Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Automatic skull-template alignment without a guidance image

Oscar Bates, Carlos Cueto, Ciaran Coleman, Cameron A. B. Smith, Lluis Guasch, Oscar Calderon Agudo

TL;DR

The paper addresses the practical hurdle of aligning skull templates to patient anatomy without using guidance images like MRI. It introduces Manifold Optimisation for Full-Waveform Inversion (MOFI), which constrains skull-template updates to a low-dimensional SE(2) manifold (rotations and translations) and leverages full-waveform acoustic data to optimise this transformation, followed by conventional FWI refinement. MOFI uses automatic differentiation to obtain the gradient with respect to the motion parameters, improving robustness to large misalignments and reducing computational burden by avoiding high-dimensional pixel-wise updates. In-silico and in vitro experiments demonstrate that MOFI can accurately recover skull-template position and orientation, enabling accurate FWI reconstructions and offering a practical alternative or complement to MRI-guided registration with potential for easier, faster clinical deployment in transcranial ultrasound tomography.

Abstract

Transcranial ultrasound must overcome the significant challenge of the human skull, limiting both imaging and therapeutic applications. While high-fidelity numerical simulations can compensate for skull-induced distortions, they require precise skull templates (typically derived from Computed Tomography, CT) and spatial alignment to the patient's anatomy. Current alignment relies on concurrent Magnetic Resonance Imaging (MRI) for registration, introducing financial, logistical, and throughput barriers. To overcome these challenges, we present Manifold Optimisation for Full-Waveform Inversion (MOFI), a method to register skull templates without using a guidance image. Our method aligns the skull template by minimising the difference between simulated and observed radio-frequency acoustic data. We demonstrate that MOFI accurately recovers the position of skull templates in silico and in vitro, offering an alternative to MRI guidance-based registration. These results indicate that MOFI has the potential to be a practical alternative to MRI-guided approaches, reducing the barriers to wider clinical adoption of transcranial ultrasound.

Automatic skull-template alignment without a guidance image

TL;DR

The paper addresses the practical hurdle of aligning skull templates to patient anatomy without using guidance images like MRI. It introduces Manifold Optimisation for Full-Waveform Inversion (MOFI), which constrains skull-template updates to a low-dimensional SE(2) manifold (rotations and translations) and leverages full-waveform acoustic data to optimise this transformation, followed by conventional FWI refinement. MOFI uses automatic differentiation to obtain the gradient with respect to the motion parameters, improving robustness to large misalignments and reducing computational burden by avoiding high-dimensional pixel-wise updates. In-silico and in vitro experiments demonstrate that MOFI can accurately recover skull-template position and orientation, enabling accurate FWI reconstructions and offering a practical alternative or complement to MRI-guided registration with potential for easier, faster clinical deployment in transcranial ultrasound tomography.

Abstract

Transcranial ultrasound must overcome the significant challenge of the human skull, limiting both imaging and therapeutic applications. While high-fidelity numerical simulations can compensate for skull-induced distortions, they require precise skull templates (typically derived from Computed Tomography, CT) and spatial alignment to the patient's anatomy. Current alignment relies on concurrent Magnetic Resonance Imaging (MRI) for registration, introducing financial, logistical, and throughput barriers. To overcome these challenges, we present Manifold Optimisation for Full-Waveform Inversion (MOFI), a method to register skull templates without using a guidance image. Our method aligns the skull template by minimising the difference between simulated and observed radio-frequency acoustic data. We demonstrate that MOFI accurately recovers the position of skull templates in silico and in vitro, offering an alternative to MRI guidance-based registration. These results indicate that MOFI has the potential to be a practical alternative to MRI-guided approaches, reducing the barriers to wider clinical adoption of transcranial ultrasound.
Paper Structure (10 sections, 13 equations, 4 figures)

This paper contains 10 sections, 13 equations, 4 figures.

Table of Contents

  1. Method
  2. Full-Waveform Inversion
  3. Manifold Optimisation for FWI
  4. Numerical Implementation
  5. Results
  6. In-silico
  7. In vitro
  8. Discussion
  9. Conclusion
  10. Acknowledgements

Figures (4)

  • Figure 1: (a) The ground truth model in the "correct" position and (b) the skull template in the "shifted" position. (c) The correct and shifted data -- respectively the black and red traces -- show a large offset. (d) The FWI gradient captures the leftward shift, but the pixel-wise details are unsatisfactory. (d) By parametrising the image using coordinates, (f) MOFI (Manifold Optimisation for FWI) can represent the update in terms of translation and rotation of the image, providing an update direction (white arrows) that captures the shift.
  • Figure 2: In-silico wavefield data are simulated from (a) the ground truth sound speed. (b) The skull template at the initial position produces (c) a failed FWI reconstruction. (d) MOFI identifies the correct translation ($\mathrm{\delta1, \delta2}$) and rotation ($\mathrm{\theta}$). (d) The MOFI aligned skull template produces (f) a successful FWI reconstruction.
  • Figure 3: MOFI optimisation on the in vitro skull template. (column 1) The parameters $(\delta1,\delta2,\theta)$ follow a smooth curve from the initial to the final values, which (column 2) corresponds to a smooth movement from the inital to the final position. (column 3, panel 1) One set of the reflected data are gradually time-shifted to earlier. (column 4, panel 2) The transmitted data are redistributed to adjacent transducers, but not time-shifted. (column 5, panel 3) The second set of reflected data are also time-shifted, but these data overshoot before setting at the correct delay.
  • Figure 4: (a) MOFI identifies a translation $\boldsymbol{\delta}=(7.2 \pm 0.2, 0.7 \pm 0.2) \, \mathrm{mm}$ and a rotation $\mathrm{\theta} = 10.71 \pm 0.07^\mathrm{o}$. (b) The skull template at the initial position produces (c) a failed FWI reconstruction. (d) The MOFI aligned skull template provides (e) a successful FWI reconstruction.