{"id":"600e4956-2aea-4e1f-b7d8-d9771e6cbf81","arxiv_id":"1908.06498","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A geodesic-map autoencoder prior added to the segmentation loss improves cardiac MRI segmentation Dice by 4.4% on clean labels and by 4.6% to 6.3% on two levels of synthetic label noise.","lead":"This paper trains an autoencoder to turn distance maps of heart structures into binary shapes, then uses its internal features to supervise a 3D segmentation network during training. The method reports Dice gains of roughly 4 to 6 points over a plain segmentor on cardiac MRI, including when training labels are artificially noised.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The auxiliary loss in Eq. (3) assumes the GAE encoder embeds soft probability maps like geodesic maps, but this domain shift is unanalyzed; the reported gains may not isolate the geodesic prior.","rationale":"The reader's weakest_assumption correctly identifies the central technical soft spot: the geodesic prior is injected through a fixed encoder that was trained on geodesic maps, yet is applied to soft probability maps. This assumption is load-bearing because the entire auxiliary loss in Eq. (3) depends on the encoder producing comparable features for both input types. If the encoder's feature space is not transferable across this domain shift, the auxiliary gradient could be arbitrary, and the observed Dice improvements might stem from implicit regularization or other artifacts rather than from the geodesic prior. The proposed diagnostic—decoding Encgae(Pi) and checking whether it reconstructs the label shape—directly tests whether the encoder's embedding of probability maps is semantically meaningful. This is a concrete, computational check that does not require retraining. I find no reason to change the reader's conditional verdict: the central hypothesis is plausible and the experimental comparison is coherent, but this unexamined assumption plus the absence of error bars and the synthetic-noise-only evaluation leave the claim conditional on further verification.","tokens_in":6850,"tokens_out":6819,"duration_ms":82141,"concrete_test":"Use a trained GAE and a trained segmentor. For a held-out set, feed each probability map Pi through the GAE encoder and then the GAE decoder to produce a reconstructed binary map R_i = Decgae(Encgae(Pi)). Compare R_i to the clean ground-truth label with Dice. If R_i is not substantially better than chance, or if the distribution of Encgae(Pi) is far from Encgae(Gi) by a simple metric such as maximum mean discrepancy, then Encgae(Pi) is not a valid shape encoding and Eq. (3)'s gradient signal is not a shape prior. If R_i accurately recovers the label shape, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is in Eq. (3): Lgae(Encgae(Pi), Encgae(Gi)) treats the fixed GAE encoder as a meaningful shape embedding for the segmentor's soft probability map Pi, even though the GAE was trained only on geodesic distance maps Gi (Section 2.2). A geodesic map is a distance transform with large dynamic range and sharp, surface-like structure; Pi is a [0,1]-valued soft map. Nothing in the training of Net_gae ensures that Encgae(Pi) lies in a region of feature space where its distance to Encgae(Gi) corresponds to shape dissimilarity. If the encoder's response to probability maps is arbitrary, the auxiliary loss can be minimized by distorting Pi in ways unrelated to segmentation quality, and the reported Dice gains would not be evidence for the geodesic prior mechanism. The paper provides no analysis of this domain shift and no ablation that isolates the contribution of the geodesic-map encoding from a generic feature-matching regularizer. The synthetic-noise limitation is acknowledged in the conclusion, but this domain-mismatch assumption is not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a weakly supervised segmentation framework for 3D cardiac MRI. A geodesic autoencoder (GAE) is first trained to reconstruct binary labels from geodesic distance maps; its encoder is then frozen and used in an auxiliary feature-matching loss Lgae(Encgae(Pi), Encgae(Gi)) that encourages the segmentor's softmax output to share bottleneck features with the label-derived geodesic maps. Experiments on the ACDC dataset with clean labels and two synthetic noise levels (L1, L2) report Dice improvements of 4.4%, 4.6%, and 6.3% over the segmentor trained without the prior, and the method also outperforms a binary-map AE variant. The authors conclude that the geodesic prior improves robustness to weak labels.","tokens_in":7142,"tokens_out":6336,"duration_ms":62114,"significance":"If the reported effects are reproducible, the work is a useful contribution to the line of work on shape-prior regularization for medical image segmentation, offering a concrete way to encode multi-object shape information in a 3D network. The design is coherent: a fixed GAE with a feature-matching loss is a simple and computationally cheap auxiliary objective, and the comparison against a binary AE baseline is the right control for the choice of geodesic over binary maps. The main limitations are experimental: no variance/significance information, an unexplained exceedance of the 'Upper boundary' under L2 noise, and no analysis of the domain shift between binary-derived geodesic maps and soft probability maps. The core mechanism is therefore not yet established.","major_comments":[{"comment":"The reported Dice values for Geodesic Prior (LV 0.873, RV 0.831, MYO 0.809; average 0.838) are all higher than the corresponding 'Upper boundary' values (0.869, 0.824, 0.801; average 0.831). Because the upper boundary is defined as the Dice of the noisy training labels against the clean ground truth, a model trained on those labels would not be expected to exceed it. The authors should explain this anomaly or justify why the comparison is meaningful; as written, it undermines the main claim of improved accuracy under noise.","section":"Table 1, Inexpert Labels (L2) rows"},{"comment":"The auxiliary loss feeds the segmentor's continuous probability map Pi through Encgae, yet Encgae was trained exclusively on geodesic maps Gi generated from binary labels. Geodesic maps are distance transforms with a large dynamic range and sharp boundaries, while Pi is a softmax output in [0,1]. The paper provides no analysis showing that Encgae(Pi) lies in the feature regime where distances to Encgae(Gi) reflect shape dissimilarity, and no ablation isolates the geodesic encoder from a generic feature-matching regularizer (e.g., replacing Encgae with a randomly initialized frozen encoder, or applying the same loss directly to Pi). Without such evidence, the reported gains may not be attributable to the geodesic prior mechanism.","section":"Section 2.2, Eq. (3)"},{"comment":"The 'prior' is learned from the same labels (clean or noisy) that supervise the segmentor. For the L1 and L2 conditions, the GAE's geodesic inputs and reconstruction targets are generated from the corrupted labels, so the auxiliary loss inherits the same label noise. Calling this a shape prior is therefore not justified. The authors should either reframe the contribution as a learned feature-matching regularizer or provide evidence that the GAE recovers clean shape statistics despite noisy inputs.","section":"Section 2.2 and Fig. 1"},{"comment":"No variance, confidence intervals, or statistical tests are reported. The average Dice differences between Geodesic Prior and Seg.Net are 0.044, 0.046, and 0.063, but with a single trained model per condition and no error bars, these differences may not be robust. The authors should report means and standard deviations over multiple runs, or provide per-volume significance testing on the 50 test volumes.","section":"Experiments, Table 1"}],"minor_comments":[{"comment":"The text 'LV, RV, Mayo' should read 'LV, RV, Myo'.","section":"Section 2.2"},{"comment":"The notation Lrecons appears in the GAE training description while Lrecon is used earlier; please unify.","section":"Section 2.2"},{"comment":"The conclusion states the approach 'outperforms the state-of-the-art methods,' but no comparison to state-of-the-art segmentation methods is included in Table 1; this overstates the results.","section":"Conclusion"},{"comment":"The sentence 'we used 80 MR images for training, the 20 images were used as validation' needs a conjunction or period and should clarify whether these counts refer to volumes or 2D slices.","section":"Implementation Details"},{"comment":"Equation (4) writes 'F(x)|ΔT(x)| = 1' but the surrounding text discusses the gradient of T; please replace ΔT with ∇T for clarity.","section":"Equation (4)"},{"comment":"The bottleneck dimension Lfeat of the GAE is never specified numerically; please report it for reproducibility.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript would benefit from code/data release; the only reported results appear to come from a single training run per condition, and the Table 1 upper-bound exceedance under L2 noise is a substantive anomaly. I would ask the editor to require the authors to address the Table 1 issue and provide either code or additional experiments on the domain-shift concern before considering the paper for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague —\n\nYou should know up front: this is a solid incremental paper, not a breakthrough. The new bit is using fast-marching geodesic distance maps as the input to an autoencoder shape prior, then using the encoder's bottleneck features as an auxiliary loss for a 3D segmentation network. Oktay et al. used binary maps, Zotti et al. used atlases; the geodesic encoding is new, and the comparison against a binary-prior baseline gives it real content. The method is coherent, and the experiments cover two noise levels plus clean labels, with both Dice and Hausdorff reported. The gains are a few points, which is plausible for a regularizer of this kind.\n\nThe soft spots are about evidence quality, not design. First, no error bars, no significance tests, and no per-case breakdown, so a 0.004 average Dice difference (expert labels: 0.853 vs 0.849) could be noise. Second, in the L2 row the geodesic prior (0.838) exceeds the paper's own 'Upper boundary' (0.831). That column is just the Dice of the noisy labels against clean ground truth; exceeding it is what a regularizer should do, but the paper doesn't explain the seeming contradiction. Third, the load-bearing assumption in Eq. (3) is that the frozen GAE encoder, trained on geodesic maps, produces meaningful embeddings for the segmentor's soft probability maps. The stress-test note is right: nothing guarantees that Encgae(Pi) lives in a region where distance to Encgae(Gi) tracks shape dissimilarity. That said, the binary-prior baseline has exactly the same domain shift and the geodesic version beats it, so the empirical case for the prior is not empty. The paper would be stronger with an ablation swapping Lgae for a generic feature-matching loss, and with a sentence acknowledging the shift.\n\nBottom line: the central claim — that a learned geodesic prior helps with noisy labels — holds up as a plausible, modest result. The missing variance and the unanalyzed domain shift keep it from being fully convincing, but it deserves a proper peer-review round rather than a desk reject. I'd send it to review, ask for error bars, a direct comparison with the atlas prior of Zotti et al., and an ablation isolating the geodesic encoding.","headline":"A modest but plausible result: geodesic distance maps as an autoencoder shape prior give a few Dice points on noisy cardiac labels; the main weaknesses are missing variance and an unanalyzed domain-shift assumption.","tokens_in":7629,"tokens_out":3242,"would_cite":false,"duration_ms":32561,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68U10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A geodesic autoencoder prior added to the loss improves 3D cardiac segmentation under clean and noisy labels by 4.4–6.3 Dice points.","keywords":["weakly supervised segmentation","geodesic prior","shape prior","autoencoder","cardiac MRI","noisy labels","deep learning","3D medical image segmentation"],"falsifier":"If the auxiliary loss were replaced by the same mean squared error against features of a permuted or random geodesic map and the Dice scores remained equally high, the reported gains would not be caused by the learned prior; alternatively, one can measure on a validation split how far the segmentor's probability-map features and the geodesic-map features actually are and check whether that distance tracks segmentation error.","tokens_in":6679,"feed_emoji":"🫀","tokens_out":5791,"duration_ms":55862,"temperature":0.7,"pith_summary":"The paper is trying to establish that a learned geodesic shape prior can make deep segmentation networks much less sensitive to weak or noisy labels. Its mechanism is an auxiliary loss: a geodesic autoencoder, trained once to convert geodesic distance maps of the annotated structures back into binary maps, supplies bottleneck features that the segmentor's own probability map is asked to match. On cardiac MRI with expert labels and two levels of simulated boundary noise, this extra term raises average Dice over the same network without the prior by 4.4, 4.6, and 6.3 points. That would matter because expert medical annotations are scarce and expensive, while imperfect labels from non-experts are abundant.","feed_headline":"Geodesic prior lifts weak-label segmentation by 6.3 Dice points","feed_subtitle":"A learned shape prior trained on geodesic maps makes segmentation networks tolerate noisy, inexpert annotations.","key_machinery":"The central object is the geodesic autoencoder (GAE): an encoder-decoder with a fully connected bottleneck, trained to reconstruct binary label maps from geodesic distance maps computed by fast marching, which solves the Eikonal equation for contour arrival time. The encoder's bottleneck feature vector captures shape and boundary information for each object, and the mean squared error between the features of the segmentor's soft probability map and the features of the geodesic map becomes the auxiliary loss term. This carries the argument because it converts weak boundary information into a shape constraint that is independent of the pixel-level segmentation loss, and because the GAE is frozen during segmentor training so the prior is a fixed geometric guide rather than a learned feature that can drift with the noisy labels.","core_discovery":"The paper's central claim is that a deep geodesic prior reduces the adverse effect of weak labels on segmentation accuracy. Concretely, the authors train a geodesic autoencoder (GAE) to map geodesic distance maps of the objects of interest (left ventricle, right ventricle, myocardium) back to their binary maps, and use the GAE's frozen encoder to compute an auxiliary loss between the segmentor's soft probability map and the corresponding geodesic map. The total loss becomes Lseg + Lgae, with Lgae a mean squared error on bottleneck features. On the ACDC cardiac MRI data, the geodesic prior outperforms both the bare segmentor and a binary-map autoencoder prior: average Dice rises from 0.809 to 0.853 with clean labels, from 0.795 to 0.841 with noise level L1, and from 0.775 to 0.838 with noise level L2.","pith_inferences":["The reported gains come from a single cardiac MRI dataset with simulated label noise; a natural next test would be whether the same prior helps with real inexpert annotations and other anatomies, where the noise structure may differ.","Because the geodesic map is derived from the same labels used for supervision, the prior may act mainly as a regularizer that suppresses boundary errors; comparing the method against other strong regularizers such as conditional random fields or adversarial losses would clarify the source of the gain.","The mechanism could be validated more directly by monitoring the feature distance Lgae during training: if the segmentor's probability-map features quickly collapse to the geodesic-map features even when the probability map is wrong, the loss may be pulling the wrong shape, which would be especially visible on donut-shaped objects like the myocardium.","The idea could be adapted to weaker supervision than noisy full masks, such as scribbles or bounding boxes, by defining the geodesic map from those available cues; that would test how much of the prior lives in the distance representation rather than in the full label."],"forward_implications":["Medical segmentation models can be trained on inexpert or automatically generated annotations and still approach the accuracy of models trained on expert labels, as long as a geodesic shape prior is available.","The relative benefit of the prior grows with label noise, from 4.4 Dice points on clean labels to 6.3 Dice points on the noisier level, suggesting shape priors matter most exactly where supervision is weakest.","The geodesic representation is more effective than a binary-map autoencoder prior, so the choice of how the shape is encoded into the prior is part of the performance.","At test time only the segmentor runs; the geodesic autoencoder is needed only during training, so the accuracy gain does not add inference cost.","The method is stated to be generic, meaning any 3D multi-object segmentation problem could in principle use the same geodesic-prior loss."],"supporting_citations":[{"why":"Supplies the ACDC cardiac MRI dataset, the expert annotations, and the test set used for all reported Dice and Hausdorff distance comparisons.","marker":"[3]"},{"why":"Provides the dense connectivity design that the segmentor's encoder-decoder architecture is built from.","marker":"[4]"},{"why":"The Tiramisu fully convolutional dense net that the paper extends from 2D to fully 3D for the segmentation network.","marker":"[5]"},{"why":"Prior work training an autoencoder on binary labels to obtain shape features for segmentation; the geodesic prior is the authors' replacement for this binary shape prior.","marker":"[9]"},{"why":"The fast marching method used to compute the geodesic distance maps from the binary labels.","marker":"[11]"}],"fun_headline_variants":["Geodesic prior adds 6.3 Dice for noisy labels","Deep geodesic prior boosts weak-label Dice by 6.3","Learned geodesic prior lifts noisy-label segmentation","Geodesic autoencoder prior hardens segmentation vs noise","Weak-label segmentation improved via geodesic shape prior"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's gain depends on the assumption that the bottleneck features extracted from the segmentor's soft probability map are comparable to those extracted from the binary-derived geodesic map, so that forcing them together conveys useful shape information; the paper does not examine how different those two feature distributions are.","fun_headline_variants_meta":{"raw":{"variants":["Geodesic prior adds 6.3 Dice for noisy labels","Deep geodesic prior boosts weak-label Dice by 6.3","Learned geodesic prior lifts noisy-label segmentation","Geodesic autoencoder prior hardens segmentation vs noise","Weak-label segmentation improved via geodesic shape prior"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000161,"raw_usage":{"total_tokens":1230,"prompt_tokens":937,"completion_tokens":293,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":212}},"tokens_in":553,"tokens_out":293,"duration_ms":3076,"temperature":1.0,"reasoning_tokens":212,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:43:13.955855+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If the auxiliary loss were replaced by the same mean squared error against features of a permuted or random geodesic map and the Dice scores remained equally high, the reported gains would not be caused by the learned prior; alternatively, one can measure on a validation split how far the segmentor's probability-map features and the geodesic-map features actually are and check whether that distance tracks segmentation error.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ACDC cardiac MRI dataset, the expert annotations, and the test set used for all reported Dice and Hausdorff distance comparisons."},{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"Provides the dense connectivity design that the segmentor's encoder-decoder architecture is built from."},{"cited_title":"In: Com- puter Vision and Pattern Recognition Workshops (CVPRW), 2017 IEEE Confer- ence on","cited_arxiv_id":null,"evidence_quote":"The Tiramisu fully convolutional dense net that the paper extends from 2D to fully 3D for the segmentation network."},{"cited_title":"IEEE transactions on medical imaging 37(2), 384–395 (2018)","cited_arxiv_id":null,"evidence_quote":"Prior work training an autoencoder on binary labels to obtain shape features for segmentation; the geodesic prior is the authors' replacement for this binary shape prior."},{"cited_title":"Journal of graphics tools 9(1), 23–34 (2004)","cited_arxiv_id":null,"evidence_quote":"The fast marching method used to compute the geodesic distance maps from the binary labels."}],"review_version":1}