Super-resolution imaging
- Algorithms:
We have extended our previous 1D network by adding a transducer-element dimension, enabling processing of multi-channel RF data in a single forward pass. The model now outputs super-resolved images directly, rather than deconvolved multi-channel RF. To preserve interpretability and generalizability, we embed differentiable beamforming layers (Delay-and-Sum and Stolt’s f-k migration) directly within the network. These algorithms are implemented in a modular, user-friendly framework allowing non-experts to easily run experiments and compare methods. The multi-channel model successfully deconvolves the lateral direction, overcoming a key limitation of the earlier single-channel model.
We are now investigating realistic 3D pressure fields, with promising preliminary results from the beamforming layer, and are exploring the benefits of hard-coding the beamforming algorithm in terms of performance, dataset size, microbubble density, and invariance to steering angle and transmission. Neural networks have been trained to localize microbubbles in raw ultrasound data for different pulse types, demonstrating that waveform-specific networks outperform others in noisy conditions (see Fig. 5).
Our next step is to generate arbitrary transmit pulses optimized for super-resolution through differentiable simulations.
- Experimental translation:
We have achieved initial experimental validation by applying our algorithms to a monodisperse bubble suspension, yielding encouraging results. To progress from prototype validation to in vitro assessment, we are finalizing a flow phantom that uses hydrodynamic instabilities to create a vortex street as the imaging target. The second version of this setup is now under final testing.
In parallel, we explored particle image velocimetry (PIV) on super-resolved images and demonstrated the critical role of point-spread function (PSF) shape. Preliminary results indicate that 2D super-resolution algorithms significantly reduce the PSF, thereby enhancing velocity estimation accuracy.
Bubbles as capillary mechanical sensors
- Numerical models:
In collaboration with CEMEF (Sophia Antipolis, France), we developed a 3D finite-element model of an ultrasound-driven, coated microbubble confined within a (hyper)elastic capillary. The simulator has been validated against free-field theory and published results. Final refinements are underway, including comparisons with experimental data.
We use this model to study how geometry influences tissue mechanics, bubble dynamics, and wave generation. By varying capillary wall thickness, length, and stiffness, we simulate realistic confinement. Unlike prior studies that neglect surrounding tissue, we find that different wave modes arise depending on wall thickness, significantly affecting microbubble response. Quantifying these effects will improve our understanding of how confinement alters bubble dynamics.
- Experiments:
We characterized the acoustic properties of tissue-mimicking materials in terms of ultrasound attenuation, optical transparency (for high-speed imaging), and stiffness. Based on these results, we developed a multi-capillary phantom embedded in a soft, tunable viscoelastic material with capillary diameters from 15 to 200 μm. The design is being refined to improve perfusion, and we are also developing a nonlinear-propagation compensation method to enhance measurement sensitivity.
Because storage and loss moduli depend strongly on frequency, we recorded ultrasound-driven oscillations of microbubbles in PAA hydrogels using ultra-high-speed imaging and extracted loss and storage moduli from resonance curves. We are extending this approach using a recent theoretical breakthrough on bubble dynamics in gels and atomic-force-microscopy-inspired techniques to determine frequency-dependent viscoelastic properties.
