{"id":"d2fa2bd7-0182-411e-a516-7997242a874b","arxiv_id":"2507.01279","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"ResNet+, a ResNet-D and CBAM hybrid, achieves up to 99.25% accuracy on lung cancer CT classification and 98.14% on histopathology, though the reported metrics contain inconsistencies.","lead":"This paper presents ResNet+, a convolutional network that combines two existing techniques, ResNet-D downsampling and CBAM attention, for classifying lung cancer and disease images. The authors report accuracy gains over plain ResNet on five public datasets, with the best results of 99.25% accuracy on a CT scan dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table II's impossible accuracy/F1 combinations invalidate the reported superiority of ResNet+; the empirical claim is not supported.","rationale":"The paper's main contribution is an architecture modification (ResNet-D + CBAM) whose value is demonstrated solely through reported test metrics. The reader's verdict identifies the key weakness: the reported evaluation is internally inconsistent. I focus on the strongest, most directly fatal evidence: Table II contains impossible accuracy/F1 pairs on a balanced test set, and the COVIDxCT rows repeat the exact precision/recall/F1 values from LCC, indicating unreliable table construction. This is not a matter of disagreement with community consensus; it is an internal mathematical contradiction. A reader cannot accept the headline 'remarkable accuracy/F1' when the numbers cannot simultaneously be true. The concrete test is feasible because the authors provide a public code repository; recomputing the LCC metrics would settle the question. I do not heavily weigh the patient-level split issue because the metric inconsistency already invalidates the claimed results, though that issue would also warrant attention if the metrics were reproduced. The verdict should remain REJECT, as the central empirical claim is unsupported.","tokens_in":13439,"tokens_out":4079,"duration_ms":46647,"concrete_test":"Run the published GitHub code on the LCC test set (5000 balanced images) to produce the confusion matrix for ResNet50+, then compute accuracy, macro precision, recall, and F1 with sklearn using the standard multiclass definitions. If the reproduced numbers match Table II, the metric definitions are non-standard or the test set is not the one described in Section III-A; if instead F1 is near 99.9%, Table II contains a major reporting error. Also verify that the COVIDxCT rows' precision, recall, and F1 are not identical to the LCC rows.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that ResNet+ achieves superior classification accuracy rests entirely on the test metrics in Table II, which contain internally impossible values. For the LCC dataset, Section III-A states a balanced test set of 5 classes x 1000 images. Yet Table II reports ResNet50+ with ACC=99.99%, PRE=81.85%, REC=84.58%, F1=79.85%. On a balanced 5000-sample test, 99.99% accuracy admits no integer number of correct predictions (5000 x 0.9999 = 4999.5), and even if rounded, it permits at most one misclassification, forcing macro precision, recall, and F1 to be approximately 99.98%, not ~80%. No standard multiclass metric definition reconciles these numbers. Moreover, the COVIDxCT rows for ResNet50+ and ResNet101+ exactly reuse the PRE/REC/F1 values from the LCC rows, a clear copy-paste artifact. Because the paper's headline numbers are mutually inconsistent, the accuracy and F1 claims cannot be taken as valid experimental evidence. The absence of a patient-level split description for IQ-OTH/NCCD is a further concern, but the impossible metrics alone are sufficient to invalidate the evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ResNet+, a modification of ResNet50 and ResNet101 that combines ResNet-D downsampling changes with Convolutional Block Attention Modules (CBAM), and evaluates it on five public lung image datasets: LC25000, IQ-OTH/NCCD, LCC, ChestXray, and COVIDx-CT. The authors report high test accuracies (e.g., 98.14% on LC25000, 99.25% on IQ-OTH/NCCD), claim reduced computational cost relative to the original ResNet series, present ablations of CBAM and ResNet-D components, and describe a 'multimodal' experiment that pools LC25000 and IQ-OTH/NCCD images. Public code is linked in the abstract.","tokens_in":13800,"tokens_out":8233,"duration_ms":89988,"significance":"If the reported results were valid, the paper would provide a simple architectural recipe—ResNet-D plus CBAM—that improves multiclass lung-image classification across several public datasets and lowers training time on some of them. The paper includes useful components: publicly available code, multiple public datasets, and a systematic ablation of CBAM and ResNet-D. However, the empirical foundation contains internally inconsistent metrics and a probable data-splitting problem, so the significance of the contribution cannot be assessed until those issues are resolved. There is no circularity in the central claim, since the results are external benchmark comparisons rather than derivations from the data.","major_comments":[{"comment":"The LCC ResNet50+ row reports ACC=99.99%, PRE=81.85%, REC=84.58%, F1=79.85% on a test set that Section III-A states is balanced with 5 classes x 1000 images. An accuracy of 99.99% would require 4999.5 correct predictions out of 5000, which is impossible; if rounded from 4999/5000 the value would be 99.98%, not 99.99%. Moreover, with at most one misclassification, macro precision, recall, and F1 would all be near 99.98%, not the 80% range. The same PRE/REC/F1 values are then repeated exactly in the COVIDxCT rows of Table II (e.g., ResNet50+ and ResNet101+), despite COVIDxCT having a differently sized, imbalanced three-class test set. These internal inconsistencies invalidate the LCC and COVIDxCT performance claims in Section IV-B.","section":"Table II, Section III-A"},{"comment":"The IQ-OTH/NCCD dataset is described as CT slices derived from 110 patients, yet the paper reports using 1336 images split into 1070/133/133 without any patient-level split. Since a patient can contribute multiple slices, an image-level random split very likely places slices of the same patient in both training and test sets, causing patient-identity leakage and inflating the 99.25% accuracy highlighted in the abstract. The authors need to specify how the split was performed at the patient level, or demonstrate that no patient appears in both sets; without this, the IQ-OTH/NCCD result is not interpretable.","section":"Section III-A (IQ-OTH/NCCD)"},{"comment":"The abstract's unqualified claim that ResNet+ 'saved computational cost compared to the original ResNet series' is contradicted by the paper's own measurements. Table IV shows that ResNet50+ and ResNet101+ have higher per-sample inference latency than the corresponding baselines on every dataset (e.g., 7.63±0.92 ms vs 6.52±0.78 ms on IQ-OTH/NCCD; 11.85±1.54 ms vs 11.07±1.33 ms on LC25000). Table III also shows greater training time for ChestXray (154.16 vs 147.39 min for ResNet50+ vs ResNet50). The cost savings appear only for some training runs on lung-cancer datasets, so the claim should be restricted to that setting or removed from the abstract.","section":"Abstract, Section IV-B, Tables III-IV"},{"comment":"The experiment described as 'multimodal' simply pools LC25000 histopathology images and IQ-OTH/NCCD CT images into one training set. No modality encoding, separate branch, modality-conditional mechanism, or modality label is provided, so a single CNN trained on the combined images cannot exploit the fact that the inputs come from different modalities. This does not support the paper's stated contribution of combining different modalities; it is at best a multi-dataset pooling experiment. The claim needs to be substantially reframed or tested with an architecture that actually models modality identity.","section":"Section IV-B (Multi-modality data, Table VI)"},{"comment":"The abstract's statement that 'The proposed model outperformed the baseline models on publicly available datasets' is not supported by the data presented. In Table II, ResNet50+ has lower accuracy than ResNet50 on COVIDxCT (78.45% vs 80.58%), and in Table VIII ResNet50+ has lower accuracy than ResNet50 on ISIC2018 (67.98% vs 68.12%). Section V itself acknowledges that ResNet50+ performed worse than its standard model on a large dataset. The conclusions should be qualified to state which datasets and metrics support the claimed improvement.","section":"Abstract, Section V"}],"minor_comments":[{"comment":"The F1-score formula is written as '2 × TP / 2 × FP + FN', which omits TP in the denominator; the correct expression is 2*TP / (2*TP + FP + FN).","section":"Eq. (6)"},{"comment":"The ChestXray dataset size is inconsistent: Section III-A says 5863 images, while Table I sums to 5216+16+624=5856 and the abstract reports n=5856. Please unify these counts.","section":"Section III-A, Table I, Abstract"},{"comment":"The text says the ResNet+ models give '~4% improvements' in ACC on ChestXray, but the differences in Table II are about 3.0 percentage points (87.98 vs 84.94 and 83.01 vs 79.97), not 4%.","section":"Section IV-B"},{"comment":"The dataset name is written inconsistently as IQ-OTHNCDD, IQ-OTH/NCCD, and IQ-OCTNCCD; please choose one spelling and use it consistently.","section":"Throughout"},{"comment":"The caption says 'Scc' stands for 'small cell lung cancer' in the LC25000 confusion matrix, but in the paper 'Scc' denotes squamous cell carcinoma; please correct the caption.","section":"Figure 4 caption"},{"comment":"The training description gives a batch size, optimizer, EMA decay, and 200 epochs, but no random seeds or exact data-split procedure are reported; adding this information would improve reproducibility beyond the public code link.","section":"Section IV-A"}],"recommendation":"reject","confidential_remarks":"The exact reuse of PRE/REC/F1 values across two completely different test sets in Table II, together with the numerically impossible accuracy/F1 combination on the balanced LCC test set, is difficult to explain as a typo. I would recommend that the editor ask for raw prediction files or checkpoints before considering any revision, because the current empirical record does not support the central claims. The self-citations [43]-[44] are confined to the limitations paragraph and are not a concern for me."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a routine architecture combination (CBAM + ResNet-D) with a systematic-but-broken evaluation. The LCC row in Table II reports 99.99% accuracy on a balanced 5000-image test set with an F1 of 79.85%. That is mathematically impossible: 99.99% accuracy permits at most one misclassification, which forces macro precision/recall/F1 to be near 99.98%. No standard metric definition reconciles those numbers. On top of that, the COVIDxCT rows for ResNet50+ and ResNet101+ exactly reuse the LCC PRE/REC/F1 values, which is a copy-paste artifact. These two facts alone destroy the paper's central claim that ResNet+ outperforms baselines.\n\nWhat is genuinely there: the combination of ResNet-D and CBAM is not novel, but the paper does run it on five public datasets, includes ablations for each component, and reports training time and inference latency. The ablations in Tables V and VII are a reasonable attempt to isolate where gains come from. The code is public. That is more than many submissions in this space do.\n\nThe soft spots are not all created equal. The impossible metrics are load-bearing and disqualifying. The absence of a patient-level split description matters too, especially for IQ-OTH/NCCD, which comes from only 110 patients; random image-level splits in that setting produce optimistic results due to leakage. The computational-cost claim is weaker than the abstract suggests: Table IV shows ResNet+ has higher per-sample inference time, and Table III shows mixed training-time results. The paper's own text half-acknowledges this, but the abstract still says \"saved computational cost,\" which is misleading.\n\nWho is this for? Someone doing a quick engineering survey of attention-plus-ResNet variants on lung images might skim the ablations, but nobody should trust the headline numbers. The internal contradictions are clear enough that a desk rejection is appropriate; sending this to referees would waste their time on a paper whose own reported metrics cannot be true.\n\nRecommendation: do not engage. If the authors fix the metric computation and redo the splits with patient-level grouping, the underlying idea might be publishable as a modest incremental study, but this version is not.","headline":"The paper's central empirical claim is invalidated by impossible metric combinations in Table II, despite some useful ablations and code.","tokens_in":14147,"tokens_out":2159,"would_cite":false,"duration_ms":25276,"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":"The paper proposes ResNet+, a ResNet variant that inserts CBAM attention into bottleneck blocks and adopts ResNet-D downsampling, and reports that it outperforms ResNet50/101 on five public lung imaging datasets, with best results of…","keywords":["lung cancer","deep learning","ResNet","CBAM","ResNet-D","medical image classification","CT imaging","histopathology images"],"falsifier":"Reproduce the LCC confusion matrix: the test set is balanced at 1,000 images per class, so an accuracy of 99.99% would mean essentially every one of the 5,000 images is correct and the macro F1 would sit near 99%, not at the reported 79.85%. Recomputing accuracy, precision, recall, and F1 from the same predictions—and checking that the 110 IQ-OTH/NCCD patients are split at the patient level—would settle whether the reported gains are real.","tokens_in":13235,"feed_emoji":"🫁","tokens_out":13859,"duration_ms":132868,"temperature":0.7,"pith_summary":"The paper tries to show that a small architectural change to a standard image-classification network can make lung-cancer screening from medical images both more accurate and cheaper. It adds a channel-and-spatial attention module (CBAM) to every bottleneck block of ResNet50/101 and replaces the downsampling path with ResNet-D's average-pooling shortcuts, calling the result ResNet+. Across five public datasets spanning histopathology slides, CT slices, and chest X-rays, the modified model is reported to beat the unmodified ResNet series on accuracy and on the F1 score, a combined precision-and-recall measure, while using less training time on the two lung-cancer sets. If those numbers hold, the recipe of attention plus lossless downsampling would be a low-cost upgrade for image-based cancer classifiers.","feed_headline":"ResNet+ hits 98.14% accuracy on lung cancer images","feed_subtitle":"Adding attention and smarter downsampling to ResNet beats original baselines on five public medical imaging datasets.","key_machinery":"The central mechanism is the ResNet+ block: a ResNet bottleneck whose shortcut path uses ResNet-D's average-pooling-then-1-by-1-convolution downsampling, and whose feature map passes through a Convolutional Block Attention Module (CBAM) before being added to the shortcut. ResNet-D alters the stem and shortcut so that stride-2 downsampling does not discard feature information, while CBAM—channel attention built from squeeze-and-excitation on pooled features, then spatial attention built from a 7-by-7 convolution on concatenated channel-pooled maps—reweights features toward informative regions. The paper's claim is that these two additions together, rather than either one alone, produce the accuracy gains and the training-time savings.","core_discovery":"The paper's central claim is that equipping ResNet50 and ResNet101 with ResNet-D's downsampling design and a CBAM attention module—the combination it calls ResNet+—improves classification accuracy across five public lung-cancer and lung-disease image datasets compared with the unmodified ResNet series. On the paper's reported numbers, the best variant reaches 98.14% accuracy and F1 on LC25000 histopathology images, 99.25% accuracy and 99.13% F1 on IQ-OTH/NCCD CT slices, and about two percentage points of accuracy gain on the five-class LCC set, with ResNet50+ also reducing training time on the two lung-cancer datasets relative to ResNet50. The paper further claims that ablations show CBAM and ResNet-D are complementary, and that the resulting model transfers to a CT-plus-pathology multimodal task and to skin-lesion classification.","pith_inferences":["The paper does not report a patient-level split for IQ-OTH/NCCD; if the 99.25% accuracy persisted under such a split, that would be strong evidence of generalization, whereas image-level splits can overstate performance on CT slices from the same patient.","The computational-cost claim is tied to one GPU and to the time until the best validation model; a fairer extension would report FLOPs or energy across hardware, since the ChestXray and large CT results do not show consistent savings.","A direct recomputation of all metrics from one confusion matrix for the LCC test set would settle the paper's central uncertainty: either accuracy and F1 reconcile, or a reporting error hides the true comparison.","Because CBAM is a generic attention module, the same architecture could be tested on other imbalanced pathology datasets; a reasonable extension is to compare training-from-scratch against fine-tuning pre-trained weights."],"forward_implications":["Applying ResNet-D plus CBAM to an existing ResNet classifier could raise accuracy on other medical image tasks without changing the data pipeline or the parameter count much.","The ablation results imply that attention alone or downsampling alone gives smaller gains, so future architecture search should treat the two modifications as a paired recipe.","The reported training-time savings mean that on small CT and histopathology sets, retraining the improved model would be faster, making higher accuracy available at lower compute cost.","Data augmentation for underrepresented classes appears sufficient to reach high accuracy on the imbalanced IQ-OTH/NCCD set, pointing to a reusable recipe for other small medical datasets.","The multimodal CT-plus-pathology experiment suggests the same architecture can fuse different imaging modalities without a redesign."],"supporting_citations":[{"why":"[13] supplies the ResNet-D downsampling modifications—average-pooling shortcuts and swapped convolution order—that ResNet+ adopts.","marker":"[13]"},{"why":"[14] provides the CBAM channel-plus-spatial attention module that ResNet+ inserts into bottleneck blocks.","marker":"[14]"},{"why":"[26] is the LC25000 histopathology dataset used for the main lung cancer evaluation.","marker":"[26]"},{"why":"[27] is the IQ-OTH/NCCD CT-slice dataset used for the second lung cancer evaluation.","marker":"[27]"},{"why":"[28] is the LCC five-class lung and colon histopathology dataset used for a large-scale test.","marker":"[28]"},{"why":"[29] is the ChestXray dataset used to test pneumonia versus normal classification.","marker":"[29]"},{"why":"[30] is the COVIDx-CT dataset used to test the model on a large CT cohort.","marker":"[30]"},{"why":"[39] is an attention-equipped ResNet baseline that ResNet+ is compared against on LC25000.","marker":"[39]"},{"why":"[40] is a hybrid CNN-transformer baseline that ResNet+ is compared against on LC25000.","marker":"[40]"}],"fun_headline_variants":["ResNet+ beats ResNet on lung cancer images","Attention module lifts lung cancer AI accuracy","ResNet+ hits 99.25% on lung CT slices","ResNet variant improves lung disease detection","ResNet+ tops baselines on five medical datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported test metrics are genuine, meaning they are computed from the same predictions and no images of the same patient appear in both the training and test sets; the LCC row of Table II, which pairs 99.99% accuracy with 79.85% F1 on a balanced test set, is not compatible with that premise.","fun_headline_variants_meta":{"raw":{"variants":["ResNet+ beats ResNet on lung cancer images","Attention module lifts lung cancer AI accuracy","ResNet+ hits 99.25% on lung CT slices","ResNet variant improves lung disease detection","ResNet+ tops baselines on five medical datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1422,"prompt_tokens":1071,"completion_tokens":351,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":687,"completion_tokens_details":{"reasoning_tokens":278}},"tokens_in":687,"tokens_out":351,"duration_ms":104501,"temperature":1.0,"reasoning_tokens":278,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:55:12.393422+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the LCC confusion matrix: the test set is balanced at 1,000 images per class, so an accuracy of 99.99% would mean essentially every one of the 5,000 images is correct and the macro F1 would sit near 99%, not at the reported 79.85%. Recomputing accuracy, precision, recall, and F1 from the same predictions—and checking that the 110 IQ-OTH/NCCD patients are split at the patient level—would settle whether the reported gains are real.","supporting_citations":[{"cited_title":"Bag of tricks for image classification with convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"[13] supplies the ResNet-D downsampling modifications—average-pooling shortcuts and swapped convolution order—that ResNet+ adopts."},{"cited_title":"Cbam: Convolutional block attention module","cited_arxiv_id":null,"evidence_quote":"[14] provides the CBAM channel-plus-spatial attention module that ResNet+ inserts into bottleneck blocks."},{"cited_title":"Lc25000 lung and colon histopathological image dataset","cited_arxiv_id":null,"evidence_quote":"[26] is the LC25000 histopathology dataset used for the main lung cancer evaluation."},{"cited_title":"The IQ-OTHNCCD Lung Cancer Dataset, 2021","cited_arxiv_id":null,"evidence_quote":"[27] is the IQ-OTH/NCCD CT-slice dataset used for the second lung cancer evaluation."},{"cited_title":"Identifying medical diagnoses and treatable diseases by image-based deep learning","cited_arxiv_id":null,"evidence_quote":"[29] is the ChestXray dataset used to test pneumonia versus normal classification."},{"cited_title":"Covidnet-ct: A tailored deep convolutional neural network design for detection of covid- 19 cases from chest ct images","cited_arxiv_id":null,"evidence_quote":"[30] is the COVIDx-CT dataset used to test the model on a large CT cohort."},{"cited_title":"Crorelu: Cross-crossing space-based visual activation function for lung cancer pathology image recognition","cited_arxiv_id":null,"evidence_quote":"[39] is an attention-equipped ResNet baseline that ResNet+ is compared against on LC25000."},{"cited_title":"Transformer-based semantic segmentation and cnn network for detection of histopathological lung cancer","cited_arxiv_id":null,"evidence_quote":"[40] is a hybrid CNN-transformer baseline that ResNet+ is compared against on LC25000."}],"review_version":1}