unlike contrast [19], contrast-to-noise ratio (CNR) [20], [21], and generalized contrast-to-noise ratio (gCNR) [22], which are relative measures of image quality comparing two specified regions, LOC can be calculated from a single region. This makes LOC better suited to automatic ALARA MI calculations. Additionally, for the gold-standard method used in this study, requiring only one region was critical to characterizing the image quality throughout the image.

2). Identifying Regions:

Regions were classified based on analysis of the swept-MI data. This method established the ground truth for which regions work well for image optimization and which ones do not. Initially, the frames of image data were divided into non-overlapping 0.5 cm (axial) by 5-line (lateral) patches. This patch width corresponds to a 2.5° angular span. For a given patch location, the mean LOC was computed in each frame. These LOC values were plotted as a function of the frame’s MI. Then, a logistic curve fit was applied to the LOC-versus-MI data. This allowed the ALARA MI, the LOC achieved at the ALARA MI, and the R² of the fit to be calculated for that region. This process was repeated for each patch of an acquisition. Examples of these data sets are shown in Fig. 2.

Fig. 2.

Examples of LOC-versus-MI data from regions that are 0.5 cm (axial) by 5 lines (lateral). The ALARA MI, the LOC achieved at the ALARA MI, and the R² of the fit are shown for each example. Subplots a-f had good fits, while subplots g and h had R² values below the threshold of 0.95. Subplots a, c, and g had high LOC asymptotes, indicating high image quality.

As was done in previous work [6], [7], the fit was performed using a five-parameter logistic function [23]. The point at which the fit reached 98% of its maximum was considered the ALARA MI. If the curve was never below 98% of its maximum value, the minimum MI used in this study was selected (0.15). Otherwise, the ALARA MI was calculated from the increasing portion of the curve. For some patches, the fit did not increase or the logistic curve fit failed, indicating that it was not a viable region for optimization. Once these calculations were completed, patches were excluded if the LOC achieved was less than the 40th percentile for that image or the R² of the fit was less than 0.95. The ALARA MI values, the LOC reached at the ALARA MI, and the R² values are all displayed for the examples in Fig. 2. Subplots g and h would both be excluded by the R² threshold, and subplots b, f, and h all have low LOC values so, depending on the overall image, they may be excluded by the LOC filter. An example result for a full image is shown in Fig. 3. One B-mode frame is shown along with the ALARA MI values and the LOC values calculated for that acquisition. Excluded regions are displayed in black in the ALARA MI and LOC images.

Fig. 3.

A B-mode image, ALARA MI values, and the LOC values reached for one acquisition. The images were divided into 0.5 cm by 5-line regions to calculate the ALARA MI. As indicated on the B-mode image, this data set includes an anterior placenta and a cross-sectional view of the fetal abdomen. Patches displayed in black on the ALARA MI image were excluded because the logistic fit failed, the LOC achieved was less than the 40th percentile for that image, or the R² of the fit was less than 0.95.

3). Testing Data Filters:

Multiple statistics were considered as ways to prospectively filter the data for regions that would create reliable ALARA MI suggestions. These statistics include average pixel brightness, standard deviation of pixel brightness, the speckle SNR, and frame-to-frame correlation. For this analysis, the constant MI data (MI = 0.8) were used because they represent the data available during live scanning. For the brightness and SNR filters, the first frame of image data with a constant MI value was used. For the frame-to-frame correlation calculation, both frames of constant MI data were used.

First, the data sets were limited to 2 cm above and below the focal depth and a 30° span. Then, the frames were divided into patches of the same size as the swept MI data, 0.5 cm (axially) by 5 lines (laterally). The image statistics were calculated within each of these patches. For brightness, the mean value (μ) within each patch was calculated and both maximum and minimum thresholds were considered as a percentile of the brightness values of the cropped image. Similarly, a maximum percentile threshold for the standard deviation (σ) within the patch was considered to help exclude structure boundaries. Speckle SNR is the ratio of the mean to the standard deviation (μ/σ). For speckle SNR and the frame-to-frame correlation coefficient, threshold values were used. The frame-to-frame correlation coefficient was used as an indicator of motion between frames. Even subtle motion can change the structure within the region in consecutive frames, and more variable results would be expected in these non-uniform border regions. All calculations were performed on beamformed, envelope-detected data.

In considering different thresholds for each filter, the primary metric that was optimized was the positive predictive value (PPV). PPV is defined as the ratio of true positives to positive results (true positives + false negatives). This statistic indicates what percent of the patches suggested by the system were also considered reliable in the ALARA MI calculations from the swept MI frames. Additionally, the positivity rate was calculated to indicate what percent of the patches passed a set of filters. The sensitivity and specificity were also investigated as part of a Receiver Operating Characteristic (ROC) curves analysis. A summary of the filters and metrics used in this study is provided in Table I.

TABLE I.

Summary of filtering methods and analysis criteria used in this work.

Filtering Method	Description	Rationale
Brightness Percentile	Mean pixel value within the patch, considered relative to the other values in the cropped image area	Brighter regions are expected to correspond to tissue, rather than fluid regions
Speckle SNR	Mean brightness of patch divided by standard deviation	Low SNR regions may reflect areas on structure borders or where tissue is not present
Frame-to-frame Correlation	Normalized cross correlation between envelope detected data of sequentially collected frames	Motion and electronic noise would both be expected to decrease frame-to-frame correlation
Standard Deviation Percentile	Standard deviation within the patch, considered relative to the other values in the cropped image area	Border regions between tissues with substantially different echogenicities are likely to have high standard deviation values relative to the rest of the image
Analysis Method	Description	Rationale
Swept-MI Data Analysis	Full frames of data were collected at each MI value and evaluated offline to determine which patches of the image worked well for optimization. This is how “positive” and “negative” patches were defined.	This was the gold standard used in this study. It is computationally intensive, and the channel data needed is not accessible on most clinical scanners.
Positive Predictive Value (PPV)	The number of acceptable patches selected divided by the total number of patches selected	A high PPV suggests that the set of filters reliably selects satisfactory regions
Positivity rate	Percent of patches selected	To ensure that enough regions are selected
Sensitivity	True Positive Rate	Indicates the rate at which the filtering method selects regions that passed the swept-MI data analysis
Specificity	True Negative Rate	Indicates the rate at which the filtering method avoids regions that failed the swept-MI data analysis

Finally, the minimum brightness percentile and speckle SNR filters were also considered for other patch sizes. Four additional sizes were used in this analysis, with lateral patch sizes between 3 and 7 lines and axial patch sizes between 0.3 and 1.5 cm. The minimum brightness percentiles used were between 40 and 80 and the minimum speckle SNR values were between 1.0 and 1.7.

C. Clinical Testing

In addition to the offline analysis described above, preliminary clinical testing was done to evaluate the region-selection method. Region selection was preformed in real time to determine what MI to use for data collection. This clinical testing was also done at the Duke University Hospital Fetal Diagnostic Center and involved recruiting ten additional patients under the protocol approved by the Duke Health IRB. As described in a proceedings paper by Huber et al. [24], all subjects were scanned with a Vantage 256 ultrasound system (Verasonics, Kirkland, WA) and a C1-6-D transducer (GE, Chicago, IL). Pulse-inversion harmonic imaging was used with a transmit frequency of 2.36 MHz, an F/2.0 aperture, and focal depths that aligned with the structure position. For this study, images were collected that contained both the fetal liver and bladder, and the focal depths used were between 6 and 8 cm. Adaptive region selection was used to locate a 1.5 cm axial by 5-line lateral region for MI sweeping and ALARA MI determination for each acquisition. The adaptive ROI selection process used a minimum brightness percentile of 60 and a minimum speckle SNR of 1.6 to grade regions of the the summed, envelope-detected data. The mean and standard deviation of the brightness values were calculated for each pixel based on a 12.5 wavelength axial by 7 wavelength lateral kernel. Then, the focal depth region of the image was searched for the 1.5 cm by 5-line area that had the most pixels meeting one of the grading criteria. Pixels meeting both criteria were counted twice. In the highest-graded region identified by the ROI selection criteria, channel data were acquired at nine sequential MI values from 0.1 to 1.2. Then, the median LOC from the data at each of these MI settings was calculated and fit with a logistic function. The MI corresponding to 98% of the maximum LOC in the logistic fit was determined to be the ALARA MI. With the ALARA MI determined, frames were collected at both the transmit voltage setting that most closely matched the ALARA MI and at an MI of 1.2. An MI of 1.2 was selected for the comparison frame because it is the highest MI used in the sweep and above the expected range of ALARA MI values [6], so it can be considered a comparison to maximum image quality. This MI value is also well within the FDA limit for MI [1]. The entire automated ALARA MI update process, from selection of the ROI to collection of the 1.2 MI frame, took approximately one second. The ROI selection part of the processing took approximately 5 ms per update. For each subject, at least five repeat acquisitions of the fetal liver and bladder view were collected. The transducer was removed from the patient’s abdomen between each acquisition to collect independent views.

To evaluate the automated region selection and ALARA imaging method, the image quality at the ALARA MI setting was compared to the image quality of the 1.2 MI frame. The percent difference in contrast, CNR, and gCNR were calculated between these two frames. To calculate these metrics, ROIs were drawn on the B-mode images such that in both frames there was a hypoechoic target region over the amniotic fluid or bladder and an echogenic background region over the fetal abdomen. ROIs were limited to within 2 cm of the transmit focus and had surface areas of at least 0.50 cm². Nine data sets were excluded because they did not have a sufficiently large hypoechoic area. This left a total of forty-two data sets to analyze.

III. RESULTS

Each filter was analyzed independently for its patch classification ability. The receiver operating characteristic (ROC) curves for each filter are shown in Fig. 4. The area under the curve (AUC) for minimum brightness percentile, minimum frame-to-frame correlation, minimum speckle SNR, maximum standard deviation percentile, and maximum brightness percentile are 0.81, 0.66, 0.51, 0.23, and 0.19, respectively. Lowest brightness had the greatest AUC and is therefore the most promising filter to use alone. Notably, maximum standard deviation and maximum brightness have AUC values less than 0.5. This suggests that these filters classify patches worse than randomly assigning the labels. Maximum brightness percentile is the counterpart to minimum brightness percentile, so the low AUC of this metric is expected given that minimum brightness performed so well. The AUC from maximum brightness is equivalent to one minus the AUC from minimum brightness. The counterpart to maximum standard deviation would be minimum standard deviation. However, since the goal of the standard deviation filter was to remove border areas, including the inverse threshold type does not make much physical sense in this case. The prevalence of amniotic fluid in the images likely contributed to the concave maximum standard deviation percentile ROC curve. These fluid regions may have low standard deviation values, and therefore pass a maximum standard deviation filter, but this area is not of interest for optimization; in the gold-standard method, it would be filtered out by the LOC threshold. A maximum standard deviation filter used in conjunction with another filter, like minimum brightness, could potentially have the desired effect. Sensitivity describes what fraction of the regions that would be good for optimization, as defined by the gold-standard method, were selected, and specificity reports the fraction of the regions that did not work well for optimization that were excluded. However, since the ALARA method just needs a small area to optimize, the most important criterion for this automated region-selection method is that suggested regions are reliable.

Fig. 4.

ROC curves for each filter considered. The filters considered are minimum brightness, minimum frame-to-frame correlation, minimum speckle SNR, maximum standard deviation, and maximum highest brightness and the corresponding area under the curve (AUC) values are 0.81, 0.66, 0.51, 0.23, and 0.19.

Further analysis considered how the filters could be used together to increase the PPV. The data were evaluated using combinations of filters, with the specific parameter ranges informed by the results from using the filters individually. The range and increment values are shown in Table II, and the results of this analysis are shown in Table III. All combinations of the threshold values were included. The goal was to maximize the PPV, so that a suggested ROI would reliably indicate a location to perform the ALARA optimization. Restrictive filters can create high PPVs, even up to 100%, but with positive condition rates that are so low that they would not be helpful in realistic conditions. Therefore, the PPV was maximized with a restriction on the overall positive condition rate. In Table III, the best results are shown for positive condition rates of at least 15, 20, and 25%. An image with filter set B of Table III is shown in Fig. 5.

TABLE II.

Parameters evaluated for ROI selection, including ranges and increments.

Parameter	Range
Minimum Brightness Percentile	0, 40–80 (step = 2)
Minimum Speckle SNR	0, 1.00–1.70 (step = 0.05)
Minimum Frame-to-frame Correlation	−1, 0.2–0.6 (step = 0.1)
Maximum Standard Deviation Percentile	40–100 (step = 10)
Maximum Brightness Percentile	80–100 (step = 10)

TABLE III.

Results from assessing the combination of filters. The best filters are shown for multiple positive condition rates (15, 20, and 25% correspond to filter sets A-C). Neither the maximum brightness percentile nor the maximum standard deviation percentile contributed to any of the best-case options.

Filter Set	Positive Predictive Value	Positive Condition Rate	Minimum Brightness Percentile	Minimum Speckle SNR	Minimum Frame-to-frame Correlation	Maximum Standard Deviation Percentile	Maximum Brightness Percentile
A	98.59%	15.06%	64	1.45	0.5	100 (off)	100 (off)
B	98.10%	20.76%	46	1.40	0.6	100 (off)	100 (off)
C	97.37%	25.39%	44	1.45	0.4	100 (off)	100 (off)

Fig. 5.

An image of a fetal abdomen with outlines showing the result of each filter and the final segmentation. The ALARA MI image is shown for comparison. For all subplots, the lateral dimension is displayed in degrees and the axial dimension is displayed in centimeters.

The positive condition rates shown in Table III were calculated as global averages across all 150 images tested (twenty-five subjects, three fetal abdomen images and three placenta images each). For some filters, the number of ROIs selected is consistent across different images and subjects. For example, using a minimum brightness of the 60th percentile will always select 40% of the possible patches. However, when we use a threshold value, like the speckle SNR filter, the rate of return is variable. This effect is visible in the positive condition rate when it is disaggregated and shown for the individual acquisitions as in Fig. 6. The standard deviation of the positive condition rates for filter sets A, B, and C are 9.88, 15.14, and 13.59%. Compared to the best filter for the 15% positive condition rate limit, the best filter for the 20% positive condition rate limit has a stronger correlation filter and weaker brightness filter, so even though there is an overall increase in the number of positive patches between conditions A and B, that is not seen for every acquisition. For example, some of the placenta data sets from Subject 20 see a decrease in their positivity rate between A and B. Furthermore, some images do not have any patches that pass all of the filters. That occurs in 10, 12, and 5% of images for filter sets A, B, and C, respectively. However, it is not necessary that all images have patches suggested for optimization for this system to be effective. If no patches are recommended, the ALARA MI calculation can wait for another frame.

Fig. 6.

Positivity rate for each data set included in the analysis, separated by subject. Filter sets A, B, and C correspond to the parameters shown in Table III. The average positivity rates were 15%, 20%, and 25%, respectively. The placenta images generally had more passed regions than the more heterogeneous fetal abdomen images.

None of the optimal combinations shown in Table III recommended using a maximum brightness threshold or maximum standard deviation percentile, so both will be removed from consideration going forward. Additionally, the 20% positivity rate (B in Table III) will be the focus of the rest of the analysis. For that set of filters, the PPV of 98.10% corresponds to 52 “false positive” regions, or regions that passed the automated ROI selection filters, but were not considered good regions for optimization using the full swept-MI data set. In this analysis, 54% of these regions were excluded based on having low LOC values, 85% were excluded based on having low R² for the logistic fit, and 38% were excluded by both. For the ten images that had both true positive and false positive regions, the mean difference in the average ALARA MI values suggested by the true positive regions and false positive regions was 0.11 and the median difference was 0.00. These results suggest that the occasional false positive region would have a weak effect on the image quality and safety of the scan.

In Table IV, the component filters are added one at a time to show their cumulative effect on the PPV and positivity rate. Each additional filter increases the PPV and decreases the positivity rate. The majority of the effect comes from the minimum brightness percentile threshold. The greatest PPV using only the minimum brightness is from setting the threshold to the 66th percentile. It produces a PPV of 92.93% and a positivity rate of 34.09%. The maximum PPV achieved by each pair of filters is shown in Table V. The positivity rate was again restricted to above 20%.

TABLE IV.

The cumulative effect of adding additional filters. Each component is added individually to show the contribution to the positive predictive value and the positivity rate.

Positive Predictive Value	Positive Condition Rate	Minimum Brightness Percentile	Minimum Speckle SNR	Minimum Frame-to-frame Correlation
91.86%	54.55%	46	-	-
95.57%	33.55%	46	1.40	-
98.10%	20.76%	46	1.40	0.6

TABLE V.

The greatest positive predictive value for each pair of filters with a positivity rate of at least 20%.

Description	Positive Predictive Value	Positive Condition Rate	Minimum Brightness Percentile	Minimum Speckle SNR	Minimum Frame-to-frame Correlation
Min. Brightness & Speckle SNR	97.05%	20.26%	54	1.60	-
Min. Brightness & Correlation	93.80%	26.52%	62	-	0.60
Speckle SNR & Correlation	89.51%	23.11%	-	1.60	0.60

From this analysis, minimum brightness percentile and minimum speckle SNR appear to be the most impactful filters. Fig. 7 shows the positivity rate and PPV as a function of minimum brightness and speckle SNR thresholds. As expected, positivity rate decreases with increasing thresholds. PPV has a more complicated response, but the greatest PPV values are near the most restrictive thresholds. When using only minimum brightness and minimum speckle SNR filters, the maximum PPV (98.64%) corresponds to a positivity rate of approximately 9%.

Fig. 7.

The relationship between the thresholds set for filters (minimum brightness percentile and minimum speckle SNR) and the positivity rate and positive predictive value.

Thus far, individual parameter sets have been selected that maximize the PPV under a given set of constraints. However, the PPV values achieved for a range of minimum brightness, speckle SNR, and correlation values show relatively small amounts of variation. In Fig. 8, different combinations of these parameters that achieve PPV and positivity rate thresholds are displayed. These results suggest that there is some flexibility in the specific threshold values when identifying regions for ALARA optimization.

Fig. 8.

The possible minimum brightness, minimum speckle SNR, and minimum correlation values that could be used to achieve desired PPV and positivity rate values. Results are shown for a PPV threshold of 95% and a positivity rate of 20% (top row) and a PPV threshold of 97% and a positivity rate of 15% (bottom row).

Using the original patch size of 0.5 cm by 5 lines, two of the best cases were tested on the remaining ten subjects. Condition B (minimum brightness = 46th percentile, minimum speckle SNR = 1.40, minimum frame-to-frame correlation coefficient = 0.6) produced a PPV of 94% and a positivity rate of 17%. No patches passed in 8 of the 60 images. With the best-case filter set using only speckle SNR and minimum brightness (1.6 and 54, respectively), the PPV was 97% with a 20% positivity rate. All images had at least one patch that passed this filter. Using only speckle SNR and minimum brightness filters showed strong performance and these metrics have the advantage of requiring just one frame of data.

Another important variable to consider is patch size. Four additional sizes were considered for the original 25 subjects and the results are shown in Table VI. Some options have shifts smaller than the patch size and therefore represent overlapping patches. Minimum brightness percentiles between 40 and 80 (increment of 5) and minimum speckle SNR values between 1.0 and 1.7 (increment of 0.1) were used. The original non-overlapping 0.5 cm by 5-line patch size is also included in this analysis for comparison. Larger patches were associated with greater PPVs, which could suggest more accurate estimates are produced when larger areas are used. An example image is shown in Fig. 9 with the ALARA MI image shown for each of the tested patch sizes. These patch sizes apply to this specific imaging configuration, but to make the results more generalizable, they can also be reported in terms of resolution cells. The lateral resolution of this system, λz/D, is 0.65 mm. The axial resolution, nλ/2, is also 0.65 mm. For these calculations, λ is the wavelength, z is the focal depth, D is the aperture size, and n is the number of cycles. For a focal depth of 6 cm, the default patch size of 5 lines laterally and 0.5 cm axially would correspond to 7.3 by 7.7 resolution cells at the focus.

TABLE VI.

The greatest positive predictive value with a positivity rate of at least 20% for each patch size.

Axial Size	Axial Shift	Lateral Size	Lateral Shift	Minimum Brightness Percentile	Minimum Speckle SNR	Positive Condition Rate	Positive Predictive Value
0.3 cm	0.3 cm	3 lines	3 lines	65	1.60	20.17%	94.06%
0.5 cm	0.5 cm	5 lines	5 lines	55	1.60	20.26%	97.05%
0.5 cm	0.1 cm	5 lines	1 line	60	1.50	21.58%	97.16%
1.0 cm	0.5 cm	7 lines	3 lines	55	1.40	20.82%	98.35%
1.5 cm	0.5 cm	5 lines	1 line	60	1.20	22.44%	98.53%

Fig. 9.

A B-mode image with corresponding ALARA MI images of varying patch sizes. These images show an anterior placenta and a fetal abdomen. The patch sizes are shown in the titles, and the shift amounts are shown on the axes.

In the preliminary clinical testing of this method, a minimum brightness percentile of 60 and a minimum speckle SNR of 1.6 was used for ROI selection. An example image from this study with the selected region outlined is shown in Fig. 10. Once a 1.5 cm by 5-line ROI was selected, swept-MI data were collected, the ALARA MI was determined, and frames of channel data were collected at the ALARA MI and an MI of 1.2. To evaluate image quality, the contrast, CNR, and gCNR of these two frames were compared. The results shown in Fig. 11 demonstrate that there is strong correlation (r = 0.96, 0.82, and 0.94 for contrast, CNR, and gCNR, respectively) between the image quality of the frame collected at an MI of 1.2 and the image quality of the frame collected at the ALARA MI. The ALARA MI values of these data sets were 0.50 +/− 0.067. On average, the ALARA MI frame showed a 5.4% decrease in contrast, a 3.4% decrease in CNR, and a 0.85% decrease in gCNR compared to the 1.2 MI frame. The standard deviations for the percent difference in image quality were 9.4% for contrast, 9.3% for CNR, and 4.6% for gCNR. This small difference in image quality is associated with a substantial decrease in acoustic exposure when using the ALARA MI value rather than the 1.2 MI setting. These results suggest that this ROI-selection method can be used with an LOC-based ALARA MI calculation to automatically implement the ALARA principle for obstetric ultrasound imaging. In the future, a full reader study should be performed to determine if the image quality at the ALARA MI is sufficient for specific clinical tasks.

Fig. 10.

An example image from preliminary clinical testing of the automated ROI-selection method. The fetal bladder is labeled in this image and the selected region is outlined in red.

Fig. 11.

Comparison between the image quality of the frame collected at an MI of 1.2 and the image quality of the frame collected at the ALARA MI. Plots are shown for three image quality metrics: contrast, CNR, and gCNR. The unity line is shown for reference, and for points that lie along this line, the image quality metrics from the two frames are equivalent, which indicates that the system would reduce exposure without affecting image quality. Points above the line show that the image quality of the 1.2 MI frame was higher, and points below the line show that the image quality of the ALARA MI frame was higher. The image quality values from the pairs of frames demonstrate good agreement, suggesting that near-maximum image quality is achieved by the ALARA MI setting.

IV. DISCUSSION

As shown in Fig. 1, this automated ROI selection system could seamlessly fit in with adaptive ALARA imaging to allow the whole process to happen behind the scenes. The ROI selection process can be performed in about 5 milliseconds, and since the MI sweep will be limited to the identified region, five lines of pulse-inversion data could theoretically be collected at sixteen MI settings in about 30 milliseconds based on acoustic propagation time. The next step, calculating the LOC, could also be done quickly. Prior work by Hyun et al. and Bottenus et al. has shown that coherence calculations can take place in real time [8], [25]. The specific speed of the implementation would depend on the computing resources available, but it would be expected to take tenths of a second. Furthermore, this processing could be done in parallel while continuing to image the patient.

For the training data, the maximum PPV with a 20% positivity rate was achieved using minimum brightness percentile, minimum speckle SNR, and minimum frame-to-frame correlation coefficient thresholds, but strong results were also achieved using only the minimum brightness and minimum speckle SNR filters. When these filters were applied to the ten patient test data set, a PPV of 94% was achieved with minimum brightness percentile, minimum speckle SNR, and minimum frame-to-frame correlation coefficient and a PPV of 97% when only using minimum brightness percentile and minimum speckle SNR. The latter option would require just a single frame of standard B-mode data. The optimal threshold values to use with the filters depend on the positivity rate limit that is set because restrictive filters can have high precision but produce few recommended regions. Furthermore, the thresholds used to identify regions that worked well for optimization (greater than 0.95 R² and the 40th percentile of LOC values) were empirically determined and comparing to a different standard would likely change the result.

This ROI-selection method relies on having high enough image quality in the frame to have meaningful differences in the assessed image characteristics. The frames used in this study (MI = 0.8) consistently met this requirement, but at low enough output settings, the image could be overwhelmed by electronic noise and the expected brightness differentiation between tissues would be obscured. To avoid this issue, a minimum MI could be imposed. Further study would be needed to identify this value.

The main limitation of this study is the amount of time associated with collecting the frames of channel data. It was necessary for the analysis to have many frames at different MI values, but it does not approximate the envisioned clinical environment for this system well. In this study, the constant MI frames were compared to results from more than one second of swept MI data. Motion and structure position during that swept MI duration would not necessarily be captured by the two subsequent frames, though having the full data set allows us to evaluate a wide range of ROI positions.

Another limitation is that implementing this ALARA imaging method would require channel data, which is not easily accessible on clinical scanners. The ROI-selection portion has the advantage of using readily available envelope-detected data, but the LOC calculations rely on channel data. Higher-end scanners are more likely to have access to the channel data stage of the processing pipeline, so implementation would be more straightforward on these systems.

This analysis was done on polar data rather than scan converted data, so that patches along a vertical column would be produced from the same transmit events. In implementing this method, any suggested patches in a column could be used together with a negligible increase in calculation time. As was done in the clinical testing, the median LOC value from several suggested patches could be used in the ALARA MI calculation. Future work will explore the optimal size for a region of interest as well as evaluate the LOC-versus-MI data produced from suggested regions in vivo with a smaller temporal separation.

Preliminary clinical testing using minimum brightness percentile and minimum speckle SNR filters for ROI selection showed good agreement in image quality between frames collected at the ALARA MI and frames collected at an MI of 1.2. In this set of data, the ALARA MI frames had much lower MI values with only slightly lower quantitative image quality measures. The CNR values of the ALARA MI frames were 3.4% lower on average. In the future, reader studies will be needed to determine if this difference in image quality is consequential. If needed for certain clinical tasks, the sonographer could still increase the acoustic output level following automatic adjustment. Additionally, clinical evaluation would be needed to assess user acceptance of the automated ALARA tool.

Ultrasound safety guidelines consistently recommend observing the ALARA principle. Particularly in obstetric imaging, if additional exposure is not producing any benefit, it should be avoided. However, studies show that the displayed output indices are not often referenced by ultrasound users [3]–[5]. By automating acoustic output adjustment to follow the ALARA principle, both patient safety and image quality could be consistently prioritized. The proposed region-selection method is based on standard B-mode images and could be incorporated into existing platforms to facilitate automated ALARA imaging. This would allow ultrasound scanners to use patient and view-specific output settings and reduce their reliance on sonographers’ setting adjustments.

This method also has applications outside of obstetric ultrasound. For instance, acoustic exposure is of particular concern for ophthalmic imaging, and this use has lower guideline acoustic output values than other types of imaging, including fetal [1], [26]. Ultrasound use has also shown effects on neural activity, so in cases where neuromodulation is not the intended outcome, automated ALARA could be used to minimize excess acoustic exposure to the central and peripheral nervous system [27]. Once this system is implemented on an ultrasound scanner, it would likely translate to other application modes. Since the FDA regulations advise following the ALARA principle for all diagnostic ultrasound [1], with further study, this tool could be used widely. Depending on the underlying range of echogenicities, the ROI selection filter settings could be tuned for specific uses. Additionally, though the rapid ROI selection method shown in this work was applied to the task of automated ALARA imaging, it could potentially have other applications, such as quantitative ultrasound, SWEI, and automated frequency selection.

V. CONCLUSION

These results support an automated ROI selection framework for LOC-based ALARA imaging. This coarse segmentation method uses only location (proximity to the focus and center of the aperture) and local statistics of envelope-detected data that are readily available on clinical ultrasound machines to serve as an indicator for where to perform ALARA optimization. Threshold values were suggested based on offline analysis of clinical data, and real-time clinical testing of the method showed that the ALARA imaging method with automated ROI selection produced substantially decreased exposure with only a small decrease in image quality.

Since calculating an ALARA MI requires only a few lines at the suggested location at each MI, the brief pause in live imaging to collect the data would be imperceptible, and the process would not interfere with the clinical ultrasound exam workflow. Past work [6] has shown that there is variability in the ALARA MI value for a given patient across imaging windows, so frequent optimization could be beneficial. The addition of this ROI-selection method would allow ALARA acoustic output updates to happen frequently and without any need for user input.

Biographies

Katelyn M. Flint (Graduate Student Member, IEEE) graduated from Vanderbilt University, Nashville, TN, USA, in 2015 with a B.E. degree in biomedical engineering. She received the M.S. degree in biomedical engineering from Duke University, Durham, NC, USA in 2018 and is currently pursuing a Ph.D. in biomedical engineering at Duke University. Her research interests include ultrasound safety, fetal imaging, and automatic optimization of ultrasound settings.

Emily C. Barré graduated from Duke University, Durham, NC, USA, with a B.S.E. degree in biomedical engineering in 2021. She is currently pursuing an M.P.H. degree at The Dartmouth Institute, Hanover, NH, USA. She is planning to matriculate to medical school in summer 2022. Her research interests include clinical imaging applications, point-of-care ultrasound, and fetal imaging.

Matthew T. Huber (Graduate Student Member, IEEE) graduated with a B.S. degree in physics in 2018 from Rhodes College, Memphis, TN, USA and a M.S. degree in biomedical engineering in 2021 from Duke University, Durham, NC, USA. He is currently a Ph.D. candidate in biomedical engineering at Duke University. His research interests focus on implementing adaptive transmit parameter adjustment in ultrasound B-mode and Doppler imaging.

Patricia J. McNally received the B.S. in Diagnostic Medical Sonography from Rochester Institute of Technology, Rochester, NY, USA in 1990. She started her career as a Sonographer in Radiology at Duke University Hospital in 1990. She currently is the Supervisor of the Duke Fetal Diagnostic Center. Her role is to manage all operational activities (staffing/personnel, protocols, policies, budget/finance) for the Fetal Diagnostic Center staff, as well as facilitate research with our Duke Biomedical Engineering partners.

Sarah C. Ellestad graduated from SUNY at Geneseo with a B.A. degree in both Psychology and Biology in 1994. After spending a year abroad working at a National Children’s Home in England, she matriculated into medical school at SUNY Health Science Center at Syracuse, graduating with her medical degree in 1999. She completed her residency in Obstetrics and Gynecology at Baystate Medical Center in 2003 and her fellowship in Maternal-Fetal Medicine at Duke University Medical Center in 2006. She is currently an Associate Professor in the Division of Maternal-Fetal Medicine at Duke University and is the Maternal-Fetal Medicine Fellowship Program Director. Her interests include fetal therapy and prenatal diagnosis.

Gregg E. Trahey (Fellow, IEEE) received the B.G.S. and M.S. degrees from the University of Michigan, Ann Arbor, MI, USA, in 1975 and 1979, respectively, and the Ph.D. degree in biomedical engineering from Duke University, Durham, NC, USA, in 1985. He served in the Peace Corps from 1975 to 1978 and was a project engineer at the Emergency Care Research Institute in Plymouth Meeting, PA, USA, from 1980 to 1982. He currently is a Professor with the Department of Biomedical Engineering at Duke University and holds a secondary appointment with the Department of Radiology at the Duke University Medical Center. His current research interests include adaptive beamforming and acoustic radiation force imaging methods.