{"id":"9d596774-8a2e-4199-9d34-13e81aebc019","arxiv_id":"1908.06933","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"DALS, a CNN-guided level-set active contour model with per-pixel parameter maps, improves lesion segmentation across MR and CT images compared with U-Net and a backbone CNN.","lead":"This paper combines a neural network with a classic 'snake' contour method to automatically find lesion boundaries in medical images. The hybrid is tested on MR and CT images of brain, liver, and lung lesions, and reported to beat plain neural networks and plain contour methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported Wilcoxon p-values are not attainable with the stated per-organ test sizes; statistical support for 'significantly more accurate' is not established.","rationale":"The method is coherent and the ablation is directionally consistent, so this is not a rejection of the approach. The code URL and architectural description are real assets. However, the central claim uses 'significantly' in a statistical sense, and the paper's own Section 3 and Section 4 numbers do not cohere: the lung CT test set is too small for the reported p<0.001 if the test is per-organ. The most likely resolution is that the Wilcoxon test was run on pooled test cases across all four organs, but then it cannot support the sentence 'superior accuracies under all metrics and in all datasets' at the per-dataset level. This matters because the margins over the CNN backbone are modest and the CIs overlap on two datasets. The reader's conditional verdict is appropriate; a definitive judgment would require the missing per-image data or a repeated-split evaluation.","tokens_in":5826,"tokens_out":10123,"duration_ms":104002,"concrete_test":"Compute the exact two-sided Wilcoxon signed-rank p-values implied by the reported per-organ Dice arrays and test sizes. The authors should supply, or the released code should reproduce, the per-image Dice scores for each organ. With the Section 3 split, lung CT has 8-9 test cases; the minimum possible two-sided p for n=9 is 0.0039, so p<0.001 cannot be true if the test is per-organ. If the exact recomputation shows no per-organ p<0.001, the paper's significance claim must be restated as pooled, and the Table 2 CIs should be treated as overlapping and inconclusive on lung/liver CT. A confirmatory check is to rerun the lung/liver CT comparison under 10 repeated random splits and report the distribution of the DALS-minus-backbone Dice difference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the statistical inference from Table 2 to the claim that DALS yields 'significantly more accurate boundaries ... in all datasets.' Section 3 fixes the split at 85/10/5, which leaves only 8-9 test images for the lung CT subset (87 images) and about 11 for liver CT. Section 4 then reports Wilcoxon paired-test p<0.001 vs U-Net and p<0.005 vs the backbone. For a paired Wilcoxon signed-rank test with n=9, the smallest possible exact two-sided p is 2/2^9 = 0.0039; p<0.001 is unattainable. Either the significance test was pooled across organs (in which case it does not independently support the per-dataset 'all datasets' sentence) or the effective test size is larger than the stated split implies. In either case, the per-organ comparison is not statistically supported: the Table 2 confidence intervals for DALS and its CNN backbone overlap on lung CT (Dice 0.869 with CI 0.092 vs 0.822 with CI 0.094) and on liver CT. A single 10% split, with no repeated resampling and no patient-level split described, is a thin basis for the central significance claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Deep Active Lesion Segmentation (DALS), a framework combining a fully convolutional encoder-decoder with a level-set active contour model (ACM). The CNN predicts a lesion probability map that initializes the signed distance function and also produces per-pixel parameter maps λ1(x,y) and λ2(x,y) used in a localized Chan-Vese style energy functional. The method is evaluated on a new Multiorgan Lesion Segmentation (MLS) dataset with brain MR, lung CT, liver CT, and liver MR images, comparing against U-Net, a manually initialized level-set ACM, and the authors' CNN backbone. The authors report improved Dice, Hausdorff distance, and Boundary F1 scores, and claim statistically significant superiority via Wilcoxon paired tests.","tokens_in":6065,"tokens_out":2405,"duration_ms":25827,"significance":"If the reported results hold, the work makes a useful contribution to medical image segmentation by demonstrating a practical method for learning per-pixel ACM energy parameters from a CNN, moving beyond scalar or hand-tuned λ values. The release of source code and the introduction of the MLS dataset are positive aspects that support reproducibility and further research. The central architectural idea is plausible and the paper includes a comparison against a strong U-Net baseline, which lends credibility to the approach.","major_comments":[{"comment":"The statistical support for the headline claim of 'significantly more accurate boundaries' is not established. The paper reports Wilcoxon paired test p-values of p<0.001 against U-Net and the level-set baseline, and p<0.005 against the CNN backbone, but with the stated split (85/10/5) the lung CT test set consists of roughly 9 images and the liver CT test set of roughly 11 images. For a paired Wilcoxon test with n=9, the smallest attainable two-sided p-value is 2/2^9 = 0.0039, so p<0.001 cannot be correct if the test was performed per organ. If the test was pooled across organs, then it does not independently support the per-dataset claim 'in all datasets' made in the same paragraph. In addition, the confidence intervals reported in Table 2 for DALS and the CNN backbone overlap on lung CT (Dice 0.869 ± 0.092 vs 0.822 ± 0.094) and on liver CT (0.846 ± 0.081 vs 0.801 ± 0.159), which further undermines the per-organ significance claim. The authors must specify the exact test procedure, provide per-organ p-values or confidence intervals of the differences, and correct the reported p-values if they are unattainable.","section":"Section 4, Table 2"},{"comment":"The evaluation rests on a single 10% test split per organ, with no repeated resampling, cross-validation, or description of patient-level splitting. For the lung CT subset, the test set is only about 8–9 images, which is too small to support general claims about 'all datasets' without additional evidence. The paper should report the exact number of test images per organ, clarify whether images from the same patient appear across train/test splits, and ideally provide leave-one-out or repeated random sub-sampling results to demonstrate that the reported gains are not an artifact of a particular split.","section":"Section 3, data split"},{"comment":"The manually initialized level-set ACM baseline is unsupervised: it does not use the training labels, while DALS is fully supervised. Comparing a supervised method against an unsupervised baseline does not isolate the contribution of the per-pixel parameter functions, because the baseline does not receive any task-specific tuning. A fairer ablation would either train the same backbone with scalar λ constants (for example, by integrating the level-set evolution into the training loop, as in the cited deep level-set works) or at least tune a scalar λ on the validation set and then compare. Without such a control, the claim that the learned parameter maps are responsible for the improvement is not convincingly supported.","section":"Section 4, Algorithm Comparison"}],"minor_comments":[{"comment":"The sentence 'we introduce Deep Active Lesion Segmentation (DALS), a fully automated segmentation framework that leverages...' contains a grammatical issue ('that leverages' is split from its antecedent by 'framework'); consider revising for readability.","section":"Abstract"},{"comment":"The exponential forms for λ1 and λ2 are presented without motivation for the specific formulas. A brief explanation of why the chosen exponential functions are appropriate, or a reference to the prior work that used them, would help readers understand the design choice.","section":"Section 2.3, Eq. (3)"},{"comment":"In the Liver CT row, the Hausdorff CI for the CNN Backbone is listed as 1.6, which is an order of magnitude larger than all other CI values in the table and appears to be a typographical error. Please check whether this should be 0.16 or another value.","section":"Table 2"},{"comment":"The Wilcoxon test description does not state whether the test is one-sided or two-sided, how ties are handled, or whether any multiple-comparison correction was applied across the four organs and three comparisons. Please provide these details.","section":"Section 4, statistics"},{"comment":"The tensor dimensions shown in Figure 2 are not fully consistent: for example, the input is [256,256,1] but the first feature block is [128,128,64], and the concatenated feature maps before the multiscale dilation block have different listed sizes. Clarify the notation to show how these dimensions are obtained.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The statistical issue in Section 4 is serious enough to require correction before publication, but it is fixable. I would also note that the MLS dataset is not released while the code is; if the authors can release the dataset or a benchmark version, the reproducibility of the work would be substantially strengthened. The paper's scope fits a MICCAI-style venue, and the architectural contribution is timely."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, unglamorous hybrid method—CNN gives an initial probability map and per-pixel weight maps for a level-set ACM—plus a new multi-organ lesion dataset. The paper is honest about what it's adding, but the statistical support for the headline claim is not there.\n\nThe genuinely new bit is the generalization of the scalar Chan-Vese lambdas to per-pixel parameter functions lambda1(x,y), lambda2(x,y), computed from the CNN probability map via an exponential transform. The multiscale dense encoder-decoder with dilation rates 2-16 is a reasonable architecture. The MLS dataset (brain MR, lung CT, liver MR, liver CT) is a useful contribution, and the source code is public.\n\nThe ablation consistently shows DALS beating its own backbone on mean Dice for every organ, and the boundary plots show the level-set refinement does something. That's real evidence.\n\nNow the soft spots, in order of severity. The significance claims don't hold up. With an 85/10/5 split, the lung CT test set is about 9 images. A paired Wilcoxon test on n=9 cannot produce p<0.001; the smallest two-sided p is roughly 0.0039. Either the test was pooled across organs, in which case it doesn't support the 'all datasets' sentence, or it's just wrong. Either way, 'significantly more accurate' is an overstatement. The confidence intervals in Table 2 overlap between DALS and the backbone for lung CT and liver CT, so the mean gain is not clearly significant on those organs. The manual level-set baseline is unsupervised and initialized by hand, so it's not a fair comparison. Also, the lambda maps in Eq. 3 are not learned network outputs; they are deterministic functions of Y_prob. Calling them 'learned parameter maps' is misleading. Finally, no dataset release, no commit hash, and the ACM evolution details (iterations, convergence criteria) are missing.\n\nThe core idea is coherent and the code makes reimplementation feasible. But the statistical evidence as reported is too thin for the claims. I'd send it out for review with a request to fix the stats and re-analyze with proper per-organ tests and ideally multiple splits. The paper is useful for anyone working on combining CNNs with active contours; just don't trust the significance asterisks.","headline":"Solid CNN+level-set hybrid with a new multi-organ dataset, but the per-organ significance claims exceed what the test sizes can support.","tokens_in":6619,"tokens_out":3693,"would_cite":false,"duration_ms":33640,"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":"Deep Active Lesion Segmentation couples a CNN encoder-decoder with a level-set active contour model, using learned per-pixel parameter maps to initialize and tune the contour, and reports significantly more accurate lesion boundaries than…","keywords":["lesion segmentation","active contour models","level sets","convolutional neural networks","medical image segmentation","per-pixel parameter maps","multiorgan dataset","MR and CT imaging"],"falsifier":"Run DALS and its CNN backbone on repeated random splits or k-fold cross-validation of each MLS organ subset; if the Dice advantage shrinks or the Wilcoxon p-value rises above 0.05 in any organ, the claimed significant boundary improvement is not supported.","tokens_in":5634,"feed_emoji":"🩻","tokens_out":7325,"duration_ms":64733,"temperature":0.7,"pith_summary":"The paper introduces a fully automatic framework that combines a convolutional neural network with a level-set active contour model for lesion segmentation in medical images. The network learns an initial probability map and per-pixel parameter maps that initialize and guide the contour evolution, replacing manually tuned scalar weights with spatially varying ones. The authors claim this yields significantly more accurate boundaries than conventional CNNs and can reliably segment lesions of various sizes. They report superior Dice, Hausdorff distance, and boundary F-measure results across brain, liver, and lung lesions in MR and CT, with the largest advantages appearing when training data are small. If the claim holds, it would remove two practical obstacles in medical image segmentation: manual contour initialization and parameter tuning.","feed_headline":"CNN and level-set contour team up to trace lesions more accurately","feed_subtitle":"A learned per-pixel parameter map lets active contours hug irregular lesion edges in MR and CT.","key_machinery":"The load-bearing mechanism is a per-pixel-parameterized level-set energy functional, a generalization of the Chan-Vese active contour energy in which the scalar constants $\\lambda_1$ and $\\lambda_2$ become functions $\\lambda_1(x,y)$ and $\\lambda_2(x,y)$ over the image domain. An encoder-decoder CNN with dense blocks and a multiscale dilated-convolution block produces the initial probability map $Y_{\\mathrm{prob}}$; the Transformer converts it into the signed distance map that initializes the contour, and the exponential formulas above convert it into the parameter maps. During evolution, these local weights let the contour bend precisely to irregular, low-contrast lesion boundaries, which is the mechanism the paper credits for its boundary accuracy.","core_discovery":"The central claim is that lesion segmentation can be made both fully automatic and boundary-precise by letting a CNN supply the initial contour and the spatially varying weights of a level-set active contour model. Concretely, the CNN outputs a probability map $Y_{\\mathrm{prob}}(x,y)$; a transformer converts it into a signed distance map $\\varphi(x,y,0)$ that seeds the contour, and the same map defines per-pixel parameters $\\lambda_1(x,y)=\\exp\\big((2-Y_{\\mathrm{prob}})/(1+Y_{\\mathrm{prob}})\\big)$ and $\\lambda_2(x,y)=\\exp\\big((1+Y_{\\mathrm{prob}})/(2-Y_{\\mathrm{prob}})\\big)$ in the energy functional. The contour then evolves by the Euler-Lagrange PDE of a localized Chan-Vese energy, so each boundary point carries its own interior and exterior weighting. On the paper's Multiorgan Lesion Segmentation dataset, DALS reports higher Dice and boundary metrics than U-Net, a manually initialized level-set model, and its own CNN backbone, and the authors state that the improvements are statistically significant.","pith_inferences":["A natural ablation not reported in the paper: replace the learned $\\lambda_1(x,y)$ and $\\lambda_2(x,y)$ maps with their spatial means while keeping the same initialization; if boundary metrics degrade but Dice does not, the parameter maps are doing the boundary work.","Because the exponential transforms amplify $Y_{\\mathrm{prob}}$ values near 0 and 1, a poorly calibrated probability map could produce extreme local weights; measuring calibration per organ would predict where DALS is most likely to fail.","The same pipeline should transfer to non-radiology segmentation tasks with irregular boundaries, such as ultrasound lesions or cells, since the level-set evolution is agnostic to image modality.","The significance claim could be stress-tested by repeated cross-validation; the reported 10 percent test splits leave as few as nine lung images, and several confidence intervals overlap with the backbone CNN."],"forward_implications":["Lesion segmentation becomes fully automatic: no manual contour initialization and no hand-picked active-contour weights, because the network supplies both.","Boundary quality should improve most where intensity contrast is low or the lesion edge is irregular, exactly the cases pure CNNs handle poorly.","The same architecture can be retrained for other organs and modalities without changing the level-set machinery.","The method should remain usable with small training sets, because the CNN only needs to produce a coarse localization and parameter maps rather than a pixel-perfect boundary."],"supporting_citations":[{"why":"Supplies the base Chan-Vese region-based energy that DALS generalizes by replacing scalar weights with parameter functions.","marker":"[1]"},{"why":"Introduces active contour models, the family of methods DALS extends.","marker":"[6]"},{"why":"Provides the localized region-based formulation used to compute interior and exterior mean intensities in a window.","marker":"[7]"},{"why":"Shows CNN-estimated active contour parameters; DALS extends this to per-pixel maps and full automation.","marker":"[3]"},{"why":"U-Net is the main CNN baseline and the comparison target for boundary accuracy.","marker":"[9]"},{"why":"Prior deep level-set approach with fixed weighting; contrast for learned per-pixel parameter maps.","marker":"[4]"},{"why":"Prior CNN plus active-contour approach requiring manual initialization; contrast for the fully automatic pipeline.","marker":"[8]"}],"fun_headline_variants":["CNN seeds active contours for sharp lesion borders","Deep active contours learn lesion boundaries","Fully automated lesion segmentation with CNN-ACMs","Hybrid model traces irregular lesion edges","DALS: neural nets guide level-set contours"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 10-percent held-out test subsets of the MLS dataset, as few as nine lung images, are representative enough to support the reported significance given that DALS's confidence intervals overlap with its own CNN backbone on several organs.","fun_headline_variants_meta":{"raw":{"variants":["CNN seeds active contours for sharp lesion borders","Deep active contours learn lesion boundaries","Fully automated lesion segmentation with CNN-ACMs","Hybrid model traces irregular lesion edges","DALS: neural nets guide level-set contours"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000843,"raw_usage":{"total_tokens":3671,"prompt_tokens":941,"completion_tokens":2730,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":2664}},"tokens_in":557,"tokens_out":2730,"duration_ms":18131,"temperature":1.0,"reasoning_tokens":2664,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:30:12.384563+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DALS and its CNN backbone on repeated random splits or k-fold cross-validation of each MLS organ subset; if the Dice advantage shrinks or the Wilcoxon p-value rises above 0.05 in any organ, the claimed significant boundary improvement is not supported.","supporting_citations":[{"cited_title":"IEEE Transactions on Image Processing 10(2), 266--277 (2001)","cited_arxiv_id":null,"evidence_quote":"Supplies the base Chan-Vese region-based energy that DALS generalizes by replacing scalar weights with parameter functions."},{"cited_title":"International Journal of Computer Vision 1(4), 321--331 (1988)","cited_arxiv_id":null,"evidence_quote":"Introduces active contour models, the family of methods DALS extends."},{"cited_title":"IEEE Transactions on Image Processing 17(11), 2029--2039 (2008)","cited_arxiv_id":null,"evidence_quote":"Provides the localized region-based formulation used to compute interior and exterior mean intensities in a window."},{"cited_title":"IEEE Transactions on Medical Imaging 36(3), 781--791 (2017)","cited_arxiv_id":null,"evidence_quote":"Shows CNN-estimated active contour parameters; DALS extends this to per-pixel maps and full automation."},{"cited_title":"In: LNCS vol","cited_arxiv_id":null,"evidence_quote":"U-Net is the main CNN baseline and the comparison target for boundary accuracy."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Prior deep level-set approach with fixed weighting; contrast for learned per-pixel parameter maps."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Prior CNN plus active-contour approach requiring manual initialization; contrast for the fully automatic pipeline."}],"review_version":1}