{"id":"03f66bf1-60c7-4e62-8eaf-d97024595bcf","arxiv_id":"2507.10490","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A confidence-weighted consistency loss between successive mini-batches improves polyp segmentation Dice by about one to three points over base training on several unseen colonoscopy datasets, but the method is an incremental variant of existing self-distillation.","lead":"This paper proposes a training trick for medical image segmentation: a model is nudged to agree with its own confident predictions from one step earlier, instead of using a second teacher model. The authors show modest gains on polyp segmentation benchmarks, especially when tested on hospitals or image sources the model never saw during training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DCSD loss is defined two mutually contradictory ways: Eq. 1 weights consistency by Dice(p_prev, y_prev), while Algorithm 2 weights by 1 - Dice; this ambiguity is the load-bearing flaw.","rationale":"The paper has genuine strengths: it evaluates on 12 datasets, includes multi-architecture ablations (TransNetR, ShallowNet, own model), and the proposed regularizer adds no test-time cost. I do not question the authors' intent; the issue is purely in the specification of the method. The reader's conditional verdict already lists the Algorithm 1/2/Eq.1 inconsistency, so I partially agree with the weakest-assumption choice: the formal generalization claim in Section 3.2 is asserted, not derived, but the empirical claim does not depend on that theory. My concrete test would settle which loss is real. If the variants behave equivalently and match the tables, the contradiction is cosmetic and the verdict can stand; if not, the reported numbers are not tied to DCSD. Also, Table 4 shows SD beating DCSD on Kvasir, ClinicDB, and ColonDB, so the 'outperforms self-distillation' statement is already limited to unseen datasets, and without error bars those gaps may be noise. I therefore keep the conditional verdict, with the explicit conditions: resolve the loss definition, release code, and report seed variance.","tokens_in":11338,"tokens_out":7359,"duration_ms":84986,"concrete_test":"Re-train the same architecture on data_c1-c5 with data_c6 as test, 30 epochs, 5 seeds, under three losses: (A) Eq. 1 using Dice(p_prev, y_prev) as weight; (B) Algorithm 2 using 1 - Dice(p_prev, y_prev); (C) Algorithm 1 with t2 fixed to 1.0. If only one variant reproduces the reported DCSD row (0.8151/0.7542), that resolves which definition the paper actually used; if none reproduce it, unstated hyperparameters drive the result. Also repeat the Table 4 ablation on ETIS, BKAI-IGH, and EndoScene with multiple seeds; if the DCSD-vs-SD differences (<=0.018 Dice) are within seed noise, the generalization claim needs tempering.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim - that confidence-weighted self-distillation improves segmentation - depends on one well-defined loss. The paper gives three incompatible definitions. Eq. 1 defines L_DCSD = (1/n) * sum Dice(p^{t-1}_i, y^{t-1}_i) * MSE(p^t_i, p^{t-1}_i), i.e. a high-Dice previous prediction is trusted more. Algorithm 2 lines 5-6 set confidence-coefficient = 1 - Dice(pre_out, pre_label) and loss = consistency * confidence-coefficient, i.e. a high-Dice previous prediction is trusted less. These are opposite weighting schemes, and both are called 'dynamic confidence'. Algorithm 1 adds a third variant: line 12 calls dcsd(out_pre, pre_out, pre_label) and line 13 multiplies by an undefined t2; the pair (pre_out, out_pre) is also used differently from Eq. 1. Since the sign of the confidence weight is the entire mechanism, the reported data_c6 numbers (0.8151 Dice) and Table 4 cross-dataset gains cannot be attributed to the described method. This is a reproducibility failure at the exact point the central claim rests, independent of the also-unproven complexity-reduction argument in Section 3.2.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Dynamic Confidence-based Self-Distillation (DCSD), a training-time regularizer for polyp segmentation. DCSD computes a consistency loss between the current iteration's predictions and the previous mini-batch's softened predictions, weighted by a 'dynamic confidence coefficient' derived from the Dice overlap of the previous prediction with the ground truth. The authors combine DCSD with a PVT-backed encoder-decoder architecture with RFB and layer aggregation. They report that DCSD outperforms the base model and conventional self-distillation on the data_c6 test set, and on three of six cross-dataset targets (EndoScene, ETIS, BKAI-IGH), with no extra test-time cost. The paper also includes ablations on soft versus hard confidence and a comparison against ten published polyp segmentation methods.","tokens_in":1635,"tokens_out":1707,"duration_ms":86391,"significance":"If DCSD is exactly as described and the gains are reproducible, it would be a low-cost regularizer applicable to any segmentation model, since it only stores the previous mini-batch and has no inference overhead. The paper's three-architecture ablation (TransNetR, ShallowNet, and the proposed model) is a useful check on generality, and the cross-dataset experiments address an important clinical need. However, the manuscript currently defines the loss in two contradictory ways, omits a hyperparameter, and overstates the state-of-the-art comparison, so the empirical contribution cannot be assessed until these issues are resolved.","major_comments":[{"comment":"The loss definition is contradictory. Eq. (1) defines L_DCSD = (1/n) Σ Dice(p^{t-1}_i, y^{t-1}_i) · MSE(p^t_i, p^{t-1}_i), so a high-Dice previous prediction receives a larger weight. Algorithm 2, lines 5–6, sets confidence-coefficient = 1 − dice_loss(pre_out, pre_label) and loss = consistency · confidence-coefficient, so a high-Dice previous prediction receives a smaller weight. These are opposite weighting schemes, and both are called 'dynamic confidence.' Algorithm 1, line 12, calls dcsd(out_pre, pre_out, pre_label) with an argument order that does not match Eq. (1), and line 13 multiplies by an undefined hyperparameter t2 whose value is never given in Section 4.1. Because the sign and form of the confidence weight are the entire mechanism of the method, the reported improvements in Tables 3–5 cannot be attributed to a well-defined DCSD loss as it stands.","section":"Section 3.1, Eq. (1); Algorithm 1 line 13; Algorithm 2 lines 5-6"},{"comment":"The theoretical analysis asserts Complexity(H_DCSD) < Complexity(H_SD) without defining either hypothesis space or proving the inequality. Since DCSD changes only the training objective and not the model family or parameterization, the hypothesis space of the trained model is not obviously smaller than that of the base model, so the claimed tighter generalization bound does not follow. This section should be removed or replaced with a rigorous argument, or explicitly labeled as an intuition rather than a theorem.","section":"Section 3.2"},{"comment":"The abstract states that DCSD 'outperforms state-of-the-art models,' but Table 2 shows that Ours ties DeepLabV3+ (ResNet101) on both Dice (0.82) and IoU (0.75) and has lower precision (0.91 vs. 0.92). The only metric on which Ours is strictly better is Recall (0.82 vs. 0.81). The claim should be toned down to 'matches or exceeds' or the comparison should be expanded to show a statistically meaningful improvement.","section":"Section 5.1 and Abstract"},{"comment":"The claim that DCSD 'achieves superior results' on cross-dataset generalization is not consistent across Table 4: DCSD is worse than SD on Kvasir (0.8985 vs. 0.9022 Dice), ClinicDB (0.8994 vs. 0.9036), and ColonDB (0.7639 vs. 0.7681). The paper reports a single run without confidence intervals or significance tests, so the gains on EndoScene, ETIS, and BKAI-IGH cannot be distinguished from noise. Please report multiple seeds and statistical testing, and discuss the datasets where DCSD underperforms SD.","section":"Section 5.2, Table 4"}],"minor_comments":[{"comment":"The word 'temparature' should be 'temperature'.","section":"Section 5.3"},{"comment":"Equation (1) uses the same index i for p^t_i and p^{t-1}_i, but Algorithm 1 compares predictions on the previous batch's images, so the indices and variables should be defined to match the algorithm.","section":"Equation (1)"},{"comment":"The input description in Algorithm 2 says 'the previous mini-batch's prediction ... is denoted as pre_out, while the current iteration's prediction ... is denoted as out_pre'; this naming conflicts with Algorithm 1 and should be aligned with Eq. (1).","section":"Algorithm 2"},{"comment":"The caption does not define the symbols B_t, θ_t, and p_t in the text; please spell out the notation.","section":"Figure 1 caption"},{"comment":"The distillation weight t2 and the temperature T used in the main experiments are not specified in the experimental setup; T=4 appears only in Table 5, and t2 is never given a value.","section":"Section 4.1"},{"comment":"The introduction misspells 'Radcliffe' as 'Radclife'.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The inconsistency between Eq. (1) and Algorithm 2 is a load-bearing flaw that prevents the results from being attributed to a well-defined method. If the authors can state which loss was actually used, provide the missing t2 value, and rerun or clarify the experiments, the empirical claims could be checked. I would also ask the editor to request the code and all hyperparameters, since the current version does not allow replication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe core idea here is actually appealing: take Shen et al.'s last-mini-batch self-distillation and weight the consistency term by how well the previous iteration's predictions matched ground truth. That is a small, sensible extension, and the paper has the right instinct—it is a training-time regularizer with zero test-time cost. The three-architecture ablation gives some real evidence that the weighting can help on unseen datasets (EndoScene, ETIS, BKAI-IGH), so I don't think this is just noise.\n\nThe problems are in the presentation and the claims. First, the loss is not pinned down. Eq. 1 weights by Dice(p^{t-1}, y); Algorithm 2 uses \"1 - diceloss(pre_out, pre_label)\". If \"diceloss\" means the standard Dice loss (1 - Dice), these are the same, but the paper never says that. Algorithm 1 multiplies by an undefined t2, and the argument order in the dcsd call is confusing. This is a reproducibility failure exactly where the mechanism lives. A referee must ask for one clean definition and all hyperparameter values. Second, the \"outperforms state-of-the-art\" headline is not supported: in Table 2, the method ties DeepLabV3+ (ResNet101) on Dice and IoU, and only leads on Recall. Third, the theoretical section (3.2) asserts rather than derives that confidence-based regularization shrinks the hypothesis space; that part is hand-waving and should be either removed or made honest. Finally, the advantage over ordinary self-distillation is inconsistent: on Kvasir and ClinicDB in Table 4, plain self-distillation does better than DCSD. That should be acknowledged.\n\nThese are fixable issues, not a dead end. The idea is cheap, plausible, and gets some empirical support. The paper deserves a serious referee, but the referee should insist on a rigorous loss definition, code or detailed hyperparameters, and error bars before it can be accepted.\n\nI'd send it to review; it is not ready as is.","headline":"A cheap, plausible confidence-weighted self-distillation trick that shows real but inconsistent gains—worth reviewing, but the paper's loss is underspecified and the SOTA claims overreach.","tokens_in":12172,"tokens_out":5493,"would_cite":true,"duration_ms":59448,"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 training-time confidence weight on self-distillation consistency losses improves polyp segmentation and generalization across datasets at zero extra test-time cost.","keywords":["Self-distillation","Confidence-based distillation","Regularization","Polyp segmentation","Medical image segmentation","Convolutional networks","Dynamic confidence","Generalization"],"falsifier":"Train the same architecture on a dataset with artificially corrupted labels (e.g., 30% flipped masks). If DCSD's confidence weighting works as claimed, the corrupted high-Dice predictions should be down-weighted and the model should degrade gracefully; if instead DCSD amplifies those errors and yields lower Dice than base training, then the confidence mechanism is not performing the claimed reweighting.","tokens_in":11131,"feed_emoji":"🎯","tokens_out":4906,"duration_ms":51234,"temperature":0.7,"pith_summary":"The paper proposes Dynamic Confidence-Based Self-Distillation (DCSD), a training-time regularizer that adds a consistency term between the model's predictions in consecutive batches, weighted by how well the earlier prediction matched the ground truth. The central claim is that this simple weighting improves segmentation accuracy and, more importantly, generalization to unseen datasets, while adding no extra computation or memory at test time. Experiments on polyp segmentation across multiple colonoscopy and wireless endoscopy datasets support that claim: the weighted version beats both a plain model and conventional self-distillation on the held-out data_c6 set and on unseen EndoScene, ETIS, and BKAI-IGH sets.","feed_headline":"Confidence-weighted training lifts segmentation accuracy","feed_subtitle":"A dynamic Dice-based confidence score helps self-distillation generalize across colonoscopy datasets at no test-time cost.","key_machinery":"The dynamic confidence coefficient is the mechanism. It is the Dice score between the previous iteration's softened prediction and the ground truth, used as a multiplicative weight on the MSE consistency term in the DCSD loss. High-Dice predictions are treated as trustworthy targets and enforced, while low-Dice predictions are down-weighted. Temperature softening (T=4) is applied to obtain soft confidence scores, and only the previous mini-batch's predictions are retained.","core_discovery":"The central discovery is that a dynamic confidence coefficient, computed as the Dice overlap between the previous iteration's temperature-softened prediction and the ground truth, can reweight the mean-squared-error consistency loss between consecutive batch predictions, and doing so consistently improves performance over both base training and unweighted self-distillation. The DCSD loss is $\\mathcal{L}_{DCSD} = \\frac{1}{n}\\sum_{i=1}^{n} \\text{Dice}(p^{t-1}_i, y^{t-1}_i) \\cdot \\text{MSE}(p^t_i, p^{t-1}_i)$. Only the previous mini-batch's predictions need to be stored during training; no teacher model or extra inference cost is required.","pith_inferences":["A natural extension is to apply DCSD to other dense prediction tasks and to noisy-label regimes; the confidence weighting may act as a self-paced regularizer that naturally down-weights unreliable early predictions.","The dependence on prior-iteration Dice with ground truth suggests that batch size and label quality are critical: with very small batches the confidence estimate is noisy, and with mislabeled ground truth the weight can systematically amplify wrong targets, an effect the paper does not quantify.","The theoretical argument that confidence-based regularization shrinks the hypothesis space is asserted rather than derived; a falsifiable test would be to measure empirical complexity, such as sharpness or effective capacity, with and without DCSD."],"forward_implications":["If DCSD is correct, any segmentation model can gain accuracy and cross-dataset generalization by adding a few lines of training code: reweight the consistency loss by the previous prediction's Dice with the label.","The improvement transfers across architectures: both established polyp segmentation models improved with DCSD over their base and self-distilled versions, so the regularizer is not tied to the proposed encoder-decoder backbone.","The method appears to generalize to unseen datasets: DCSD raised Dice to 89.54 on EndoScene, 71.21 on ETIS, and 81.41 on BKAI-IGH compared with base and self-distillation baselines.","Because no extra computation or memory is used during inference, the approach could be deployed in real-time colonoscopy settings without changing runtime."],"supporting_citations":[{"why":"The self-distillation from the last mini-batch baseline that DCSD extends and improves upon.","marker":"[32]"},{"why":"Provides the multi-centre data_c1 to data_c6 benchmark and training/testing protocol used for the main comparison.","marker":"[4]"},{"why":"Supplies the cross-dataset experimental protocol of training on Kvasir-SEG and CVC-ClinicDB and testing on unseen datasets.","marker":"[22]"},{"why":"The Pyramid Vision Transformer backbone used in the proposed architecture.","marker":"[3]"},{"why":"TransNetR, one of the two polyp segmentation models used in the ablation study to test DCSD's transferability.","marker":"[5]"},{"why":"ShallowNet, the other polyp segmentation model used in the ablation study to test DCSD's transferability.","marker":"[6]"},{"why":"Introduces the temperature-based soft target mechanism that DCSD adapts for computing soft confidence scores.","marker":"[26]"}],"fun_headline_variants":["Confidence-weighted self-distillation improves polyp segmentation","Dynamic confidence distill: better segmentation, zero test-time cost","Self-distillation with dynamic confidence for robust polyp segmentation","Confidence-based self-distillation strengthens polyp segmentation","Power of certainty: confident self-distillation improves segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a prediction's Dice against the ground truth in the previous iteration is a reliable signal of which predictions are safe to use as distillation targets; if early predictions are miscalibrated or the batch is small, the weighting can reinforce errors instead of preventing them.","fun_headline_variants_meta":{"raw":{"variants":["Confidence-weighted self-distillation improves polyp segmentation","Dynamic confidence distill: better segmentation, zero test-time cost","Self-distillation with dynamic confidence for robust polyp segmentation","Confidence-based self-distillation strengthens polyp segmentation","Power of certainty: confident self-distillation improves segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1426,"prompt_tokens":886,"completion_tokens":540,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":465}},"tokens_in":502,"tokens_out":540,"duration_ms":7583,"temperature":1.0,"reasoning_tokens":465,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:30:06.908324+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same architecture on a dataset with artificially corrupted labels (e.g., 30% flipped masks). If DCSD's confidence weighting works as claimed, the corrupted high-Dice predictions should be down-weighted and the model should degrade gracefully; if instead DCSD amplifies those errors and yields lower Dice than base training, then the confidence mechanism is not performing the claimed reweighting.","supporting_citations":[{"cited_title":"Self-distillation from the last mini-batch for consistency regulariza- tion","cited_arxiv_id":null,"evidence_quote":"The self-distillation from the last mini-batch baseline that DCSD extends and improves upon."},{"cited_title":"A multi-centre polyp detection and segmentation dataset for generalisability assessment","cited_arxiv_id":null,"evidence_quote":"Provides the multi-centre data_c1 to data_c6 benchmark and training/testing protocol used for the main comparison."},{"cited_title":"Transnetr: Transformer- basedresidualnetworkforpolypsegmentationwithmulti-centerout- of-distribution testing","cited_arxiv_id":null,"evidence_quote":"TransNetR, one of the two polyp segmentation models used in the ablation study to test DCSD's transferability."},{"cited_title":"Shallow attention network for polyp segmentation","cited_arxiv_id":null,"evidence_quote":"ShallowNet, the other polyp segmentation model used in the ablation study to test DCSD's transferability."}],"review_version":1}