{"id":"e14f5a77-4fdb-48f1-bee8-4297af27ca7b","arxiv_id":"2505.12998","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new open dataset of 2,500 3D focused-ultrasound-through-skull simulations and a conditioning U-Net that predicts pressure fields with about 2.5 mm median focal error.","lead":"This paper releases TFUScapes, a public dataset of 2,500 simulated ultrasound pressure fields through 125 skull models, and DeepTFUS, a neural network that predicts those fields from skull images and transducer position. If the dataset is valid, it gives the transcranial focused ultrasound community a shared benchmark and a fast route to treatment planning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (1c) is undefined for soft-tissue HU values and as written assigns the maximum attenuation at the low end of the HU range, contradicting the stated water-like brain model; this affects every simulated ground-truth field.","rationale":"The reader's weakest assumption was that pseudo-CT plus simplified Eq. (1) mappings may not represent real patient acoustics; I agree that this is the main risk to the clinical/realism claim. My stress-test sharpens it to an internal, checkable ambiguity in Eq. (1c): as written it is undefined for the stated soft-tissue HU range and, under plausible clipping, assigns brain attenuation an order of magnitude too high. This is more directly load-bearing than pseudo-CT fidelity alone because it does not depend on outside data or consensus: it can be settled by reading the released code. The paper should be credited for open-sourcing the dataset and configs, which makes this test feasible, and for candidly listing pseudo-CT uncertainty in Section 5. If the code reveals a correct water/bone mapping with proper clipping, the concern dissolves and the remaining issue is the acknowledged pseudo-CT fidelity. If the code confirms the mapping bug, the ground-truth pressure fields and all reported metrics are affected, so the dataset would need regeneration or a major caveat. Because either resolution is possible without discarding the engineering contribution, keep the conditional verdict rather than accept or reject.","tokens_in":10861,"tokens_out":10113,"duration_ms":110953,"concrete_test":"Inspect the released TFUScapes construction code (YAML config generation and medium setup) and evaluate Eq. (1) at HU=0, 300, and 2000. Determine whether HU is clipped and what α values are actually assigned to soft-tissue/brain voxels. Then rerun one 500 kHz k-Wave configuration twice: once with the released mapping and once with a physically consistent water-like soft-tissue α≈0.1–0.2 dB/(MHz cm) and bone α increasing with HU up to ≈8.7 dB/(MHz cm) at HU=2000. If the focal position shifts by ≳1 mm or the peak amplitude changes by >10%, the reported median focal error 2.45 mm and peak-pressure error 16.6% are not robust to the mapping ambiguity and the dataset's ground-truth fields need regeneration or re-analysis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires TFUScapes simulations to be usable ground truth. The paper states that soft tissue and brain (HU≈0) are modeled approximately as water (Section 4.1), but Eq. (1c) is written with HUmin=300 and HUmax=2000. For HU=0 the radicand (HU−HUmin)/(HUmax−HUmin) is negative, so α(x) is undefined unless values are clipped or special-cased. If HU<300 is clipped to HUmin, every brain voxel receives α=8.7 dB/(MHz cm), not water-like attenuation (~0.1–1 dB/(MHz cm)); if another clipping is used, the paper does not say. Since Eq. (1) is the mechanism by which pseudo-CT values become the acoustic medium for the k-Wave runs, this ambiguity is not a detail of the neural surrogate: it propagates into the amplitudes and focal distortions that both the dataset and the evaluation metrics treat as truth. The paper's own Limitations section (Section 5) flags pseudo-CT fidelity, but not this immediate implementation question. The open release of the dataset and configuration files makes the question directly checkable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TFUScapes, a public dataset of 2,500 full-wave 3D transcranial focused ultrasound (tFUS) simulations generated with the k-Wave solver from pseudo-CT volumes derived from T1-weighted MRI scans of 125 subjects. Each subject is paired with 20 bowl-transducer configurations at 500 kHz with random focal lengths and apertures in the 55–75 mm range. The paper also presents DeepTFUS, a transducer-conditioned 3D U-Net that predicts normalized pressure fields from a cropped pseudo-CT volume and transducer surface coordinates, using Fourier positional encoding, attention pooling, dynamic convolutions, FiLM, and cross-attention. Training uses a weighted MSE loss plus a gradient consistency loss. On a held-out test set of 30 subjects, the full model achieves relative_l2 of about 41%, a median focal position error of 2.45 mm, and a median maximum-pressure error of 16.63%. The dataset and configuration files are publicly released.","tokens_in":11056,"tokens_out":5048,"duration_ms":51730,"significance":"If the simulations are physically sound, TFUScapes would be a valuable public resource for data-driven tFUS planning, and DeepTFUS demonstrates a plausible path toward near-instant pressure-field estimation. The paper's strengths include the use of a standard solver (k-Wave), release of the dataset and code, detailed configuration-file pipeline, and explicit discussion of limitations. The main significance is therefore contingent on the correctness of the acoustic medium construction; the reported predictive accuracy is meaningful only if the underlying simulations are trustworthy, which the manuscript does not yet establish.","major_comments":[{"comment":"Equation (1c) is undefined for Hounsfield units below HUmin=300 because the square root of (HU−HUmin)/(HUmax−HUmin) becomes negative when HU<300. The text in Section 4.1 states that soft tissue and brain matter (HU≈0) are modeled approximately as water. If values below HUmin are clipped to HUmin, then every soft-tissue and brain voxel receives α=αmax=8.7 dB/(MHz cm), which is not water-like and would strongly alter focal amplitudes and phase. If a different clipping or water assignment is used, it is not specified. Since this mapping defines the acoustic medium for every k-Wave simulation in the dataset, this is not a cosmetic issue: it propagates into all ground-truth fields, the trained model, and all reported metrics. The authors must correct the formula, specify the exact handling of HU<300, and confirm that the released configuration files are consistent with the stated water-like brain model.","section":"Section 3.1, Eq. (1c)"},{"comment":"The dataset is not validated against real CT-based simulations or experimental measurements. The paper's central claim is that TFUScapes contains 'anatomically realistic' full-wave simulations, but the acoustic medium is derived entirely from pseudo-CT (an MR-to-pCT model) with a simplified HU-to-property mapping and a water-like brain model. The Limitations section acknowledges uncertainty, but the manuscript provides no quantitative comparison—e.g., a subset of cases re-simulated with real CT, or comparison with the Aubry et al. benchmark cases (Ref. [2])—to support the fidelity of the generated ground truth. Without such validation, the dataset's value as a benchmark for physics-based tFUS planning is unclear. Please add at least one comparison experiment or state explicitly and more prominently that the dataset is not validated for physical accuracy.","section":"Section 5 and Section 4.1"},{"comment":"The ablation results do not consistently support the full architectural design. Removing FiLM yields the lowest max_pressure_error (14.65±12.59), removing dynamic convolutions gives nearly identical focal_position_error to the full model (2.98±2.03 vs. 2.89±2.14), and DeepTFUStiny is competitive across all metrics. No statistical tests, confidence intervals, or repeated runs are reported. The text's claim that DeepTFUS achieves 'the best focal position accuracy' is therefore not substantiated. Please report variance across seeds or statistical significance, and temper the architectural conclusions accordingly.","section":"Table 1"}],"minor_comments":[{"comment":"There are several typos: 'Y AML' in Section 1, 'vmodels' in Section 2.2, and 'Fourier postional encoding' in Section 4.2.","section":"Throughout"},{"comment":"The Fourier encoding formula is ambiguous as typeset: 'sin(20πti), cos(20πti),..., sin(2n−1πti)' likely means sin(2^{k}πt_i) for k=0,...,n−1; please use explicit superscript notation.","section":"Section 3.2, Eq. (3)"},{"comment":"The variable 'ppp' (points-per-period) is used but not defined in terms of ppw and the CFL number; please provide the exact relation used to compute Δt.","section":"Section 4.1"},{"comment":"The description of dynamic convolutions is underspecified: it is not clear how many dynamic kernels are generated, what their spatial size is, or how they are applied to the U-Net feature maps at each resolution.","section":"Section 3.2"},{"comment":"The dataset release would benefit from explicit specification of the file format (beyond '.npz'), the exact contents of the metadata (e.g., transducer geometry, HU-to-property parameters per simulation), and the software license terms.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The Eq. (1c) issue is the critical point. If the released code confirms the formula as written without special-casing HU<300, the dataset's physical validity is compromised, and the paper would need to regenerate or re-validate the simulations. The absence of any comparison to real CT or experimental data is a second concern that, combined with the Eq. (1c) issue, makes the central claims of 'anatomically realistic' and 'high-fidelity' unsubstantiated until addressed. The paper is potentially a useful contribution to the community, so I recommend a major revision with the opportunity to correct and validate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know this paper is first and foremost a dataset paper. The TFUScapes resource — 2,500 full-wave 3D tFUS simulations from 125 subjects — is genuinely new and openly released. That alone makes it a worthwhile contribution; prior work was 2D or small-scale and closed. The DeepTFUS model is a sensible baseline: U-Net with transducer embeddings via Fourier features, FiLM, dynamic convs and cross-attention. It's a re-combination of known components, not a novel architecture, but it's applied cleanly and ablated honestly. The ablation shows FiLM removal actually improves max-pressure error, which the authors report without spin. Good.\n\nThe soft spots: first, the attenuation mapping in Eq. (1c) is undefined for HU < 300. Since brain tissue is around HU≈0, the radicand goes negative. If the code clips to HUmin, every brain voxel gets α=8.7 dB/(MHz cm) — the maximum, not water. That would corrupt every simulated ground-truth field. This is directly checkable because the dataset and configs are public, and it must be clarified before I'd trust the ground truth. Second, the ground truth is unvalidated against experiment: pseudo-CT rather than real CT, brain treated as water, single 500 kHz frequency. The authors state this limitation, but it's more than a caveat — it means the dataset's absolute pressure values and focal distortions carry unknown bias. Third, the abstract says 'input 3D CT volumes' when the model actually takes pseudo-CT from T1 MRI; that's an inconsistency. Finally, 'high-fidelity wavefields' in the abstract is an overstatement given 41% relative L2 error, though the focal metrics (median 2.45 mm error) are reasonable.\n\nNone of this kills the paper. The dataset fills a real gap, and the mapping issue is likely a fixable clarification. I'd send this to peer review, with the expectation of major revision: explain or correct Eq. (1c), provide the exact HU-to-property code path, run a small validation against real CT if possible, and tighten the abstract. The authors have been candid; they just need to close the loop.\n\nRead this if you work on ultrasound simulation or data-driven medical imaging. It's a solid engineering contribution, not a breakthrough, but a useful one.","headline":"TFUScapes fills a genuine gap with an open 3D tFUS simulation dataset, but the attenuation mapping has an ambiguity that must be resolved before the ground truth is trustworthy.","tokens_in":11645,"tokens_out":3555,"would_cite":true,"duration_ms":36243,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces a public dataset of 2,500 full-wave transcranial ultrasound simulations and a neural model that predicts the 3D pressure field from skull anatomy and transducer geometry, localizing the focal spot within a median of…","keywords":["transcranial focused ultrasound","deep learning surrogate","pressure field prediction","pseudo-CT","3D U-Net","transducer conditioning","full-wave simulation","focal spot localization"],"falsifier":"Generate a test set where ground truth comes from real CT-derived acoustic parameters (or from hydrophone measurements through an ex vivo human skull) and run DeepTFUS with the same 55–75 mm bowl transducer at 500 kHz; the claim would be falsified if median focal position error rises above roughly one wavelength (3 mm) or median peak-pressure error exceeds about 25% on that real-bone benchmark.","tokens_in":10646,"feed_emoji":"🧠","tokens_out":6496,"duration_ms":62608,"temperature":0.7,"pith_summary":"Transcranial focused ultrasound (tFUS) targets deep brain structures with millimeter precision, but planning each patient requires running full-wave acoustic simulations through the skull that take hours. This paper aims to replace that bottleneck with a learned predictor: it builds TFUScapes, a public dataset of 2,500 full-wave 3D pressure-field simulations across 125 MRI-derived skull models with randomized transducer geometries, and trains DeepTFUS, a 3D U-Net conditioned on the transducer position and shape, to output the normalized steady-state pressure field directly from the skull volume and transducer coordinates. On held-out subjects the model finds the focal spot with a median error of 2.45 mm and estimates the peak pressure with a median error of 16.6%, at a tiny fraction of the solver's runtime. The claim, if it holds, is that data-driven surrogates can make patient-specific tFUS planning fast and scalable enough for routine clinical use.","feed_headline":"AI predicts skull ultrasound focus within 2.45 mm","feed_subtitle":"A public dataset of 2,500 simulations trains a model that localizes the focal spot and estimates peak pressure.","key_machinery":"The load-bearing mechanism is transducer-aware conditioning of a 3D U-Net (an encoder–decoder convolutional network with skip connections). The transducer's active surface is represented as a set of 3D points, each encoded by a Fourier positional embedding and processed by MLP layers; an attention-weighted sum pools the points into a single permutation-invariant embedding. That embedding is injected into the U-Net three ways: dynamic convolutions in the encoder, feature-wise linear modulation (FiLM, a learned per-channel scale-and-shift) in the decoder, and bi-directional cross-attention between the embedding and CT features at each resolution. A weighted mean-squared-error loss, with voxel weights exponential in pressure, plus a gradient-consistency loss, focuses training on the focal region and preserves spatial structure.","core_discovery":"The central discovery is that a conditional deep network can approximate the full-wave pressure field of a transcranial focused ultrasound transducer well enough to locate the focus and estimate its amplitude, given only a pseudo-CT skull volume and the transducer's 3D coordinates. The paper builds the largest openly released dataset of its kind—2,500 full-wave simulations at 500 kHz through 125 anatomically realistic skull models, each with 20 transducer placements—and shows that the proposed model, DeepTFUS, achieves a median focal position error of 2.45 mm and a median maximum-pressure error of 16.6% on 30 unseen subjects. The ablation study establishes that the spatially weighted loss, not the architectural embellishments, is the main driver of localization accuracy: moving from plain L2 to the weighted loss reduces median focal error from 4.64 mm to 2.55 mm, while removing FiLM actually improves peak-pressure error. The paper therefore argues that simulation-grade accuracy is achievable at inference speed, with the residual error being acceptable for planning.","pith_inferences":["A decisive validation the paper does not perform is testing on real CT-derived acoustic models; if pseudo-CT systematically underestimates bone density, the 2.45 mm error may not survive the transfer to actual patient bone.","The network outputs a normalized field, so converting predictions into absolute acoustic intensities or thermal dose requires a separate calibration step against the solver's source amplitude.","Because the conditioning mechanism is decoupled from the U-Net backbone, the same architecture could be retrained for transcranial imaging or phase-aberration correction, where the input–output pair has the same geometry.","The ablation suggests the model is over-parameterized: removing FiLM improved peak-pressure error, so a leaner architecture might match or beat the full model at lower inference cost."],"forward_implications":["Patient-specific tFUS planning could shrink from hours of numerical simulation to seconds of network inference, enabling real-time re-planning during a procedure.","The public TFUScapes dataset gives the community a common benchmark to compare future data-driven tFUS solvers, which did not previously exist at this scale.","The loss-function result indicates that any competitive model should weight high-pressure voxels heavily, not just minimize global L2 error.","The transducer-conditioning design is general: the same embedding machinery could accept different aperture sizes and curvatures within the trained range, and extensions could cover other source geometries.","The reported median localization error of 2.45 mm is below one wavelength at 500 kHz (3 mm in water), suggesting the predicted field is usable for focus targeting."],"supporting_citations":[{"why":"Supplies the pseudo-spectral solver used to generate every simulated pressure field in the dataset.","marker":"[28]"},{"why":"Supplies the MR-to-pseudo-CT conversion that turns T1 MRI volumes into the acoustic tissue maps.","marker":"[32]"},{"why":"Defines the Hounsfield-unit-to-acoustic-property mappings (density, sound speed, attenuation) used in Equation 1.","marker":"[33]"},{"why":"Defines the focal position error and maximum pressure error metrics used to evaluate the model.","marker":"[2]"},{"why":"Provides the 3D U-Net backbone that the transducer-aware conditioning modules extend.","marker":"[21]"},{"why":"Supplies the 125 manually skull-stripped T1-weighted MRI scans that seed the dataset.","marker":"[19]"},{"why":"Introduces the Fourier positional encoding used to embed transducer coordinates.","marker":"[27]"},{"why":"Introduces dynamic convolutions, one of the three transducer-conditioning mechanisms.","marker":"[6]"},{"why":"Introduces feature-wise linear modulation (FiLM), the decoder conditioning mechanism used in the model.","marker":"[18]"}],"fun_headline_variants":["Skull-adaptive AI targets ultrasound within 2.45 mm","DeepTFUS: AI predicts skull focus with 2.45 mm precision","AI model estimates ultrasound focus from CT in 2.45 mm","Largest tFUS dataset powers fast, accurate AI focus prediction","Fast AI predicts brain ultrasound focus for skull targeting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that pseudo-CT volumes made from T1 MRI, combined with the simplified Hounsfield-unit-to-acoustic-property formulas, faithfully capture how real skull bone bends and absorbs ultrasound; if the synthetic skull properties are systematically wrong, the simulated training data and everything learned from it will not reflect true patient acoustics.","fun_headline_variants_meta":{"raw":{"variants":["Skull-adaptive AI targets ultrasound within 2.45 mm","DeepTFUS: AI predicts skull focus with 2.45 mm precision","AI model estimates ultrasound focus from CT in 2.45 mm","Largest tFUS dataset powers fast, accurate AI focus prediction","Fast AI predicts brain ultrasound focus for skull targeting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000422,"raw_usage":{"total_tokens":2219,"prompt_tokens":1047,"completion_tokens":1172,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":1083}},"tokens_in":663,"tokens_out":1172,"duration_ms":9053,"temperature":1.0,"reasoning_tokens":1083,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:21:47.871719+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a test set where ground truth comes from real CT-derived acoustic parameters (or from hydrophone measurements through an ex vivo human skull) and run DeepTFUS with the same 55–75 mm bowl transducer at 500 kHz; the claim would be falsified if median focal position error rises above roughly one wavelength (3 mm) or median peak-pressure error exceeds about 25% on that real-bone benchmark.","supporting_citations":[{"cited_title":"k-wave: Matlab toolbox for the simulation and reconstruction of photoacoustic wave fields","cited_arxiv_id":null,"evidence_quote":"Supplies the pseudo-spectral solver used to generate every simulated pressure field in the dataset."},{"cited_title":"Pseudo-cts from t1-weighted mri for planning of low-intensity transcranial focused ultrasound neuromodulation: An open-source tool","cited_arxiv_id":null,"evidence_quote":"Supplies the MR-to-pseudo-CT conversion that turns T1 MRI volumes into the acoustic tissue maps."},{"cited_title":"Bric tus simulation tools","cited_arxiv_id":null,"evidence_quote":"Defines the Hounsfield-unit-to-acoustic-property mappings (density, sound speed, attenuation) used in Equation 1."},{"cited_title":"Benchmark problems for transcranial ultrasound simulation: Intercomparison of compressional wave models","cited_arxiv_id":null,"evidence_quote":"Defines the focal position error and maximum pressure error metrics used to evaluate the model."},{"cited_title":"The preprocessed connectomes project repository of manually corrected skull-stripped t1-weighted anatomical mri data","cited_arxiv_id":null,"evidence_quote":"Supplies the 125 manually skull-stripped T1-weighted MRI scans that seed the dataset."},{"cited_title":"Dynamic convolution: Attention over convolution kernels","cited_arxiv_id":null,"evidence_quote":"Introduces dynamic convolutions, one of the three transducer-conditioning mechanisms."}],"review_version":1}