{"id":"624130c8-0929-41a2-81d2-45cf8ec86525","arxiv_id":"2412.05968","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The paper reports a 0.71M-parameter encoder-decoder network that combines multi-scale convolutions, focal modulation attention, and spatial feature refinement, with reported dice scores of 86.44%, 87.88%, and 84.22% on DRIVE, STARE, and CHASE_DB.","lead":"LVS-Net is a compact neural network with 0.71 million parameters that segments retinal blood vessels and also separates arteries from veins. The authors report accuracy comparable to or better than much larger models on three public retinal datasets, while using less compute and memory.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported dice scores may be inflated by test-set threshold tuning and non-standard splits; the 'outperforms' claim is not yet supported.","rationale":"The reader's weakest assumption correctly identifies the evaluation protocol as the most load-bearing condition for the paper's central claim. My analysis confirms that the paper does not establish that the dice-maximizing threshold is chosen without access to test labels, and the 80/20 train/validation description for CHASE_DB and STARE leaves no separate test set, making it likely that the reported dice are computed on validation images after threshold tuning on those same images. The internal metric inconsistencies across abstract, body, and table add weight to this concern because they indicate the numerical results were not carefully cross-checked. I therefore agree with the reader's assessment: the architecture is plausible and the lightweight parameter count is credible, but the performance superiority claim requires a verification of the evaluation protocol before it can be accepted. Since the reader already assigned a CONDITIONAL verdict, I do not recommend changing the verdict; the concern is real but does not by itself warrant rejection, as the authors could resolve it by clarifying or correcting the protocol and rerunning the comparisons.","tokens_in":17959,"tokens_out":2222,"duration_ms":22431,"concrete_test":"Obtain or re-train LVS-Net from the published architecture, then evaluate it under a protocol where the decision threshold is fixed at 0.5 (or tuned only on a held-out validation split that is never used to compute reported metrics). Use the standard DRIVE test split and clearly defined held-out splits for CHASE_DB and STARE, matching the split used by each baseline in Table II. If the resulting dice scores drop by more than 1-2 points on any dataset, or if LVS-Net no longer exceeds the compared baselines, the central performance claim is not substantiated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that LVS-Net outperforms existing models rests on the reported dice values, but the evaluation protocol described in the paper allows those values to be optimistically biased. Section III-A states that 'F1-thresholding is used that maximizes the dice score' after the final sigmoid, but it does not state that the threshold is selected on validation data disjoint from the reported test set. Section IV-B says only that 80% of images were used for training and 20% for validation, with no separate test set mentioned for CHASE_DB and STARE; Table I similarly lists CHASE_DB and STARE with no testing split. If the threshold is tuned on the same images used to compute the reported dice, the numbers are partially fitted to the test ground truth. On small datasets (CHASE_DB has 28 images, STARE has 20), threshold optimization on a 20% validation set can inflate dice by several points relative to a fixed 0.5 threshold or a threshold chosen on a truly held-out set. This concern is reinforced by internal inconsistencies: the abstract reports CHASE_DB dice as 84.22%, Table II reports 84.78%, and the conclusion reports 82.10%; the body text reports STARE dice as 84.78% while Table II reports 87.88%. Such discrepancies suggest the reported numbers were not carefully validated. Without code, weights, or a precise protocol statement, the comparison to baselines in Table II is not reproducible, and the 'outperforms existing models' conclusion is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes LVS-Net, a lightweight encoder-decoder network for retinal vessel segmentation. The encoder uses multi-scale convolutional blocks; the bottleneck combines a Focal Modulation Attention Module (FMAM) and a Spatial Feature Refinement Block (SFRB); the decoder upsamples with transposed convolutions and skip connections that also pass through SFRB. The model is evaluated for vessel segmentation on DRIVE, CHASE_DB, and STARE and for artery/vein segmentation on RITE. The authors report 0.71 million parameters, 2.74 MB memory, and 29.60 GFLOPs, and claim state-of-the-art dice scores of 86.44%, 84.22%, and 87.88% on DRIVE, CHASE_DB, and STARE, respectively.","tokens_in":18287,"tokens_out":5177,"duration_ms":45909,"significance":"If the evaluation protocol is sound and the reported numbers are reproducible, LVS-Net would be a practically useful lightweight baseline for retinal feature segmentation; the 0.71M parameter count is genuinely low, and the ablation study in Table IV gives a useful decomposition of the architectural components. The paper does not ship code, weights, or a precise test-set definition, and the reported metrics contain internal inconsistencies, so the central 'outperforms existing models' claim is not currently supported. The architectural idea is defensible, but the quantitative evidence needs to be corrected and the protocol made explicit before the claims can be assessed.","major_comments":[{"comment":"The sentence 'F1-thresholding is used that maximizes the dice score' does not state whether the threshold is chosen on a validation set disjoint from the test images or on the test ground truth itself. If the threshold is optimized on the test labels, the reported dice values are optimistically biased, especially on CHASE_DB (28 images) and STARE (20 images), where a few threshold choices can change the score by several percentage points. The authors must specify the threshold-selection protocol and confirm that test labels were not used for any model or threshold selection.","section":"Section III-A"},{"comment":"For CHASE_DB and STARE, the text states only that 80% of images were used for training and 20% for validation, and Table I lists no testing images for these datasets. Without a defined held-out test split, the reported test-set numbers and the comparison to published baselines are ill-defined. The authors must report the exact split, the number of test images, and how the published baselines were evaluated under the same protocol.","section":"Section IV-B and Table I"},{"comment":"The reported CHASE_DB dice score appears as 84.22% in the abstract, 84.78% in the body text and Table II, and 82.10% in the conclusion; the STARE dice score appears as 84.78% in Section IV-D and 87.88% in the abstract and Table II. These internal inconsistencies mean the headline quantitative claims are not reliable as stated and must be reconciled with a single, corrected set of results.","section":"Abstract, Section IV-D, Table II, Conclusion"},{"comment":"The 'AUC' formula in Eq. (26) is not the standard ROC-AUC; it is dimensionally inconsistent and cannot be interpreted as a probability because it multiplies TP and TN and divides by nested sums. Since Section IV-D uses AUC values (0.993, 0.997, 0.998) as evidence of superiority over other models, the metric must be defined correctly and all AUC values must be recomputed with the standard definition.","section":"Section IV-C, Eq. (26)"},{"comment":"The RITE results are internally inconsistent: the average accuracy of 98.44% cannot be reconciled with artery accuracy 97.13% and vein accuracy 99.75%, and the average dice of 81.34% cannot be reconciled with artery dice 75.46% and vein dice 71.18% unless the background class is included in the average, which is not stated. The table needs a clear definition of how the 'Average' column is computed and corrected row values.","section":"Table III"}],"minor_comments":[{"comment":"The first contribution bullet says 'Introducing LA V-Net' but the model is named LVS-Net elsewhere; this appears to be a typo.","section":"Section I"},{"comment":"The dice formula is written with 'TP+TP' in the numerator and denominator; it is mathematically equivalent to 2TP/(2TP+FP+FN) but should be simplified for readability and to avoid confusion.","section":"Section IV-C, Eq. (22)"},{"comment":"The CHASE_DB dataset is described as 'CHASE DB-DB' in the first sentence of Section IV-A and as 'CHASE DB' elsewhere; the name should be standardized.","section":"Section IV-A"},{"comment":"Equation (4) uses 'Re' where 'ReLU' is intended, and the notation for convolution (C), transposed convolution (T), and the SFRB/FMAM operations G and F is not consistently defined at first use.","section":"Eqs. (4)-(6) and Fig. 1"},{"comment":"The dash in the 'Testing' column for CHASE_DB and STARE is unexplained; if no separate testing split is used, this should be stated explicitly in the table caption or text.","section":"Table I"},{"comment":"The ablation row 'MLU + CBAM in Skip Connections' reports a lower dice (80.86%) than the Lightweight U-Net baseline (82.06%), while the text says CBAM 'substantially improves' performance; a brief explanation would help the reader interpret the ablation.","section":"Section IV-F, Table IV"},{"comment":"No code, trained weights, or public implementation are provided, which, combined with the protocol ambiguities, prevents independent verification of the reported results.","section":"Availability"}],"recommendation":"major_revision","confidential_remarks":"The reference list contains a large number of self-citations to the authors' own lightweight segmentation networks (e.g., refs [22], [25], [43], [44], [45], [81]). This is not a reason for rejection, but the authors should ensure that statements such as 'existing methods cannot detect multi-retinal features with state-of-the-art results' are calibrated against the cited prior work. The internal metric inconsistencies and the ambiguous threshold-selection protocol suggest the paper was not carefully proofread; the editor may wish to require a statement of code/data availability and a precise evaluation protocol before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the architecture is a credible combination of existing ideas, and the ablation study is the best part. But the paper does not support its headline claim that LVS-Net outperforms existing models, because the reported metrics are internally inconsistent and the evaluation protocol allows optimistic bias.\n\nWhat is actually new here is modest but real: a lightweight U-Net variant that uses multi-scale convolutional blocks, a focal modulation attention module, and spatial feature refinement blocks in the bottleneck and skip connections. The ablation in Table IV, which builds up from a plain lightweight U-Net to the full model, is a genuine strength. At 0.71M parameters and 2.74 MB, the model is genuinely small.\n\nThe problems are in the evaluation. First, the dice values are not stable within the paper: CHASE_DB dice appears as 84.22% in the abstract, 84.78% in the body text and Table II, and 82.10% in the conclusion; STARE dice appears as 84.78% in the text and 87.88% in Table II. That alone is cause for caution. Second, the paper says 'F1-thresholding is used that maximizes the dice score' but never states that the threshold is selected on a held-out set. For CHASE_DB (28 images) and STARE (20 images), the 80/20 split leaves only a handful of validation images, and tuning the threshold on those can inflate dice by several points compared to a fixed threshold or a properly separated test set. No separate test set is mentioned for those datasets. Third, Eq. 26's AUC formula is dimensionally nonsense; it cannot be the ROC-AUC used to produce the ROC curves. Fourth, there is no code, no weights, and no error bars, so the comparison to baselines in Table II is not independently checkable.\n\nIf the authors can clarify the threshold selection, report on standard splits or a held-out test set, and fix the internal numbers, the performance claim could be plausible. As written, the evidence is not convincing. This is a paper for someone cataloging lightweight retinal vessel segmentation architectures; it is not a benchmark to rely on.\n\nI'd send it to peer review because the architecture and ablation merit referee time and the issues are fixable. But I'd expect a major revision and likely changed numbers.","headline":"Plausible lightweight architecture with a useful ablation, but inconsistent numbers and a suspicious test protocol leave the performance claim unverified.","tokens_in":18798,"tokens_out":3292,"would_cite":false,"duration_ms":28703,"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 0.71-million-parameter network is claimed to out-segment larger retinal vessel models.","keywords":["retinal vessel segmentation","lightweight encoder-decoder","focal modulation attention","spatial feature refinement block","fundus image analysis","artery vein segmentation","DRIVE dataset","STARE dataset"],"falsifier":"Re-running the evaluation with a threshold fixed from training data alone and with the standard test splits for STARE and CHASE_DB would settle the claim; if the dice score then drops below the best compared baseline on any dataset, the 'outperforms' conclusion fails.","tokens_in":17779,"feed_emoji":"👁️","tokens_out":7217,"duration_ms":62191,"temperature":0.7,"pith_summary":"The paper introduces LVS-Net, a lightweight encoder-decoder network for segmenting retinal blood vessels in fundus images, aimed at early disease screening in settings with limited hardware. The authors claim that with only 0.71 million parameters, 2.74 MB of memory, and 29.60 GFLOPs, the model outperforms larger published networks on three standard public datasets, reaching dice scores of 86.44% on DRIVE, 84.22% on CHASE_DB, and 87.88% on STARE. The design combines multi-scale convolutional blocks in the encoder with focal modulation attention and spatial feature refinement at the bottleneck, plus refinement blocks along skip connections and decoder stages. If the reported evaluation is sound, this would establish that high-accuracy vessel segmentation is feasible at a small enough footprint for portable and real-time retinal screening, including artery/vein classification.","feed_headline":"0.71M-parameter net reports top retinal vessel dice scores","feed_subtitle":"LVS-Net keeps memory near 3 MB while claiming dice over 86% on DRIVE and near 88% on STARE.","key_machinery":"The central mechanism is the combination of the Focal Modulation Attention Module (FMAM) and the Spatial Feature Refinement Block (SFRB). FMAM aggregates context through depth-wise convolutions at multiple levels plus global average pooling, then modulates each query token by gated element-wise multiplication. SFRB is a residual block that concatenates max-pooled and average-pooled features, weights them with a sigmoid attention coefficient from global average pooling, and adds back the input. These blocks are placed at the bottleneck and, for SFRB, at every skip connection and decoder stage, so that multi-scale vessel details survive downsampling and are refined during upsampling.","core_discovery":"LVS-Net is a lightweight encoder-decoder network that segments retinal vessels and also separates arteries from veins. The encoder uses 1x1 and 3x3 convolutions at three scales, the bottleneck applies Focal Modulation Attention followed by a Spatial Feature Refinement Block, and every decoder upsampling stage and skip connection passes through SFRB before concatenation. The final sigmoid output is thresholded by choosing the F1 threshold that maximizes dice score, and training uses dice loss with Adam. On DRIVE, STARE, and CHASE_DB the authors report accuracy, dice, jaccard, sensitivity, and specificity numbers that exceed those of compared baselines, including U-Net, G-Net Light, Attention U-Net, MultiResNet, BCD-UNet, SegNet, U-Net++, FR-UNet, and RetinaLiteNet, while the model has 0.71M parameters, 2.74 MB memory, and 29.60 GFLOPs. On RITE the model reports average dice 81.34% for arteries/veins/background with higher individual artery and vein dice than listed baselines.","pith_inferences":["A direct next step the paper does not run is cross-dataset evaluation (train on DRIVE, test on STARE and CHASE_DB), which would test whether the reported dice gains survive domain shift between fundus cameras.","The bottleneck attention and refinement blocks could transfer to other elongated-structure segmentation tasks, such as coronary angiography or neural fiber tracing, where thin-structure preservation is the limiting factor.","Because the reported memory footprint is 2.74 MB before quantization, 8-bit quantization could plausibly bring the deployed model under 1 MB; the paper does not report quantized performance."],"forward_implications":["If the reported numbers hold, accurate vessel segmentation no longer requires a heavy model: a network under 3 MB can reach dice scores above 84% on all three standard datasets.","The same lightweight architecture also separates arteries from veins on RITE, so a single small model can support multi-feature retinal screening.","The ablation study's stepwise gains indicate that multiscale convolutions, SFRB at skip connections and bottleneck, and FMAM at the bottleneck each contribute to capturing thin vessels; the final configuration is the sum of these additions.","At 29.60 GFLOPs and 0.71M parameters, the model sits between the smallest lightweight baselines and larger U-Net variants, giving clinicians a concrete footprint target for portable screening devices."],"supporting_citations":[{"why":"Supplies the Focal Modulation Attention module placed at the model bottleneck.","marker":"[46]"},{"why":"Provides the spatial attention design used in the Spatial Feature Refinement Block.","marker":"[47]"},{"why":"Source of the DRIVE dataset images and their vessel and artery/vein ground-truth annotations.","marker":"[48]"},{"why":"Source of the CHASE_DB images used for one of the three vessel-segmentation evaluations.","marker":"[49]"},{"why":"Source of the STARE images and expert vessel annotations used in the third evaluation.","marker":"[50]"},{"why":"The U-Net architecture that serves as the principal baseline and comparison point.","marker":"[75]"},{"why":"A lightweight baseline (G-Net Light, 0.39M parameters) whose accuracy-efficiency trade-off LVS-Net claims to exceed.","marker":"[43]"},{"why":"Supplies the RITE dataset used for the artery/vein multi-class segmentation experiment.","marker":"[74]"}],"fun_headline_variants":["0.71M-param LVS-Net posts top retinal vessel dice","Lightweight LVS-Net beats heavy models on retinal vessels","Tiny net, huge dice: LVS-Net for retinal vessels","Retinal vessel segmentation with just 0.71M parameters","LVS-Net: 0.71M params, high dice on three retinal datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the evaluation protocol being valid: the dice-maximizing threshold applied after the sigmoid must be selected on validation data rather than test ground truth, and the 80/20 image-level splits used for CHASE_DB and STARE must yield test sets comparable to those used by the published baselines.","fun_headline_variants_meta":{"raw":{"variants":["0.71M-param LVS-Net posts top retinal vessel dice","Lightweight LVS-Net beats heavy models on retinal vessels","Tiny net, huge dice: LVS-Net for retinal vessels","Retinal vessel segmentation with just 0.71M parameters","LVS-Net: 0.71M params, high dice on three retinal datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000321,"raw_usage":{"total_tokens":1854,"prompt_tokens":1036,"completion_tokens":818,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":721}},"tokens_in":652,"tokens_out":818,"duration_ms":7827,"temperature":1.0,"reasoning_tokens":721,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:09:02.071478+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running the evaluation with a threshold fixed from training data alone and with the standard test splits for STARE and CHASE_DB would settle the claim; if the dice score then drops below the best compared baseline on any dataset, the 'outperforms' conclusion fails.","supporting_citations":[{"cited_title":"Focal modulation networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the Focal Modulation Attention module placed at the model bottleneck."},{"cited_title":"Spatial attention for multi-scale feature refinement for object detection,","cited_arxiv_id":null,"evidence_quote":"Provides the spatial attention design used in the Spatial Feature Refinement Block."},{"cited_title":"A manually- labeled, artery/vein classified benchmark for the drive dataset,","cited_arxiv_id":null,"evidence_quote":"Source of the DRIVE dataset images and their vessel and artery/vein ground-truth annotations."},{"cited_title":"Robust retinal vessel segmentation via locally adaptive derivative frames in orientation scores,","cited_arxiv_id":null,"evidence_quote":"Source of the CHASE_DB images used for one of the three vessel-segmentation evaluations."},{"cited_title":"Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response,","cited_arxiv_id":null,"evidence_quote":"Source of the STARE images and expert vessel annotations used in the third evaluation."},{"cited_title":"U-Net: Convolutional net- works for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"The U-Net architecture that serves as the principal baseline and comparison point."},{"cited_title":"G-net light: A lightweight modified google net for retinal vessel segmentation,","cited_arxiv_id":null,"evidence_quote":"A lightweight baseline (G-Net Light, 0.39M parameters) whose accuracy-efficiency trade-off LVS-Net claims to exceed."},{"cited_title":"Automated separation of binary overlapping trees in low-contrast color retinal images,","cited_arxiv_id":null,"evidence_quote":"Supplies the RITE dataset used for the artery/vein multi-class segmentation experiment."}],"review_version":1}