BrainPuzzle: Hybrid Physics and Data-Driven Reconstruction for Transcranial Ultrasound Tomography

TL;DR

The paper addresses the challenge of obtaining quantitative transcranial ultrasound images due to skull-induced speed-of-sound contrasts and limited clinical aperture. It introduces BrainPuzzle, a two-stage hybrid framework that first uses physics-driven time-reversal acoustics to generate migration fragments, then employs a transformer-based encoder–decoder with a graph-based attention unit to fuse fragments into a quantitative SoS map. Experiments on synthetic brain phantoms show BrainPuzzle outperforms full-waveform inversion and various baselines, especially under partial-aperture conditions, and demonstrate robustness to fragment perturbations and noise. This hybrid approach offers a path toward practical, real-time, noninvasive brain imaging at point-of-care settings by leveraging both physical modeling and data-driven reconstruction.

Abstract

Ultrasound brain imaging remains challenging due to the large difference in sound speed between the skull and brain tissues and the difficulty of coupling large probes to the skull. This work aims to achieve quantitative transcranial ultrasound by reconstructing an accurate speed-of-sound (SoS) map of the brain. Traditional physics-based full-waveform inversion (FWI) is limited by weak signals caused by skull-induced attenuation, mode conversion, and phase aberration, as well as incomplete spatial coverage since full-aperture arrays are clinically impractical. In contrast, purely data-driven methods that learn directly from raw ultrasound data often fail to model the complex nonlinear and nonlocal wave propagation through bone, leading to anatomically plausible but quantitatively biased SoS maps under low signal-to-noise and sparse-aperture conditions. To address these issues, we propose BrainPuzzle, a hybrid two-stage framework that combines physical modeling with machine learning. In the first stage, reverse time migration (time-reversal acoustics) is applied to multi-angle acquisitions to produce migration fragments that preserve structural details even under low SNR. In the second stage, a transformer-based super-resolution encoder-decoder with a graph-based attention unit (GAU) fuses these fragments into a coherent and quantitatively accurate SoS image. A partial-array acquisition strategy using a movable low-count transducer set improves feasibility and coupling, while the hybrid algorithm compensates for the missing aperture. Experiments on two synthetic datasets show that BrainPuzzle achieves superior SoS reconstruction accuracy and image completeness, demonstrating its potential for advancing quantitative ultrasound brain imaging.

Figures (13)

• Figure 1: BrainPuzzle workflow. Stage 1: physics-driven RTM/TRA converts multi-view channel data into migration fragments under partial-aperture, low-SNR conditions. Stage 2: a transformer+GAU network fuses fragments and outputs a quantitative speed-of-sound (SoS) map. Movable partial-array sweeps provide multi-view coverage.
• Figure 2: 3D rendering of the phantom. Outer skin shown for context. Axial 2D slices are extracted along the $z$-axis for channel simulation and reconstruction.
• Figure 3: (a) Element count vs. slice index (full array). Number of transducers $N$ varies with head circumference; yellow band marks slices used in the dataset (indices 10–159), with $N\in[358,585]$. (b) Full-transducer channel data (2D slice with dashed red line at (a) ). Shape $(T,N_s,N_r)=(5001,512,512)$ for time, sources, and receivers. One element transmits per shot; echoes are recorded on all 512 receivers. Sampling $\Delta t=5\times10^{-8}$ s (0.25 ms total). Here $N_s=N_r=512$ for this slice (varies across slices; see Fig. \ref{['fig:sample_num']}).
• Figure 4: Stage-2 network of BrainPuzzle. Multi-view TRA fragments are embedded (with positional and view encodings) and passed through a shared transformer encoder; a lightweight super-resolution upsampler restores spatial detail; a graph-based attention unit (GAU) performs cross-fragment fusion and adaptive weighting; a decoder and reconstruction head output the quantitative SoS map. The architecture naturally fuses partial-array acquisitions by aggregating fragments from multiple probe positions.
• Figure 5: Partial-array acquisition schematic. A movable low-count transducer set (e.g., 50 elements) is placed at successive head positions to collect multi-view channel data. For each view, one element transmits while the others receive; repeating across sweeps yields partial-aperture coverage with improved coupling and reduced hardware cost.