{"id":"ec1bdc03-ff3c-4c9a-9696-da92462bf875","arxiv_id":"2501.09116","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A lightweight distance-map regression network and a distance-weighted Dice loss improve small object segmentation in medical images.","lead":"This paper presents a deep learning method that adds a small distance-map regression network to an existing segmentation network to improve the segmentation of small tumors, such as liver tumors, in CT images. The method also uses a new loss that weights errors near object boundaries using distance maps.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The frozen LR-Net is trained only on hard binary masks but is applied to soft probability maps during M-Net training; the claimed 'rigorous mapping' does not cover this distribution shift, and the paper provides no evidence that it is benign.","rationale":"The central claim is an empirical one, and the paper presents consistent tables that support a modest improvement. However, the proposed mechanism depends on the frozen LR-Net producing meaningful distance-map regressions from soft probability maps, a regime it never saw during pre-training. This is not a disagreement with consensus; it is a testable train/test mismatch internal to the method. Table 2 row 4, where unfreezing the LR-Net degrades performance, makes the fixed binary-trained LR-Net the critical component, but no experiment characterizes its response on soft inputs. If the proposed test shows large soft-input regression error, then the 'rigorous mapping' claimed in Section 2.2 is not in effect during training, and the improvement would need a different explanation, such as the auxiliary loss acting mainly as a regularizer, which would weaken the stated contribution. If soft-input error is comparable to binary-input error, the assumption is benign and the framework is plausible. The reader's conditional verdict is appropriate: the missing evidence is an ablation or calibration check, not a fundamental contradiction. I would keep the verdict conditional, pending this test and the already-flagged need for error bars and released code.","tokens_in":6861,"tokens_out":11099,"duration_ms":122107,"concrete_test":"On the LiTS validation set, take the frozen LR-Net from Section 3 and feed it (a) the ground-truth binary masks and (b) the softmax probability maps produced by a trained M-Net without LR-Net; compute smooth-L1 error against the GT NI-DM for both. If the soft-input error is more than roughly 2x the binary-input error, the LR-Net does not generalize from hard to soft inputs, and the claimed distance-map supervision during joint training is not faithful. This single comparison directly tests the distribution-shift assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism (Section 2.2) is that a frozen LR-Net, pre-trained to regress NI-DM from ground-truth binary masks, provides reliable distance-map supervision when fed soft probability maps from the M-Net during joint training. This is a train/test distribution shift: during pre-training the LR-Net sees only hard one-hot inputs, while during M-Net training it receives continuous softmax outputs in (0,1). The paper asserts that the rigorous mapping between mask and distance map 'can guarantee the correspondence' between M-Net outputs and LR-Net outputs, but a function learned on hard inputs need not behave sensibly on soft inputs. The regression loss could in principle be minimized by the M-Net producing probability maps that exploit LR-Net's unvalidated soft-input response rather than by learning true shape. Table 2 row 4, where unfreezing the LR-Net degrades DC from 0.748 to 0.628, is indirect evidence that the regression objective is easy to game and that the frozen, binary-trained LR-Net is load-bearing; yet no experiment characterizes LR-Net's soft-input behavior. If LR-Net's distance predictions on soft maps are inaccurate, the claimed shape-aware supervision is not actually delivered, and the reported 0.751 DC improvement over WCE, GDS, Tversky, Focal Tversky, and ExpLog is not explained by the stated mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a segmentation framework that couples a main segmentation network (M-Net) with a lightweight distance-map regression network (LR-Net). The LR-Net is pre-trained on ground-truth binary masks to regress a normalized inverse distance map (NI-DM), then frozen; during joint training it receives soft probability maps from the M-Net and produces distance-map predictions, which are supervised by a smooth-L1 regression loss. In addition, the authors introduce a MapDice loss that weights a Dice-like term by the ground-truth distance map. Experiments on the MICCAI 2017 LiTS dataset and on a clinical CECT dataset report seven metrics and compare the method with several imbalance-oriented losses (WCE, GDS, Tversky, Focal Tversky, Exp-Log) as well as two earlier distance-map regression approaches. The best configuration reaches a Dice per case of 0.751 on the clinical dataset, outperforming the compared baselines, and the authors claim state-of-the-art or competitive performance.","tokens_in":7199,"tokens_out":7508,"duration_ms":72189,"significance":"If the mechanism is validated, the idea of pre-training a lightweight network to compute distance maps from masks and then freezing it as a differentiable shape-aware module is attractive: it can be plugged into any segmentation network and provides geometric supervision beyond binary masks. The paper also provides a useful comparison of several distance-map variants and imbalance losses on a clinical dataset. However, the support for the central claims is weakened by several definitional issues, an unreported balancing parameter, and the lack of any characterization of the LR-Net's behavior on soft probability inputs. The paper is not circular, and the empirical comparison is a genuine strength, but the current presentation leaves load-bearing points underspecified.","major_comments":[{"comment":"Equation (1) as written does not produce values in [0,1] as claimed. For a boundary voxel where D(x)=0, the formula gives φ(x) = (M+1)/M > 1, where M is the local maximum distance; the range is [1/M, 1+1/M]. If the intended normalization is instead (M−D(x))/M or (M+1−D(x))/(M+1), then the text and all downstream experiments need to be corrected, because the NI-DM is both the regression target for the LR-Net and the penalty map in MapDice. Please also clarify the notation: the variable x is used both for the evaluation point and as the index in the maximum over the connected component C.","section":"§2.1, Eq. (1)"},{"comment":"Equation (2) is not a well-defined scalar loss: it sums over classes but not over spatial positions, leaving the expression as a per-voxel tensor. Moreover, the term (p_c×φ_c+ε)/(p_c+φ_c+ε) evaluates to approximately 1 for background voxels where p_c=0 and φ_c=0, so unless the spatial sum is specified, the loss contains a large constant offset. Please provide the complete voxel-wise formulation and clarify how ε and the class sum are applied.","section":"§2.3, Eq. (2)"},{"comment":"The total loss includes α·L_smoothL1, but α is never defined, and its value is not reported anywhere. This is a free parameter tuned on the 13-case LiTS validation split, so the reported comparisons are not reproducible. The claim that α is 'verified to be effective for the performance improvement' is also unsupported by any ablation or sensitivity study. Please report the value of α and, ideally, a small sensitivity analysis.","section":"§2.3 and Table 1"},{"comment":"The training pipeline pre-trains the LR-Net on ground-truth binary masks and then freezes it, but during joint training the LR-Net receives soft probability maps from the M-Net. The paper asserts that the rigorous mapping trained on binary inputs 'can guarantee the correspondence' between M-Net outputs and LR-Net outputs, but no evidence is provided that the LR-Net generalizes from hard to soft inputs. The large drop when the LR-Net is unfrozen (DC 0.628 vs. 0.748, Table 2, rows 4 and 5) shows that the regression objective is sensitive to the training scheme, but it does not characterize the soft-input behavior. Please add a direct experiment on the LR-Net's soft-input predictions (e.g., distance error or visual comparison), or retrain the LR-Net with soft targets, to justify the frozen-binary design.","section":"§2.2 and Table 2, row 4"},{"comment":"All results are reported as single runs with no error bars or significance tests. The LiTS validation set contains only 13 cases, and the differences among the top configurations in Table 1 (e.g., 0.7463 vs. 0.7445) are small enough to be within noise. In addition, the clinical dataset section states that 'all the cases are used for testing' but does not describe how the model was trained or selected for this dataset; please clarify whether the LiTS-trained checkpoint was applied directly, and report standard deviations or confidence intervals over at least a few runs.","section":"§3, Tables 1–2"}],"minor_comments":[{"comment":"The abstract and conclusion claim to outperform 'other existing state-of-the-arts,' but the experimental comparison is limited to five loss functions and two earlier regression-based methods. No recent segmentation networks are compared, so the SOTA claim is broader than the evidence.","section":"Abstract and §3.2"},{"comment":"The text cites baseline results of MNet+LDice (0.6581) and MNet+MapDice (0.6856), but these rows are not present in Table 1, which makes the comparison difficult to follow.","section":"Table 1"},{"comment":"The clinical dataset description says all 137 cases are used for testing but does not state whether any clinical data were used for training or validation; please specify the training/evaluation protocol for this dataset.","section":"§3, Dataset"},{"comment":"There are several typos: 'During reference' should be 'During inference,' 'indicates indicates' is repeated, 'reslults' should be 'results,' 'Tvserky' should be 'Tversky,' 'I-DM is derived from the I-DM' should presumably read 'I-DM is derived from the O-DM,' and the Figure 3 caption repeats 'Fig. 3.'","section":"§2.2 and Figure 3"},{"comment":"In Eq. (1), the connected component C is not formally defined in the notation; please define it as a set of voxels belonging to one object component in the binary mask.","section":"§2.1"},{"comment":"The LiTS leaderboard submission is reported without a date or version; since leaderboards change over time, this result should be described as a single submission rather than a stable state-of-the-art comparison.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the empirical comparison is useful, but the manuscript currently has several load-bearing gaps: the distance-map normalization in Eq. (1) is inconsistent with the stated [0,1] range, Eq. (2) is not a well-defined scalar loss, the value of α is missing, and the frozen LR-Net's behavior on soft inputs is unvalidated. These are fixable with additional experiments and a careful rewrite, so I do not see grounds for rejection. The statistical weakness (no error bars, 13-case validation set) should also be addressed explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an incremental but useful combination for imbalanced medical image segmentation. The new piece is a separate regression network (LR-Net) trained on ground-truth masks, then frozen, that maps the M-Net's soft probability map to a distance map; the MapDice loss uses the ground-truth distance map as a spatial penalty. Reported Dice gains on small liver tumors are consistent across seven metrics and two datasets, and that is a real plus.\n\nWhat is actually new: the specific pipeline—pre-training the LR-Net on hard masks, freezing it, and supervising the M-Net with both a smooth-L1 regression loss on the predicted distance map and the distance-weighted Dice—is not in the cited distance-transform work. The comparisons against multi-branch and cascaded regression baselines are appropriate and show the design choices matter.\n\nSoft spots, in proportion. First, the mechanism is under-verified. The LR-Net is trained only on binary masks but applied to soft probability maps; a function learned on hard inputs need not behave sensibly on continuous inputs. The paper asserts the rigorous mapping guarantees correspondence, but that is an assertion, not a result. The drop from 0.748 to 0.628 when the LR-Net is unfrozen suggests the regression objective can be gamed. This does not sink the empirical claim, but it means the stated explanation for the gains is not established. An analysis of LR-Net outputs on soft maps, or a comparison with a version trained on soft masks, would close the gap.\n\nSecond, the 'outperforms state-of-the-arts' claim is broader than the evidence. The comparison set is loss functions and two distance-map variants, not recent full segmentation systems. The LiTS leaderboard entry helps, but it is one table row.\n\nThird, no error bars, no significance test, and alpha is not reported. The LiTS validation set is 13 cases, so Table 1 differences are not necessarily stable. These are medium-size issues for an empirical paper.\n\nMinor: Eq. (1) uses D(x) for both the raw and normalized distance, and there are typos; nothing that affects the method.\n\nVerdict: the central empirical claim is likely defensible, but the paper needs revision before I would trust the mechanism. It deserves a serious referee; I would not desk reject it.","headline":"Incremental but plausible distance-map regression scheme for small tumor segmentation; the mechanism is under-tested, but the empirical gains justify a careful review.","tokens_in":7655,"tokens_out":3397,"would_cite":false,"duration_ms":34626,"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":"Distance-map regression branch lifts small-tumor segmentation Dice to 0.751.","keywords":["distance map regression","medical image segmentation","liver tumor segmentation","class imbalance","shape-aware loss","Dice loss","lightweight regression network","CT imaging"],"falsifier":"Train the same pipeline but replace the learned LR-Net with a direct, differentiable computation of the norm-inverse distance map from the soft probability map, or feed the LR-Net with soft probability maps sampled early in training; if Dice on the clinical dataset stays at 0.751 or the regression error on soft maps is uncorrelated with segmentation quality, the learned frozen mapping is not the actual source of the improvement.","tokens_in":6700,"feed_emoji":"🩻","tokens_out":9322,"duration_ms":76649,"temperature":0.7,"pith_summary":"The paper sets out to show that distance maps—not just binary masks—should supervise segmentation of small anatomical objects such as liver tumors. It proposes a two-part framework: a lightweight regression network (LR-Net) that is pretrained on ground-truth masks to map a mask to a normalized inverse distance map and then frozen, plus a MapDice loss that uses the ground-truth distance map as a spatial penalty map. The authors argue that this injects geometric shape information that ordinary class-imbalance losses lack. On the LiTS validation set and a clinical tumor dataset, the combination raises per-case Dice from 0.647 with plain Dice to 0.751 with the full method, outperforming WCE, GDS, Tversky, Focal Tversky, and ExpLog baselines.","feed_headline":"Frozen distance-map regressor lifts tumor Dice to 0.751","feed_subtitle":"Pairing a frozen mask-to-distance network with a distance-weighted Dice loss beats WCE, Tversky, and focal loss baselines.","key_machinery":"The norm-inverse distance map (NI-DM) is the central representation: for each voxel it records the Euclidean distance to the nearest object boundary, then inverts and normalizes each connected component so boundary-adjacent voxels receive the highest values. The LR-Net is a lightweight U-shaped regression network, pretrained separately on ground-truth binary masks with a smooth L1 loss and then frozen, so that it can convert the main network's soft probability map into a predicted distance map in a differentiable way. The MapDice loss, $L_{\\mathrm{MapDice}} = 1 - \\sum_c \\frac{2(p_c \\varphi_c) + \\varepsilon}{p_c + \\varphi_c + \\varepsilon}$, multiplies the predicted probability for each class by the ground-truth distance map $\\varphi_c$, turning boundary proximity into a pixel-wise penalty. Together these pieces make shape and boundary information explicit in the training signal without changing the architecture of the existing segmentation network.","core_discovery":"The central claim is that the rigorous mapping between a binary segmentation mask and its distance map can be learned once, by a small regression network, and then reused as a frozen geometric critic for any segmentation network. During training, the main network's soft probability map is fed into the frozen LR-Net, whose smooth-L1 regression loss pushes the whole system to produce masks whose implied distance maps match the true ones; at the same time, the MapDice loss reweights Dice by the boundary-emphasizing norm-inverse distance map. The paper reports that unfreezing the LR-Net degrades performance, which it reads as evidence that the pretrained, frozen mapping supplies a stable shape prior rather than a trainable module. The best configuration uses the norm-inverse distance map with ReLU activation, and the framework remains competitive on the public LiTS leaderboard while improving markedly over imbalance-focused losses on the clinical dataset.","pith_inferences":["Because the LR-Net is trained on binary ground-truth masks but evaluated on soft probability maps, its usefulness depends on that distribution shift being benign; a targeted test would be to train it on soft labels or to calibrate the main network's probabilities before regression.","The boundary-emphasizing weight map suggests MapDice could be combined with or compared against explicit boundary losses such as Hausdorff-style penalties; the paper does not test that combination.","The plug-in design implies the same frozen LR-Net could transfer to other small-object segmentation tasks such as pancreas or lymph-node lesions, but the paper only reports liver-tumor evidence.","A natural ablation would measure the LR-Net's regression error on the main network's soft maps at different training epochs to see whether the geometric signal is actually flowing through the frozen branch."],"forward_implications":["The frozen LR-Net is a plug-in: any classification-based segmentation backbone can receive distance-map supervision without architectural changes.","At inference time the LR-Net is removed, so the improved training signal costs nothing at deployment.","Distance-map variants matter: norm-inverse distance maps with ReLU activation outperform raw, inverse-only, sign, and sigmoid-mapped distance maps.","The MapDice loss and the regression loss are complementary; removing either drops Dice, and unfreezing the LR-Net drops Dice from 0.748 to 0.628 on the clinical dataset.","On the LiTS leaderboard submission the method reaches Dice per case 0.679 and Dice global 0.830, showing the gain transfers to a public benchmark."],"supporting_citations":[{"why":"Supplies the LiTS challenge dataset used for validation and the distance-map-boost line of work this method extends.","marker":"[12]"},{"why":"Defines the LiTS benchmark and the seven evaluation metrics (Dice, VOE, RVD, ASSD, MSD, RMSD) used in all comparisons.","marker":"[3]"},{"why":"Weighted cross-entropy baseline for imbalanced segmentation that the proposed method outperforms.","marker":"[13]"},{"why":"Generalized Dice Score baseline, a class re-weighting loss compared in Table 2.","marker":"[4]"},{"why":"Tversky loss baseline for precision-recall tradeoff compared in Table 2.","marker":"[14]"},{"why":"Focal Tversky loss baseline that combines imbalance re-weighting with focal weighting.","marker":"[1]"},{"why":"Exponential Logarithmic loss baseline for highly unbalanced object sizes.","marker":"[16]"},{"why":"Multi-branch distance-map regression baseline that the tandem design outperforms.","marker":"[5]"},{"why":"Cascaded distance-transform regression baseline that the tandem design outperforms.","marker":"[2]"}],"fun_headline_variants":["Frozen distance-map critic lifts tumor Dice to 0.751","Shape-aware loss via distance maps beats focal loss","Lightweight LR-Net learns shape prior for small tumors","Distance-map regression improves imbalanced segmentation","Tumor Dice gains from frozen geometric prior"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the LR-Net, pretrained on perfect binary masks, will still produce useful distance-map regression signals when it is fed the M-Net's soft probability maps during joint training; the paper's own ablation shows unfreezing the LR-Net hurts, so the method relies on this mismatch being benign.","fun_headline_variants_meta":{"raw":{"variants":["Frozen distance-map critic lifts tumor Dice to 0.751","Shape-aware loss via distance maps beats focal loss","Lightweight LR-Net learns shape prior for small tumors","Distance-map regression improves imbalanced segmentation","Tumor Dice gains from frozen geometric prior"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1433,"prompt_tokens":910,"completion_tokens":523,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":449}},"tokens_in":526,"tokens_out":523,"duration_ms":6750,"temperature":1.0,"reasoning_tokens":449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:09:51.448523+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same pipeline but replace the learned LR-Net with a direct, differentiable computation of the norm-inverse distance map from the soft probability map, or feed the LR-Net with soft probability maps sampled early in training; if Dice on the clinical dataset stays at 0.751 or the regression error on soft maps is uncorrelated with segmentation quality, the learned frozen mapping is not the actual source of the improvement.","supporting_citations":[{"cited_title":"In: Medical Imaging with Deep Learning (2020)","cited_arxiv_id":null,"evidence_quote":"Supplies the LiTS challenge dataset used for validation and the distance-map-boost line of work this method extends."},{"cited_title":"IEEE transactions on medical imaging 25(11), 1451–1461 (2006)","cited_arxiv_id":null,"evidence_quote":"Generalized Dice Score baseline, a class re-weighting loss compared in Table 2."},{"cited_title":"In: International Workshop on Machine Learning in Medical Imaging","cited_arxiv_id":null,"evidence_quote":"Tversky loss baseline for precision-recall tradeoff compared in Table 2."},{"cited_title":"In: 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019)","cited_arxiv_id":null,"evidence_quote":"Focal Tversky loss baseline that combines imbalance re-weighting with focal weighting."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Exponential Logarithmic loss baseline for highly unbalanced object sizes."},{"cited_title":"Medical physics 46(12), 5637–5651 (2019)","cited_arxiv_id":null,"evidence_quote":"Multi-branch distance-map regression baseline that the tandem design outperforms."},{"cited_title":"Computer Vision and Image Understanding 189, 102809 (2019)","cited_arxiv_id":null,"evidence_quote":"Cascaded distance-transform regression baseline that the tandem design outperforms."}],"review_version":1}