{"id":"e687253e-48a0-4683-8b66-a9629036f1c4","arxiv_id":"2505.12418","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MEDL combines class-aware evidential fusion from two networks with an uncertainty-ranked curriculum to generate pseudo-labels for semi-supervised medical segmentation, reporting gains on five datasets though with incomplete statistical reporting.","lead":"This paper couples two deep networks that exchange uncertainty-weighted evidence to label unlabeled medical scans, then trains from confident to difficult voxels. If the numbers hold, it would improve semi-supervised medical segmentation, but the reported tables are incomplete and internally inconsistent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 3 is not a valid Dempster-style fusion: it omits conflict normalization and maps disagreement into confident arbitrary masses, so the pseudo-labels and reliability masks behind the reported gains are uncalibrated.","rationale":"I looked for the condition that must hold for the central claim: the fused pseudo-labels must be more reliable than either subnet's own prediction and must carry a trustworthy uncertainty for the curriculum. Eq. 3 is where that condition is implemented, and it is not a valid evidential combination. The reader's weakest assumption focused on unmeasured complementarity between the two sub-networks; I agree that complementarity is never directly measured, but the deeper problem is that even with complementary networks the fusion formula is mathematically incomplete, omits conflict, and has no stated normalization denominator. This is not a disagreement with consensus; it is an internal consistency issue with Eq. 3, and the manuscript itself flags the missing normalization by saying the masses 'are supposed to be normalized' without specifying how. Because no code is released, the reader cannot determine whether the experiments used normalization after Eq. 3, and the formula's behavior in the conflicting case makes the reliability mask unable to detect the most dangerous pseudo-label errors. I therefore keep the CONDITIONAL verdict: acceptance should require a corrected, derived fusion rule with explicit normalization, a toy verification of its belief-function properties, and a re-run of at least one headline comparison using that corrected rule. My concern does not change the verdict category, but it adds a specific technical condition.","tokens_in":16356,"tokens_out":8813,"duration_ms":92658,"concrete_test":"Analytical plus experimental check: (1) Feed the two-class example N1=[0.9,0.1], u1=0 and N2=[0.1,0.9], u2=0 into Eq. 3; compute total mass, the normalized pseudo-label, and its fused uncertainty. Show that total mass=0.18 with zero uncertainty before normalization. (2) Rerun the 5%-labeled LA experiment (same folds) with three variants: Eq. 3 exactly as written, standard Dempster combination with conflict normalization, and simple probability averaging, each over at least 5 random seeds. If the Dice differences among variants are within seed noise, the headline gain is not caused by CAEF; if Eq. 3 beats Dempster, the gain is an artifact of the unnormalized ad hoc coefficient rather than evidential theory.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the CAEF fusion in Eq. 3. The paper calls it an 'improved' evidential fusion, but the formula is not a Dempster combination and is internally incomplete. With the paper's own definitions, |C_n|=1, |C_K|=K, c=1/(K+1), and u_i=b_{Ni}(C_K), the unnormalized fused mass for K classes sums to 1 - (K/(K+1))(u_1+u_2-2u_1u_2), not 1 whenever either network is uncertain. The sentence 'the fused probability mass assignments are supposed to be normalized' does not fix this: no normalization constant or conflict term is given. More seriously, empty-set/conflict mass is never created, so two networks that are both highly confident but disagree produce b(C_K)=u_1u_2 approximately 0 and, after normalization, an arbitrary class pseudo-label with near-zero uncertainty. The reliability mask in Eq. 4 therefore cannot down-weight precisely the voxels where fusion is most unreliable. The central '5% labeled beats 20%' result depends on these pseudo-labels and on the curriculum that sorts voxels by this fused uncertainty; if the fusion is not a calibrated belief function, the claimed advantage is not attributable to evidential fusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes MEDL, a semi-supervised medical image segmentation framework in which two architecturally different networks produce evidential predictions. For unlabeled data, a class-aware evidential fusion (CAEF) rule, Eq. (3), combines the belief masses of the two networks into pseudo-labels; an uncertainty-based reliability mask, Eq. (4), filters these pseudo-labels; and an asymptotic Fisher-information evidential learning (FIE) schedule, Eqs. (6)--(9), gradually shifts training from confident to uncertain voxels. For labeled data, a similar uncertainty-weighted loss is applied. The method is evaluated on LA, Pancreas-CT, TBAD, ACDC, and BraTS2018, reporting state-of-the-art results, including a headline claim that 5% labeled data on TBAD outperforms competing methods that use 20% labeled data. The paper's central technical contribution is the CAEF formula and the uncertainty-based curriculum built on it.","tokens_in":16755,"tokens_out":10637,"duration_ms":102037,"significance":"If the theoretical core were sound, the proposed framework would be a useful contribution to semi-supervised medical segmentation, particularly because it combines evidence from heterogeneous network architectures and uses uncertainty to schedule learning. The manuscript has notable strengths: it evaluates on five public benchmarks, compares against many recent methods, includes component ablations, and reports voxel-level uncertainty reasoning. However, the load-bearing fusion equation is not a valid normalized evidential combination and the reliability measure built on it is ill-defined, so the claimed advantages of \"evidential fusion\" are not established as written. The empirical tables also contain internal inconsistencies that must be corrected before the results can be assessed. The central idea may be salvageable, but the current presentation is not technically reliable.","major_comments":[{"comment":"Equation (3) is not a normalized evidential fusion and cannot produce valid belief masses as written. With K classes, writing p_in = b_Ni(C_n), u_i = b_Ni(C_K), and c = |C_n|/(|C_n|+|C_K|), the sum of the unnormalized fused masses over all K+1 focal elements is T = u_1 u_2 + sum_n p_1n p_2n + c(u_1 + u_2 - 2u_1 u_2), which is not identically 1. In the extreme case p_1 = (1,0,...,0), p_2 = (0,1,0,...,0), u_1 = u_2 = 0, the sum is 0, so no normalized pseudo-label exists. The sentence in the text saying that the fused probability mass assignments are \"supposed to be normalized\" does not supply the normalization constant or a conflict-mass term. In addition, |C_K| is stated as K in Eq. (3) but as K-1 in Eq. (2), which changes the coefficient c. Since CAEF is the source of the pseudo-labels and of the uncertainty b(C_K) used in Eq. (4), the central 5%-vs-20% claim in Sec. III-B rests on an uncalibrated fusion operation.","section":"II-C, Eq. (3)"},{"comment":"Equation (4) defines reliability as R = exp(b(C_K)) / sum_{n=0}^{N-1} zeta_n log_2 zeta_n, but no guard is given for a zero denominator, and the zeta_n are the unnormalized masses produced by Eq. (3). When the fused mass is concentrated on one class, the entropy term approaches zero and R becomes unbounded; when the two networks are confident but disagree, b(C_K) is approximately zero, so the disagreement is not penalized. The index N is also not defined and is inconsistent with the K-class notation used elsewhere. The reliability mask that filters pseudo-labels therefore cannot be relied on to down-weight exactly the voxels where fusion is most unreliable.","section":"II-C, Eq. (4)"},{"comment":"Several quantities required to reproduce the curriculum are unspecified. The amplitude Xi in Eq. (6) is never given; psi_1(alpha_vn) in Eq. (7) is not defined; lambda_2 in the Fisher-information term is not specified; and lambda_GWU in Eq. (9) is called \"self-adaptive\" but no update rule is provided. The text says that a Gaussian warming-up function controls lambda_GWU following reference [47], but the concrete schedule is absent. Table VII varies only (lambda_1, lambda_2), so the sensitivity of the method to the other hyperparameters is unknown. Moreover, the manuscript asserts in Sec. II-C that the two different architectures generate complementary evidence, but it reports no per-subnet Dice or disagreement statistics; without such measurements, the pseudo-label gain cannot be distinguished from simple ensembling or from the uncertainty-based weighting alone.","section":"II-D, Eqs. (6), (7), (9)"},{"comment":"The empirical claims are not fully supported by the tables as printed. In Table II, the LA rows for URPC and MC-Net are identical at the 5% and 10% labeled ratios (Dice 86.92/87.62, Jaccard 77.03/78.25, 95HD 11.13/10.03, ASD 2.28/1.82), which is implausible and suggests a copying error. In Table III, the Ours rows report ASD values larger than the corresponding 95HD values (e.g., RV: ASD 7.28 vs 95HD 1.70; Myo: ASD 2.87 vs 95HD 0.98), which is geometrically impossible under the standard definitions of these metrics; the formatting also runs numbers together, making the table difficult to read. The authors should provide corrected tables and verify every reported value.","section":"III-B, Tables I--III"},{"comment":"The headline statement that \"the proposed method achieves superior performance using only 5% of labeled data compared to other methods using 20% of labeled data\" is made specifically for the TBAD dataset in Sec. III-B, but the abstract and the contributions section claim general state-of-the-art performance. In Table II, at the 5% labeled ratio MEDL does not uniformly dominate: on ACDC, Ours has 95HD = 4.80 versus Co-BioNet's 1.11 and BCP's 1.90; on Pancreas-CT, Ours has 95HD = 8.75 versus Co-BioNet's 5.43. The paper should state precisely on which datasets, ratios, and metrics MEDL is superior, rather than claiming to \"far outperform previous state of the arts.\"","section":"III-B, abstract and contributions"}],"minor_comments":[{"comment":"The heading contains a typo: \"Backgoround\" should be \"Background.\"","section":"II-A"},{"comment":"The cardinality of the multiple-objective set C_K is inconsistent: Eq. (2) says its cardinality is K-1, while Eq. (3) uses |C_K| = K; this should be reconciled throughout the derivation.","section":"II-B, Eq. (2)"},{"comment":"The notation L^I_{j,v} and L^{fiel}_{j,v} appears to mix; the subscript j is used for both sample index and the FIE loss, and the symbol psi_1 is never defined.","section":"II-D, Eq. (7)"},{"comment":"The text says the final output is the mean of both subnet outputs, but earlier sections describe interaction and fusion of predictions; the relation between training-time fusion and test-time averaging should be clarified.","section":"III, inference"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be an extended version of the authors' BIBM 2024 paper [12], which has the same title; I recommend the editor ask the authors to state the overlap and the novel contribution relative to that publication. The reference list contains a large number of self-citations, which is not itself a problem but should be checked for whether the positioning against prior work is sufficiently critical."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is essentially the authors' BIBM 2024 paper reposted to arXiv with more experiments. The core method, MEDL, and the CAEF fusion rule are the same as in [12]. The new content is experimental: five benchmarks, more baselines, ablations. That is useful, but the central fusion equation (Eq. 3) is not a valid Dempster combination. As written, the fused masses do not sum to 1. The text notes the masses 'are supposed to be normalized' but gives no normalization constant and no conflict mass. This matters because the pseudo-labels and the reliability mask in Eq. 4 both depend on the fused uncertainty. If that uncertainty is miscalibrated, then the headline claim (5% labeled data beating other methods at 20%) is not attributable to evidential fusion.\n\nWhat the paper does well: the experiments are on public datasets, the comparisons are extensive, and the ablation study gives some credit to the individual components. The general idea — fuse evidential outputs from two architecturally different networks and use the uncertainty to order the learning from easy to hard — is plausible and worth testing. The method is described in enough detail to reimplement, except for a few missing pieces.\n\nThe soft spots, in order. Eq. 3 is the load-bearing one: it omits normalization, never creates a conflict mass, so two networks that strongly disagree will produce a fused pseudo-label with near-zero uncertainty. Eq. 4 divides by Shannon entropy without guarding the zero-entropy case. The amplitude parameter Xi in Eq. 6 is never specified. The tables contain obvious typos and unreadable concatenated entries (for example '0.931.111.02' in Table I, '7.281.70' in Table III), and no error bars or multi-seed runs are reported anywhere. No code is released. The paper also does not state clearly what it adds over the authors' own BIBM 2024 paper [12]; the contribution section says 'far outperforms' but does not give a delta relative to [12].\n\nOn citations: the reference list is dominated by the authors' own prior work. That is not automatically a flaw, since the method builds directly on those papers, but it does mean the novelty has to be judged relative to [5], [6], [12], and the authors do not make that case.\n\nWho this is for: someone in semi-supervised medical segmentation who wants to try uncertainty-weighted fusion. They will get a plausible recipe, but they would have to supply the missing normalization, choose the missing parameters, and reimplement without code.\n\nMy recommendation: I would not cite this version. A serious editor should still send it to referees, because the empirical scope is substantial and the core idea is fixable. But the referees should require a corrected normalization, specified parameters, proper tables, multi-seed statistics, and an explicit diff against [12]. Without those, the central claim is unsupported.","headline":"A re-packaged BIBM 2024 paper whose central fusion equation is mathematically incomplete, so the headline 5%-beats-20% claim is not yet supported.","tokens_in":17219,"tokens_out":5074,"would_cite":false,"duration_ms":48336,"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":"Mutual evidential deep learning fuses two heterogeneous networks' evidence into pseudo-labels and uses uncertainty-ranked training to segment medical images from very few labeled volumes.","keywords":["semi-supervised learning","medical image segmentation","evidential deep learning","pseudo-label generation","uncertainty quantification","curriculum learning","Fisher information","class-aware evidential fusion"],"falsifier":"Take a held-out portion of the unlabeled pool, have each sub-network produce its own pseudo-label and the CAEF rule produce a fused pseudo-label, and compare all three against manual ground truth; if the fused pseudo-label does not beat the better of the two individual networks on most classes, the evidential fusion is not the source of the gain. Alternatively, replace the two heterogeneous backbones with two identical copies of VNet; if the reported advantage disappears, architectural complementarity is confirmed as the driver.","tokens_in":16150,"feed_emoji":"🩻","tokens_out":8750,"duration_ms":75598,"temperature":0.7,"pith_summary":"The paper claims that semi-supervised medical image segmentation can be made far more label-efficient by treating two collaborating networks as sources of evidence rather than as soft-label voters. It proposes mutual evidential deep learning (MEDL), in which a VNet and a 3D-ResVNet each emit evidential predictions, a class-aware evidential fusion rule combines them into pseudo-labels for unlabeled volumes, and a Fisher-information-based evidential loss schedules training from certain voxels to uncertain ones. A reliability mask built from the fused uncertainty suppresses low-quality pseudo-labels before they are used as supervision. On five public datasets, the paper reports that with only 5% labeled data MEDL matches or beats prior methods trained with 20%, and that each added component raises Dice in ablation. Label scarcity is the main bottleneck in medical segmentation, so a mechanism that safely exploits unlabeled scans would matter even if the exact gains vary across settings.","feed_headline":"Two networks fuse evidence to label scans with 5% labeled data","feed_subtitle":"Fusing two heterogeneous networks with an uncertainty curriculum beats prior methods at low annotation budgets.","key_machinery":"The load-bearing object is the class-aware evidential fusion rule (Eq. 3), a Dempster-combination-style product of the two sub-networks' belief masses with an added uncertainty-interaction term scaled by the class-count ratio $|C_n|/(|C_n|+|C_K|)$, which prevents the joint uncertainty mass from dominating as the number of classes grows. This fusion produces both pseudo-labels and the per-voxel uncertainty used in the reliability mask (Eq. 4), which suppresses low-confidence voxels before loss computation. The second mechanism is the asymptotic Fisher information-based evidential learning (FIE) weighting, $\\omega(q,v)=\\Xi\\tanh(\\psi(h(v))\\zeta(q))+1$, which ranks voxels by uncertainty and shifts training weight from certain to uncertain voxels over epochs, combined with the Fisher-information evidential loss in Eq. 7 that avoids penalizing potentially mislabeled classes in high-uncertainty voxels.","core_discovery":"The central claim is that heterogeneous-network disagreement can be turned into a training signal rather than a nuisance. Each sub-network produces per-voxel belief masses over segmentation classes plus an uncertainty mass; the class-aware evidential fusion (CAEF) reallocates these masses through a Dempster-combination-style interaction, producing fused pseudo-labels for unlabeled data and fused evidence for labeled data. The paper further claims that ranking voxels by the fused uncertainty and weighting the evidential loss so that training starts on reliable voxels and gradually includes harder ones—while avoiding over-penalization of mislabeled classes—improves both robustness and final Dice. Experiments across left atrium MRI, pancreas CT, aortic dissection CTA, cardiac MRI, and brain tumor MRI report state-of-the-art results, with the strongest comparisons on TBAD where 5% labeled data surpasses prior methods at 20%.","pith_inferences":["A direct test of the complementarity assumption would be to measure the error overlap between the two sub-networks on unlabeled data; if their mistakes are largely correlated, CAEF is mostly averaging rather than correcting, and the gains should shrink.","The class-count discounting coefficient suggests a general principle for evidential fusion: as the number of candidate classes grows, uncertainty mass should be down-weighted during combination, which could transfer to multi-class classification beyond segmentation.","The same uncertainty ranking that drives the curriculum could be reused for active learning, proposing the most informative voxels or slices for human annotation instead of only consuming them as pseudo-labels.","Replacing both sub-networks with two identical copies of the same architecture in an ablation would isolate whether architectural diversity, rather than simple ensembling, is what makes the fused pseudo-labels reliable."],"forward_implications":["If the reported results hold, a 5% labeled training set can yield segmentation quality comparable to prior methods using 20%, which would cut annotation cost roughly fourfold across modalities such as CTA, MRI, and CT.","The fusion rule only assumes complementary evidence from two differently built networks, so the mechanism should transfer to other backbone pairs beyond VNet and 3D-ResVNet.","The uncertainty-scheduled curriculum removes the need for a fixed pseudo-label confidence threshold, since training automatically starts from reliable voxels and incorporates harder ones as epochs progress.","Applying the same evidential fusion to labeled data, not just pseudo-labels, is reported to add further gains, suggesting the method improves supervised learning itself rather than only semi-supervised exploitation.","The component ablation attributes measurable Dice increases to each part of the pipeline, indicating that fusion, reliability masking, and curriculum weighting each contribute independently."],"supporting_citations":[{"why":"Defines the Dirichlet evidence view of belief masses and uncertainty that MEDL's per-voxel evidence is built on.","marker":"[29]"},{"why":"Supplies Dempster's combination rule from which the class-aware fusion in Eq. 3 evolves.","marker":"[30]"},{"why":"The mutual correction framework used as the baseline and starting point for the ablation study.","marker":"[34]"},{"why":"Source of the reliability-mask strategy applied to pseudo-labels before loss computation.","marker":"[38]"},{"why":"Uncertainty-informed prototype consistency method used as a comparison on TBAD and in the optimization objective.","marker":"[43]"},{"why":"Provides the Fisher-information evidential loss that the paper adapts with voxel-level curriculum weights.","marker":"[45]"},{"why":"Supplies the Gaussian warming-up function for the self-adaptive loss weight and serves as a comparison method.","marker":"[47]"},{"why":"Evidential inference learning baseline that MEDL contrasts against in the ACDC experiments.","marker":"[63]"},{"why":"Evidential-enhanced tri-branch consistency method compared on ACDC with 14 labeled scans.","marker":"[48]"}],"fun_headline_variants":["Evidential fusion teaches two nets to label with 5% data","Uncertainty curriculum turns network disagreement into pseudo-labels","Mutual evidential nets beat prior at low annotation budgets","Heterogeneous nets merge evidence, cut labeled data to 5%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole gain rests on the two networks making different, complementary mistakes on unlabeled data, so their fused pseudo-labels are better than either network's own; the paper asserts this complementarity but reports no per-network accuracy or error-overlap measurement to confirm it.","fun_headline_variants_meta":{"raw":{"variants":["Evidential fusion teaches two nets to label with 5% data","Uncertainty curriculum turns network disagreement into pseudo-labels","Mutual evidential nets beat prior at low annotation budgets","Heterogeneous nets merge evidence, cut labeled data to 5%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1451,"prompt_tokens":931,"completion_tokens":520,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":448}},"tokens_in":547,"tokens_out":520,"duration_ms":5253,"temperature":1.0,"reasoning_tokens":448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:34:23.876788+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out portion of the unlabeled pool, have each sub-network produce its own pseudo-label and the CAEF rule produce a fused pseudo-label, and compare all three against manual ground truth; if the fused pseudo-label does not beat the better of the two individual networks on most classes, the evidential fusion is not the source of the gain. Alternatively, replace the two heterogeneous backbones with two identical copies of VNet; if the reported advantage disappears, architectural complementarity is confirmed as the driver.","supporting_citations":[{"cited_title":"Evidential deep learning to quantify classification uncertainty,","cited_arxiv_id":null,"evidence_quote":"Defines the Dirichlet evidence view of belief masses and uncertainty that MEDL's per-voxel evidence is built on."},{"cited_title":"Upper and lower probabilities induced by a multivalued mapping,","cited_arxiv_id":null,"evidence_quote":"Supplies Dempster's combination rule from which the class-aware fusion in Eq. 3 evolves."},{"cited_title":"Mcf: Mutual correction framework for semi-supervised medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"The mutual correction framework used as the baseline and starting point for the ablation study."},{"cited_title":"Upcol: Uncertainty-informed prototype consistency learning for semi- supervised medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Source of the reliability-mask strategy applied to pseudo-labels before loss computation."},{"cited_title":"Upcol: Uncertainty-informed prototype consistency learning for semi- supervised medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Uncertainty-informed prototype consistency method used as a comparison on TBAD and in the optimization objective."},{"cited_title":"Uncertainty estimation by fisher information-based evidential deep learning,","cited_arxiv_id":null,"evidence_quote":"Provides the Fisher-information evidential loss that the paper adapts with voxel-level curriculum weights."},{"cited_title":"Shape-aware semi-supervised 3d semantic segmentation for medical images,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian warming-up function for the self-adaptive loss weight and serves as a comparison method."},{"cited_title":"EVIL: evi- dential inference learning for trustworthy semi-supervised medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Evidential inference learning baseline that MEDL contrasts against in the ACDC experiments."},{"cited_title":"An evidential- enhanced tri-branch consistency learning method for semi-supervised medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Evidential-enhanced tri-branch consistency method compared on ACDC with 14 labeled scans."}],"review_version":1}