Optimizing B-lines on Lung Ultrasound – an In-vitro to In-vivo Pilot Study with Clinical Implications

Abstract

Purpose:

B-lines on lung ultrasound (US) are the hallmark of pulmonary edema. It is unknown if ultrasound machine settings or probe type matter.

Methods:

We created an in-vitro gelatin model. Using lung presets as baseline, five blinded investigators assessed the impact of 32 distinct settings on B-line visibility based on a Likert-Scale (LS) from 0–10 (<5 worse, >5 better) separately for two probes. The experiment was then repeated in-vivo in a patient with known pulmonary edema.

Results:

Based on a multivariable regression LS-ratings were similar when comparing the in-vitro vs. in-vivo experiment (P=0.16; partial R²=0.2%) and when using the curvilinear vs. linear probe (P=0.69; partial R²=0.02%) but significantly different across machine settings (P<0.0001; partial R²=34.4%).

Conclusions:

Limited by its pilot character, our study suggests that 1.) certain US-machine settings heavily impact B-line visibility, with no clear difference between probes; 2.) in-vitro models are a valid and practical alternative to more challenging patient-based research; 3.) there is significant potential to improve B-line visibility and thus diagnostic yield in the clinical setting by using lung presets, centering the focal zone at the pleural line and increasing the distal time gain compensation, most of which are (in our experience) rarely done.

BACKGROUND

The high air content in normal lungs generally prevents penetration of ultrasound waves beyond the pleura. Thus, lung ultrasound (US) is largely dependent on artifacts. Clinically, one of the most important artifacts are B-lines, which are the hallmark of “alveolar interstitial syndrome” comprising pulmonary edema, pneumonia and pulmonary fibrosis.[1,2] B-lines are defined as vertical hyperechoic structures arising from the pleural line, with a discrete laser-like appearance, extending all the way to the bottom of the screen without fading (>17–18cm) and moving synchronously with lung sliding.[1,2] Clinically, B-lines are particularly useful to diagnose and monitor pulmonary edema, as they have similar diagnostic accuracy as BNP[3,4], experimentally appear even before the occurrence of hypoxemia[5], correlate with the amount of extravascular lung water (EVLW)[6], and rapidly disappear as EVLW decreases under therapy[7,8]. To be clinically useful, the evaluation for B-lines needs to be reproducible across different examiners (i.e. they should count the same number of B-lines) and several studies have evaluated how to increase inter-rater agreement by using different counting techniques[9], assessing different anatomic locations[10] or with classification systems which define a positive result[10,11].

From the physical receipt of the reflected ultrasound waves until the appearance of an image on the screen, the US signal undergoes significant processing and filtering (Figure 1)[12], most of which has been developed to improve image quality of the more commonly examined solid organs. While several authors[9,13–15] acknowledge the likely importance of ultrasound machine settings for the ability to detect artifacts, their impact on B-lines has not been systematically studied. Similarly, while in clinical practice the curvilinear (“abdominal”) probe is probably used most often, and except for one small study showing no difference between a curvilinear vs. cardiac probe,[8] the impact of probe type on B-lines has not been formally studied. In fact, many studies do not specify the ultrasound probe used, and even less provide any comments about the applied ultrasound settings.

FIGURE 1. Processing and Filtering of the Ultrasound Signal on its Path from Probe to Screen.

Created based on Ali et al. 2008[12] and Powers et al. 2011[20].

We hypothesized that ultrasound machine settings and probe type (linear vs. curvilinear) do affect visibility of B-lines. We performed a pilot study testing this hypothesis by creating an in-vitro model followed by in-vivo validation.

METHODS

Phantom and Study Design

A Philips Sparq® was selected for all experiments. Even though some of the settings are trademarked terms for proprietary processing, they are typical of generic processing found on competitive modern systems (e.g. SonoCT is spatial compounding).[16] The system allows users to make various setting changes, some of which are binary (e.g. SonoCT on vs off), categorical (e.g. Smoothing 1, 2, 3 or 4 frames) or near-continuous in nature (e.g. location of focal zone from most proximally to most distally in approximately 0.5cm increments). Users can also switch between different “presets” which change several of the aforementioned settings at once, but apparently also affect “hidden” settings that cannot be changed by the user (see discussion). By collapsing continuous variables into three to five approximately equidistant categories we identified a total of 30 distinct machine settings that were evaluated separately for the curvilinear and the linear probe after selecting “lung” as the underlying presets (see e-Table E1 for details). In addition, we tested two variants of the “abdominal” presets using only the curvilinear probe. In the following “machine settings” refers to these 32 distinct machine states.

Inspired by two studies trying to elucidate the physical origin of the B-line artifact by using different configurations of air bubbles in liquid media,[17,18] we created an in-vitro model using an inverted plastic cup submerged in a 3.5%-gelatine/0.1%-soap solution and then filled a single layer of air bubbles followed by a partial second layer to create scattered air bubble tetrahedrons (air bubbles were instilled by applying steady pressure to an air-filled syringe connected to a small tube ending under the plastic cup resulting in air bubbles with a diameter of approximately 3mm)[17]. Small 1-cent coins on the side of the plastic cup served as weights preventing the uplift of the air-filled cups while the model hardened over a few hours in a refrigerator set at 40°F (4°C). To simulate lung sliding, the hardened model was then placed on a rocker table set at 1.5rpm with the ultrasound probe fixed above (Figure 2).

FIGURE 2. Model Creation and Experimental Set-up.

Air-bubbles were instilled under a plastic cup submerged in a gelatin-soap solution (A) in order to create a single followed by a partial second layer (B); the hardened model was then placed on a rocking table (C) with the probe fixed above (D).

For each type of probe, a 6-second video was recorded using lung presets as baseline (with a few select manual setting changes to optimize the baseline comparator image: gain at 80; focal zone centered at the air bubble layer/pleura [by default focal zone is set at mid-screen; thus for our baseline the focal zone was set at ~1.5cm in-vitro vs ~3cm in-vivo, respectively]; depth set at 18cm for the curvilinear - and 10cm for the linear probe), followed by 6-second videos each showing the result of univariably changing one of the 32 above described machine settings at a time.

An online survey was created (SurveyMonkey Inc., San Mateo, CA, USA) displaying the baseline video side by side with a synchronized video reflecting a setting change, allowing each investigator (two critical-care board certified physicians [FS, RD] and three internal medicine residents [CS, AM, BS]) to independently rate the latter on a 10-point Likert Scale (LS; 0 much worse, 5 same, 10 much better). Each survey started with a test screen to help calibrate screen settings to ensure consistency. Videos were devoid of any information regarding the applied setting to ensure blinded review (e-Appendix E1).

To see if findings from our in-vitro model could be translated into clinical practice (in-vivo) we repeated the above described process using a patient with B-lines from known pulmonary edema with the probe fixed in the longitudinal orientation over the right 3^rd intercostal space in the midclavicular line, i.e. “zone 1” of the BLUE protocol[14]). The combined (in-vivo+in-vitro) data served as the base for below described statistical analysis.

The study was approved by the Boston University Medical Center Institutional Review Board (H-34443). The examined patient provided written informed consent prior to participation.

Analysis

We used linear regression as an omnibus test (F-test) to assess the impact of experimental phase (in-vitro vs. in-vivo), probe type (curvilinear vs. linear) and machine settings on LS-ratings while adjusting for potential confounding by the different raters. To assess what percentage of the total variance is explained by each variable we further assessed the partial R-squares (pR²) for each variable. Because this analysis showed significantly different LS-ratings based on machine settings but comparable results for the two experimental phases and probe types, we then created a forest plot showing the mean LS-rating with 95%-confidence interval for each machine setting based on the entire dataset. Settings were judged to be significantly worse or better than the baseline setting if the upper or lower confidence interval excluded the null (5.0), respectively. Post hoc the data were further explored using forest plots after stratifying by experimental phase and probe type.

Regression analysis was performed in JMP 10.0.0 (SAS Institute Inc., Cary, NC, USA) using two-sided P-values <0.05 to judge statistical significance. Forest plots were created in Microsoft Excel Spreadsheets (2008 for Mac; Version 12.3.6).

RESULTS

For the main analysis, five investigators each rated 123 videos in comparison to the corresponding baseline video (615 videos in total). Based on the multivariable linear regression model LS-ratings were similar when comparing the in-vitro vs. in-vivo experiment (P=0.16; pR² = 0.2%) and the curvilinear vs. linear probe (P=0.69; pR² = 0.02%), but were significantly different across machine settings (P<0.0001; pR² = 34.4%) and reviewers (P<0.0001; pR² =6.7%). Figure 3 provides a visual example.

FIGURE 3. Examples of the Impact of Machine Setting Changes on B-line Visibility.

“Baseline” denotes lung presets with all machine settings left at their default mode, except for manual changing gain to 80, centering the focal zone at the layer of the air bubbles/pleura (by default focal zone is set at mid-screen; thus for our baseline the focal zone was set at 1.5cm in-vitro vs 3cm in-vivo, respectively), and setting depth at 18cm for the curvilinear - and 10cm for the linear probe; “Focal Zone at 2/3” denotes that focal zone was set at approximately 2/3 depth (i.e. ~12cm for the curvilinear – and ~6.5cm for the linear probe) with all other settings unchanged from “Baseline”. More TGC” denotes that the time gain compensation was increased in the mid- and even more so in the far field (vs no difference across the entire field by default) with all other settings unchanged from “Baseline”. Note that the default frequency for the curvilinear (4MHz) and linear (8MHz) probe was held constant across the shown setting changes.

As compared to our baseline settings, 41% (13/32) of machine setting changes significantly worsened B-line visibility whereas only 9% (3/32) of changes improved it (Figure 4). However, note that our baseline included manual centering of the focal zone at the air-bubble layer/pleura which thus reflects a change that significantly improves B-line visibility (e-Figure E1). Some of the continuously modifiable settings showed visual trends (Figure 4): ratings progressively worsened with increased distance of the focal zone from the air-fluid interface; the more focal zones were added; and the more the dynamic range was lowered.

FIGURE 4. Forest Plot showing mean rating + 95% Confidence Interval for each Machine Setting (based on combined in-vivo and in-vitro data).

§low/default/high frequency is 2/4/6 and 4/8/12MHz for curvilinear and linear probe, respectively. Abbreviations: TGC time gain compensation. For each setting n=20, except “More TGC distally” (n=15; erroneously not recorded during in-vitro experiment using the linear probe) and abdomen presets (n=10; evaluated with curvilinear probe only).

Exploratory subgroup analysis confirmed overall similar LS-ratings when stratifying by probe type (e-Figure E2) or in-vitro vs. in-vivo experiment (e-Figure E3).

There were signs of good construct validity and intra-rater reliability; there was statistically significant inter-rater variability but clinically the effect size was small and main analysis adjusted for inter-rater variability (e-Table E2).

DISCUSSION

This pilot study had three main results: First, certain ultrasound machine settings heavily impact B-line visibility. Second, B-line visibility is not significantly affected by probe type (curvilinear vs. linear). Third, findings in the proposed in-vitro model closely mimicked the findings in the one in-vivo patient studied.

A recent study using a static sponge-model reported somewhat discrepant results for a few tested machine settings (e.g. smoother B-lines using harmonics), but assessments were limited by mere qualitative, unblinded, assessment without in-vivo validation.[19] The impact of machine settings on B-lines that we observed in our study is largely biophysically plausible (detailed discussion in e-Appendix E2):[20] B-lines can experimentally be produced by air-bubble tetrahedrons,[18,17] likely due to reverberation or resonance of US-waves in the water column enclosed by air. One of the largest effects we found was that appropriate placement of the focal zone on the origin of the artifact (i.e. bubble-layer in-vitro vs pleura in-vivo) greatly increased its visibility, which seems logical as the focal zone is the part of the ultrasound beam with the highest energy density. Conversely, settings that shift ultrasound energy towards the mid/far field were generally associated with worse LS-ratings: e.g. harmonics (which relatively increases mid-field signals) was associated with lower LS-ratings. On the other hand, time gain compensation (TGC) which amplifies delayed US waves, led to significantly improved B-line visibility. This is of direct clinical relevance as the definition of B-lines includes the requirement that the phenomenon extends to the end of the screen, but in our study, we saw that simply changing the mid-to-distal TGC can determine whether a B-line meets this criterion or not (Figure 3 baseline vs “more TGC”). Also clinically interesting to note is that “autoscan”, a proprietary image-optimization algorithm, failed to improve B-line visibility. We further noted that lung presets resulted in a significant better B-line visibility than the abdomen presets, but pre-sets may vary substantially across different machines, thus, since we evaluated only one device, it is difficult to draw firm conclusions in this regard.

An interesting incidental observation was that abdomen vs lung presets resulted in a significantly worse image (even after adjusting all settings which can be manipulated by the ultrasound user to the same level as the baseline lung presets), revealing the presence of “hidden” settings which cannot readily be modified by users, but do substantially impact image quality.

The second main finding of this study was that quantitatively there was no clear difference in B-line visibility between the curvilinear vs. linear probe, with probe type explaining less than 1% of the total variation among LS-ratings. Qualitatively images from the linear probe did appear somewhat superior to those from the curvilinear probe in-vitro but not in-vivo (Figure 3). Although not directly comparable to our study due to the use of different frequencies, Mallamaci et al. also showed good agreement between a linear “cardiac” (3.0MHz) and curvilinear “renal” probe (3.5MHz).[8] Further, the first description of B-lines by Dr. Lichtenstein was done using cardiac probes of 3.0–3.5MHz with “2.5-, 5-, and 7.5-MHz probes [being] equally effective”, [1] a frequency range essentially comprising both our linear (8MHz) and curvilinear (6MHz) probe. These days Dr. Lichtenstein advocates for the use of a 5MHz microconvex probe[14], but it is rarely available in clinical practice and so far no formal comparisons have been performed.

Third, except for minor individual differences (e-Figure E3) findings in the in-vitro model were highly concordant with in-vivo findings (with experimental phase explaining less than 1% of the total variation among LS-ratings). Given the many challenges and costs of performing in-vivo research on patients, our model thus may provide a great alternative for future research on B-lines. We suspect that future studies exploring modifications such as changes in the gelatin/soap concentration or bubble size may further improve the generalizability of in-vitro-based B-line research.

The strengths of our study include that raters were blinded, minimizing the risk of differential measurement bias; good intra-rater reliability, signs of good construct validity, relatively narrow confidence intervals, robustness of results in subgroup analyses, and physical plausibility of findings increasing internal validity; and the validation of in-vitro findings in a patient from the target population thus increasing external validity.

This was an experimental proof-of-concept pilot study to test our hypotheses that ultrasound machine settings matter and that model-based research on B-lines is both feasible and potentially translatable to bed-side patient care. While we are encouraged by the relative clear “signal”, one of the study’s main limitations is that results are based on the evaluation of only one model and validation in only one patient. Furthermore, we assessed only settings on one ultrasound machine (chosen due to its availability and use in our clinical practice) which limits generalizability to other machines, although many of the settings are common to ultrasound machines in general (e-Table E1). Moreover, there was significant inter-rater variability; compared to machines settings it explained only a relatively small amount of the observed variability in B-line visibility ratings (34% vs 7%) and we adjusted for it in regression analyses, but discrepancies between ultrasound operators may limit the generalizability of our findings. Another issue is that we performed multiple comparisons in this study. Given the pilot nature of this study we decided against adjustment of P-values or confidence intervals and instead performed global omnibus tests first before delving into individual comparisons. However, despite that one would still expect 1/20 or in our case a total of approximately 5 false positive results (i.e. confidence intervals to spuriously exclude the null [LS of 5]), which may explain some of the variation in the exploratory Figures E1 and E2 which were based on even smaller sample sizes than for instance Figure 4. On the other hand, given the pilot nature and small sample size of this study we also ignored the paired nature of some of the data which may have led in part to overly wide and thus conservative confidence intervals.

Future research should further validate our approach of using in-vitro data for in-vivo inferences by using more models and patients, test different machines and probes, and assess setting changes in a multivariable fashion (i.e. changing several settings at the same time).

CONCLUSION

To our knowledge this experimental pilot study is the first study showing in-vivo that ultrasound machine settings can heavily impact B-line visibility and presents an in-vitro model which closely mimicked the findings in the one in-vivo patient studied. Whether B-lines are assessed manually or detected automatically via machine algorithms[21,13], optimized machine settings resulting in better B-line visibility (thus increasing the signal-to-noise-ratio) should ultimately translate in better diagnostic yield and reproducibility. Limited by small sample size and evaluation of a single ultrasound machine, our results suggest that in clinical practice B-line visibility may be improved by centering the focal zone at the pleural line, increasing distal TGC, minimizing the number of focal zones and removing settings designed to attenuate artefacts.

Abbreviations:

BNP

Brain Natriuretic Peptide

LS

Likert Scale

pR²

partial R-square

TGC

time gain compensation

US

ultrasound