Simulated transcranial ultrasound of human language cortex: verifying heard speech needs far less data and signal than decoding it

1. INTRODUCTION

• No task-evoked cortical response had yet been imaged through the temporal bone, despite prior transcranial Doppler measurements of task-evoked blood-flow velocity in large arteries.Prior functional-ultrasound reports through the intact adult skull covered a breath-hold response in arteries around the midbrain, not cortex.Source: §1 · p. 1
• A decoder must produce heard text from an open set, whereas a verifier scores a candidate supplied from outside.The paper asks what a simulated language-cortex signal could support and whether the answer depends on the readout.Source: §1 · p. 1
• On shared anatomical voxels, the study compares a published decoder with two verifiers under simulated transcranial noise, finding that verifiers need 16 to over 30 times less training and retain steeper learning under noise.The fMRI data came from three participants; clean-data learning rates were similar, but the decoder slowed and stalled under simulated transcranial physics.Source: §1 · p. 1

2.1. Data and readouts

• Three participants’ public fMRI data covered 82 training stories (15.25 h) and a repeated held-out story; every readout used the same anatomy-selected voxels.The readouts were Tang et al.’s decoder, encode-then-compare, and a contrastive verifier; the decoder used the authors’ settings and favorable word timing.Source: §2.1 · pp. 1–2
• Figure 1 maps the left-temple beam and a 5-cm voxel path, then shows one 2-s speech signal and one-probe SNR change by depth.Panel C reports the median and 10th–90th percentile per 4-mm bin across participants.Source: Figure 1 · p. 2

2.2. Evidence in bits and goals

• Verifier evidence ranks the true segment among candidate segments, while decoder evidence compares its text against 4,000 brain-free texts; the verifier goal is 1 bit.The decoder goal is the published decoder’s median-window evidence after 15 h of clean data, 4.6 to 5.7 bits; learning speed is measured per tenfold increase in training time.Source: §2.2 · p. 2

2.3. Transcranial noise model

• The simulation degrades each voxel’s real fMRI time series with Gaussian noise until its SNR matches a power-Doppler prediction for a temple probe.It models a 2-MHz, 20-mm-aperture probe, attenuation and receiver, clutter, and reverberation noise; a second probe assigns each voxel to the nearer probe.Source: §2.3 · p. 2

K SNR0 q

• Figure 2 reports that verifiers need far less data and keep learning under noise that stops the decoder, including under optimistic simulated-ultrasound physics.Panels A–F plot evidence as training data grow; G shows training needed for the goal, and H the largest uniform noise at which it is reached.Source: Figure 2 · p. 3
• The model adds voxel-wise noise to bring real fMRI SNR down to its simulated prediction; repeated test-story scans calibrate the noise, while readouts use single trials.The added noise has fMRI noise’s spatial structure, is white in time, and is independent across stories and repeats.Source: K SNR0 q · p. 3
• The link-budget simulation does not model wave propagation and has not been validated against task-evoked transcranial recordings, which do not yet exist.The authors present it as a first step and note that attenuation dominates per-voxel signal-to-noise through the temporal bone.Source: K SNR0 q · p. 3
• The simulated recording retains fMRI’s 2-s scans and whole-cortex coverage, but assumes an idealized probe and yields an upper bound for one probe.Assumptions that could improve the result include window-quality bone under the probe, head-on sound, continuous recording, and all reachable cortex; omitted factors include coded pulses and contrast agent.Source: K SNR0 q · p. 3
• Under median physics, the verifier analysis also raises every voxel’s target SNR by 6, 12, or 18 dB to test the effect of added gain.These verifier runs use 49 minutes to 15 hours of training.Source: K SNR0 q · p. 3

3.1. What the temporal window delivers

• At 5 cm depth, the simulated echo returns about 10 dB below receiver noise, and a 2-s scan’s speech signal is 22 to 24 dB below fMRI’s.Under median physics, one probe predicts SNR reductions of −23 to −25 dB across readout voxels.Source: §3.1 · p. 4

3.2. Data needed, with and without uniform noise

• In clean data, verifiers reached 1 bit with 11 to 12 min of training, while the decoder needed 3.2, 6.2 and 6.6 h.All readouts gained 2.6 to 4.2 bits per tenfold more data; the decoder’s best-predicted voxel set did not explain the gap.Source: §3.2 · p. 4
• With added noise, verifiers withstood 13 to over 17 dB more noise than the decoder across participants.From −6 dB onward, the decoder never reached its goal.Source: §3.2 · p. 4

3.3. Simulated ultrasound

• With optimistic simulated physics, both verifiers reached their goal, while the decoder gained only 0.3 to 1.0 bits per tenfold more training.The decoder’s slope confidence interval included zero in three of six cases, all with one probe.Source: §3.3 · p. 4

3.4. How much stronger the signal must be

• 12 dB of added SNR was enough for the contrastive verifier to reach 1 bit under median physics; 6 dB fell short.At 18 dB, it reached the goal in 0.9 h or less.Source: §3.4 · p. 4, Figure 3 · p. 4

• The authors hypothesize that sentence-confirming evidence may lie in coarse, redundant cortical-response components that survive noise, whereas text generation depends on finer components removed first.They note that the verifiers’ advantage holds wherever the goals are placed because their learning slopes diverged.Source: §4 · p. 4
• Scaling the published breath-hold signal gives about −40 to −31 dB per 2-s sample, an order-of-magnitude consistency check rather than validation.The estimate is based on a global deep-artery response recorded with coded excitation and log-compressed units, and is good to about ±10 dB.Source: §4 · p. 4
• The study used data from three people and fMRI as a blood-volume proxy; verification tested gist rather than paraphrases or negations.The decoder also received the best word timing.Source: §4 · p. 4

• Ethical approval was not required for this secondary analysis of de-identified open data.The original data were collected under the original institution’s IRB approval.Source: §5 · p. 5