{"id":"89d181a5-4399-45be-b03e-95ade4894969","arxiv_id":"2506.14816","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper applies the authors' existing concatenated ConvNeXt-EfficientNet architecture to falcon endoscopy images, reporting 98% average accuracy, though the evaluation conflates validation and test data.","lead":"This paper reports a hybrid AI model that sorts falcon endoscopy images into normal, liver disease, or aspergillosis, with 98% accuracy on a 610-image dataset. A generalist might read it as a test of whether off-the-shelf neural networks can screen birds cheaply, but the evidence behind the number is currently hard to verify.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 98% test accuracy is not credible because the 20% used for hyperparameter tuning appears to be the same 20% used for reporting test metrics; no separate held-out test set is described.","rationale":"The central claim is empirical: the concatenated model achieves 98% test accuracy, beating the two baselines. For that claim to hold, the 98% must be measured on a set not used for model selection. The paper conflates validation and test: Section IV explicitly uses the remaining 20% for internal validation and for hyperparameter fine-tuning, while Section III-A's pseudocode treats the same 20% as the testing set. No separate, untouched test set is described anywhere in the manuscript. If the same 20% was used to tune hyperparameters and to report the headline accuracy, the figure is an optimistically biased estimate and the comparison with the baselines is not an apples-to-apples test. The additional inconsistency between the stated dataset size (610 images, 20% = 122) and the reported test-set size (126 images) in Section IV-B reinforces that the paper's data accounting is unreliable. The absence of code, data, and repeated-seed statistics prevents any independent verification. These issues directly undermine the paper's central claim, so the REJECT verdict is appropriate and unchanged.","tokens_in":6166,"tokens_out":5215,"duration_ms":56049,"concrete_test":"Obtain the exact train/validation/test split used by the authors (or re-split the 610-image dataset into a stratified 60/20/20 split), retrain all three models with hyperparameter tuning performed only on the validation set, and report test-set metrics on the untouched 20%. If the proposed model's test accuracy falls below 98%, or below the EfficientNet baseline, the central claim is refuted. As a secondary check, verify the test-set size: 20% of 610 images is 122, not 126; any run reporting 126 must specify how the four additional images were obtained.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV states \"the remaining 20% for internal validation\" and \"hyperparameter fine-tuning was executed using the validation set,\" while the pseudocode in Section III-A defines only a training (80%) and testing (20%) split. No separate test set is ever described. If the same 20% of images was used to tune hyperparameters and to report the 98% accuracy in Table I and the conclusion, the number is not an unbiased estimate of generalization, and the comparison against the ConvNeXt and EfficientNet baselines is not a fair test. This concern is corroborated by an internal inconsistency: 20% of the stated 610 images is 122, but Section IV-B says the confusion matrices are based on 126 test images. The paper does not explain the discrepancy, which further undermines the split accounting. Additionally, the \"Accuracy\" column in Table I lists per-class accuracy values; the paper never defines how this is computed from the confusion matrix, nor whether the reported average (0.98) is a macro average. Without the underlying confusion matrices (Fig. 4-5) and code/data, the 98% figure is unverifiable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid deep-learning model that concatenates features from ConvNeXt and EfficientNet to classify falcon endoscopy images into three classes: Normal, Liver Disease, and Aspergillosis. The authors report an average test accuracy of 98% for the hybrid model, compared with 87% for ConvNeXt and 94% for EfficientNet, using a dataset of 610 images from the Sharjah Falcon Clinic. The manuscript includes training curves, confusion matrices, and an ROC curve, and concludes that the concatenated model surpasses both individual architectures and traditional diagnostic methods.","tokens_in":6351,"tokens_out":5297,"duration_ms":57783,"significance":"If the reported accuracy were obtained on a genuinely held-out test set, this would be a modest but useful application of transfer learning to a niche veterinary imaging problem. The comparison against two strong baselines is a reasonable starting point, and the paper gives credit for reporting confusion matrices and training curves. However, the claimed 4-point improvement over EfficientNet rests on a small, single-split, closed dataset with no statistical analysis, and the evaluation protocol is ambiguous. The architectural contribution is also incremental, reusing the concatenation scheme from the authors' prior brain-tumor classification work [2]. The current significance is therefore unverified and largely empirical.","major_comments":[{"comment":"The evaluation protocol is ambiguous: the pseudocode in Section III-A describes only a training (80%) and testing (20%) split, while Section IV states that the remaining 20% is used for internal validation and that hyperparameter fine-tuning was executed using the validation set. No separate held-out test set is ever described. If the same 20% of images was used both to tune hyperparameters and to compute the test metrics in Table I, the 98% figure is not an unbiased estimate of generalization, and the comparison against ConvNeXt and EfficientNet is not a fair test.","section":"Section III-A and Section IV"},{"comment":"There is an internal inconsistency in the number of test images: 20% of the stated 610 images is 122, but Section IV-B says the confusion matrices are based on 126 test images. The paper does not explain where the extra four images come from or whether the split was not exactly 80/20. This discrepancy undermines the split accounting and makes the reported metrics unverifiable.","section":"Section IV-B"},{"comment":"The Accuracy column in Table I lists per-class accuracy values and an average, but the manuscript never defines per-class accuracy or states whether the average is a macro average. Equation (1) defines only overall accuracy. Without a precise definition and without the underlying per-class counts, the reported average of 0.98 cannot be checked from the confusion matrices.","section":"Table I and Section IV-A"},{"comment":"The claim that the hybrid model outperforms traditional diagnostic methods is unsupported by the experiments: the study compares only ConvNeXt and EfficientNet and includes no traditional diagnostic modalities such as veterinarian interpretation, radiography, or laboratory tests. This overclaim should be removed or supported by a direct comparative experiment.","section":"Abstract and Conclusion"},{"comment":"No error bars, confidence intervals, or repeated-run statistics are reported for any of the models. With a single 80/20 split and only 610 images, the difference between 0.98 and 0.94 may be within run-to-run variability, so the claimed superiority of the proposed model over EfficientNet is not statistically established.","section":"Section IV"}],"minor_comments":[{"comment":"There are typographical issues with quotation marks in the abstract and introduction, such as \"Normal'\", \"Liver Disease'\", and \"'Aspergillosis'\", where opening quotes are missing.","section":"Section I"},{"comment":"The text says \"Figure 4 and 5\" and \"Figure 6 show\"; the first should be \"Figures 4 and 5\" and the second should be \"Figure 6 shows\".","section":"Section IV-B"},{"comment":"The word \"Specificity\" is misspelled as \"Specif icity\" in Equation (4).","section":"Equation (4)"},{"comment":"The paper states that each of the three classes contains 203 images, which sums to 609, but the total dataset is stated as 610 images; the remaining image is not accounted for.","section":"Section IV"},{"comment":"The block diagram is not accompanied by a layer-wise description of the concatenation, such as how the ConvNeXt and EfficientNet features are fused and which dense layers follow the concatenation; specifying this would improve reproducibility.","section":"Fig. 2 and Section III-A"},{"comment":"Reference [2] is the authors' own prior paper on a similar concatenated architecture for brain-tumor classification; the manuscript should state explicitly what is new in the present work beyond applying that architecture to a different dataset.","section":"Reference [2]"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is not currently supported because of the validation/test conflation and the 122-versus-126 test-image inconsistency. I recommend major revision rather than outright rejection because the issue could in principle be resolved by re-running the evaluation with a properly separated validation and test split and reporting the necessary statistics. If the authors cannot provide such a split or the underlying data, the paper should be rejected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a straightforward reuse of the authors' prior concatenated ConvNeXt-EfficientNet architecture, applied to a new, small falcon endoscopy dataset. The new dataset and the three-class problem (Normal, Liver, Aspergillosis) are genuinely new and could be useful to avian veterinarians, but the reported 98% test accuracy cannot be trusted as an unbiased number because the paper never cleanly separates validation and test.\n\nWhat it does well: it assembles a labeled dataset of 610 endoscopy images from a falcon clinic, and it compares the hybrid against the two base architectures. The performance table is coherent in showing the hybrid above the baselines, and the authors are transparent about their prior use of the same concatenation in [2]. That kind of direct extension is fine, provided the evaluation is sound.\n\nThe soft spots are real and load-bearing. The pseudocode in Section III-A defines only an 80/20 train/test split. Section IV then says the remaining 20% was used for 'internal validation' and that hyperparameter tuning was done on the validation set, while also reporting metrics on a 'test set' that is never separately described. The numbers also do not reconcile: 20% of 610 images is 122, but the confusion matrices are said to cover 126 test images. On top of that, the 'Accuracy' column in Table I lists per-class values without any definition; that is unusual and makes the reported average uninterpretable without the raw confusion matrices. No code or data are provided, and there are no error bars or repeated-seed statistics. None of this proves the method is bad, but the evaluation as written does not support the conclusion's claim of 'testing accuracies of 98%' as a reliable estimate.\n\nWho is this for? Avian veterinarians and applied-ML folks working on small medical imaging datasets. It is a niche contribution, not a methodological advance. A serious referee could help the authors tighten the split protocol, define the metrics, and release code/data. With that, the paper would be a modest but legitimate addition to the literature. As is, it should go to peer review rather than desk reject, because the dataset is new and the claim is potentially valuable but unverified.\n\nRecommendation: send to peer review, with instructions to the authors to clarify the split, provide the actual test set composition, define per-class accuracy, and share code and data.","headline":"New dataset and domain, but the 98% claim rests on a split protocol the paper never actually describes.","tokens_in":6900,"tokens_out":2875,"would_cite":false,"duration_ms":30486,"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":"This paper claims that concatenating ConvNeXt and EfficientNet into one classifier labels falcon endoscopy images as normal, liver disease, or aspergillosis with 98% reported test accuracy, beating the two base models by 4 and 11 points.","keywords":["falcon disease detection","ConvNeXt","EfficientNet","concatenated model","endoscopy image classification","aspergillosis","liver disease","transfer learning"],"falsifier":"Rebuild the model on the same 610 images using three disjoint splits—train, validation for tuning, and a test set seen only once at the end. If the test accuracy on that untouched split falls well below 98%, or no longer clearly beats ConvNeXt and EfficientNet alone, the paper's central claim fails. A cheaper check is to inspect whether any of the 126 test images used for the reported confusion matrices also appear in the 20% set described as 'internal validation.'","tokens_in":5953,"feed_emoji":"🦅","tokens_out":6294,"duration_ms":60076,"temperature":0.7,"pith_summary":"This paper tries to establish that a hybrid deep-learning model, built by concatenating the ConvNeXt and EfficientNet image classifiers, can sort falcon endoscopy images into three categories—normal, liver disease, and aspergillosis—more accurately than either architecture on its own. On a dataset of 610 labeled images from the Sharjah Falcon Clinic, the authors report an average test accuracy of 98% for the hybrid, against 87% for ConvNeXt and 94% for EfficientNet. The significance of the claim, if it holds, is that automated image triage could give falconers and veterinarians a fast, consistent second opinion for two serious avian conditions. The paper also frames the result as a step toward AI-assisted avian healthcare beyond falcons.","feed_headline":"Falcon disease AI hits 98% test accuracy","feed_subtitle":"The joint network beats ConvNeXt and EfficientNet on all three falcon health classes.","key_machinery":"The load-bearing object is the concatenation operation itself: a model architecture whose final feature representations from ConvNeXt and from EfficientNet are merged and passed to a classifier with three output classes. The paper sets the input at 128×128 pixels and trains the joined model over 50 epochs on 80% of the data, using transfer learning to initialize the two branches. ConvNeXt supplies fine-grained visual pattern extraction; EfficientNet supplies efficient scaling and robust feature reuse. The mechanism works by letting each branch's strengths compensate for the other's weaknesses, which is why the authors expect the combined score to beat either branch alone.","core_discovery":"The central claim is that a \"concatenated AI model\"—the feature outputs of ConvNeXt and EfficientNet joined into one classifier, fine-tuned on endoscopy images—classifies Normal, Liver Disease, and Aspergillosis with a reported 98% average testing accuracy, 0.97 precision, 0.98 recall, and 0.98 F1-score. Per-class results in the paper's Table I show the hybrid reaching 0.96, 0.99, and 0.98 accuracy for Normal, Liver, and Aspergillosis respectively, with the weakest per-class cell being Liver recall at 0.89 for EfficientNet and 0.83 for ConvNeXt. The authors argue this hybrid outperforms traditional diagnostic methods and the two base models on the same test images. In the authors' words, the concatenated model \"achieves testing accuracies of 98%, surpassing other deep learning models considered in this study.\"","pith_inferences":["I would infer that this concatenation recipe is a general template for veterinary image triage, not a falcon-specific finding, but the paper only demonstrates it on one dataset from one clinic.","I would not expect the 98% figure to transfer unchanged to images from other clinics, cameras, or lighting conditions, because the paper reports no cross-clinic evaluation.","The clearest testable extension is to run the identical three-way split with a validation set that is never shown during hyperparameter search; if accuracy stays near 98%, the claim is solid, and if it drops, the published number was inflated by tuning on the test set."],"forward_implications":["If the 98% accuracy holds, the model can automatically label the great majority of falcon endoscopy images as normal, liver disease, or aspergillosis without a specialist viewing each frame.","The hybrid's gains over ConvNeXt (87%) and EfficientNet (94%) imply that combining the two feature extractors extracts information neither branch captures alone on this dataset.","The paper's transfer-learning approach means the same two-branch architecture can be fine-tuned on new disease classes with far fewer images than training from scratch would require.","A veterinary clinic could use the model as a screening step, flagging suspected aspergillosis or liver disease for rapid confirmatory tests."],"supporting_citations":[{"why":"Supplies the concatenated ConvNeXt-EfficientNet architecture, previously used for brain-tumor classification, that the paper adapts to falcon endoscopy images.","marker":"[2]"},{"why":"Defines Aspergillosis as the fungal respiratory disease the model must separate from normal and liver-disease cases.","marker":"[6]"},{"why":"Provides the evaluation metrics (accuracy, precision, recall, F1) the paper uses to rank the hybrid against the base models.","marker":"[10]"}],"fun_headline_variants":["Hybrid AI spots falcon diseases at 98% accuracy","ConvNeXt+EfficientNet duo tops falcon disease tests","Falcon disease AI: hybrid model hits 98% accuracy","Concatenated model wins falcon ailment classification","AI blends two nets to diagnose falcon diseases"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported 98% accuracy depends on the test images being truly unseen: if the same 20% of images used to tune hyperparameters was also used to compute the final metrics, the score is not an honest estimate.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid AI spots falcon diseases at 98% accuracy","ConvNeXt+EfficientNet duo tops falcon disease tests","Falcon disease AI: hybrid model hits 98% accuracy","Concatenated model wins falcon ailment classification","AI blends two nets to diagnose falcon diseases"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000471,"raw_usage":{"total_tokens":2309,"prompt_tokens":876,"completion_tokens":1433,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":1364}},"tokens_in":492,"tokens_out":1433,"duration_ms":11013,"temperature":1.0,"reasoning_tokens":1364,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:38:58.150741+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rebuild the model on the same 610 images using three disjoint splits—train, validation for tuning, and a test set seen only once at the end. If the test accuracy on that untouched split falls well below 98%, or no longer clearly beats ConvNeXt and EfficientNet alone, the paper's central claim fails. A cheaper check is to inspect whether any of the 126 test images used for the reported confusion matrices also appear in the 20% set described as 'internal validation.'","supporting_citations":[{"cited_title":"Panthakkan, S","cited_arxiv_id":null,"evidence_quote":"Supplies the concatenated ConvNeXt-EfficientNet architecture, previously used for brain-tumor classification, that the paper adapts to falcon endoscopy images."},{"cited_title":"Tarello, ``Etiologic agents and diseases found associated with clinical aspergillosis in falcons,'' International Journal of Microbiology, vol","cited_arxiv_id":null,"evidence_quote":"Defines Aspergillosis as the fungal respiratory disease the model must separate from normal and liver-disease cases."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the evaluation metrics (accuracy, precision, recall, F1) the paper uses to rank the hybrid against the base models."}],"review_version":1}