{"id":"f2a8a8f8-c5db-48f0-86f3-b634a3631240","arxiv_id":"2507.12485","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Applying quantum transfer learning to a weak CNN raises reported dementia detection accuracy on OASIS-2 from 73 percent to 91 percent, but the comparison lacks error bars and uses test-set-based model selection.","lead":"The authors replaced the final layers of a weak brain-scan classifier with a small quantum circuit and report that the hybrid model detects dementia more accurately than the classical version, about 91 percent versus 73 percent test accuracy. This is an early demonstration of quantum transfer learning on medical images, but the evaluation has statistical weaknesses that make the exact gain uncertain.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.2 picks the 6-qubit/4-repetition QTL config by test accuracy from 24 grid-search runs, then reports that same test set's 91.29% as headline; test-set-based model selection can inflate the quantum gain, so the central claim needs a validation-based selection protocol.","rationale":"I agree with the reader's CONDITIONAL verdict, but my focus is narrower: the reader's weakest_assumption emphasizes the unseeded, unstratified split, whereas I see the explicit test-set-based grid-search selection (Section 4.2) as the single most load-bearing flaw. The random split concern is real but secondary; even a clean split cannot fix the selection bias. The paper is otherwise a straightforward application of the authors' prior QTL method, with no released code or formal verification, but those are not the decisive issue. The decisive question is empirical: is 91.29% an estimate of a chosen model's performance or the maximum of 24 noisy estimates? The proposed re-analysis with validation-based selection directly answers that. If it survives, the central claim is credible enough for a venue like GLSVLSI; if not, the paper should be revised or rejected. Since the reader already conditions on exactly this kind of methodological repair, my verdict does not move: still CONDITIONAL with no adjustment needed.","tokens_in":7511,"tokens_out":6403,"duration_ms":74866,"concrete_test":"Re-run the full protocol with a three-way split: train on the 70% training subjects, choose the best qubit count and repetition count (and CTL initializations) on a validation split of 15-20% of training subjects, then evaluate only the chosen model on the untouched test subjects. Repeat at least 5 times with different stratified, subject-level random seeds. If the best-on-validation QTL accuracy does not exceed CTL by the claimed margin with non-overlapping intervals, the central 'boost' claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the model-selection protocol in Section 4.2: the authors run a grid search over 8 qubit counts × 3 repetition counts = 24 QTL configurations and choose the 6-qubit, 4-repetition setup 'based on the test accuracies recorded in the ideal simulation.' The headline 91.29% is then measured on the same test set for that selected configuration. Because the test set has already been used to choose the winner, the final accuracy is an optimistically biased maximum, not an unbiased estimate of QTL performance. The bias is material here: the authors themselves note that performance was 'almost equivalent' across configurations, so differences among the 24 runs may be dominated by noise, and the maximum is exactly what would be inflated. The comparison to CTL is also affected because the reported CTL is the 'best-performing model among them' after several Glorot initializations (Section 3.4), though the main selection burden falls on QTL. This issue is independent of the unseeded/stratified split concern: even a perfect holdout split does not protect against selecting the model on that holdout. Without a separate validation set, repeated seeds, and error bars, the claimed 5.7-point improvement over classical fine-tuning cannot be distinguished from selection noise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a quantum transfer learning (QTL) pipeline for binary dementia detection on the OASIS-2 MRI dataset. A weak classical CNN baseline is first trained; its convolutional layers are frozen and its final dense layers are replaced by a dressed quantum network consisting of a classical pre-net, angle embedding, a variational quantum circuit with varying qubit counts and ansatz repetitions, and a classical post-net. The authors report that the best QTL configuration (6 qubits, 4 repetitions) reaches 91.29% test accuracy and 0.94 recall, versus 73.10% for the baseline and 86.36% for classical fine-tuning, and they present results from an IonQ Forte-1 noisy simulator to support robustness claims. The paper also reports a grid search over 24 QTL configurations and notes the occurrence of barren plateaus.","tokens_in":7783,"tokens_out":3181,"duration_ms":40885,"significance":"If the central claim were established, the paper would provide a useful empirical demonstration that a quantum fine-tuning head can improve a deliberately weak classical model on a medical imaging task, with a realistic NISQ noise model. The comparison against a classical fine-tuned counterpart on a public benchmark is a sensible experimental design, and the explicit reporting of the grid-search range and the barren-plateau observation are helpful. However, the reported quantitative gains are not currently reliable: the headline configuration is selected on the test set, all metrics come from single runs without error bars, and the noisy-simulator accuracy is higher than the noiseless accuracy for the same configuration. The paper does not provide code, random seeds, or a validation-based selection protocol, so the central claim is not yet reproducible or statistically supported.","major_comments":[{"comment":"The model selection protocol uses the test set twice. The authors state that they chose the 6-qubit, 4-repetition setup 'based on the test accuracies recorded in the ideal simulation' after a grid search over 24 configurations, and then report the accuracy of that same configuration on the same test set. Selecting the maximum over 24 runs on the test set biases the reported 91.29% upward, and the bias is especially concerning because the authors themselves note that performance was 'almost equivalent' across configurations, meaning the maximum may reflect selection noise rather than a genuinely superior model. A separate validation split should be used for selection, with the test set evaluated only once for the final reported numbers; alternatively, nested cross-validation should be used.","section":"Section 4.2, Table 3"},{"comment":"The data split is not described with enough detail to rule out leakage or imbalance. The test set is formed by a random 30% of subjects excluding IDs 1 and 2, but no random seed is reported and no stratification by class or by subject-level image count is described. Because IDs 1 and 2 contribute the largest numbers of images and are placed entirely in the training set, the test set may not be representative of the full dataset. The authors should report the seed, the class distribution of the train and test sets, and ideally perform subject-level stratified cross-validation to confirm that all reported accuracies are stable.","section":"Section 3.1"},{"comment":"The noisy-simulator result (91.29% test accuracy) is higher than the noiseless ideal-simulation result for the same 6-qubit, 4-repetition configuration (90.50%). Under a depolarizing noise model one would expect degradation, not improvement, unless the discrepancy is within run-to-run variation or stems from a different inference procedure. This anomaly is load-bearing for the paper's robustness claim and should be explained, ideally with repeated runs under both simulators and with error bars.","section":"Section 4.2, Table 3"},{"comment":"All reported metrics appear to come from a single training run per configuration. The classical fine-tuning baseline is described as the 'best-performing model' among several Glorot initializations, but the number of initializations and the spread of results are not given. Given the paper's own observation of flat optimization landscapes and near-equivalent QTL performance across configurations, the claimed 5.71-percentage-point improvement of QTL over CTL cannot be distinguished from seed noise without repeated initializations, standard deviations, or confidence intervals.","section":"Section 4.2 and 3.4"}],"minor_comments":[{"comment":"The text says 'Performance improved along all the other matrices by roughly 20%'; 'matrices' should be 'metrics'.","section":"Section 4.1"},{"comment":"The notation for the tanh output range is written as '[1, -1]' and then '[π/2, -π/2]'; the conventional ordering is '[-1, 1]' and '[-π/2, π/2]'.","section":"Section 3.3"},{"comment":"The phrase 'cross-validated and trained the QTL-based models' in Section 3.5 is ambiguous, and the earlier statement in Section 3 that the authors 'assess the overall performance by 4-fold cross-validation, training on the entire dataset, and validating on a separate test dataset' appears internally contradictory; the exact protocol should be stated once and consistently.","section":"Section 3.4"},{"comment":"The caption describes 'An ansatz' with Rz gates, a C-NOT ring, and controlled Ry gates, but the text says two different ansatzes were used; it should be clarified which ansatz is shown and how the second one differs.","section":"Figure 4"},{"comment":"The description of Figure 2 says demented and non-demented images from IDs other than 1 and 2 are 'shown in the third column,' but the figure is not self-explanatory; labeling the columns directly would improve clarity.","section":"Section 3.1"},{"comment":"The paper would benefit from releasing the code and specifying all random seeds, simulator versions, and the exact preprocessing steps for the OASIS-2 images, as these details are necessary for reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is a short empirical study whose central numerical claim is undermined by test-set-based model selection and the absence of error bars. The issues are fixable in principle by rerunning the experiments with a validation split, multiple seeds, and a clarified noisy-simulator protocol, so I do not recommend rejection on the merits of the research direction. However, the authors should be pressed to provide the missing protocol details and to report whether the QTL-vs-CTL advantage survives under a statistically honest evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper applies the authors' existing quantum transfer learning (QTL) pipeline from Ref. [1] to the OASIS-2 dementia dataset. The genuinely useful part is the setup: they intentionally start from a weak CNN baseline, compare against classical fine-tuning, and include a noisy-device simulation. The subject-level split that keeps the two dominant subjects (IDs 1 and 2) entirely in the training set shows real awareness of leakage, and acknowledging barren plateaus in a short conference paper is fair.\n\nBut the central claim—that QTL boosts a weak classifier by 24% over baseline and 5.7% over CTL—is not supported by the reported evaluation. The stress-test concern is correct and load-bearing: Section 4.2 selects the 6-qubit/4-repetition configuration after a grid search over 24 options, using test accuracies on the same test set that later produces the headline number. The authors even say performance was 'almost equivalent' across configurations, which means the gap between the chosen run and the runner-up is likely noise, and picking the max after seeing the test set inflates the gain. The CTL comparison is also weakened by selecting the best of several Glorot initializations on the same test set. With single runs, no error bars, and no separate validation set, the 5.7-point quantum-specific advantage is indistinguishable from selection noise.\n\nA second red flag: the noisy-simulator accuracy (91.29%) comes out higher than the noiseless result (90.50%) for the same configuration. That is not impossible, but it is the kind of result that needs explanation before it can be believed. The missing seed and unclear cross-validation description in Section 3 add to the unreliability.\n\nThe paper is honest about its limits in places, and there is no circularity beyond the selection bias. But as it stands, the metric that matters—whether QTL really improves upon a reasonable classical baseline—is unproven. The novelty is modest: this is the authors' own method applied to a known benchmark, and prior hybrid QML dementia papers already exist.\n\nWho gets value from this? A reader interested in seeing how QTL pipelines are being adapted to medical imaging might skim it for the architectural details. But I would not cite the empirical numbers.\n\nRecommendation: the paper deserves a serious referee only if the venue expects heavy revision. The evaluation can be fixed with repeated seeds, a held-out validation set for hyperparameter selection, stratified subject-level splits, error bars, and a proper explanation of the noisy-vs-noiseless discrepancy. Without those, the central claim is not credible.\n\nBest,\n[You]","headline":"A routine application of the authors' own QTL pipeline to OASIS-2, with a useful weak-baseline setup but an evaluation protocol that cannot support the headline accuracy claim.","tokens_in":8342,"tokens_out":1748,"would_cite":false,"duration_ms":23308,"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":"Quantum transfer learning can lift a weak dementia classifier from 73 percent to 91 percent accuracy on MRI data.","keywords":["dementia detection","quantum transfer learning","variational quantum circuit","hybrid quantum-classical model","MRI classification","OASIS-2 dataset","noisy simulation","binary classification"],"falsifier":"Rerun the same experiment with a stratified train/test split by subject, a fixed seed, and a class-balanced test set; if the QTL accuracy falls to the level of classical fine-tuning, the central claim fails.","tokens_in":7314,"feed_emoji":"🧠","tokens_out":4453,"duration_ms":49810,"temperature":0.7,"pith_summary":"This paper tries to show that a weak classical convolutional network for dementia detection can be made much more accurate by replacing its final dense layers with a trained quantum circuit, a strategy called quantum transfer learning. On the OASIS-2 MRI dataset, the best hybrid configuration reaches about 91 percent test accuracy and 94 percent recall, compared with about 73 percent for the untuned baseline and about 86 percent for classical fine-tuning. The authors also report that this performance survives simulation with a realistic noisy device model. If true, this would mean that lightweight classical models need not be discarded when they underperform: a quantum readout head can upgrade them while leaving the convolutional feature extractor frozen.","feed_headline":"Quantum transfer learning boosts dementia MRI detection to 91%","feed_subtitle":"A 6-qubit circuit readout beats the 73 percent baseline and classical fine-tuning, with 94 percent recall.","key_machinery":"The load-bearing component is the dressed quantum network (DQN), a hybrid layer that replaces the classical dense head. A pre-net maps the CNN's latent vector to the chosen number of qubits, a tanh activation and scaling by pi/2 prepare the values for angle embedding, and a parameterized quantum circuit built from Rz rotations, a ring of CNOT entangling gates, and controlled Ry rotations processes the embedded state. Measurements of the circuit give expectation values that a post-net turns into a single logit for binary classification. The same frozen convolutional feature extractor is used throughout, so the comparison isolates the effect of the quantum head against the classical fine-tuned head.","core_discovery":"The central claim is that quantum transfer learning can transform a suboptimal classical model into a better dementia classifier. Concretely, the paper reports that a 6-qubit, 4-repetition variational circuit, used as the classification head of a frozen small convolutional network, achieves 91.29 percent test accuracy, 0.8750 precision, 0.9412 recall, 0.9069 F1, and 0.9232 AUC on the OASIS-2 binary dementia/non-dementia task. The comparable classically fine-tuned model achieves 86.36 percent accuracy and 0.8739 recall, and the untuned baseline achieves 73.10 percent accuracy and 0.7394 recall. The paper further claims the quantum model remains reliable under a depolarizing noisy-device simulation, showing resilience without explicit error mitigation.","pith_inferences":["Editorial inference: the classically fine-tuned head was retrained from scratch after freezing the convolutional layers, so part of the quantum gain may come from reinitialization rather than from quantum mechanics; a control experiment that reinitializes the classical head with the same protocol would isolate the quantum contribution.","Editorial inference: the single random, non-stratified 30-percent test split means the reported margins could shift if the split is unlucky; testing across multiple seeds would show whether the quantum advantage is stable.","Editorial inference: because the paper observes a barren-plateau-like flat landscape, scaling the circuit to more qubits may erase the accuracy gain; a natural extension is to test whether the advantage persists with 10 or more qubits or with a different ansatz.","Editorial inference: the slight drop in AUC for QTL versus classical fine-tuning (0.9232 versus 0.9399) suggests the quantum model is not uniformly better; its claim is specifically about threshold-based metrics."],"forward_implications":["A weak, resource-constrained classical model can be upgraded by swapping only its final layer for a trained quantum circuit, without retraining the convolutional feature extractor.","The high recall of 0.94 means fewer demented patients are missed, which matters more than raw accuracy in a screening setting.","Because the gain persists under a depolarizing noise model, near-term quantum hardware without full error correction may be usable for this kind of hybrid classification.","The reported numbers give a concrete benchmark: 6 qubits and 4 circuit repetitions outperform both the classical baseline and classical fine-tuning on this dataset."],"supporting_citations":[{"why":"Supplies the quantum transfer learning methodology, including the hybrid classical-quantum model design, that the paper adapts for dementia detection.","marker":"[1]"},{"why":"Provides the preprocessed OASIS-2 brain MRI image collection used for training and testing the models.","marker":"[11]"},{"why":"Companion dataset source describing the OASIS-2 images and their preparation for machine learning.","marker":"[12]"},{"why":"Defines the barren plateau phenomenon that the paper says it observed during QTL training.","marker":"[7]"},{"why":"Supplies the data encoding patterns, including the angle embedding used in the dressed quantum network.","marker":"[14]"}],"fun_headline_variants":["Quantum transfer learning lifts dementia detection to 91%","6-qubit circuit powers dementia MRI detection to 91%","Quantum model beats classical fine-tuning in dementia screening","Noise-resistant quantum transfer learning for dementia detection","Quantum twist turns weak model into 91% dementia detector"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every reported accuracy depends on the assumption that the randomly chosen 30 percent of subjects used for testing is balanced and representative; the split used no fixed random seed and no stratification, and the two most image-heavy subjects were kept entirely in training.","fun_headline_variants_meta":{"raw":{"variants":["Quantum transfer learning lifts dementia detection to 91%","6-qubit circuit powers dementia MRI detection to 91%","Quantum model beats classical fine-tuning in dementia screening","Noise-resistant quantum transfer learning for dementia detection","Quantum twist turns weak model into 91% dementia detector"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000463,"raw_usage":{"total_tokens":2284,"prompt_tokens":881,"completion_tokens":1403,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":1326}},"tokens_in":497,"tokens_out":1403,"duration_ms":12496,"temperature":1.0,"reasoning_tokens":1326,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:24:29.990549+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the same experiment with a stratified train/test split by subject, a fixed seed, and a class-balanced test set; if the QTL accuracy falls to the level of classical fine-tuning, the central claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the quantum transfer learning methodology, including the hybrid classical-quantum model design, that the paper adapts for dementia detection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the preprocessed OASIS-2 brain MRI image collection used for training and testing the models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Companion dataset source describing the OASIS-2 images and their preparation for machine learning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the barren plateau phenomenon that the paper says it observed during QTL training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the data encoding patterns, including the angle embedding used in the dressed quantum network."}],"review_version":1}