Automatic B-lines quantification with lung ultrasound in mechanically ventilated patients

Abstract

Background

If ultrasound dedicated software for the identification and quantification of interstitial syndrome is equal or superior to experts observation is still being explored. The primary objective of this study was to compare the performance of an automatic method based on a dedicated software algorithm with that of expert clinicians in terms of B-lines quantification.

Methods

Adult patients with acute respiratory failure who were deeply sedated, paralysed and mechanically ventilated were prospectively enrolled within 48 hours from ICU admission and underwent lung ultrasound by a trained physician. For each patient, 12 clips were stored, one for each thoracic region. The clips were then analyzed automatically using a built-in automatic recognition system and visually by two experts. A score of 0 was assigned if fewer than two B-lines were detected, a score of 1 if three or more B-lines were detected and a score of 2 if coalescent B-lines were detected.

Results

A total of 40 patients were included in the study. Out of the 480 clips analysed, 54.2% had a lung score of 0, 6.8% had a lung score of 1, 8.8% had a lung score of 2 and 20.2% had a lung score of 3. The agreement between the experts in detecting the scores 0, 1, 2 was excellent (ICC = 0.82) in 383 (79.8%) clips. The agreement between the automatic system and the experts was poor (ICC = 0.33), with a significant difference (p < 0.001).

Conclusions

In mechanically ventilated patients with acute respiratory failure, an automatic B-lines quantification system did not present an acceptable accuracy compared to expert consensus for B-lines counting, probably due to the coexistence of consolidation and pleural effusion.

Introduction

Traditionally, the application of ultrasound to the lung was excluded until 1989, when Lichtenstein et al. demonstrated for the first time that ultrasound could be used to inspect the pleural surface [1, 2]. Since then, lung ultrasound (LUS) imaging became a standard tool in patients with acute respiratory failure [1, 3]. In addition, LUS had been shown to improve the diagnostic accuracy and treatment of critically ill patients with pulmonary and cardiovascular diseases [3–7].

Patients with acute respiratory failure typically present with a combination of normal lung areas and lung areas with an interstitial pattern, with or without consolidation/atelectasis and pleural effusion [3, 8]. Unlike chest X-rays or CT scans, ultrasound cannot penetrate the lung; therefore, it can only assess the pleural surface and the artifacts associated with the pleural line [9, 10]. The main sonographic signs detectable with LUS are represented by B-lines, alveolar consolidations and pleural effusions [2]. In particular, multiple, well-separated B-lines suggest a moderate decrease in lung aeration, while coalescent B-lines correspond to a marked decrease in lung aeration and a consolidation reflects a complete loss of aeration of lung parenchyma [3, 10]. B-lines should be indicative of the presence of elements with a significant acoustic impedance gradient (e.g., fluid and air) [1, 10]. The presence of B-lines has been found to be associated with pulmonary interstitial syndrome, and it has been hypothesised that they originate from the resonance or reverberation secondary to the increased thickness of the interlobular septa [2, 11, 12]. However, subsequent data showed that considering the origin of B-lines solely from a specific anatomical structure was an over-simplification of the phenomena [10]. Furthermore, the visualisation of B-lines can be modified according to the setting of the ultrasound machine, such as the imaging frequency and the type and angle of the probe [10, 13]. However, the variability in the number or the characteristics of visible B-lines has been showed to have no influence on the final diagnosis [13]. B-lines has been identified in several lung diseases, such as ARDS, interstitial lung diseases, pulmonary fibrosis, cardiogenic pulmonary edema and COVID-19 [4, 9, 10, 14, 15].

Artificial intelligence has been applied in the field of radiology mainly in retrospective studies and in less than 2% of real time environments in intensive care [16]. While LUS is currently usually performed by radiologists or trained clinicians in order to reduce interobserver variability, the possibility of having automatic algorithms that can interpret LUS images and identify normal areas from interstitial syndrome could ensure a faster diagnosis, independently from the skill of a physician [17–23].

Morales et al. demonstrated that applying semiautomatic imaging software to screen patients with hypotension identified up to 55% with reduced cardiac output and B-lines patterns indicative of pulmonary edema [24]. Similar data showed an acceptable correlation between automatic LUS imaging and physician assessment of B-lines [25, 26]. The potential of ultrasound-dedicated software for the identification and quantification of interstitial syndrome as compared with expert observation is still being explored.

Aim of this study was to compare B-lines quantification by a dedicated software algorithm compared with the evaluation of expert clinicians in mechanically ventilated patients.

Materials and methods

Study design

This is a prospective observational study conducted in an intensive care unit from December 2023 to December 2025. All adult patients (> 18 years old) deeply sedated, paralyzed and mechanically ventilated in volume controlled with acute respiratory failure (defined as a PaO₂/FiO₂ < 300) were considered eligible for the study. Exclusion criteria were history of lung surgery and presence of pneumothorax, subcutaneous emphysema and a body mass index higher than 40 kg/m².

The accountable ethical committee approved the study protocol, and all the procedures were performed in accordance with the Helsinki Declaration.

Lung imaging acquisition

A detailed demonstration and instruction of the protocol to the operator were performed before the study. Within 48 hours from ICU admission a trained physician performed a standardized ultrasound lung evaluation according to Bouhemad et al [27] in patients in supine position and ventilated in volume-controlled mode. Twelve thoracic areas were analyzed, six for each hemithorax, two (cranial and caudal) for each position: anterior, lateral and posterior [28]. A Mindray TEX20 equipped with both a linear (L14-6Ns, bandwidth: 3.5 – 16 MHz, FOV 38.2mm) and a convex (SC6-1s, bandwidth 1.2 – 6.0 MHz, FOV 71°, convex radius 60mm) probe was used. The probe was applied longitudinally to the intercostal space. For each area, a 10-secondlong clip was obtained and stored. Examinations were performed using a standardized acquisition preset across patients: depth was fixed at 6 cm (increased to 8 cm only when required to include the pleural line and the far field), overall gain was set to 50% and kept constant during each examination, and the focal zone was positioned at 1cm below the pleural line. All analyses were performed post hoc on the stored clips.

Automatic lung imaging analysis

Each clip was analyzed by means of a built-in B-line automatic recognition system, (Smart B-lines®, Mindray Medical International LimitedTM), that automatically quantified the number of B-lines for each scan. The software combines a clinically trained machine learning algorithm and image feature extraction both in real time and non-real time conditions to calculate the B-line count, percent of the intercostal scan area, B-line score and distance of B-lines in a chosen video clip. For tuned B-line detection, the software detects the pleural line, delineates the acoustic shadow of the rib as the edge of ROI, and returns the following output parameters: B-line count (coalescent B-lines were considered when B-line width exceeded a specified threshold and were separately marked), B-line distance (based upon automatic quantification of the distance between two adjacent B-lines and the average distance of all the B-lines of a specific clip), B-line area ratio (indicating the ratio of the total B-line width to the ROI width), and B-line score. Specifically, the score was applied as follows: 0, (less than 2 B-lines), 1 (3 or more B-lines), 2 (in presence of coalescent B-lines, whatever the number of B-lines).

Visual lung imaging analysis

The clips were visually analyzed by two experts (a pulmonologist and a radiologist); who were blinded to each other.

Lung score of 0 if less than 2 B-lines are present, 1 in the presence of thin B-lines and 2 if a coalescent B-lines pattern and 3 for consolidation area [27, 29]. After this initial assessment, two experts discussed and reached consensus on the cases where they did not agree in the first round. After consensus was reached, all clips with a score of 3 were excluded.

Statistical analysis

Continuous data are expressed as median [inter-quartile range], while categorical data are expressed as percentage (number). The agreement between the two experts in their first assessment was evaluated by intraclass correlation coefficient (ICC). After reaching consensus, the expert lung score was compared with the automatic score by ICC. The agreement analysis between the experts was performed on the 0–3 points scale score, while the agreement between the automatic score and the expert consensus was performed after excluding clips where a consolidation was detected by the two experts. A Pearson’s correlation coefficient was used to determine the association between the scores from automatic score and from the expert consensus in each lung area. A p value < 0.05 was considered as statistically significant. All analyses were performed by RStudio (R Core Team (2022); R Foundation for Statistical Computing, Vienna, Austria).

Results

Forty patients were enrolled; a total of 480 clips were obtained and analyzed.

Study population

The baseline characteristics of the study population are reported in Table 1. As shown, the median age of patients was 63 [54 – 72] years; 25 (62.5%) of patients were male. The most common diagnosis for intensive care admission was acute respiratory failure (16 patients, 40%), with a median PaO₂/FiO₂ of 190 [165 – 230]. The median SAPS II and SOFA scores were 36 [27 – 47] and 6 [3 – 8], respectively.

Table 1.

Baseline characteristics of the study population.

	n = 40
Age, years	63 [54 – 72]
Male sex, % (n)	62.5 (25)
Weight, kg	75 [70 – 85]
Height, cm	170 [165 – 175]
BMI, kg/m2	26 [24 – 27]
ICU admission diagnosis, % (n)	15 (6)
ARDS	40 (16)
Acute respiratory failure	10 (4)
Sepsis or septic shock	10 (4)
Post operative monitoring	25 (10)
Other
SAPS II	36 [27 – 47]
APACHE III score	11 [6 – 16]
SOFA score	6 [3 – 8]
Study day
Tidal volume, mL	480 [440 – 500]
Respiratory rate, bpm	18 [16 – 22]
Minute ventilation, L/min	8.8 [7.1 – 10.2]
PEEP, cmH2O	8 [5 – 8]
FiO2	0.50 [0.40 – 0.60]
PaO2/FiO2	190 [165 – 230]
PaCO2, mmHg	42 [39 – 47]

Lung scores

Out of the 480 clips analyzed, 260 (54.2%) presented a lung score of 0, 81 (16.8%) presented a lung score of 1, 42 (8.8%) presented a lung score of 2 and 97 (20.2%) presented a lung score of 3 (p < 0.001) according to the expert consensus (Figure 1).

Figure 1.

Distribution of interstitial syndrome scores (0 to 3) according to the expert consensus system.

Agreement between the experts

The agreement between the two experts in detecting the scores 0, 1, 2 and 3 was evaluated as good (ICC = 0.73) (Table S1 and Figure S1 of Supplementary Material). After excluding consolidated patterns (score = 3), the agreement between the two experts in detecting the scores 0, 1 or 2 was evaluated as excellent (ICC = 0.82) in 383 (79.8%) clips (Table S1, Figure S2 of Supplementary Material).

Agreement between the automatic system and the experts

The total number of clips with lung scores of 0, 1 or 2 according to the expert consensus was 383 (79.8% of the total number of clips). The percentages of clips assigned a score of 0, 1 or 2 were significantly different (p < 0.001) (Figure 2, Table 2). The agreement between the automatic score and the experts in detecting a score of 0, 1 or 2 was evaluated to be poor (ICC = 0.33). For an ultrasound score of 0, 1 and 2 there was a significant difference between the automatic system and the experts (p<0.001, Figure 2).

Figure 2.

Distribution of interstitial syndrome scores according to the automatic system and the expert consensus.

Table 2.

Lung ultrasound assignment by the Automatic System (AS) and the consensus between the experts.

Lung ultrasound score	Automatic System	Expert Consensus	p
0, % (n)	55.1 (211)	67.9 (260)	<0.001
1, % (n)	44.9 (172)	21.1 (81)	0.001
2, % (n)	0 (0)	11.0 (42)	<0.001

Agreement according to lung area

Table 3 shows Spearman’s correlation coefficients according to lung area between the expert consensus and the automatic system lung score for each lung area. Although not statistically different, the correlation showed a positive antero-posterior gradient, being the highest in posterior areas, although the correlation remained weak. Correlation was also similar for the right and left lung (Table 4).

Table 3.

Spearman’s correlation coefficients for agreement between the expert consensus and the automatic system according to lung area.

Lung area	Spearman correlation coefficient (95% CI)
Anterior (1+2)	0.13 (0.02; 0.28)
Lateral (3+4)	0.25 (0.09; 0.40)
Posterior (5+6)	0.33 (0.13; 0.51)

Table 4.

Spearman’s correlation coefficient and absolute accordance rate between the expert consensus and the automatic system lung score according to the patient’s side.

Lung area	Spearman correlation coefficient (95% CI)
Right side	0.18 (0.03; 0.30)
Left side	0.19 (0.05; 0.29)

Discussion

This study found no good correlation between an automatic system and expert observers for B-lines counting in mechanically ventilated patients.

In addition to clinical examination and laboratory data, a rapid but accurate evaluation of the radiologic lung imaging is essential for a prompt diagnosis, a tailored treatment and to improve the outcome in patients with acute respiratory failure [30]. Nowadays, chest radiography is the first diagnostic lung imaging technique used to diagnose acute respiratory failure [3, 8]. However, lung ultrasound is more accurate than bedside radiography at detecting the presence of alveolar interstitial syndrome, pneumonia, pulmonary edema or pleural effusion [3, 4, 12, 31]. In addition, lung ultrasound techniques do not require the patient to be taken to the radiology department and are entirely non-invasive and safe [3, 8–10].

In normal lung conditions, ultrasound does not produce any visible image of the lung parenchyma, as the ultrasound beam is not reflected due to the presence of similar acoustic impedance [32]. However, when the amount of air decreases due to increased lung density (e.g., lung edema), an acoustic mismatch occurs and the ultrasound beam is reflected, leading to the visualization of typical artifacts [2, 11]. Among these artifacts, B-lines are defined as discrete, laser-like, vertical hyperechoic images originating from the pleural line and extending to the bottom of the screen, moving synchronously with ventilation, without fading. The artifact phenomena depend on ultrasound frequency, focal depth, frame rate, gain and operator [10, 33]. Currently, there is no consensus on the optimal ultrasound machine settings for visualizing these images [9, 10].

In clinical practice, a lung ultrasound score, based on the presence of these artifacts, is usually computed by analyzing all lung fields to assess the severity of the lung disease and to monitor the changes during the hospitalization [27, 29, 31]. A strong correlation was found between higher scores in patients with heart failure, pulmonary edema and ARDS [34].

Several examinations protocols for lung ultrasound have been proposed, ranging from simple to complex and involving multiple examinations and complicated calculations of the B-lines score [10]. In this study, a 12-lung ultrasound examination was used, as it is widely used in critically ill patients, being able to assess all the lung fields [28]. In addition, the 12-zone LUS was correlated with the severity of lung disease and with the spatial distribution of the lesions within the lung [8]. The inter-observer agreement between the two expert observers was found to be quite good (i.e. ICC 0.73) for all lung fields examined; similarly to those previously reported [10].

The rapid detection of a lung interstitial syndrome by ultrasound is essential for an accurate and reliable diagnosis in critically ill patients. A possible limitation of lung ultrasound is that the identification and quantification of B-lines require a specific training for the physicians and remains operator dependent [10, 23]. Therefore, the availability of automatic software embedded in ultrasound machines should promote a wider and faster use of lung ultrasound evaluation, mainly by non-expert physicians, ensuring adequacy of these findings in supporting clinical decisions.

The first data on the feasibility of applying an automatic B-lines scoring system was published by Brattain et al. in 2013 [18]. Subsequent studies demonstrated the accuracy of different software algorithms for automatic calculation of B-lines computation in comparison to expert observers [25, 35–37].

In this study, we evaluated an ultrasound machine equipped with an automatic software that detects B-lines on the screen and counts them in real time, generating a LUS score ranging from 0 to 2. We did not find an acceptable correlation between the automatic B-lines computation and the expert observers. Considering the lung regions evaluated, the automatic software tended to underestimate the number of B-lines, especially in the less dependent regions (anterior compared with lateral and posterior). In addition, our patient population consisted of patients with acute respiratory failure from various causes, resulting in a heterogeneous range of lung patterns. This factor could have contributed to the relatively low clinical accuracy of the automatic system, as compared to previous studied.

The study suffers from several limitations. First, the analysis was restricted to mechanically ventilated and sedated patients, the cohort was small and therefore the results’ generalizability is limited, mainly due to a selection bias. Second, due to the software design, detection of pleural effusion and areas of reduced aeration, often coexisting with B-lines in patients with acute respiratory failure [38–40], was not efficient, hindering the software performance and the agreement observed. Third, treating all clips as independent, although deriving from fewer patients, may have influenced statistical significance, overestimating clinical accuracy.

Conclusions

In the present study an automatic B-lines quantification did not show an acceptability clinical accuracy compared to expert consensus, probably due to the coexistence of consolidation and pleural effusion typically present in patients with acute respiratory failure.

APPENDIX. SUPPLEMENTARY MATERIALS

Table S1.

Lung ultrasound assignment by the experts.

Lung ultrasound score	Expert 1	Expert 2
0, % (n)	59.2 (284)	54.2 (260)
1, % (n)	10.0 (48)	16.9 (81)
2, % (n)	3.7 (18)	8.7 (42)
3, % (n)	27.1 (130)	20.2 (97)

Figure S1

Distribution of lung ultrasonography scores according to the experts.

Figure S2

Distribution of interstitial syndrome scores according to the experts.

Availability of data and material

The study data will be available from the Corresponding Authors upon reasonable request.