{"id":"65d00952-cd67-4c40-8570-4f3e77fac173","arxiv_id":"2412.06045","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Dilated balanced cross entropy loss, which weights pixels by the inverse of dilated class areas, performs as well as Dice plus cross entropy on three medical segmentation tasks.","lead":"A new loss function for medical image segmentation weights each class by the inverse of its dilated mask area, so pixels near object boundaries get higher weight. In tests on polyp, skin lesion, and multi-organ segmentation, it matches the performance of the widely used Dice plus cross entropy loss.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Small margins over CE and per-dataset dilation radius without a validation protocol leave the central claim conditional; test-set tuning could explain the reported gains.","rationale":"Good-faith summary: the method is clearly specified (Eqs. 2-5), the dilation caps extreme weights for tiny objects, and the qualitative and external consistency with prior loss-function studies is plausible. The experiments show a large improvement over balanced CE and close parity with Dice+CE. My concern is not with the mechanism but with the empirical support for the superiority component of the claim. The margins over CE are small, the hyperparameter R is dataset-specific, and no validation procedure or variance estimate is reported. Figure 5 demonstrates that performance varies strongly with R, so without a pre-registered or validation-based R choice the comparison is at risk of selection bias. The reader's verdict already marks this as conditional; my analysis agrees. I would not move to ACCEPT without the proposed check, nor to REJECT, because the method has a coherent rationale and the observed parity with Dice+CE is consistent across three datasets. The concrete test above would settle whether the R selection is a genuine threat or a benign implementation detail.","tokens_in":7358,"tokens_out":10027,"duration_ms":112063,"concrete_test":"Re-run the polyp segmentation experiment (Table 1) with a fixed protocol: split Kvasir-SEG into train/validation (e.g., 80/20), select R on the validation split, retrain with 5 random seeds using the selected R, and report mean±std on CVC-ClinicDB. If the validation-selected R is not 20, or the mean DBCE-minus-CE mDice gap is within seed noise, the claimed superiority over CE is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DBCE outperforms CE and matches Dice+CE (Sec. 4). The evidence is three single-seed runs with small margins: mDice gains over CE are +0.61 (polyp), +0.97 (skin), and +0.58 (Synapse). The only method-specific hyperparameter, dilation radius R, is set \"empirically\" per dataset (Sec. 3.2), and Fig. 5 shows performance is strongly R-dependent with a clear optimum near R=20. The paper reports no validation-based selection rule, no multiple seeds, and no confidence intervals. If R was selected by inspecting test-set performance, the reported DBCE numbers are optimistically biased relative to CE and Dice+CE, and the small observed advantages may not be reproducible under an honest tuning protocol. Because the practical contribution is precisely that a weighted CE can replace Dice+CE, this unresolved selection bias is the load-bearing weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pixel-wise weighting scheme for cross-entropy loss in medical image segmentation. Instead of weighting each class by the inverse of its global frequency, as in balanced CE, the method dilates each class mask with a disk structuring element, computes the inverse of the dilated area, and assigns each pixel the maximum of these inverse-area weights across classes. The resulting dilated balanced CE (DBCE) loss is compared against CE, balanced CE, and Dice+CE on three tasks: polyp segmentation (Kvasir-SEG to CVC-ClinicDB), skin lesion segmentation (ISIC 2018), and Synapse multi-organ segmentation. The reported results show that DBCE outperforms CE by small margins (mDice gains of +0.61, +0.97, +0.58 on the three tasks) and is close to Dice+CE, with the conclusion that a properly weighted CE can match a region-based loss.","tokens_in":7549,"tokens_out":3681,"duration_ms":37110,"significance":"If the reported results hold, the paper would provide a simple and inexpensive alternative to region-based losses: a weighted cross-entropy loss whose weights are derived from dilated masks, applicable to both binary and multi-class segmentation. The method is easy to implement, and the experiments span three commonly used public datasets with a standard U-Net backbone, which lends practical relevance. The paper also honestly discusses the sensitivity of the loss to the dilation radius in Figure 5. However, the significance is tempered by the absence of uncertainty quantification and by the small, possibly within-noise improvements over CE. The core claim is that DBCE is a generalizable replacement for Dice+CE, but the evidence is conditional on the choice of the dilation radius and on single-seed evaluation.","major_comments":[{"comment":"The dilation radius R is the only method-specific hyperparameter, and it is set \"empirically\" per dataset (R=20 for polyp and skin lesion, R=10 for Synapse). Figure 5 shows a pronounced performance optimum near R=20 for the polyp task, with a large drop for other radii. The manuscript does not describe a validation-based selection procedure (e.g., a held-out split or cross-validation). If R was chosen by inspecting test-set performance, the reported gains over CE in Tables 1-3 are optimistically biased, and the method's advantage over CE is conditional on dataset-specific tuning. The paper should either report a principled tuning rule (e.g., pick R on a validation split) or show results for a fixed, transferable R across all datasets.","section":"Sec. 3.2, Fig. 5"},{"comment":"All experiments are based on a single training run per loss, and no error bars, confidence intervals, or significance tests are reported. The central claim that DBCE outperforms CE and matches Dice+CE rests on small differences: e.g., Table 1 shows mDice of 88.85 for DBCE vs. 88.24 for CE, and Table 3 shows mDice of 81.68 for DBCE vs. 81.75 for Dice+CE. These differences are within a few tenths of a point and cannot be distinguished from run-to-run variation without at least multiple seeds and a paired test. Please report mean and standard deviation over multiple seeds (3-5) and, if possible, per-fold comparisons for the Synapse dataset.","section":"Sec. 3.3, Tables 1-3"},{"comment":"The aggregate claim that DBCE produces \"similar results\" to Dice+CE hides notable organ-level differences. For example, DBCE is 0.95 points worse on Aorta (90.84 vs. 91.79) and 1.54 points worse on Pancreas (66.54 vs. 68.08), while being 1.33 points better on Spleen (91.12 vs. 89.79). These differences are not discussed. If the conclusion is that DBCE is comparable to Dice+CE across organs, the organ-wise pattern should be addressed; if not, the conclusion should be restricted to the aggregate metric.","section":"Table 3"},{"comment":"The wording \"produces superior results compared to CE\" appears in the Conclusion and is also implied in the Abstract. Given the small margins (mDice gains of 0.6-1.0 points) and the absence of statistical support, this is stronger than the evidence warrants. Either provide significance testing or revise the claim to \"comparable to, or slightly better than, CE\" to avoid overstating the finding.","section":"Sec. 4, Conclusion"}],"minor_comments":[{"comment":"In the abstract, \"improved my modifying\" should read \"improved by modifying\".","section":"Abstract"},{"comment":"The mathematical typesetting of Equations (2) through (5) appears garbled in the manuscript, particularly the normalization in Eq. (4); please provide a clean, unambiguous formulation of the weight map.","section":"Eqs. (2)-(5)"},{"comment":"Please state in the caption or text what metric is plotted on the vertical axis (presumably mDice on the test set) and clarify whether the curve represents a single run or an average.","section":"Fig. 5"},{"comment":"The training details list epoch counts (600, 200, 300) but not batch size or whether the same training budget was kept for all losses; these details are needed for reproducibility.","section":"Sec. 3.2"},{"comment":"Reference [5] appears twice in the text: once for the U-Net architecture (Ronneberger et al.) and once for Bertels et al. on optimizing the Dice score; please fix the citation numbering.","section":"References"},{"comment":"The qualitative examples are illustrative, but the selection criteria for the displayed slices are not stated; please specify whether they are selected randomly or by other criteria.","section":"Figs. 2-4"}],"recommendation":"major_revision","confidential_remarks":"The paper tackles a practical and timely problem, and the proposed loss is simple enough to be adopted if the empirical claims hold. My main concern is the lack of a validation-based tuning protocol for the dilation radius and the absence of any uncertainty quantification; both are fixable within a revision. I would not reject on novelty grounds alone, but the current evidence is too fragile to support the strong concluding claims. If the authors can provide multiple seeds and a principled tuning rule, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: the paper proposes DBCE, which replaces class-frequency weights in balanced CE with inverse dilated-area weights, combined by a max across classes. The idea is simple and the motivation is sound: put weight on the object plus a rim around it, where false positives concentrate. I haven't seen this exact weighting in the literature, so it is a modest but real variant.\n\nThe evaluation is honest in scope: three datasets (polyp, skin, multi-organ), same network and training conditions across losses, and balanced CE plus Dice+CE as baselines. The qualitative figures show the expected effect—balanced CE gives many boundary false positives, and DBCE reduces them. The paper also shows a dilation-radius sweep with an optimum at R=20, which is more than many loss papers do.\n\nSoft spots: all results are single runs with no error bars or significance tests. The gains over CE are small—+0.61, +0.97, +0.58 mDice across the three datasets—so the differences could easily be noise. Second, the per-dataset dilation radius is set \"empirically\" without a described validation protocol. Figure 5 shows strong dependence on R; if R was chosen by looking at test performance, the comparison is biased, and the small advantage over CE may not survive a clean tuning rule. The paper is transparent about the radii, but not about how they were selected. Third, no code is released, which makes reproduction harder. The Synapse results show DBCE matching Dice+CE on some organs and losing on others, consistent with \"similar\" but not a clear win.\n\nI don't think the central argument collapses. A weighted CE with spatially aware weights can plausibly imitate Dice+CE, and the experiments point that way. But the reported advantage over CE is small and unquantified, and the one hyperparameter is tuned without a stated validation split. This is a legitimate empirical loss-design paper, and I would send it to review, but with a strong request for multiple seeds, confidence intervals, validation-based radius selection, and code. With those additions it would be a solid incremental contribution to the segmentation loss literature.\n\nThis is for people working on loss functions for imbalanced segmentation; it gives a clear description of a new weighting and a useful baseline. It won't change the field, but it's worth a look after the statistical gaps are addressed.","headline":"A simple new weighting scheme for cross-entropy that performs comparably to Dice+CE, but single-run results and per-dataset tuning of the one hyperparameter leave the practical claim conditional.","tokens_in":7995,"tokens_out":2420,"would_cite":false,"duration_ms":23927,"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 weighted cross-entropy loss that also weights each object's surrounding pixels can match the performance of Dice+CE in medical segmentation.","keywords":["dilated balanced cross entropy loss","weighted cross entropy","class imbalance","medical image segmentation","polyp segmentation","skin lesion segmentation","multi-organ segmentation","loss function design"],"falsifier":"Select $R$ by cross-validation on a validation split for each task, then compare DBCE against plain CE and Dice+CE on the same held-out test sets; if DBCE no longer beats CE or matches Dice+CE, the claim is settled false. A more mechanism-specific test would measure the false-positive rate in the ring around each object, since DBCE should cut it relative to balanced CE.","tokens_in":7185,"feed_emoji":"🩺","tokens_out":8163,"duration_ms":76305,"temperature":0.7,"pith_summary":"Sparse foreground objects and large backgrounds make medical segmentation a class-imbalance problem. The proposed dilated balanced cross-entropy (DBCE) loss reweights plain cross-entropy not by raw class frequency but by the area of each class mask after dilation, so small objects and the ring of background around them receive high weight. On three medical segmentation tasks the paper reports that DBCE beats both plain CE and balanced CE and matches the popular Dice+CE combination. The intended upshot is that a weighted CE loss with the right spatial weighting can handle imbalance as well as a region-based loss.","feed_headline":"Dilated weighting lifts cross entropy to Dice-level accuracy","feed_subtitle":"Weighting each object plus its surrounding pixels lets plain cross-entropy keep up with popular region-based losses.","key_machinery":"The machine is the pixel-wise weight map $M = \\max_c W_c$, where $W_c$ is the dilated, area-normalized mask of class $c$. Given binary ground-truth masks $B_c$ and a disk structuring element of radius $R$, each mask is dilated to $B'_c$, normalized by its area to $W_c = B'_c / \\sum B'_c$, and the per-pixel maximum over classes is the loss weight. The CE loss map is then multiplied pointwise by $M$ and summed. This construction carries the argument: it moves balancing weight from the object to its uncertain boundary zone, bounds the weight of tiny classes through dilation, and lets the smallest object dominate where masks overlap.","core_discovery":"The central claim is that balanced cross-entropy fails because it over-weights the object and under-weights the surrounding background, and that this can be fixed by dilating each class mask before computing balancing weights. In DBCE, each binary class mask is dilated with a disk-shaped structuring element, the dilated mask is normalized by its own area, and at every pixel the loss weight is the maximum of these normalized values across classes; this weight map is multiplied into the CE loss map. Because the max operation makes the smallest object dominate overlapping dilated regions, and because dilation sets a floor on the effective area of tiny classes, the weighting is both stronger for small structures and less noisy than inverse-frequency balancing. The paper's experimental claim is that, with only the dilation radius as an added hyperparameter, this weighting outperforms CE and balanced CE on polyp, skin-lesion, and multi-organ segmentation, and is similar to Dice+CE.","pith_inferences":["The paper leaves implicit that the dilation radius could be annealed over training: a large radius early would stabilize learning, and shrinking it later would sharpen boundary emphasis; this is a testable extension the authors do not report.","Because the weight map is built from normalized dilated masks, DBCE behaves like a soft distance-to-object attention; connecting it to boundary losses or distance-transform losses could yield variants the paper does not discuss.","A portability test the paper does not run is to fix one radius across many datasets or choose $R$ by validation rather than by inspecting test performance; until such a rule exists, the claimed advantage over plain CE remains conditional on dataset-specific tuning.","The weighting recipe is class-agnostic, so it could transfer to non-medical sparse-object segmentation, such as aerial or satellite imagery, though the paper only evaluates medical benchmarks."],"forward_implications":["With the right spatial weighting, a weighted cross-entropy loss can match a region-based loss like Dice+CE on binary and multiclass medical segmentation.","Per-sample weights can be computed on the fly from each label map, removing the need to precompute class frequencies over the whole training set.","The method specifically lowers false positives in the area around objects, the failure mode that plain balanced CE worsens.","The dilation radius $R$ is a single interpretable knob: $R=0$ reproduces balanced CE, large $R$ erases the balancing effect, and intermediate values give the reported improvement.","DBCE reaches Dice-comparable accuracy without changing the network, data, or training schedule, so the reported gains are attributable to the loss function."],"supporting_citations":[{"why":"Documents that balanced CE generally underperforms plain CE in medical segmentation and reviews loss options, setting up the problem the paper addresses.","marker":"[1]"},{"why":"Introduces the weighted/balanced CE formulation and the segmentation architecture that DBCE modifies.","marker":"[4]"},{"why":"Supplies the claim that balanced CE may be improved by a better choice of weighting, the direct motivation for DBCE.","marker":"[5]"},{"why":"Reports that polyp segmentation errors concentrate around the polyp boundary, the spatial motivation for dilating the mask.","marker":"[7]"},{"why":"Provides the Kvasir-SEG polyp training set used in the first evaluation.","marker":"[8]"},{"why":"Provides the CVC-ClinicDB polyp test set used for cross-dataset evaluation.","marker":"[9]"},{"why":"Provides the ISIC 2018 skin lesion data used in the second evaluation.","marker":"[10]"},{"why":"Supplies the Synapse data split of 18 training and 12 test scans used in the multi-organ experiment.","marker":"[12]"},{"why":"A recent medical segmentation study that uses the Dice+CE combination, which DBCE is compared against.","marker":"[15]"},{"why":"A widely used segmentation model whose Dice+CE loss variant serves as a strong popular baseline in the comparison.","marker":"[19]"}],"fun_headline_variants":["Dilating class masks boosts cross entropy to Dice+CE level","DBCE: cross entropy catches up with Dice+CE","Dilation-based weighting lifts CE to Dice+CE match","Simple dilation rescues balanced cross entropy for segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's advantage rests on choosing the dilation radius $R$ per dataset, and the paper sets $R$ empirically (20 pixels for two tasks, 10 for the third) without reporting a validation-based selection rule; if $R$ was chosen to fit the test results, the reported improvements over plain cross-entropy may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Dilating class masks boosts cross entropy to Dice+CE level","DBCE: cross entropy catches up with Dice+CE","Dilation-based weighting lifts CE to Dice+CE match","Simple dilation rescues balanced cross entropy for segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001039,"raw_usage":{"total_tokens":4394,"prompt_tokens":988,"completion_tokens":3406,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":3338}},"tokens_in":604,"tokens_out":3406,"duration_ms":21399,"temperature":1.0,"reasoning_tokens":3338,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:03:17.909231+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Select $R$ by cross-validation on a validation split for each task, then compare DBCE against plain CE and Dice+CE on the same held-out test sets; if DBCE no longer beats CE or matches Dice+CE, the claim is settled false. A more mechanism-specific test would measure the false-positive rate in the ring around each object, since DBCE should cut it relative to balanced CE.","supporting_citations":[{"cited_title":"saliency maps from physicians","cited_arxiv_id":null,"evidence_quote":"Provides the CVC-ClinicDB polyp test set used for cross-dataset evaluation."},{"cited_title":"In: Medical Image Analysis","cited_arxiv_id":null,"evidence_quote":"Documents that balanced CE generally underperforms plain CE in medical segmentation and reviews loss options, setting up the problem the paper addresses."},{"cited_title":"In: Medical Image Computing and Computer-Assisted Intervention–MICCAI (2015)","cited_arxiv_id":null,"evidence_quote":"Introduces the weighted/balanced CE formulation and the segmentation architecture that DBCE modifies."},{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention–MICCAI (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies the claim that balanced CE may be improved by a better choice of weighting, the direct motivation for DBCE."},{"cited_title":"In: Multimed ia Modeling: 26th International Con- ference, MMM (2020)","cited_arxiv_id":null,"evidence_quote":"Provides the Kvasir-SEG polyp training set used in the first evaluation."},{"cited_title":"Multi-scale hierarchical v ision transformer with cascaded at- tention decoding for medical image segmentation","cited_arxiv_id":null,"evidence_quote":"A recent medical segmentation study that uses the Dice+CE combination, which DBCE is compared against."},{"cited_title":"In: IEEE transactions on medical imag- ing","cited_arxiv_id":null,"evidence_quote":"A widely used segmentation model whose Dice+CE loss variant serves as a strong popular baseline in the comparison."}],"review_version":1}