{"id":"68534434-3a46-42ec-abe3-b6630334b2db","arxiv_id":"1908.05770","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A weakly supervised segmentation method that uses ADMM to alternate between neural network training and exact discrete optimization of size and boundary constraints, improving Dice scores on cardiac and prostate MRI.","lead":"This paper trains a neural network for medical image segmentation using only a small set of labeled pixels, combined with two rules: the organ's size must fall in a known range and its boundary should be smooth. The authors show that enforcing these rules with exact discrete optimization, instead of continuous penalties, improves accuracy and constraint satisfaction on cardiac and prostate MRI.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The near-full-supervision result is achieved only at ε=0, where the size bound uses the exact per-volume ground-truth foreground volume; this is a strong annotation, not a weak prior.","rationale":"The reader's CONDITIONAL verdict is well-supported. The paper's internal derivations for the two discrete proposal updates are correct, and the comparison with the continuous penalty baseline is informative. The most load-bearing issue, however, is evaluative rather than mathematical: the headline 'near full supervision' result is attained only when the size bound is the exact ground-truth volume, which is a strong per-image annotation. This does not invalidate the method, but it sharpens the condition under which the central claim holds. The verdict should remain CONDITIONAL, with the condition made explicit: the reported near-full-supervision performance assumes exact per-volume foreground size information.","tokens_in":14691,"tokens_out":6760,"duration_ms":71622,"concrete_test":"Re-run Table 2's headline CRF+size configuration using ε=10% and ε=20% bounds derived from a training-set-only anatomical prior (e.g., the mean foreground fraction) rather than per-volume Sreal, with a fixed hyperparameter grid, early-stopping rule, and held-out test split. If the Dice scores do not stay within about two points of the fully supervised baselines, the near-full-supervision claim depends on exact ground-truth size.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 defines [Smin,Smax] = [(1−ε)Sreal, (1+ε)Sreal], so the best-performing CRF+size numbers in Table 2 — LV 0.901 vs 0.927, RV 0.730 vs 0.870, prostate 0.807 vs 0.873 — are obtained when the size constraint is the exact ground-truth foreground size of every volume. That is a much stronger piece of supervision than the paper's 'weak annotations' framing suggests. At ε=10% the same configuration drops to 0.884/0.719/0.795, and at ε=40% to 0.879/0.691/0.771. The advertised 'performance near full supervision' is therefore conditional on providing the exact target size, which is a per-volume ground-truth quantity. The ADMM convergence caveat in Section 3.4.1 is real but secondary: empirical stability can be assessed directly, whereas the ε=0 setting changes what supervision the headline result actually uses.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses weakly supervised medical image segmentation by training a CNN under two discrete priors: a lower/upper bound on foreground size and a boundary-length (weighted Potts) regularization term. Because these terms are non-differentiable and discrete, the authors split the training objective with ADMM (Eq. 8), introducing two binary 'proposal' vectors per image that are decoupled from the continuous network output. The network parameters are updated by mini-batch SGD (Section 3.4.1); the CRF-regularized proposal is updated by solving a submodular graph-cut problem (Section 3.4.2); and the size-constrained proposal is updated by ranking pixels, since the subproblem is an equal-weight knapsack (Section 3.4.3). Experiments on three tasks (cardiac LV and RV from ACDC; prostate from PROMISE12) compare 'CRF only', 'size only', and 'CRF+size' variants against the continuous penalty baseline of Kervadec et al. [10] and a fully supervised model, reporting Dice scores, constraint-satisfaction curves, and sensitivity to the ADMM penalty parameter. The central quantitative claim is that at tight size bounds (epsilon=0) the method approaches fully supervised performance (e.g., LV 0.901 vs 0.927).","tokens_in":1867,"tokens_out":2227,"duration_ms":138767,"significance":"If the results hold, the paper makes a useful methodological contribution: it is, to my knowledge, the first to plug exact discrete optimizers (submodular graph cuts and cardinality-constrained ranking) into CNN training through ADMM variable splitting, and the two proposal updates are derived cleanly in Eqs. (10)-(13), each with a polynomial-time solver that is globally optimal for the subproblem. The paper also reports a complete experimental protocol (architectures, learning schedules, grid-searched hyperparameters), including an ablation of the two priors (Tables 2-3) and a constraint-satisfaction analysis (Fig. 6), which is good for reproducibility and goes beyond single-configuration reporting. The main qualification is that the 'near full supervision' headline is achieved only when the size constraint encodes the exact per-volume ground-truth foreground size (epsilon=0), which is a strong annotation; at epsilon=10-40% the gains over the baseline are smaller but still consistent for the CRF+size variant.","major_comments":[{"comment":"The paper's strongest quantitative claim, 'performance near to full supervision,' is conditioned on epsilon=0, where the size bounds are [S_real, S_real], i.e., the exact per-volume ground-truth foreground size is supplied to the method at training time. This is a per-volume ground-truth quantity, not a weak prior; the same configuration drops to 0.884 (LV), 0.719 (RV), and 0.795 (prostate) at epsilon=10%, and to 0.879/0.691/0.771 at epsilon=40%. The abstract and conclusion should state this conditioning explicitly, and the discussion should be rebalanced around the epsilon=10-40% regime, where the contribution of the discrete formulation is still visible but more modest.","section":"Section 4.1 / Table 2"},{"comment":"The text states that the mini-batch gradient-descent update of the network parameters 'corresponds to a proximal variant of ADMM' citing references [30,31], and Section 3.5 further claims that 'the only requirement for convergence is that the update of network parameters decreases the overall loss [29].' For the non-convex network objective, the conditions of the cited proximal-ADMM analyses (Lipschitz smoothness of the smooth part, sufficient decrease of the augmented Lagrangian) are not verified, and the authors themselves note in Section 6 that strong penalty parameters cause the discrete proposals to oscillate. Please either provide empirical convergence evidence (e.g., augmented-Lagrangian or constraint-violation curves over epochs) or explicitly frame the outer loop as a heuristic whose stability is demonstrated empirically in Figures 3 and 6.","section":"Section 3.4.1 / Algorithm 1"},{"comment":"There is a sign inconsistency in the size-constrained proposal update. Minimizing (1/2)||y_tilde - (s(theta) + u_tilde)||^2 over binary y_tilde, the problem stated in Eq. (12) and following from the augmented term in Eq. (8), is equivalent to maximizing sum_p (s_p(theta) + u_tilde_p - 1/2) y_tilde_p, not sum_p (s_p(theta) - u_tilde_p - 1/2) y_tilde_p as printed in Eq. (13). The printed sign also conflicts with the multiplier update u_tilde := u_tilde + (s - y_tilde) in Algorithm 1. Since the optimality of the size update is one of the two 'guarantee of solution optimality' claims of the paper, this inconsistency should be resolved (it may be a typo) rather than left as is.","section":"Section 3.4.3, Eqs. (12)-(13)"},{"comment":"The reported numbers are single runs with no variance or significance tests, which does not by itself support the adjective 'significant' used in the summary of Section 5.1; moreover, the epsilon-sweep for the CRF+size variant is non-monotonic (RV 0.734 at epsilon=20% exceeds 0.730 at epsilon=0%, and prostate 0.809 at epsilon=20% exceeds 0.807 at epsilon=0%), which is unexplained. At minimum, the paper should report repeated runs or paired significance tests for the headline comparisons and add a sentence addressing the non-monotonicity.","section":"Section 5 / Table 2"}],"minor_comments":[{"comment":"The arXiv abstract says experiments were run 'on a benchmark cardiac segmentation dataset,' while the full-text abstract and Section 4 describe two datasets (ACDC and PROMISE12); the abstracts should be aligned.","section":"Abstract"},{"comment":"Figure numbering is inconsistent: a figure captioned with annotation examples is labeled Figure 2, but the text says 'Figure 3 shows examples of annotations and ground truth,' and Figure 3 is also used later for the validation Dice curves; please renumber the figures.","section":"Section 4.1 / Figures 2-3"},{"comment":"The training loop returns parameters after the final epoch, but Section 5.1 reports that the highest validation accuracy is obtained around epoch 50; please specify the model-selection rule used for the numbers in Table 2 (final epoch versus best validation epoch).","section":"Section 5.1 / Algorithm 1"},{"comment":"There are minor typos in the dataset description: 'well-defined defined pathologies' and 'The images resolution ranges' should be corrected.","section":"Section 4.1"},{"comment":"The claim of 'negligible computational overhead' for the once-per-epoch proposal updates would be more convincing with a short empirical note on wall-clock training time, given that each CRF update is a max-flow solve on 256x256 images.","section":"Section 3.5"}],"recommendation":"major_revision","confidential_remarks":"The epsilon=0 issue is a matter of framing rather than concealment, since all epsilon values are reported in Table 2; my main recommendations are to fix the sign inconsistency in Section 3.4.3, temper the abstract's 'near full supervision' claim, and add convergence or significance evidence for the two evaluative claims. The paper is otherwise a solid methods contribution by the standards of the area, and I see no grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Know this paper for the ADMM splitting idea: SGD updates the network, while two discrete subproblems handle size constraints and boundary-length regularization. The discrete solvers are genuine contributions - graph cuts for the Potts term, and a simple ranking method for the size-constrained knapsack. The math in Eqs. (10)-(13) checks out, and the experiments show consistent improvements over the continuous penalty baseline on all three tasks. That part is solid.\n\nThe 'near full supervision' headline needs a qualifier, and it is a real one. The best numbers (LV 0.901 vs 0.927, RV 0.730 vs 0.870, prostate 0.807 vs 0.873) all come at ε=0, where the size bound [Smin,Smax] is exactly the ground-truth foreground size of each volume. That is not a weak annotation; it is the target size. At ε=10% the same CRF+size configuration drops to 0.884/0.719/0.795, and at ε=40% to 0.879/0.691/0.771. The paper is transparent about this in Table 2 and in the text, but the abstract's 'performance near full supervision' does not carry the caveat. The method still beats the penalty baseline at looser bounds, so the core approach is not undermined, but the strongest claim is conditional on exact size knowledge.\n\nTwo secondary concerns. First, the ADMM convergence discussion is thin: the paper calls the SGD parameter update a proximal ADMM step citing [30,31], but does not verify the required conditions (sufficient decrease, Lipschitz smoothness) for this non-convex setting. Empirically the training looks stable, so this is a theory gap rather than a demonstrated failure. Second, the hyperparameters (µ, λ, σ) are tuned per task on the validation sets, and the reported numbers are best-case. That is standard practice, but it weakens the comparative claims slightly. No code or splits are released, which hurts reproducibility but is not a fatal flaw for a 2019 preprint.\n\nThe citation pattern is fair. They build on Pathak et al. and Kervadec et al., and they cite the discrete-continuous ADMM line of work. The 'first' claim is a bit assertive given references [27,28], but they do distinguish their weakly-supervised training setting.\n\nWho gets value: anyone doing weakly supervised medical segmentation, especially with size priors or boundary regularization. The method is clean, the exposition is clear, and the improvements over a strong penalty baseline are credible. I would send it to peer review. My recommendation is a revise that explicitly states the ε=0 setting is exact-size supervision in the abstract and adds an honest caveat about the unverified ADMM convergence conditions.","headline":"A solid ADMM-based method for injecting discrete size and boundary priors into CNN training, with correct discrete solvers and consistent gains over penalty baselines, but the advertised 'near full supervision' result depends on giving the network the exact ground-truth foreground size.","tokens_in":15432,"tokens_out":2206,"would_cite":true,"duration_ms":22775,"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":"A CNN can be trained with discrete size and boundary-length constraints via ADMM, reaching near-fully-supervised Dice scores on weakly annotated medical images.","keywords":["weakly supervised segmentation","discrete constraints","ADMM","size constraints","boundary length regularization","medical image segmentation","convolutional neural networks","graph cuts"],"falsifier":"Reset the size bounds to values derived from an independent anatomical prior rather than from the ground-truth volume and record the augmented Lagrangian after every epoch; if the Dice advantage over the penalty baseline disappears when the true size is unknown, or if the Lagrangian fails to decrease across epochs, the central claim is undermined.","tokens_in":14505,"feed_emoji":"🩺","tokens_out":8339,"duration_ms":76714,"temperature":0.7,"pith_summary":"This paper argues that weakly supervised segmentation networks should be trained by enforcing constraints on hard binary labels, not on soft probability maps, and shows a concrete way to do it: split the training objective with the alternating direction method of multipliers (ADMM) into continuous network updates and discrete proposal updates. On cardiac and prostate MRI, with only a few foreground and background seeds per volume, the method reports Dice scores well above a continuous squared-penalty baseline and close to fully supervised training (left ventricle 0.901 versus 0.927 at the tightest size bound). The paper also shows that combining a discrete size constraint with boundary-length (weighted-Potts) regularization is what makes the gain robust, since size alone under-segments complex structures and CRF regularization alone has a shrinking bias. If the method works as reported, it gives practitioners a way to inject anatomical priors into deep segmentation without pixel-level annotations.","feed_headline":"Discrete ADMM nearly closes weak-to-full supervision gap","feed_subtitle":"Splitting network updates from discrete size and boundary proposals beats continuous penalties on cardiac and prostate MRI.","key_machinery":"The driving object is an ADMM-augmented Lagrangian with two auxiliary binary proposal vectors per image, $\\hat{y}_i$ and $\\tilde{y}_i$, tied to the network output by equality constraints $s_i(\\theta)=\\hat{y}_i$ and $s_i(\\theta)=\\tilde{y}_i$. The regularized proposal $\\hat{y}_i$ is updated by minimizing a weighted Potts (boundary-length) energy with graph cuts, which is globally optimal for submodular pairwise terms; the size proposal $\\tilde{y}_i$ is updated by solving a cardinality-constrained selection problem—pick between $S_{\\min}$ and $S_{\\max}$ pixels with largest utility—whose optimum is a simple ranking. The network parameters are updated by SGD, interpreted as a proximal ADMM update, and the multipliers are updated in the standard ADMM way. This machinery is what lets constraints act on hard labels rather than soft probabilities, avoiding gradient saturation and allowing large, optimal discrete steps.","core_discovery":"The central claim is that discrete size constraints and boundary-length regularization can be imposed on a CNN during training through an ADMM variable-splitting scheme, and that doing so beats the standard continuous penalty approach. Two binary proposal vectors are introduced for each image, one coupled to a pairwise boundary-length regularizer and one coupled to a constraint on the number of foreground pixels, with equality constraints to the network's soft predictions enforced through scaled Lagrange multipliers. Each proposal update is solved exactly: the regularized one as a submodular graph-cut and max-flow problem, the size-constrained one by sorting pixel utilities and selecting the top S_min up to S_max pixels. The network parameters are updated with ordinary mini-batch SGD, which the paper treats as a proximal ADMM step. The result, on left and right ventricle and prostate MRI with sparse seed annotations, is consistent gains over the penalty baseline in Dice, faster and more stable constraint satisfaction, and near-fully-supervised performance when the size bound is tight.","pith_inferences":["An untested but natural extension is to replace the two proposal types with other discrete priors the paper cites, such as star-shape, compactness, or histogram constraints, provided each can still be solved to optimality in its proposal update.","The headline epsilon=0 result uses bounds computed from the ground-truth foreground volume, so transferring the method to a genuinely unseen organ would require deriving bounds from population statistics; the reported 10-40% relaxation results suggest the method degrades gracefully as bounds loosen.","Because the paper identifies oscillation risk when the ADMM penalty is large, adapting the penalty parameter per epoch, as standard ADMM allows, could stabilize training and is directly testable with the reported setup.","The observed early-best validation peak hints that the discrete proposals act as a strong curriculum signal; a testable variant would anneal the proposal influence over training to retain the early boost while avoiding late-training instability."],"forward_implications":["If the central claim holds, weakly supervised cardiac and prostate segmentation can be trained from a few seeds per volume to within roughly 3 Dice points of full supervision on the left ventricle when the target size is known.","The discrete ADMM formulation satisfies size constraints more stably than the continuous penalty method, with predicted-to-real size ratios staying inside the imposed bounds during training.","Combining size constraints with boundary-length regularization is necessary for complex, narrow structures like the right ventricle, where size alone is insufficient and CRF regularization alone over-shrinks.","The same variable-splitting recipe can be applied to other discrete priors without changing the network training loop, at small extra cost: one max-flow and one sort per image per epoch.","The best validation accuracy occurs early in training, around epoch 50, before convergence, which suggests early stopping is beneficial for this scheme."],"supporting_citations":[{"why":"Supplies the continuous penalty baseline and the weak-annotation setup that the proposed method must beat.","marker":"[10]"},{"why":"Provides the graph-cut and max-flow algorithm used to solve the CRF-regularized proposal update to global optimality for submodular Potts energies.","marker":"[15]"},{"why":"Gives the ADMM framework, scaled-multiplier formulation, and the convergence requirement cited for the alternating updates.","marker":"[29]"},{"why":"Cited to justify treating a first-order SGD parameter update as a proximal ADMM step in the non-convex setting.","marker":"[30]"},{"why":"Second source cited for proximal ADMM convergence in linearly constrained nonconvex minimization, backing the network-update approximation.","marker":"[31]"},{"why":"Provides the ACDC benchmark dataset used for left- and right-ventricle segmentation experiments.","marker":"[33]"},{"why":"Provides the PROMISE12 prostate MRI dataset used for the third segmentation task.","marker":"[34]"},{"why":"Defines the ENet architecture used as the segmentation network on the cardiac dataset.","marker":"[35]"}],"fun_headline_variants":["Discrete ADMM constraints beat continuous penalties in weak seg","ADMM with discrete size and boundary priors near full supervision","Exact discrete proposals in ADMM outperform smooth penalties","Weak supervision segmentation via discrete-constrained ADMM","ADMM discrete constraints approach full supervision in MRI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on treating a fixed number of mini-batch SGD steps as a proximal ADMM update that decreases the augmented Lagrangian, and on computing the tightest reported size bounds from the exact ground-truth foreground volume, so the headline \"weak\" supervision includes the true target size.","fun_headline_variants_meta":{"raw":{"variants":["Discrete ADMM constraints beat continuous penalties in weak seg","ADMM with discrete size and boundary priors near full supervision","Exact discrete proposals in ADMM outperform smooth penalties","Weak supervision segmentation via discrete-constrained ADMM","ADMM discrete constraints approach full supervision in MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1178,"prompt_tokens":854,"completion_tokens":324,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":247}},"tokens_in":470,"tokens_out":324,"duration_ms":3703,"temperature":1.0,"reasoning_tokens":247,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:05:41.846718+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reset the size bounds to values derived from an independent anatomical prior rather than from the ground-truth volume and record the augmented Lagrangian after every epoch; if the Dice advantage over the penalty baseline disappears when the true size is unknown, or if the Lagrangian fails to decrease across epochs, the central claim is undermined.","supporting_citations":[{"cited_title":"Kervadec, J","cited_arxiv_id":null,"evidence_quote":"Supplies the continuous penalty baseline and the weak-annotation setup that the proposed method must beat."},{"cited_title":"Boykov, O","cited_arxiv_id":null,"evidence_quote":"Provides the graph-cut and max-flow algorithm used to solve the CRF-regularized proposal update to global optimality for submodular Potts energies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited to justify treating a first-order SGD parameter update as a proximal ADMM step in the non-convex setting."},{"cited_title":"Bernard, A","cited_arxiv_id":null,"evidence_quote":"Provides the ACDC benchmark dataset used for left- and right-ventricle segmentation experiments."},{"cited_title":"Litjens, R","cited_arxiv_id":null,"evidence_quote":"Provides the PROMISE12 prostate MRI dataset used for the third segmentation task."}],"review_version":1}