{"id":"d5a54428-772f-4298-8470-0219e31d44d7","arxiv_id":"2501.03836","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Inserting SCConv into YOLOv9 raises mAP50 by 0.3% on Br35H and 0.5% on a private brain tumor dataset, but no error bars, code, or data are provided.","lead":"A team added a module called SCConv to the YOLOv9 object detector and tested it on MRI brain tumor images. It reports a small improvement in detection accuracy, 0.3% on a public dataset and 0.5% on a private one, compared with the unmodified detector.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-run mAP50 gains of 0.003–0.005 are within likely run-to-run variance; without repeated seeds or error bars the central improvement claim is unsupported.","rationale":"The reader's verdict is CONDITIONAL, and the weakest assumption identified is exactly the one I consider load-bearing: a single training run per model cannot support a 0.003–0.005 mAP50 improvement. The paper's own tables (Table 3 and Table 4) show the gain is tiny, and the evaluation section provides no error bars or statistical tests. Br35H contains only 101 test images, so the metric itself is noisy; the custom dataset is unreleased and its labeling quality is unverified. My stress-test pass therefore confirms, rather than overturns, the reader's concern. I do not find an additional independent fatal flaw: the architectural description, while garbled, is consistent with inserting SCConv at layer 37, and the parameter counts (Table 5) are plausible for YOLOv9 plus an added module. The main issue is evidentiary, not logical: the reported numbers are compatible with a real small gain, but they are also compatible with chance variation. The requested verification — repeated seeds, variance reporting, release of code and data, and comparison against the cited state-of-the-art detectors — would settle this. Since the reader already conditioned acceptance on these requirements, the verdict should remain CONDITIONAL rather than being moved to ACCEPT or REJECT; the paper should not be rejected outright because the proposed module could still yield a genuine improvement, but it is not yet established. My recommendation is therefore UNCHANGED: the reader's conditional verdict is appropriate.","tokens_in":8071,"tokens_out":1461,"duration_ms":15652,"concrete_test":"Run SCC-YOLO and YOLOv9 (and ideally SE-YOLOv9) on Br35H with at least 5 different random seeds, keeping the same hyperparameters, data splits, and evaluation code. Report mean and standard deviation of mAP50, and compute a paired bootstrap or permutation test over per-image AP scores (or, at minimum, a t-test over per-run mAP50 values). If the 0.003 Br35H difference is within one standard deviation or the p-value exceeds 0.05, the claimed improvement is not statistically distinguishable from noise and the central claim fails. For the custom dataset, the same multi-seed test should be repeated, and the dataset should be released or replaced with a public benchmark to allow independent verification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that inserting SCConv into YOLOv9 improves detection accuracy: mAP50 rises from 0.954 to 0.957 on Br35H (Table 3) and from 0.855 to 0.860 on the custom Brain_Tumor_Dataset (Table 4) — improvements of 0.003 and 0.005, respectively. These differences are small enough that they may be explained by training stochasticity. The manuscript reports exactly one training run per model configuration (Section 3.1) and provides no variance estimates, confidence intervals, or significance tests. In object detection, run-to-run mAP50 variation from random weight initialization, data shuffling, and augmentation order is commonly on the order of 0.005–0.01 or larger for small test sets; Br35H has only 101 test images, making the 0.003 gain especially fragile. Additionally, the custom dataset is unreleased, was labeled with Labelimg by the authors, and no label verification or inter-annotator agreement is reported; the claim of 'state-of-the-art performance' is also not supported because the cited RCS-YOLO and BGF-YOLO baselines were not evaluated under the same protocol. The load-bearing assumption is therefore that a single-run 0.3–0.5% mAP50 difference is a real, reproducible improvement. If YOLOv9's run-to-run variance is comparable to the reported gain, the central contribution collapses to a null result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SCC-YOLO, a modification of YOLOv9 in which the SCConv module is inserted after the 37th layer, and evaluates it on the public Br35H dataset and a custom Brain_Tumor_Dataset. The authors report that SCC-YOLO improves mAP50 by 0.3% on Br35H (0.957 vs. 0.954) and by 0.5% on the custom dataset (0.860 vs. 0.855) compared with plain YOLOv9, and also compare with a variant incorporating the SE attention mechanism. They conclude that SCC-YOLO achieves state-of-the-art performance in brain tumor detection.","tokens_in":8368,"tokens_out":3400,"duration_ms":29625,"significance":"If the reported gains were statistically robust, the contribution would be a marginal, incremental improvement to a well-established detector, with no new algorithmic ideas: SCConv is a published module, and the paper tests a single insertion location. The paper does provide direct head-to-head comparisons under a fixed hyperparameter protocol and reports standard detection metrics, which is useful. However, the core quantitative claim rests on differences of 0.003 and 0.005 mAP50 from a single training run, which are within typical run-to-run variation for YOLO training, especially on a test set of only 101 images (Br35H). The lack of error bars, multiple seeds, or significance testing means the central claim is not yet supported.","major_comments":[{"comment":"The central claim that SCC-YOLO improves mAP50 by 0.003 on Br35H and 0.005 on the custom dataset is based on a single training run per configuration. No error bars, confidence intervals, or significance tests are reported. Given the small test set (101 images for Br35H) and the known stochasticity of YOLO training, these differences are within the range of run-to-run variation; the mAP50:95 differences (0.001 and 0.002) are even smaller. Please provide results from multiple random seeds, variance estimates, or a standard significance test to support the improvement claim.","section":"Section 4, Tables 3 and 4"},{"comment":"The custom Brain_Tumor_Dataset is not publicly released, was labeled by the authors using Labelimg, and no annotation verification or inter-annotator agreement is reported. The main result on this dataset therefore cannot be independently reproduced or audited. The authors should either release the dataset or provide a detailed annotation protocol, label statistics, and external validation to support the claim that the dataset is correctly labeled and representative.","section":"Section 2.1 and Section 4 (Brain_Tumor_Dataset)"},{"comment":"The claim that SCC-YOLO achieves 'state-of-the-art performance in brain tumor detection' is not supported by the experiments, because the cited state-of-the-art methods RCS-YOLO and BGF-YOLO were not evaluated under the same protocol on the same data splits. The comparison is only against YOLOv9 and an SE-augmented variant. Please either include the relevant baselines or temper the claim to a direct comparison with YOLOv9.","section":"Abstract, Section 4, and Conclusion"},{"comment":"The text states that 'with such improved accuracy, precision parameters increased,' but Table 3 shows that precision decreases from 0.926 (YOLOv9) to 0.922 (SCC-YOLO) on Br35H. This internal inconsistency must be corrected; if the authors intend to emphasize mAP50, the text should not claim a precision increase that the table contradicts.","section":"Section 4, third paragraph"}],"minor_comments":[{"comment":"The dataset resolution is given as 139x132 pixels in the Introduction and 133x132 in Section 2.1; please unify the numbers.","section":"Section 2.1 and Introduction"},{"comment":"The regression loss is listed as 'SSD' in Table 2; this appears to be a typo, as the paper does not otherwise mention the SSD loss. Please clarify which YOLOv9 loss function was actually used.","section":"Section 3.1, Table 2"},{"comment":"The captions of Figures 3 and 4 are informal ('The architecture of Spatial Reconstruction Unit' and 'The architecture of Channel Reconstruction Unit'); please make them informative and reference them in the text where the units are first described.","section":"Figures 3 and 4"},{"comment":"The English in Section 2.4 is very difficult to follow (e.g., 'Once the feature channel descriptor is obtained during the'squeeze' stage, it will then proceed through the the process of adaptively adjusting...'). A thorough language revision is needed throughout the manuscript.","section":"Section 2.4"},{"comment":"There are multiple typos and inconsistent terms, such as 'Intergration' in Section 2.3, 'Evironment' in Section 3.1, 'Gobal Average Pooling' in Section 2.4, and 'Preciseness' used instead of 'Precision' in Section 4; please correct these.","section":"General"},{"comment":"The manuscript does not mention code or model availability; please state whether the implementation will be released to support reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is very short and the incremental contribution is small; the main scientific issue is the lack of any uncertainty quantification for the central accuracy improvement. The reported gains are tiny and, without repeated seeds or release of the custom dataset, the result is not verifiable. I would encourage the editor to require the authors to provide multi-seed statistics and a public dataset or a very detailed annotation protocol before considering publication. The paper's scope and depth are closer to a workshop paper than a full journal article."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a YOLOv9 plus SCConv plug-in for brain tumor MRI detection, and the only reported gain over the plain detector is 0.003–0.005 mAP50 from a single run per model. That difference is smaller than typical run-to-run variance, so the central claim doesn't survive contact with the evidence as presented.\n\nWhat's worth crediting: the choice of baseline is sensible — they compare against plain YOLOv9 and an SE-attention variant, and the tables are consistent with the abstract. Inserting SCConv at layer 37 is a concrete, describable modification, and the architecture table gives parameters, layers, and gradients. If the improvement were real, it would be a marginally better detector for a specific medical imaging task.\n\nThe soft spots are more serious. No error bars, no repeated seeds, no significance test; Br35H has only 101 test images, so a 0.003 mAP50 shift is within the noise floor. The custom dataset is unreleased, labeled by the authors with no verification or inter-annotator agreement. The 'state-of-the-art' claim is unsupported because the cited RCS-YOLO and BGF-YOLO baselines are not evaluated under the same protocol. The writing is severely garbled in places — the method section is at times unintelligible, and the conclusion repeats itself. That makes it hard to verify exactly what was done.\n\nOn balance, the paper has a clear hypothesis, but the load-bearing evidence is a single run. This is fixable: run multiple seeds, report variance, release code and the custom dataset, and benchmark against the SOTA baselines they cite. As it stands, it reads like an incomplete engineering report rather than a result I would rely on.\n\nFor a reading group, it could be useful as a cautionary example of small-effect reporting without uncertainty. I wouldn't cite it. If I were the editor, I'd desk reject this version — the claim is too fragile and the missing resources too fundamental. If the authors resubmit with proper experimental rigor, it's worth a look.","headline":"Single-run mAP50 gains of 0.003–0.005 without error bars leave the central claim unverified; the paper is a thinly specified plug-in experiment, not yet a contribution.","tokens_in":8928,"tokens_out":2758,"would_cite":false,"duration_ms":24726,"reading_group":"maybe","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 reports that adding the SCConv module to YOLOv9 raises brain-tumor detection mAP50 by 0.3 percentage points on Br35H and 0.5 percentage points on a custom MRI dataset.","keywords":["brain tumor detection","YOLOv9","SCConv","object detection","MRI","mAP50","attention mechanism","medical imaging"],"falsifier":"Retrain YOLOv9, SE-YOLOv9, and SCC-YOLO on both datasets with several random seeds and compare the spread of mAP50 values; if the SCC-YOLO distribution overlaps the YOLOv9 distribution, the claimed improvement is not established.","tokens_in":7865,"feed_emoji":"🧠","tokens_out":6467,"duration_ms":55409,"temperature":0.7,"pith_summary":"This paper tries to establish that a small architectural plug-in, SCConv, improves the YOLOv9 object detector for brain-tumor detection in MRI. The authors report mAP50 rising from 0.954 to 0.957 on the public Br35H dataset and from 0.855 to 0.860 on their own Brain_Tumor_Dataset, and they interpret these gains as evidence that reducing spatial and channel redundancy helps tumor feature learning. They also compare against YOLOv9 with SE attention, which scores lower than plain YOLOv9, and conclude that their SCC-YOLO is state of the art for this task. A careful reader would care because the proposed change is a drop-in module, so even a small stable gain would transfer easily to other detection pipelines.","feed_headline":"SCConv nudges YOLOv9 brain-tumor detection up 0.3–0.5 points","feed_subtitle":"A plug-in redundancy-reduction module edges out plain YOLOv9 and SE attention on two MRI datasets.","key_machinery":"SCConv is the load-bearing component: a plug-and-play module composed of two blocks in series, the Spatial Reconstruction Unit (SRU) and the Channel Reconstruction Unit (CRU). The SRU separates and reconstructs feature maps to remove spatial redundancy; the CRU splits, transforms, and fuses channels to remove channel redundancy. Inserted at layer 37 of YOLOv9, it is meant to let the network learn from MRI features with fewer redundant computations, and the paper attributes the mAP50 improvements to this reduction in redundancy.","core_discovery":"On the paper's own terms, SCC-YOLO is YOLOv9 with the SCConv module inserted after the 37th layer of the network head. SCConv is a plug-and-play combination of a Spatial Reconstruction Unit (SRU) and a Channel Reconstruction Unit (CRU) that is designed to cut redundancy in intermediate feature maps. The reported outcome is a 0.003 mAP50 gain on Br35H (0.957 versus 0.954) and a 0.005 mAP50 gain on the custom dataset (0.860 versus 0.855), alongside mixed changes in precision and recall. The paper takes these numbers as evidence that SCConv extracts tumor-relevant features more efficiently than plain YOLOv9, and that the channel-only SE attention mechanism is comparatively unsuited to this medical detection task.","pith_inferences":["Editorial inference: a repeated-seed benchmark would determine whether margins of 0.003 to 0.005 in mAP50 are stable; the paper reports one run per configuration, so the practical size of the gain is not yet pinned down.","Editorial inference: if the gain replicates, the same layer-37 SCConv insertion could be tested at comparable depths in other YOLO versions and on other small-object medical detection tasks, since the module is designed to be plug-and-play.","Editorial inference: the state-of-the-art claim is stronger than what the experiments directly show, because the comparisons cover YOLOv9 and SE-YOLOv9 rather than the RCS-YOLO and BGF-YOLO detectors cited as prior work."],"forward_implications":["On Br35H, SCC-YOLO improves mAP50 from 0.954 to 0.957 and mAP50:95 from 0.751 to 0.752 while raising recall from 0.939 to 0.943.","On the custom dataset, SCC-YOLO improves mAP50 from 0.855 to 0.860 and mAP50:95 from 0.631 to 0.633, with precision and recall close to YOLOv9's.","Because SCConv is inserted at a single layer without changing training hyperparameters, the reported gain is a drop-in improvement rather than a retrained-from-scratch architecture.","The SE-attention variant scores below plain YOLOv9 on both datasets, which the paper uses to argue that channel-only attention is a worse fit than SCConv's combined spatial and channel reconstruction for brain-tumor images."],"supporting_citations":[{"why":"Supplies the base detector, YOLOv9 with Programmable Gradient Information, which SCC-YOLO modifies.","marker":"[19]"},{"why":"Supplies the inserted SCConv module, whose SRU and CRU units are the paper's main added mechanism.","marker":"[24]"},{"why":"Supplies the public Br35H MRI benchmark used for one of the two evaluations.","marker":"[23]"},{"why":"Defines the SE attention baseline the paper compares against, arguing channel-only attention underperforms.","marker":"[25]"},{"why":"Provides RCS-YOLO, the prior brain-tumor detector on Br35H that motivates the YOLO-based setting.","marker":"[15]"},{"why":"Provides BGF-YOLO, the YOLOv8-based prior detector cited as demonstrating YOLO feasibility for brain tumors.","marker":"[16]"}],"fun_headline_variants":["SCConv boost: 0.3–0.5% mAP for brain tumor detection","YOLOv9 + SCConv improves brain tumor mAP by 0.3–0.5","Plug-in SCConv nudges YOLOv9 brain tumor mAP up","SCC-YOLO: 0.3–0.5% mAP gain in brain tumor detection","Redundancy cut: SCConv boosts YOLOv9 tumor spotting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported comparison assumes that a single training run per model captures the true performance difference; a 0.003 to 0.005 mAP50 gap could be within normal run-to-run variation if not confirmed across seeds.","fun_headline_variants_meta":{"raw":{"variants":["SCConv boost: 0.3–0.5% mAP for brain tumor detection","YOLOv9 + SCConv improves brain tumor mAP by 0.3–0.5","Plug-in SCConv nudges YOLOv9 brain tumor mAP up","SCC-YOLO: 0.3–0.5% mAP gain in brain tumor detection","Redundancy cut: SCConv boosts YOLOv9 tumor spotting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000724,"raw_usage":{"total_tokens":3214,"prompt_tokens":877,"completion_tokens":2337,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":2220}},"tokens_in":493,"tokens_out":2337,"duration_ms":16561,"temperature":1.0,"reasoning_tokens":2220,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:45:45.334862+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain YOLOv9, SE-YOLOv9, and SCC-YOLO on both datasets with several random seeds and compare the spread of mAP50 values; if the SCC-YOLO distribution overlaps the YOLOv9 distribution, the claimed improvement is not established.","supporting_citations":[{"cited_title":"Hamada, Br35h::Brain tumor detection 2020, Kaggle, 2021, https://www.kaggle","cited_arxiv_id":null,"evidence_quote":"Supplies the public Br35H MRI benchmark used for one of the two evaluations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides RCS-YOLO, the prior brain-tumor detector on Br35H that motivates the YOLO-based setting."},{"cited_title":"F., & Phan, R","cited_arxiv_id":null,"evidence_quote":"Provides BGF-YOLO, the YOLOv8-based prior detector cited as demonstrating YOLO feasibility for brain tumors."}],"review_version":1}