{"id":"dcb504d0-eaa9-4cde-bc6d-50b4d479d6f5","arxiv_id":"2504.12652","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"AdaptoVision combines residual, depthwise, and hierarchical skip connections for image classification, but its state-of-the-art claims are contradicted by its own comparison tables and no code is provided.","lead":"AdaptoVision is a new convolutional neural network that combines residual connections, depthwise convolutions, and multi-scale skip connections, and reports accuracy on CIFAR, Flowers, Cars, BreakHis, and ISIC. The paper claims state-of-the-art results on medical images, but internal tables contradict those claims and no code is released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BreakHis 70/30 split may be tile-level, causing train/test leakage that inflates reported SOTA; patient-level split needed.","rationale":"The reader's weakest assumption is correct and is the single most load-bearing issue. The entire medical-image contribution hangs on the BreakHis numbers, and the tiling protocol creates a concrete mechanism for leakage if the split is not image/patient-level. The paper's silence on this point, combined with the absence of code, makes the reported accuracies unverifiable. I find no additional concern that changes the verdict; the internal inconsistencies (e.g., ImageNet text vs. Table 7, CIFAR text vs. Table 3) further support rejection but are secondary to the leakage risk.","tokens_in":11580,"tokens_out":3644,"duration_ms":34914,"concrete_test":"Obtain the exact BreakHis split used in Section 3.1.1 and verify whether any source image contributes tiles to both training and test. If so, rerun the binary and 8-class experiments with a patient-level split (all tiles from a given patient in one partition) and compare accuracy per magnification. A drop of more than a few points below the best baseline in Table 4 would confirm leakage and invalidate the SOTA claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim for BreakHis (Section 4, Table 4) rests on a 70/30 split of tiled images (Section 3.1.1). Tiles are generated with a sliding window of step size 4 pixels and size 126x128, so adjacent tiles from the same source image overlap substantially. If the 70/30 split is applied to the tiled corpus rather than to source images or patients, training and test sets will contain near-duplicate views of the same tissue region. This can inflate accuracy to the reported 99%+ levels, making the 'state-of-the-art' claim an artifact of data leakage. The paper never states the split granularity, and no code is released to check it. Competing methods in the same table (e.g., Yari et al. at 100%) also undermine the SOTA claim, but the unresolvable leakage risk is the load-bearing flaw.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AdaptoVision, a CNN architecture combining enhanced residual units, depth-wise separable convolutions, and hierarchical skip connections. The authors claim state-of-the-art results on BreakHis and ISIC 2019, and accuracy comparable to heavier models on CIFAR-10/100, Flowers-102, and Stanford Cars, while using significantly fewer parameters and FLOPs. The paper reports experiments on these datasets and compares with various baselines.","tokens_in":11806,"tokens_out":4327,"duration_ms":37793,"significance":"If the reported results were valid, a lightweight CNN achieving near-SOTA across multiple benchmarks would be a useful contribution to efficient image recognition. However, the paper's own tables contradict the central claims: the ImageNet result is far below every baseline listed, and the BreakHis SOTA claim is not supported by the numbers in the same table. The reported efficiency gains rest on an incorrect complexity comparison. No code or detailed experimental protocol is provided, further limiting the value of the empirical results.","major_comments":[{"comment":"The text states that the proposed method 'consistently outperforms existing ConvNets' on ImageNet, but Table 7 reports 66.70% top-1 and 86.60% top-5, while every listed baseline achieves at least 76.0% top-1. Moreover, EfficientNet-B0 achieves 77.1% top-1 with 5.3M parameters and 0.3B FLOPs, while AdaptoVision uses 6.7M parameters and 4.9G FLOPs; the proposed model is therefore neither more accurate nor more efficient than this baseline. This contradiction invalidates the paper's efficiency-accuracy trade-off claim.","section":"Section 4, Table 7"},{"comment":"The BreakHis evaluation protocol is not described at the level needed to rule out data leakage. The paper states that images were tiled with a sliding window (step 4 pixels, tile size 126x128) and that the dataset was split 70/30, but it does not state whether the split is at the patient, image, or tile level. Because adjacent tiles overlap heavily, a tile-level split would place near-duplicate crops in both training and test sets and could inflate the reported accuracies to the 98-99% range. Without code or clarification, the SOTA claims in Tables 4 and 5 are not verifiable.","section":"Section 3.1.1, Tables 4 and 5"},{"comment":"The claim that AdaptoVision 'achieved state-of-the-art accuracy in all three magnification images' is contradicted by the same table: Yari et al. report 100% at 40X and 100X, and 99.26% magnification-independent, whereas the proposed method reports 99.37%, 98.67%, 99.33%, 99.17%, and 99.43%. Even accepting the reported numbers, the proposed method is not the top performer in the table, so the SOTA claim is unsupported.","section":"Section 4, Table 4"},{"comment":"The complexity comparison is not credible: the table caption says input size 224x224, while the text says 32x32; more importantly, ResNet-18 is listed at 768 GFLOPs, which is roughly 400 times larger than the standard published value (~1.8 GFLOPs at 224x224). With an inflated baseline, the claim that AdaptoVision 'reduces FLOPs by nearly 4x' has no quantitative support. The parameter count comparison (ResNet-18 11.7M vs. proposed 6.6M) is also less favorable than the 'more than 4x' claim suggests.","section":"Section 3.4, Table 2"}],"minor_comments":[{"comment":"The table number 'Table 1' is used both for the dataset summary in Section 3.1 and for the architecture configuration in Section 3.3; renumber to avoid ambiguity.","section":"Section 3.3 and Table 1"},{"comment":"AlexNet is cited as Hinton et al. [2012], but the canonical citation is Krizhevsky et al. (2012); several references have incomplete bibliographic data (e.g., 'dvvf', 'bhjbjh', 'Unknown Booktitle').","section":"References"},{"comment":"The text references 'Figure 2: Exploring Accuracy...', but no actual figure appears in the manuscript; check whether the figure is missing or the reference is erroneous.","section":"Figure 2"},{"comment":"The claim of SOTA accuracy on ISIC 2019 is based on Table 6, which lists only one baseline and no statistical significance; the claim should be qualified accordingly.","section":"Section 4.1, ISIC 2019"}],"recommendation":"reject","confidential_remarks":"This manuscript is far from the standards of a peer-reviewed venue: the central claims are contradicted by its own tables, the experimental protocol for BreakHis is underspecified, and the presentation has many citation and formatting errors. I see no straightforward revision that would address the ImageNet contradiction and the incorrect complexity numbers, since these require re-running the experiments and correcting the analysis. The paper would need to be substantially reworked before it could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper combines known building blocks into a new arrangement, but its own tables contradict its headline claims. The ImageNet section says the model 'consistently outperforms' existing ConvNets, yet Table 7 shows 66.7% top-1 against baselines at 76-80%. That is not a minor discrepancy; it is the central claim collapsing. The BreakHis 'SOTA' claim also ignores Yari et al.'s 100% at 40X and 100X in the same table.\n\nWhat's genuinely new: the specific recipe—enhanced residual units, depthwise separable convs in a residual Block-2, and hierarchical skip connections—is not in the cited literature. The paper also makes an effort to report params and FLOPs across many datasets. But without an ablation, a precise layer-by-layer spec, or released code, the novelty is unverifiable.\n\nThe soft spots are serious. The CIFAR-10 text reports 98% after transfer learning, but Table 3 lists 95.3%. The Flower/Cars text also doesn't match the table. The BreakHis tiling creates a real leakage risk: sliding window with step 4 pixels over 126x128 tiles produces heavily overlapping crops, and the 70/30 split granularity is never stated. If the split is applied to tiles rather than patients, the 99%+ numbers are artifacts. The paper should say whether it's patient-level, image-level, or tile-level; it doesn't. There are also citation artifacts ('bhjbjh' as a venue) and a complexity comparison whose FLOPs numbers (ResNet-18 at 768G?) don't match reality.\n\nThis paper doesn't support its central claims. The internal contradictions alone would justify rejection, and the leakage risk makes the medical result unreliable. The authors seem to have put effort into running experiments, but the reporting is too sloppy to trust. If they release code, specify the architecture, fix the tables/text mismatches, and use a patient-level split for BreakHis, the underlying idea might be worth another look. As it stands, a desk reject is appropriate.\n\nRecommendation: don't send to peer review in this state; encourage the authors to address the above and resubmit. Not something I'd bring to reading group, and I wouldn't cite it.","headline":"The architecture is a plausible mixture of known techniques, but the paper's own tables contradict its SOTA claims and the BreakHis evaluation likely leaks; desk-reject.","tokens_in":12334,"tokens_out":6062,"would_cite":false,"duration_ms":55559,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a 6.6-million-parameter CNN, AdaptoVision, can match or beat much larger models on standard benchmarks and reach top reported accuracy on breast-tumor images, all without pretrained weights.","keywords":["AdaptoVision","efficient CNN","image classification","depth-wise separable convolution","enhanced residual unit","hierarchical skip connection","BreakHis","CIFAR"],"falsifier":"Re-run the BreakHis evaluation with a patient-level split (all tiles of each patient in one set only) and compare the reported 99.43% magnification-independent accuracy; if accuracy falls markedly—say below 95%—the SOTA claim is inflated by tile leakage.","tokens_in":11366,"feed_emoji":"🩺","tokens_out":6719,"duration_ms":60027,"temperature":0.7,"pith_summary":"This paper claims that a deliberately compact convolutional network, AdaptoVision, can match or exceed much larger models on standard benchmarks (95.3% on CIFAR-10, 85.77% on CIFAR-100) and can reach top reported accuracy on the BreakHis breast-tumor dataset (99.43% magnification-independent), all without pretrained weights. The motivation is practical: if true, high-accuracy image recognition does not require billion-parameter models or transfer learning, which matters for phones, edge devices, and clinical settings with limited compute. The authors attribute the gain to three architectural choices—enhanced residual units, depth-wise separable convolutions in a residual block, and hierarchical skip connections—and support the claim with parameter and FLOP counts far below ResNet-18 and comparable to MobileNetV2.","feed_headline":"No-pretraining 6.6M CNN tops breast-tumor benchmark","feed_subtitle":"AdaptoVision reports 99.43% on BreakHis and 95.3% on CIFAR-10 with a fraction of ResNet's compute.","key_machinery":"The load-bearing components are (1) the Enhanced Residual Unit (ERU), which replaces a shallow residual mapping with a deeper transformation chain $T(x)=W^{(4)}*\\sigma(W^{(3)}*B_2(\\sigma(W^{(2)}*\\sigma(W^{(1)}*x))))$; (2) Block-2, a residual-compatible inception-like unit using $1\\times1$ pointwise convolutions around a depth-wise convolution, costing $O(h\\cdot w\\cdot c\\cdot k^2)$; (3) hierarchical skip connections that sum feature maps from prior encoder blocks after projection, keeping gradient flow across depth; and (4) a global average pooling plus reshape aggregation before the final projection. The complexity formula $O_{ERU}=2h w c^2 k^2$, together with hardware-aware constraints, is the quantitative backbone for the efficiency claim.","core_discovery":"The paper's central claim is that a from-scratch CNN using enhanced residual units, a depthwise-separable 'Block-2', and long-range skip connections can achieve accuracy competitive with or better than much larger pretrained models while using roughly 6.6 million parameters and 4.9 GFLOPs. On BreakHis binary classification the paper reports 99.43% accuracy across magnifications, above every comparison listed; on ISIC 2019 it reports 95.30%; on CIFAR-10/100 it reports 95.30%/85.77%, 'comparable' to strong baselines though below the best listed at 98-99.5% on CIFAR-10. The authors frame the result as evidence that architectural efficiency and information flow, not scale or pretraining, can carry competitive classification.","pith_inferences":["Extending the paper's logic, the architecture's efficiency suggests it could serve as a backbone for object detection or segmentation, where parameter counts multiply quickly; the paper does not test this.","The strongest test of the BreakHis claim would be re-running the pipeline with a patient-level split; if accuracy stays near 99%, the generalization claim is substantially stronger than the tile-level protocol currently supports.","The CIFAR numbers are presented as 'comparable' rather than state-of-the-art; a fair reader inference is that the architecture trades a few accuracy points for large efficiency gains, and the useful comparison is accuracy per FLOP rather than raw accuracy."],"forward_implications":["If the claims hold, high-accuracy classifiers can be trained from scratch on consumer hardware, removing the reliance on pretrained ImageNet weights.","The reported parameter and FLOP reductions imply that similar accuracy can be reached with an order-of-magnitude less compute than ResNet-18, which matters for real-time and embedded deployment.","On BreakHis, the architecture would offer a single magnification-independent model that beats specialized medical pipelines without transfer learning.","The same design principles—deep residual transforms plus depthwise separable operations—could transfer to other fine-grained medical or benchmark tasks."],"supporting_citations":[{"why":"Supplies the ResNet identity skip connection that the ERU extends and the complexity comparison measures against.","marker":"He et al. [2016]"},{"why":"Supplies the MobileNetV2 depthwise-separable-convolution design and the efficiency baseline for parameter and FLOP comparison.","marker":"Sandler et al. [2019]"},{"why":"Introduces the BreakHis dataset whose 70/30 protocol the breast-tumor SOTA claim is evaluated on.","marker":"Spanhol et al. [2016]"},{"why":"Provides the CIFAR-10 and CIFAR-100 benchmark datasets used for the claimed comparable accuracy.","marker":"Krizhevsky et al. [2009]"},{"why":"DenseNet is cited as the dense-connectivity approach that the proposed hierarchical skip connections build on.","marker":"Huang et al. [2017]"},{"why":"Its 98-99% BreakHis results are the closest prior numbers the paper directly competes with under a 70/30 protocol.","marker":"Alkassar et al. [2021]"},{"why":"Supplies per-magnification BreakHis baselines that appear in the comparison table.","marker":"Gour et al. [2020]"}],"fun_headline_variants":["6.6M-param CNN hits 99.4% on breast cancer","From-scratch CNN beats pretrained on BreakHis","Lightweight CNN matches giants without pretraining","99.43% on BreakHis with just 6.6M parameters","Efficient CNN: 6.6M params, competitive on medical and CIFAR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The BreakHis 99%+ accuracies assume that the 70/30 train/test split keeps tiles originating from the same source image from appearing in both sets; the paper never states or justifies the split level.","fun_headline_variants_meta":{"raw":{"variants":["6.6M-param CNN hits 99.4% on breast cancer","From-scratch CNN beats pretrained on BreakHis","Lightweight CNN matches giants without pretraining","99.43% on BreakHis with just 6.6M parameters","Efficient CNN: 6.6M params, competitive on medical and CIFAR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00098,"raw_usage":{"total_tokens":4115,"prompt_tokens":853,"completion_tokens":3262,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":3168}},"tokens_in":469,"tokens_out":3262,"duration_ms":23864,"temperature":1.0,"reasoning_tokens":3168,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:26:20.266855+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the BreakHis evaluation with a patient-level split (all tiles of each patient in one set only) and compare the reported 99.43% magnification-independent accuracy; if accuracy falls markedly—say below 95%—the SOTA claim is inflated by tile leakage.","supporting_citations":[],"review_version":1}