The human skull remains a major obstacle for transcranial ultrasound, as it disrupts ultrasound waves and limits both imaging and therapeutic applications. Numerical simulations can help correct these effects, but they require accurate skull templates and precise alignment with patient anatomy. A recent analysis published in Ultrasound in Medicine & Biology focuses on manifold optimisation for full-waveform inversion, or MOFI, as a way to align skull templates without a separate guidance image. The method uses acoustic data to reduce the difference between simulated and observed ultrasound signals, offering an alternative to magnetic resonance imaging-guided registration.
Acoustic Data Guide Skull Alignment
Transcranial ultrasound has potential roles in imaging and therapy, including neuromodulation, drug delivery and ablation. Magnetic resonance imaging and computed tomography remain central neuroimaging tools, but they rely on large, expensive and immobile equipment located in specialised facilities. Ultrasound offers a safer, faster and more portable option, but transcranial imaging remains difficult because the skull weakens, scatters and distorts the wavefield.
The main challenge lies in the skull’s structure and acoustic behaviour. Its high impedance and heterogeneous composition degrade image quality and can affect therapeutic targeting. Full-waveform inversion offers one way to recover intracranial structure by using measured acoustic signals to reconstruct acoustic properties. However, the strong contrast between bone and soft tissue makes this optimisation difficult and can lead to convergence problems.
Template-based approaches reduce the complexity by starting from prior skull information, usually obtained from computed tomography or magnetic resonance imaging. Their clinical usefulness depends on placing the skull template accurately. MOFI focuses on that alignment step. Instead of using a concurrent magnetic resonance image, it compares simulated and observed radiofrequency acoustic data. The approach uses both reflected and transmitted signals, allowing the ultrasound data themselves to guide the skull template into the correct position.
A Simpler Optimisation Path
Conventional full-waveform inversion updates a large sound-speed image pixel by pixel. This structure supports detailed modelling, but it can be inefficient when the main problem is not the local value of each pixel but the global position of the skull template. A template that has the right shape and approximate acoustic properties can still fail if it is shifted or rotated relative to the anatomy.
MOFI changes the optimisation target from many individual image values to a smaller set of motion parameters. The implemented version uses rigid movement, covering translation and rotation. This makes the update more physically meaningful for the alignment task. The method moves the sound-speed image as a whole, rather than trying to correct a misplaced skull by changing many separate pixels.
The approach uses a smooth mathematical structure to represent this movement. This helps preserve geometric relationships while allowing the optimisation to search for the correct alignment. The transformed template is generated by mapping coordinates from the updated image back to the original image and estimating the sound-speed values. The usual full-waveform inversion gradient then feeds into an update of the motion parameters.
Must Read: AI Strengthens CT Detection of Brain Haemorrhage
This reduced formulation improves robustness when the initial misalignment is large. Once the main position and orientation are corrected, standard full-waveform inversion can still be used to refine smaller details. The method therefore acts as a coarse alignment step that prepares the model for more reliable reconstruction.
Testing Shows Improved Reconstruction
In simulated testing, the skull template started with the correct shape and sound speed but the wrong position and orientation. Conventional full-waveform inversion failed to reconstruct the brain from this starting point. MOFI then adjusted the template by estimating its translation and rotation. The final position closely matched the known displacement, and the aligned template enabled a successful reconstruction.
The laboratory testing was more challenging because several experimental factors were uncertain. The response of the transducer was not known precisely, the physical model was simplified, and the starting template had imperfectly known properties. The laboratory setup also involved materials with different densities and visible elastic effects in the observed data. These conditions made the alignment problem closer to a practical ultrasound setting than the simulated case.
During laboratory optimisation, the skull template moved smoothly from its initial position towards its final alignment. Reflected signals shifted towards earlier arrival times, while transmitted energy redistributed across neighbouring receivers. At the end of the process, the simulated and observed data aligned more closely. The initial template placement produced a failed reconstruction, but the MOFI-aligned template supported a successful reconstruction.
The approach still has defined limits. MOFI remains affected when predicted and observed data are too far out of phase, although it tolerates larger mismatch than the usual full-waveform inversion requirement. The current formulation uses a standard loss function, while other loss functions could widen the range of successful alignment. The physics also assume homogeneous density and omit viscous or elastic effects. Future extension to three-dimensional in vivo data would clarify how these factors influence performance.
MOFI demonstrates that a skull sound-speed template can be aligned automatically using acoustic data alone. The method reduces dependence on manual alignment and offers an alternative or complement to magnetic resonance imaging-guided registration for transcranial ultrasound applications. By focusing on physically meaningful movement rather than many pixel-level updates, MOFI improves the starting conditions for full-waveform inversion and supports successful reconstruction after initial failure. Remaining work centres on phase mismatch, loss formulation, simplified physics and extension to three-dimensional in vivo datasets.
Source: Ultrasound in Medicine and Biology
Image Credit: iStock
References:
Bates O, Cueto C, Coleman C et al. (2026) Automatic Skull-Template Alignment Without a Guidance Image. Ultrasound in Medicine and Biology: In Press.
