{"id":"90321784-254e-4ea8-b537-3ff0dc51cfbe","arxiv_id":"2505.14716","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A PCA-plus-4-qubit-amplitude-encoding feature fusion pipeline with an SVM classifier reaches 99% accuracy on a public multi-region bone X-ray dataset, matching a transfer-learning baseline while cutting reported feature extraction time by 82%.","lead":"This paper proposes a hybrid pipeline that compresses X-ray images with PCA, encodes the result in a simulated 4-qubit quantum circuit, and merges both feature sets for fracture classification, reporting 99% accuracy on a public dataset. It is relevant as a potential low-cost alternative to deep transfer learning, but the quantum circuit definition and the comparison baseline are underspecified.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantum advantage rests on an uncontrolled comparison: classical SVM uses 8 PCA features, hybrid uses 16 fused features; added dimensions alone could explain the reported gains.","rationale":"The reader's weakest_assumption identifies test-set leakage; that is a valid concern and contributes to the REJECT verdict. I choose the uncontrolled comparison because it attacks the specific causal claim 'quantum features improve accuracy' independently of how the data are split. Even with a perfectly clean 80/20 split, Table 2 would still not distinguish quantum enrichment from dimensionality enrichment: the classical baseline uses 8 PCA features while the hybrid uses 16 fused features, and the ablation repeats the same mismatch. The paper gives no matched-dimensionality control, no error bars, and only a vague '10-15%' ablation claim. This is an internal logic problem in the argument as written, not a matter of outside consensus. A single controlled run—classical SVM with 16 PCA features on the same test set—would settle whether the quantum feature extractor is actually load-bearing. If that run matches 99%, the central quantum-advantage claim collapses even under a clean split; if it does not, the quantum claim gains support, though the unspecified circuit and split inconsistencies would still need addressing before the paper could be considered reproducible.","tokens_in":10334,"tokens_out":5128,"duration_ms":60245,"concrete_test":"On the same held-out test partition used for Table 2, train a classical SVM on a 16-dimensional feature vector consisting of the top 16 PCA components of the preprocessed images (or, alternatively, 8 PCA features plus 8 random nonlinear features such as random Fourier features). Compare accuracy, F1, Cohen's kappa, and ROC area to the Hybrid-Quantum SVM row in Table 2. If the 16-PCA SVM reaches ~99% accuracy and kappa ~0.98, then the quantum features are not the cause of the reported margin and the central claim requires revision. If the 16-PCA SVM remains near 83%, the quantum feature contribution would be supported, and the remaining concerns would shift to circuit reproducibility and data leakage.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim of quantum advantage is not established because the comparison in Table 2 and the ablation study are confounded. Section III-E defines the hybrid feature vector as 8 PCA features fused with 8 quantum features, giving 16 dimensions. Section IV states that the 'classical' baseline SVM is trained 'using only PCA-derived features,' i.e., 8 dimensions. Thus the classical-vs-hybrid comparison simultaneously changes two variables: the number of features (8 vs 16) and the feature construction (linear PCA projections vs PCA plus nonlinear quantum map). The reported jump from 83% to 99% accuracy and kappa from 0.65 to 0.98 is exactly the kind of gain that can come from adding informative dimensions or nonlinearity, without any quantum contribution. The ablation bullet 'when the quantum feature extraction module was removed, accuracy dropped by approximately 10-15%' repeats the same confound and gives no exact numbers, standard deviations, or matched-dimensionality control. Therefore the paper's strongest claim—that hybrid quantum SVM beats classical SVM 'by a clear margin in every possible parameter'—does not follow from the reported experiments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid quantum-classical pipeline for X-ray fracture detection. It reduces each image to 8 PCA features, encodes them into a 4-qubit amplitude-encoding circuit to produce 8 quantum features, fuses the two sets into a 16-dimensional vector, and trains classical ML classifiers (SVM, Random Forest, KNN, Gradient Boosting, Decision Tree). On the publicly available 9,463-image Bone Fracture Multi-Region X-Ray Dataset, it reports 99% accuracy with SVM, an 82% reduction in feature-extraction time relative to a MobileNet-LGBM transfer-learning baseline, and claims that hybrid quantum SVM beats classical SVM by a clear margin (99% vs 83% accuracy, Cohen's kappa 0.98 vs 0.65). An ablation bullet states that removing the quantum module lowers accuracy by approximately 10-15%.","tokens_in":10560,"tokens_out":4582,"duration_ms":47441,"significance":"If the claims were substantiated, the paper would be of interest to both the quantum-machine-learning and medical-imaging communities, since a small 4-qubit feature-enrichment module could improve classical classifiers while reducing computational cost versus deep transfer learning. The paper has notable strengths: it uses a public dataset, reports concrete metrics for several classifiers (Table 1), includes a classical-versus-hybrid comparison (Table 2), and provides wall-clock timing comparisons (Table 3). However, the central claim of quantum advantage is not currently established because the hybrid/classical comparison changes both feature dimensionality and feature construction, the quantum circuit is not specified, the data-split description is inconsistent, and the ablation study is not reported with enough detail to isolate the quantum contribution. These issues are load-bearing and prevent verification of the headline results.","major_comments":[{"comment":"The classical SVM baseline is trained on 'only PCA-derived features' (8 dimensions), while the hybrid SVM uses the fused 16-dimensional vector (8 PCA + 8 quantum). The comparison therefore changes two variables at once: feature dimensionality (8 vs 16) and feature construction (linear PCA vs PCA plus a nonlinear quantum map). The reported jump from 83% to 99% accuracy and from 0.65 to 0.98 kappa could plausibly result from the additional 8 informative dimensions alone, with no quantum contribution. To support the paper's central claim, the authors must include a matched-dimensionality classical baseline, e.g., an SVM on 16 PCA components, or on 8 PCA features concatenated with 8 classical nonlinear features. Without such a control, the statement that the hybrid approach 'is beating the classical SVM approach by a clear margin in every possible parameter' is not supported by the reported experiment.","section":"Section III-E and Section IV, Table 2"},{"comment":"The quantum feature extraction circuit is not defined. The text states that amplitude encoding is used with a 4-qubit scheme and that 'the data undergoes various quantum operations such as rotation gates (RX, RY) and entanglement mechanisms', but it does not provide the circuit diagram, the gate sequence, the entanglement pattern, whether the rotation angles are fixed or variational, or which observables are measured. Because the quantum circuit is the central novel component of the pipeline, this omission makes the method impossible to reproduce or audit. A precise specification of the circuit, including the mapping from the 8 PCA features to the 4-qubit state, is required.","section":"Section III-D"},{"comment":"The data split is described inconsistently across the paper. Section III-A states that the dataset is split 80/10/10 into training, validation, and testing (7,570/946/947 images), while Section IV says an 80/20 train-test split is used. Section III-B further states that preprocessing, including data augmentation and Canny edge detection, is completed before the split. If augmented or filtered versions of the same original image appear in both the training and test partitions, the reported 99% accuracy could be inflated. The authors must clarify the actual split, state how the validation set was used, and confirm that the train/test split was performed before any augmentation or filtering. This is essential to the credibility of the headline accuracy.","section":"Section III-A and Section IV (Experiment Setup)"},{"comment":"The ablation study does not provide enough information to assess the quantum contribution. The bullet 'when the quantum feature extraction module was removed, accuracy dropped by approximately 10-15%' gives no exact accuracy values, no standard deviations, and no indication of the feature dimensionality of the ablated model. It also repeats the confound described above: removing the quantum module reduces the feature vector from 16 to 8 dimensions, so the observed drop could be due to dimensionality alone. The authors should report the ablation with matched feature dimensionality (e.g., an 8-PCA-feature model versus a 16-feature model that replaces quantum features with classical features) and provide exact metrics, ideally with repeated runs.","section":"Section IV (Ablation Study)"}],"minor_comments":[{"comment":"The final sentence is an incomplete grammatical construction: '...into a 16-dimensional vector and then classifying with different machine learning models achieving 99% accuracy...'. Please revise to form a complete sentence.","section":"Abstract"},{"comment":"The text says the basis states are of 'log2 n qubits', but the correct expression is log2 N qubits, where N is the number of features. Please correct the notation.","section":"Section III-D, Eq. (1)"},{"comment":"The confusion matrices are not described with axis labels or class names, making it difficult to interpret the counts. Please add labels and, if possible, report the underlying counts in the text.","section":"Figures 3 and 4"},{"comment":"The MCVS column reports mean cross-validation values but no standard deviations. Since the text defines MCVs as repeated random splits, report the variability to allow assessment of stability.","section":"Table 1"},{"comment":"The conclusion states training times are reduced 'by up to 30%', but Table 2 shows 0.11 sec versus 0.07 sec (about 36% reduction) and Table 3 reports an 82% reduction in feature-extraction time. Please reconcile these numbers and state clearly which time comparison is intended.","section":"Section V (Conclusion) versus Table 2 and Table 3"},{"comment":"Please include the dataset version and access date for the Kaggle link. In addition, the claim of being 'on par with state-of-the-art transfer learning models' is supported only by comparison with a single model [16]; consider phrasing this as 'the transfer-learning baseline reported in [16]'.","section":"Reference [28]"},{"comment":"The statement that the dataset size is 'relatively limited' for DL methods is questionable given 9,463 images. Please justify this claim or remove it.","section":"Section III-F"},{"comment":"The text says 'Python 3.0'; this should be specified as a concrete version or 'Python 3.x' to avoid confusion.","section":"Section IV (Experiment Setup)"}],"recommendation":"major_revision","confidential_remarks":"The paper's biggest vulnerability is that the quantum advantage claim may be an artifact of the 8-versus-16-dimensional feature comparison. Before resubmission, the authors should run a matched-dimensionality classical baseline (e.g., SVM on 16 PCA components) and report the result; if that baseline already reaches near 99%, the quantum contribution is not demonstrated. I also urge the editors to require the full quantum circuit specification and a clear description of the split/preprocessing order, as these are needed for reproducibility. The paper appears to have been prepared quickly and would benefit from careful editing, but the underlying idea is not inherently unsound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper on the hybrid quantum-classical fracture pipeline. The application is new: combining PCA with a 4-qubit amplitude encoding circuit and feeding the fused 16-dimensional vector into classical classifiers on the public multi-region X-ray dataset is not something I've seen before. They also compare several classifiers, report confusion matrices, and show a 99% accuracy for SVM that matches the transfer learning baseline in [16]. That's real work, and the result isn't obviously fake.\n\nThe claimed quantum advantage, however, does not follow from their experiments. The classical SVM baseline uses only the 8 PCA features; the hybrid uses 8 PCA plus 8 quantum features. That changes two variables at once: feature count and feature construction. A classical SVM with 16 PCA features might well get the same jump from 83% to 99%. The ablation study has the same problem and gives no exact numbers or standard deviations either. This is not a minor issue — it's the entire basis for saying quantum features help.\n\nThe split description is also inconsistent. Section III-A says 80/10/10 train/validation/test, while Section IV says 80/20 train/test. Preprocessing, including augmentation and edge detection, is described as happening before the split. If any augmented or filtered version of a training image leaks into the test set, the accuracy is inflated. There's no code, no error bars, and the circuit itself is never specified: no gate sequence, no entanglement structure, no amplitude encoding map beyond the general equation. Without the circuit, the 'quantum feature extraction' is a black box, and the 82% time reduction is unverifiable.\n\nThe citation pattern is fine — each component is attributed to prior work. The paper is coherent and the authors are not sloppy in an obvious way; they just haven't given reviewers enough to check the central claim. The public dataset helps, but the experimental design needs a matched-dimensionality control, a full circuit definition, and a consistent, leakage-proof split.\n\nThis paper is for someone curious about whether simulated quantum feature enrichment can work in medical imaging with tiny feature vectors, but they should treat the headline results as unverified. I wouldn't cite it in my own work, and I wouldn't spend a whole reading group on it unless the goal is to discuss confounded baselines. For peer review: the topic is timely and the flaws are fixable in principle, so I would send it to a careful referee rather than desk reject — but the referee should be instructed to demand the missing details before taking the quantum advantage claim seriously.","headline":"A novel application with a plausible-sounding result, but the quantum advantage claim is unsubstantiated because the comparison is confounded and the circuit is never defined.","tokens_in":11053,"tokens_out":3260,"would_cite":false,"duration_ms":36047,"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":"A 16-dimensional fused PCA-plus-quantum feature vector lets a plain SVM match deep transfer learning at 99% accuracy on bone X-rays while cutting feature-extraction time by 82%.","keywords":["Bone fracture detection","X-ray diagnosis","Principal Component Analysis","Quantum feature extraction","Amplitude encoding","Hybrid quantum-classical pipeline","Support vector machine","Medical image classification"],"falsifier":"Re-run the pipeline with the split performed before any augmentation, filtering, or edge detection, and verify by image hashing that no training image or any transformed copy of it appears in the test set; then recompute SVM accuracy and Cohen's kappa. As a second check, fit a classical SVM on the eight PCA features alone under the same leakage-free split: if its accuracy already approaches 99%, the claimed contribution of the quantum module collapses.","tokens_in":10128,"feed_emoji":"🦴","tokens_out":8301,"duration_ms":83745,"temperature":0.7,"pith_summary":"Automated fracture screening usually means heavy deep networks trained on large annotated image sets. This paper proposes a far lighter route: compress each X-ray with PCA to eight numbers, feed those numbers through a simulated four-qubit quantum amplitude-encoding circuit, and fuse the eight classical values with eight quantum-derived values into a 16-feature vector for an ordinary classifier such as an SVM. On a public multi-region bone X-ray dataset, this pipeline is reported to reach 99% accuracy, matching the transfer-learned baseline it is compared with, while cutting mean feature-extraction time by 82%. The paper's thesis is that the quantum half of the vector supplies nonlinear structure that pure PCA lacks, letting a 16-feature SVM outperform a classical PCA-only SVM (99% versus 83% accuracy). If the result holds, accurate fracture screening no longer requires a deep-learning stack or large annotated corpora.","feed_headline":"16 features, 99% accuracy: quantum-classical bone X-ray diagnosis","feed_subtitle":"PCA plus a four-qubit quantum circuit matches transfer learning while cutting extraction time by 82 percent.","key_machinery":"The load-bearing object is the 4-qubit quantum amplitude-encoding circuit used for feature enrichment. Eight PCA-reduced features are normalized and mapped onto the amplitudes of a four-qubit state, $|\\psi\\rangle = \\sum_{i=0}^{N-1} x_i |i\\rangle$, then passed through rotation gates (RX, RY) and entanglement operations before being measured to yield eight quantum-enhanced features. These eight values are concatenated with the eight PCA values to form the 16-dimensional dual-source feature vector that feeds the classical classifiers. The circuit is what the paper credits for the accuracy jump: removing the quantum module is reported to drop accuracy by 10–15 percent, and the SVM's gain from 83% to 99% is attributed to the structure these measured features add on top of PCA.","core_discovery":"The paper's central claim is that a dual-source feature vector—eight principal components of an X-ray image plus eight features measured from a 4-qubit amplitude-encoding circuit—is enough to classify bone fractures at 99% accuracy on the Bone Fracture Multi-Region X-Ray Dataset. With this representation, an SVM achieves 99% accuracy, an F1-score of 0.99, and a Cohen's kappa of 0.98, while a classical SVM on PCA features alone reaches 83% accuracy and 0.65 kappa; the paper describes the hybrid SVM as beating the classical SVM \"by a clear margin in every possible parameter.\" The same fused pipeline matches the 99% accuracy reported by the MobLG-Net transfer-learning baseline while reducing mean feature-extraction time from 67.5 seconds to 12.1 seconds. The intended conclusion is that quantum amplitude encoding contributes genuine, non-classically-redundant information to the feature set, and that this enrichment is obtainable within a 4-qubit simulator constraint.","pith_inferences":["An inference the authors leave implicit: since the quantum circuit is simulated on a classical workstation, the 83-to-99 point gain is a claim about feature representation, not about a hardware speedup; quantum hardware is not needed to reproduce the result.","A testable extension they do not run: replace the quantum module with a classical nonlinear expansion of the same eight PCA features (for example, polynomial or random Fourier features) at equal dimension; if that also reaches 99%, the measured gain is not specific to quantum encoding.","If the 99% figure survives a strictly leakage-free split, the same PCA-plus-amplitude-encoding-plus-fused-SVM recipe is a natural candidate for other scarce-data medical imaging tasks, such as chest or pediatric X-ray screening, where annotation costs are the main bottleneck."],"forward_implications":["A clinic with only a CPU workstation could run fracture screening: the reported SVM training time is 0.07 seconds and evaluation time is 0.04 seconds.","If a 16-feature fused vector really matches deep transfer learning at 99%, the need for large annotated datasets and GPU training is removed for this task, lowering barriers in resource-limited settings.","The reported 82% reduction in feature-extraction time and roughly 15% reduction in inference time suggest the hybrid representation could serve real-time decision support where transfer-learned CNNs are too slow.","The ablation finding—accuracy falls 10–15% without the quantum module—means the quantum features, not the PCA features, are the active ingredient to optimize in later work."],"supporting_citations":[{"why":"Supplies the Bone Fracture Multi-Region X-Ray Dataset that every accuracy and timing figure is measured on.","marker":"[28]"},{"why":"Defines the MobLG-Net transfer-learning baseline whose 99% accuracy and 67.5-second extraction time the hybrid pipeline claims to match and reduce.","marker":"[16]"},{"why":"Justifies PCA as the dimensionality-reduction step that compresses images to eight features before quantum encoding.","marker":"[21]"},{"why":"Provides the theoretical basis for quantum-enhanced feature spaces making classification easier for classical models.","marker":"[24]"},{"why":"Grounds the use of quantum circuits inside machine-learning pipelines, the paper's overall strategy.","marker":"[22]"},{"why":"Cited as evidence that amplitude encoding is effective in medical imaging, supporting the 4-qubit encoding choice.","marker":"[36]"}],"fun_headline_variants":["4-qubit circuit raises X-ray fracture diagnosis to 99% accuracy","Hybrid quantum AI reads bone X-rays at 99% accuracy, 82% faster","PCA plus 4 qubits: fracture diagnosis hits 99% like transfer learning","Quantum-classical pipeline matches deep learning on bone X-rays"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 99% figure stands or falls on the test images being truly unseen: the paper preprocesses, augments, and edge-detects the images before splitting the data, and it reports the split inconsistently (80/10/10 in the methodology, 80/20 in the experiments), so any transformed copy of a training image that lands in the test set would inflate the reported accuracy.","fun_headline_variants_meta":{"raw":{"variants":["4-qubit circuit raises X-ray fracture diagnosis to 99% accuracy","Hybrid quantum AI reads bone X-rays at 99% accuracy, 82% faster","PCA plus 4 qubits: fracture diagnosis hits 99% like transfer learning","Quantum-classical pipeline matches deep learning on bone X-rays"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001256,"raw_usage":{"total_tokens":5115,"prompt_tokens":883,"completion_tokens":4232,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":4150}},"tokens_in":499,"tokens_out":4232,"duration_ms":33952,"temperature":1.0,"reasoning_tokens":4150,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:27:15.821707+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the pipeline with the split performed before any augmentation, filtering, or edge detection, and verify by image hashing that no training image or any transformed copy of it appears in the test set; then recompute SVM accuracy and Cohen's kappa. As a second check, fit a classical SVM on the eight PCA features alone under the same leakage-free split: if its accuracy already approaches 99%, the claimed contribution of the quantum module collapses.","supporting_citations":[{"cited_title":"Papers With Code","cited_arxiv_id":null,"evidence_quote":"Supplies the Bone Fracture Multi-Region X-Ray Dataset that every accuracy and timing figure is measured on."},{"cited_title":"Novel transfer learning based bone fracture detection using radiographic images,","cited_arxiv_id":null,"evidence_quote":"Defines the MobLG-Net transfer-learning baseline whose 99% accuracy and 67.5-second extraction time the hybrid pipeline claims to match and reduce."},{"cited_title":"Principal component analysis: a review and recent developments,","cited_arxiv_id":null,"evidence_quote":"Justifies PCA as the dimensionality-reduction step that compresses images to eight features before quantum encoding."},{"cited_title":"Supervised learning with quantum-enhanced feature spaces,","cited_arxiv_id":null,"evidence_quote":"Provides the theoretical basis for quantum-enhanced feature spaces making classification easier for classical models."},{"cited_title":"Brain tumor diagnosis using quantum convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Cited as evidence that amplitude encoding is effective in medical imaging, supporting the 4-qubit encoding choice."}],"review_version":1}