{"id":"421e69a4-1069-4edb-8003-4614d1c6cccb","arxiv_id":"2507.07011","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Combining EfficientNetB0 and ResNet50 with transfer learning yields an 88 percent accuracy claim for four-class brain tumor MRI classification, but the reported metrics are internally inconsistent and the model is not reproducible from the paper.","lead":"This paper reports a deep learning model that combines two standard image recognition networks to classify brain tumors from MRI scans. The authors claim 88 percent accuracy, but the paper contains conflicting result numbers and omits code and test details.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Final dataset in §III-A is self-contradictory: if only figshare images remain, the four-class no-tumor results in Table III cannot exist; this blocks the central accuracy claim.","rationale":"The reader's weakest-assumption analysis correctly identifies the absence of a defined held-out test set and early stopping on validation loss as serious threats to generalization. I agree those are real problems. However, the more fundamental and more decisive flaw is in the dataset description itself: §III-A cannot simultaneously be true that the final dataset consists only of figshare images and that the four-class evaluation includes a no-tumor class taken from Br35H. Since the central claim is about four-class accuracy, an internally consistent description of the input data is a precondition; without it, no amount of code or metric recomputation can anchor the 88% figure to a specific public dataset. The proposed check therefore targets the dataset composition and the internal consistency of the reported metrics. If the authors can provide the exact per-class counts and the confusion matrix, and those counts include a nonzero no-tumor class while the confusion-matrix accuracy equals 88%, then this particular objection would be resolved; the held-out test-set concern would still remain. Since the reader already rejected the paper on closely related grounds, my verdict remains REJECT, but the justification should center on the unresolved dataset contradiction as well as the evaluation-protocol problem.","tokens_in":8949,"tokens_out":7364,"duration_ms":80327,"concrete_test":"Request the exact final dataset composition and the confusion matrix underlying Figure 11, or reconstruct both from the stated sources after applying the stated removals; then recompute overall accuracy as trace(CM)/total validation samples and the weighted F1 from Table III using the per-class validation counts. If the no-tumor count is zero, or if the recomputed accuracy does not equal the reported 88% while also reconciling the 93.2% validation accuracy, the four-class central claim is unsupported by the described data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an 88% four-class accuracy on a public brain MRI dataset. That claim requires a dataset containing all four classes and a defined evaluation split. §III-A undermines the dataset condition: it first says the study uses 7023 images from figshare, SARTAJ, and Br35H, with no-tumor images taken from Br35H; then it says SARTAJ images were removed and 'The final dataset used for this study consists of images from figshare.' In the paper's own description, the figshare collection contributes the three tumor classes and Br35H supplies no-tumor, so a figshare-only dataset cannot produce the no-tumor row in Table III or the four-class confusion matrix in Figure 11. If Br35H no-tumor images were in fact retained, then the final dataset is not figshare-only and the true total and per-class counts are unspecified. Either reading leaves the input to the experiment not uniquely determined, so the reported 88% cannot be mapped to a reproducible dataset. The reader's concern about the missing held-out test set is valid and compounds this: Section V reports validation accuracy values of 93.2% and 94.7% as well as an overall accuracy of 88%, and the paper never defines a test split independent of the early-stopping criterion. No code or trained weights are provided to resolve any of these ambiguities. The most load-bearing gap is therefore not the architecture but the absence of an internally consistent definition of the exact four-class dataset and split that produced the headline numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DeepBrainNet, a hybrid deep learning model combining EfficientNetB0 and ResNet50 with transfer learning for four-class brain tumor classification from MRI images. The authors report an overall accuracy of 88%, a weighted F1-score of 88.75%, and a macro AUC-ROC of 98.17% on a publicly available dataset, and claim that this consistently outperforms existing state-of-the-art methods. The manuscript describes dataset collection, preprocessing, fuzzy C-means feature selection, model architecture, training details, and a comparative analysis.","tokens_in":9259,"tokens_out":2824,"duration_ms":32683,"significance":"If the reported results were reproducible and properly validated, the model would be a modest but potentially useful engineering contribution to automated brain tumor classification. The paper's strengths are that it uses publicly available data, provides training details, includes a confusion matrix and ROC curves, and compares with two prior systems. However, the contribution is incremental—a standard late-fusion ensemble of two pretrained CNNs—and the experimental validation as presented is not rigorous enough to support the central claims of consistent superiority and clinical potential.","major_comments":[{"comment":"The dataset description is internally contradictory. The section first states that the study uses 7023 images from figshare, SARTAJ, and Br35H, with no-tumor images taken from Br35H. It then states that SARTAJ images were removed and 'The final dataset used for this study consists of images from figshare.' If the figshare collection contains only the three tumor classes (as is standard for this dataset), then a figshare-only dataset cannot yield the no-tumor results reported in Table III and the four-class confusion matrix in Figure 11. If Br35H no-tumor images were retained, then the final dataset is not figshare-only and the exact per-class counts and total are not specified. Either way, the input dataset for the experiments is not uniquely defined, which makes the reported 88% accuracy unreproducible.","section":"§III-A (Dataset)"},{"comment":"The central accuracy claim is reported inconsistently and without a defined evaluation split. The Abstract states 88% accuracy, Table IV lists 89% for DeepBrainNet, and Section V reports a validation accuracy of 93.2% (and later 94.7%) for the same model. The manuscript never defines a held-out test set; the models are selected by early stopping on validation loss, so the reported metrics appear to be validation-set metrics that are optimistically biased. Without a clearly separated test partition, the claim that the model 'generalizes well to unseen data' is unsupported.","section":"§V (Results) and Abstract/Table IV"},{"comment":"Fuzzy C-means (FCM) feature selection is described in detail and Algorithm 1 is provided, but FCM never appears in the results section, no ablation study is performed, and there is no explanation of how the selected features are integrated into the classification pipeline. As written, the feature-selection component has no observable effect on the empirical results, so its role in the claimed performance is unclear.","section":"§III-D (Feature Selection)"},{"comment":"The comparison in Table IV is not adequately grounded. The table claims 'same dataset used in this study,' but given the ambiguity in §III-A, it is unclear which dataset and split were used for the cited baselines [14] and [15]. No details are provided on the preprocessing, augmentation, or evaluation protocol for the baselines, and no error bars or multiple-run statistics are given. The claim that DeepBrainNet 'consistently outperforms' prior work is therefore not supported by the reported evidence.","section":"Table IV (Comparative Analysis)"}],"minor_comments":[{"comment":"The description of EfficientNetB0 incorrectly states that it combines EfficientNet and VNet for segmentation; this paper addresses classification, and the quoted text appears to have been reused from a segmentation study.","section":"§IV-A"},{"comment":"The variables in the resizing equation are inconsistently named: 'Iresized = f(I original, Dnew)' but the definition list mentions 'Original image' rather than 'I original'.","section":"Equation (1)"},{"comment":"The membership update equation in Algorithm 1 has a typesetting error: 'uij = db(xi, vj)Pc k=1 db(xi, vk)' is missing the division symbol, making the formula unreadable.","section":"Algorithm 1"},{"comment":"Reference [4] cites 'arXiv:2304.100394', which appears to contain an extra digit; the correct ID appears to be arXiv:2304.10039.","section":"Reference [4]"},{"comment":"The bar chart in Figure 3 is not annotated with axis labels or class counts, making it difficult to verify the class distribution described in the text.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as an incomplete engineering report. The most serious problems—the self-contradictory dataset description and the absence of a defined test split—are fixable in principle, but they require re-running the experiments with a clearly specified protocol and reporting consistent metrics. The authors should also verify that the described FCM component is actually used in the experiments, or remove it from the methodology. I recommend that the editor require a revised manuscript that addresses these issues before any further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a standard transfer-learning application with an unsupported headline. The architecture is a parallel concatenation of EfficientNetB0 and ResNet50; the paper's own related-work table lists similar ensembles (AlexNet+VGG19, ResNet50+DenseNet201+InceptionV3), so novelty is incremental at best. To its credit, the manuscript is readable and documents preprocessing (resizing, normalization, augmentation, cropping), defines the evaluation metrics, and presents per-class results rather than only a single accuracy number. The FCM feature-selection algorithm is given, although the results section never says whether FCM was actually used in the final pipeline.\n\nThe soft spots are not minor. The dataset description in §III-A is internally contradictory: it says the final dataset is 'images from figshare' after removing SARTAJ, but the no-tumor class was taken from Br35H. A figshare-only dataset cannot produce the no-tumor row in Table III or the confusion matrix in Figure 11. If Br35H images were retained, the true dataset is not figshare-only and the per-class counts are unstated. Either way, the input to the experiment is not reproducible.\n\nThe numbers are also inconsistent: 88% in the abstract, 89% in Table IV, 93.2% validation accuracy in Section V, 94.7% 'best validation accuracy' elsewhere. There is no defined test split. Early stopping is based on validation loss, and the reported metrics appear to be validation-set metrics, so they are optimistically biased. No code, weights, or seed are provided. The claim that DeepBrainNet 'consistently outperforms' SOTA is contradicted by the paper's own Table I, which lists accuracies of 98.7–99.39%. Finally, Section IV-A describes EfficientNetB0 as if it were a combined EfficientNet/VNet segmentation model, which is a different architecture from what was trained; the fusion of the two backbones is not described.\n\nWho gets value from this? Maybe a reader who wants a template for a simple two-backbone fine-tuning script. The paper does not advance the state of the art, and its central claim cannot be mapped to a defined dataset and split. I would desk reject this in its current form. If the authors resubmit with a consistent dataset description, a clear held-out test set, and code, it could be a minor application note, but the current manuscript is not reproducible enough for peer review.","headline":"A standard two-backbone transfer-learning application whose headline accuracy is unsupported by an internally contradictory dataset description and inconsistent metric reporting.","tokens_in":9790,"tokens_out":2184,"would_cite":false,"duration_ms":23449,"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":"DeepBrainNet fuses EfficientNetB0 and ResNet50 to classify four brain MRI tumor categories, claiming 88% accuracy, 88.75% weighted F1, and a 98.17% macro AUC.","keywords":["DeepBrainNet","brain tumor detection","MRI classification","EfficientNetB0","ResNet50","transfer learning","depthwise separable convolutions","fuzzy C-means"],"falsifier":"Re-run the architecture on the same 7,023-image data with a strictly disjoint test split and report test accuracy and weighted F1; if test accuracy falls below 88% by more than split sampling error, the central claim is not supported. A cheaper control is an ablation training EfficientNetB0 and ResNet50 alone under identical settings: the hybrid must beat its better single branch, otherwise the fusion is not what carries the result.","tokens_in":8741,"feed_emoji":"🧠","tokens_out":8877,"duration_ms":94012,"temperature":0.7,"pith_summary":"This paper is an engineering claim: a hybrid network called DeepBrainNet, built by running EfficientNetB0 and ResNet50 in parallel and fine-tuning both on a public brain MRI dataset, can sort scans into glioma, meningioma, pituitary tumor, and no tumor at 88% accuracy. It reports a weighted F1 of 88.75% and a macro AUC-ROC of 98.17%, which it reads as evidence that the model separates tumor types reliably while staying computationally cheap. Transfer learning from ImageNet supplies the starting weights, depthwise separable convolutions in the EfficientNetB0 branch cut parameters, and residual connections in ResNet50 keep the deeper branch trainable. The stated purpose is a practical decision-support tool for radiologists, not a new biological mechanism.","feed_headline":"88% accuracy reported for four-way brain tumor MRI model","feed_subtitle":"Fusing EfficientNetB0 and ResNet50 with transfer learning beats single-backbone baselines across four MRI classes.","key_machinery":"The machine doing the work is a parallel hybrid: an EfficientNetB0 branch, whose mobile inverted bottleneck (MBConv) blocks use depthwise separable convolutions to cut parameters and computational cost, joined with a ResNet50 branch that is initialized from ImageNet weights and fine-tuned, so residual connections keep gradients flowing during transfer learning. Around this sit a preprocessing chain (resizing, CLAHE contrast enhancement, histogram equalization, margin cropping), fuzzy C-means feature selection that retains features with high membership across clusters, and training-time augmentation with rotations, flips, zooms, shears, brightness changes, and shifts. Early stopping on validation loss and dynamic learning-rate reduction stabilize training.","core_discovery":"The central claim of the paper is that a single deep network can get the best of both efficiency and depth by fusing EfficientNetB0's mobile inverted bottleneck blocks with a ResNet50 branch that carries over ImageNet features. Trained with on-the-fly augmentation, dropout, early stopping, and fuzzy C-means feature selection, the hybrid reports 88% overall accuracy on a 7,023-image four-class MRI dataset. The class-level report shows the strongest separation for the no-tumor class (precision 0.946) and the weakest for meningioma (F1 0.808). The paper presents these numbers as outperforming prior deep-learning classifiers on the same benchmark and as evidence that an architecture tuned for efficiency on natural images transfers well to medical imaging.","pith_inferences":["Because no separate held-out test set is described, the 88% figure should be read as a validation-based estimate; a clean test split may land lower.","The paper never compares the hybrid against its own two backbones run alone, so the claimed fusion advantage is not yet isolated; an ablation with identical training would settle it.","Tumor and no-tumor images come from different source repositories, so some of the discriminative signal may be source-specific rather than tumor-specific; testing on an external dataset with matched acquisition is the cleanest check.","The comparison table's baselines [14] and [15] have no corresponding entries in the reference list, so the claim that DeepBrainNet beats those systems cannot be independently verified from this manuscript alone."],"forward_implications":["If the reported accuracy holds on a clean test set, DeepBrainNet could serve as a first-pass four-class triage tool in settings where an MRI radiologist is not immediately available.","A macro AUC near 0.98 means the model ranks classes well even where hard labels are missed, so threshold tuning could trade sensitivity against specificity per tumor type.","The roughly 24 million trainable parameters and the EfficientNetB0 branch's depthwise separable convolutions keep the model light enough for deployment on clinical workstations rather than large GPU clusters.","Per-class results point to meningioma as the most error-prone category, so future data collection or class weighting should target that class first."],"supporting_citations":[{"why":"Defines the EfficientNet architecture and supplies the MBConv and depthwise separable convolution backbone used in the hybrid's first branch.","marker":"[11]"},{"why":"Compares transfer-learning methods for brain tumor detection and motivates fine-tuning a pretrained network on MRI data.","marker":"[3]"},{"why":"Employs deep learning and transfer learning for accurate brain tumor detection; frames the transfer-learning approach and supplies tumor-class illustrations.","marker":"[2]"},{"why":"EfficientNetB1 plus U-Net baseline for tumor classification and segmentation; a comparison point the paper argues it improves on.","marker":"[4]"},{"why":"Custom CNN baseline reporting 99% tumor classification accuracy; used as a comparison point.","marker":"[5]"},{"why":"Stacked ensemble of AlexNet and VGG19 baseline reporting 98.70% accuracy; used as a related-work comparison.","marker":"[6]"},{"why":"Introduces CLAHE, the contrast-enhancement preprocessing step applied before training.","marker":"[7]"},{"why":"Provides the fuzzy C-means algorithm used for feature selection in the pipeline.","marker":"[9]"},{"why":"Baseline accuracy score of 88.47% in Table IV that DeepBrainNet claims to beat.","marker":"[14]"},{"why":"Baseline accuracy score of 85.30% in Table IV that DeepBrainNet claims to beat.","marker":"[15]"}],"fun_headline_variants":["Fused CNN nets hit 88% on brain tumor MRI","Hybrid EfficientNet-ResNet50 brain tumor detector: 88%","MRI tumor AI: 88% accuracy via net fusion","Deep Brain Net: 88% on brain MRI with dual backbones"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on treating the validation set, which was also used for early stopping and model selection, as a measure of generalization; if that set is not a cleanly separated test set, the 88% accuracy figure is not a supported estimate of performance on new scans.","fun_headline_variants_meta":{"raw":{"variants":["Fused CNN nets hit 88% on brain tumor MRI","Hybrid EfficientNet-ResNet50 brain tumor detector: 88%","MRI tumor AI: 88% accuracy via net fusion","Deep Brain Net: 88% on brain MRI with dual backbones"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1358,"prompt_tokens":981,"completion_tokens":377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":303}},"tokens_in":597,"tokens_out":377,"duration_ms":4310,"temperature":1.0,"reasoning_tokens":303,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:48:55.397479+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the architecture on the same 7,023-image data with a strictly disjoint test split and report test accuracy and weighted F1; if test accuracy falls below 88% by more than split sampling error, the central claim is not supported. A cheaper control is an ablation training EfficientNetB0 and ResNet50 alone under identical settings: the hybrid must beat its better single branch, otherwise the fusion is not what carries the result.","supporting_citations":[{"cited_title":"Efficientnet: Rethinking model scaling for convolutional neu- ral networks","cited_arxiv_id":null,"evidence_quote":"Defines the EfficientNet architecture and supplies the MBConv and depthwise separable convolution backbone used in the hybrid's first branch."},{"cited_title":"Brain tumour detection from images and compar- ison with transfer learning methods and 3- layer CNN","cited_arxiv_id":null,"evidence_quote":"Compares transfer-learning methods for brain tumor detection and motivates fine-tuning a pretrained network on MRI data."},{"cited_title":"Employing deep learn- ing and transfer learning for accurate brain tumour detection","cited_arxiv_id":null,"evidence_quote":"Employs deep learning and transfer learning for accurate brain tumor detection; frames the transfer-learning approach and supplies tumor-class illustrations."},{"cited_title":"Brain tumor multi classification and segmentation in MRI images using deep learning","cited_arxiv_id":"2304.10039","evidence_quote":"EfficientNetB1 plus U-Net baseline for tumor classification and segmentation; a comparison point the paper argues it improves on."},{"cited_title":"Enhancing brain tumour classification in MRI scans with a multi-layer customized convolutional neural network ap- proach","cited_arxiv_id":null,"evidence_quote":"Custom CNN baseline reporting 99% tumor classification accuracy; used as a comparison point."},{"cited_title":"Automated brain tumour iden- tification in biomedical radiology images: A multi-model ensemble deep learning approach","cited_arxiv_id":null,"evidence_quote":"Stacked ensemble of AlexNet and VGG19 baseline reporting 98.70% accuracy; used as a related-work comparison."},{"cited_title":"Realization of the contrast limited adaptive histogram equalization (CLAHE) for real-time image enhancement","cited_arxiv_id":null,"evidence_quote":"Introduces CLAHE, the contrast-enhancement preprocessing step applied before training."},{"cited_title":"Comparison of Fuzzy C-Means, Fuzzy Kernel C-Means, and Fuzzy Kernel Robust C-Means to Classify Thalassemia Data","cited_arxiv_id":null,"evidence_quote":"Provides the fuzzy C-means algorithm used for feature selection in the pipeline."}],"review_version":1}