{"id":"6b1dc89b-9a83-4bfc-8c59-89b96c52aef1","arxiv_id":"2501.16992","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"FedEFM uses differentiable Earth Mover's Distance inside federated knowledge distillation to train an endovascular foundation model from heterogeneous, privacy-protected X-ray data silos.","lead":"This paper proposes FedEFM, a federated learning method that trains an endovascular imaging foundation model across hospitals without sharing patient data, using Earth Mover's Distance to handle situations where each hospital has different types of X-ray data. It reports large accuracy gains over standard federated learning baselines on catheter and guidewire segmentation and classification tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The EMD coefficient in Eq. 2 is undefined as written: Eq. 5's transport LP is infeasible for unnormalized feature maps, and Eq. 3 scores the complementary cost to the one optimized in Eq. 5.","rationale":"The reader's rejection was based on empirical inconsistencies; my pass finds that the theoretical object at the center of the method is also not well-formed. I agree with the reader's choice of weakest assumption: the EMD weight is load-bearing. My concern is slightly more fundamental: it is not merely that the metric might not be task-relevant, but that the LP in Eq. 5 is infeasible or dimensionally inconsistent, and Eq. 3 and Eq. 5 define different objectives. If a released implementation normalizes feature maps or uses a different solver, the equations must be corrected; without such a correction, the algorithm cannot be run as written. I also note the reader's other points are independently serious: Table II reports 98.2% while the 100% unseen-data ablation reports 84.9%; Table I lists VESSEL12, DRIVE, and SenNet as X-ray datasets though they are CT, retinal color, and phase-contrast tomography; and EISimulation appears in both foundation training and downstream evaluation despite the statement that training datasets are not used downstream. Any one of these would prevent verification of the SOTA claim. Because the central argument is unsupported both formally and empirically, the REJECT verdict stands; no change is needed.","tokens_in":12396,"tokens_out":5844,"duration_ms":58474,"concrete_test":"Reproduce the forward pass of Algorithm 1 for one silo: take a ViT trained as theta_i and a neighbor-trained theta_hat_i->j, feed the same batch, extract U and V, and attempt to solve Eq. 5 with a standard LP solver. Record whether a feasible x_tilde exists, whether total masses sum_p u_p and sum_q v_q match, and whether Eq. 3 evaluated at x_tilde equals the optimal LP objective of Eq. 5. If the LP is infeasible on any batch, or if the Eq. 3 and Eq. 5 objectives disagree, then the EMD weight used in Eq. 2 is not the quantity defined in the paper, and the reported 98.2%/97.5% gains cannot be attributed to the stated mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a differentiable EMD weight in Eq. 2 lets a silo learn from neighbor-trained teachers. This weight is not defined by the equations given. In Eq. 5, U and V are sets of C-dimensional embeddings u_p, v_q, but the decision variables x_pq are scalars; the equalities sum_p x_pq = v_q and sum_q x_pq = u_p are therefore dimensionally inconsistent unless u_p and v_q are intended to be scalar masses, which is never stated. Even treating each feature vector as a scalar mass, feasibility requires total supply sum_p u_p to equal total demand sum_q v_q; for two independently trained networks on the same input this equality will generically fail, so the LP has no solution and EMD(theta_i, theta_hat_i->j) is undefined. Moreover, Eq. 3 defines the reported EMD as sum_pq (1 - c_pq) x_tilde_pq while Eq. 5 minimizes sum_pq c_pq x_tilde_pq; the same optimal flow cannot maximize and minimize complementary objectives, so the quantity called EMD is not the optimum of the stated LP. Finally, the KKT Jacobian in Eq. 9 requires invertibility of J_x g; no argument is given, and for an infeasible or degenerate LP this matrix is singular. Since Eq. 2 multiplies the entire distillation gradient by this scalar, if the scalar is undefined or arbitrary the method's mechanism for learning from neighbors has no stated implementation, independent of the experimental numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedEFM, a federated learning scheme for training an endovascular foundation model. In each communication round, a silo's weights are transmitted to a neighbor, trained locally on the neighbor's data as an 'overseas expert,' and transferred back; the returned weights are then used as knowledge-distillation teachers at the original silo, with the distillation gradient scaled by a differentiable Earth Mover's Distance (EMD) between feature representations (Eqs. 2-11). The authors claim state-of-the-art classification accuracy (98.2% in centralized-style federated learning, 97.5% in decentralized federated learning) and improved downstream segmentation and classification over CLIP, SAM, and LVM-Med, together with robustness to unseen data.","tokens_in":12808,"tokens_out":5095,"duration_ms":47445,"significance":"If the proposed mechanism were well defined and the evaluations were sound, the idea of training a domain-specific foundation model across hospital silos without sharing raw data would be valuable for privacy-sensitive medical imaging. The paper also contributes a new robotic phantom dataset (EIPhantom) and labels for simulation data, which are potentially useful assets. However, the central EMD-based update is not defined by the equations as written, the headline accuracies are internally inconsistent, and the experimental corpora do not support the claimed endovascular X-ray setting. These issues prevent the result from being evaluated as stated, so the contribution cannot currently be accepted on the basis of the manuscript.","major_comments":[{"comment":"The EMD coefficient that scales the distillation gradient in Eq. (2) is not a well-defined quantity as written. In Eq. (5), u_p and v_q are C-dimensional embeddings, so the constraints sum_p x_pq = v_q and sum_q x_pq = u_p equate scalar decision variables with vectors; if instead u_p and v_q are intended to be scalar masses, this is never stated and contradicts the definition U,V in R^{H x W x C}. Even in the scalar-mass reading, feasibility requires sum_p u_p = sum_q v_q, which will generically fail for representations from two independently trained networks on the same input. In addition, Eq. (3) defines the reported EMD as sum_{p,q} (1 - c_pq) x_tilde_pq, while Eq. (5) minimizes sum_{p,q} c_pq x_tilde_pq; a single optimal flow cannot simultaneously maximize one objective and minimize its complement. Finally, Eq. (9) requires invertibility of J_x g, for which no argument is given and which fails for infeasible or degenerate linear programs. Since Eq. (2) multiplies the entire distillation gradient by this scalar, the proposed learning-from-neighbors mechanism has no stated implementation.","section":"Section III-B, Eqs. (3)-(5), (9)"},{"comment":"The numerical claim for the 100% unseen-data setting is internally inconsistent. Section IV-B states that the default setup is maintained at 100% unseen label corpus, and Table II reports FedEFM accuracy of 98.2% for CFL and 97.5% for DFL. Section IV-D, discussing Fig. 4, describes the same 100% unseen condition and reports an accuracy of 84.9% for 'our approach.' No explanation is offered for this large discrepancy, and no experimental details (task, dataset split, number of silos, evaluation metric) are given for Fig. 4, so the reader cannot tell which number is correct or what was actually measured.","section":"Table II vs. Fig. 4"},{"comment":"The training corpora do not match the paper's characterization. Table I labels all corpora as 'X-ray datasets related to endovascular intervention,' but VESSEL12 is lung CT vessel segmentation, DRIVE is retinal fundus photography, SenNet is phase-contrast tomography of human organs, and Medical Decathlon contains multiple modalities (including CT and MRI) unrelated to endovascular X-ray. More importantly, EISimulation appears both in the 'Federated Foundation Training' column of Table I and as a downstream fine-tuning benchmark in Table III and Fig. 5, directly contradicting the statement in Section IV-A that 'the datasets used to train the foundation model are not being used in downstream endovascular understanding tasks.' This makes the EISimulation downstream results circular for evaluating the foundation model.","section":"Section IV-A, Table I, Table III"},{"comment":"The experimental protocol is too underspecified to support the headline claim. The paper does not state the number of hospital silos, the dataset partition used to create the 100% unseen condition, the number of communication rounds K, the hyperparameters beta and T, or the evaluation split for the accuracy numbers. Given that the central claim is a new state-of-the-art result, these missing details are not merely presentational; they prevent the reader from checking whether the comparisons in Table II are fair or whether the reported numbers reflect the proposed method as defined.","section":"Section IV-B, Table II"}],"minor_comments":[{"comment":"The entry 'Drive [49] 8.028' appears to be a typo for 8,028 frames; if not, the frame count should be clarified.","section":"Section IV-A, Table I"},{"comment":"The first affiliation contains the typo 'Deparment' instead of 'Department.'","section":"Author affiliation"},{"comment":"The sentence about FW-Net contains a duplicated word: 'the authors in in [25] proposed...' should be 'the authors in [25] proposed...'.","section":"Section II"},{"comment":"The displayed softmax formula appears garbled: it should read Q^tau_k = exp(l_k/T) / sum_k exp(l_k/T), with an explicit division and summation.","section":"Section III-C, Eq. (12)"},{"comment":"The figure caption and text do not specify which task and metric are reported, nor whether these are classification accuracies on a single split; error bars or standard deviations over multiple runs should be provided.","section":"Fig. 4"},{"comment":"The citation for STAR [57] points to a paper on betweenness centrality, which does not appear to correspond to the STAR federated learning method; please verify the reference.","section":"References [57] and [58]"}],"recommendation":"reject","confidential_remarks":"For the editor: the paper's relationship to the DeepEMD work [42] is not clearly delineated; the differentiable EMD machinery appears to be imported from prior work, and the paper does not explain what is new beyond applying it in a federated distillation loop. Given the unresolved formal and empirical inconsistencies, I do not see a path to acceptance within the current manuscript's scope. A substantial rewrite that redefines the method, corrects the data characterization, and re-executes the experiments could change this assessment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's central mechanism, the EMD-weighted distillation, is not actually well-defined in the equations, and two of its headline numbers disagree for the same setting. That's a reject in current form, even though the problem it tackles is real and the authors put together a new dataset.\n\nThe good part: training a foundation model across hospital silos with genuinely disjoint data types is an important privacy-motivated problem. The EIPhantom collection (4,700 images) is a concrete contribution, and the paper compares against a reasonable set of FL baselines. The idea of using a distributional distance to weight which neighboring teacher's gradient should be trusted is plausible.\n\nNow the soft spots. First, the math in Section III-B doesn't type-check. In Eq. 5, u_p and v_q are C-dimensional feature vectors from H×W×C maps, but x_pq is a scalar decision variable. The constraints sum_p x_pq = v_q and sum_q x_pq = u_p equate a scalar to a vector, so the LP has no meaning as written. Even if you treat each feature as a scalar mass, feasibility requires sum_p u_p = sum_q v_q, which fails for any two independently trained networks. The KKT Jacobian in Eq. 9 then has no guarantee of invertibility, since the LP may be infeasible or degenerate. This isn't a notation typo; it means Algorithm 1's weighting factor is undefined. The stress-test note also mentions that Eq. 3 uses (1-c_pq) while Eq. 5 minimizes c_pq—that part is less damning because the total flow is constant if the LP is feasible, but the feasibility problem already sinks it.\n\nSecond, the numbers. The validation section says the default setup is 100% unseen label corpus, and Table II reports FedEFM at 98.2% in CFL. In Figure 4, the same setting (100% unseen) shows 84.9%. That's an unexplained 13-point gap. No error bars or hyperparameters are given, so a reader can't tell which is real.\n\nThird, the data description is unreliable. Table I lists VESSEL12 (CT), Drive (retinal), and SenNet (phase-contrast tomography) as 'X-ray datasets' for foundation training. They are not X-ray. Also, EISimulation appears in both the training group and the downstream evaluation, contradicting the explicit claim that training datasets are not used downstream.\n\nWho this is for: people working on federated learning for medical imaging might read it as an example of what the data issues are, but the results can't be trusted. A serious referee should look at it, but the authors need to fix the EMD derivation and re-run the experiments with consistent settings. I would not cite the current version.","headline":"FedEFM's core EMD formulation is ill-defined and the headline results are internally inconsistent, so the paper should be returned for major revision despite a real problem and a useful dataset.","tokens_in":13268,"tokens_out":4853,"would_cite":false,"duration_ms":42801,"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":"FedEFM claims that a federated distillation algorithm using differentiable Earth Mover's Distance lets each silo learn from data types it never sees, and that the resulting weights fine-tune into top-performing endovascular models.","keywords":["federated learning","foundation model","endovascular intervention","catheter and guidewire segmentation","knowledge distillation","Earth Mover's Distance","unseen data","X-ray imaging"],"falsifier":"Replace the EMD scalar in the update rule with a constant, or with a random permutation of the same scalar values, keeping all other training conditions identical; if accuracy on the 100% unseen-data setting stays near 98%, the transport-based weighting is not what carries the result.","tokens_in":12168,"feed_emoji":"🩺","tokens_out":9957,"duration_ms":81679,"temperature":0.7,"pith_summary":"The paper sets out to train a foundation model for endovascular X-ray images without centralizing patient data. Its central claim is that a federated distillation scheme using differentiable Earth Mover's Distance lets each hospital silo learn from its neighbors' data even when the data types at one silo are completely absent from another, a situation the authors call the unseen data problem. If this holds, hospitals could pool knowledge across human, animal, phantom, and simulation X-ray data while only exchanging model weights, and the resulting weights would initialize downstream catheter and guidewire segmentation as well as line-position classification. The paper reports that with a fully unseen label corpus the method reaches 98.2% accuracy in client-server federated learning and 97.5% in decentralized federated learning, and that fine-tuning these weights beats CLIP, SAM, and LVM-Med on the evaluated tasks.","feed_headline":"Federated X-ray model hits 98% accuracy without sharing patient images","feed_subtitle":"Each hospital keeps its images; distance-weighted distillation still learns data types it never saw.","key_machinery":"The central object is a differentiable Earth Mover's Distance computed between two dense feature maps $U$ and $V$ of the same input produced by the local student and the transferred teacher. It solves a linear program that transports mass between the feature vectors, with unit cost $c_{pq} = 1 - u_p^T v_q / (\\|u_p\\|\\|v_q\\|)$, and then uses the KKT conditions and a Jacobian argument to differentiate the optimal transport plan with respect to the model parameters. The resulting scalar $EMD(\\theta_i, \\hat{\\theta}_{i\\to j})$ scales the distillation gradient in the local update, so a silo trusts a neighbor's transferred weights more when the two feature distributions align and less when they diverge; this gating is what the paper credits for stable learning under unseen data.","core_discovery":"On the paper's own terms, the discovery is that the divergence that normally breaks federated training when silos hold disjoint data types can be tamed by weighting each distillation gradient with the Earth Mover's Distance between the local model's feature map and the feature map of the neighbor-trained 'overseas expert.' With that weighting, FedEFM maintains high accuracy even when 100% of each silo's data are unseen elsewhere, whereas the federated baselines collapse. Fine-tuning the federated weights also improves segmentation and classification over existing foundation models, which the paper takes as evidence that the federated training has produced genuinely reusable knowledge rather than a task-specific compromise.","pith_inferences":["A testable extension the paper does not run is to replace the EMD scalar with a constant or a simpler similarity measure; if accuracy is unchanged, the specific transport geometry would not be the active ingredient, only some form of gradient scaling.","Because the method exchanges only weights, the same EMD-gated distillation could be carried to other privacy-sensitive medical imaging domains where hospitals hold disjoint or partially overlapping label sets, such as pathology or endoscopic video.","The reported gains are on phantom, simulation, and benchmark classification data; live clinical fluoroscopy would be the stronger test of whether the learned representations survive motion blur, low contrast, and unseen anatomy."],"forward_implications":["Federated foundation-model training no longer requires silos to share a common label set, because the EMD weighting operates on feature distributions rather than on matching labels.","Hospitals can contribute human, animal, phantom, and simulation X-ray data while sharing only weights, which directly addresses the privacy barrier that the paper identifies as blocking medical foundation models.","The trained weights are reusable initializations: fine-tuning U-Net, TransUNet, SwinUNet, or ViT from FedEFM weights improves downstream segmentation and classification relative to training from scratch.","The paper's ablation shows the EMD weighting matters: removing it drops accuracy from 98.2% to 84.7% in client-server and from 97.5% to 72.4% in decentralized settings."],"supporting_citations":[{"why":"This vision-language model is the baseline that FedEFM outperforms in downstream classification and segmentation.","marker":"[7]"},{"why":"This medical-image foundation model provides the comparison baseline and evaluation protocol for fine-tuning.","marker":"[9]"},{"why":"This benchmark supplies one of the large-scale endovascular X-ray datasets used in the federated foundation training.","marker":"[14]"},{"why":"This work supplies the knowledge-distillation loss with temperature-softened outputs that FedEFM extends.","marker":"[40]"},{"why":"This work contributes the differentiable Earth Mover's Distance machinery for matching structured feature sets.","marker":"[42]"},{"why":"This work defines the decentralized federated-learning topology and serves as the DFL baseline.","marker":"[44]"},{"why":"This is the FedAvg baseline in client-server federated learning, which FedEFM reports improving from 80.9% to 98.2%.","marker":"[54]"},{"why":"This segmentation foundation model is a downstream fine-tuning baseline for comparison.","marker":"[55]"}],"fun_headline_variants":["Federated model learns unseen data via Earth Mover's Distance","Federated X-ray model uses EMD to handle unseen data types","Distance-weighted distillation lets federated model learn new data","Federated endovascular model: EMD overcomes unseen data in X-ray","Federated X-ray AI: distance metric solves unseen data challenge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole mechanism depends on the Earth Mover's Distance between the local and transferred feature maps being a meaningful measure of task-relevant agreement; if that distance does not track what the neighbor actually knows, the gradient weighting is just an arbitrary scaling and the unseen-data robustness claim has no foundation.","fun_headline_variants_meta":{"raw":{"variants":["Federated model learns unseen data via Earth Mover's Distance","Federated X-ray model uses EMD to handle unseen data types","Distance-weighted distillation lets federated model learn new data","Federated endovascular model: EMD overcomes unseen data in X-ray","Federated X-ray AI: distance metric solves unseen data challenge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000681,"raw_usage":{"total_tokens":3046,"prompt_tokens":851,"completion_tokens":2195,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":2103}},"tokens_in":467,"tokens_out":2195,"duration_ms":14716,"temperature":1.0,"reasoning_tokens":2103,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T05:14:07.247657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the EMD scalar in the update rule with a constant, or with a random permutation of the same scalar values, keeping all other training conditions identical; if accuracy on the 100% unseen-data setting stays near 98%, the transport-based weighting is not what carries the result.","supporting_citations":[{"cited_title":"Learning transferable visual models from natural language supervision,","cited_arxiv_id":null,"evidence_quote":"This vision-language model is the baseline that FedEFM outperforms in downstream classification and segmentation."},{"cited_title":"Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching,","cited_arxiv_id":null,"evidence_quote":"This medical-image foundation model provides the comparison baseline and evaluation protocol for fine-tuning."},{"cited_title":"CathAction: A Benchmark for Endovascular Intervention Understanding","cited_arxiv_id":"2408.13126","evidence_quote":"This benchmark supplies one of the large-scale endovascular X-ray datasets used in the federated foundation training."},{"cited_title":"Distilling the knowledge in a neural network,","cited_arxiv_id":null,"evidence_quote":"This work supplies the knowledge-distillation loss with temperature-softened outputs that FedEFM extends."},{"cited_title":"Deepemd: Few-shot image classification with differentiable earth mover’s distance and structured classifiers,","cited_arxiv_id":null,"evidence_quote":"This work contributes the differentiable Earth Mover's Distance machinery for matching structured feature sets."},{"cited_title":"Throughput-optimal topology design for cross-silo federated learning,","cited_arxiv_id":null,"evidence_quote":"This work defines the decentralized federated-learning topology and serves as the DFL baseline."},{"cited_title":"Communication-efficient learning of deep networks from decentral- ized data,","cited_arxiv_id":null,"evidence_quote":"This is the FedAvg baseline in client-server federated learning, which FedEFM reports improving from 80.9% to 98.2%."},{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"This segmentation foundation model is a downstream fine-tuning baseline for comparison."}],"review_version":1}