{"id":"6c289762-95f4-4056-8662-5cd4c112a52c","arxiv_id":"1908.03339","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Hyper Vision Net reports training Dice of 0.9552 and validation Dice of 0.8967 for kidney tumor segmentation on KiTS19, but claims of superiority lack any baseline comparison.","lead":"This paper presents Hyper Vision Net, a U-Net variant for kidney and kidney tumor segmentation in CT scans that combines coordinate convolution, residual blocks, attention units, and decoder supervision layers. The authors report high Dice scores on the KiTS19 benchmark but provide no comparison to any other segmentation method on the same validation split.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation Dice are not evidence of state-of-the-art performance: no same-split baseline, no specified output head, and the abstract's superiority claim cites training-set scores.","rationale":"The reader's weakest assumption correctly identifies the central weakness: the validation Dice scores are presented as evidence of state-of-the-art performance without any same-split baseline comparison, and the multiple supervised output heads are not tied to a specified inference protocol. My review confirms this and adds two supporting observations: the abstract's superiority claim is explicitly based on training Dice rather than validation or test performance, and Section 4.2 contains a textual conflation of training and validation values, which further underscores the need for a precise evaluation protocol. This is an evidence problem rather than a conceptual one: the proposed architecture is coherent, the attention-layer ablation is a reasonable internal experiment, and the reported validation scores are plausible. The appropriate remedy is the reader's conditional requirement that the authors provide same-split baselines and specify which output head and postprocessing produce the reported numbers. Therefore I recommend no change to the existing CONDITIONAL verdict.","tokens_in":5200,"tokens_out":6686,"duration_ms":73239,"concrete_test":"Re-implement the architecture from Sections 3.2-3.3, train on the same 32175/13790 image split with the stated preprocessing, Adam schedule, and loss, then evaluate each of the three supervised outputs (two hyper vision layer outputs and the fused attention output) on the same 13790 validation slices using one documented inference rule (for example, argmax over the three output channels with no postprocessing). Train a vanilla U-Net baseline on the same split and protocol. If no output head reproduces the Table 1 validation values (0.8967 tumor, 0.9535 kidney), or if the U-Net baseline matches or exceeds them, the reported scores and the superiority claim are not verified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (abstract and Section 4.2) is that Hyper Vision Net is superior to state-of-the-art kidney tumor segmentation methods. The only quantitative evidence is Table 1, which reports 0.9552/0.9633 training Dice and 0.8967/0.9535 validation Dice for tumor/kidney with the attention layer. Three facts make this evidence insufficient for the claim. First, no baseline method is evaluated on the same split: Table 1 compares only the proposed model with and without attention, so the stated superiority over state-of-the-art is asserted rather than demonstrated. Second, the architecture trains three supervised outputs (Section 3.2 and Section 4.1): two hyper vision layer outputs and the fused final output. The paper does not identify which of these outputs is used to compute the reported Dice scores, so the numbers are not reproducible from the description. Third, the abstract explicitly bases the superiority claim on the training Dice score; training-set performance does not establish generalization, and the KiTS19-style evaluation would require a held-out comparison. The text in Section 4.2 also contains an internal inconsistency, attributing the validation values (0.8967, 0.9535) to training and then repeating the same values as a separate result, while the conclusion reports the training maximum. Because the comparison class is absent and the inference protocol is unspecified, the central superiority claim is not supported by the evidence as presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Hyper Vision Net, a U-Net-based architecture for kidney and kidney tumor segmentation in CT images from the KiTS19 dataset. The main additions are a coordinate convolutional layer at the input, residual blocks in the encoder/decoder, two 'hyper vision' layers with deep supervision in the decoder, and a CBAM-style attention unit. The authors report Dice scores for the model with and without the attention layer on the KiTS19 training and validation subsets (with attention: 0.9552/0.9633 training and 0.8967/0.9535 validation for tumor/kidney) and claim that the method is superior to state-of-the-art segmentation methods. The paper contains no comparison with any external baseline on the same split, does not specify which of the three supervised outputs is used for the reported scores, and contains an internal inconsistency in Section 4.2 where validation numbers are described as training results.","tokens_in":5505,"tokens_out":4886,"duration_ms":45321,"significance":"If the architecture were validated with appropriate comparisons, the combination of coordinate convolution, deep supervision, and attention within a U-Net is a plausible incremental contribution to medical image segmentation, and the use of a public dataset is commendable. The with/without attention ablation provides some internal evidence for the utility of the attention unit. However, the central claim of state-of-the-art superiority is not supported by the reported experiments: no same-split baseline is evaluated, the inference protocol is under-specified, and the abstract bases the headline claim on training-set Dice, which does not measure generalization. The paper therefore needs substantial additional experimental work before its significance can be assessed.","major_comments":[{"comment":"The central claim that Hyper Vision Net is superior to state-of-the-art methods is unsupported because no external baseline is evaluated on the same training/validation split. Table 1 only compares the proposed model with and without the attention layer; there is no comparison to U-Net, nnU-Net, or any published KiTS19 result on the same data partition. The abstract's reliance on the training Dice value of 0.9552/0.9633 for the superiority claim is particularly problematic, as training-set performance does not establish generalization. Please add at least one standard baseline (e.g., vanilla U-Net) evaluated under identical conditions and base the headline claim on validation or an independent test set.","section":"Section 4.2, Table 1, Abstract"},{"comment":"The inference head used to compute the reported Dice scores is not specified. The architecture produces three supervised outputs (two hyper vision layer outputs and the fused final output), each passed through a softmax layer under ground-truth supervision. The paper does not state which of these outputs generated the numbers in Table 1, nor how the final segmentation is derived (e.g., argmax over the three-class map). Without this information, the results are not reproducible from the description. Please clarify exactly which output head was evaluated and describe the post-processing steps.","section":"Section 3.2, Section 4.1"},{"comment":"The text in Section 4.2 is internally inconsistent with Table 1. The sentence 'during training, the proposed method achieves the Dice score of 0.8967 and 0.9535 for the tumor region and kidney region by involving attention layer' attributes the validation numbers (as listed in Table 1) to training, and the next sentence repeats the same values as a separate result. This makes it unclear which subset corresponds to which numbers and must be corrected to match the table.","section":"Section 4.2"},{"comment":"No error bars, multiple runs, or statistical tests are reported. Single-run Dice scores for medical segmentation can vary by several tenths of a point depending on initialization and data shuffling, so the observed differences—e.g., 0.9552 vs. 0.8967 for tumor Dice with vs. without attention—cannot be judged statistically meaningful without an estimate of variance. Please report the mean and standard deviation over at least three random seeds and state the number of runs.","section":"Section 4.2, Table 1"}],"minor_comments":[{"comment":"In Section 3.2, the text refers to 'the encoder and decoder structure of vanilla U-Net [13]', but reference [13] is Shen et al., DeepContour, not U-Net; the U-Net citation is [11]. Please correct the citation.","section":"Section 3.2, References"},{"comment":"The figure numbering is duplicated: Fig.3 is used both for the residual block structure in Section 3.2 and for the qualitative segmentation outputs in Section 4.2. Please renumber the figures sequentially.","section":"Section 3.2, Section 4.2"},{"comment":"The author affiliations are inconsistent: the author line lists superscripts 1, 2, and 3, but affiliation 3 has no corresponding author superscript, and the link between the authors and the affiliations is ambiguous. Please correct the affiliation markers.","section":"Title page"},{"comment":"The Dice loss formula in Eq. (2) has unclear summation indices (the limits 'i=0' and 'n=k' are not standard) and the denominator lacks parentheses, which makes the expression hard to parse. Please rewrite it in standard mathematical notation.","section":"Equation (2)"},{"comment":"The abstract states that the dataset includes 45,964 images, but Section 2 reports 32,175 training plus 13,790 validation images, which sum to 45,965. Please check and correct this one-image discrepancy.","section":"Abstract, Section 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is quite thin for a journal submission and would benefit from a much more thorough experimental validation, including same-split baselines and multi-run statistics. Also note that reference [15] (SkeletonNet) shares an author with this paper, and the self-citation is not explicitly flagged; this is a minor citation-pattern concern but worth keeping in mind during any revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a modest, unremarkable U-Net variant. The combination of CoordConv, residual blocks, CBAM-style attention, and decoder supervision may be new, but each piece is published, and the paper never shows the proposed model beating anything on the same split. The validation Dice values themselves are plausible—tumor 0.8967, kidney 0.9535 on KiTS19—but they are reported without error bars, without a baseline, and without specifying which of the three supervised outputs produced them.\n\nWhat I credit: the authors use the public KiTS19 split, describe the architecture in enough detail that a knowledgeable reader could reimplement it, and they are transparent about the training setup (Adam, learning rate schedule, batch size, 500 epochs). The qualitative images look reasonable. The with-attention vs without-attention internal comparison suggests the attention unit helps on training and validation, which is a minor but honest sanity check.\n\nNow the soft spots, in proportion. The biggest is that the abstract's superiority claim is built on the training Dice (0.9552/0.9633), which is not evidence of generalization. The validation Dice are lower and are the only held-out numbers, but no state-of-the-art method is evaluated on the same split, so 'comparable to SOTA' is not demonstrated. Second, Section 3.2 describes three supervised outputs—two hyper vision layers and the fused final output—but neither the text nor the table says which output is used to compute the reported Dice. That is a reproducibility gap. Third, Section 4.2 has an internal mismatch: it calls 0.8967/0.9535 training values, then repeats them as validation values, while Table 1 lists those numbers for validation and lists different values for training. This looks like a copy-paste error, but it needs fixing. Minor: reference [13] for U-Net actually points to DeepContour, the coordinate conv reference is correct, and no code is provided.\n\nIs this a serious contribution? As a scientific claim, no—it is an incremental engineering recipe without a fair comparison. But I would not call it dishonest; the authors clearly ran the experiments. The central flaw is not the architecture but the absence of a same-split baseline and a clearly specified inference protocol. That is fixable, but as submitted it does not support the title's implication of superiority.\n\nFor peer review: I would desk reject this version at a strong venue and ask the authors to first add baselines (at least a standard U-Net and a recent KiTS19 top method trained and evaluated on the same split), specify the inference head, report repeated runs with standard deviations, and release code. With those additions, the paper could be a modest workshop or short-paper contribution. Until then, I would not send it to a serious referee.","headline":"A routine U-Net variant with a plausible but unverified Dice claim; the central comparison to state of the art is not actually made.","tokens_in":5981,"tokens_out":2806,"would_cite":false,"duration_ms":29982,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a U-Net variant with coordinate convolution, supervised decoder layers, and channel-spatial attention reaches 0.9552 tumor Dice and 0.9633 kidney Dice on KiTS19 training images.","keywords":["Kidney tumor segmentation","U-Net","Coordinate convolution","Attention unit","KiTS19","Dice score","CT imaging","Deep learning"],"falsifier":"Train a plain U-Net with comparable depth, loss, and preprocessing on the 32,175 training images and evaluate on the same 13,790 validation images; if that baseline reaches or exceeds the reported 0.8967 tumor Dice, the paper's claim that Hyper Vision Net is superior is contradicted.","tokens_in":5051,"feed_emoji":"🩺","tokens_out":10409,"duration_ms":90692,"temperature":0.7,"pith_summary":"The paper seeks to show that a modified U-Net architecture, named Hyper Vision Net, can accurately segment kidney tumors and kidneys in the KiTS19 abdominal CT dataset, and that the reported Dice scores make it competitive with or better than existing segmentation methods. The proposed change is architectural: a coordinate convolution layer at the input, two supervised decoder \"hyper vision\" layers that refine fine details, and a channel-and-spatial attention unit applied to the fused decoder features. If correct, the work would offer a concrete recipe for improving fine-grained tumor segmentation in CT images, which matters for surgical planning and tumor assessment. The authors report training Dice scores of 0.9552 (tumor) and 0.9633 (kidney) with the attention unit, and a sharp drop without it, down to 0.8186 training tumor Dice.","feed_headline":"U-Net variant with attention scores 0.955 tumor Dice on KiTS19","feed_subtitle":"CoordConv, supervised decoder layers, and channel/spatial attention lift kidney-tumor segmentation on CT scans.","key_machinery":"The load-bearing mechanism is the combination of three components inside a U-Net encoder-decoder. First, a coordinate convolution layer prepends two coordinate channels to the input, letting the network use spatial coordinate information. Second, two \"hyper vision\" decoder layers produce intermediate segmentations that are each supervised with the ground truth and fused with the final encoder output; this is intended to refine even very small tumor regions. Third, a convolutional block attention module (CBAM) applies channel attention then spatial attention to the fused feature map, and its output is added back before the final softmax. The loss is the sum of categorical cross-entropy and separate Dice losses for the tumor and kidney channels.","core_discovery":"The central claim is that a U-Net architecture augmented with a coordinate convolution layer, two supervised decoder \"hyper vision\" layers, and a channel-and-spatial attention unit accurately segments kidney and tumor regions in the KiTS19 CT dataset. On the authors' split, the model with the attention unit achieves Dice scores of 0.9552 for tumor and 0.9633 for kidney on the 32,175-image training set, and 0.8967 and 0.9535 on the 13,790-image validation set. The paper credits the attention unit with most of the gain, because dropping it lowers training tumor Dice to 0.8186 and validation tumor Dice to 0.84. The stated conclusion is that this makes Hyper Vision Net superior to previous segmentation methods in terms of Dice score.","pith_inferences":["An implication the authors do not spell out is that the reported validation tumor Dice of 0.8967 is the more conservative estimate of real-world performance, since the training score is 0.9552; applying the model to new scans would likely land closer to the validation number.","The architecture's emphasis on spatial coordinate information could transfer to other small, low-contrast structures such as pancreatic lesions, where precise boundaries are hard to learn from intensity alone.","Because the paper works with 2D slices, a natural extension is to replace the 2D convolutions with 3D ones and keep the same attention-and-supervision structure, which would use inter-slice context and may reduce false positives.","The claim of state-of-the-art performance would be directly testable by running the same network and a plain U-Net on the same split; the paper does not include that comparison, so that test remains open."],"forward_implications":["If the scores hold, U-Net variants for abdominal CT segmentation can be expected to gain accuracy from decoder supervision layers and a channel-and-spatial attention module.","The attention ablation (training tumor Dice drops from 0.9552 to 0.8186) indicates that most of the reported accuracy comes from the attention unit rather than from the coordinate layer alone.","The loss formulation—cross-entropy plus a Dice term per target channel—can be transferred to other segmentation problems with severe class imbalance.","A direct practical consequence is that the method, as described, can be run on a single GPU with 256x256 slices, making it accessible for clinical research settings."],"supporting_citations":[{"why":"Supplies the U-Net encoder-decoder structure that the Hyper Vision Net modifies.","marker":"[11]"},{"why":"Provides the KiTS19 CT dataset of 300 patients used for training and validation.","marker":"[12]"},{"why":"Introduces the coordinate convolution layer that adds two spatial coordinate channels to the input.","marker":"[14]"},{"why":"Source of the \"hyper vision\" supervised decoder layers used to refine small regions.","marker":"[15]"},{"why":"Defines the channel-and-spatial attention module applied to the fused decoder features.","marker":"[16]"},{"why":"Supplies the residual blocks used to deepen the encoder and decoder.","marker":"[17]"}],"fun_headline_variants":["Attention unit lifts kidney tumor Dice to 0.955","Hyper Vision Net: attention-driven tumor Dice hits 0.955","CoordConv + attention: 0.955 tumor Dice on KiTS19","Supervised decoder + attention boost tumor Dice to 0.955"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the validation Dice scores it reports can be compared directly with published state-of-the-art results on KiTS19, even though it does not evaluate any baseline method on the same training/validation split.","fun_headline_variants_meta":{"raw":{"variants":["Attention unit lifts kidney tumor Dice to 0.955","Hyper Vision Net: attention-driven tumor Dice hits 0.955","CoordConv + attention: 0.955 tumor Dice on KiTS19","Supervised decoder + attention boost tumor Dice to 0.955"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000335,"raw_usage":{"total_tokens":1828,"prompt_tokens":888,"completion_tokens":940,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":865}},"tokens_in":504,"tokens_out":940,"duration_ms":9521,"temperature":1.0,"reasoning_tokens":865,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:15:36.772061+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a plain U-Net with comparable depth, loss, and preprocessing on the 32,175 training images and evaluate on the same 13,790 validation images; if that baseline reaches or exceeds the reported 0.8967 tumor Dice, the paper's claim that Hyper Vision Net is superior is contradicted.","supporting_citations":[{"cited_title":"U -net: Convolutional networks for biomedical image segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net encoder-decoder structure that the Hyper Vision Net modifies."},{"cited_title":"The KiTS19 Challenge Data: 300 Kidney Tumor Cases with Clinical Context, CT Semantic Segmentations, and Surgical Outcomes","cited_arxiv_id":null,"evidence_quote":"Provides the KiTS19 CT dataset of 300 patients used for training and validation."},{"cited_title":"SkeletonNet: Shape Pixel to Skeleton Pixel","cited_arxiv_id":"1907.01683","evidence_quote":"Source of the \"hyper vision\" supervised decoder layers used to refine small regions."},{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the residual blocks used to deepen the encoder and decoder."}],"review_version":1}