Keywords blood flow - localization microscopy - microbubble contrast agents - super-resolution
ultrasound - lymph node micro-vessel
Introduction
The lymph node (LN) is a key route for cancer metastasis and LN status is one of the
most important indicators of prognosis in patients who are diagnosed with cancer.
Currently, various modalities are used to characterize LNs, including ultrasound (US),
magnetic resonance imaging (MRI), computed tomography (CT), and positron emission
tomography (PET). Liao [1 ] has reported that the sensitivities of existing modalities US, MRI, CT, and PET
for differentiating between benign and malignant LNs are low (<=66%). Invasive procedures
such as LN dissection and excisional sentinel LN biopsy are routinely used for the
diagnosis and management of cancer patients, despite the fact that many cancer patients
do not have LN metastases at the time of diagnosis [2 ]
[3 ]. Besides the cost of such invasive procedures, they can also cause complications
like infection, lymphedema, and sensory loss [2 ]
[4 ]. Hence, there is a pressing clinical need for an advanced noninvasive imaging modality
to accurately characterize LN diseases.
Angiogenesis plays a fundamental role in the development of the chaotic and irregular
vessel structure [5 ] and can contribute to the early detection, diagnosis, and prognosis of cancer [6 ]. From a physiological point of view, vessels whose function is of great significance
for tissue integrity often lie at the perfusion level. Hence, noninvasive imaging
of microvasculature can potentially address the aforementioned clinical need in LN
characterization.
Contrast-enhanced ultrasound (CEUS) has emerged as a valuable tool for imaging flow
and tissue perfusion in vivo [7 ]. CEUS has been used for imaging human lymph systems, e.g., for identifying sentinel
LNs [8 ], and for distinguishing benign and malignant LNs [9 ]. However, due to various confounding factors in CEUS imaging [10 ], objective and quantitative clinical assessment of the images is challenging, and
inconsistent results have been reported in different studies [11 ]. Furthermore, existing clinical CEUS has limited spatial resolution, making characterization
of LNs (some <1cm) difficult.
Optical super-resolution has revolutionized the field of fluorescence microscopy [12 ]. Its acoustic counterpart, super-resolution ultrasound (SRUS), localizes and tracks
individual microbubbles, enabling the quantification of microvascular structure and
flow far beyond the wave diffraction resolution limit [13 ]. Notably, studies include the first demonstration of super-localization in vivo
[14 ] and the first demonstration of super-resolution by separating two closely positioned
structures as well as super-resolved velocity mapping [15 ]
[16 ]. A number of following studies have demonstrated the feasibility of SRUS in vivo
in animal models [17 ]
[18 ]
[19 ]
[20 ]
[21 ]
[22 ] including tumor models [17 ]
[21 ]
[22 ], in healthy volunteers [23 ], and in breast cancer patients [21 ]. Recently phase change nanodroplets have also been used for real-time SRUS [24 ].
Volumetric SRUS images of LN vascular structure and flow dynamics have been demonstrated
in a scanning 2D imaging approach on a rabbit model [19 ], offering a spatial resolution of < 30 microns. Noninvasive SRUS imaging with such
resolution in deep tissue has great potential to improve the management of, e.g.,
cancer patients. However, the use of SRUS for human LN imaging to distinguish different
pathologies has not been reported.
In this study we hypothesized that the application of SRUS imaging provides quantitative
markers to distinguish metastatic LNs from benign (reactive) ones in patients with
lymphadenopathy.
Materials and methods
Clinical data acquisition
CEUS data acquisition was performed on patients (Supplementary Table 1 ) with suspected lymphadenopathy between March 2019 and September 2019 at the Second
Affiliated Hospital, Zhejiang
University School Of Medicine. Ethical approval was provided by the University Institutional
Ethics Board. Informed consent was signed by the participants before the ultrasound
examination.
Ultrasound examinations were performed using a clinical scanner (Mindray Resona 7S)
which provides live dual images of B-mode and CEUS. An L11–3U transducer was used
with a transmit central frequency of 5.6 MHz. The patients were asked to breath normally.
B-mode and color Doppler ultrasound was first used to identify the target node, and
a long-axis view without a major artery nearby which otherwise can cause significant
pulsatile motion. Then 1.2ml SonoVue (Bracco, Milan, Italy) microbubbles were administered
intravenously as a 2-second bolus followed by a flush of saline. Clinical CEUS data
were acquired for at least 80s after the injection, using a frame rate of 20Hz in
CEUS mode and a mechanical index (MI) of 0.085. The bubble concentration was adjusted
by a flash sequence on the system at an MI of 0.553 at 40s after injection, and a
second flash was given if bubble concentration was judged high and individual bubbles
could not be distinguished by visual assessment 60s after injection. After CEUS acquisition,
ultrasound-guided core needle biopsy was conducted according to the routine clinical
management protocol. The pathological results received from biopsy were generated
for comparison. The datasets for 10 out of 44 patients were included for producing
SRUS images through post-processing of the CEUS images as they met the following conditions:
(1) there is no large or significant out-of-plane motion; (2) the corresponding pathological
diagnosis results were available; (3) the length of useable video data with appropriate
bubble concentration was >= 30s.
Post-processing for SRUS image
SRUS images were obtained through off-line post-processing of the acquired video sequences.
After motion correction [23 ], an intensity threshold was estimated by maximum entropy to remove the noise. To
detect microbubbles in the images, only regions with features including shape, intensity,
and size that were inside the predefined ranges were accepted as a single microbubble
signal. The ranges of these features were determined by the variations of individual
bubble signals measured in a lab phantom using a similar ultrasound setting. These
detected microbubbles were localized and tracked over frames and accumulated to generate
SRUS images. Each centroid position of a detected microbubble was represented by a
Gaussian profile in the localization density map with a standard deviation of 10 µm,
given by the localization precision.
The tracking algorithm found the best correlated bubble signals within a search window
between neighboring frames [13 ]. For each bubble signal identified in frame n, the intensity cross-correlation was
calculated between that bubble and each of those bubbles found in frame n +1 within
a search window. The bubbles in different frames with the maximum normalized cross-correlation
above an empirically determined threshold of 0.65 were paired. The search window size
was defined according to the imaging frame rate and the maximum flow velocity of interest.
Given the limited frame rate of 20Hz for the clinical system, and as we were mainly
interested in the micro-vessels where the flow velocity is typically several millimeters
per second, 600 micrometers was set as the maximum search range so that flow velocities
of up to 12 mm/s could be tracked.
The detected bubbles and tracks were accumulated to generate the SRUS images and flow
velocity and direction maps. The SRUS processing pipeline is described in supplementary Fig. S1 .
Quantifications from super-resolution ultrasound image
Vessel density
Vessel density is defined as the ratio of the total microvascular area in the binary
LN SRUS microvascular map and the total area of the region of interest (ROI). The
ROI is a manually selected full 2D cross-section of the LN. The binary SRUS microvascular
map was generated by detecting the pixels in the SRUS map with an intensity higher
than a localization threshold of 0.5.
Vessel spatial complexity
Fractal dimension is a ratio providing a statistical index of complexity for a given
structure/pattern. It is hypothesized that the microvascular structure associated
with cancer would be more complex than that of normal tissue, and hence has a higher
fractal dimension. The fractal dimension of the LN microvasculature was used to quantify
the complexity of the microvascular geometry and estimated by applying the frequently
used box-counting method developed by Russel et al. [25 ].
Flow velocity and direction
Since microbubbles remain within the vascular space and have similar flow dynamics
as red blood cells [26 ], the tracking of individual microbubbles was made to generate super-resolved maps
of blood flow velocity and direction of the microvasculature.
Local Flow Direction Irregularity (LFDI)
The blood flow in a malignant tumor is more likely to be disorganized. Therefore,
a measure of how irregular local micro-flow is could be a promising marker for malignancy.
This was only made possible by the high-resolution images of microvascular flow dynamics
afforded by SRUS. We define Local Flow Direction Irregularity (LFDI) as the variance
of flow direction within a defined local region/window:
Var (angle ) = E [(angle – E (angle ))2 ]
where E denotes the mathematical expectation, and angle denotes the flow direction
obtained through super-resolved velocity mapping. The LFDI was calculated block-by-block
for each non-overlapping 2 mm×2 mm block regions within the segmented LN ROI. A block
size of 2 mm was defined according to the difference between micro-metastases (less
than or equal to 2 mm) and macro-metastases (greater than 2 mm).
The mean and standard deviations of the different measures were generated from 4 reactive
and 6 metastatic LN acquisitions. Student’s two sample t-test was used to test the
statistical significance of differences in the different image markers between metastatic
and reactive LNs.
Results
[Fig. 1 ] shows B-mode, CEUS maximum intensity projection (MIP), and SRUS images of two sample
LNs. The LN ROI is indicated by the green line.
Fig. 1 Images of two sample LNs (A, B, C are from a reactive lymph node; D, E, F from a metastatic lymph node). A, D : Conventional B-mode ultrasound, LN region of interest (ROI) is manually segmented
from the original B-mode image as indicated by the green contour. B, E : Maximum Intensity Projection (MIP) images of the LNs. C, F : Super-resolution binary LN vessel map.
Supplementary Fig. S2 illustrates that SRUS image contains more detailed morphological information which
is not visible in the MIP image. Figs. S2B and C show higher magnification of the
same ROI from the MIP and SRUS velocity map. The same ROI is indicated with the white
box in Fig. S2A.
Supplementary Fig. S3 shows a comparison of a classic Doppler vascular image and the SRUS image of the
same reactive LN. As can be seen in the color Doppler image, only major vessels with
fast blood flow (up to 6.7 cm/s) are visualized. In the corresponding SRUS velocity
map, blood flow with a much slower velocity can be detected.
Figs. 2A and 2C show the SRUS direction maps for the same two LNs as Fig 1. The corresponding
LFDI maps are also visualized in Figs. 2B and 2D. As can be seen from the figure,
the reactive LN shown in [Fig. 2 ]B has a lower LFDI and appears more homogeneous than that of the metastatic LN shown
in [Fig. 2 ]D.
Fig. 2
A, C : Super-resolution flow direction maps of two LNs, where color codes the angle of
blood flow direction; B, D : Local Flow Direction Irregularity (LFDI) maps show a higher degree of irregularity
in the metastatic LN in D compared to the reactive LN in B .
[Fig. 3 ] displays measurements of micro-vessel density, fractal dimension, mean velocity,
and LFDI for the reactive and metastatic LNs. As can be seen, reactive LNs have a
significantly lower LFDI (1000 ± 376) than that of metastatic LNs (1600 ± 388, P =
0.0465). While the estimated micro-vessel density has a lower average value (31 ±
10%) in reactive LNs than in metastatic LNs (42 ± 5%), no statistically significant
difference was found. For the results of fractal dimension, reactive LNs have a similar
value (1.75 ± 0.04) as that of metastatic LNs (1.79 ± 0.05). The measured blood flow
velocity has a slightly higher average (2.0 ± 0.7) for reactive LNs than for metastatic
LNs (1.8 ± 0.5), although the difference is again not statistically significant.
Fig. 3 Quantification results in reactive (n=4) and metastasis (n=6) LNs: A : micro-vessel density. B : fractal dimension. C : mean speed and D : Local Flow Direction Irregularity (LFDI). Significant difference between reactive
and metastatic LNs was only found in the LFDI (P < 0.05.).
Discussion
Abnormal hemodynamics in tumor-associated vasculature can be a valuable imaging marker
clinically, but currently there is a lack of an effective tool to evaluate this in
vivo in deep tissue. Since SRUS is able to generate a microvascular morphology and
flow velocity map with resolution at the microscopic level, it allows the analysis
of local hemodynamics in greater detail.
In this study we have, for the first time as far as we are aware, applied SRUS to
the evaluation of LNs in patients with lymphadenopathy using a clinical scanner. An
image marker derived from the SRUS flow maps, Local Flow Direction Irregularity (LFDI),
shows a statistically significant difference between reactive and metastatic LNs,
even at a very low sample number (10 patients in total). This is encouraging and consistent
with our expectation of increased heterogeneity in hemodynamics in tumor-associated
vasculature [27 ]. It should be noted that in Opacic [21 ] an entropy measure was used to characterize the local microvascular flow direction,
similar to the LFDI measure presented in this study.
A number of other quantitative image markers derived from the SRUS images including
micro-vessel density, micro-flow velocity, and vessel fractal dimension were also
generated. As shown in [Fig. 3 ], the estimated micro-vessel density has a lower average value in reactive LNs than
in metastatic LNs. This is consistent with the positive correlation between micro-vessel
density and tumor occurrence which have been reported in previous studies [28 ]. However, the difference is not statistically significant, likely due to the small
number of samples and the different types of malignancies and locations of LNs used
in this study. The data here contains five different types of malignancies from various
body locations. Furthermore, the different time points since the onset of the cancer
for each patient is also a confounding factor. This has increased the variability
of the data within each group. Previous studies using spectral Doppler sonography
have shown increased flow velocity in reactive LNs [29 ]. Furthermore, it was reported that the overall blood flow velocity tends to be lower
in tumors than in normal tissues [30 ]. These are consistent with the data shown here.
SRUS imaging offers visualization and quantification of individual micro-vessels,
not achievable in conventional modalities. These novel measures open new avenues for
characterizing metastatic LNs in order to inform patient management. The gold-standard
procedure for LN staging is still surgical excision and histopathological assessment,
despite the fact that many patients with cancer do not have LN metastases at the time
of surgery, and that the procedure can result in complications. Our results demonstrate
that SRUS as a noninvasive imaging modality has potential to assist LN diagnosis.
2D imaging with a single slice of an LN which is a 3D structure certainly poses a
risk of missing a micro-metastasis. Multiple 2D slice acquisitions would reduce such
a risk, but this will reduce the time spent on each acquisition given the limited
persistence of the bubble signals in vivo, and as a result the number of localizations
and signal-to-noise ratio for each individual 2D SRUS image may be reduced. However,
given that SRUS imaging requires low bubble concentration, it is possible to change
from a bolus to an infusion so that a sufficient amount of data from multiple planes
can be acquired. Further clinical studies should be done to optimize data acquisition
for multiple 2D slices.
One limitation of this study is the low number of patients. In this study out of the
44 patient datasets, 34 were not used for SRUS processing, mainly due to out-of-plane
or large motion (14 datasets), and also due to the usable video length with appropriate
bubble concentration being less than 30s (20 datasets). The data acquisition protocol
should be further optimized. A slightly longer acquisition time or acquisition at
a later time after injection could generate data with more appropriate bubble concentration
for SRUS. More training for operators as well as future full 3D imaging technologies
would help address the out-of-plane motion issue. Further development of more advanced
post-processing would also help to make use of data with higher bubble concentrations
and greater motion.
Another limitation of the study is that the super-resolved flow velocity tracking
is limited to a flow velocity of ~12mm/s or less. This is primarily due to the limited
imaging frame rate available to the clinical system being used. A higher frame rate
system will allow flows in a broader range to be tracked. Additionally, while the
flow velocities were calculated based on bubble pairs in this study, velocity estimation
using the Kalman filter based on multiple frames [21 ] may provide a more robust estimation. A further limitation of this study is that
while we have clinical diagnosis of the LNs in this study through core needle biopsy
and pathology as part of the patients’ clinical management, we do not have sufficient
information to be able to generate a spatial correspondence between the pathology
and the SRUS images.
Conclusion
In this study we have, for the first time as far as we are aware, applied SRUS to
the evaluation of LNs in patients with lymphadenopathy. Local microvascular flow direction
irregularity has been shown to be a promising marker for distinguishing metastatic
LNs.
