{"id":"3ead0777-98f2-4cf5-90f9-8dc4e64909e4","arxiv_id":"2607.24529","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Token-conditioned NCA adapters in a lightweight encoder–decoder raise ultrasound segmentation Dice and boundary metrics on HC18/CCA/PSFHS and stay competitive under African domain shift.","lead":"LANCANet is a U-Net-style ultrasound segmenter that inserts token-conditioned Neural Cellular Automata adapters to refine features with global structure tokens. It posts strong Dice on fetal-head and carotid benchmarks and holds up under African domain shift, aiming at portable ultrasound.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The accuracy-and-efficiency claim fails specifically against SegFormer: LANCANet's DSC advantage is statistically indistinguishable (Table VI, p=0.39/0.24; HC18 p=0.46) while it is strictly worse on every efficiency axis, and SegFormer beat it on both external sets.","rationale":"The reader's weakest_assumption identified the same soft spot — the efficiency-for-POCUS framing undercut by Table V — and I agree it is the load-bearing concern. I sharpen it in two ways. First, the problem is not merely that LANCANet is slower than UNeXt/EGE-UNet/LB-UNet; it is that the strongest baseline, SegFormer, dominates it on all three efficiency axes while being statistically indistinguishable on accuracy (by the paper's own Table VI) and beating it on both external datasets. That combination directly negates the conjunctive strongest_claim rather than just softening it. Second, the efficiency measurement itself has internal inconsistencies (constant GFLOPs across T in Table VIII despite iterative NCA compute; FPS contradicting latency in the same table; a text/table mismatch on T=4 latency), so the efficiency comparison may not even be reliably measured. I do not recommend moving the verdict from CONDITIONAL: the accuracy contribution is well-supported by ablations (Table VII shows the Token Adapter and NCA blocks are necessary), external African validation is genuinely valuable and unusual for this venue, non-significant results are disclosed rather than hidden, and the Discussion partially concedes the latency issue. The appropriate resolution is exactly what the reader prescribed: temper the efficiency/deployment language, fix the Table VIII profiling, and ship permanent code. If the multi-seed test shows the SegFormer gap is pure noise, the novelty rests on the NCA-adapter mechanism itself (novelty 5.5 seems fair) rather than on any practical deployment advantage.","tokens_in":20096,"tokens_out":2405,"duration_ms":83005,"concrete_test":"Two-part check. (1) Re-profile Table VIII with a standard FLOP counter (fvcore/thop) and wall-clock timing at T=1,2,4,8: if GFLOPs does not scale with T or the FPS/latency mismatch persists, the efficiency numbers in Table V are unreliable and must be re-measured for all baselines. (2) Re-run LANCANet (T=2 and T=4) vs SegFormer with 5 random seeds on HC18/CCA/PSFHS, paired per-image DSC tests, plus latency on one actual edge target (e.g., Jetson Orin Nano or a POCUS-grade ARM CPU). If the DSC gap stays within seed noise and edge latency remains >2.5x SegFormer's, the dual accuracy+efficiency claim should be narrowed to an accuracy claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is conjunctive: competitive/superior accuracy AND suitability for resource-constrained deployment versus recent lightweight CNN/transformer methods. The paper's own tables break both conjuncts against one baseline. Table VI shows the average-DSC gain over SegFormer is not significant (p=0.39 DSC, p=0.24 HD95; HC18 p=0.46, PSFHS PS p=0.51, FH p=0.33), so \"superior\" reduces to a 0.33-point mean-DSC edge within noise. Meanwhile Table V shows LANCANet is strictly worse than SegFormer on efficiency: 6.70M vs 3.71M params, 27.69 vs 10.42 GFLOPs, 3.49 vs 12.15 FPS — in fact the slowest and largest model in the entire efficiency table, behind even UNet (33.26 FPS). SegFormer also won both external datasets (KEN-FH 92.90 vs 89.63 DSC; AFR-FH 93.01 vs 90.64), weakening the \"trained-from-scratch robustness\" selling point, though the authors are transparent about SegFormer's ImageNet pretraining. A secondary internal inconsistency compounds the efficiency reporting: Table VIII holds GFLOPs constant at 27.69 across T=1,2,4,8 even though NCA iterations must scale compute, and the FPS column contradicts the latency column (251.49 ms at T=1 implies ~3.98 FPS, not 3.49; 286.34 ms at T=4 implies ~3.49, not 3.64), and the text cites 274.70 ms at T=4 where the table says 286.34. The accuracy results and ablations look sound and honestly reported (non-significant p-values are disclosed), so this is a framing/positioning problem, not a correctness error — but it is load-bearing for the abstract and conclusion sentences that assert maintained computational efficiency for resource-constrained clinical deployment.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper proposes LANCANet, an ultrasound segmentation network that inserts token-conditioned Neural Cellular Automata (NCA) adapters into the two deepest encoder levels of a lightweight ConvMixer/transformer U-shaped architecture. Structure tokens are extracted by adaptive pooling and attention, mean-pooled, and injected into the NCA state via FiLM-style (γ, β) modulation before each of T iterative NCA updates. The method is evaluated on HC18, CCA, and PSFHS against nine baselines, with external validation on two African fetal-head datasets (KEN-FH, AFR-FH) under domain shift, paired t-tests, module and T-step ablations, and a parameter/GFLOP/CPU-FPS efficiency comparison. The authors report the best DSC on HC18 (96.62%) and CCA (92.86%), the best pubic-symphysis DSC on PSFHS (79.00%), competitive external performance despite training from scratch, and claim a favorable accuracy–efficiency balance suitable for resource-constrained deployment.","tokens_in":20607,"tokens_out":3350,"duration_ms":114036,"significance":"If the results hold, the work is a useful contribution: it is, to my knowledge, the first integration of token-conditioned NCA adapters (extending the AdaNCA idea of [24]) into ultrasound segmentation, and the evaluation is unusually complete for this literature — three in-domain datasets spanning fetal head, carotid, and multi-class intrapartum anatomy, two genuinely external African test sets under population/device shift, paired t-tests with non-significant results honestly disclosed in red, module and iteration-count ablations, and an anonymized code release. The accuracy numbers themselves (Tables I–III) are credible and the ablations support the claim that the NCA refinement is doing real work (Table VII: removing the Token Adapter or NCA TGB costs 1–2 DSC points; removing the Lite Transformer costs little). The weakness is not the accuracy evidence but the efficiency and \"superiority\" framing: the paper's own Table V shows LANCANet is the largest and slowest model in the comparison, and its headline advantage over SegFormer is statistically indistinguishable (Table VI). These are framing/positioning problems on top of sound empirical work, plus one table (VIII) with internal","major_comments":[{"comment":"The efficiency claim is contradicted by the paper's own Table V. The abstract ('minimal computational cost'), §V-E ('remaining a compact and computationally efficient model'), and the Conclusion ('favorable balance ... for resource-constrained clinical deployment') frame LANCANet as lightweight and deployment-ready. Table V shows the opposite within the comparison set: LANCANet has the most parameters (6.70M, vs 3.71M for SegFormer, 1.47M for UNeXt, 0.05–0.32M for EGE-UNet/LB-UNet/MK-UNet), the second-highest GFLOPs (27.69, vs 10.42 for SegFormer), and the lowest throughput (3.49 FPS — slower even than UNet at 33.26 FPS and ~3.5× slower than SegFormer at 12.15 FPS). The statement in §V-E that LANCANet has 'a comparable model size' to SegFormer is not supported (1.8× the parameters, 2.7× the GFLOPs, 3.5× the latency). The paper does not need to win on efficiency to be valuable — its contr","section":"§V-E, Table V; Abstract; Conclusion"},{"comment":"The headline 'superior' claim over the strongest baseline is not statistically supported, and the strongest baseline also wins the external evaluation. Table VI reports the average-DSC gain over SegFormer as non-significant (p=0.39 DSC, p=0.24 HD95), and Table IV shows non-significant differences on HC18 (p=0.46), PSFHS PS (p=0.51), and FH (p=0.33). Meanwhile Table III shows SegFormer is best on both external sets (KEN-FH 92.90 vs 89.63 DSC; AFR-FH 93.01 vs 90.64). The authors deserve credit for disclosing these p-values, and the from-scratch vs ImageNet-pretrained asymmetry is a legitimate mitigating point — but it cuts against, not for, the abstract's 'competitive or superior' and the Conclusion's framing. Since SegFormer is also cheaper on every efficiency axis (Comment 1), the paper's differentiator against this one baseline currently rests on within-noise in-domain differences. The","section":"§V-A/§V-B, Tables III, IV, VI"},{"comment":"Table VIII is internally inconsistent in ways that suggest measurement or reporting errors in the efficiency numbers. (i) GFLOPs are reported as a constant 27.69 for T=1,2,4,8, but the NCA update (Eq. 4) is applied T times over the feature grid, so compute must grow with T unless the GFLOPs figure excludes the NCA iterations — if so, that must be stated, since the same 27.69 figure is used in Table V as the model's headline complexity. (ii) The FPS and latency columns are mutually inconsistent: 251.49 ms at T=1 implies ~3.98 FPS, not the reported 3.49; 286.34 ms at T=4 implies ~3.49 FPS, not the reported 3.64; and T=1 and T=2 report identical FPS (3.49) despite a 34 ms latency difference. (iii) The text cites 'latency rises from 274.70 ms at T=4' but Table VIII says 286.34 ms at T=4 (274.70 matches no row). These numbers feed directly into the efficiency claims of Comment 1, so they must","section":"§V-F2, Table VIII"},{"comment":"The default configuration and the ablation conclusion disagree. §IV-B sets T=2 as the default and Tables I–III report all main results at T=2, but §V-F2 concludes 'T=4 offers the best balance between segmentation performance and inference efficiency on the PSFHS dataset' — indeed at T=4 the FH DSC rises from 85.63 to 90.15 and average DSC from 82.32 to 84.17, which would change several Table II rankings (e.g., FH DSC 90.15 would be the best in the table rather than mid-pack). Either the main PSFHS results should be reported at the authors' own recommended operating point, or the discrepancy should be explained (e.g., T=2 generalizes better on HC18/CCA — if so, show it). As it stands, the reader cannot tell which configuration the paper is actually proposing.","section":"§IV-B vs §V-F2, Tables II and VIII"}],"minor_comments":[{"comment":"§IV-A: 'we conducted experiments on four ultrasound datasets' is followed by a list of five (HC18, CCA, PSFHS, KEN-FH, AFR-FH).","section":"§IV-A"},{"comment":"Table I, HC18 row for LB-Unet: HD95 is reported as 2.94 (37.46); a mean of 2.94 with std 37.46 and that being far below all other methods' HD95 is implausible — likely a typo for 29.4 or 32.94. Please verify.","section":"Table I"},{"comment":"Table V, EGE-UNet: 0.05M parameters with 39.71 GFLOPs is an unusual ratio (higher GFLOPs than LANCANet at 125× fewer parameters); please verify the GFLOPs figure, especially since Fig. 9's x-axis only spans 0–40 GFLOPs and this point sits at the edge.","section":"Table V / Fig. 9"},{"comment":"§III-B2, Structure Token Extraction: N is used for both H×W ('flatten ... into N=H×W spatial cells') and the token count ('T ∈ R^{B×N×C}, where N=S²'). Please disambiguate the notation.","section":"§III-B2"},{"comment":"§III-B2: 'we extract a compact structural representation via adaptive pooling with output size S×S to get structure tokens T ∈ R^{B×N×C}' — the tensor named T collides with the refinement-step count T used throughout (Eq. 4, Table VIII). Rename one of them.","section":"§III-B2"},{"comment":"Eq. (2): the projections Q, K, V are applied but the dimensions of the attention output relative to the adaptive-pooled tokens (S² vs H×W queries) are not fully specified — a sentence clarifying whether tokens or cells are the queries would help reproduction.","section":"Eq. (2)"},{"comment":"The running header reads 'AUTHOR et al.: TITLE' and affiliation/author placeholders appear unanonymized in the arXiv text but the code link is anonymized (anonymous.4open.science); please make the anonymization policy consistent with the venue's review model.","section":"Throughout"},{"comment":"Efficiency is reported only as CPU FPS on a server-class AMD EPYC 9654P. Even a single additional number (e.g., Jetson Orin or a mid-range mobile SoC latency) would substantially strengthen, or honestly bound, the POCUS-deployment motivation.","section":"§V-E"},{"comment":"The unified training protocol is a strength, but SegFormer's ImageNet initialization (§IV-D) breaks the 'controlled comparison' symmetry for exactly the baseline that matters most; a from-scratch SegFormer row in Tables I–III would make the comparison fully controlled.","section":"§IV-D / Tables I–III"}],"recommendation":"major_revision","confidential_remarks":"The empirical work is solid and more complete than most papers in this space (three in-domain datasets, two external African cohorts, disclosed non-significant p-values, code release). The problem is positioning: the manuscript is written as an efficiency paper but the measurements describe an accuracy paper with a slow model, and the one baseline that beats it on both efficiency and external generalization (SegFormer) is statistically tied with it in-domain. I see no integrity issue — the authors disclose the inconvenient p-values — but the abstract/conclusion need to be brought into line with Tables III, V, and VI, and Table VIII's efficiency numbers look mis-measured or mis-transcribed. Novelty over AdaNCA [24] is incremental but adequate for an application-focused venue."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a competent methods paper that composes token-conditioned NCA adapters (structure tokens + FiLM + Med-NCA-style updates) on a Lite ConvMixer backbone for ultrasound. The in-domain accuracy gains on HC18/CCA and on the hard pubic-symphysis class are real and honestly tabled. The efficiency-for-resource-constrained-deployment claim is the soft spot and should be dialed back before anyone treats the abstract as settled.\n\nWhat is actually new is the finished system, not any single ingredient. Med-NCA, AdaNCA, FiLM, and ConvMixer-style blocks are prior. Putting token-guided NCA as a refinement adapter inside a U-shaped ultrasound segmenter, with T-step and module ablations plus external African tests (KEN-FH, AFR-FH), is a legitimate engineering contribution for the POCUS/obstetric niche. Multi-dataset protocol, paired t-tests with non-significant p-values disclosed, and clear boundary metrics (HD95/ASD) are done properly. Code link is there (anonymous for now).\n\nSoft spots, in proportion: Table V undercuts the efficiency narrative. LANCANet is the largest and slowest model in their own CPU comparison (6.7M params, 27.7 GFLOPs, ~3.5 FPS) versus UNeXt/EGE/LB-UNet and even SegFormer. Average DSC edge over SegFormer is not significant (p≈0.39), and SegFormer wins both external sets—though the authors are transparent that SegFormer used ImageNet init. Table VIII holding GFLOPs fixed across T while latency moves, plus small FPS/latency inconsistencies, is sloppy reporting, not a correctness failure. KEN-FH is small. None of that invents the Dice numbers; it weakens the conjunctive “accurate and efficient for edge” pitch.\n\nWho it is for: people building lightweight US segmenters and anyone who cares about NCA-as-adapter rather than NCA-as-backbone. Worth a serious referee if efficiency claims are tempered and permanent code/data splits land. I would engage the work, not build a project around the efficiency story as written.\n\nRecommendation: send to peer review; ask for honest repositioning on compute and a clean efficiency table.","headline":"Solid ultrasound segmentation architecture paper with real multi-site evidence; the accuracy story holds, the “lightweight for POCUS” framing does not.","tokens_in":21523,"tokens_out":574,"would_cite":false,"duration_ms":17864,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Token-guided cellular automata refine ultrasound features enough to beat recent lightweight segmenters on fetal head and carotid images while staying compact enough for bedside use.","keywords":["Neural Cellular Automata","Ultrasound","Segmentation","Lightweight Model","Point-of-Care Ultrasound","Token FiLM","Fetal head","Domain shift"],"falsifier":"Run the same CPU latency and Dice comparison on a representative handheld ultrasound device or edge board: if LANCANet is slower or less accurate than the lighter baselines under that fixed power and memory budget, the efficiency-plus-accuracy claim fails.","tokens_in":21115,"feed_emoji":"🩺","tokens_out":780,"duration_ms":18245,"temperature":0.7,"pith_summary":"Point-of-care ultrasound needs accurate organ outlines on devices that cannot run heavy neural nets. This paper argues that the missing piece is not a bigger backbone but a cheap iterative refiner: structure-aware tokens condition Neural Cellular Automata so local cell updates improve weak boundaries without large compute. The resulting network, LANCANet, matches or exceeds recent lightweight CNN and transformer segmenters on fetal-head, carotid, and pubic-symphysis datasets, and stays competitive on African fetal-head scans it never saw in training. A sympathetic reader cares because that combination—better contours plus a design aimed at low-resource clinics—is exactly what bedside biometric tools need.","feed_headline":"Token-guided cellular automata top lightweight ultrasound segmenters","feed_subtitle":"Fetal head and carotid Dice lead the pack; the design targets bedside devices under domain shift","key_machinery":"Token-conditioned NCA adapters: structure tokens summarize global anatomy, Token FiLM modulates the cellular state, and shared local NCA update rules iteratively refine features for T steps—preserving backbone semantics while sharpening boundaries at limited cost.","core_discovery":"LANCANet shows that attaching token-conditioned NCA adapters to a lightweight encoder–decoder yields competitive or superior ultrasound segmentation versus recent efficient CNN- and transformer-based methods, with top Dice on HC18 (96.62%) and CCA (92.86%) and best pubic-symphysis Dice on PSFHS (79.00%), while remaining usable under domain shift when trained from scratch.","pith_inferences":["If token FiLM is the main carrier of global context, similar adapters might lift other noisy local-update models (e.g., diffusion or recurrent refiners) in ultrasound without full attention stacks.","The external African results suggest token-NCA may reduce reliance on large-scale natural-image pretraining for cross-device fetal biometry.","A natural next measurement is on-device FPS and battery draw; the paper’s CPU numbers alone do not settle bedside feasibility."],"forward_implications":["Bedside fetal-head and carotid tools can adopt token-NCA refinement instead of larger transformers when training data are limited.","Pubic-symphysis segmentation in labor ultrasound can improve without abandoning a compact architecture.","Models trained only on standard public sets can still transfer usefully to multi-country African fetal scans without ImageNet pretraining.","Ablation of iteration count becomes a practical knob: moderate T improves boundaries; too many steps waste latency."],"fun_headline_variants":["Token-conditioned NCA adapters lift lightweight US segmentation Dice","LANCANet leads HC18 and CCA Dice with token-guided NCA refinement","Structure-aware tokens steer NCA for bedside ultrasound boundaries","Scratch-trained LANCANet holds fetal-head Dice under African domain shift","Token FiLM NCA refinement tops efficient US segmenters on HC18/CCA"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The accuracy gains still leave the model efficient enough for real point-of-care and edge devices, even though measured size and CPU speed are worse than several lighter baselines.","fun_headline_variants_meta":{"raw":{"variants":["Token-conditioned NCA adapters lift lightweight US segmentation Dice","LANCANet leads HC18 and CCA Dice with token-guided NCA refinement","Structure-aware tokens steer NCA for bedside ultrasound boundaries","Scratch-trained LANCANet holds fetal-head Dice under African domain shift","Token FiLM NCA refinement tops efficient US segmenters on HC18/CCA"]},"model":"grok-4.5","effort":"low","cost_usd":0.001875,"raw_usage":{"total_tokens":959,"prompt_tokens":861,"num_sources_used":0,"completion_tokens":78,"cost_in_usd_ticks":18748000,"prompt_tokens_details":{"text_tokens":861,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":20,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":861,"tokens_out":78,"duration_ms":2398,"temperature":1.0,"reasoning_tokens":20,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T12:18:39.688554+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same CPU latency and Dice comparison on a representative handheld ultrasound device or edge board: if LANCANet is slower or less accurate than the lighter baselines under that fixed power and memory budget, the efficiency-plus-accuracy claim fails.","supporting_citations":[],"review_version":1}