{"id":"d85f1575-6461-4669-8529-2ace8d8f9d5d","arxiv_id":"1908.01279","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A LinkNet-34 model with transfer learning, augmentation, and a BCE-plus-Dice loss reaches DICE 78.8% for liver tumors on 3DIRCADb and 67.4% tumor DICE on KiTS-2019.","lead":"This paper reports that a standard segmentation network with a carefully tuned training routine reaches competitive liver and kidney tumor segmentation scores on two public CT datasets. It argues that training details can be as important as choosing a complex architecture.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Outperforming AHCNet is unsupported unless the 78.8% tumor DICE is shown to use the same tumor-size evaluation mode (any-size vs >100-px) as the comparison; the paper reports two modes but only one number.","rationale":"The paper's central claim is that a customized training pipeline can make a relatively simple CNN competitive with more complex architectures, and specifically that LinkNet-34 outperformed AHCNet. The numerical basis for that claim is the tumor DICE comparison in Table 1: 78.8% versus 73.4%. The load-bearing condition is that these two numbers are computed under comparable evaluation protocols. The paper's own §2.2 introduces two distinct evaluation modes for 3DIRCADb — tumors larger than 100-pixel area and tumors of any size — but Table 1 reports only one unlabeled 78.8%. If that number comes from the more favorable large-tumor mode, the outperformance claim is unsupported. This is a concrete, fixable omission rather than a fundamental flaw, so the reader's CONDITIONAL verdict remains appropriate. The concrete test of reporting both evaluation modes for the same model would settle the question directly. I agree with the reader's weakest assumption; the reader identified the same ambiguity and the same load-bearing comparison.","tokens_in":7623,"tokens_out":4686,"duration_ms":45452,"concrete_test":"Contact the authors or re-run the trained LinkNet-34 (or retrain from §2.2) on 3DIRCADb and compute tumor DICE separately for (a) tumors with area >100 pixels and (b) tumors of any size. Also compute liver DICE with per-case mean±std. If (a) yields 78.8 and (b) is materially lower, recompute AHCNet on the same (a) subset or revise the conclusion. If (b) also yields 78.8, the ambiguity is resolved and the outperformance claim is supported, subject to confirming AHCNet's protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In §2.2 the authors state that validation on 3DIRCADb was done in two ways: tested on tumors larger than 100-pixel area and on tumors of any size. Yet Table 1 lists a single tumor DICE of 78.8% with no indication of which mode produced it. The Conclusion then claims that LinkNet-34 outperformed AHCNet [22] on liver and liver tumor segmentation. The liver comparison (96.2 vs 95.9) is within a plausible noise band, and the tumor comparison (78.8 vs 73.4) is the actual evidence for 'outperformed.' If 78.8 was computed only on the subset of tumors with area larger than 100 pixels, the comparison to AHCNet's 73.4, whose evaluation protocol is not shown to be identical, is not a valid superiority claim. This is not a question of consensus; it is an internal ambiguity: the paper itself defines two evaluation protocols but publishes only one result. Additionally, no error bars or per-case statistics are provided, and the 'fraction of time' claim is not backed by any runtime measurement. The primary fix is to specify the evaluation mode and, ideally, report both modes for all methods compared.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript describes a segmentation system for liver, liver tumors, and kidney tumors from CT volumes. The method uses LinkNet-34 with an ImageNet-pretrained ResNet-34 encoder, a differential learning-rate schedule, a binary cross-entropy plus negative log-Dice loss, and data augmentation. Training is performed on the LiTS dataset and evaluation on 3DIRCADb, reporting a liver DICE of 96.2% and a tumor DICE of 78.8%; the same pipeline is applied to the KiTS-2019 challenge, reporting a kidney DICE of 96.38% and a tumor DICE of 67.38%. The paper's central claim is that a relatively simple CNN can match or beat more complex compound CNNs when attention is paid to the training pipeline rather than to the architecture.","tokens_in":7880,"tokens_out":4110,"duration_ms":42736,"significance":"If the results hold, the paper provides a useful data point: a straightforward 2D U-Net variant with transfer learning can reach competitive liver and liver-tumor segmentation on a public validation set and obtains a credible mid-ranking KiTS result. The evaluation design is a strength: training and validation are separated across public datasets, which avoids circularity, and the KiTS result was produced by the challenge's blinded test process. The contribution is, however, incremental. The reported advantage over AHCNet is not yet established at the level of statistical or protocol-wise rigor, and the 'fraction of time' claim lacks any supporting measurement.","major_comments":[{"comment":"The paper defines two evaluation protocols for 3DIRCADb ('tumors larger than 100-pixel area' and 'tumors of any size') but reports only a single tumor DICE of 78.8%. Because the Conclusion's claim of outperforming AHCNet (73.4%) rests on this number, the missing protocol specification is load-bearing. Please report DICE for both evaluation modes and verify that the comparison methods' numbers were computed under the same mode; otherwise the superiority claim is unsupported.","section":"§2.2 and §3, Table 1"},{"comment":"No error bars, confidence intervals, or per-case statistics are reported for any method in Table 1. The liver DICE gap of 96.2 versus 95.9 is small and likely within patient-level noise, so the statement that the method 'outperformed at least one compound CNN [22] for liver and liver tumor segmentations' is not established by the data as presented. Add variance estimates such as per-volume DICE standard deviations, and ideally a paired or bootstrap test for the comparison.","section":"§3, Table 1"},{"comment":"The claim that the pipeline achieves competitive results 'in fraction of time what would normally required for more complex CNNs' is not supported by any measured training or inference time. Please include hardware, epochs, and wall-clock measurements for the proposed method and, where possible, for the comparison methods, or remove the runtime claim from the conclusion.","section":"§4, Conclusion"}],"minor_comments":[{"comment":"The word 'prepossessing' should read 'preprocessing'.","section":"§1.2"},{"comment":"The dataset name is written both as '3DIRCAD' and '3DIRCAB'; use '3DIRCADb' consistently throughout.","section":"§2.1 and Table 1"},{"comment":"The row for H-DenseUNet is ambiguous because '[18]' and 'H-DenseUNet' appear on adjacent lines with a single set of metrics; align the method names with their metric columns.","section":"Table 1"},{"comment":"The 2.5D input variant is mentioned but no details are given for the number of adjacent slices tested or the final configuration; a sentence with the chosen stack size would help reproducibility.","section":"§2.2"},{"comment":"The paper uses 'DICE' and 'Dice' interchangeably; unify the notation.","section":"General"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a solid, modest technical report. A LinkNet-34 with ImageNet-initialized ResNet-34 encoder, reduced encoder learning rate, and BCE-minus-log-Dice loss gives liver DICE 96.2% and tumor DICE 78.8% on 3DIRCADb when trained on LiTS, and kidney/tumor DICE 96.38/67.38% on KiTS-19. Those numbers are plausible and the evaluation is on held-out public data. The paper's claim that a simple network plus careful training can match more complex architectures is worth taking seriously.\n\nWhat's genuinely new is the specific combination and results, not the components. Nothing here invents a loss, a network, or a training trick. But that's fine; the value is as a reproducible baseline.\n\nThe soft spots are real. The most important is in Section 3/Table 1: the paper says validation on 3DIRCADb was done in two modes—tumors larger than 100-pixel area and tumors of any size—but reports a single tumor DICE of 78.8%. The conclusion uses that number to claim it outperformed AHCNet's 73.4. If 78.8 came from the favorable subset, the comparison is not valid as stated. That's an internal ambiguity, not a literature dispute. Fix is simple: report both modes or clearly state which one was used, and ideally do the same for the comparison methods.\n\nBeyond that, there are no error bars or per-case statistics, so the 96.2 vs 95.9 liver comparison is within noise. The 'fraction of time' claim in the conclusion has no runtime data behind it. No code is provided, and hyperparameters are incomplete. These are standard weaknesses for a technical report, but they limit how much I'd lean on the conclusions.\n\nOn the whole the central argument is not broken. The paper does what it says: it shows a straightforward pipeline can reach competitive scores on public benchmarks. The claims beyond that are overreach, not fraud. I'd send this to a specialty short-paper or workshop track, not a flagship, and require the evaluation-mode ambiguity to be fixed. The kernel is a useful baseline for anyone working on liver or kidney tumor segmentation.","headline":"A modest, useful baseline for liver and kidney tumor segmentation whose headline superiority claim is muddied by an unreported evaluation-mode choice.","tokens_in":8413,"tokens_out":2402,"would_cite":false,"duration_ms":24012,"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 standard LinkNet-34 trained carefully on LiTS matched or beat more complex networks for liver and kidney tumor segmentation in CT.","keywords":["automatic segmentation","liver tumor","kidney tumor","CT images","convolutional neural networks","LinkNet-34","transfer learning","DICE score"],"falsifier":"Compute the paper's tumor DICE on 3DIRCADb under both stated evaluation modes, tumors larger than 100-pixel area and tumors of any size, and also compute AHCNet's DICE on the same set of test slices; if the 78.8% figure comes only from the larger-tumor subset while AHCNet's 73.4% was computed on any size, the claimed superiority would not hold.","tokens_in":7452,"feed_emoji":"🩻","tokens_out":7124,"duration_ms":65452,"temperature":0.7,"pith_summary":"This paper argues that for automatic liver and kidney tumor segmentation in CT images, the training pipeline can matter as much as the CNN architecture. The authors took an off-the-shelf LinkNet-34 with a pretrained ResNet-34 encoder and added a simple set of learning techniques: a reduced learning rate for the encoder, a loss combining binary cross entropy with a negative log-DICE term, and data augmentation. DICE is a standard overlap measure between predicted and true masks, with 100% being perfect. Trained on the LiTS dataset and validated on 3DIRCADb, the method reached 96.2% liver DICE and 78.8% liver tumor DICE, which the authors read as outperforming at least one compound attention network, AHCNet. The same recipe applied to the KiTS-2019 challenge scored 96.38% kidney and 67.38% tumor DICE, placing 38th out of 106 submissions. If the result holds, well-resourced bespoke architectures are not required for competitive tumor segmentation; a disciplined training procedure on a simple network can get there at a fraction of the cost.","feed_headline":"Off-the-shelf CNN beats AHCNet on liver tumor DICE","feed_subtitle":"Report: 78.8% liver-tumor DICE on 3DIRCADb, plus a top-38 kidney score, with a tuned LinkNet-34.","key_machinery":"The central object is LinkNet-34, a U-Net-style encoder-decoder segmentation network with ResNet-34 as its feature encoder, using the implementation of [26]. The method keeps the ImageNet-pretrained encoder weights unfrozen but lowers the encoder's learning rate by a factor of 10 relative to the randomly initialized decoder; the loss is binary cross entropy minus the logarithm of the DICE coefficient (Eq. 1). Training uses 58,638 image-mask pairs from the LiTS dataset (131 patients), and validation is on the 3DIRCADb dataset; a 2.5D variant that stacks adjacent slices was tried but did not improve DICE. This combination of a simple 2D CNN and a careful training recipe, rather than a bespoke deep architecture, is the mechanism the paper credits for its results.","core_discovery":"The central claim, stated on the paper's own terms, is that a customized training pipeline can make an off-the-shelf LinkNet-34 competitive with, and in one comparison better than, substantially more complex compound CNNs for liver and kidney tumor segmentation. On the 3DIRCADb validation set, the authors report liver DICE = 96.2% and tumor DICE = 78.8%, compared with AHCNet's 95.9% and 73.4% in their Table 1; both methods were trained on LiTS and evaluated on 3DIRCADb. When the same pipeline was applied to the KiTS-2019 challenge, it produced kidney DICE = 96.38%, tumor DICE = 67.38%, composite DICE = 81.88%, and 38th place among 106 submissions. The paper presents no new architecture; instead it attributes the gains to a combination of known techniques: transfer learning from a pretrained encoder, a 10x lower learning rate on that encoder, a BCE-minus-log-DICE loss, and data augmentation.","pith_inferences":["A fair reading of Table 1 should treat the 78.8% versus 73.4% comparison as conditional on the evaluation protocol: the paper does not state whether its tumor DICE used the 'any size' or 'larger than 100-pixel area' mode, and prior methods may have used different protocols; rescoring both sets of masks on the same subset would make the comparison decisive.","Because the same training recipe produced strong scores on two organs with different contrast characteristics, the recipe is a plausible candidate for other low-contrast lesion segmentation tasks; that is my extension, not a claim tested in the paper.","The negative result for 2.5D input suggests a testable hypothesis: in this regime, augmentation and loss shape matter more than inter-slice context; an ablation that drops the log-DICE term while keeping augmentation fixed would isolate that contribution."],"forward_implications":["On the paper's own results, a single off-the-shelf LinkNet-34 trained on LiTS reaches 96.2% liver and 78.8% tumor DICE on 3DIRCADb, exceeding AHCNet's 95.9% and 73.4% in the same table.","The differential learning-rate scheme, with the encoder's learning rate reduced by a factor of 10, is reported to converge faster and to a better solution than uniform training, so transfer learning need not freeze the encoder.","Because data augmentation improved accuracy while 2.5D multi-slice input did not, slice-stacking is not necessary to achieve these numbers.","The same pipeline transfers to a different organ: on KiTS-2019 it achieved 96.38% kidney and 67.38% tumor DICE, finishing 38th of 106 submissions."],"supporting_citations":[{"why":"Supplies the LinkNet-34 encoder-decoder architecture whose segmentation results the paper reports.","marker":"[24]"},{"why":"Supplies the ResNet-34 architecture used as the pretrained encoder.","marker":"[25]"},{"why":"Supplies the reference implementation of LinkNet-34 used in the experiments.","marker":"[26]"},{"why":"Supplies the LiTS training dataset with 58,638 image-mask pairs.","marker":"[20]"},{"why":"Supplies the 3DIRCADb validation dataset on which the liver and tumor DICE scores are computed.","marker":"[15]"},{"why":"Defines AHCNet, the compound CNN the paper compares against and claims to outperform.","marker":"[22]"},{"why":"Supplies the KiTS-2019 challenge dataset and ranking used for the kidney evaluation.","marker":"[23]"}],"fun_headline_variants":["Tuned training lifts off-the-shelf CNN to beat AHCNet on liver DICE","Simple training beats complex net for CT tumor segmentation","Off-the-shelf LinkNet-34 tops AHCNet on liver tumor DICE","78.8% liver tumor DICE via tuned off-the-shelf CNN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the DICE values in Table 1 were measured under compatible rules, but the paper never states whether its 78.8% tumor score counts all tumors or only those larger than 100 pixels.","fun_headline_variants_meta":{"raw":{"variants":["Tuned training lifts off-the-shelf CNN to beat AHCNet on liver DICE","Simple training beats complex net for CT tumor segmentation","Off-the-shelf LinkNet-34 tops AHCNet on liver tumor DICE","78.8% liver tumor DICE via tuned off-the-shelf CNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000877,"raw_usage":{"total_tokens":3798,"prompt_tokens":957,"completion_tokens":2841,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2759}},"tokens_in":573,"tokens_out":2841,"duration_ms":20093,"temperature":1.0,"reasoning_tokens":2759,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:16:59.562096+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the paper's tumor DICE on 3DIRCADb under both stated evaluation modes, tumors larger than 100-pixel area and tumors of any size, and also compute AHCNet's DICE on the same set of test slices; if the 78.8% figure comes only from the larger-tumor subset while AHCNet's 73.4% was computed on any size, the claimed superiority would not hold.","supporting_citations":[{"cited_title":"LinkNet: Exploiting encoder repre- sentations for efficient semantic segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the LinkNet-34 encoder-decoder architecture whose segmentation results the paper reports."},{"cited_title":"De ep Residual Learning for Image Recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the ResNet-34 architecture used as the pretrained encoder."},{"cited_title":"Kalin in and Vladimir I","cited_arxiv_id":null,"evidence_quote":"Supplies the reference implementation of LinkNet-34 used in the experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LiTS training dataset with 58,638 image-mask pairs."},{"cited_title":"and others","cited_arxiv_id":null,"evidence_quote":"Supplies the 3DIRCADb validation dataset on which the liver and tumor DICE scores are computed."},{"cited_title":"AHCNet: An Applica- tion of Attention Mechanism and Hybrid Connection for Liver Tumor Segmentation in CT Volumes","cited_arxiv_id":null,"evidence_quote":"Defines AHCNet, the compound CNN the paper compares against and claims to outperform."}],"review_version":1}