{"id":"84ee9585-973c-492f-bfe6-fdcc25a1f93c","arxiv_id":"2502.04682","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A concatenated ConvNeXt-EfficientNet model reportedly classifies falcon endoscopy images into three health classes with 98% accuracy, but the evidence is a single unshared run.","lead":"This paper combines two established image-classification networks, ConvNeXt and EfficientNet, into one model that sorts falcon endoscopy images into 'normal,' 'liver disease,' or 'aspergillosis.' The authors report 98% test accuracy, but the dataset and code are not released and the comparison has no error bars, so the result cannot be independently checked.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 98% claim rests on an unseeded, image-level 80/20 split of 610 images; without falcon-level separation, near-duplicate frames from the same bird can leak across train/test and inflate the reported accuracy.","rationale":"The reader's REJECT is the right call. My stress-test did not find a reason to raise or lower it; it found the same load-bearing weakness. The strongest claim is a comparative accuracy claim, and the only evidence is one unseeded image-level split of 610 images without bird-level grouping. This makes the failure mode concrete: endoscopic videos naturally yield multiple frames per falcon, so train and test sets can share the same individual, and the network could exploit bird-specific visual signatures. The paper neither reports bird counts nor offers a corrected split. Until the authors demonstrate falcon-level generalization or release code/data enabling an independent check, the 98% accuracy and superiority claim is not supported. No additional major concern is needed; the internal repeated paragraph and duplicate reference are signs of sloppiness but do not alter the verdict.","tokens_in":5441,"tokens_out":3878,"duration_ms":40367,"concrete_test":"Request the falcon dataset with per-image falcon (bird) identifiers and split by falcon, e.g., group k-fold with falcon as the group, so all images from one bird are either in training or test but never both. Retrain the concatenated model under the paper's reported settings and compare the falcon-level test accuracy against the reported 98%. Also rerun the original random split 10 times with different seeds and report mean +- SD for the proposed, ConvNeXt, and EfficientNet models. If the falcon-level split drops accuracy materially (say, more than 3-5 points) or removes the margin over EfficientNet, leakage is confirmed. If bird IDs are unavailable, compute image similarity (e.g., nearest-neighbor Euclidean distance in the penultimate-layer embedding) between train and test images; near-duplicate cross-split pairs would corroborate leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim is that the concatenated ConvNeXt-EfficientNet model 'achieves testing accuracies of 98%, surpassing other deep learning models considered in this study' (Section V), supported by Table I. The dataset is 610 labeled endoscopy images, split 80/20 randomly once (Sections III-IV). What has to be true for the claim to be evidence is that the test set contains only images independent of the training set, in particular no endoscopy frames from the same falcon appearing in both splits. The paper never reports the number of individual falcons, whether multiple frames per falcon are present, or any grouping in the split. In endoscopic image collections, multiple frames from one bird are common; an image-level random split then allows the model to memorize bird-specific or session-specific artifacts (illumination, scope angle, background, lesion texture) rather than disease-general features. That would inflate accuracy for every model, and could easily create a spurious margin for the concatenated model. The lack of repeated splits, cross-validation, or confidence intervals means the 98% figure has no variance estimate; a single 126-image test set cannot support a superiority claim. The text also calls the 20% split both 'internal validation' and 'test set', so the reported test set may have been used for hyperparameter tuning. The repeated paragraph in Section IV and the duplicate references [1]/[2] are presentation defects, but the split design is the load-bearing correctness risk.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a hybrid deep learning model that concatenates feature representations from ConvNeXt and EfficientNet to classify falcon endoscopic images into three classes ('Normal', 'Liver', 'Aspergillosis'). The authors use 610 images from a falcon clinic, with an 80/20 train/test split, and report that the concatenated model reaches about 98% test accuracy, compared to 87% for ConvNeXt and 94% for EfficientNet. The paper includes confusion matrices, a ROC curve, and per-class precision/recall/F1 metrics. The central claim is that the concatenated model is superior to the standalone architectures and to 'traditional methods', though no non-deep baseline is actually evaluated.","tokens_in":5772,"tokens_out":5796,"duration_ms":54467,"significance":"If the reported accuracy were robust, the work would provide a practical baseline for AI-assisted avian endoscopy, and the concatenation approach is simple enough to be adopted easily. The paper has some virtues: it uses real clinical images, reports class-specific metrics, and gives a pseudocode-level description of the pipeline. However, the evidence as presented is not yet sufficient to establish the superiority claim. The lack of falcon-level separation in the random split, the conflation of validation and test sets, and the absence of any variance estimate are serious threats to the validity of the headline number. These are standard and fixable evaluation issues, but they must be addressed before the claim can be accepted.","major_comments":[{"comment":"The paper states that the dataset was 'randomly divided' into 80% training and 20% test, but it does not report the number of individual falcons or whether images from the same falcon can appear in both partitions. Endoscopy acquisitions commonly contain multiple frames per bird; if the split is image-level, the 98% test accuracy may be inflated by bird-specific or session-specific artifacts. Please report the falcon-level distribution and repeat the evaluation with a split that groups all frames from each falcon into one partition, e.g., leave-one-falcon-out cross-validation.","section":"Section IV, dataset split"},{"comment":"The text says the remaining 20% was used for 'internal validation' and that 'hyperparameter fine-tuning was executed using the validation set,' yet it also refers to the same portion as the 'test set' and reports '126 test images' in the confusion matrix. If hyperparameters were chosen on the test set, the reported accuracy is an optimistic selection result, not an independent test statistic. Please clarify whether the reported numbers come from a fully held-out test set and, if not, provide an independent test evaluation.","section":"Section IV, validation/test terminology"},{"comment":"The superiority claim rests on a single 80/20 split with no repeated runs, confidence intervals, or cross-validation. With only 126 test images, the difference between the reported 98% and 94% accuracies corresponds to a handful of examples, and without variance estimates one cannot judge whether it is significant. Please report per-class confidence intervals and results over multiple seeds or k-fold cross-validation.","section":"Table I and Section V, statistical support"},{"comment":"The proposed 'concatenated' model is described only by the pseudocode bullet 'construct the concatenated deep learning model incorporating features from ConvNeXt and EfficientNet.' The fusion layer design, feature dimensions, optimizer, learning rate, data augmentation, and class-balance handling are not reported. Without these details, the 98% result cannot be reproduced or attributed to the concatenation rather than to unspecified training choices. Please provide a precise architecture description and a complete training protocol.","section":"Section III and Section IV, architecture and training details"}],"minor_comments":[{"comment":"References [1] and [2] are identical; please replace the duplicate with the actual reference for the deep-learning falcon-disease paper discussed in Section II.","section":"References"},{"comment":"Section IV contains the same paragraph twice, starting with 'To evaluate the model's performance...' – please remove the duplicate.","section":"Section IV"},{"comment":"Section IV refers to 'Figure 7x' for the dataset distribution; the correct figure is Fig. 3.","section":"Section IV"},{"comment":"The abstract and conclusion claim that the concatenated model outperforms 'traditional methods', but Table I only compares against the two standalone CNN models; please either add a traditional baseline or soften the claim.","section":"Abstract and Conclusion"},{"comment":"The description of ConvNeXt as 'developed by Google Research' is inaccurate; it should be credited to Liu et al. (Meta AI) and cited properly.","section":"Section II, ConvNeXt description"},{"comment":"Figure 2 shows the concatenated architecture but the text never explains the fusion layer or the dimensions of the concatenated feature vector; please add an explicit description.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the paper is far below the standard of a rigorous journal article in its current form. The recommendation of major_revision reflects the fact that the evaluation flaws are, in principle, addressable if the authors have bird-level metadata. If they cannot provide a falcon-level split or an independent test set, I would advise that the paper be rejected rather than go through a lengthy revision. The duplicated references and repeated paragraph also suggest that the manuscript needs careful editorial rewriting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a thin empirical paper whose central accuracy claim is not backed by the evaluation design. The core problem is the split.\n\nWhat's actually here: a real veterinary problem, a genuine dataset from Sharjah Falcon Clinic, and a straightforward head-to-head comparison of a feature-concatenated ConvNeXt+EfficientNet against both backbones. That is worth something. The authors do not falsify numbers; the reported metrics are internally consistent, and the method is a reasonable thing to try.\n\nBut the evidence is too weak to carry the \"superior performance\" conclusion. The dataset is 610 endoscopy images, split 80/20 at the image level, and the paper never says how many individual falcons contributed images. If frames from the same bird appear in both training and test, the model can memorize per-bird artifacts (illumination, scope angle, lesion texture) and the 98% figure is inflated. The text also calls the 20% split both \"internal validation\" and \"test set\" and says hyperparameter tuning was done on the validation set, so the final test numbers may not be independent. There are no repeated splits, no confidence intervals, no external validation. The literature review is weak, and reference [2] is a duplicate of [1]. Section IV contains a repeated paragraph.\n\nNone of this means the method is nonsense. It is a standard transfer-learning pipeline with a generic feature-concatenation trick. The flaws are fixable: report the number of birds, do a falcon-level split, run multiple seeds, and show error bars. If the authors can do that, there is a usable veterinary screening tool hiding here. As it stands, the paper is not publishable.\n\nMy recommendation: send it to peer review only if the venue demands those revisions. The current submission would not survive, but the underlying experiment is real and the topic is underserved. I would not cite it in its current form, and I would not bring it to reading group except as an example of image-level leakage in medical imaging. If you do review it, ask for bird-level cross-validation and insist on code and data release.","headline":"The 98% claim rests on an unseeded image-level split of 610 images with no bird-level separation; the approach is standard, the evaluation is not yet rigorous enough to support the claim.","tokens_in":690,"tokens_out":661,"would_cite":false,"duration_ms":26943,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a concatenation of ConvNeXt and EfficientNet feature representations classifies falcon endoscopy images into Normal, Liver, and Aspergillosis with 98% average test accuracy, beating both standalone backbones.","keywords":["falcon disease classification","ConvNeXt","EfficientNet","concatenated model","transfer learning","endoscopy images","avian veterinary AI","aspergillosis detection"],"falsifier":"Count, for each individual falcon in the collected endoscopy images, how many images contribute to the training set and how many to the test set. If any bird contributes images to both sides, retrain the same model with a bird-exclusive split; if the test accuracy falls materially below 98%, the reported superiority of the concatenated model is an artifact of image-level leakage rather than a genuine diagnostic gain.","tokens_in":5278,"feed_emoji":"🦅","tokens_out":8437,"duration_ms":71142,"temperature":0.7,"pith_summary":"This paper claims that joining two pretrained convolutional networks, ConvNeXt and EfficientNet, into one model accurately classifies falcon endoscopy images into three health categories: normal, liver disease, and aspergillosis. On a dataset of 610 images, the concatenated model reaches 98% average test accuracy, roughly 11 points above ConvNeXt alone and 4 points above EfficientNet alone. The authors argue that the two architectures extract complementary features, so combining them captures a wider range of diagnostic cues than either network on its own. If the result holds, it offers a practical recipe for veterinary disease screening from endoscopic images with a relatively small labeled dataset.","feed_headline":"Two-model fusion hits 98% for falcon disease classification","feed_subtitle":"Merging ConvNeXt and EfficientNet beats each network alone on a three-class endoscopy task.","key_machinery":"The central object is the concatenated AI model: the feature representations produced by ConvNeXt, a modern convolutional backbone built around depthwise-separable convolutions, and EfficientNet, a backbone whose compound scaling balances depth, width, and resolution, are joined before a final three-class classifier. The input layer is fixed to 128 by 128 endoscopy images, and transfer learning initializes both branches from pretrained weights before fine-tuning on the falcon dataset. The concatenation step is what carries the argument: it allows the classifier to see both backbones' representations at once, and the paper attributes the accuracy gain to this combined feature set.","core_discovery":"The central claim is that a model formed by concatenating the feature representations of ConvNeXt and EfficientNet classifies falcon endoscopy images into Normal, Liver, and Aspergillosis with 98% average test accuracy, surpassing the 87% of ConvNeXt and 94% of EfficientNet. The paper reports per-class F1 scores of 0.96, 0.99, and 0.98 for the three classes, with a training accuracy of 99.65% and a validation accuracy of 98.50% over 50 epochs. The authors interpret the result as evidence that the two backbones capture complementary image features, and that the concatenated model is more reliable for this diagnostic task than either model alone.","pith_inferences":["If the accuracy gain is real, the same concatenation recipe is a cheap thing to try on any small medical-imaging dataset with two pretrained convolutional backbones; the paper's numbers suggest the benefit may transfer without task-specific tuning.","The natural stress test is a bird-exclusive split: the authors could report performance when no individual falcon contributes images to both training and test, which would separate genuine generalization from image-level memorization.","Because the paper compares only against its own two backbone models on one split, it leaves open whether the 98% figure would survive comparison with published bird-disease classifiers or with a single-backbone model trained with the same data augmentation and tuning."],"forward_implications":["On the reported dataset, the concatenated model attains 98% average test accuracy, so a veterinarian screening endoscopy images could expect roughly 98 of every 100 images to be labeled correctly.","The hybrid beats both standalone backbones by a clear margin, 98% versus 87% and 94% average accuracy, so combining features is more useful on this task than either architecture alone.","Per-class results, with F1 scores of 0.96 for normal, 0.99 for liver, and 0.98 for aspergillosis, imply the model is not simply memorizing the majority class and does separate the two disease classes from normal tissue and from each other.","Since training used 80% of only 610 images with transfer learning, the result suggests pretrained backbones can be adapted to a niche veterinary domain with a few hundred labeled examples."],"supporting_citations":[],"fun_headline_variants":["Fused AI nets hit 98% on falcon disease diagnosis","ConvNeXt+EfficientNet fusion tops 98% for falcon illnesses","AI fusion classifies falcon liver, aspergillosis at 98%","Falcon disease AI: concatenated models reach 98% accuracy","Two-model AI beats singles for falcon health checks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the single random 80/20 split of 610 images gives an honest test set, with no images of the same falcon appearing in both training and test; the paper does not report bird-level separation or cross-validation.","fun_headline_variants_meta":{"raw":{"variants":["Fused AI nets hit 98% on falcon disease diagnosis","ConvNeXt+EfficientNet fusion tops 98% for falcon illnesses","AI fusion classifies falcon liver, aspergillosis at 98%","Falcon disease AI: concatenated models reach 98% accuracy","Two-model AI beats singles for falcon health checks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000904,"raw_usage":{"total_tokens":3831,"prompt_tokens":832,"completion_tokens":2999,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":448,"completion_tokens_details":{"reasoning_tokens":2905}},"tokens_in":448,"tokens_out":2999,"duration_ms":18803,"temperature":1.0,"reasoning_tokens":2905,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T21:52:06.171945+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count, for each individual falcon in the collected endoscopy images, how many images contribute to the training set and how many to the test set. If any bird contributes images to both sides, retrain the same model with a bird-exclusive split; if the test accuracy falls materially below 98%, the reported superiority of the concatenated model is an artifact of image-level leakage rather than a genuine diagnostic gain.","supporting_citations":[],"review_version":1}