{"id":"978e934f-088a-455c-87f7-63857f6fff1f","arxiv_id":"2411.15086","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An unsupervised QUBO-based segmentation pipeline using quantum annealing and variational circuits performs comparably to supervised UNet on small mammography crops, with claimed speedups over Gurobi.","lead":"The paper tests quantum annealing and quantum-inspired solvers for unsupervised segmentation of mammography images, framing the task as a QUBO optimization problem. On 42-by-42 image crops from the INbreast dataset, it reports that the quantum approaches roughly match a classical solver and narrow the gap to supervised neural networks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 15's QUBO lacks any unary/data term; for nonnegative edge weights its global minima are the all-zero and all-one masks, so the written objective cannot produce the reported segmentations, and the alpha sweep is vacuous as a consequence.","rationale":"The paper is clearly written and the general pipeline is motivated, but the mathematical core of the method is not sound. The reader correctly identified that the smoothness term in Eq. 15 is redundant with the min-cut term for binary variables, making the alpha sweep meaningless. However, the more load-bearing problem is that the entire objective has no unary term, so the constant labelings are global minima with objective value zero, while every actually segmented mask has positive energy. No optimizer minimizing Eq. 15 could produce the reported Dice/IoU values. This is an internal mathematical inconsistency, not merely a disagreement with external consensus, and it directly invalidates the central claim that quantum annealing and VQAs match classical optimizers on a meaningful segmentation objective. Since the reader already recommended REJECT and this deeper concern reinforces that conclusion, the verdict should remain unchanged.","tokens_in":14546,"tokens_out":8786,"duration_ms":86306,"concrete_test":"Analytic check: for any image, evaluate Eq. 15 at the all-zero and all-one masks; both give 0, and any nonconstant mask has positive energy because every crossing edge contributes W_ij > 0. This proves the global minima of the written QUBO are the trivial masks. Then compute the energy of the masks reported in Table 2; if they have positive energy, they are not solutions of Eq. 15, demonstrating that the experiments solved a different objective. A minimal re-run: implement Eq. 15 exactly with alpha = 0.1 in Gurobi on one 42x42 image and confirm the returned binary solution is all-constant; if it is not, the implementation must contain an undocumented term.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central QUBO in Eq. 15 is L = sum W_ij[(x_i + x_j - 2 x_i x_j) + alpha(1 - (x_i + x_j - 1)^2)]. For binary x_i, the parenthesized smoothness term equals exactly x_i + x_j - 2 x_i x_j, so L = (1 + alpha) * sum W_ij(x_i + x_j - 2 x_i x_j). Thus alpha only rescales the objective and cannot influence the argmin; the alpha sweep in Section 2.2.1 is vacuous. More seriously, this objective contains no unary/data term tied to pixel intensities, no source/sink, and no label-balance constraint. All edge weights W_ij are Gaussian similarities (Eq. 10), hence nonnegative. Consequently L(0,...,0) = L(1,...,1) = 0, and every nonconstant labeling has L > 0. The global minimizers are the trivial empty and full masks, not the ROI masks whose Dice/IoU near 0.84/0.74 are reported in Table 2. The same triviality affects the VQA relaxation in Eq. 20. The experiments therefore cannot be minimizing the written QUBO; either an undisclosed unary term or constraint was used, or the reported masks are not optima of Eq. 15. This undermines the central speed and accuracy claims. The reader's alpha-redundancy objection is correct but is a symptom of this deeper misspecification. Eq. 16's c_i and Q_ii do combine to the correct linear coefficient, so the inconsistency there is secondary.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end unsupervised pipeline for segmenting breast masses in mammograms. It combines a quantum-inspired image transformation (Sect. 4.2), a QUBO formulation of a min-cut plus smoothness objective (Eq. 15), and several solvers: simulated annealing, quantum annealing, and variational quantum circuits, benchmarked against Gurobi, Otsu thresholding, and supervised UNet/ResUNet models on the INbreast dataset. The central claims are that the quantum annealing and VQA results are comparable to the classical Gurobi optimizer, that quantum annealing is an order of magnitude faster, and that the unsupervised pipeline approaches the accuracy of supervised UNet-based methods.","tokens_in":14899,"tokens_out":4363,"duration_ms":44883,"significance":"If the claims were correct, this would be a meaningful demonstration of quantum annealing and variational circuits for medical image segmentation, with a public dataset and standard metrics. The experimental scope is appropriate, and the quantum-inspired image representation is evaluated independently on supervised models, which is a useful self-contained contribution. However, the paper ships no reproducible code or machine-checked derivations, and the core QUBO objective as written is algebraically degenerate: the smoothness term is identical to the min-cut term for binary variables, and the objective has no data-attachment term, so its global minima are the empty and full masks rather than the reported ROI segmentations. Because the reported performance numbers and speed comparisons rest on this ill-posed objective, the main claims are not supported as written.","major_comments":[{"comment":"For binary variables, (1 - (x_i + x_j - 1)^2) = x_i + x_j - 2 x_i x_j, exactly the same expression as |x_i - x_j|. Therefore the term multiplied by alpha is identical to the min-cut term, so alpha merely rescales the total objective and cannot change the optimal binary solution. The alpha sweep in Sect. 2.2.1 (\"Effect of alpha\") is vacuous, and the statement that alpha=0.1 provides the best balance is not supported. Furthermore, the QUBO coefficients in Eq. (16) do not follow from Eq. (15): with the stated Q_ij = -2(1+alpha) W_ij and the standard x^T Q x convention, the effective cross-coefficient would be -4(1+alpha) W_ij, not -2(1+alpha) W_ij, while the claimed linear coefficient (2alpha+1) sum_j W_ij together with Q_ii = -alpha sum_j W_ij only reproduces (alpha+1) sum_j W_ij after absorbing the diagonal. The algebraic mismatch must be resolved before any experimental comparison is meaningful.","section":"Sect. 4.3.1, Eq. (15)"},{"comment":"The objective in Eq. (15) contains no unary or data-attachment term tied to pixel intensities, no source/sink term, and no label-balance constraint. Since all edge weights W_ij are Gaussian similarities (Eq. 10) and hence nonnegative, L(0,...,0) = L(1,...,1) = 0, while every nonconstant binary labeling has strictly positive cost. Thus the global minima of the written objective are the trivial empty and full masks, not the ROI masks whose Dice/IoU values are reported in Table 2. The same issue affects the VQA relaxation in Eq. (20). The experiments therefore cannot be minimizing the written QUBO; either an undisclosed unary term or constraint was used, or the reported masks are not optima of Eq. (15). This undermines the central accuracy and speed claims, including the comparison against Gurobi and the claim that quantum annealing is an order of magnitude faster.","section":"Sect. 4.3.1, Eq. (15), and Sect. 2.2.2, Table 2"},{"comment":"Several key hyperparameters are selected on the same test set used to report final scores: alpha is chosen by sweeping over the test set (Sect. 2.2.1), p=0.9 is chosen after \"thorough experimentation\" (Sect. 4.2), T=0.3 for the VQA warm start is fixed without a separate validation procedure, and \\hat{sigma} = 0.5 std(z) is data-dependent. This is mild test-set fitting, but it matters because the reported Dice/IoU differences among QA, VQA, and Gurobi are small (Table 2), and the paper provides no confidence intervals or statistical tests. At minimum, the authors should state which images were used for hyperparameter selection and which for scoring, or use a nested validation split.","section":"Sects. 2.1, 2.2.1, 4.2"},{"comment":"The execution-time comparison is reported for a single 42x42 grid (1764 pixels) and for D-Wave with 2000 annealing runs, but the paper does not specify whether the reported quantum annealing time includes embedding, QPU access, or post-processing, and Figure 5 appears to lack units and error bars. The claim that quantum annealing is an order of magnitude faster than Gurobi therefore lacks the detail needed to assess fairness and reproducibility. The authors should report the exact timing methodology, the number of images, and the variance across runs, and should clarify whether Gurobi is given the same warm-start or optimality-tolerance settings.","section":"Sect. 2.2.3 and Sect. 4.3.3"}],"minor_comments":[{"comment":"There is a typo: \"computational resourcess\" should be \"computational resources\".","section":"Abstract"},{"comment":"The text says \"optimizes both the Dice ans IoU scores\"; \"ans\" should be \"and\".","section":"Eq. (3)"},{"comment":"The notation for the quantum-inspired transformation is difficult to parse: the inner-product-like expression I_{ij} times the state overlaps is not defined precisely, and the rendering of Eq. (4) is garbled. The authors should rewrite the definition with standard bra-ket notation or explicit scalar functions.","section":"Sect. 4.2, Eqs. (4)-(6)"},{"comment":"Table 1 reports UNET on original images with Dice 0.911, while Table 2 reports UNET with Dice 0.85; the discrepancy is not explained. The authors should clarify whether the two tables use different image resolutions, different test subsets, or different training configurations.","section":"Tables 1 and 2"},{"comment":"The text refers to a \"Potts model\" smoothness penalty, but the penalty 1 - delta(x_i, x_j) is identical to the min-cut term for binary variables. The authors should either introduce a genuinely different smoothness term (e.g., based on higher-order interactions or a label-cost penalty) or remove the claim that the two objectives are competing.","section":"Sect. 4.3.1, Eq. (15)"}],"recommendation":"reject","confidential_remarks":"The stress-test concern from the reader is correct and, in my reading, the more severe issue is that Eq. (15) has no data term, making the empty and full masks the exact global minima. This is not a presentation issue; it means the reported experiments cannot be reproducing the written objective. The alpha-redundancy objection is a symptom of the same missed identity. I see no way to repair the manuscript within a normal revision cycle, because the entire experimental comparison would need to be redone with a corrected QUBO formulation (likely including a unary data term and a non-redundant smoothness term). The quantum-inspired image representation section is a reasonable standalone contribution, but it does not salvage the paper's central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis one should not be taken at face value. The central QUBO, Eq. 15, has a serious algebraic identity: for binary variables, the smoothness term exactly equals the min-cut term. Alpha just rescales the whole objective and cannot influence the argmin, so the alpha-sweep experiment in Section 2.2.1 is vacuous. Worse, the objective has no unary or data term. With nonnegative Gaussian edge weights, the global minima are the all-zero and all-one masks. The reported Dice/IoU scores around 0.84/0.74 cannot come from minimizing this objective. Either an undisclosed unary term exists or the experiments used a different objective. That is not a minor typo; it undermines the paper's core claim that quantum annealing and VQA solve this segmentation QUBO.\n\nWhat is genuinely new is the end-to-end unsupervised pipeline: the Konar-style quantum-inspired image transform feeding a QUBO min-cut formulation, evaluated on INbreast mammography crops with a broad side-by-side of simulated annealing, D-Wave QA, VQA, Gurobi, Otsu, UNet, and ResUNet. The writing is clear, the dataset is public, and Table 1's finding that the quantum-inspired transform speeds up UNet convergence is a useful empirical observation. Code availability is promised, which is good practice.\n\nThe soft spots extend beyond the QUBO bug. The timing comparison pits QA against Gurobi, a general-purpose solver, when a specialized graph-cut solver such as Boykov-Kolmogorov is the natural baseline; an order-of-magnitude speedup claim is misleading without that comparison. Hyperparameters alpha, sigma, p, T, and mu are tuned on the test set, and no error bars are reported. The experiments use 42x42 downsampled images, far from clinical resolution, so the \"comparable to supervised UNet\" claim is narrow. The stress-test note is correct: the alpha-redundancy objection is a symptom of the deeper misspecification—no unary term exists.\n\nThe empirical scaffolding looks genuine, but the load-bearing math does not hold. The right response is a major revision: either add a proper data term and show the derivation, or state the actual objective that was minimized. If the implementation includes an unstated unary term, that must be disclosed before the timing and hyperparameter claims can be fairly assessed.\n\nThis paper deserves a serious referee because the topic is timely and the comparison is broad, but I would not cite it in its current form.","headline":"The paper's QUBO objective is algebraically degenerate—alpha is redundant and the global minima are trivial—so the reported segmentation results cannot be optima of the written objective.","tokens_in":15453,"tokens_out":5517,"would_cite":false,"duration_ms":50799,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"According to the paper, quantum annealing can segment mammograms about as accurately as a classical solver while running roughly ten times faster, and the unsupervised pipeline's masks land close to supervised UNet quality.","keywords":["quantum annealing","image segmentation","QUBO","unsupervised learning","mammography","variational quantum circuits","medical image analysis","quantum-inspired image representation"],"falsifier":"Evaluate Eq. 15 over all binary labelings of a small graph at $\\alpha=0.1$ and $\\alpha=100$: since for binary $x_i,x_j$ one has $1-\\delta(x_i,x_j)=x_i+x_j-2x_ix_j$, the loss is $(1+\\alpha)$ times the min-cut term, so the optimal labeling is identical at both values, which would refute the claimed trade-off that the paper uses to set $\\alpha=0.1$.","tokens_in":14304,"feed_emoji":"🩺","tokens_out":13011,"duration_ms":115747,"temperature":0.7,"pith_summary":"The paper proposes a complete unsupervised pipeline for segmenting lesions in mammograms, built around a quantum-inspired preprocessing step and a quadratic unconstrained binary optimization (QUBO) formulation of the segmentation task. It claims that solving that QUBO with quantum annealing or with variational quantum circuits gives masks whose Dice and IoU scores match a classical numerical solver, that quantum annealing runs about ten times faster than that solver in their experiments, and that the unsupervised masks land close to those of a supervised U-shaped network (UNet) and its stronger residual variant. If these claims hold, quantum annealing would offer a label-free route to medical-image segmentation with accuracy comparable to supervised deep learning, which matters in settings where annotated datasets are scarce or expensive. The paper also reports that its quantum-inspired image representation sharpens lesion boundaries and cuts the training epochs of supervised models by about a third when used as their input.","feed_headline":"Quantum annealing matches classical solver at 10x speed for mammograms","feed_subtitle":"No labels, no training: an unsupervised quantum pipeline lands close to UNet accuracy on breast-cancer images.","key_machinery":"The load-bearing object is the segmentation QUBO: for binary pixel labels $x_i$, the cost sums over neighboring pixels $W_{ij}[(x_i+x_j-2x_ix_j)+\\alpha(1-\\delta(x_i,x_j))]$, where $W_{ij}$ is a Gaussian similarity weight, $\\delta(x_i,x_j)$ is the Kronecker delta (one when the two labels match), the first term cuts weakly similar pairs apart, and the $\\alpha$ term is intended to encourage smooth masks. The second piece is the quantum-inspired image transform, a single-pass filter that reweights each pixel by its local contrast and neighborhood intensity sum through a multilevel sigmoid, producing the enhanced image that feeds the graph. The QUBO is solved three ways: simulated annealing, quantum annealing on a physical processor, and a variational circuit that amplitude-encodes one pixel per basis state with an ancilla qubit and treats the measured probability $|\\beta_i|^2$ as the pixel's class score. The classical baselines are a classical optimization solver, a thresholding method, and two supervised neural networks.","core_discovery":"On the paper's own terms, the central discovery is that a segmentation mask can be recovered without any labeled data by mapping a quantum-enhanced image to a weighted grid graph, writing a min-cut-with-smoothness objective as a QUBO, and minimizing that QUBO with quantum annealing or a variational circuit. On 42x42 pixel mammography crops, quantum annealing reaches Dice/IoU of 0.84/0.74 and the variational approach 0.83/0.73, essentially matching the classical solver's 0.84/0.74, while clearly beating the histogram-thresholding baseline (0.75/0.62). Those numbers sit just below the supervised UNet (0.85/0.75) and ResUNet (0.89/0.81), which is the paper's evidence that unsupervised quantum optimization can approach supervised state-of-the-art. The average execution time of quantum annealing is reported as an order of magnitude shorter than the classical solver's, with lower variance, which the paper presents as the concrete practical advantage of the quantum step.","pith_inferences":["Beyond the paper, the claimed speedup is tied to the tested image size and solver settings; a fairer comparison would report quantum annealing with embedding and sampling overhead against the classical solver under matched optimality tolerances, and repeat the timing on larger grids.","Beyond the paper, the quantum-inspired transform may be doing much of the accuracy work; running the same QUBO pipeline with and without that transform, and with the classical solver in place of quantum annealing, would attribute the gain between preprocessing and solver.","Beyond the paper, the variational circuit's logarithmic qubit count suggests a natural scaling test: if real quantum hardware removes the simulation overhead, the same encoding could be pushed to much larger segmentation grids, where the classical solver's runtime grows steeply."],"forward_implications":["Quantum annealing can serve as an unsupervised segmentation optimizer for small mammography crops, matching the classical solver's accuracy while reducing average execution time by roughly an order of magnitude.","The full unsupervised pipeline reaches Dice/IoU scores close to those of supervised UNet on this dataset, implying that comparable mask quality can be obtained without expert annotations or network training.","The quantum-inspired preprocessing alone improves supervised training: UNet/ResUNet trained on the transformed images converge in about 30 epochs instead of 45, with slightly higher IoU.","The variational circuit uses only a logarithmic number of qubits for the pixel indices, so if runtime on real hardware drops, it is the most scalable of the quantum options considered here."],"supporting_citations":[{"why":"Supplies the mammography images and expert ROI masks that all methods are scored against.","marker":"[32]"},{"why":"The classical QUBO solver used as the accuracy reference and the runtime baseline for the claimed speedup.","marker":"[23]"},{"why":"Defines the supervised UNet baseline whose Dice/IoU the unsupervised pipeline is claimed to match.","marker":"[17]"},{"why":"Defines the residual UNet baseline that sets the stronger supervised anchor in the comparison.","marker":"[51]"},{"why":"Supplies the min-cut-as-QUBO formulation that this work adapts to mammography segmentation.","marker":"[41]"},{"why":"Introduces the qubit-efficient variational quantum approach whose amplitude-encoding and ansatz the paper's VQA follows.","marker":"[42]"},{"why":"Source of the quantum-inspired image transformation that the paper adapts as its preprocessing step.","marker":"[27]"},{"why":"Defines the histogram-thresholding baseline that the quantum methods are shown to outperform.","marker":"[9]"}],"fun_headline_variants":["Quantum annealing speeds unsupervised mammogram segmentation 10x","Unsupervised quantum segmentation matches classical at 10x speed","Quantum annealing hits UNet-level accuracy on mammograms without labels","Mammogram segmentation: quantum annealing 10x faster, no training data","Quantum pipeline segments breast scans unsupervised, near UNet accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the smoothness term in Eq. 15 is genuinely different from the min-cut term, so that the tuning parameter $\\alpha$ really changes which segmentation is best; if the two terms coincide for binary pixels, the $\\alpha$-tuning story and the reported coefficients do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Quantum annealing speeds unsupervised mammogram segmentation 10x","Unsupervised quantum segmentation matches classical at 10x speed","Quantum annealing hits UNet-level accuracy on mammograms without labels","Mammogram segmentation: quantum annealing 10x faster, no training data","Quantum pipeline segments breast scans unsupervised, near UNet accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1366,"prompt_tokens":1023,"completion_tokens":343,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":257}},"tokens_in":639,"tokens_out":343,"duration_ms":3963,"temperature":1.0,"reasoning_tokens":257,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:30:50.839984+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate Eq. 15 over all binary labelings of a small graph at $\\alpha=0.1$ and $\\alpha=100$: since for binary $x_i,x_j$ one has $1-\\delta(x_i,x_j)=x_i+x_j-2x_ix_j$, the loss is $(1+\\alpha)$ times the min-cut term, so the optimal labeling is identical at both values, which would refute the claimed trade-off that the paper uses to set $\\alpha=0.1$.","supporting_citations":[{"cited_title":"Gurobi Optimizer Reference Manual (2024)","cited_arxiv_id":null,"evidence_quote":"The classical QUBO solver used as the accuracy reference and the runtime baseline for the claimed speedup."},{"cited_title":"Q-Seg: Quantum Annealing-Based Unsupervised Image Segmentation","cited_arxiv_id":"2311.12912","evidence_quote":"Supplies the min-cut-as-QUBO formulation that this work adapts to mammography segmentation."}],"review_version":1}