MUST 2022

Jun 27, 2022, 12:00 PM → Jun 29, 2022, 5:00 PM Europe/London

Auditorium, First Floor, Module 16 (National Physical Laboratory)

The 3rd International Workshop on Medical Ultrasound Tomography (MUST 2022)

Location: London, UK

Dates: 27-29 June 2022

Type: In-person event

The third addition of the International Workshop on Medical Ultrasound Tomography was held in London from 27th to 29th June 2022. The conference was hosted by UK's National Physical Laboratory in collaboration with University College London and Imperial College London.

MUST 2022 is a small, intimate workshop, providing a great opportunity to discuss the latest advances in ultrasound computed tomography, with topics ranging from system design to image reconstruction. A full list of authors and talks presented can be found here.

Invited Speakers

Nicole Ruiter	Jeroen Tromp	James Wiskin	Jeroen Veltman

MUST 2022 Workshop Program.pdf

Hosting MUST 2023 Example MUST Host Proposal (NPL's 2021 Submission).pdf Instructions for MUST Hosting Proposals.pdf

Monday, June 27

12:00 PM Welcome Lunch

Introduction to NPL and Opening Remarks

1 Clinical use of ultrasound computed tomography.

As a radiologist ultrasound is mainly used as a low threshold, first step, confirm or rule out imaging modality. In combination with the correct clinical information it can be applied at bed side to support clinical decisions, also by clinicians. Besides the use as a diagnostic tool ultrasound is also used for image guided interventions.
In daily practice 2d handheld ultrasound is standard of care. Acquisition is user dependant, interpretation real time and documentation is only a selection of the exam.
The use of 3d ultrasound imaging overcomes the aforementioned limitations making it a more robust tool. However it does not improve it’s imaging properties. In breast ultrasound however the additional imaging plane that is created does help in the recognition of lesions.
Improving the diagnostic performance of ultrasound can however be achieved using additional methods like ultrasound contrast or photoacoustic tissue characteristics.
For instance in prostate or breast imaging MRI is currently the gold standard of imaging. However MRI is not the modality of choice for both prostate or breast, ultrasound is. In bore MRI guided biopsies can be performed for both organs but they are time consuming and uncomfortable for patients. This is where target biopsies supported by image fusion plays a key-role. Using image fusion the lesions detected on MRI are biopsied on whole organ 3d ultrasound imaging. This can also technically be done with the assistance of a robot arm.

Speaker: Dr Jeroen Veltman (University of Twente)

2:00 PM Tea and Coffee Break

2 Ultrasound Tomography for Breast Cancer Screening and Diagnosis

Ultrasound tomography (UST) is an imaging modality that implements solutions of the acoustic wave equation to deliver 3D imaging of various tissue properties in support of breast cancer imaging. Our group has led a series of studies to test the efficacy of UST in a clinical setting, culminating in a recently completed, multi-center, national trial aimed that resulted in pre-market approval from the FDA for breast cancer screening. The screening trial demonstrated that adding UST to mammography improved sensitivity by 20% and specificity by 8% for women with dense breasts. The improved performance helps fill a gap for women who are at higher risk for breast cancer and for whom mammography performs poorly.

Results from studies relating to breast cancer screening, diagnostics, risk assessment and treatment monitoring are also presented. These results suggest a possible role for UST, not only in screening but also in the management of breast cancer.

The challenges of translating UST from bench to bedside in a regulatory setting and the associated commercialization of the technology are also described.

Finally, future prospects for applications of UST and potential collaborations are also discussed.

Speaker: Prof. Neb Duric

3 Reconstruction of scatter characteristics using 3D Ultrasound Tomography

Conventional reflectivity imaging with Ultrasound Tomography (USCT) reconstructs qualitative images proportional to the magnitude of the impedance gradient. We propose a method to additionally recover the scatter characteristics of each reconstructed voxel, i.e. whether a reflection is diffuse or specular. This additional information may be used to discriminate different tissue types as it can be hypothesized that plain surfaces of e.g. cysts reflect the ultrasound different than rough surfaces of e.g. spiculated masses.

We recover scatter characteristics based on 3D Synthetic Aperture Focusing Technique (SAFT). In conventional SAFT, the signal amplitude at a certain time of flight for a given emitter-voxel-receiver-combination is added up for each voxel. The novel approach now separates the incoming and outgoing energy in each voxel according to the incident direction e of the emitted wave and the direction to the receiver r. To reduce memory requirements, the directions to emitter resp. receivers are discretized: per voxel a set of n direction vectors d_i with i in [1,n] (Figure 1(a)) is created for emitters and receivers leading to a 2D scatter map for each voxel with size n x n. The energy is assigned to a d_(i,s) and. d_(i,r) based on a threshold on the angle of e and r to d_(i,s) and d_(i,r) respectively. The 2D scatter map can be interpreted as the distribution of energy which has been introduced from a certain direction and which has been reflected into a particular direction.

For validation we developed a simulation based on the Phong reflection model in order to simulate signal data for scatterers with different characteristics. For visualization we show plots of the obtained 2D scatter maps in Figure 1(b) of the voxel at which the scatterer (Figure 1(a)) was located. From the maps true point scatterers, diffuse scattering objects and specular reflecting objects can be distinguished: The more concentrated the energy, the more specular the scatterer is reflecting. Furthermore, we performed several experiments to visualize the multidimensional data for the entire volume by calculating characteristic numbers such as the maximum of the scatter map divided by the mean per voxel, see Figure 1(c).

In a next step we are planning to analyze the method with simulated data of more complex objects and finally with experimental data. The novel multidimensional data, which can be reconstructed, may be used to differentiate tissues in future, e.g. by appropriate visualization for doctors or using machine learning.

Speaker: Torsten Hopp (Karlsruhe Institute of Technology)

4 In vivo measurements of the bulk ultrasonic attenuation coefficient of breast tissue using a novel phase-insensitive receiver

Speaker: Daniel Sarno (National Physical Laboratory)

5 A 3D Lung imaging using ultrasound computed tomography

There are limited or no tools for real-time imaging indicating the pulmonary function in patients going through mechanical ventilation because of respiratory failure. A bedside imaging system allows visualising the lung in real-time cab be lifesaving. Real-time images could provide early warnings of developing pulmonary pathologies in real-time, thereby reducing the incidence of complications and improving patient outcomes. A low-frequency ultrasound computed tomography (USCT) is an imaging technique with the potential to provide real-time non-ionising pulmonary monitoring in the ICU setting. In this presentation, we show the results of a new phantom study with a 32-senor array of USCT sensors in two rings of 16 sensors allowing for real-time 3D imaging of air volume in a simple lung phantom. Lung USCT application is dynamical imaging, so many transducers generally used in other medical applications of USCT may provide challenges in data collection and image reconstruction. However, the proposed USCT system is both low and computationally cost-effective, making it a good candidate for bedside imaging. We provide results of air volume reconstruction using a multi-modality USCT dynamical reconstruction algorithm.

Speaker: Prof. Manuchehr Soleimani (University of Bath)

3:40 PM Tea and Coffee Break

6 Towards ultrasound full-waveform inversion for in-vivo whole-body slice imaging of a mouse

We present current work towards a comprehensive full-waveform reconstruction using in-vivo ultrasound data of a mouse. The data was acquired with a trimodal transmission-reflection optoacoustic ultrasound imaging platform \cite{Lafci} that allows for the joint imaging of optical absorption, ultrasound reflectivity and speed of sound of the tissue. The measurement device consists of a cylindrical array of 512 transducers that collect reflection and transmission ultrasound data. The band-passed filtered data between 0.5 MHz and 6 MHz contains enough low frequency components that can be used in a full-waveform inversion (FWI), which is a powerful imaging technique to create high-resolution models of tissue and bone structure using non-invasive ultrasonic measurements.
The method is capable of using both transmission and reflection data providing quantitative images of tissue parameters such as the speed of sound, density or attenuation, as well as reflectivity images to reveal the shape and location of tissue interfaces.\
Full-waveform inversion is a non-convex problem, which typically requires an initial model that already captures the acoustic properties of the tissue reasonably well to ensure convergence toward meaningful solutions.
To make optimal use of the available transmission and reflection data, we design a multi-stage inversion strategy.
First, we use reverse-time migration (RTM) on a homogeneous initial model to obtain the reflectivity distribution of the tissue structure.
This provides valuable information to fine-tune parameters of subsequent FWI iterations. For instance, it can be used to define a region of interest and space-dependent regularization.
We then perform a time-domain visco-acoustic full-waveform inversion using the spectral-element method to discretise the wave equation.
To promote relevant information such as the separation of reflected and transmitted waves, we employ different strategies to pre-process and select data and to compare synthetic signals with measurements.
Since the cost per FWI iteration is proportional to the number of emitters used, we implement source-encoding as a sampling strategy, which allows us to use the superimposed wavefield from simultaneous emitters directly in an inversion and therefore reduces the number of wave propagation simulations that need to be computed.
Lastly, the quantitative model of the sound speed and density distribution provided by FWI is used to obtain a refined reflectivity image with RTM.
This combination of reflection and transmission information is crucial for further processing steps such as segmentation and classification.

Speaker: Ines Ulrich (ETH Zürich)

7 Transcranial Ultrasound Computed Tomography with Minimal Prior Knowledge Using Optimal Transport

Transcranial ultrasound imaging using waveform-based inversion techniques is an emerging high resolution brain imaging modality, but often relies on incorporating considerable prior knowledge to avoid the significant cycle skipping effects introduced by the skull. In this work, we present a strategy employing both reverse time migration and optimal transport which is able to mitigate cycle skipping while using a starting model that assumes no prior knowledge of any tissue or bone structures within the medium. Numerical examples are shown in silico using a realistic brain phantom to demonstrate the effectiveness of this strategy.

This inversion workflow involves first approximating the major impedance contrasts within the medium using reflection data. These reflectors are mapped to an otherwise uniform sound speed model to approximate the two highest contrast portions of the domain, namely the skin and the skull. This strategy allows for these major structures to be incorporated within the starting model without requiring one to have explicit knowledge of these tissues a priori. Given that the spectral-element method is utilized for modeling the propagation of the ultrasound waves, the soft tissue-bone interfaces can be explicitly meshed within the spatial discretization to further enhance the quality of the starting model.

Once this initial model has been constructed, a full-waveform inversion approach is used in conjunction with a graph-space formulation of an optimal transport misfit functional. This misfit functional has been shown in the context of geophysical inverse problems to be excellent at dealing with heavily cycle skipped data. This approach allows for potentially significant mismatches in both phase and amplitude between the observed and simulated measurements to be accounted for.

Speaker: Patrick Marty (ETH Zürich)

8 Wave-based Ultrasound Transmission Tomography Using the Paraxial Approximation in 2D and 3D

3D Ultrasound Computed Tomography (3D USCT) is a promising technique for early breast cancer diagnosis. The 3D USCT III device built at Karlsruhe Institute of Technology allows reconstruction of quantitative tissue parameters like speed of sound and attenuation. We investigated the paraxial approximation of the Helmholtz equation as forward model for reconstruction of speed of sound and attenuation images for this full 3D device.
An example for the setup of the method in 2D is shown in Fig. 1 (A) as ring transducer with water as a background medium and an object placed inside the region of interest (ROI). The field propagation from an emitter to receivers is divided into three parts: (i) from the emitter to z = -ROI pixel line it is done with the Green’s function; (ii) within the ROI the field is transmitted with the paraxial forward operator slice by slice; (iii) from the pixel line z = ROI the field is propagated with the paraxial approximation in one step to the receivers. Our backprojection method allows calculating the field from the receivers to the z = ROI line, which is the start line of the reconstruction.
We realized the forward solution, backprojection and reconstruction in 2D. The reconstructions were evaluated with data simulated using the k-Wave toolbox, resulting in the mean error of the speed of sound of 12.6 m/s for a pixel size of 0.3 mm. The ground truth phantom and the reconstructed speed of sound in 2D and can be seen in Fig. 1 (B,C) correspondingly.
In a next step, simulation and reconstruction were realized in 3D, based on the geometry of the USCT III device. Reconstruction was implemented by the L-BFGS algorithm. The reconstruction was initially evaluated using the paraxial approximation to simulate data. Reconstructed sound speed in 3D is presented in Fig. 1 (D,E) for coronal and transverse planes correspondingly. The average error of the reconstructed 3D speed of sound map is 0.3 m/s for a pixel size of 1.5 mm. The backprojection method in 3D based on neural networks is currently under development.
We present initial results of using the paraxial approximation for the reconstruction of the speed of sound maps in 3D.

Speaker: Prof. Hartmut Gemmeke (Karlsruhe Institute of Technology)

Conference Dinner

Tuesday, June 28

9:00 AM Breakfast

9 Full Waveform Inversion, from geosciences to medical imaging

The concept of imaging based on the full physics of seismic wave propagation was introduced in seismology approximately 35 years ago. Thanks to modern numerical methods and high-performance computers, seismic Full Waveform Inversion (FWI) has finally come to fruition in the past decade. Today, FWI is used across nine orders of frequency and wavelengths, from megahertz frequencies and millimeter wavelengths in ultrasound medical imaging and non-destructive testing to millihertz frequencies and thousand-kilometer wavelengths in seismology. The ultimate goal of FWI is to use every wiggle in a time series to map an object, be it the Earth or the Sun, a rock sample, or a body part. The purpose of this talk is to give an overview of the challenges and opportunities for FWI in medical imaging, in light of the geosciences state of the art.

Speaker: Prof. Jeroen Tromp (Princeton University)

10:20 AM Tea and Coffee Break

10 First imaging results with the new 3D USCT III setup at KIT

Based on our earlier experience with the first clinically applicable 3D USCT system, we now realized our new pseudo-randomly sampled 3D USCT device (3D USCT III). It contains 2304 transducers arranged in a semispherical aperture (Fig. 1 (a)). The aperture diameter and shape, the transducer opening angle, the bandwidth and active area of the transducers as well as the front-end electronics and data acquisition system were improved. The system is expected to increase the field of view (FOV), the image contrast and considerably reduce measurement and read out time. In this work we present first imaging results obtained with the new system.

After acquiring approx. 10 million A-Scans using all emitter-receiver-combinations, we reconstructed reflectivity images using synthetic aperture focusing technique. For signal processing we applied matched filtering, local maximum detection and convolution with an optimal pulse of adaptable width. Transmission images were reconstructed using straight and bent ray-based tomography in an algebraic reconstruction with total variation regularization. To assess the image quality we imaged several phantoms including a steel sphere and metal thread for reflectivity imaging and a custom built phantom made from gelatin with spherical inclusions of different size made from PVC. Furthermore we imaged first volunteers.

Resulting images are shown in Fig.1 (b) and (c). Initial assessment of the point spread function in reflectivity images using the metal thread phantom resulted in nearly isotropic maximum resolution of 0.26 mm. The mean reconstructed speed of sound of the main body of the gelatin phantom was 1531.3 m/s compared to 1530 m/s measured as ground truth. The mean attenuation averaged over frequency was 0.62 dB/cm compared to 0.42 dB/cm measured as ground truth. For the largest inclusion of diameter 2.2 cm, the speed of sound was 1436.1 m/s (1430 m/s ground truth) and the attenuation 6.1 dB/cm (5.9 dB/cm ground truth). The smallest inclusion of 0.8 cm diameter is visible both in sound speed and attenuation imaging.

In conclusion, the speed of sound and attenuation images show promising quality and good quantitative reproduction of the ground truth values given the ray-based reconstruction. Reflectivity images show improved contrast over the previous 3D USCT system due to pseudorandom sampling and a PSF near the theoretical resolution limit. We expect to increase the image quality further after successfully completing system calibration. Currently we are working on preparation of the system for a clinical study.

Speaker: Torsten Hopp (Karlsruhe Institute of Technology)

11 Crosstalk-free source encoding

In the context of wavespeed and attenuation maps reconstruction using Full Waveform Inversion with explicit time domain solvers, we developed a cross-talk free source encoding method. This enables image computations within minutes while data from hundreds of different sources are assimilated. Instead of involving a wave simulation number proportional to the number of sources at each iteration, the proposed method only requires two "super" wave simulations per iteration. Our source encoding method consists of capturing simultaneously the monochromatic behavior of different sources at specific frequencies. This is achieved by running a "super" wave simulation until it reaches steady state. Individual contributions of each source to the "super" steady wavefield can then be deblended by taking advantage of trigonometric orthogonality. Rather than capturing the behavior of the full spectrum of a given source, only a few frequencies per source are considered. Thanks to frequency redundancy, this decimation does not dramatically affect sensitivity kernel quality. On the other hand, assimilation of data coming from numerous different sources dramatically improves the resulting kernel, which translates into significantly faster convergence. Another benefit is the statistical reduction of the impact on convergence of data noise. We show how to build measurements based on the Fourier coefficients of the full data time series, such as waveform, phase and amplitude. We evaluate the relative convergence of each associated cost function, in their standard or in their double difference formulation. 2D and 3D results of wavespeed and attenuation reconstructions are presented.

Speaker: Etienne Bachmann

12 High-Performance Full-Waveform Inversion using Stride and Devito: a Practical View

Background

Ultrasound computed tomography techniques such as full-waveform inversion (FWI) have the potential to produce high-resolution, 3D images of tissues such as the breast, the limbs, or the adult human brain. However, adoption of these techniques is hindered by the fact that tomography algorithms are computationally demanding, and the codes that exist are closed source, difficult to maintain and slow to adapt to new research.

Here, we give a practical introduction to Stride, an open-source library for ultrasound modelling and tomography that provides flexibility and scalability together with production-grade performance; and to Devito, a domain-specific language for the automatic generation of optimised, architecture-specific finite-difference (FD) code.

Methods

Ultrasound tomography is computationally expensive because, for realistic 3D problems, it requires the solution of thousands of partial-differential equations (PDEs) and the storage of hundreds of gigabytes of memory at every iteration in order to estimate billions of parameters.

Stride provides a modular, end-to-end definition of the imaging process using state-of-the-art FWI algorithms in 2D and 3D, allowing users to easily and rapidly prototype their own inversion scripts and the flexibility to redefine every step. Stride integrates tightly with Devito, allowing us to write high-level symbolic differential equations in Python, from which high-performance FD code is automatically generated targeting both CPUs (OpenMP) and GPUs (OpenMP and OpenACC). We also provide Mosaic, an actor-based parallelisation library that lets users prototype FWI workflows in a local workstation and then efficiently scale them to hundreds of nodes in an HPC cluster without any code changes. A standard file format for the storage and exchange of ultrasound tomography data is also proposed.

We show how our software stack can be used to perform FWI reconstructions in 2D –with a numerical breast phantom (Figure 1-A) and an experimental tissue-mimicking phantom (Figure 1-C)– and 3D –with a numerical model of the adult human head (Figure 1-E). We test the numerical accuracy of the acoustic solver when compared to an analytical solution, and we showcase the scaling capabilities of in an HPC cluster.

Results

Stride produces state-of-the-art reconstructions of the breast (Figure 1-B), experimental phantoms (Figure 1-D), and the brain (Figure 1-F) with only a few lines of code. We validate its modelling accuracy by comparing it with an analytical solution of the acoustic wave equation (Figure 1-G), and validate Mosaic's scaling capacity by running on up to 128 nodes on an HPC cluster (Figure 1-H).

Speaker: Dr Carlos Cueto (Imperial College London)

13 Using uncertainty to estimate imaging errors induced by variable density, attenuation, and position

Ultrasound tomography relies on reconstruction algorithms which use approximate reconstruction physics. One promising example is full-waveform inversion (FWI), which can produce high-accuracy, high-resolution images of patient anatomy. Because the relationship between sound-speed and density is unknown, most FWI algorithms use an acoustic, constant-density approximation to simplify the inversion. Unfortunately, this approximation is emphatically untrue in the head, where the skull has a significantly higher density than surrounding tissues and supports elastic wave propagation. Failure to account for density and shear-mode propagation introduces artefacts into the reconstruction, but no reliable artefact detection methods exist. Instead, clinical practitioners are expected to identify and mitigate artefacts themselves.
To resolve this problem, we introduce a variance estimator into the reconstruction problem. The estimator is derived using a Gaussian approximation of the posterior (i.e the reconstruction) and stochastic variational inference. Despite the simplicity of the posterior approximation, the estimator is representative of image quality, requires less than 1% of the computational overhead, and needs no additional observational data compared to a mean-only image.
In this work, we also show that the estimator can identify artefacts induced by physics that violate the assumptions of the inversion algorithm. Figure 1(c) shows a constant-density FWI reconstruction of an in-silico dataset, where the in-silico dataset also adopts the constant-density approximation. Figure 1(d) shows the variance estimate is low when the reconstruction is correct. Figure 1(e) shows a constant-density FWI reconstruction of an in-silico dataset but, now, the dataset contains the density inhomogeneity shown in Figure 1(b). The failure of the constant-density approximation induces an artefact, and Figure 1(f) shows the variance estimate higher where artefact is located.

Speaker: Mr Oscar Bates (Imperial College London)

12:00 PM Lunch

14 3D Ultrasound tomography (volography): Theory, History and Clinical applications

We review some history and give a theoretical basis for the 3D low frequency ultrasound tomography (LFUT) algorithm, i.e. volography, including the evolution from CT based algorithms to nonlinear large scale minimization and performance optimization. We explain why 2D algorithms are insufficient for clinical applications, indicate relevant timing results and discuss the congruence to training a convolutional neural network (CNN) with Lie symmetries and the resulting efficiency that leads to reconstructions in clinically relevant times, making this method ideal as a high resolution imaging technology for low resource environments and underserved populations.

We summarize clinical applications including breast imaging, early detection and monitoring of breast cancer, breast density measurements, functional ultrasound tomography (FUT), knee and orthopedic imaging, pediatric and whole body imaging. We give examples and review the concomitant refraction corrected reflection algorithm, critical to obtaining sub-mm resolution. We show FUT enables doubling time estimation, calcium location in ducts or masses and introduce sequential calcium scoring and its clinical implications for monitoring cancer and disease.

We show the quantitative accuracy of speed of sound estimation for various tissues using literature values for ligaments, cartilage, tendons, muscle, skin and fat, in the presence of bone and verify its ability to monitor Duchenne MD, or sports injuries in humans or animals.
The consistent high resolution and quantitative accuracy is shown and specific breast cancer cases are reviewed. Detailed comparison of knee images with MRI indicate the improved contrast and spatial resolution. Fusion of speed of sound and reflection yield sub-mm resolution quantitative orthopedic images.

The low cost, simple training, lack of ionizing radiation or contrast agents in LFUT volography are discussed. The system is easily converted to a portable platform to serve Low Resource Environments. Over 13000 breast and related scans have been performed and extensive training sets have been culled from these. Clinical trials are reviewed showing improved area under the ROC curve when compared to X-ray mammography breast cancer screening.
Doubling time estimation, calcification and tumor functional imaging are shown and discussed. Ductal and Glandular individual segmentation is shown over time and correlated with hormone levels in volunteers, indicating high spatial and contrast resolution. Quantitative estimates of spatial/contrast resolution are reviewed.

The clinical advantages of 3D LFUT volography as a tested technology is summarized. We conclude this technology is safe and ideally suited for clinical deployment in diverse situations.

We summarize clinical applications including breast imaging, early detection and monitoring of breast cancer, breast density measurements, functional ultrasound tomography (FUT), knee and orthopedic imaging, pediatric and whole body imaging. We give examples and review the concomitant refraction corrected reflection algorithm, critical to obtaining sub-mm resolution. We show FUT enables doubling time estimation, calcium location in ducts or masses and introduce sequential calcium scoring and its clinical implications for monitoring cancer and disease.

We show the quantitative accuracy of speed of sound estimation for various tissues using literature values for ligaments, cartilage, tendons, muscle, skin and fat, in the presence of bone and verify its ability to monitor Duchenne MD, or sports injuries in humans or animals.
The consistent high resolution and quantitative accuracy is shown and specific breast cancer cases are reviewed. Detailed comparison of knee images with MRI indicate the improved contrast and spatial resolution. Fusion of speed of sound and reflection yield sub-mm resolution quantitative orthopedic images.

The low cost, simple training, lack of ionizing radiation or contrast agents in volography are discussed. The system is easily converted to a portable platform to serve Low Resource Environments. Over 13000 breast and related scans have been performed and extensive training sets have been culled from these. Clinical trials are reviewed showing improved area under the ROC curve.

Doubling time estimation, calcification and tumor functional imaging are shown and discussed. Ductal and Glandular individual segmentation is shown over time and correlated with hormone levels in volunteers, indicating high spatial and contrast resolution. Quantitative estimates of spatial/contrast resolution are reviewed.

The clinical advantages of 3D UT/volography as a tested technology is summarized. We conclude this technology is safe and ideally suited for clinical deployment in diverse situations.

Speaker: James Wiskin (QT Imaging)

wiskin_fig1.png

2:00 PM Tea and Coffee Break

15 Quantification of Speed-of-Sound with Preclinical Multimodal Transmission-Reflection Optoacoustic Ultrasound (TROPUS) system

Introduction
The multimodal transmission-reflection optoacoustic ultrasound (TROPUS) platform has been developed for multi-parametric detection and characterization of diseases in mouse models. The system consists of a circular transducer array (512 elements, 5 MHz) and provides coregistered images with multiple contrasts simultaneously, including multi-spectral optoacoustic tomography (MSOT), reflection ultrasound computed tomography (RUCT), and speed-of-sound using transmitted ultrasound waves (TUCT). In this work, we describe the speed of sound (SoS) reconstruction algorithm and rapid quantification of the SoS values obtained from the transmitted ultrasound (US) waves.

Methods
The transmitted US signals collected from 171 elements on the opposite side of each transmitting element of the TROPUS system are used to reconstruct quantitative SoS values of the tissues using a modified full-wave inversion (FWI) algorithm. The main challenge of the implementation is to enable rendering of quantitative reconstructions in a matter of minutes so that it can be used routinely. Reference waveforms are obtained from acquisitions in water (i.e. without any sample in the FOV). The signals obtained from the samples are decomposed as the sum of scaled and time-shifted versions of the reference waveforms, and the time-of-flight (TOF) for each emitter-receiver pair is obtained as the minimum of the time shifts obtained from this decomposition.

In the iterative image reconstruction process, the space between each US emitter and receiver is simultaneously sampled along multiple paths using a GPU code, and the computed TOF values are used to create estimated signals using the reference waveforms. Then, a conjugate gradient-descent algorithm is used to iteratively vary the SoS values in the defined image grid to minimize the mean-square error between the estimated waveforms and the actual measurements. This process is repeated until the cost function converges. Such a procedure requires less than 5 minutes per image slice.

Results/Discussion
As an example of a quantitative in-vivo study with TROPUS using SoS imaging, 4 NAFLD and 7 control mice were imaged. The livers were segmented using RUCT images, and SoS values of the liver were quantified in the TUCT images. As expected, the SoS decreased in NAFLD livers due to accumulation of lipids which are known to have lower SoS compared to healthy liver tissues (Figure 1).

Conclusions
SoS/TUCT images reconstructed with the proposed modified FWI algorithm provided quantitative results of lipid accumulation in a few minutes, indicating that the proposed approach is suitable for practical use in biomedical studies.

Speaker: Prof. Joaquin L. Herraiz (Complutense University of Madrid, Spain)

16 Deep-Learning-Driven Low Frequency Extrapolation for Full-Waveform Inversion Breast Imaging

Full-waveform inversion (FWI) is a promising ultrasound computed tomography (USCT) alternative to X-ray mammography for breast cancer screenings. This reconstruction method can recover high quality acoustic models at sub-wavelength resolutions by modelling the full wave propagation of ultrasonic waves to minimise the misfit between numerically generated and observed USCT data. This allows clinically useful images to be recovered while using frequencies low enough to image the whole breast without using ionising X-rays, which pose a non-negligible risk of inducing breast cancer. However, FWI tends towards local minima when the phase shift between synthetic and observed data exceeds a half-cycle. Medical transducers do not typically image at frequencies that meet this criterion for breast imaging problems. We therefore propose the application of a U-Net based low frequency extrapolation convolutional neural network (CNN) to overcome bandwidth limited hardware.

Imaging was performed using the dual P4-1 probe system seen in Figure 1b. These probes are among the lowest frequency medical probes available. Despite this, P4-1 data appears to be bandlimited, as seen by the unsuccessful P4-1 data reconstructions (see Figure 1c and 1d). Conversely, synthetic data generated using a broadband source wavelet resulted in successful reconstructions (see Figure 1c), suggesting this USCT data meets the half-cycle criterion. The CNN was trained to predict the broadband equivalent of P4-1 input data by supervised learning. Training data consisted of ~7 million paired bandlimited (input) and broadband (target) chunks of USCT data (256 times samples, 96 traces). These were sampled from synthetic datasets from imaging random slices of 3D numerical breast models. 15 additional test datasets were generated using breast slices not used during training. Experimental imaging data of a CIRS breast phantom was acquired using the dual-probe system.

Synthetic extrapolated FWI reconstructions were found to closely match broadband FWI reconstructions in all 15 test cases (see Figure 1c). The mean root mean square (RMS) difference between true breast models and the bandlimited, broadband and extrapolated reconstructions were found to be 34.76 ± 2.67 ms-1, 9.35 ± 1.17 ms-1 and 8.61 ± 0.93ms-1 respectively. Given that the RMS for broadband and extrapolated cases were comparable and lower than the bandlimited RMS, this suggests that the CNN predicted low frequency content was sufficient to meet the half-cycle criterion. Experimental breast phantom FWI reconstructions closely matched CT images of the phantom (see Figure 1d). This suggests that the low frequency extrapolation was also successful in the experimental case.

Speaker: Thomas Robins (Imperial College London)

17 Reconstructions of experimental low-frequency ultrasound tomography data collected on a circular belt of Tonpilz transducers

The low-frequency range of 10 to 750 kHz holds promise for pulmonary imaging since it has been shown (Rueter et al. are Ultraschall in der Medizin, 2009) that acoustic waves in this frequency range penetrate lung tissue. An adaptation of the Tonpilz transducer design aimed at an operation frequency between 50 and 200 kHz, a beam angle of at least 38°, and external dimensions of 25 mm diameter by 10 mm of thickness was developed. The novel sensors were calibrated for this application and found optimal efficiency near 125 kHz. A flexible belt of adjustable length that can fit up to 32 transducers was built and used to collect ultrasonic data in a ring array configuration. The data was collected with a Verasonics Vantage 64 Low-Frequency Research Ultrasound system and 32 transducers. For this preliminary study, the system was programmed to transmit a sinusoidal signal of frequency 125 kHz on one transducer at a time while listening to the received signal on the remainder of the transducers. Here we present the transducer belt, experimental configuration, and results from preliminary tests on an agar phantom with inhomogeneities. Two-dimensional transmission travel-time tomography was used to reconstruct the sound velocity. The images demonstrated the ability of this system to detect inhomogeneities accurately.

Speaker: Mr Andre Vieira Pigatto (Colorado State University)

18 Multimodality imaging for preclinical oncology applications using a PET registered Ultrasonography (PETRUS) device

PETRUS (Positron Emission Tomography Registered with Ultrasonography) [1-3] is a new hybrid in vivo imaging instrument ideally suited for the development of preclinical oncology applications using laboratory rodents. This device allows for the simultaneous monitoring of several major cancer hallmarks, in particular proliferation, dysregulation of glucose metabolism, changes in tissue composition, and angiogenesis. For this purpose, PETRUS employs Positron Emission Tomography (PET) to study glucose metabolism dysregulation, Computed Tomography (CT) to study tumor morphology, and Ultrafast-Ultrasound-Imaging (UUI) to (i) analyze the vascular architecture by Ultrasensitive Doppler [1], (ii) obtain local sound speed (SS) maps and (iii) elasticity maps to characterize tissue composition by analyzing speckle pattern disparities in ultrafast B-mode multi-angle planar sequences. In this work, we describe the methods developed for the ultrasonic component of PETRUS, applied to the study of a mouse model of paraganglioma tumors (using n=3 mice).
The ultrasound component of PETRUS consists of a clinical UUI scanner (Aixplorer, Supersonic Imagine, France) connected to a a custom-made ultralight probe (Vermon, France) with 15 MHz central frequency, 128 transducer elements and 100 µm pitch. The probe is attached through a 35 cm long hollow carbon arm (Polyplan Composites, France) to a six-degree-of-freedom high-precision micromotor (Hexapod H811, Physik Instrumente, Germany) to operate inside the PET/CT gantry.
To obtain the SS volumes, planar sequences at -5/0/5 degrees were repeated in each transverse plane of the tumor using the robot. Acquisitions were launched during the animal's respiratory pause, using an external monitor. Speckle tracking between the acquired planes was performed using a mixed optical-flow and local-phase algorithm. An iterative algorithm was applied to minimize the disparity between the speckle locations by varying the local SS map during the beamforming algorithm employed, which consisted of a delay and sum method, with heterogeneous SS map. The elasticity maps were obtained using micro-vibrators controlled with a pulse generator and tissue deformation analysis. The Doppler maps were obtained by means of the ultrasensitive Doppler technique using spatiotemporal filtering by singular value decomposition [1].
Fig.1 shows a fusion of Bmode-Ultrasensitive-Doppler-SS in a cross-section of the tumor. All parameters obtained are being analyzed for tumor phenotype characterization. This may be crucial in redefining or confirming treatment or elucidating the best time to combine different precision medicine drugs.
[1]Provost et al., Nat.BE 2018;[2]Pérez-Liva et al.,MIB 2020;[3]Facchin et al.,Theranostics 2020.

Speaker: Dr Mailyn Pérez-Liva (Universidad Complutense de Madrid)

Discussion Session

Wednesday, June 29

9:00 AM Breakfast

19 MUBI Research Platform: Recent Advances and New Developments

The Multi-Modal Ultrasound Breast Imaging platform (MUBI) is a joint development of the Spanish National Research Council (CSIC) and the Complutense University of Madrid (UCM). It is composed of a ring of 16 transducers with 128 elements each, with a diameter of 22 cm, and working with a central frequency of 3.2 MHz. The total number of elements is 2048, and the electronic system can operate in single-element emission-reception or phased-array modes. It is intended to be a flexible platform for multi-modal ultrasound imaging research, combining reflection, transmission, and elasticity imaging, mainly oriented to breast diagnosis. In this work, research advances on the system are presented: Development of a fully operational prototype to be used with patients, design and manufacturing of calibration and anthropomorphic phantoms and design of an elastography device based on a waveform generator controlled vibrator. The work describes the design and laboratory preliminary tests of the new system components.

Speaker: Jorge Fernandez Cruza (Spanish National Research Council (CSIC))

20 Towards In-Vivo Multi-Perspective Bistatic Ultrasound Imaging Using Aberration Correction

Abdominal ultrasound imaging (US) is used to monitor rupture risk of abdominal aortic aneurysms. However, assessment of aortic geometry and wall deformation using conventional US is limited by the lateral lumen-wall contrast and resolution. We therefore introduce ultrafast multi-perspective (MP) bistatic imaging to improve aortic US. In MP bistatic imaging, two curved array transducers receive simultaneously on each transmit event. The advantage of such bistatic US imaging was investigated in US simulations and in an experimental study on ex vivo porcine aortas. Using MP bistatic imaging, the wall-lumen contrast-to-noise ratio was improved by up to 8 dB in vessel wall regions between transducers (Fig. 1a). This improved the accuracy of strain estimates in US simulations and resulted in more homogeneous strain results in ex vivo experiments (Fig 1b).

In vivo, MP image fusion is hampered by wavefront aberrations, caused by the strong speed-of-sound variations between muscle and fat in the abdominal wall. This limits abdominal ultrasound image quality by introducing distortions of the imaged structures, especially at deep imaging locations, such as the aorta. In MP US, these aberrations can be even more severe, because image features from different probes can misalign severely. We developed a generic algorithm for aberration correction in delay-and-sum (DAS) beamforming to improve image quality for both single-perspective (SP) and MP US. The method employs aberration corrected wavefront arrival times based on a speed-of-sound estimate derived from the image data. Two wavefront models are compared. The first model is based on a straight ray (SR) approximation, and the second model on the Eikonal equation, which is solved by a multi-stencils fast marching (MSFM) method. Their accuracy for abdominal imaging was evaluated in acoustic simulations and phantom experiments involving tissue-mimicking and ex vivo porcine tissue that were placed on top of a commercial CIRS phantom. The lateral resolution was improved by up to 90% in simulations and up to 65% in experiments compared to standard DAS, in which MSFM-DAS outperformed SR-DAS. Moreover, successful MP image fusion in the presence of aberration was shown, yielding a better overlap and reduction in position error of the wires in the CIRS phantom by up to 85% (Fig 1c).

In conclusion, a more complete understanding of aortic geometry and wall motion can be retrieved by using ultrafast MP bistatic imaging. Moreover, results show that in vivo MP image fusion can be enabled by aberration correction using modelled arrival times in DAS.

Speaker: Vera van Hal (Eindhoven University of Technology)

21 Ray-based inversion accounting for scattering for biomedical ultrasound tomography

Most approaches to sound speed estimation in full waveform ultrasound tomography fit computationally expensive wave equation solvers to the measured time series. The computational cost of inversions therefore becomes a significant hurdle, especially for 3D tomography. This problem is exacerbated when using a sparse rotating detector array as the efficiency of such solvers does not change as the number of receivers decreases, but scales only with the number of emitters. Furthermore, it can be challenging to incorporate the directional responses of the ultrasound transducers into these models efficiently, often a key requirement to reduce model mismatch. Here we propose an inversion approach [1] that uses the full measured waveform but fits an efficient forward model based on two-point ray tracing [2]. It scales with both the number of sources and receivers, and can straightforwardly account for sensors’ directional response. Whereas time-of-flight based inversion approaches that use rays only account for refraction, this approach also accounts for first-order scattering through the use of a Hessian matrix in the inversion, and for geometric spreading through the use of Green’s law. Frequency-dependent (power law) absorption and dispersion are also accounted for in the model. The approach is based on a second-order iterative minimisation of the difference between the measurements and a model based on a ray-approximation to the heterogeneous Green’s function. Using a second-order iterative minimisation scheme, applied stepwise from low to high frequencies, the effects of scattering are incorporated into the inversion. The method is demonstrated using ultrasound data simulated using the k-Wave toolbox [3] and for a realistic breast phantom [4].

[1] A Javaherian, B T Cox, 2021, Ray-based inversion accounting for scattering for biomedical ultrasound tomography,, Inverse Problems, 37 (11), 115003.
[2] A Javaherian, F Lucka and B T Cox, 2020, Refraction-corrected ray-based inversion for three-dimensional ultrasound tomography of the breast, Inverse Problems, 36 125010.
[3] B E Treeby and B T Cox, 2010, k-Wave: MATLAB toolbox for the simulation and reconstruction of photoacoustic wave fields, J. Biomed. Opt. 15 021314.
[4] Y Lou, W Zhou, T P Matthews, C M Appleton and M A Anastasio, 2017, Generation of anatomically realistic numerical phantoms for photoacoustic and ultrasonic breast imaging J. Biomed. Opt. 22 041015.

Speaker: Dr Ashkan Javaherian (UCL)

Poster Session

10:20 AM Tea and Coffee Break

22 Comparison of Phase-Insensitive and Phase-Sensitive Ultrasound Absorption Tomography in the Frequency Domain

Ultrasound tomography for acoustic absorption relies on accurately measuring the energy loss of the propagated waves. Phase-sensitive sensors are susceptible to phase-cancellation which can result in artifacts in the absorption reconstruction, especially if the sensors are sparse or large. Recently, ultrasound tomography has been demonstrated using phase-insensitive sensors. Ultrasound absorption tomography of a 2D breast phantom is investigated for both phase-sensitive and phase-insensitive sensors in the frequency domain. Synthetic data is used with Gaussian noise added to avoid inverse crime. Both linear and non-linear reconstruction methods are considered. The linear reconstructions are Jacobian based, using an LSQR solver for Tikhonov regularisation and a primal-dual solver for total variation regularisation. The non-linear reconstructions use the L-BFGS-B algorithm with a simple adaptive step size method, where both Tikhonov and total variation priors are implemented through their gradients. The total variation function is smoothed using the Green approximation to make it
differentiable over its entire domain. Reconstruction quality is compared using mean-squared error and contrast metrics.

Speaker: Santeri Kaupinmäki (University College London)

23 Multi-perspective bistatic 2-D and 3-D ultrasound acquisitions and strain imaging of abdominal aortas

Ultrasound imaging is a safe and accessible modality to visualize the abdomen. It is frequently used to assess the geometry of the abdominal aorta, which can provide clinicians with patient-specific information of aortic aneurysms. The anisotropy in resolution and contrast, caused by the limited aperture size and refraction, degrade image quality and restrict the estimation and precision of local wall deformation and mechanical properties. Expanding the imaging aperture with an additional transducer can improve the lateral resolution and extent the angular coverage of the vessel wall. This study demonstrates a dual-transducer system for aortic strain imaging that combines multi-perspective ultrasound transmits with a simultaneous receive (bistatic imaging).

The acquisition sequence consists of fast interleaved transmits of diverging waves that can be received by both transducers. For 3-D imaging this was realized with the use of sparse random apertures on two matrix arrays. After registration, multi-perspective ultrasound images were fused and 2-D axial displacement fields were compounded, discarding all lateral tracking data. Strain estimates were compared for ex vivo porcine aortas in a mock loop of the abdomen and systemic circulation. In 3-D, aortic wall contrast was measured and spatial resolution was quantified in a phantom containing point sources, normal background scattering, and different contrast lesions.

Compounding of multi-perspective axial displacements reduced motion tracking errors with a factor 10 compared to conventional tracking of focused scanline images. Consequently, strain precision and resolution increased, leading to more homogeneous circumferential strain patterns and enabling measurements of local radial strain at high resolution wall segments which can be further extended with the inclusion of trans-probe signals. In 3-D, coherent fusion of multi-perspective signals reduced the volumetric speckle size by 66% at a depth of 8 cm and wall-lumen contrast of the aorta increased by 4.5 dB. Future work includes multi-perspective strain estimations in 3-D, in vivo measurements, and mechanical characterization of patient-specific local wall properties of the abdominal aorta.

Speaker: Mr Hein de Hoop (Eindhoven University of Technology)

24 Characterisation of Bone through Guided Circumferential Lamb-Type Waves

One of the most debilitating fragility fractures brought on by osteoporosis is a hip fracture, for which 1 in 3 patients die within 12 months of occurrence. The current gold standard of osteoporosis detection measures bone mineral density (BMD) by dual-energy X-ray absorptiometry (DEXA) and is specific but not sensitive enough. Ultrasound (US) is a promising alternative to BMD as the multi-scaled structural features of bone have a physical relationship with scattering and speed of US propagation. To date, the majority of US characterisation of bone has been carried out in peripheral regions (e.g. calcaneus, radius, tibiae) with longitudinal guided waves, which is then correlated with femoral BMD.

We propose a novel method using circumferential Lamb-type waves to directly characterise femurs, with a simplified cylindrical digital phantom simulated with the finite element package Pogo. A directional guided wave is excited through spatiotemporal transducers, travels through the cortical layer of the bone and the leaky wave is captured by the same transducers. This presentation will show that, given bone geometry, the inverse problem can be solved to obtain the local material property and thickness. Extension of the technique to porosity will also be discussed.

Speaker: Mr Aaron Chung (Imperial College London)

12:00 PM Lunch

Lab Tours

25 Ultrasound Tomography: Technical Challenges and System Design

A clinically applicable ultrasound tomography system should produce data that both provides optimal imaging results for diagnosis and is at the same time suitable for clinical use. Clinical applicability includes patient safety and comfort, high patient throughput, rapid image reconstruction, and low cost of acquisition and operation.

The technical challenges of building and operating such a device are due to the large size of a complex object being imaged, e.g. the female breast, compared to the wavelength of the ultrasound. A large number of ultrasound transducers is required to image the object, which need to be as identical as possible. In order to approximate spherical waves (3D systems) or cylindrical (2D systems), the individual transducers have to be very small, resulting in low sound level pressures and a low signal-to-noise ratio. The large number of ultrasound transducers that have to be recorded in parallel leads to a large number of parallel channels and a high data rate in order to avoid patient motions with the shortest possible data acquisition times. Due to the complex interaction of ultrasound with tissue, reconstruction algorithms for high image quality are complex and time consuming.

This paper discusses these challenges, presents the different available hardware setups and how they tackle these challenges, and provides an outlook on future developments.

Speaker: Dr Nicole Valerie Ruiter (KIT)

3:00 PM Tea and Coffee Break

26 Acoustic propagation in weakly nonlinear regime using ray tracing approximation with applications in HIFU

High Intensity Focused Ultrasound (HIFU) is a therapy that uses ultrasound waves to non-
invasively destroy malignant cells inside the human body. The technique works by sending a
high-energy beam of ultrasound into the tissue using a focused transducer. Numerically mod-
elling HIFU presents a problem due to nonlinear effects leading to the formation of harmonics
of the source frequency. Each significant harmonic requires a finer grid to resolve, rapidly
increasing computational complexity. We look to use the weakly non-linear ray theory frame-
work to reduce the nonlinear PDE in Rd to a set of one dimensional PDEs. We construct rays
emanating from the transducer on which we calculate the phase of the waves via the Eikonal
equation. In ray coordinates the amplitude can be found by solving the nonlinear transport
equation along the ray. This equation can be transformed into the Burger’s equation which
we then solve and transform back to obtain the amplitude along each ray.

Speaker: Matt Foster (UCL)

27 Automatic aperture localization in free-hand multi-perspective ultrasound imaging

The image quality in conventional reflection ultrasound is strongly restricted by the small aperture of the ultrasound transducer. This limits the field-of-view and angular coverage of image features. Full-view tomography systems solve this problem, but are generally limited to certain medical applications and anatomic sites.

An efficient tradeoff that increases the effective aperture but maintains the flexibility and cost efficiency of reflection ultrasound is multi-perspective ultrasound (MPUS). Here, two or more standard ultrasound transducers that partially surround the imaged region are combined by individually transmitting ultrasonic waves while all transducers receive on each transmit event (bistatic imaging).

We envision a future dual probe ultrasound system that allows for free-hand movement of one transducer while constantly registering and stitching the acquired volumes of both probes in space and time. However, to reconstruct bistatic image data, precise knowledge about the relative positions of the two probes is inevitable.

Approaches to apply probe localization based on image registration from different views are strongly impeded by speckle decorrelation and common image features can hardly be registered. It is, however, intuitive to believe that receive data of a wave that was transmitted by one and received by another probe contains information about the relative probe positions. Here, we present two different approaches to perform such a localization.

The first approach leverages redundancies in the Radon domain of receive data for different transmit angles in plane wave acquisitions. This is achieved by relating displacements in the Radon domain between different transmit angles to the relative probe position. The approach has been assessed for 2D data and achieved an accuracy of 2.5±1.1λ in a phantom study and an accuracy of 1.8±0.7λ in a simulation study in pure speckle.

The second approach relies on an optimization of coherence in the trans-probe image when approaching the correct relative position. We show that the problem converges nicely in 2D and discuss some challenges of the method in 3D. By including prior knowledge of the imaged geometry, we then show that this approach can achieve sub-wavelength accuracy in 3D simulation data (0.66±0.17λ). We see a great potential for this approach in free hand dual probe imaging, but also for large area, multi-aperture CMUT patches and for calibration of full-view tomography setups with rotating arrays.

Speaker: Hans-Martin Schwab (TU Eindhoven)

28 ADMM-based full-waveform inversion (FWI) for photoacoustic tomography (PAT)

This study deals with methodological development in photoacoustic tomography (PAT). PAT aims at recovering spatially varying absorption coefficients (AC), which act as an internal passive source induced by electromagnetic radiation. The acoustic waves that are triggered by this source are recorded by acoustic sensors surrounding the target and can be inverted to reconstruct the AC. In this study, we recast the AC reconstruction as a source location problem that is tackled by the so-called nonlinear full-waveform inversion (FWI) method. We design a robust and computationally-efficient formulation of FWI in the frequency domain with the alternating direction method of multipliers (ADMM). The ADMM-based FWI relies on an augmented Lagrangian function, which depends on the wavefields, the sound of speed (SOS), the ACs and the Lagrange multipliers. ADMM solves this multivariate problem iteratively with alternating directions for primal variables (wavefields, SOS, and ACs) and dual variables (Lagrange multipliers) until convergence. The practical implementation of this procedure raises however two major issues. First, the SOS model cannot be reconstructed from the photoacoustic source alone due to insufficient illumination and hence cannot be processed as an optimization variable. Therefore, we need a good apriori knowledge of SOS to reconstruct AC. Second, the ill-posedness of PAT prevents assigning a degree of freedom to each sample discretizing the medium. To overcome the later issue, we design a parsimonious parametrization of the ACs by finding regions with equal AC from the segmentation of the SOS map. By doing so, we end up with an ADMM loop that embeds two least-squares problems for wavefields and ACs. First, the ACs are kept fixed, and the wavefields are reconstructed with a data assimilation. Then, the wavefields are kept fixed, and the ACs are updated when a total-variation regularization is applied to tighten the search space of the problem. One difficulty with the proposed algorithm is the requirement for a good approximation of the SOS model. To address the first issue, we propose to build the SOS and intrinsic attenuation a priori by applying another ADMM-based FWI on an ultrasound dataset obtained by sending and recording acoustic waves from the device surrounding the target. The numerical results on a neonate head model show significantly better results than classical time-reversal algorithms.

Speaker: Dr Hossein Aghamiry (University Cote d'Azur - CNRS - IRD - OCA, Geoazur, Valbonne, France)

Closing Remarks