Detecting pulmonary nodules by using ultrasound multiple scattering

Abstract

Although X-Ray Computed Tomography (CT) is widely used for detecting pulmonary nodules inside the parenchyma, it cannot be used during video-assisted surgical procedures. Real-time, non-ionizing, ultrasound-based techniques are an attractive alternative for nodule localization to ensure safe resection margins during surgery. Conventional ultrasound B-mode imaging of the lung is challenging due to multiple scattering. However, the multiple scattering contribution can be exploited to detect regions inside the lung containing no scatterers. Pulmonary nodules are homogeneous regions in contrast to the highly scattering parenchyma containing millions of air-filled alveoli. We developed a method relying on mapping the multiple scattering contribution inside the highly scattering lung to detect and localize pulmonary nodules. Impulse response matrices were acquired in ex-vivo pig and dog lungs using a linear array transducer to semi-locally investigate the backscattered field. Extracting the multiple-scattering contribution using singular-value decomposition and combining it with a depression detection algorithm allowed us to detect and localize regions with less multiple scattering, associated with the nodules. The feasibility of this method was demonstrated in five ex-vivo lungs containing a total of 20 artificial nodules. Ninety-five percent of the nodules were detected. Nodule depth and diameter significantly correlated with their ex-vivo CT-estimated counterparts (R = 0.960, 0.563, respectively).

I. INTRODUCTION

Ultrasound imaging of biological tissues is based on the concept of echolocation where ultrasound pulses are transmitted towards a target and reflected back to be received by the ultrasound probe. This reflection is due to the difference in acoustic impedance between the target and its surrounding environment. The time delay between the transmission and reception of ultrasound waves can be converted to the distance travelled to reach the target, based on an a priori assumption on the speed of sound in the medium, allowing images to be created. A critical prerequisite for using echolocation is that there is linearity between travel time and target distance. That is not the case in lungs where this linearity is lost due to the complex microstructure of the parenchyma.

The lung is a complex and highly scattering medium. Due to the presence of numerous alveoli in the parenchyma, which are filled with air when the lung is inflated, there is a complex distribution of large amounts of strong scatterers inside the lung parenchyma. Ultrasound scattering is reinforced by the strong acoustic impedance difference between air and alveolar tissue. Alveoli cause ultrasound waves to encounter multiple scattering events within lung tissue, which results in a loss of linearity between time of flight and target distance from transducer, and also causes aberrations that distort the wave front of the focused beam.^1–5 There is also significant absorption in the lung. Therefore, conventional ultrasound imaging techniques cannot be utilized to produce ultrasound images of the lung that would accurately render its structure. As a consequence, conventional ultrasound imaging is unable to render images of the parenchyma that would allow us to detect and localize pulmonary nodules inside the lung.

Lung imaging and pulmonary nodule detection are generally done using chest radiography and X-Ray Computed Tomography (CT). Although these modalities have limitations, chest CT plays an important role in lung cancer diagnosis and has clear advantages over ultrasound imaging.⁶ CT allows pulmonary nodule detection and provides an accurate evaluation of the size and location of pulmonary nodules, as well as a detailed evaluation of the pleural surface itself.⁶ However, CT suffers from a lack of portability, a lack of real-time capability, high associated costs, and ionizing radiation, leading to ultrasound gaining popularity and becoming an option for lung evaluation. Ultrasound plays a role in the diagnosis of acute respiratory failure and has shown higher accuracy in comparison with radiography for diagnosing consolidation and pleural effusion.^7–9 Due to large amounts of fluid buildup in the critically ill lung, conventional ultrasound images of the lung show distinctive tissue-like, white regions called “white lung,” which are a well-known sign of lung consolidation. Conventional lung ultrasonography can play a role in the diagnosis of acute respiratory distress syndrome (ARDS), pneumonia, pulmonary fibrosis, and pulmonary edema by reading signs such as B-line vertical artifacts and tissue like regions due to lung consolidation, sometimes enabling diagnosis.^10–12 However, interpretation of these signs is highly subjective, operator-dependent, and frequency dependent. The qualitative nature of these methods affirms the importance of a more quantitative approach of utilizing ultrasound methods for lung evaluation. These artifacts, especially B-lines, do not provide direct information about the lung structure or the parenchyma structure. Conventional ultrasonography fails to produce a real image of the lung, and hence cannot be used to localize lesions inside the lung. Interventional laparoscopic ultrasound may be used to detect lung nodules during surgery. However, ultrasound laparoscopic images result in poor contrast, and the smaller tumors cannot be detected. In this paper, we present a quantitative approach exploiting multiple scattering of ultrasound waves, which is used as a source of contrast.

Multiple scattering and ultrasound propagation in scattering media is a complex topic which has been investigated over the last 40 years. In an infinite, homogeneous medium, the wave propagates in a known direction. On the contrary, when multiple scattering is involved, the original direction of propagation is eventually lost and the ultrasonic wave can enter a radially growing diffusive regime.^13–16 The wave diffusion is characterized by the diffusion constant (D), which depends on the emitted wave wavelength and scatterers' size and distribution and can be related to scattering mean free path (SMFP) in the lung.¹⁷ The diffusion constant and SMFP have been used as quantitative parameters to evaluate the structure of the parenchyma,¹⁸ and the SMFP was shown to be correlated to the severity of pulmonary fibrosis.¹⁹ This diffusive regime can be characterized by analyzing the coherent and incoherent waves separately. Both the backscattered coherent and incoherent waves have been proved useful to characterize disordered media.^17–26

Methods have been developed to image highly scattering and complex media. A full matrix capture was developed²⁷ (FMC) and combined it with the total focusing method (TFM), which has shown good results in imaging complex media. Due to the fact that multiple scattering alters conventional imaging techniques, a method was proposed in Ref. 28 for the separation of the single and multiple scattering contributions of backscattered waves. This work played a considerable role in the further development of imaging techniques in random scattering media. After the separation of single and multiple scattering, the DORT (decomposition of the time-reversal operator) method can be applied and combined with a SSF (single scattering filter) for removing any multiple scattering contribution.⁵ Shahjahan et al.^29,30 then improved the algorithm developed by Aubry and Derode by utilizing the DORT and TFM approaches with a multiple scattering filter (MSF) to image polycrystalline media and detect notches or cracks. In media where multiple scattering is dominant, the wave diffuses slowly, and the wave propagation can be characterized by the diffusion constant, which can be calculated by analyzing the coherent or incoherent contributions. Previous work by Aubry and Derode,²² Mohanty et al.,^18,31 and Du et al.²⁶ demonstrated that by extracting the incoherent intensity of the backscattered wave, the diffusion constant can be calculated to characterize heterogeneous media such as the lung parenchyma and melamine sponge foams. Mohanty et al.³¹ also showed that using smaller sections of an ultrasound linear array (subarrays), and mapping the diffusion constant along the transducer axis, scattering heterogeneities in a porous medium can be detected, allowing us to identify the presence of anechoic lesions in a complex medium. Mohanty et al.³² also proposed a methodology for detecting nodules by mapping the diffusion constant on a two-dimensional (2-D) map and combined it by a depression detection filter (DDF) previously developed by Winslow.³³

In this study, we propose a new quantitative approach based on separating the multiple and single scattering contributions of backscattered waves for the detection and localization of pulmonary nodules embedded inside ex-vivo lungs. Pulmonary nodules are small solidified masses inside the lung parenchyma. Although pulmonary nodules are detected by CT scans, surgeons cannot rely on CT images to localize these nodules in real time during surgery. During surgery, surgeons may use their fingers to palpate the nodules which is sometimes difficult. Small, non-visible and non-palpable nodules are challenging to localize during Video-Assisted Thoracoscopic Surgery (VATS) or robotic surgery. It is critical that the surgeon is able to locate the nodule in real time to ensure resection with safe margins. Since imaging the lung parenchyma using B-mode is elusive due to multiple scattering, we propose a new algorithm that leverages the contrast in the amount of multiple scattering inside nodules compared to the surrounding area. The contrast between the absence of scatterers inside the nodule and the presence of large amounts of scatterers around the nodule can be revealed by separating the multiple and single scattering contributions, and mapping only the single scattering. A depression detection method is then applied to localize the nodule. In this paper, we first present the algorithm and then evaluate the method by injecting artificial nodules inside ex-vivo lungs of five animals. The results are validated by comparison with CT scans of animal lungs.

II. METHODS

A. Data acquisition

For experimental data acquisition we used a 128-element linear transducer array with a central frequency of 5.2 MHz (Philips ATL L7–4) operated by Verasonics Vantage 128 (Verasonics, Kirkland, WA). The transducer was placed directly on the surface of the lung with approximately 3 mm of ultrasound gel between the probe and the lung. A two-cycle Gaussian pulse with central frequency of 5.2 MHz was emitted from each element of the array consecutively and received by all of the 128 elements every time. This resulted in a three-dimensional Inter-element Response Matrix (IRM) whose size was 128*128*N, N being the number of time samples for each time trace. The IRM has a unique reciprocity feature in which $h_{ij}\left(t\right)=h_{ji}(t)$, where $h_{ij}\left(t\right)$ is the signal transmitted from element i and received by element j.³² It is therefore a symmetric matrix. The sampling rate used for the data acquisition is 62.5 MHz. In Fig. 1, the emitted pulse and received backscattered wave has been depicted.

FIG. 1.

(a) Transmitted pulse into the scattering region. (b) Schematic example of wave path through the medium showing multiple scattering events. (c) Received backscattered signal from multiple scattering medium detected by probe.

B. Data processing and separating the multiple scattering contribution

In order to detect and localize pulmonary nodules, our algorithm produces a map in which the nodule size and location is shown with respect to the linear transducer. This map is an image which shows depth, lateral position, and size of the detected nodule. The first step of this algorithm is to separate the multiple scattering from single scattering contribution using singular value decomposition, as proposed in Ref. 28 and described later. The second step is to apply Winslow³³ depression detection technique to identify a closed contour, corresponding to the nodule in the map. The single scattering contribution shows a deterministic coherent behavior along the anti-diagonal of the IRM, which allows isolating it from the multiple scattering contribution using singular value decomposition.

Once the IRM was acquired, it was divided into sub-IRMs which consist of a small section of the IRM acquired by a sub-array. It is physically equivalent to having a smaller array translated along the lung surface. The size of a given sub-IRM is P * P * N, where P = 41in the present study. P will be called the lateral resolution in the following. For example, a sub-IRM with central element number 30 and lateral resolution of 41, $H_{41}\left(t\right)$, consists of IRM signals acquired using elements from 10 to 50. The lateral resolution should be an odd number so that P + 3 can be a factor of 4, as required for the rotation.²⁸ A lateral resolution of 41 leads to 88 sub-IRMs with respective central elements ranging from 21 to 108. The elements on the edges of the array (elements from 1 to 20 and 109 to 128) have no representation by any sub-IRMs. They cannot be used for backscattered intensity calculation and image generation. The schematic of these inaccessible regions is shown in Fig. 2.

FIG. 2.

(Color online) Picture showing Sub-IRMs within the IRM and inaccessible regions.

After splitting the IRM, all of the following steps were performed on each of the 88 sub-IRMs. Each time trace was truncated into overlapping time windows of 1 μs width leading to, for each time window T, $\mathrm{k}\left(T,t\right)=h_{ij}\left(T+t\right)\hspace{0.17em}*\hspace{0.17em}{W}_R\left(t\right)$, where with $W_R\left(t\right)=1$ for t = [0, 1 μs] and W_R(t) = 0 elsewhere. A short Fast Fourier transform (FFT) provided a response matrix in the frequency domain, $K\left(T,f\right)$ for each time window and each sub-IRM. Each element of the response matrix is represented as k_i (T, f) corresponding to the responses at the frequency f and time T between the emitter location (i = X_E) and the receiver location (j = X_R). As described by Aubry and Derode,²⁸ the K matrix at a given time window T and frequency exhibits coherence along its anti-diagonals. We can use this feature as a key to separate the response matrix into two contributing terms: one for single scattering (SS) and the other for multiple scattering (MS),

$$k_{ij}\left(T,f\right)=k_{ij}^S\hspace{0.17em}\left(T,f\right)+k_{ij}^M\hspace{0.17em}\left(T,f\right).$$

We used the following three steps to separate the SS and MS contributions to the response matrix. Step 1: since the deterministic coherence of the single scattering signals is along the antidiagonals, the matrix was rotated in order for the coherence to manifest itself along the columns. Step 2: using singular value decomposition (SVD), the rotated matrices were decomposed into two subspaces. A signal subspace on one hand, where the matrix is characterized by important correlations between its lines and/or columns, and a noise subspace on the other hand, where the matrix exhibits randomness without any correlations between its entries. Using a threshold in the eigen values to determine the rank of separation as described in Ref. 28 enables separating the signal and noise subspaces. Step 3: It is now possible to distinguish the two decomposed matrices as single scattering and multiple scattering by reversing the first step rotation to reconstruct two response matrices corresponding respectively to the SS and MS contributions.

Once the IRM is decomposed into two terms for SS and MS, it becomes possible to use multiple scattering as a source of contrast. To do so, it is necessary to calculate the backscattered intensity for each time window and each sub-IRM for SS and MS, respectively. The backscattered intensity I (X, T) is calculated, separately for SS and MS contributions, by integrating the squared values of the response matrix $k_{{X_{E}X}_R}\hspace{0.17em}(T,f)$ for all emitter/receiver couples (denoted by ${⟨\left|k(T,f)\right|}^2⟩$ in the following equation) that are separated by the same distance X = |XE-XR|:

$$I^{S,M}={⟨{\left|k_{ij}^{S,M}(T,f)\right|}^2⟩}_{f,\left\{\right(i,j\left)\right|m=j-i\}}.$$

C. Nodule detection and mapping

The method described previously enables computing the backscattered intensities for SS and MS separately as a function of time window and of the distance between emitter and receiver couples for the 88 sub-IRMs. Plotting the standard deviation of normalized SS intensity function for each sub-IRM and for each time window allowed us to create a map of this parameter, with the depth corresponding to the time window number and the lateral position corresponding to the sub-IRM number. By applying a depression detection algorithm to this map, a depression in the standard deviation of intensity of single scattering could be identified. This depression is expected to be associated with the location of the nodule shown in Fig. 3.

FIG. 3.

(Color online) Left, map of standard deviation of normalized SS intensity. Right, rendered map after depression detection. Color scales are in arbitrary units.

The position of the central element corresponds to the lateral position axis in the generated map. The time window can be converted into depth by using an effective speed of wave propagation. For this study, we arbitrarily fixed the effective speed of sound propagation as 1.5 mm/μs. The calibration was performed by fitting the depth of an artificial nodule implanted in a pig lung to the depth of the same nodule, in the same lung, measured by CT.

By fitting a Gaussian curve to single scattering and multiple scattering intensities, separately, at each time window, we calculated the standard deviation of the curve with respect to the distance X between emitter and receiver. Each point in the map has a value that is related to the standard deviation of SS intensity function. After applying median filtering and a 2-D Gaussian filter on the composed image, we generated a map in which a depression can be observed. It corresponds to a region where the amount of multiple scattering is substantially lower and is therefore expected to correspond to a nodule containing no scatterers.

D. Validation

The proposed lung ultrasound (LUS) method was validated using experimental data on ex-vivo pig and dog lungs in which artificial nodules were implanted. A total of five animal lungs, including two dogs and three pigs, were used for this study. They were recovered from animals euthanized for other reasons. Lungs were cannulated through the pulmonary artery and flushed with 2 liters of cold Custodiol (Essential Pharmaceuticals, LLC), clamped at end-inspiration and stored at 4 °C for study the following day.

To simulate pulmonary nodules, various amounts (0.2–1.0 ml) of dental impression compounds e.g., O-Bite, Honigum Pro, (DMG Chemisch-Pharmazeutische, Hamburg, Germany), Aquasil Ultra (Dentsply Sirona, Charlotte, NC), Imprint Bite (3 M, St. Paul, MN), and Splash! (Den-Mat Holdings, LLC, Lompoc, CA) were injected through a 14GA catheter directly into the apex of one of the lobes of the lung. Initially, we injected Vaseline as nodules into the lung. We also tried using acrylic latex caulk or silicone caulk for making nodules, but they remained liquid. After experimenting with different materials, we used dental impression compounds e.g., O-Bite due to its dimensional stability, resistance, and short setting time for placing inside lungs. A total of 22 nodules were implanted in the five lungs (Fig. 4). After implanting the nodule, the access point was closed off with a plastic umbilical cord clamp to avoid any air leaks.

FIG. 4.

(Color online) (a) Data acquisition from pig lung. (b) Artificial nodule excised from the lung.

For all experiments, two IRMs were acquired as described previously and used as an input for the nodule detection method. Each IRM acquisition led to a rendered nodule image. The nodule diameter and nodule depth were measured on the rendered images, and averaged over the two images. The size and depth of nodules then get averaged to ensure using both data points. Each ultrasonic acquisition happens in less than quarter second (∼100 ms).

After IRM acquisitions, the lung block had an ex-vivo CT scan and was then dissected to confirm the location and diameter of each artificial nodule. The diameter was also calculated via CT by averaging the diameters of the nodule measured in both the coronal and axial planes. The distance of the nodules from the pleural surface was also measured using the NIH software ImageJ on CT images.

III. RESULTS

Out of the 22 implanted nodules, one could not be detected (neither by CT nor by ultrasound), and was excluded from the study. Two more nodules were implanted in an airway instead of the lung parenchyma. This results in a nodule with an elongated shape, which also had to be excluded from the analysis. Out of the remaining 19 properly implanted nodules, 18 were detected, which corresponds to a 95% detection rate.

During these experiments, the deepest detected nodule was implanted 20 mm from the pleural edge, and the smallest detected nodule was 4.7 mm in diameter, as measured by CT. Figure 5 shows examples of coronal and axial CT scans of an ex-vivo pig lung. The artificial nodule is visible on both planes, and its diameter can be measured. Table I summarizes the results obtained with this new method compared to metrics obtained with CT scan.

FIG. 5.

(Color online) (a) Coronal CT scan and (b) axial CT scan images of an ex-vivo pig lung with implanted nodules. Nodule size measured in two planes and averaged, distance to the pleural edge measured by CT. (c) Magnified CT image of nodule with respect to probe. (d) Generated map of the detected nodule in the same lung, at the same location as the nodule shown in the CT.

TABLE I.

A summary of comparison between CT scan and ultrasound results.

Detection	18 /19 (95%)		Pearson correlation (R)	P-value
Method	CT scan (average ± standard deviation)	LUS (average ± standard deviation)
Nodule diameter (mm)	11.14 ± 3.60	9.44 ± 1.73	0.5623	0.015
Nodule depth (mm)	11.57 ± 4.11	11.45 ± 4.44	0.960	3.10 ×10 −10

Out of the 18 detected nodules, significant correlations were observed between the diameters obtained by CT and by scattering contrast (R = 0.56, p = 0.015). Similarly, significant correlations were observed between the nodule depth, which are defined as distance from the lung pleural surface, obtained by CT, and by scattering contrast (R = 0.960, p < 0.0001). We can see the correlation between US algorithm numbers and CT image values in the plots for Nodule size and location in Fig. 6.

FIG. 6.

(Color online) Linear correlation between US algorithm results and CT images for Nodule Depth (right) and Diameter (left).

IV. DISCUSSION

The method proposed in this paper is a novel technique leveraging ultrasound scattering as a source of contrast, to detect a solitary pulmonary nodule inside the lung parenchyma. This method is tested on artificial nodules implanted in ex-vivo pig and dog lungs. The approach discussed in this paper combined the depression detection method with the separation of multiple scattering contribution from single scattering contribution of backscattered waves, transmitted using a linear array transducer. The multiple and single scattering contributions are separated using singular value decomposition. In this article, we demonstrate that the method can be used to detect homogenous regions in the lung parenchyma, where multiple scattering and single scattering intensities are different from the surrounding area, which is highly heterogeneous and therefore highly scattering. Artificial lesions of various diameter embedded at various depths in the lung were detected. Their size and depth as measured by the proposed method correlated well with their CT counterparts. One of the major strengths of this algorithm is the potential to be developed as a fast and point-of-care method to detect pulmonary nodules during surgery.

The results show we can utilize the method in detecting nodules in the aerated lung. Two nodules had to be excluded from the study because they were implanted in an airway, which resulted in a lesion with an elongated shape. Excluding these nodules is justified because solitary pulmonary nodules are usually round, and, although lesions do occur in large airways, they would be seen and biopsied using a bronchoscope, and would likely need a lobectomy.

To ensure perfect contact between probe and lung surface, which is usually curved, we need to apply some pressure to the lung. On the other hand, too much pressure on the harvested lung can lead to the lung being compressed and deformed, which also may lead to change of air volume. This might affect the measurement of nodule depth. However, the distribution of depths for all nodules, as measured by this novel ultrasound method was not significantly different from the distribution of depths as measured by CT, according to a student's t-test. This could be related to the arbitrary choice of effective speed of sound that might impact the absolute estimation of nodule depth. In this study, we assumed 1.5 ms/μs. This parameter will need to be calibrated further in future studies on human aerated lungs for more accurate results.

Due to high scattering properties of air in alveoli in the lung the penetration of ultrasound through the lung is limited. High ultrasound attenuation is also likely to limit the depth of detectable nodules. In the present study, the deepest nodule was implanted at a depth of 20 mm, and we were able to detect it. Further studies will be necessary to fully investigate the maximum depth for nodule detection. It is possible that using lower frequencies might allow the detection of deeper nodules. However, pulmonary wedge resection using VATS is usually performed for relatively superficial nodules. The deeper nodules are usually removed via a lobectomy. As a consequence, a potential depth limitation may not affect the application of the proposed algorithm to guide wedge resection using VATS.

The resolution of the final rendered nodules image is another limitation. The size and location of the nodule detected showed some discrepancies in comparison with CT measurement. Particularly, the correlation between diameter as estimated by CT and by scattering contrast, although significant, could be improved. One of the avenues for improvement is in the application of the depression detection algorithm. As described in prior work,^32,33 a threshold is applied to determine the size of the depression. The choice of the threshold can impact the estimation of the nodule diameter. It will have to be calibrated in future studies.

Another potential source of error could be a bias in the estimation of the distance from the pleural surface by CT since the lung needs to be reinflated right before imaging. Additionally, during the ultrasound acquisition, we have to ensure perfect contact between the probe and lung pleural surface by slightly pressing the probe, which may result in a slight compression of the lung. Due to small leakage of air from the harvested lung and ongoing consumption of oxygen, keeping the lung at the exact volume between ultrasound acquisition and CT imaging is impossible. However, as described previously, the distribution of depths for all nodules, as measured by this novel ultrasound method was not significantly different from the distribution of depths as measured by CT, according to a student's t-test.

It should be noted that for the VATS application, a preoperative CT will give the approximate location of the nodule. The purpose of the present method is not to localize the nodule in the entire lung, but to accurately determine its location based on a priori knowledge from the preoperative CT. We will also be interested in measuring the location of the nodule with respect to a surgical stapler, in order to evaluate the resection margin and inform the surgeon's decision on firing the stapler. Additional experiments are planned to determine co-localization of the rendered nodule image with an image of a surgical stapler. An intra-thoracic ultrasound probe that is small enough to insert during minimally invasive surgery will be necessary for this technology to be used clinically.

Despite these limitations, the present methodology is showing very promising results. There is currently no real time imaging method to guide pulmonary wedge resection. The proposed method can be implemented with an endoscopic ultrasound probe to be inserted inside the chest. We will focus next on calibrating the effective speed of sound, calibrating the threshold for depression detection, and on the real-time implementation of the method. It will also be interesting to generate a map of the nodule that would be fused to an image of the lung so the nodule position relative to the surgical stapler can be evaluated.

V. CONCLUSION

By taking advantage of the fact that the lung parenchyma has a highly scattering structure in contrast with pulmonary nodules, which can be considered homogenous regions with uniform properties, we proposed an algorithm to distinguish the contrast in multiple scattering between the healthy parenchyma and a pulmonary nodule. We successfully implemented a novel algorithm for detecting pulmonary nodules which were embedded in ex-vivo animal lungs, based on a singular value decomposition method for the separation of multiple scattering and single scattering, which we combined with a depression detection method. This enabled the generation of 2-D a map of the detected nodule with respect to the position of the linear transducer array. The algorithm was used to detect and localize artificial nodules implanted in dog and pig lungs ex-vivo. The method was validated by comparison with CT images of animal lungs. These first results are highly promising and the comparison between lung ultrasound method and CT results showed strong and significant correlations.