{"id":"ccc43bf2-701b-4e1b-a0c8-a91afb449c3f","arxiv_id":"2507.08688","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An MRI-only pipeline that generates synthetic skull CT and combines it with fast angular-spectrum solvers can plan transcranial focused ultrasound targets with reported sub-millimeter lateral accuracy and about 90% less computation time than full-wave simulation.","lead":"The authors trained a neural network to create synthetic CT scans from MRI, then used those synthetic scans in fast ultrasound simulations to plan focused ultrasound brain treatments without exposing patients to radiation. If accurate, the approach could make repeated treatment planning safer and much faster.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eqs. 7–8 invert the HU-to-density/sound-speed mapping; if literal, higher-HU bone gets negative density and sound speed, so every simulated acoustic field is computed on nonphysical skull maps.","rationale":"The central contribution is an end-to-end pipeline whose reported accuracy is defined through the HU-to-acoustic conversion in Section 2.2. As written, both linear conversions have decreasing slopes, so higher-HU bone is assigned lower density and lower sound speed, with negative values at the upper end. This is more load-bearing than the reader's identified weakest assumption about 17-subject sCT generalization, because even a perfect sCT would fail if the acoustic parameter maps are nonphysical. The same sign error contaminates the gold-standard rCT maps, so the reported rCT-versus-sCT consistency checks are not sufficient to detect it. I therefore disagree with the reader's identification of the weakest assumption. The runtime inconsistency (34 s in the abstract vs. 345 ± 85 s in Figure 8E) is a secondary issue that also supports the need for released code and artifacts, but Eq. 7–8 is decisive. Unless the authors confirm the printed equations are typographical errors and verify the simulations with corrected formulas, the central claim of accurate, CT-free tFUS planning is not supported.","tokens_in":17381,"tokens_out":5641,"duration_ms":64024,"concrete_test":"Re-derive Eqs. 7–8 from Marsac et al. (reference 41) and recompute one skull's density and sound-speed volumes from the same HU data using both the printed formulas and the corrected monotone formulas. Then rerun the kWave-rCT pipeline for one deep and one shallow target with both conversions. If the printed formulas yield any negative density or sound-speed voxel, or if focal position shifts by more than the reported sub-millimeter and 0.2-normalized-pressure thresholds when using the corrected formulas, the central accuracy claim fails unless the manuscript is corrected.","verdict_should_be":"REJECT","load_bearing_attack":"Section 2.2 defines the physical link between imaging and acoustics. Eq. 7 reads ρ = ρmin + (ρmin − ρmax)(HU−HUmin)/(HUmax−HUmin) and Eq. 8 reads c = cmin + (cmin − cmax)(ρ−ρmin)/(ρmax−ρmin), with ρmin = 1000 kg/m³, ρmax = 3200 kg/m³, cmin = 1480 m/s, cmax = 4050 m/s. At the upper HU end, these formulas give ρ = 2ρmin − ρmax = −1200 kg/m³ and c < 0, which is physically impossible for any tissue. The standard Marsac-model forms use (ρmax − ρmin) and (cmax − cmin), so the signs appear inverted. Because the same conversion is applied to rCT and sCT, the reported rCT-versus-sCT agreement cannot reveal this error: both pipelines would be biased together. If the sign error is real, the CT-based gold standard is not a valid reference and all targeting, FWHM, and pressure-error numbers are computed on nonphysical maps. If it is only a typesetting or OCR mistake, the authors must state the corrected equations and release code; otherwise the central claim of accurate tFUS planning is unsupported as written.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fully MRI-based pipeline for transcranial focused ultrasound (tFUS) planning: a 3D U-Net with transformer blocks generates synthetic CT (sCT) from T1-weighted MRI, and the resulting skull acoustic property maps are used in three simulation pipelines (full-wave k-Wave, hybrid k-Wave/Angular Spectrum Method, and Rayleigh–Sommerfeld/Angular Spectrum Method). The authors compare sCT-based simulations against real-CT-based k-Wave 'gold standard' simulations on five test skulls, reporting sub-millimeter targeting deviation, consistent focal widths, normalized pressure errors below 0.2, and roughly 90% reduction in simulation time for the hybrid methods. The manuscript claims this enables fast, accurate, radiation-free tFUS planning.","tokens_in":17646,"tokens_out":4756,"duration_ms":52554,"significance":"If correct, the proposed framework would be a practical step toward CT-free, rapid tFUS treatment planning, leveraging open-source tools (k-Wave, MONAI, ANTs) and offering a clear time advantage over full-wave simulation. The paper includes a quantitative comparison of six pipelines, which is a useful design. However, the central claims are currently undermined by a likely sign error in the density/sound-speed conversion equations, by an abstract that overstates the targeting accuracy and misreports the RS-ASM runtime, and by evaluation on only five subjects with no external validation. These issues are fixable in principle, but they must be addressed before the results can be considered physically meaningful.","major_comments":[{"comment":"The Hounsfield-unit-to-acoustic-property conversion is written with inverted signs. Eq. (7) gives rho = rho_min + (rho_min - rho_max)*(HU-HU_min)/(HU_max-HU_min), so at HU = HU_max the density is 2*rho_min - rho_max = 2000 - 3200 = -1200 kg/m^3. Eq. (8) similarly gives c = 2*c_min - c_max = 2960 - 4050 = -1090 m/s at the upper end. These are unphysical values, and because the same conversion is applied to both rCT and sCT, the reported rCT-versus-sCT agreement cannot reveal the error. Every simulated pressure field, focal metric, and timing result is therefore computed on nonphysical skull maps, and the k-Wave 'gold standard' is itself invalid. If this is a typographical error and the code uses the standard forms (rho_max - rho_min) and (c_max - c_min), the authors must correct the equations and provide the corrected code or the corrected numerical results; otherwise the central claim of accurate tFUS planning is unsupported as written.","section":"Section 2.2, Eqs. (7)-(8)"},{"comment":"The abstract's claim of 'sub-millimeter targeting deviation' is contradicted by the reported quantitative results. Figure 8B shows longitudinal (axial) deviations of 1.5-2.6 mm for all pipelines; only transverse errors (Figure 8D) are below 0.7 mm. The abstract should either state 'sub-millimeter transverse deviation' or also report the axial errors. Additionally, the abstract states that RS-ASM reduced simulation time to 34 s, but Figure 8E reports 345 +/- 85 s for RS-ASM; 34 s would correspond to ~99% time savings, not the claimed ~90%. The abstract's runtime number is inconsistent with the results by an order of magnitude.","section":"Abstract and Figure 8"},{"comment":"Table 1 reports MAE values for N=5 test skulls without standard deviations, confidence intervals, or any statistical test comparing the proposed method with the pseudo-CT baseline. With only five samples, the improvements in head, brain, and skull MAE may not be statistically meaningful. Furthermore, the table header says 'Skull MAE (0<=HU<=2000)' but the text in Section 3.1 refers to 'skull region (HU > 2000)', and the training loss uses a bone mask of HU [100,1500]. These inconsistent HU thresholds make the skull MAE result difficult to interpret and should be reconciled.","section":"Section 3.1, Table 1"},{"comment":"The 'targeting deviation' results are relative to k-Wave simulations using real CT, not to an absolute ground truth such as hydrophone measurements, ex vivo skull experiments, or clinical outcomes. The paper therefore demonstrates consistency between simulation pipelines, not absolute tFUS targeting accuracy. The authors should explicitly state this distinction and temper the conclusion that the framework provides 'precise tFUS targeting' in an absolute sense. Additionally, the evaluation uses only five test skulls from the same dataset used for training, with no external cohort, different field strengths, or different MRI protocols, so generalizability to clinical populations remains unestablished.","section":"Section 3.2.3"}],"minor_comments":[{"comment":"Section numbering is inconsistent: Section 1.1.3 is followed by Section 2.2 ('MRI-Derived Acoustic Parameter Estimation') and Section 2.3; these should be renumbered to maintain a coherent hierarchy (e.g., 1.2, 1.3).","section":"Section numbering"},{"comment":"In the brain SSIM loss, the mask is multiplied only with the SSIM fraction, so outside the mask the loss equals 1 - 0 = 1 (or 1 - mask*SSIM), rather than ignoring non-brain voxels. Consider using masked averaging over the brain region or applying the mask to both the numerator and the denominator consistently.","section":"Eq. (5)"},{"comment":"The phrase 'fully CT free simulation framework' is inaccurate because the sCT model is trained on CT and the gold standard uses real CT; the framework is CT-free only at inference. Please clarify this to avoid overstatement.","section":"Abstract"},{"comment":"The third highlight contains a typo: 'approachs' should be 'approaches'.","section":"Highlights"},{"comment":"The sentence 'We developed a deep learning pipeline to sCT images from paired T1-weighted MRI data' is missing a verb; it should be 'to generate sCT images'.","section":"Section 1.1.1"},{"comment":"The notation 'Skull MAE (0<=HU<=2000)' is inconsistent with the text's 'HU > 2000' and with the bone mask range used in training ([100,1500]); please define the HU bins precisely and use the same definitions throughout.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Eqs. (7)-(8) is likely a typo rather than a deliberate modeling choice, but it is load-bearing because all acoustic simulations use that conversion. I would ask the authors to correct the equations, rerun the simulations if necessary, and release the conversion code. The abstract also needs correction for the sub-millimeter claim and the RS-ASM runtime. The small sample size and lack of external validation should be addressed or at least more strongly caveated. The work has practical value if these issues are resolved, but the current manuscript is not suitable for publication in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — this paper integrates MRI-derived synthetic CT with angular-spectrum accelerators for tFUS planning and compares six pipelines against a k-Wave/real-CT gold standard. That systematic comparison is the real contribution; each component is known but someone needed to run the combinations on the same skulls. The authors also cite the relevant sCT and ASM literature honestly. If the numbers held up, the workflow would be clinically useful.\n\nThey don't hold up as written. Section 2.2, Eqs. 7–8, contain inverted signs in the HU-to-density and density-to-sound-speed conversions. The standard Marsac forms use (ρmax−ρmin) and (cmax−cmin); here they have (ρmin−ρmax) and (cmin−cmax). At the upper HU end the literal equations give density −1200 kg/m³ and negative sound speed. That is nonphysical. Because the same conversion is applied to both rCT and sCT, the reported rCT/sCT agreement cannot catch the error — both pipelines share the same flawed skull map. So every absolute targeting, FWHM, and pressure-error number is computed on nonphysical tissue properties. If this is a typesetting mistake, the authors need to say so and rerun; if not, the central claim is unsupported.\n\nOther issues are less severe but still real. Training on 17 subjects, testing on 5, no external cohort. No error bars on Table 1 MAE, and the Table 1 header 'Skull MAE (0≤HU≤2000)' contradicts the text saying skull region is HU>2000. The abstract claims 'sub-millimeter targeting deviation' while Figure 8B shows mean longitudinal errors of 1.5–2.6 mm; only transverse errors are sub-mm. The abstract also says RS-ASM runs in 34 s, while Figure 8E reports 345±85 s — a factor of ten. No code or data are released.\n\nThis is a fixable paper, not a hopeless one. The six-pipeline comparison is worth preserving, and the idea of testing sCT against rapid solvers on the same skulls is right. But the physical mapping has to be corrected and the simulations rerun, and the internal inconsistencies resolved before any of the accuracy claims can be trusted.\n\nI'd send it to peer review — a good referee will catch the sign error and the data problems. But I would not cite it or use its numbers in my own work until the authors address the equations and release artifacts.","headline":"A useful six-pipeline comparison undermined by a load-bearing sign error in the HU-to-acoustic conversion that makes the reported absolute accuracy figures invalid as written.","tokens_in":18229,"tokens_out":3025,"would_cite":false,"duration_ms":32270,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MRI-only brain ultrasound planning matches CT accuracy.","keywords":["transcranial focused ultrasound","synthetic CT","T1-weighted MRI","angular spectrum method","Rayleigh-Sommerfeld diffraction","deep learning","treatment planning","neuromodulation"],"falsifier":"Take the trained synthetic-CT network and apply it to T1-weighted images acquired on a different scanner vendor, field strength, or patient age group, comparing the output skull Hounsfield units against paired real CT; if the skull-region mean absolute error rises substantially above the reported ~178 HU or if full-wave simulations through the sCT shift the focus by more than 1 mm, the central claim fails. A physical test would be to measure the pressure field with a calibrated hydrophone through an ex vivo human skull using phase corrections computed from the sCT, and check that the measured focal position and width match the sub-millimeter and ~3.3–3.8 mm values claimed.","tokens_in":17177,"feed_emoji":"🧠","tokens_out":7205,"duration_ms":72914,"temperature":0.7,"pith_summary":"This paper claims that transcranial focused ultrasound (tFUS) treatment planning can be done entirely from MRI, without CT, by synthesizing a CT-like image of the skull from a T1-weighted MRI and feeding it into fast acoustic simulation. The authors train a deep-learning network to produce synthetic CT (sCT) Hounsfield units from MRI, then combine the sCT with two accelerated propagation methods that reduce simulation time from roughly 55 minutes to about 3 minutes or 34 seconds. Across five skull models, sCT-based simulations kept targeting within a millimeter of CT-based gold standard, produced focal widths of 3.3–3.8 mm, and stayed below 0.2 normalized pressure error. If these results hold in clinic, patients could receive repeated, radiation-free tFUS planning that fits into a single session.","feed_headline":"MRI-only brain ultrasound planning matches CT accuracy","feed_subtitle":"Synthetic CT from T1-weighted MRI keeps targeting within a millimeter and cuts simulation time by up to 94 percent.","key_machinery":"The argument rests on two pieces of machinery. The first is a synthetic-CT generator: an encoder-decoder U-Net with four strided 3D convolutions, two vision-transformer blocks, and subpixel upsampling, trained with a loss that blends global, bone-masked, and brain-masked mean absolute error with global and brain SSIM terms. Converting the generated Hounsfield units into density, sound speed, and attenuation via the Marsac bone model turns the sCT into an acoustic medium for simulation. The second is the hybrid propagation scheme: both accelerated pipelines first compute the transducer's field at a plane near the skull (one using a full-wave solver, the other using a Rayleigh–Sommerfeld diffraction integral), then propagate that plane wavefield through the skull with the angular spectrum method, which decomposes the field into plane waves and applies a phase factor in the spatial-frequency domain. This split avoids time-stepping through the skull, which is why the runtime drops by roughly 90–94% while keeping focal geometry comparable to full-wave simulation.","core_discovery":"The central discovery is that a 3D U-Net with transformer blocks, trained on 17 paired T1-weighted MRI and CT volumes, can generate a synthetic CT whose Hounsfield-unit values—especially in the skull—are accurate enough to drive acoustic simulations for transcranial focused ultrasound. The paper validates this by comparing six pipelines: full-wave and hybrid angular-spectrum solvers, each run on real CT and on synthetic CT. The sCT-based full-wave fields reproduce the CT-based fields with normalized pressure differences below 0.2 outside the focal zone, and the hybrid solvers (k-Wave plus angular spectrum, and Rayleigh–Sommerfeld plus angular spectrum) keep sub-millimeter lateral targeting error while cutting runtime to 187 s and 34 s, respectively, from about 3320 s for the gold-standard solver. The authors conclude that MRI-derived sCT combined with rapid solvers enables fast, accurate, and radiation-free tFUS planning.","pith_inferences":["Beyond the paper, the same sCT-plus-rapid-solver stack could be tested on MR-guided blood–brain barrier opening or histotripsy sessions, where repeated planning per session makes the radiation-free workflow most valuable.","Beyond the paper, the ~34-second RS-ASM runtime suggests that interactive planning—where a clinician moves the transducer and sees the focus update in near real time—is now within reach, a use case the paper does not demonstrate.","Beyond the paper, the claim would be strengthened by a physical check: measuring the focal field through an ex vivo skull with a hydrophone using sCT-derived phase corrections, to confirm the simulated sub-millimeter targeting survives real propagation."],"forward_implications":["tFUS treatment planning can be performed with MRI alone, eliminating CT scanning and its ionizing radiation for patients who need repeated or longitudinal neuromodulation sessions.","Simulation time falls from about 55 minutes to roughly 3 minutes with the k-Wave–angular-spectrum hybrid and to about 34 seconds with the Rayleigh–Sommerfeld hybrid, making per-patient planning practical in a clinical workflow.","Deep targets such as the thalamus and shallower 4-cm targets are both simulated with sub-millimeter lateral targeting error and focal widths of 3.3–3.8 mm, consistent with CT-based gold standard.","Because sCT-based and CT-based pipelines agree in focal geometry, the sCT pipeline can serve as a drop-in replacement for CT in existing tFUS planning systems."],"supporting_citations":[{"why":"Supplies the conversion from CT Hounsfield units to skull density, sound speed, and attenuation used to turn sCT into an acoustic model.","marker":"[41]"},{"why":"Prior open-source T1-to-pseudo-CT method that this network builds on (subpixel upsampling) and is compared against in the accuracy table.","marker":"[54]"},{"why":"Prior demonstration that deep-learning sCT can replicate CT accuracy for ultrasound therapy planning, motivating the sCT approach.","marker":"[55]"},{"why":"Prior validation of MRI-derived sCT for tFUS planning in independent test cases, which this paper extends by adding rapid solvers.","marker":"[60]"},{"why":"Provides the angular-spectrum rapid beam simulation framework that the hybrid propagation methods are built on.","marker":"[63]"},{"why":"Introduces the hybrid angular spectrum method for transcranial phase correction, the basis of the kW-ASM pipeline.","marker":"[65]"},{"why":"The U-Net architecture that the synthetic-CT generator is based on.","marker":"[74]"},{"why":"The full-wave solver used as the gold-standard reference for all accuracy comparisons.","marker":"[76]"}],"fun_headline_variants":["MRI-only tFUS planning: sub-mm accuracy, no CT, in seconds","Synthetic CT from MRI matches CT targeting for ultrasound","Radiation-free brain ultrasound planning: MRI-only matches CT","Fast, precise tFUS planning without CT using MRI-derived sCT","MRI-based sCT makes focused ultrasound targeting fast and exact"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes that T1-weighted MRI carries enough information, after skull stripping, for the trained network to reconstruct skull geometry and bone density in a new patient, even though cortical bone is nearly silent on T1-weighted MRI and the network was trained on only 17 subjects.","fun_headline_variants_meta":{"raw":{"variants":["MRI-only tFUS planning: sub-mm accuracy, no CT, in seconds","Synthetic CT from MRI matches CT targeting for ultrasound","Radiation-free brain ultrasound planning: MRI-only matches CT","Fast, precise tFUS planning without CT using MRI-derived sCT","MRI-based sCT makes focused ultrasound targeting fast and exact"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000497,"raw_usage":{"total_tokens":2458,"prompt_tokens":987,"completion_tokens":1471,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":1383}},"tokens_in":603,"tokens_out":1471,"duration_ms":16957,"temperature":1.0,"reasoning_tokens":1383,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:12:30.115640+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained synthetic-CT network and apply it to T1-weighted images acquired on a different scanner vendor, field strength, or patient age group, comparing the output skull Hounsfield units against paired real CT; if the skull-region mean absolute error rises substantially above the reported ~178 HU or if full-wave simulations through the sCT shift the focus by more than 1 mm, the central claim fails. A physical test would be to measure the pressure field with a calibrated hydrophone through an ex vivo human skull using phase corrections computed from the sCT, and check that the measured focal position and width match the sub-millimeter and ~3.3–3.8 mm values claimed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the conversion from CT Hounsfield units to skull density, sound speed, and attenuation used to turn sCT into an acoustic model."},{"cited_title":"& Hus�nx, R","cited_arxiv_id":null,"evidence_quote":"Prior open-source T1-to-pseudo-CT method that this network builds on (subpixel upsampling) and is compared against in the accuracy table."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior validation of MRI-derived sCT for tFUS planning in independent test cases, which this paper extends by adding rapid solvers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the angular-spectrum rapid beam simulation framework that the hybrid propagation methods are built on."},{"cited_title":"& Jing, Y","cited_arxiv_id":null,"evidence_quote":"Introduces the hybrid angular spectrum method for transcranial phase correction, the basis of the kW-ASM pipeline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The U-Net architecture that the synthetic-CT generator is based on."},{"cited_title":"S., Dalca, A","cited_arxiv_id":null,"evidence_quote":"The full-wave solver used as the gold-standard reference for all accuracy comparisons."}],"review_version":1}