{"id":"044abd94-0346-48f7-b990-0eebd83515ef","arxiv_id":"2607.18283","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Frozen DINOv2 + LoRA + YOLO head, trained with FedAvg, localizes the fetal corpus callosum at mAP@50 0.857 with 8.5x lower communication than full fine-tuning.","lead":"This paper presents FedCC, a federated-learning system that finds the corpus callosum in fetal ultrasound images using a frozen DINOv2 backbone with lightweight LoRA adapters and a YOLO-style head. It reports strong detection performance on a new three-center dataset while exchanging only 2.9 million trainable parameters per round, about one-eighth the cost of full fine-tuning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline C3 advantage may be statistical noise: the test set for Client 3 has only 3 patients and no repeated runs/error bars, so the claimed LoRA superiority over full fine-tuning is not yet established.","rationale":"The reader's verdict CONDITIONAL is appropriate. My stress-test identifies a more specific load-bearing concern than the reader's weakest assumption: rather than the absence of a held-out center (an external validity issue), the central claim is internally fragile because the per-client test sets, especially C3, contain very few patients (3) and no repeated runs or error bars are reported. This directly affects the reliability of the reported mAP advantage on C3 and hence the average. The reader did note missing error bars but framed the weakest assumption as generalizability to unseen centers. I partially agree: the held-out-center issue is a real limitation, but the more immediate threat to the central claim is statistical. The paper has strengths: a broad comparison across backbones, adaptation strategies, and aggregation methods, and a plausible parameter-count reduction. The internal data inconsistency in Section 3.3 vs Table 2 and the suspicious repeated Precision/Recall/F1 values in Table 5 reinforce the need for public artifacts and careful reporting, but they do not by themselves disprove the method. The recommended verdict remains CONDITIONAL: the claim is plausible but not yet verified without repeated splits, confidence intervals, and consistent data descriptions.","tokens_in":23283,"tokens_out":10387,"duration_ms":103203,"concrete_test":"Re-run the FedAvg DINOv2base comparison (Proposed vs Full FT vs Freeze) under K=5 or more patient-level resampled train/validation/test splits, stratifying by client and keeping all frames of a patient in one split. Report the mean and standard deviation of C3 mAP@50 and average mAP@50, plus a paired bootstrap 95% confidence interval for the Proposed minus Full FT difference. If the C3 difference interval includes zero, the central claim that LoRA outperforms full fine-tuning on the hardest client is not statistically supported. Also correct the Section 3.3 vs Table 2 Client 3 split discrepancy before recomputation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparative claim rests heavily on Client 3, where the proposed DINOv2+LoRA under FedAvg reaches mAP@50 0.788, versus 0.579 for full fine-tuning and 0.507 for freezing (Table 4). This 0.2+ margin on the 'most challenging client' drives both the average (0.857 vs 0.721) and the narrative of robustness across heterogeneous domains. However, Table 2 shows Client 3's test set comprises only 710 images from 3 patients (Section 3.3 text even gives inconsistent numbers: 683 images/3 patients). With only 3 test patients, all frames from the same patient are highly correlated, so the effective sample size for the C3 metric is essentially 3. The paper reports no repeated federated runs, no seeds, and no confidence intervals anywhere. Thus the reported C3 gap — and the average mAP gap that depends on it — could be a sampling artifact rather than a genuine property of the method. This concern is internal to the evaluated data, not just external generalization, and it directly threatens the headline claim that LoRA adaptation outperforms full fine-tuning. The parameter-count arithmetic (2.9M vs 24.4M) is credible, but the empirical performance advantage needs statistical support.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedCC, a federated learning framework for corpus callosum (CC) localization in fetal ultrasound. It combines a frozen DINOv2 backbone with LoRA adapters and a lightweight YOLO-style detection head, exchanging only LoRA/head parameters across clients. On a self-collected multi-center dataset of 10,970 frames from 58 patients at three sites, the authors report an average mAP@50 of 0.857 and an F1-score of 0.803 under FedAvg, outperforming full fine-tuning and encoder-freezing baselines, while reducing trainable parameters from 24.4M to 2.9M (an 8.5x parameter reduction). The paper includes extensive comparisons across backbones (DINOv2, SAM, UltraSAM, UltraFedFM), adaptation strategies, and aggregation methods.","tokens_in":23578,"tokens_out":6239,"duration_ms":53803,"significance":"If the empirical claims hold, the paper provides a practical and privacy-preserving recipe for adapting large visual foundation models in federated medical imaging: a frozen backbone with low-rank adapters and a compact detection head can match or exceed full fine-tuning at a fraction of the communication cost. The newly collected multi-center fetal US dataset and the systematic comparison against multiple foundation-model baselines are useful contributions. However, the central comparison currently rests on point estimates without any measure of variance, and the most challenging client's test set is very small; the communication-efficiency claim is inferred from parameter counts rather than measured. These issues need to be addressed before the comparative conclusions are fully convincing.","major_comments":[{"comment":"The headline result (DINOv2base+LoRA under FedAvg: mAP50 0.857 vs 0.721 for full FT) is based on a single run per configuration. No seeds, repeated runs, confidence intervals, or significance tests are reported anywhere. The largest per-client gap is on Client 3 (mAP50 0.788 vs 0.579), but Client 3's test set has only 3 patients (Table 2 reports 710 images, while Section 3.3 text gives 683 images). Since frames from the same patient are highly correlated, the effective sample size for this client-level metric is effectively 3. The observed advantage on C3, and hence the average gap, could be a sampling artifact. Please provide variance estimates (e.g., multiple federated runs with different seeds, per-patient bootstrap confidence intervals) or temper the 'outperforms' claim.","section":"§5, Table 4; §3.3, Table 2"},{"comment":"There is an internal inconsistency in the data split for Client 3. Section 3.3 states 2,165 training / 472 validation / 683 test images, while Table 2 reports 1,971 / 445 / 710 for the same client. This discrepancy makes it unclear which split was actually used to produce Table 4, and it must be corrected for reproducibility.","section":"§3.3 vs. Table 2"},{"comment":"Several table entries are arithmetically inconsistent. In Table 4, the FedProx DINOv2base-Proposed row reports F1 values exactly equal to the Precision values (0.752 / 0.922 / 0.683), which is inconsistent with the stated Precision/Recall pairs (e.g., C1 Precision=0.752, Recall=0.739 gives F1≈0.745, not 0.752). In Table 5, the Precision, Recall, and F1 columns are identical for both the 3-scale and 2-scale detection-head rows, which is implausible and likely a table-generation error. These errors undermine confidence in the numerical results and must be corrected and re-verified.","section":"§5, Tables 4 and 5"},{"comment":"The abstract and discussion claim an '8.5x reduction in communication cost' based solely on the trainable parameter count (2.9M vs 24.4M). In federated learning, per-round communication also includes optimizer state, aggregation overhead, encoding, and potential secure-aggregation costs; actual bytes transferred are not measured. The claim is defensible as a parameter-count ratio, but it should be stated as such, with a measurement of actual communication or a caveat that it is an approximation.","section":"Abstract; §6; Figure 6"}],"minor_comments":[{"comment":"Equation numbering restarts at (1) in Section 4.2 after already being used in Section 3.1. Renumber equations sequentially throughout the manuscript.","section":"§4.2"},{"comment":"The paper appropriately acknowledges the simulated federated setting and the lack of an external held-out center. However, the abstract and Section 6 state 'robustness across heterogeneous domains' as a finding. Since all test sets are drawn from the same three sites used for training, the claim should be qualified as robustness to the domain shift present among the participating sites, not to unseen centers.","section":"§6, Limitations"},{"comment":"The caption says ground truth is in red and predictions in green, but the figure appears to be in color and may not be accessible to color-blind readers. Consider using different line styles or adding a grayscale-visible distinction.","section":"Figure 5"},{"comment":"Typo: 'Aknowledgements' should be 'Acknowledgements'.","section":"§8"},{"comment":"The statement that code will be made public upon acceptance is useful, but consider releasing at least the evaluation protocol and hyperparameter configuration for reproducibility even before acceptance.","section":"Code Availability"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is sound and the experimental sweep is extensive, but the statistical grounding is currently too weak for the central comparative claim. The internal inconsistencies in the data split and tables are fixable, but the Client 3 test-set size and the absence of any uncertainty quantification are more fundamental. I would be comfortable with acceptance after the authors provide variance-aware results and recalibrate the communication-cost and generalization claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a legitimate application paper—federated adaptation of a frozen DINOv2 backbone with LoRA and a YOLO head for corpus callosum bounding-box detection in fetal ultrasound, plus a new 10,970-frame, 58-patient, 3-site dataset. It does not introduce new machinery; the contribution is the combination and the empirical comparison. The headline result—LoRA under FedAvg reaching 0.857 mAP@50 and 0.803 F1, beating full fine-tuning (0.721/0.733) and freezing—is plausible, but I would not treat it as established yet.\n\nCredit where due: the dataset is a real asset, the comparison tables are extensive across backbones (DINOv2 variants, SAM, UltraSAM, UltraFedFM, YOLO26), and the parameter arithmetic (2.9M vs 24.4M trainable, 8.5x communication reduction) is sound as a proxy. The paper also acknowledges its main limitations in Section 6: simulated FL with full participation, normal pregnancies only, bounding-box detection only, and a three-site sample.\n\nThe soft spots are exactly the ones the reader flagged, and they land. First, no repeated seeds, no confidence intervals. The per-client C3 numbers come from a test set of 3 patients (710 images). Those frames are highly correlated, and the 0.2 mAP advantage of LoRA over full FT on C3 drives the average. Without variance estimates, that gap could be sampling noise. This is a load-bearing weakness in the central comparative claim, not a cosmetic one. Second, there is an internal inconsistency between Section 3.3's C3 split (2,165/472/683) and Table 2 (1,971/445/710). Table 2 sums to the stated total of 3,126, so the text numbers look like a typo, but the paper should not leave both in. Third, there is no truly held-out center; all test frames come from the same three sites used for training. The \"robustness across heterogeneous domains\" claim is therefore within-domain robustness, and the paper's own limitations paragraph says external validation is needed. Fourth, code and data are unavailable, so the numbers cannot be independently verified; the dataset is not even accessible as a benchmark yet.\n\nWho this is for: researchers working on FL + PEFT for medical imaging, especially ultrasound. They will find the empirical setup and the comparison matrix useful, and the dataset is valuable if it ever becomes available. It deserves a serious referee but needs substantial revision: repeated runs with error bars, a consistent data description, and a toned-down generalization claim.\n\nI would not desk-reject this; I would send it to review, but I would expect the reviewers to demand statistics and artifacts.","headline":"A useful FL+PEFT application to fetal CC localization with a new dataset, but the central LoRA advantage rests on unrepeated, small-sample numbers and needs statistical support before it can be trusted.","tokens_in":24121,"tokens_out":2935,"would_cite":false,"duration_ms":26223,"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":"Adapting a frozen DINOv2 backbone with low-rank LoRA modules under federated averaging reaches 0.857 mAP@50 for corpus callosum detection in fetal ultrasound, beating full fine-tuning while cutting transmitted parameters from 24.4M to 2.9M.","keywords":["federated learning","fetal ultrasound","corpus callosum localization","foundation models","parameter-efficient fine-tuning","low-rank adaptation","DINOv2","domain shift"],"falsifier":"Hold out an entire fourth clinical site or a never-seen ultrasound scanner, train FedCC only on the original three clients, and measure mAP@50 on that held-out site; if the score drops far below the in-federation average or below a full-fine-tuning baseline on the same held-out data, the robustness-across-domains claim is refuted.","tokens_in":23170,"feed_emoji":"🧠","tokens_out":7702,"duration_ms":67497,"temperature":0.7,"pith_summary":"The paper proposes FedCC, a federated-learning system that locates the corpus callosum in fetal ultrasound images, and argues that the right way to adapt a foundation model across hospitals is to freeze the backbone entirely and share only small low-rank adapters and a light detection head. In a simulated three-site federation using FedAvg, the frozen DINOv2 backbone with LoRA adapters reached an average mAP@50 (mean precision at 50% box overlap) of 0.857 and F1-score of 0.803, outperforming both full fine-tuning (0.721/0.733) and encoder freezing (0.715/0.704). The same configuration lifted the hardest client from 0.579 to 0.788 mAP@50 and cut per-round transmitted parameters from 24.4M to 2.9M, an 8.5x communication saving. The paper also reports that natural-image pretraining outperformed an additional fetal-ultrasound pretraining step, and that the federated model beat its own centralized counterpart trained on pooled data. If correct, this makes privacy-preserving multi-center training of fetal neurosonography tools substantially cheaper and more robust to scanner differences.","feed_headline":"Fetal-brain detector beats full fine-tuning with 8.5x fewer parameters","feed_subtitle":"LoRA adapters on a frozen DINOv2 reach 0.857 mAP@50 for corpus callosum across three ultrasound sites, without sharing images.","key_machinery":"The machinery is the pairing of a frozen DINOv2 backbone (a self-supervised vision transformer pretrained on diverse images) with LoRA adapters—low-rank matrices A and B inserted into the query and value projections so each weight update is ΔW=BA with rank r=8—plus a compact single-scale head built from YOLOv8's SPPF, convolution, C3f, and detection layers. Only the LoRA adapters and head weights are optimized and exchanged; the backbone stays on each client and is never transmitted, which is what produces the 8.5x communication reduction. The single-scale head matters as much as the adapters: replacing it with 2- or 3-scale heads dropped average mAP@50 from 0.857 to 0.757 and 0.741, respect","core_discovery":"FedCC's central claim is that parameter-efficient federated adaptation—a frozen self-supervised ViT backbone, LoRA matrices inserted into attention query and value projections, and a single-scale YOLO-style head with only 2.9M trainable parameters—is the best operating point for corpus callosum detection under inter-site domain shift. Under FedAvg, this configuration achieved 0.857 average mAP@50 and 0.803 F1-score on the three-site test split, ahead of full fine-tuning (0.721 mAP@50, 0.733 F1) and encoder freezing (0.715, 0.704). The improvement was largest on Client 3, the hardest domain, where mAP@50 rose to 0.788 versus 0.579 under full fine-tuning. The authors explain this as LoRA const","pith_inferences":["A direct test of the mechanism would compare FedCC against centralized training with explicit domain-balanced sampling; if the gap persists, the federated aggregation itself—not just seeing more data—is the active ingredient.","The same recipe likely transfers to other small fetal-brain landmarks (cavum septi pellucidi, cerebellar vermis) and to other ultrasound targets, since nothing in the architecture is corpus-callosum-specific apart from the data.","The current evidence cannot separate 'federated training is good' from 'three sites are not enough to expose a held-out domain': all test frames come from the same three sites, and the paper's Section 6 acknowledges real deployments may involve asynchronous updates and client dropout, which are untested.","A modest extension would measure the actual privacy/security properties—e.g., whether shared LoRA parameters can be inverted to recover image-level information—since the paper claims privacy preservation but provides no attack analysis."],"forward_implications":["Hospitals can collaboratively train a corpus-callosum detector without sharing ultrasound images, annotations, or backbone weights; only 2.9M adapter/head parameters move per round.","The 8.5x communication cut makes foundation-model adaptation practical in bandwidth-limited clinical networks, not just research clusters.","Federated aggregation with LoRA can outperform centralized training on pooled data for heterogeneous, imbalanced multi-site data, suggesting aggregation acts as structured multi-domain training.","For this task, the breadth of large-scale natural-image pretraining outweighs a subsequent fetal-ultrasound self-supervised pretraining step in the federated setting.","LoRA's built-in low-rank regularization may remove the need for proximal terms such as FedProx, simplifying federated deployment."],"fun_headline_variants":["Federated fetal-brain AI: 8.5x fewer params, higher accuracy","LoRA on frozen DINOv2: 8.5x fewer params for fetal brain detection","Private federated learning: 8.5x smaller model, better fetal brain AI","Fetal ultrasound: federated LoRA hits 0.857 mAP with 8.5x fewer params"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a simulated federation of three Italian sites with full client participation, synchronized rounds, and test sets drawn from those same three sites represents real privacy-preserving multi-center deployment; if a truly held-out scanner or protocol is added, the reported 0.857 mAP@50—especially the 0.788 on the hardest site—may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Federated fetal-brain AI: 8.5x fewer params, higher accuracy","LoRA on frozen DINOv2: 8.5x fewer params for fetal brain detection","Private federated learning: 8.5x smaller model, better fetal brain AI","Fetal ultrasound: federated LoRA hits 0.857 mAP with 8.5x fewer params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001633,"raw_usage":{"total_tokens":6411,"prompt_tokens":909,"completion_tokens":5502,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":5411}},"tokens_in":653,"tokens_out":5502,"duration_ms":35851,"temperature":1.0,"reasoning_tokens":5411,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T09:20:56.140160+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out an entire fourth clinical site or a never-seen ultrasound scanner, train FedCC only on the original three clients, and measure mAP@50 on that held-out site; if the score drops far below the in-federation average or below a full-fine-tuning baseline on the same held-out data, the robustness-across-domains claim is refuted.","supporting_citations":[],"review_version":1}